Makefile

@@ -86,13 +86,19 @@ proxy-image:
 
 # The whole gate. Raises a database, runs everything against it, and takes it down again --
 # including when the tests fail, which is why the teardown is not conditional.
+#
+# **One package at a time (-p 1), and it is not about speed.** The live tests reach one bus, and on
+# it they assert, read and remove the mesh's own objects -- streams and consumers with fixed names,
+# because those names are the mesh's and a test cannot choose others. Two packages doing that at once
+# is one deleting a consumer the other is reading through, and the failure lands in whichever test
+# was reading, as "no response from stream". That reads as a bug in the code under test.
 check: fmt vet postgres
-	@go test ./... ; status=$$? ; $(MAKE) postgres-stop ; exit $$status
+	@go test -p 1 ./... ; status=$$? ; $(MAKE) postgres-stop ; exit $$status
 
 # Without a database the live tests skip rather than fail, so this is the honest subset and not
-# the gate.
+# the gate. Serialised for the same reason check is: a bus may be configured even when a store is not.
 test:
-	go test ./...
+	go test -p 1 ./...
 
 vet:
 	go vet ./...

cmd/mesh-builder/main.go

@@ -29,10 +29,10 @@ import (
 	"os/signal"
 	"strings"
 	"syscall"
-	"time"
 
 	amqp "github.com/rabbitmq/amqp091-go"
 
+	"github.com/novox/mesh-controller/internal/broker"
 	"github.com/novox/mesh-controller/internal/builder"
 	"github.com/novox/mesh-controller/internal/link"
 )
@@ -115,72 +115,63 @@ func run() error {
 	ctx, stop := signal.NotifyContext(context.Background(), syscall.SIGINT, syscall.SIGTERM)
 	defer stop()
 
-	conn, err := dial(credential)
-	if err != nil {
-		// Not quoted back: the URL carries this builder's broker password.
-		return fmt.Errorf("cannot reach the broker: %w", err)
-	}
-	defer conn.Close()
-	channel, err := conn.Channel()
-	if err != nil {
-		return err
-	}
-	defer channel.Close()
-
-	if _, err := channel.QueueDeclare(link.BuildQueue, true, false, false, false, nil); err != nil {
-		return err
-	}
-	// One at a time. A build machine that took five requests at once would run five container
-	// builds against one runtime and finish all of them slower than it would have finished the
-	// first — and the queue is what shares work between machines, so nothing is lost by it.
-	if err := channel.Qos(1, 0, false); err != nil {
-		return err
-	}
-
-	// Not auto-acknowledged. A request acknowledged on arrival is a build that vanishes if this
-	// process dies mid-way, with nobody waiting on it ever hearing why.
-	requests, err := channel.ConsumeWithContext(ctx, link.BuildQueue, "mesh-builder",
-		false, false, false, false, nil)
+	machine, err := takeWorkFrom(credential, on)
 	if err != nil {
 		return err
 	}
+	defer machine.Close()
 
 	fmt.Fprintf(os.Stderr, "building for the mesh, publishing to %s\n", registry)
 	publisher := builder.Registry{Address: registry, Run: builder.Command}
 
-	for {
-		select {
-		case <-ctx.Done():
-			fmt.Println("stopping")
-			return nil
-		case delivery, ok := <-requests:
-			if !ok {
-				return fmt.Errorf("the broker closed the connection")
-			}
-			answer(ctx, channel, publisher, on, workspace, delivery)
-		}
+	return machine.Take(ctx, func(ctx context.Context, work link.Build) {
+		answer(ctx, publisher, on, workspace, work)
+	})
+}
+
+// takeWorkFrom opens this machine's link to whichever bus the mesh is on.
+//
+// **One place chooses**, as everywhere else the bus change went (novox/hq ADR 0116 step 5): a build
+// machine told about both would take work from one and answer on the other, and every log line would
+// say it was fine.
+func takeWorkFrom(credential Credential, on string) (link.BuildMachine, error) {
+	address, onNATS, err := broker.OnNATS()
+	if err != nil {
+		return nil, err
 	}
+	if err := broker.MustBeOneBus(credential.URL, address); err != nil {
+		return nil, err
+	}
+	if onNATS {
+		js, err := broker.Dial(address)
+		if err != nil {
+			return nil, fmt.Errorf("cannot reach the bus at %s: %w", address, err)
+		}
+		return link.MachineOverNATS(js, on), nil
+	}
+
+	conn, err := dial(credential)
+	if err != nil {
+		// Not quoted back: the URL carries this builder's broker password.
+		return nil, fmt.Errorf("cannot reach the broker: %w", err)
+	}
+	channel, err := conn.Channel()
+	if err != nil {
+		conn.Close()
+		return nil, err
+	}
+	return link.MachineOverCurrent(conn, channel, on), nil
 }
 
 // answer does one build and says what happened, whichever way it went.
-func answer(ctx context.Context, channel *amqp.Channel, publisher builder.Publisher,
-	on, workspace string, delivery amqp.Delivery) {
+func answer(ctx context.Context, publisher builder.Publisher, on, workspace string, work link.Build) {
+	request := work.Request()
 
-	// **First thing, and to stdout.** A build request that arrives and produces no visible line
-	// until it either finishes or fails is indistinguishable from one that never arrived — which
-	// cost a long diagnosis against a running mesh, chasing "the handler never fired" when the
-	// truth was only that the handler said nothing until the end.
-	fmt.Fprintf(os.Stderr, "a build request arrived (%d bytes)\n", len(delivery.Body))
-
-	var request link.BuildRequest
-	if err := json.Unmarshal(delivery.Body, &request); err != nil {
-		// Unreadable. Acknowledged and dropped rather than requeued: a message this builder
-		// cannot parse will not become parseable by being delivered again, and requeueing it
-		// would put it in front of every real request for ever.
-		fmt.Fprintf(os.Stderr, "a request could not be read and was dropped: %v\n", err)
-		_ = delivery.Ack(false)
-		return
-	}
+	// **First thing, and to stdout.** A build request that arrives and produces no visible line until
+	// it either finishes or fails is indistinguishable from one that never arrived — which cost a long
+	// diagnosis against a running mesh, chasing "the handler never fired" when the truth was only that
+	// the handler said nothing until the end.
+	fmt.Fprintf(os.Stderr, "a build request arrived for %s\n", request.Repository)
 
 	result := link.BuildResult{
 		ID: request.ID, Repository: request.Repository, Path: request.Path,
@@ -199,8 +190,8 @@ func answer(ctx context.Context, channel *amqp.Channel, publisher builder.Publis
 	var built builder.Result
 	if err == nil {
 		// The package-registry credential is a build input, so it is resolved before the clone: a
-		// build that could not have resolved its dependencies is refused in front of the reason,
-		// not after a clone that then fails at npm ci.
+		// build that could not have resolved its dependencies is refused in front of the reason, not
+		// after a clone that then fails at npm ci.
 		built, err = builder.Build(ctx, builder.Command, publisher,
 			request.Repository, request.Path, request.Ref, workspace, request.Held, npmrc,
 			forgeFrom(),
@@ -230,67 +221,19 @@ func answer(ctx context.Context, channel *amqp.Channel, publisher builder.Publis
 		}
 	}
 
-	body, err := json.Marshal(result)
-	if err != nil {
-		fmt.Fprintf(os.Stderr, "cannot report a build: %v\n", err)
-		_ = delivery.Ack(false)
+	if err := work.Announce(ctx, result); err != nil {
+		// Said, not fatal: the build happened. A build reported as failed because announcing it
+		// failed is a lie about work that was done — and the request stays unsettled below only if
+		// nothing was said at all, so another machine can try.
+		fmt.Fprintf(os.Stderr, "cannot say what came of a build: %v\n", err)
 		return
 	}
 
-	// Always through the exchange, whether or not somebody is waiting.
-	//
-	// **Never the default exchange.** Permission there is granted per exchange rather than per
-	// queue, so a builder allowed to use it could publish into any node's queue — the privilege a
-	// build machine most obviously should not have. An asker binds its own reply queue to this
-	// key and filters by correlation; a control plane that records builds is bound to it too, so
-	// a result nobody asked for is still kept rather than reported into the void.
-	publishCtx, cancel := context.WithTimeout(ctx, 30*time.Second)
-	defer cancel()
-	if err := channel.PublishWithContext(publishCtx, link.Exchange, link.KeyBuilt, false, false,
-		amqp.Publishing{
-			ContentType:   "application/json",
-			CorrelationId: result.ID,
-			Body:          body,
-		}); err != nil {
-		fmt.Fprintf(os.Stderr, "cannot answer a build request: %v\n", err)
+	// Settled only once the outcome is away, so a machine that dies before answering leaves the work
+	// for another rather than losing it.
+	if err := work.Done(); err != nil {
+		fmt.Fprintf(os.Stderr, "the outcome is away and the request could not be settled: %v\n", err)
 	}
-	// **And announced, which is a different act from answering.** The reply goes to whoever asked
-	// and is correlated to their request; this says to the whole mesh that a module now exists at
-	// a commit, and the catalogue places it in the module graph (novox/hq ADR 0072). A build
-	// nobody asked for still has to be announced, or the graph knows less than the registry does.
-	//
-	// Only on success: a failed build produced no module-version, and announcing one would put
-	// something in the graph that was never made.
-	if result.Failed == "" && result.Commit != "" {
-		announced := map[string]any{
-			"module": moduleOf(result.Manifest), "commit": result.Commit,
-			"repository": result.Repository, "path": result.Path, "ref": result.Ref,
-			"manifest": json.RawMessage(result.Manifest), "against": result.Against,
-			"made": result.Made,
-		}
-		if err := link.EmitEvent(publishCtx, channel, link.KeyModuleBuilt, "builder", on, announced); err != nil {
-			// Said, not fatal: the build happened and was answered. A module the catalogue has not
-			// heard of is a gap somebody can close; a build reported as failed because announcing
-			// it failed is a lie about work that was done.
-			fmt.Fprintf(os.Stderr, "  built, but could not announce it: %v\n", err)
-		}
-	}
-
-	// Acknowledged only once the answer is away, so a builder that dies before answering leaves
-	// the request for another machine rather than losing it.
-	_ = delivery.Ack(false)
-}
-
-// moduleOf reads the module's name out of the manifest it just built, which is the only place it is
-// authoritative — the request named a repository and a path, not a module.
-func moduleOf(manifest json.RawMessage) string {
-	var named struct {
-		Module string `json:"module"`
-	}
-	if err := json.Unmarshal(manifest, &named); err != nil {
-		return ""
-	}
-	return named.Module
 }
 
 // packagesFrom is where a build resolves the mesh's own published packages — the SDK above all

cmd/mesh-controller/build.go

@@ -395,7 +395,13 @@ func buildOne(ctx context.Context, source buildSource, path, ref string, wait ti
 	}
 	fmt.Println()
 
-	result, err := link.RequestBuild(ctx, server.Channel(), request, wait)
+	ask, err := askOver(server)
+	if err != nil {
+		return err
+	}
+	defer ask.Close()
+
+	result, err := ask.Submit(ctx, request, wait)
 	if err != nil {
 		return err
 	}
@@ -472,7 +478,13 @@ func buildAndShow(ctx context.Context, source buildSource, path, ref string, wai
 	}
 	defer server.Close()
 
-	result, err := link.RequestBuild(ctx, server.Channel(), link.BuildRequest{
+	ask, err := askOver(server)
+	if err != nil {
+		return err
+	}
+	defer ask.Close()
+
+	result, err := ask.Submit(ctx, link.BuildRequest{
 		ID:         fmt.Sprintf("%s-%d", "build", time.Now().UnixNano()),
 		Repository: repository, Path: path, Ref: ref,
 		Held: heldBy(ctx),
@@ -557,3 +569,22 @@ func heldBy(ctx context.Context) map[string]string {
 	}
 	return routed
 }
+
+// askOver opens the way a build is asked for, on whichever bus the mesh is on.
+//
+// **One place chooses**, as everywhere else the bus change went (novox/hq ADR 0116 step 5). On the bus
+// the mesh runs on today this needs the controller's own connection, so it is handed one; on the bus
+// being built it dials, because a build request is a one-shot and holds nothing else.
+func askOver(server *link.Server) (link.Builders, error) {
+	address, onNATS, err := broker.OnNATS()
+	if err != nil {
+		return nil, err
+	}
+	if err := broker.MustBeOneBus(os.Getenv(broker.AMQPVarName), address); err != nil {
+		return nil, err
+	}
+	if onNATS {
+		return link.BuildsOverNATS(address)
+	}
+	return link.BuildsOverCurrent(server.Channel()), nil
+}

cmd/mesh-controller/main.go

@@ -114,6 +114,8 @@ func run() error {
 		return planCommand(ctx, args[1:])
 	case "push":
 		return pushCommand(ctx, args[1:])
+	case "rollout":
+		return rolloutCommand(ctx, args[1:])
 	case "seats":
 		return seatsCommand(ctx, args[1:])
 	case "seat":

cmd/mesh-controller/modules.go

@@ -257,6 +257,21 @@ func moduleCommand(ctx context.Context, args []string) error {
 			return err
 		}
 
+		// Which bus this mesh is on. A module gets a credential for exactly one, and the two are
+		// made in entirely different ways: on the bus the mesh runs on today an account is a
+		// management call, and on the bus being built it is a row the next composition writes into
+		// the server's user list (novox/hq design 25 §4).
+		busAddress, onNATS, err := broker.OnNATS()
+		if err != nil {
+			return err
+		}
+		if err := broker.MustBeOneBus(os.Getenv(broker.AMQPVarName), busAddress); err != nil {
+			return err
+		}
+		if onNATS {
+			return issueOnTheNewBus(ctx, inv, m, *forNode, busAddress)
+		}
+
 		management, err := broker.ManagementFromEnvironment()
 		if err != nil {
 			return err
@@ -567,3 +582,77 @@ func mayIssue(m catalogue.Manifest) error {
 	}
 	return nil
 }
+
+// issueOnTheNewBus gives an assigned module its credential on the bus being built.
+//
+// **Three things differ from a management call, and each is the point of the move.** The credential
+// is minted into the mesh's records and becomes usable at the next composition, so there is no
+// server to be reachable for this to work. The password travels beside the address rather than inside
+// it, because the runtime's contract already separates them and a credential embedded in a URL is one
+// that leaks into every log line that prints a connection. And the module's durable consumer is
+// derived from what it declared rather than declared by name, so a module cannot ask for delivery of
+// something it did not say it consumes.
+func issueOnTheNewBus(ctx context.Context, inv *inventory.Inventory, m catalogue.Manifest,
+	node, busAddress string) error {
+
+	user := broker.Principal{Kind: broker.KindModule, Node: node, Module: m.Module}.Username()
+	password, err := inv.MintBusPassword(ctx, inventory.BusUser{
+		Username: user, Kind: inventory.BusModule, Node: node, Module: m.Module,
+	})
+	if err != nil {
+		return err
+	}
+
+	// Where the module is told to find the bus, and what certificate it must present. The same pair
+	// a node is told, for the same reason: a mesh's bus presents its own certificate, in no public
+	// trust store, so an address alone fails at TLS.
+	known, err := broker.FromEnvironment()
+	if err != nil {
+		return fmt.Errorf("cannot deliver a credential without knowing where the bus is: %w", err)
+	}
+	reachable, err := brokerReachableAt(ctx, inv, known, node)
+	if err != nil {
+		return err
+	}
+	held, err := json.Marshal(struct {
+		URL         string `json:"url"`
+		Fingerprint string `json:"fingerprint,omitempty"`
+		Node        string `json:"node"`
+		Module      string `json:"module"`
+		User        string `json:"user"`
+		Password    string `json:"password"`
+	}{
+		URL: "nats://" + reachable, Fingerprint: known.Fingerprint,
+		Node: node, Module: m.Module, User: user, Password: password,
+	})
+	if err != nil {
+		return err
+	}
+	if err := inv.AcceptSecretForModule(ctx, node, m.Module, "broker", string(held)); err != nil {
+		return err
+	}
+
+	// And how it hears what it consumes. Derived from its declaration, and only when it declared
+	// something: a module that consumes nothing needs no consumer, and creating one would be a
+	// durable subscription nobody reads.
+	if consumer, needed := broker.ConsumerFor(broker.Principal{
+		Kind: broker.KindModule, Node: node, Module: m.Module,
+		Emits: m.Emits, Consumes: m.Consumes, Serves: m.Tools,
+	}); needed {
+		js, err := broker.Dial(busAddress)
+		if err != nil {
+			return fmt.Errorf("the credential is minted and the mesh cannot reach the bus to create "+
+				"how %s hears what it consumes: %w", m.Module, err)
+		}
+		defer js.Close()
+		if err := js.EnsureConsumer(consumer); err != nil {
+			return err
+		}
+	}
+
+	fmt.Printf("bus user %s minted for %s, scoped to what it emits and consumes\n", user, m.Module)
+	fmt.Printf("  sealed to %s. It arrives with the next push — `push %s` to send it\n", node, node)
+	fmt.Printf("  and it works once the bus has been told: the user list is composed into the " +
+		"machine holding mesh-broker\n")
+	return nil
+}

cmd/mesh-controller/operator.go

@@ -2,12 +2,15 @@ package main
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"flag"
 	"fmt"
 	"os"
 	"strings"
 
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
 	"github.com/novox/mesh-controller/internal/secrets"
 )
 
@@ -26,9 +29,25 @@ import (
 //	operator key make [--out <file>]    make a keypair: private half to the file, public half printed
 //	operator key set <public>           tell the mesh which key to seal to
 //	operator key show                   the public key, its fingerprint, and what it can recover
-const operatorUsage = "operator key make [--out <file>] | operator key set <public> [--replace] | operator key show"
+const operatorUsage = "operator key make [--out <file>] | operator key set <public> [--replace] | " +
+	"operator key show | operator issue <name> --invokes <tool,tool|*> | operator revoke <name> | " +
+	"operator list"
 
 func operatorCommand(ctx context.Context, args []string) error {
+	if len(args) == 0 {
+		return errors.New(operatorUsage)
+	}
+	// The people who may reach the mesh's tools (design 25 §7). Beside the operator's key because
+	// both answer "who, other than a machine, may do something here" — and a person reading this
+	// command's usage is asking exactly that.
+	switch args[0] {
+	case "issue":
+		return personIssue(ctx, args[1:])
+	case "revoke":
+		return personRevoke(ctx, args[1:])
+	case "list":
+		return personList(ctx)
+	}
 	if len(args) < 2 || args[0] != "key" {
 		return errors.New(operatorUsage)
 	}
@@ -160,3 +179,125 @@ func readPrivateKey(path string) (string, error) {
 	}
 	return strings.TrimSpace(string(raw)), nil
 }
+
+// personIssue gives somebody a credential for the mesh's tools, and prints it once.
+//
+// **Printed, not stored.** The mesh keeps a hash and nothing else, so this is the only moment the
+// credential exists anywhere but on the workstation that will use it — the same contract a token has,
+// and for the same reason: a credential recoverable from the mesh's store has the store's blast
+// radius.
+func personIssue(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("operator issue", flag.ContinueOnError)
+	invokes := set.String("invokes", "", "the tools this person may call, comma-separated, or * for every one")
+	if err := set.Parse(args); err != nil {
+		return err
+	}
+	if set.NArg() != 1 {
+		return errors.New("operator issue <name> --invokes <tool,tool|*>")
+	}
+	name := set.Arg(0)
+	if *invokes == "" {
+		return errors.New(
+			"say what this person may call: --invokes mesh-catalog.catalog_tools,gitea.repo_create, " +
+				"or --invokes '*' for an administrator")
+	}
+	var tools []string
+	for _, t := range strings.Split(*invokes, ",") {
+		if t = strings.TrimSpace(t); t != "" {
+			tools = append(tools, t)
+		}
+	}
+
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	if err := inv.RecordPerson(ctx, inventory.Person{Name: name, Invokes: tools}); err != nil {
+		return err
+	}
+	// Refused here rather than at the next composition, where it would stop the whole file being
+	// written for everybody. A name that cannot be part of a subject is one the server would read as
+	// a wider permission than anybody granted.
+	if _, err := broker.PermissionsFor(broker.Principal{
+		Kind: broker.KindPerson, Module: name, Invokes: tools, PasswordHash: "x",
+	}); err != nil {
+		return err
+	}
+
+	user := broker.Principal{Kind: broker.KindPerson, Module: name}.Username()
+	password, err := inv.MintBusPassword(ctx, inventory.BusUser{Username: user, Kind: inventory.BusPerson})
+	if err != nil {
+		return err
+	}
+
+	where, err := broker.FromEnvironment()
+	if err != nil && !errors.Is(err, broker.ErrNotConfigured) {
+		return err
+	}
+	held, err := json.Marshal(struct {
+		URL         string   `json:"url"`
+		Fingerprint string   `json:"fingerprint,omitempty"`
+		User        string   `json:"user"`
+		Password    string   `json:"password"`
+		Person      string   `json:"person"`
+		Invokes     []string `json:"invokes"`
+	}{
+		URL: "nats://" + where.Address, Fingerprint: where.Fingerprint,
+		User: user, Password: password, Person: name, Invokes: tools,
+	})
+	if err != nil {
+		return err
+	}
+
+	fmt.Printf("issued %s, who may call %s\n", name, strings.Join(tools, ", "))
+	fmt.Println("  this is the only time the credential is printed; the mesh keeps a hash")
+	fmt.Println("  it works once the bus has been told, which is the next push to the machine holding mesh-broker")
+	fmt.Println()
+	fmt.Println(string(held))
+	return nil
+}
+
+func personRevoke(ctx context.Context, args []string) error {
+	if len(args) != 1 {
+		return errors.New("operator revoke <name>")
+	}
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	if err := open.inventory.ForgetPerson(ctx, args[0]); err != nil {
+		return err
+	}
+	// **Revoked at the next composition, not now.** The bus's users are a file, so a credential stops
+	// working when the file no longer names it. Said plainly, because "revoked" that still works for
+	// another minute is worth knowing about.
+	fmt.Printf("%s is forgotten, and their credential stops working at the next composition — "+
+		"push the machine holding mesh-broker to make it so\n", args[0])
+	return nil
+}
+
+func personList(ctx context.Context) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+
+	people, err := open.inventory.People(ctx)
+	if err != nil {
+		return err
+	}
+	if len(people) == 0 {
+		fmt.Println("nobody but machines reaches this mesh")
+		return nil
+	}
+	for _, p := range people {
+		fmt.Printf("%-20s %s\n", p.Name, strings.Join(p.Invokes, ", "))
+	}
+	return nil
+}

cmd/mesh-controller/plan.go

@@ -621,13 +621,20 @@ func renderingFor(ctx context.Context, open *stores, node string,
 	if err != nil {
 		return catalogue.Rendering{}, inventory.Node{}, err
 	}
+	// The bus's user list, for the machine that runs the bus. Composed per push rather than kept,
+	// because it is a function of the mesh's records and a kept copy could disagree with them.
+	busUsers, err := composeBusUsers(ctx, inv, plan.Modules)
+	if err != nil {
+		return catalogue.Rendering{}, inventory.Node{}, err
+	}
 	return catalogue.Rendering{
 		Settings: settings, Generators: gens, Grants: grants, Needed: needed, Ports: ports,
 		Certificate: certificate, Authority: authority, Mesh: private, Names: names,
 		Machines: machines,
-		Suffix:   overlay.Suffix(), MeshRange: meshRange, Accounts: accounts, Foundation: foundation, Kept: kept,
-		Adopted: record.Adopted,
-		Given:   given, Taken: taken, Seats: seats, ArtifactStore: artifactStore, Built: built,
+		Suffix:   overlay.Suffix(), MeshRange: meshRange, Accounts: accounts, Foundation: foundation,
+		Kept: kept, Adopted: record.Adopted,
+		Given: given, Taken: taken, Seats: seats, ArtifactStore: artifactStore, Built: built,
+		BusUsers: busUsers,
 	}, record, nil
 }
 
@@ -1187,6 +1194,76 @@ func portsOn(
 	return out, nil
 }
 
+// composeBusUsers is the bus's user list, for a push to the machine that runs the bus.
+//
+// Empty for every other machine, and for every machine while the mesh is on the bus it runs on
+// today — where accounts are a management call and there is no file to write.
+//
+// **Composed on each push, never kept.** The list is a function of the mesh's records (who exists,
+// what runs where, what each declares), and a stored copy would be a second account of who may reach
+// the bus, able to disagree with the records while both looked internally consistent (ADR 0043).
+//
+// A user the mesh has never minted a password for is **left out and said**, not written as a user
+// without one — the composer refuses that, because a user with no password is a user anybody is. That
+// is an ordinary situation with an obvious remedy (`module issue`, or enrolling), so the push carries
+// the rest rather than failing: a bus that is missing one module's user is a mesh where that module
+// cannot connect, and a bus with no file at all is a mesh where nothing can.
+func composeBusUsers(ctx context.Context, inv *inventory.Inventory,
+	onThisNode []catalogue.Manifest) (string, error) {
+
+	// **Not gated on which bus the controller is on, and that was a bug.** It read "compose this only
+	// once the mesh is on the new bus" — which cannot work, because the server needs its user list
+	// *before* anything moves onto it. Step 2 of the change is exactly that: the server stands in the
+	// mesh carrying nothing, on its own ports, while every node is still on the old bus (novox/hq
+	// ADR 0116). Under the old gating that step could not happen: the module would come up, find no
+	// accounts file, and its entrypoint would wait for one the controller had decided not to write.
+	//
+	// So the question is only whether this machine runs the module that asked for the file. A mesh
+	// that never moves has written a user list nothing reads, which costs a few hundred bytes on one
+	// node; the reverse cost a step that cannot be taken.
+	//
+	// Asked of what this push resolves to rather than of the seat's holder mesh-wide: the file is a
+	// resource of that module, so the question is whether it is here.
+	holdsTheBus := false
+	for _, m := range onThisNode {
+		if m.BusUsers != "" && m.ClaimsSeat("mesh-broker") {
+			holdsTheBus = true
+		}
+	}
+	if !holdsTheBus {
+		return "", nil
+	}
+
+	records, err := inv.BusRecords(ctx)
+	if err != nil {
+		return "", err
+	}
+	users, err := broker.Users(records)
+	if err != nil {
+		return "", err
+	}
+	kept, err := inv.BusUsers(ctx)
+	if err != nil {
+		return "", err
+	}
+	hashes := make(map[string]string, len(kept))
+	for name, u := range kept {
+		hashes[name] = u.PasswordHash
+	}
+	filled, missing := broker.WithPasswords(users, hashes)
+	if len(missing) > 0 {
+		fmt.Printf("the bus's user list leaves out %d user(s) the mesh has minted no credential "+
+			"for: %s. Each is a user that cannot connect until one is issued\n",
+			len(missing), strings.Join(missing, ", "))
+	}
+	if len(filled) == 0 {
+		return "", fmt.Errorf(
+			"this machine runs the bus and not one user has a credential, so the composed list " +
+				"would refuse every connection in the mesh")
+	}
+	return broker.ComposeAccounts(filled)
+}
+
 // foundationPortsFor is the broker port, kept only when a module resolved onto this node listens
 // on it (novox/hq issue: the broker opening leaked onto every node). The foundation opening
 // exists to WIDEN the broker's `from: mesh` port to from-anywhere, because a machine enrolling is

cmd/mesh-controller/push.go

@@ -77,12 +77,34 @@ func serve(ctx context.Context) error {
 			"reconnect. Set %s and %s.\n", broker.AddressVar, broker.CertificateVar)
 	}
 
-	work := link.Enrolment{Inventory: inv, Identity: ident, Management: management, Broker: known}
+	// **Which bus this mesh is on, read once** (novox/hq ADR 0116 step 5). Both clients ship; both
+	// being live is refused, because a mesh half on each is one where a declaration goes out on one
+	// and the report comes back on the other, and every component logs success while it happens.
+	busAddress, onNATS, err := broker.OnNATS()
+	if err != nil {
+		return err
+	}
+	if err := broker.MustBeOneBus(os.Getenv(broker.AMQPVarName), busAddress); err != nil {
+		return err
+	}
+
+	work := link.Enrolment{Inventory: inv, Identity: ident, Management: management, Broker: known,
+		OnNATS: onNATS}
 	server, err := link.Connect(work, work)
 	if err != nil {
 		return err
 	}
 	defer server.Close()
+
+	// The bus's own objects, asserted on every start. **Not created once at genesis**: a stream
+	// somebody deleted, a mesh raised from a restored backup, or a bus whose data directory was
+	// replaced all have records and no objects — and a node whose consumer is missing hears nothing
+	// while everything else about it looks correct.
+	if onNATS {
+		if err := raiseTheBus(ctx, inv, busAddress); err != nil {
+			return err
+		}
+	}
 	// And build results nobody was waiting for. A build triggered any other way than `build`
 	// would otherwise be reported into the void, which is the same as not reporting it.
 	server.Records(builds{inv})
@@ -142,7 +164,7 @@ func declare(ctx context.Context, args []string) error {
 	}
 	defer server.Close()
 
-	if err := link.Declare(ctx, server.Channel(), ident, node, raw, 15*time.Second); err != nil {
+	if err := link.Declare(ctx, link.OverCurrent{Channel: server.Channel()}, ident, node, raw, 15*time.Second); err != nil {
 		return err
 	}
 	fmt.Printf("sent %s a signed declaration (%d bytes)\n", node, len(raw))
@@ -310,7 +332,7 @@ func pushCommand(ctx context.Context, args []string) error {
 		if err != nil {
 			return err
 		}
-		if err := link.Declare(ctx, server.Channel(), ident, s.node, body, 15*time.Second); err != nil {
+		if err := link.Declare(ctx, link.OverCurrent{Channel: server.Channel()}, ident, s.node, body, 15*time.Second); err != nil {
 			return err
 		}
 		// After it is away, not before. A digest recorded for something that failed to send would
@@ -393,7 +415,7 @@ func pushCommand(ctx context.Context, args []string) error {
 					return declarationWith(held, open, node, plan, settings, gens, Allocating)
 				},
 				func(s readyNode, body []byte) error {
-					if err := link.Declare(ctx, server.Channel(), ident, s.node, body,
+					if err := link.Declare(ctx, link.OverCurrent{Channel: server.Channel()}, ident, s.node, body,
 						15*time.Second); err != nil {
 						return err
 					}
@@ -617,7 +639,7 @@ func sendTo(ctx context.Context, open *stores, names []string) error {
 		if err != nil {
 			return err
 		}
-		if err := link.Declare(ctx, server.Channel(), ident, s.node, body, 15*time.Second); err != nil {
+		if err := link.Declare(ctx, link.OverCurrent{Channel: server.Channel()}, ident, s.node, body, 15*time.Second); err != nil {
 			return err
 		}
 		record, err := inv.NodeByName(ctx, s.node)
@@ -670,3 +692,57 @@ func wouldSend(ctx context.Context, open *stores,
 	}
 	return out, nil
 }
+
+// raiseTheBus asserts the streams and consumers the mesh's own traffic needs.
+//
+// **Every start, and it says what it did.** The objects are the mesh's, created by nothing else —
+// the controller is their only writer (design 25 §3) — so a mesh that came up without them is one
+// where nodes connect, authenticate, and hear nothing. Said rather than silent for the reason the
+// first line of `serve` is said: a log that is quiet on success and loud on failure reads as broken
+// when it is working.
+func raiseTheBus(ctx context.Context, inv *inventory.Inventory, address string) error {
+	js, err := broker.Dial(address)
+	if err != nil {
+		return fmt.Errorf("the mesh is on the bus at %s and this control plane cannot reach it: %w",
+			address, err)
+	}
+	defer js.Close()
+
+	// **Its own user, before anything else.** The controller's account is created by the installer at
+	// a bootstrap password, before there is a controller to mint one — so nothing recorded a hash for
+	// it, and the first composition would leave the writer out of the file it was writing. Recorded
+	// only if absent: a credential the mesh minted since is the one that counts.
+	// **Its own user, before anything else it does here.** The controller's account is created by the
+	// installer at a bootstrap password, before there is a controller to mint one — so nothing
+	// recorded a hash for it, and the first composition would leave the writer out of the file it was
+	// writing: a bus nothing can connect to, produced by the thing connected to it. Recorded only if
+	// absent, so a restart cannot put the bootstrap credential back over a rotated one.
+	if user, password, _ := broker.CredentialIn(address); user != "" && password != "" {
+		if err := inv.SeedBusUser(ctx, inventory.BusUser{
+			Username: user, Kind: inventory.BusController,
+		}, password); err != nil {
+			return fmt.Errorf("cannot record the credential this control plane is using: %w", err)
+		}
+	}
+
+	nodes, err := inv.Nodes(ctx)
+	if err != nil {
+		return err
+	}
+	names := make([]string, 0, len(nodes))
+	for _, n := range nodes {
+		names = append(names, n.Name)
+	}
+	if err := broker.Raise(js, names); err != nil {
+		return err
+	}
+	// The work queues of the mesh's own roles (novox/hq ADR 0121). The queue before the holder,
+	// deliberately: work queues until somebody arrives to do it, so assigning a build machine a week
+	// after something started asking for builds flushes the backlog instead of having lost it.
+	if err := broker.RaiseSeats(js, inventory.MeshSeats(), nil); err != nil {
+		return err
+	}
+	fmt.Printf("the bus at %s has its streams, and %d machine(s) can hear a declaration\n",
+		address, len(names))
+	return nil
+}

cmd/mesh-controller/rollout.go

@@ -0,0 +1,193 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Moving the mesh's own traffic to the bus being built (novox/hq ADR 0116 step 5).
+//
+// **The whole mesh moves at once, so there is nothing to inspect afterwards.** Every seam ships both
+// transports and every one of them chooses by a single fact; this is the step that flips it. That
+// shape is deliberate — steps 1 to 4 leave every node where it is, so the cost of being wrong stays
+// bounded until here — and it means the useful work is almost all in the checking.
+//
+// So `rollout check` is the command that matters and the one that can be run any number of times
+// against a mesh that is serving. It answers from records: what is missing, and what would happen.
+// `rollout` itself refuses unless the check is clean.
+//
+// **The old broker is not switched off by this.** It stays an ordinary provider of `amqp` for whatever
+// else uses it — on this installation, a whole automation layer that has nothing to do with the mesh
+// ([ADR 0119](../../02-DECISIONS/0119-amqp-is-a-provision-not-the-bus.md)). Only the mesh's own
+// traffic moves, which is why this is survivable at all: what breaks if it goes wrong is the mesh's
+// ability to change things, not the services its modules are serving.
+
+const rolloutUsage = "rollout check | rollout --confirm"
+
+func rolloutCommand(ctx context.Context, args []string) error {
+	switch {
+	case len(args) == 1 && args[0] == "check":
+		return rolloutCheck(ctx)
+	case len(args) == 1 && args[0] == "--confirm":
+		return errors.New(
+			"the rollout itself is not built yet: `rollout check` answers whether it could run, and " +
+				"what is missing. Moving every node at once is the one step with nothing to inspect " +
+				"afterwards, so it is not being written before the check it depends on has been run " +
+				"against a real mesh")
+	default:
+		return errors.New(rolloutUsage)
+	}
+}
+
+// rolloutCheck says whether the mesh could move, and what would happen if it did.
+func rolloutCheck(ctx context.Context) error {
+	open, err := openStores(ctx)
+	if err != nil {
+		return err
+	}
+	defer open.Close()
+	inv := open.inventory
+
+	state, err := readinessOf(ctx, inv)
+	if err != nil {
+		return err
+	}
+
+	fmt.Println("the bus this mesh would move to")
+	if state.TheBus == "" {
+		fmt.Printf("  nothing names one (%s is unset)\n", broker.NATSVar)
+	} else {
+		standing := "not answering"
+		if state.ServerStanding {
+			standing = "answering"
+		}
+		fmt.Printf("  %s — %s\n", state.TheBus, standing)
+	}
+	fmt.Println()
+
+	fmt.Println("what would move")
+	for _, step := range broker.WhatMoves(state) {
+		fmt.Printf("  %s\n", step)
+	}
+	fmt.Println()
+
+	why := notReadyOf(state)
+	if len(why) == 0 {
+		fmt.Println("nothing is missing: this mesh could move its bus.")
+		fmt.Println()
+		fmt.Println("Read `what would move` above once more before running it. Every node moves at the")
+		fmt.Println("same moment and there is no half-moved state to look at afterwards.")
+		return nil
+	}
+	fmt.Printf("not ready — %d thing(s) to do first:\n", len(why))
+	for i, w := range why {
+		fmt.Printf("  %d. %s\n", i+1, w)
+	}
+	return nil
+}
+
+// readinessOf gathers what the mesh knows about its own ability to move.
+//
+// Reads and one dial, and nothing is written. Safe to run on a mesh that is serving, which is the
+// point: the answer is only useful if it can be had without committing to anything.
+func readinessOf(ctx context.Context, inv *inventory.Inventory) (broker.Readiness, error) {
+	state := broker.Readiness{
+		Credentialled:       map[string]bool{},
+		ModuleCredentialled: map[string]bool{},
+		// The old broker keeps its other clients on this installation, and saying so is how the plan
+		// stops reading as a retirement.
+		OldBusHasOtherClients: true,
+	}
+
+	address, _, err := broker.OnNATS()
+	if err != nil {
+		return state, err
+	}
+	state.TheBus = address
+	if address != "" {
+		// One dial, briefly. "Is it answering" is the one fact records cannot hold, and a mesh about
+		// to move onto a server that is not there should hear it here rather than afterwards.
+		if conn, err := nats.Connect(broker.BareAddress(address), nats.Timeout(5*time.Second)); err == nil {
+			state.ServerStanding = true
+			conn.Close()
+		}
+	}
+
+	nodes, err := inv.Nodes(ctx)
+	if err != nil {
+		return state, err
+	}
+	kept, err := inv.BusUsers(ctx)
+	if err != nil {
+		return state, err
+	}
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		return state, err
+	}
+
+	for _, n := range nodes {
+		state.Nodes = append(state.Nodes, n.Name)
+		_, has := kept[broker.Principal{Kind: broker.KindNode, Node: n.Name}.Username()]
+		state.Credentialled[n.Name] = has
+
+		assigned, err := inv.Assigned(ctx, n.Name)
+		if err != nil {
+			return state, err
+		}
+		for _, module := range assigned {
+			m, known := shelf[module]
+			if !known {
+				continue
+			}
+			// The machine that holds the bus seat is the one that would be sent the user list.
+			if m.BusUsers != "" && m.ClaimsSeat("mesh-broker") {
+				state.Holder = n.Name
+				state.AccountsComposed = wasSentTheUserList(ctx, inv, n.Name)
+			}
+			// A module that never speaks needs no credential, so it is not counted as missing one.
+			if !speaksOnTheBus(m) {
+				continue
+			}
+			named := n.Name + "/" + module
+			state.Modules = append(state.Modules, named)
+			_, hasOne := kept[broker.Principal{
+				Kind: broker.KindModule, Node: n.Name, Module: module,
+			}.Username()]
+			state.ModuleCredentialled[named] = hasOne
+		}
+	}
+	return state, nil
+}
+
+// speaksOnTheBus says whether a module reaches the bus at all.
+//
+// A third of the catalogue never does (novox/hq ADR 0120), and counting those as missing a credential
+// would bury the ones that matter under a list nobody can act on.
+func speaksOnTheBus(m catalogue.Manifest) bool {
+	return len(m.Emits) > 0 || len(m.Consumes) > 0 || len(m.Tools) > 0 ||
+		len(m.DefinesSeats) > 0 || len(m.Uses) > 0 || len(m.Claims) > 0
+}
+
+// wasSentTheUserList says whether the machine holding the bus has had a declaration since the user
+// list became part of one.
+//
+// Read from what the mesh recorded sending rather than asked of the machine: a machine that is away
+// has still been sent it, and this question is about whether the mesh did its part.
+func wasSentTheUserList(ctx context.Context, inv *inventory.Inventory, node string) bool {
+	digest, err := inv.Outstanding(ctx, node)
+	return err == nil && strings.TrimSpace(digest) != ""
+}
+
+// notReadyOf is the readiness reasoning, named here so a test can reach it without the command's
+// printing. The reasoning itself is the broker package's, where it is pure.
+func notReadyOf(state broker.Readiness) []string { return broker.NotReady(state) }

cmd/mesh-controller/rollout_test.go

@@ -0,0 +1,93 @@
+package main
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Whether a mesh could move its bus, read from a real store.
+//
+// The readiness reasoning has its own tests; this is about the gathering — that the question is
+// answered from what the mesh actually holds, on a store with machines and modules in it, without
+// writing anything.
+
+func TestReadinessIsGatheredFromWhatTheMeshHolds(t *testing.T) {
+	inv := inventory.ForTest(t)
+	ctx := context.Background()
+
+	// A mesh mid-change: two machines, the bus module on one of them, a module that speaks and a
+	// module that never does.
+	for _, m := range []catalogue.Manifest{
+		{Module: "nats", Version: "1", BusUsers: "/var/lib/nats-module/conf/accounts.conf",
+			Claims: []catalogue.Claim{{Name: "mesh-broker", Scope: catalogue.ScopeMesh}}},
+		{Module: "gitea", Version: "1", Tools: []string{"repo_create"}},
+		{Module: "wallpaper", Version: "1"},
+	} {
+		if err := inv.RegisterModule(ctx, m, inventory.Source{Repository: "/r"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	for _, n := range []string{"anchor", "laptop"} {
+		if _, err := inv.AddNode(ctx, n); err != nil {
+			t.Fatal(err)
+		}
+	}
+	for _, a := range [][2]string{{"anchor", "nats"}, {"anchor", "gitea"}, {"laptop", "wallpaper"}} {
+		if _, err := inv.Assign(ctx, a[0], a[1]); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	state, err := readinessOf(ctx, inv)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	if state.Holder != "anchor" {
+		t.Errorf("the machine holding the bus reads as %q", state.Holder)
+	}
+	if len(state.Nodes) != 2 {
+		t.Errorf("machines read as %v", state.Nodes)
+	}
+	// **A module that never speaks is not counted as missing a credential.** A third of the catalogue
+	// never reaches the bus, and listing those would bury the ones that matter.
+	for _, m := range state.Modules {
+		if strings.HasSuffix(m, "/wallpaper") {
+			t.Errorf("a module that never speaks was counted: %v", state.Modules)
+		}
+	}
+	if len(state.Modules) != 2 {
+		t.Errorf("modules that speak read as %v; expected the bus module and the one with a tool",
+			state.Modules)
+	}
+
+	// Nothing has been minted, so it is not ready — and it says so about each machine by name.
+	why := notReadyOf(state)
+	if len(why) == 0 {
+		t.Fatal("a mesh where nothing has a credential was reported ready to move")
+	}
+	said := strings.Join(why, "\n")
+	for _, name := range []string{"anchor", "laptop"} {
+		if !strings.Contains(said, name) {
+			t.Errorf("the refusal does not name %s: %s", name, said)
+		}
+	}
+
+	// Mint for one machine and it drops out of the complaint, which is how somebody works through it.
+	if _, err := inv.MintBusPassword(ctx, inventory.BusUser{
+		Username: "node.laptop", Kind: inventory.BusNode, Node: "laptop",
+	}); err != nil {
+		t.Fatal(err)
+	}
+	state, err = readinessOf(ctx, inv)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !state.Credentialled["laptop"] {
+		t.Error("a machine that was minted a credential still reads as having none")
+	}
+}

go.mod

@@ -1,19 +1,23 @@
 module github.com/novox/mesh-controller
 
-go 1.25.0
+go 1.26.0
 
 require (
 	github.com/jackc/pgx/v5 v5.10.0
 	github.com/rabbitmq/amqp091-go v1.14.0
-	golang.org/x/crypto v0.55.0
+	golang.org/x/crypto v0.57.0
 )
 
 require (
 	github.com/jackc/pgpassfile v1.0.0 // indirect
 	github.com/jackc/pgservicefile v0.0.0-20240606120523-5a60cdf6a761 // indirect
 	github.com/jackc/puddle/v2 v2.2.2 // indirect
-	golang.org/x/net v0.57.0 // indirect
-	golang.org/x/sync v0.22.0 // indirect
-	golang.org/x/sys v0.47.0 // indirect
-	golang.org/x/text v0.41.0 // indirect
+	github.com/klauspost/compress v1.20.0 // indirect
+	github.com/nats-io/nats.go v1.54.0 // indirect
+	github.com/nats-io/nkeys v0.4.16 // indirect
+	github.com/nats-io/nuid v1.0.1 // indirect
+	golang.org/x/net v0.58.0 // indirect
+	golang.org/x/sync v0.23.0 // indirect
+	golang.org/x/sys v0.48.0 // indirect
+	golang.org/x/text v0.42.0 // indirect
 )

go.sum

@@ -9,6 +9,14 @@ github.com/jackc/pgx/v5 v5.10.0 h1:VhSvgU2jSli8o3AqIEOTJr7rZwAEUVo4E4XhR94Zfr0=
 github.com/jackc/pgx/v5 v5.10.0/go.mod h1:mal1tBGAFfLHvZzaYh77YS/eC6IX9OWbRV1QIIM0Jn4=
 github.com/jackc/puddle/v2 v2.2.2 h1:PR8nw+E/1w0GLuRFSmiioY6UooMp6KJv0/61nB7icHo=
 github.com/jackc/puddle/v2 v2.2.2/go.mod h1:vriiEXHvEE654aYKXXjOvZM39qJ0q+azkZFrfEOc3H4=
+github.com/klauspost/compress v1.20.0 h1:a3C1ke2ohxFymNlb2HWAHjDeKCI90scRskErZkR0ezA=
+github.com/klauspost/compress v1.20.0/go.mod h1:LUdAzn7YLVvxLpc7y3V1m40wESHTgc1422pwwBSKYuI=
+github.com/nats-io/nats.go v1.54.0 h1:vsXoOxjHp/GmPUN+EcI7uOf/uB+iAP+kEsAFNQN0yzA=
+github.com/nats-io/nats.go v1.54.0/go.mod h1:y+DZoD1oBOYfZTU681eTUiUjI0vbqYGixNVFHcjHJ0k=
+github.com/nats-io/nkeys v0.4.16 h1:rd5oAuLOb8mnAycB0xleuEBNS1pVVnN0fv/FF34Eypg=
+github.com/nats-io/nkeys v0.4.16/go.mod h1:llLgWoI0o4z/Q57q2R1kHfmocyhGV6VG/U18Glg1Afs=
+github.com/nats-io/nuid v1.0.1 h1:5iA8DT8V7q8WK2EScv2padNa/rTESc1KdnPw4TC2paw=
+github.com/nats-io/nuid v1.0.1/go.mod h1:19wcPz3Ph3q0Jbyiqsd0kePYG7A95tJPxeL+1OSON2c=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/rabbitmq/amqp091-go v1.14.0 h1:RSaT7aOKt/OrkVUyswPDW29lnRz9psuGmfZFBmLqLek=
@@ -22,14 +30,24 @@ go.uber.org/goleak v1.3.0 h1:2K3zAYmnTNqV73imy9J1T3WC+gmCePx2hEGkimedGto=
 go.uber.org/goleak v1.3.0/go.mod h1:CoHD4mav9JJNrW/WLlf7HGZPjdw8EucARQHekz1X6bE=
 golang.org/x/crypto v0.55.0 h1:+KWHjbgOaAQ66dh/YlkZKHlz9ZUlq61AFirAR9ntP8M=
 golang.org/x/crypto v0.55.0/go.mod h1:uq0V9dE/fzQuJtbnL+2EhWOE63vo164FY8xqEnV9xis=
+golang.org/x/crypto v0.57.0 h1:3ZVCjf8Ggz7zneR/EHRVx68Ctf+2pmIMP2UFhh9cC6M=
+golang.org/x/crypto v0.57.0/go.mod h1:Fdz0i5U6CoizGwLda9DttjSk6qlZo25zYNtR+ycvuZA=
 golang.org/x/net v0.57.0 h1:K5+3DljvIuDG9/Jv9rvyMywYNFCQ9RSUY6OOTTkT+tE=
 golang.org/x/net v0.57.0/go.mod h1:KpXc8iv+r3XplLAG/f7Jsf9RPszJzdR0f58q9vGOuEU=
+golang.org/x/net v0.58.0 h1:ynWG7rqYi4ccpTEuPZ2QGWHktVEM9DMCj9yzDE0Q7To=
+golang.org/x/net v0.58.0/go.mod h1:YwCddHnFlT7eLQqVprV19OnhLGtc5xOKgE0RyqgfWAU=
 golang.org/x/sync v0.22.0 h1:SZjpbeLmrCk4xhRSZFNZW5gFUeCeFgjekvI/+gfScek=
 golang.org/x/sync v0.22.0/go.mod h1:9xrNwdLfx4jkKbNva9FpL6vEN7evnE43NNNJQ2LF3+0=
+golang.org/x/sync v0.23.0 h1:KameEIfc1IkluZyXWLn39Wd4tURc6GbCiISGiZm2bQk=
+golang.org/x/sync v0.23.0/go.mod h1:sUUOizhqBxiL6pEWpqNLUiaJn1ShEbZ6BBqskPbjZm0=
 golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
 golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/sys v0.48.0 h1:bbX/i/6MgT9BVLM9RT1thmxL04yeTAhbEz4SyadbXoo=
+golang.org/x/sys v0.48.0/go.mod h1:hNLxWAXmnKAxqDtdwIYC4bM9oQPEecfsnNMuSxOs3og=
 golang.org/x/text v0.41.0 h1:vz/seA0lnX87Othu2f/0L24RcgrXD9/YFTSuGjj3rH8=
 golang.org/x/text v0.41.0/go.mod h1:jvf1O8ajNzZqhSrQBPbutR/EB83Cc0CFrezNQIwbb5M=
+golang.org/x/text v0.42.0 h1:JbOZXgfeCPU9gacVtYliJqOhD+zhrEqK4LfdpmlUZqI=
+golang.org/x/text v0.42.0/go.mod h1:ojzP1Z+2QtioaF8DTtO8K5q7JWVVYwZKenzujK0Zd0E=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=

internal/broker/agreement.go

@@ -0,0 +1,125 @@
+package broker
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// Do the emitters and the consumers of a catalogue agree?
+//
+// **The check that was missing** (novox/hq 04-ISSUES/127). Every manifest was individually
+// well-formed and every derivation individually correct, and no cross-module subscription in the
+// mesh matched anything: a consumer's declaration derived into a namespace nobody publishes to.
+// Nothing failed, because a subscription that matches nothing is not an error — it is silence.
+//
+// The comparison has to be over the whole catalogue, because the two halves live in different
+// manifests, and it cannot simply demand that every consumed event have a live emitter: a module
+// may be installed long before the one whose events it wants. So the rule is narrower and still
+// catches this: **where the emitter is present, it must emit what the consumer asked for.**
+
+// AConsumer is one module's interest in another's events, as this check needs it.
+type AConsumer struct {
+	Module   string
+	Consumes []string
+}
+
+// AnEmitter is one module's events.
+type AnEmitter struct {
+	Module string
+	Emits  []string
+}
+
+// Disagreements are the consumed events whose emitter is in the catalogue and does not emit them.
+//
+// Returned as sentences rather than as structs: every one of them is read by a person deciding
+// whether a manifest or a catalogue is wrong, and a pair of names without the reason is a puzzle.
+func Disagreements(emitters []AnEmitter, consumers []AConsumer, seats []DeclaredSeat) []string {
+	emits := map[string]map[string]bool{}
+	for _, e := range emitters {
+		if emits[e.Module] == nil {
+			emits[e.Module] = map[string]bool{}
+		}
+		for _, name := range e.Emits {
+			emits[e.Module][name] = true
+		}
+	}
+	// A seat's events are published by its holder under the seat's name, so a consumer naming the
+	// seat is naming something real even though no module declares it as its own.
+	for _, s := range seats {
+		if len(s.Emits) == 0 {
+			continue
+		}
+		if emits[s.Name] == nil {
+			emits[s.Name] = map[string]bool{}
+		}
+		for _, name := range s.Emits {
+			emits[s.Name][name] = true
+		}
+	}
+
+	var out []string
+	for _, c := range consumers {
+		for _, pattern := range c.Consumes {
+			emitter, event, named := strings.Cut(pattern, ".")
+			// Every event from everyone, or every event from one module: both are deliberate and
+			// neither names a particular event to check.
+			if !named || emitter == "*" || emitter == catalogueTheRest || event == catalogueTheRest {
+				continue
+			}
+			known, present := emits[emitter]
+			if !present {
+				// Not installed here, which is ordinary: a module lives in its own repository and
+				// may be registered later. Nothing to compare, so nothing to say.
+				continue
+			}
+			if matchesAny(event, known) {
+				continue
+			}
+			out = append(out, fmt.Sprintf(
+				"%s consumes %q and %s emits %s — so that subscription would match nothing, and "+
+					"nothing would report it",
+				c.Module, pattern, emitter, listOf(known)))
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// matchesAny says whether one of an emitter's event names satisfies a consumer's pattern.
+func matchesAny(pattern string, emitted map[string]bool) bool {
+	want := strings.Split(pattern, ".")
+	for name := range emitted {
+		if matches(want, strings.Split(name, ".")) {
+			return true
+		}
+	}
+	return false
+}
+
+func matches(pattern, name []string) bool {
+	for i, part := range pattern {
+		if part == catalogueTheRest {
+			return i < len(name)
+		}
+		if i >= len(name) {
+			return false
+		}
+		if part != "*" && part != name[i] {
+			return false
+		}
+	}
+	return len(pattern) == len(name)
+}
+
+func listOf(names map[string]bool) string {
+	if len(names) == 0 {
+		return "nothing"
+	}
+	out := make([]string, 0, len(names))
+	for n := range names {
+		out = append(out, n)
+	}
+	sort.Strings(out)
+	return strings.Join(out, ", ")
+}

internal/broker/agreement_catalogue_test.go

@@ -0,0 +1,236 @@
+package broker
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// **Do the catalogue's emitters and consumers agree?**
+//
+// This is the check whose absence let issue 127 stand: every manifest was individually well-formed,
+// every derivation individually correct, and no cross-module subscription in the mesh matched
+// anything. A subscription that matches nothing is not an error — it is silence — so nothing
+// anywhere reported it.
+//
+// It compares what one manifest asks to hear against what another says it emits. It cannot demand
+// that every consumed event have a live emitter, because a module lives in its own repository and
+// may be registered long before the one whose events it wants. Where the emitter *is* here, it must
+// emit what the consumer asked for.
+func TestTheCataloguesEmittersAndConsumersAgree(t *testing.T) {
+	emitters, consumers, seats := theCataloguesEvents(t)
+
+	if bad := Disagreements(emitters, consumers, seats); len(bad) > 0 {
+		t.Fatalf("%d subscription(s) in the catalogue would match nothing:\n  %s",
+			len(bad), strings.Join(bad, "\n  "))
+	}
+}
+
+// And the check itself catches the thing it exists for, so it cannot pass by doing nothing.
+func TestTheAgreementCheckCatchesASubscriptionThatMatchesNothing(t *testing.T) {
+	bad := Disagreements(
+		[]AnEmitter{{Module: "builder", Emits: []string{"built"}}},
+		[]AConsumer{{Module: "mesh-catalog", Consumes: []string{"builder.finished"}}},
+		nil)
+	if len(bad) != 1 {
+		t.Fatalf("a consumer asking for an event its emitter does not emit was not caught: %v", bad)
+	}
+	if !strings.Contains(bad[0], "builder.finished") || !strings.Contains(bad[0], "built") {
+		t.Fatalf("the report names neither what was asked for nor what is emitted: %s", bad[0])
+	}
+
+	// A module that is not here is not a disagreement: it may be registered later.
+	if bad := Disagreements(nil,
+		[]AConsumer{{Module: "plex", Consumes: []string{"sonarr.download.completed"}}}, nil); len(bad) != 0 {
+		t.Fatalf("a consumer whose emitter is not installed was reported: %v", bad)
+	}
+
+	// A wildcard over emitters is deliberate and names no particular event to check.
+	if bad := Disagreements([]AnEmitter{{Module: "sonarr", Emits: []string{"download.completed"}}},
+		[]AConsumer{{Module: "plex", Consumes: []string{"*.download.completed"}}}, nil); len(bad) != 0 {
+		t.Fatalf("a wildcard over emitters was reported: %v", bad)
+	}
+
+	// A consumer of a role's event whose role does not emit it is caught, which is what stops the
+	// catalogue check above from passing by knowing nothing about roles.
+	if bad := Disagreements(nil,
+		[]AConsumer{{Module: "mesh-catalog", Consumes: []string{"mesh-build-machine.finished"}}},
+		[]DeclaredSeat{{Name: "mesh-build-machine", Emits: []string{"built"}}}); len(bad) != 1 {
+		t.Fatalf("a consumer of a role event the role does not emit was not caught: %v", bad)
+	}
+
+	// An event published under a seat's name is real even though no module declares it as its own.
+	if bad := Disagreements(nil,
+		[]AConsumer{{Module: "watcher", Consumes: []string{"mesh-artifact-store.image.pushed"}}},
+		[]DeclaredSeat{{Name: "the-artifact-store", Emits: []string{"image.pushed"}}}); len(bad) != 0 {
+		t.Fatalf("an event a seat emits was reported as matching nothing: %v", bad)
+	}
+}
+
+func theCataloguesEvents(t *testing.T) ([]AnEmitter, []AConsumer, []DeclaredSeat) {
+	t.Helper()
+	root := filepath.Join("..", "..", "..", "mesh-catalog", "modules")
+	entries, err := os.ReadDir(root)
+	if err != nil {
+		t.Skipf("catalogue sibling not present: %v", err)
+	}
+	var emitters []AnEmitter
+	var consumers []AConsumer
+	// The mesh's own roles, which emit under the seat's name rather than any module's (novox/hq
+	// ADR 0121). Without these the check skips every consumer of a role's event as "the emitter is
+	// not installed" — which is how it passed vacuously the first time one existed.
+	var seats []DeclaredSeat
+	for _, own := range catalogue.SeatsWithAProtocol() {
+		seats = append(seats, DeclaredSeat{Name: own.Name, Accepts: own.Accepts, Emits: own.Emits})
+	}
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		raw, err := os.ReadFile(filepath.Join(root, e.Name(), "module.json"))
+		if err != nil {
+			continue
+		}
+		var m struct {
+			Module   string   `json:"module"`
+			Emits    []string `json:"emits"`
+			Consumes []string `json:"consumes"`
+			Seats    []struct {
+				Name  string   `json:"name"`
+				Emits []string `json:"emits"`
+			} `json:"seats"`
+		}
+		if err := json.Unmarshal(raw, &m); err != nil {
+			t.Fatalf("%s: %v", e.Name(), err)
+		}
+		if len(m.Emits) > 0 {
+			emitters = append(emitters, AnEmitter{Module: m.Module, Emits: m.Emits})
+		}
+		if len(m.Consumes) > 0 {
+			consumers = append(consumers, AConsumer{Module: m.Module, Consumes: m.Consumes})
+		}
+		for _, s := range m.Seats {
+			seats = append(seats, DeclaredSeat{Name: s.Name, Emits: s.Emits})
+		}
+	}
+	if len(emitters) == 0 {
+		t.Skip("no manifests found beside this checkout")
+	}
+	return emitters, consumers, seats
+}
+
+// **Do the derived subjects meet, not just the names?**
+//
+// The check above compares what a consumer asks for against what an emitter says it emits, by name. It
+// passed while the catalogue's subscription pointed at `mesh.mod.mesh-build-machine.event.built` — a
+// module namespace for a role's event, which no emitter owns. The names agreed; the subjects did not,
+// and the graph stayed empty.
+//
+// So this compares the thing that actually has to match: the subject a consumer subscribes against the
+// subject an emitter publishes. It is the last place the two halves can be held together, because
+// after this the server is the only thing that knows and it says nothing — a subscription that matches
+// nothing is silence.
+func TestTheCataloguesDerivedSubjectsMeet(t *testing.T) {
+	emitters, consumers, seats := theCataloguesEvents(t)
+
+	// Every subject something publishes: a module's own events, and the events of every role.
+	published := map[string]bool{}
+	for _, e := range emitters {
+		for _, name := range e.Emits {
+			published["mesh.mod."+e.Module+".event."+name] = true
+		}
+	}
+	for _, s := range seats {
+		for _, name := range s.Emits {
+			published["mesh.seat."+s.Name+".event."+name] = true
+		}
+	}
+
+	byName := map[string]DeclaredSeat{}
+	for _, s := range seats {
+		byName[s.Name] = s
+	}
+
+	var lonely []string
+	for _, c := range consumers {
+		principal := Principal{Kind: KindModule, Node: "one", Module: c.Module, PasswordHash: "x"}
+		for _, want := range c.Consumes {
+			emitter, event, named := strings.Cut(want, ".")
+			if named {
+				if s, isASeat := byName[emitter]; isASeat {
+					principal.Watches = append(principal.Watches,
+						Seat{Name: s.Name, Emits: []string{event}})
+					continue
+				}
+			}
+			principal.Consumes = append(principal.Consumes, want)
+		}
+		perms, err := PermissionsFor(principal)
+		if err != nil {
+			t.Fatalf("%s: %v", c.Module, err)
+		}
+		for _, subject := range perms.Subscribe {
+			if !strings.Contains(subject, ".event.") {
+				continue
+			}
+			if reaches(subject, published) {
+				continue
+			}
+			// A wildcard over emitters reaches whatever arrives later, and an emitter that is not
+			// installed is ordinary — both are already excused by the check above, so only a subject
+			// that can never match anything gets here.
+			if strings.Contains(subject, "*") || strings.Contains(subject, ">") {
+				continue
+			}
+			lonely = append(lonely, c.Module+" subscribes "+subject+", which nothing publishes")
+		}
+	}
+	if len(lonely) > 0 {
+		sort.Strings(lonely)
+		t.Fatalf("%d subscription(s) derive to a subject no emitter owns:\n  %s",
+			len(lonely), strings.Join(lonely, "\n  "))
+	}
+}
+
+// And it catches the thing it exists for: a role's event read as a module's.
+func TestTheDerivedSubjectCheckCatchesARolesEventReadAsAModules(t *testing.T) {
+	published := map[string]bool{"mesh.seat.mesh-build-machine.event.built": true}
+	// What the derivation produced before a consumed seat name was resolved as one.
+	if reaches("mesh.mod.mesh-build-machine.event.built", published) {
+		t.Fatal("a module namespace was treated as reaching a role's event, which is the bug")
+	}
+	// And the corrected one does reach it.
+	if !reaches("mesh.seat.mesh-build-machine.event.built", published) {
+		t.Fatal("the role's own subject does not reach the role's event")
+	}
+}
+
+// reaches says whether a subscribed subject admits any published one.
+func reaches(subject string, published map[string]bool) bool {
+	for p := range published {
+		if admitsSubject(strings.Split(subject, "."), strings.Split(p, ".")) {
+			return true
+		}
+	}
+	return false
+}
+
+func admitsSubject(pattern, subject []string) bool {
+	for i, token := range pattern {
+		if token == ">" {
+			return i < len(subject)
+		}
+		if i >= len(subject) {
+			return false
+		}
+		if token != "*" && token != subject[i] {
+			return false
+		}
+	}
+	return len(pattern) == len(subject)
+}

internal/broker/derived.go

@@ -0,0 +1,237 @@
+package broker
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// Streams and consumers derived from what modules declare.
+//
+// The mesh's own four exist before any module does (streams.go). Everything here is the other
+// half: a seat's stream comes into being when the module declaring it is **registered**, and a
+// consumer when a module is **assigned** — which is why ADR 0116's task 1.4 had to be narrowed to
+// the foundation set. Neither has happened at genesis.
+//
+// All of it is a pure function of declarations. The controller is still the only writer; this is
+// only what it writes.
+
+// A Consumer is a durable subscription the controller creates on a module's behalf. A module
+// declares what it reacts to, never how delivery works, so it does not name these and cannot
+// misconfigure them.
+type Consumer struct {
+	Name   string
+	Stream string
+	// Filters are the subjects this consumer receives. One consumer per module with several
+	// filters, rather than one per consumed event: its ack subject is derived from its name, and
+	// a module with five consumers would need five ack permissions to ack its own deliveries.
+	Filters []string
+	// Queue is the queue group, set for a seat's worker so that "exactly one holder" survives a
+	// seat later being relaxed to several. Authority and delivery are kept separate on purpose.
+	Queue string
+	// Push asks the server to deliver to a subject rather than wait to be pulled.
+	//
+	// For the mesh's own consumer, where the controller wants every message to arrive in the one
+	// loop it already runs: pulling would mean a second goroutine fetching batches and handing
+	// them over, and a loop that acts on one message at a time is the property the store window
+	// depends on. A queue group implies this, because a group has nothing to pull from.
+	Push bool
+	// AckWaitSeconds before an unacknowledged delivery is redelivered.
+	AckWaitSeconds int
+	// MaxDeliver before the message is dead-lettered; zero for the mesh's default.
+	MaxDeliver int
+	Why        string
+}
+
+// seatStreamName is the stream holding a seat's inbound work. Named after the seat rather than
+// the module holding it, because the holder can change and the queued work must not care — which
+// is the whole reason a caller addresses a seat instead of a module.
+func seatStreamName(seat string) string { return "SEAT_" + upperSnake(seat) }
+
+// SeatStreams is one work queue per declared seat, created when the declaring module is
+// registered rather than when it is assigned.
+//
+// **The stream exists before anyone holds the seat, and that is the point.** Work queues until a
+// holder appears, so installing the telegram module a week after something started sending to it
+// flushes the backlog instead of having lost it. A stream created at assignment would make "the
+// holder is not here yet" mean "your messages are gone".
+func SeatStreams(seats []DeclaredSeat) []Stream {
+	sorted := append([]DeclaredSeat(nil), seats...)
+	sort.Slice(sorted, func(i, j int) bool { return sorted[i].Name < sorted[j].Name })
+
+	var out []Stream
+	for _, s := range sorted {
+		if len(s.Accepts) == 0 {
+			// A seat that only emits and serves needs no stream: its events ride EVENTS and its
+			// tools are core request/reply, which is never persisted.
+			continue
+		}
+		retain := s.RetainSeconds
+		if retain == 0 {
+			retain = 7 * 24 * 60 * 60
+		}
+		out = append(out, Stream{
+			Name:      seatStreamName(s.Name),
+			Subjects:  []string{"mesh.seat." + s.Name + ".accept.>"},
+			Retention: RetentionWorkQueue,
+			MaxAge:    retain,
+			Why: fmt.Sprintf("work submitted to the %s seat; one holder consumes it, and it "+
+				"queues while nobody does", s.Name),
+		})
+	}
+	return out
+}
+
+// A DeclaredSeat is a seat as the catalogue knows it. Mirrored here rather than imported so this
+// package stays free of the catalogue's own types — the same reason the host mirrors the
+// contracts instead of importing the sdk.
+type DeclaredSeat struct {
+	Name    string
+	Accepts []string
+	// Emits are the verbs the seat's holder publishes under the seat's own name. An event about a
+	// role belongs here rather than in the holder's namespace, because the name then outlives
+	// whoever fills it (novox/hq ADR 0121, 04-ISSUES/127).
+	Emits []string
+	// Serves are the verbs the holder answers, request and reply.
+	Serves        []string
+	RetainSeconds int
+}
+
+// ConsumerFor is the durable consumer a module's declarations imply, or false when it subscribes
+// to nothing and needs none.
+//
+// One per module, with every consumed subject as a filter, because its ack permission is derived
+// from its name: a module with a consumer per event would need an ack permission per consumer,
+// and the permission list would stop being derivable from the declaration.
+func ConsumerFor(p Principal) (Consumer, bool) {
+	// A module that reacts to anything — a module's events or a role's (novox/hq ADR 0121). Watching
+	// a role was missing here, so the one module that does it got no consumer at all: it started,
+	// connected, and its graph stayed empty with nothing anywhere reporting why.
+	if p.Kind != KindModule || (len(p.Consumes) == 0 && len(p.Watches) == 0) {
+		return Consumer{}, false
+	}
+	perms, err := PermissionsFor(p)
+	if err != nil {
+		return Consumer{}, false
+	}
+	// Events, wherever they live: a module's own namespace, and the namespace of any role it watches
+	// (novox/hq ADR 0121). Tool subjects and inboxes are subscribed directly and are not a consumer's
+	// business, which is why this is a filter and not the whole list.
+	var filters []string
+	for _, s := range perms.Subscribe {
+		if strings.Contains(s, ".event.") {
+			filters = append(filters, s)
+		}
+	}
+	if len(filters) == 0 {
+		return Consumer{}, false
+	}
+	sort.Strings(filters)
+	return Consumer{
+		Name:           consumerDurable(p),
+		Stream:         consumerStream(p),
+		Filters:        filters,
+		AckWaitSeconds: 30,
+		MaxDeliver:     5,
+		Why:            "what " + p.Module + " declared it consumes; after max-deliver it dead-letters",
+	}, true
+}
+
+// HolderConsumerFor is the worker a seat's holder gets on that seat's work queue.
+//
+// **A queue group even though the seat guarantees one holder.** The seat is *authority* — who may
+// be the telegram sender — and the queue group is *delivery*. Tie delivery to the seat and the
+// day somebody allows two holders for throughput, every message is processed twice with nothing
+// reporting it. Kept separate, relaxing one changes nothing about the other.
+func HolderConsumerFor(node, module string, seat DeclaredSeat) (Consumer, bool) {
+	if len(seat.Accepts) == 0 {
+		return Consumer{}, false
+	}
+	return Consumer{
+		Name:           "SEAT_" + upperSnake(seat.Name) + "_worker",
+		Stream:         seatStreamName(seat.Name),
+		Filters:        []string{"mesh.seat." + seat.Name + ".accept.>"},
+		Queue:          "holders",
+		AckWaitSeconds: 60,
+		MaxDeliver:     5,
+		Why: fmt.Sprintf("%s on %s holds %s; it acknowledges after the work is done, so a "+
+			"crash mid-work redelivers rather than loses", module, node, seat.Name),
+	}, true
+}
+
+// NodeConsumer is the durable consumer a node reads its own declaration through.
+//
+// **Derived from a node existing, and created by the controller, because a host cannot create it.**
+// A host's account may subscribe its own declaration subject and publish its own ack subject, and
+// reaches no part of the JetStream API — which is correct (the controller is the only writer of
+// consumer definitions, design 25 §3) and means the consumer must be waiting before the host binds
+// to it. Named after the node, because the node's ack grant is `$JS.ACK.NODES.<node>.>` and a
+// consumer named anything else is one the host cannot acknowledge a delivery from.
+//
+// **No max-deliver, and a long ack wait.** A declaration is settled only after the node has applied
+// it and reported, which is minutes on a machine pulling images; and a declaration the mesh cannot
+// get a node to accept is not one to dead-letter, because the stream keeps only the newest per node
+// anyway — so there is exactly one message per node to redeliver, for as long as that node is away.
+func NodeConsumer(node string) Consumer {
+	return Consumer{
+		Name:           node,
+		Stream:         "NODES",
+		Filters:        []string{"mesh.node." + node + ".declare"},
+		Push:           true,
+		AckWaitSeconds: 300,
+		Why: "how " + node + " hears what it should be; last-per-subject, so a node that was away " +
+			"gets exactly the current declaration and nothing older",
+	}
+}
+
+// AssertNodeConsumers brings every known node's declaration consumer into being.
+//
+// Asserted on start as well as created at enrolment, for the reason the streams are: a mesh raised
+// from a restored backup, or one whose bus was recreated, has node records and no consumers, and a
+// node whose consumer is missing hears nothing while everything else about it looks correct.
+func AssertNodeConsumers(e Ensurer, nodes []string) error {
+	for _, n := range nodes {
+		if err := e.EnsureConsumer(NodeConsumer(n)); err != nil {
+			return fmt.Errorf("asserting how %s hears its declaration: %w", n, err)
+		}
+	}
+	return nil
+}
+
+// AllOverlaps reports subject filters claimed by more than one stream, across the mesh's own and
+// every derived one.
+//
+// NATS refuses an overlapping stream rather than merging it (verified against nats-server 2.10:
+// "subjects overlap with an existing stream"), so this is not a subtle divergence — it is a
+// registration that fails. Catching it here names both streams, before a half-applied mesh does.
+func AllOverlaps(seats []DeclaredSeat) []string {
+	all := append(MeshStreams(), SeatStreams(seats)...)
+	seen := map[string]string{}
+	var clashes []string
+	for _, s := range all {
+		for _, subject := range s.Subjects {
+			if first, ok := seen[subject]; ok {
+				clashes = append(clashes, fmt.Sprintf("%s and %s both claim %s", first, s.Name, subject))
+				continue
+			}
+			seen[subject] = s.Name
+		}
+	}
+	sort.Strings(clashes)
+	return clashes
+}
+
+// upperSnake makes a stream name from a seat name. NATS stream names may not contain a dot,
+// a space or a wildcard, and a hyphen is legal but reads badly beside the mesh's own.
+func upperSnake(s string) string {
+	out := []rune(s)
+	for i, r := range out {
+		switch {
+		case r >= 'a' && r <= 'z':
+			out[i] = r - 32
+		case r == '-' || r == '.':
+			out[i] = '_'
+		}
+	}
+	return string(out)
+}

internal/broker/derived_test.go

@@ -0,0 +1,155 @@
+package broker
+
+import (
+	"strings"
+	"testing"
+)
+
+func telegramSeat() DeclaredSeat {
+	return DeclaredSeat{Name: "telegram-sender", Accepts: []string{"send"}}
+}
+
+// The stream exists from registration, not assignment: work queues until a holder appears, so
+// installing the module a week later flushes the backlog rather than having lost it.
+func TestASeatGetsAWorkQueueOfItsOwn(t *testing.T) {
+	got := SeatStreams([]DeclaredSeat{telegramSeat()})
+	if len(got) != 1 {
+		t.Fatalf("expected one stream, got %d", len(got))
+	}
+	s := got[0]
+	if s.Retention != RetentionWorkQueue {
+		t.Fatalf("a seat's inbound queue retains as %q; one holder must take each message once", s.Retention)
+	}
+	if s.Subjects[0] != "mesh.seat.telegram-sender.accept.>" {
+		t.Fatalf("filters on %v", s.Subjects)
+	}
+}
+
+// A seat that only emits and serves needs no stream: its events ride EVENTS and its tools are
+// core request/reply, which is never persisted.
+func TestASeatThatAcceptsNothingGetsNoStream(t *testing.T) {
+	if got := SeatStreams([]DeclaredSeat{{Name: "announcer"}}); len(got) != 0 {
+		t.Fatalf("a seat with no inbound work got %d stream(s)", len(got))
+	}
+}
+
+// Retention belongs to whoever owns the namespace, and a seat owns its own.
+func TestASeatsRetentionIsItsOwn(t *testing.T) {
+	s := SeatStreams([]DeclaredSeat{{Name: "slow", Accepts: []string{"work"}, RetainSeconds: 30 * 24 * 60 * 60}})
+	if s[0].MaxAge != 30*24*60*60 {
+		t.Fatalf("the seat's declared retention was not used: %d", s[0].MaxAge)
+	}
+	d := SeatStreams([]DeclaredSeat{telegramSeat()})
+	if d[0].MaxAge == 0 {
+		t.Fatal("a seat that declares no retention got an unbounded queue")
+	}
+}
+
+// NATS refuses an overlapping stream outright, so a clash here is a registration that fails.
+func TestNoDerivedStreamOverlapsTheMeshsOwn(t *testing.T) {
+	seats := []DeclaredSeat{telegramSeat(), {Name: "licensing-master", Accepts: []string{"report"}}}
+	if c := AllOverlaps(seats); len(c) != 0 {
+		t.Fatalf("overlapping filters: %v", c)
+	}
+}
+
+// One consumer per module, with every consumed subject as a filter — because its ack permission
+// is derived from its name, and a consumer per event would need an ack permission per consumer.
+func TestAModuleGetsOneConsumerCarryingEveryFilter(t *testing.T) {
+	c, ok := ConsumerFor(Principal{Kind: KindModule, Node: "one", Module: "audit",
+		Consumes: []string{"shop.order.placed", "billing.invoice.sent"}, PasswordHash: "x"})
+	if !ok {
+		t.Fatal("a module that consumes got no consumer")
+	}
+	if len(c.Filters) != 2 {
+		t.Fatalf("expected both subjects as filters, got %v", c.Filters)
+	}
+	perms, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "audit",
+		Consumes: []string{"shop.order.placed"}, PasswordHash: "x"})
+	ack := "$JS.ACK." + c.Stream + "." + c.Name + ".>"
+	found := false
+	for _, p := range perms.Publish {
+		if p == ack {
+			found = true
+		}
+	}
+	if !found {
+		t.Fatalf("the consumer is named %q but the ack permission is %v; a module could not ack "+
+			"its own deliveries", c.Name, perms.Publish)
+	}
+}
+
+// A module that subscribes to nothing needs no consumer, and creating one would leave an object
+// nothing reads and everything has to maintain.
+func TestAModuleThatConsumesNothingGetsNoConsumer(t *testing.T) {
+	if _, ok := ConsumerFor(Principal{Kind: KindModule, Node: "one", Module: "shop",
+		Emits: []string{"order.placed"}, PasswordHash: "x"}); ok {
+		t.Fatal("a pure emitter got a consumer")
+	}
+}
+
+// The seat is authority and the queue group is delivery. Tie them together and the day somebody
+// allows two holders, every message is processed twice with nothing reporting it.
+func TestAHoldersWorkerUsesAQueueGroupAnyway(t *testing.T) {
+	c, ok := HolderConsumerFor("one", "telegram", telegramSeat())
+	if !ok {
+		t.Fatal("the holder of a seat with inbound work got no worker")
+	}
+	if c.Queue == "" {
+		t.Fatal("the worker is not in a queue group, so a second holder would double-process")
+	}
+	if c.Stream != "SEAT_TELEGRAM_SENDER" {
+		t.Fatalf("the worker reads %q, not the seat's own stream", c.Stream)
+	}
+	if c.MaxDeliver == 0 {
+		t.Fatal("a failing worker would redeliver forever rather than dead-letter")
+	}
+}
+
+// The stream is named after the seat, not its holder: the holder can change and the queued work
+// must not care.
+func TestASeatsStreamIsNamedAfterTheSeat(t *testing.T) {
+	name := seatStreamName("telegram-sender")
+	if strings.Contains(name, "telegram-sender") {
+		t.Fatalf("%q keeps characters a stream name may not hold", name)
+	}
+	if name != "SEAT_TELEGRAM_SENDER" {
+		t.Fatalf("unexpected stream name %q", name)
+	}
+}
+
+// A node hears its declaration through a consumer only the controller can make.
+//
+// The three things that would each break it silently: a name other than the node's is one the host
+// cannot acknowledge a delivery from, because its ack grant is derived from the node's name; a
+// filter other than its own declaration subject is a node reading another's; and a pull consumer is
+// one the host cannot bind a channel to without creating something, which it has no authority for.
+func TestANodesDeclarationConsumerIsWhatItsOwnGrantAllows(t *testing.T) {
+	c := NodeConsumer("anchor")
+	if c.Name != "anchor" {
+		t.Fatalf("named %q, so the node cannot ack from it: its grant is $JS.ACK.NODES.anchor.>", c.Name)
+	}
+	if c.Stream != "NODES" {
+		t.Fatalf("on stream %q rather than the one declarations live in", c.Stream)
+	}
+	if len(c.Filters) != 1 || c.Filters[0] != "mesh.node.anchor.declare" {
+		t.Fatalf("filters %v, which is not this node's own declaration and nothing else", c.Filters)
+	}
+	if !c.Push {
+		t.Fatal("pulled, which a host cannot do: pulling needs the JetStream API and a host reaches none of it")
+	}
+	if c.MaxDeliver != 0 {
+		t.Fatalf("max-deliver %d: a declaration a node has not taken yet is not one to dead-letter, "+
+			"because the stream holds exactly one per node", c.MaxDeliver)
+	}
+
+	// And the grant the node actually gets has to match, or none of the above matters.
+	perms, err := PermissionsFor(Principal{Kind: KindNode, Node: "anchor"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Without the ack grant every declaration a node receives is redelivered for ever; without the
+	// subscribe grant its consumer delivers to nobody.
+	has(t, perms.Publish, "$JS.ACK.NODES."+c.Name+".>")
+	has(t, perms.Subscribe, c.Filters[0])
+}

internal/broker/genesis_template_test.go

@@ -0,0 +1,126 @@
+package broker
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"regexp"
+	"sort"
+	"strings"
+	"testing"
+
+	"golang.org/x/crypto/bcrypt"
+)
+
+// **The first user list the installer carries must be the one the controller would compose.**
+//
+// At genesis there is no mesh to write the bus's user list, so the installer carries one: the
+// controller's own account, at a bootstrap password, the way the store is reached at
+// `postgres:bootstrap` (novox/hq design 25 §4, task 1.7). It is written by hand in a template and
+// derived in code here, which is two statements of one fact — so this compares them.
+//
+// Getting it wrong is the worst kind of silent: a controller whose carried permissions are narrower
+// than the ones it derives comes up, connects, and is refused on the first thing it tries, with an
+// authorisation error that names a subject and not the template that forgot it. And a mesh cannot be
+// raised twice to find out.
+func TestTheInstallersFirstUserListIsWhatTheControllerWouldCompose(t *testing.T) {
+	accounts := theCarriedAccounts(t)
+
+	want, err := PermissionsFor(Principal{Kind: KindController, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	carriedPub := subjectsIn(accounts, "publish")
+	carriedSub := subjectsIn(accounts, "subscribe")
+
+	if diff := missing(want.Publish, carriedPub); len(diff) > 0 {
+		t.Errorf("the installer's user list does not let the controller publish %v — it would come up "+
+			"and be refused on the first thing it tried", diff)
+	}
+	if diff := missing(want.Subscribe, carriedSub); len(diff) > 0 {
+		t.Errorf("the installer's user list does not let the controller subscribe %v", diff)
+	}
+	// And nothing wider than what it derives, or genesis quietly grants a privilege the composition
+	// takes away again on the first push.
+	if diff := missing(carriedPub, want.Publish); len(diff) > 0 {
+		t.Errorf("the installer's user list lets the controller publish %v, which it does not derive", diff)
+	}
+	if diff := missing(carriedSub, want.Subscribe); len(diff) > 0 {
+		t.Errorf("the installer's user list lets the controller subscribe %v, which it does not derive", diff)
+	}
+
+	// The credential is the bootstrap one and the hash really is of it, because a hash of something
+	// else is a controller that cannot log in to the bus it was just given.
+	hash := regexp.MustCompile(`\$2[aby]?\$[0-9]+\$[A-Za-z0-9./]{53}`).FindString(accounts)
+	if hash == "" {
+		t.Fatal("the installer's user list carries no password hash")
+	}
+	if err := bcrypt.CompareHashAndPassword([]byte(hash), []byte("bootstrap")); err != nil {
+		t.Fatalf("the carried hash does not verify the bootstrap credential the template also carries: %v", err)
+	}
+}
+
+// theCarriedAccounts is the accounts file the installer's template writes at genesis.
+func theCarriedAccounts(t *testing.T) string {
+	t.Helper()
+	path := filepath.Join("..", "..", "..", "mesh-host", "examples", "foundation-first-node-nats.lock")
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Skipf("the host's checkout is not beside this one: %v", err)
+	}
+	// The template is JSON with line comments, which is how every one of them is written.
+	var lines []string
+	for _, l := range strings.Split(string(raw), "\n") {
+		if !strings.HasPrefix(strings.TrimSpace(l), "//") {
+			lines = append(lines, l)
+		}
+	}
+	var bundle struct {
+		Resources []map[string]any `json:"resources"`
+	}
+	if err := json.Unmarshal([]byte(strings.Join(lines, "\n")), &bundle); err != nil {
+		t.Fatalf("the template is not readable: %v", err)
+	}
+	for _, r := range bundle.Resources {
+		if r["id"] == "bus-accounts" {
+			content, _ := r["content"].(string)
+			if content == "" {
+				t.Fatal("the template's accounts file is empty, so the bus would refuse every connection")
+			}
+			return content
+		}
+	}
+	t.Fatal("the template carries no accounts file, so a mesh raised from it has a bus nobody may use")
+	return ""
+}
+
+// subjectsIn reads one allow-list out of a composed accounts file.
+func subjectsIn(accounts, which string) []string {
+	found := regexp.MustCompile(which + `: \{ allow: \[([^\]]*)\]`).FindStringSubmatch(accounts)
+	if len(found) != 2 {
+		return nil
+	}
+	var out []string
+	for _, part := range strings.Split(found[1], ",") {
+		if s := strings.Trim(strings.TrimSpace(part), `"`); s != "" {
+			out = append(out, s)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// missing is what is in want and not in got.
+func missing(want, got []string) []string {
+	have := map[string]bool{}
+	for _, g := range got {
+		have[g] = true
+	}
+	var out []string
+	for _, w := range want {
+		if !have[w] {
+			out = append(out, w)
+		}
+	}
+	return out
+}

internal/broker/jetstream.go

@@ -0,0 +1,150 @@
+package broker
+
+import (
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/nats-io/nats.go"
+)
+
+// The JetStream side of the controller: the one place the mesh's streams and consumers are
+// actually created.
+//
+// Everything that decides *what* they are is pure and lives beside this (streams.go, derived.go).
+// This is only the part that talks to a server, kept small on purpose: a bug in a subject filter
+// should be findable in a unit test, and only a bug in "did the server accept it" should need one
+// running.
+
+// A JetStream is a connection to the bus, as the controller uses it.
+type JetStream struct {
+	conn *nats.Conn
+	js   nats.JetStreamContext
+}
+
+// Dial connects and returns the controller's JetStream handle.
+func Dial(url string, opts ...nats.Option) (*JetStream, error) {
+	// A name, because a connection nobody can identify in the server's own monitoring is one
+	// nobody can attribute a problem to.
+	opts = append(opts, nats.Name("mesh-controller"), nats.Timeout(10*time.Second))
+	conn, err := nats.Connect(url, opts...)
+	if err != nil {
+		return nil, fmt.Errorf("connecting to the bus at %s: %w", url, err)
+	}
+	js, err := conn.JetStream()
+	if err != nil {
+		conn.Close()
+		return nil, fmt.Errorf("the bus at %s has no JetStream: %w", url, err)
+	}
+	return &JetStream{conn: conn, js: js}, nil
+}
+
+// Conn is the connection itself, for what the mesh keeps off JetStream on purpose — a heartbeat,
+// a tool call — where a lost message is answered by the next one or by a timeout the caller
+// already handles (design 25 §3).
+func (j *JetStream) Conn() *nats.Conn { return j.conn }
+
+// Context is the JetStream handle, for subscribing to what the consumers above define.
+func (j *JetStream) Context() nats.JetStreamContext { return j.js }
+
+func (j *JetStream) Close() {
+	if j.conn != nil {
+		j.conn.Close()
+	}
+}
+
+// EnsureStream creates the stream if it is absent and brings it to match if it is present.
+//
+// **Idempotent, because the controller asserts on every start** rather than creating once at
+// genesis: a stream somebody deleted, or a mesh raised from a restored backup, has to converge
+// rather than run without the guarantee its messages assume.
+//
+// An update, not a delete and recreate. Recreating would discard every message the stream holds
+// and every consumer's position in it — which for CONTROL means the pushes being held through a
+// store restart, exactly the guarantee the stream exists for.
+func (j *JetStream) EnsureStream(s Stream) error {
+	want := &nats.StreamConfig{
+		Name:              s.Name,
+		Subjects:          s.Subjects,
+		Retention:         retentionOf(s.Retention),
+		MaxAge:            time.Duration(s.MaxAge) * time.Second,
+		MaxMsgsPerSubject: int64(s.MaxMsgsPerSubject),
+		Description:       s.Why,
+	}
+	if s.Retention == RetentionLastPerSubject {
+		// Last-per-subject is a limits stream with one message kept per subject, not a
+		// retention policy of its own — the state shape, spelled the way the server spells it.
+		want.Retention = nats.LimitsPolicy
+		want.MaxMsgsPerSubject = 1
+		want.MaxAge = 0
+	}
+
+	switch _, err := j.js.StreamInfo(s.Name); {
+	case err == nil:
+		if _, err := j.js.UpdateStream(want); err != nil {
+			return fmt.Errorf("bringing stream %s to match: %w", s.Name, err)
+		}
+		return nil
+	case errors.Is(err, nats.ErrStreamNotFound):
+		if _, err := j.js.AddStream(want); err != nil {
+			return fmt.Errorf("creating stream %s: %w", s.Name, err)
+		}
+		return nil
+	default:
+		return fmt.Errorf("asking about stream %s: %w", s.Name, err)
+	}
+}
+
+// EnsureConsumer creates or updates one durable consumer.
+//
+// Explicit acknowledgement throughout: a consumer that acknowledges on delivery cannot redeliver
+// work its holder died in the middle of, which is the whole difference between a queue and a
+// firehose.
+func (j *JetStream) EnsureConsumer(c Consumer) error {
+	want := &nats.ConsumerConfig{
+		Durable:        c.Name,
+		AckPolicy:      nats.AckExplicitPolicy,
+		AckWait:        time.Duration(c.AckWaitSeconds) * time.Second,
+		MaxDeliver:     c.MaxDeliver,
+		DeliverGroup:   c.Queue,
+		DeliverSubject: "",
+		Description:    c.Why,
+	}
+	switch len(c.Filters) {
+	case 0:
+	case 1:
+		want.FilterSubject = c.Filters[0]
+	default:
+		want.FilterSubjects = c.Filters
+	}
+	// A queue group needs a delivery subject: a pull consumer has no group, and declaring one
+	// without the other is refused by the server with a message that does not say which half is
+	// missing.
+	if c.Queue != "" || c.Push {
+		want.DeliverSubject = "_DELIVER." + c.Name
+	}
+
+	switch _, err := j.js.ConsumerInfo(c.Stream, c.Name); {
+	case err == nil:
+		if _, err := j.js.UpdateConsumer(c.Stream, want); err != nil {
+			return fmt.Errorf("bringing consumer %s on %s to match: %w", c.Name, c.Stream, err)
+		}
+		return nil
+	case errors.Is(err, nats.ErrConsumerNotFound):
+		if _, err := j.js.AddConsumer(c.Stream, want); err != nil {
+			return fmt.Errorf("creating consumer %s on %s: %w", c.Name, c.Stream, err)
+		}
+		return nil
+	default:
+		return fmt.Errorf("asking about consumer %s on %s: %w", c.Name, c.Stream, err)
+	}
+}
+
+func retentionOf(r Retention) nats.RetentionPolicy {
+	switch r {
+	case RetentionWorkQueue:
+		return nats.WorkQueuePolicy
+	default:
+		return nats.LimitsPolicy
+	}
+}

internal/broker/jetstream_test.go

@@ -0,0 +1,71 @@
+package broker
+
+import (
+	"os"
+	"testing"
+)
+
+// Against a real server, because the questions here are all "does the server accept this" —
+// which a mock would answer by agreeing with whatever this file already believes.
+//
+// Skipped unless MESH_TEST_NATS names one, so the ordinary suite stays fast and offline:
+//
+//	docker run -d --rm --name t -p 14222:4222 nats:2.10-alpine -js
+//	MESH_TEST_NATS=nats://127.0.0.1:14222 go test ./internal/broker/ -run TestAgainstARealServer
+func TestAgainstARealServer(t *testing.T) {
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	js, err := Dial(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer js.Close()
+
+	t.Run("the mesh's own streams are accepted", func(t *testing.T) {
+		if err := AssertMeshStreams(js); err != nil {
+			t.Fatal(err)
+		}
+	})
+
+	t.Run("asserting again changes nothing and fails nothing", func(t *testing.T) {
+		if err := AssertMeshStreams(js); err != nil {
+			t.Fatalf("the second assertion failed, so the controller cannot restart: %v", err)
+		}
+	})
+
+	t.Run("a seat's work queue is accepted beside them", func(t *testing.T) {
+		seats := []DeclaredSeat{{Name: "telegram-sender", Accepts: []string{"send"}}}
+		for _, s := range SeatStreams(seats) {
+			if err := js.EnsureStream(s); err != nil {
+				t.Fatal(err)
+			}
+		}
+		if c := AllOverlaps(seats); len(c) != 0 {
+			t.Fatalf("overlaps the server would refuse: %v", c)
+		}
+	})
+
+	t.Run("a module's consumer is accepted and is idempotent", func(t *testing.T) {
+		c, ok := ConsumerFor(Principal{Kind: KindModule, Node: "one", Module: "audit",
+			Consumes: []string{"shop.order.placed", "billing.invoice.sent"}, PasswordHash: "x"})
+		if !ok {
+			t.Fatal("no consumer derived")
+		}
+		if err := js.EnsureConsumer(c); err != nil {
+			t.Fatal(err)
+		}
+		if err := js.EnsureConsumer(c); err != nil {
+			t.Fatalf("the second assertion failed: %v", err)
+		}
+	})
+
+	t.Run("a holder's worker is accepted with its queue group", func(t *testing.T) {
+		c, _ := HolderConsumerFor("one", "telegram",
+			DeclaredSeat{Name: "telegram-sender", Accepts: []string{"send"}})
+		if err := js.EnsureConsumer(c); err != nil {
+			t.Fatal(err)
+		}
+	})
+}

internal/broker/nats.go

@@ -0,0 +1,547 @@
+// Composing the bus's own configuration.
+//
+// An account is *composed*, never called for: the controller writes accounts, users and
+// per-subject permissions into one file the host keeps current, and the server reloads it in
+// place (novox/hq ADR 0106 — never through a management API; design 25 §4).
+//
+// Everything here is pure. Given the principals, it returns the file's text — so the whole of the
+// mesh's authority model is testable as strings, with no server.
+//
+// **Permissions are per subject, so a module's own name is the server's to enforce.** ADR 0042
+// reserves a module's origin — it publishes only under its own name — and here that is a refusal
+// rather than something a library promises.
+package broker
+
+import (
+	"errors"
+	"fmt"
+	"regexp"
+	"sort"
+	"strings"
+)
+
+// A Kind is what a principal is, which decides the shape of its authority rather than its
+// contents: a module's comes from its declaration, a host's from its node, and the controller's
+// and the enrolment user's are fixed.
+type Kind string
+
+const (
+	KindModule     Kind = "module"
+	KindNode       Kind = "node"
+	KindController Kind = "controller"
+	KindEnrolment  Kind = "enrolment"
+	// KindPerson is somebody reaching the mesh's tools from a workstation (design 25 §7). Its
+	// authority is a list of tools and nothing else — not control, not declarations, not builds,
+	// and no ability to answer anything, because a person asks.
+	KindPerson Kind = "person"
+)
+
+// Seat is a role on the bus as a principal relates to it: the subjects it accepts, and those it
+// emits (novox/hq ADR 0118, design 29 §5).
+type Seat struct {
+	Name     string
+	Accepts  []string
+	Emits    []string
+	Serves   []string
+	Versions []string // protocol versions served beside the current one; empty for v1 only
+}
+
+// A Principal is one user of the bus. Its permissions are derived from what it declares and
+// nothing else (novox/hq ADR 0043), over the three namespaces of design 29 §2: its own, the seats
+// it holds, and the seats it uses.
+type Principal struct {
+	Kind   Kind
+	Node   string
+	Module string
+
+	Emits    []string
+	Consumes []string
+	Serves   []string
+
+	Holds []Seat
+	Uses  []Seat
+	// Watches are seats whose events this principal consumes. Separate from Consumes because a
+	// role's event lives under the seat's namespace and not a module's, and this package cannot tell
+	// a seat's name from a module's by looking at it — whoever resolved the declaration can, and
+	// does (novox/hq ADR 0121).
+	//
+	// **Found by a consumer reading nothing.** The catalogue consumes the build machine's outcome;
+	// with that name read as a module's, its subscription pointed at `mesh.mod.mesh-build-machine.…`,
+	// a namespace no such module owns. Every service started and the graph stayed empty.
+	Watches []Seat
+
+	// Invokes are the tools a person may call, as `<module>.<tool>`; a single `*` is every tool,
+	// for an administrator. Only meaningful for KindPerson.
+	//
+	// **A list, not a role.** A person is not a module and holds no seat: nothing is addressed
+	// to them, nothing is delivered to them, and they have no durable consumer to acknowledge.
+	// What they have is permission to ask.
+	Invokes []string
+
+	// PasswordHash is the bcrypt hash the mesh minted. The plaintext is sealed to the principal
+	// and never appears here: this file is written to a node's disk and read by a server, and a
+	// secret that can be read from a configuration file is a secret with a wider blast radius
+	// than the one it protects (novox/hq design 29 §10).
+	PasswordHash string
+}
+
+// meshSeatsTheControllerUses are the roles the mesh's own flows submit work to. Named rather than
+// derived from the seat set: the controller is not a module and declares no `uses`, so its side of a
+// seat has to be stated, and a list is what makes "which roles does the mesh itself talk to" answerable.
+var meshSeatsTheControllerUses = []string{"mesh-build-machine"}
+
+// enrolmentPrefix is the space every enrolling node's user and inbox live under, so the one place the
+// controller may answer an enrolment is derived from the same constant the user is named from.
+const enrolmentPrefix = "enrol"
+
+// safeSubject refuses anything that would change the meaning of a subject rather than sit inside
+// one. A name carrying a dot would silently widen a permission by adding a token; a name carrying
+// `>` or `*` would widen it to a wildcard, which is the whole authority model gone.
+var safeSubject = regexp.MustCompile(`^[A-Za-z0-9_-]+$`)
+
+// Username is how a principal is named to the server. The node is part of it, so the same module
+// on two machines holds two users, each sealed to its own — the rule management.go already
+// applies, kept.
+func (p Principal) Username() string {
+	switch p.Kind {
+	case KindPerson:
+		return "person." + p.Module
+	case KindModule:
+		return p.Node + "." + p.Module
+	case KindNode:
+		return "node." + p.Node
+	case KindController:
+		return "controller"
+	case KindEnrolment:
+		// Per token, not one shared user. **The inbox is the reason**: with a single `enrolment`
+		// user every machine enrolling at once could read every other's answer, and an answer
+		// carries that node's credentials sealed to it. Design 25 §6 says the inbox a token
+		// derives, and a permission belongs to a user, so the user is per token.
+		//
+		// Named after the node, which **is** the token's id: a token is issued for a node record,
+		// the mesh holds one live claim per record, and the node's name is the one identifier both
+		// sides already have before anything else is agreed. It is also exactly what the other
+		// transport does, where the account is named after the node and the secret is its password.
+		return enrolmentPrefix + "." + p.Node
+	}
+	return ""
+}
+
+// inbox is a principal's own reply space. No user is ever granted a bare `_INBOX.>` (design 25
+// §4): with one account, inbox privacy is the permission list or it is nothing, so each user's
+// inbox is derived from its own identity and its permissions name that prefix and no other.
+func (p Principal) inbox() string { return "_INBOX." + p.Username() + ".>" }
+
+// Permissions is what a principal may publish and subscribe, and whether it may answer.
+type Permissions struct {
+	Publish   []string
+	Subscribe []string
+	// AllowResponses lets a principal reply to a request it received, on the reply subject that
+	// request carried, once.
+	//
+	// **This is what makes scoped inboxes possible at all**, and design 25 §4 did not say it. If
+	// every user's inbox is private to it, a module serving a tool cannot publish the answer —
+	// the answer goes to the *caller's* inbox, which the responder has no permission for. The two
+	// ways out are granting responders `_INBOX.>`, which is precisely the blanket grant §4
+	// refuses, or this: the server itself permits one reply to the subject of a message the user
+	// actually received, and nothing else. The authority is bounded by having been asked.
+	AllowResponses bool
+}
+
+// PermissionsFor derives a principal's authority. Pure, and the only place authority is decided:
+// a permission that cannot be derived from a declaration is a permission nobody can explain.
+func PermissionsFor(p Principal) (Permissions, error) {
+	for _, part := range []struct{ what, value string }{
+		{"node", p.Node}, {"module", p.Module},
+	} {
+		if part.value == "" {
+			continue
+		}
+		if !safeSubject.MatchString(part.value) {
+			return Permissions{}, fmt.Errorf(
+				"%q cannot be part of a subject: a permission is a subject pattern, and this would widen it", part.value)
+		}
+	}
+
+	var pub, sub []string
+	switch p.Kind {
+	case KindController:
+		// The controller owns the mesh's own traffic and the streams. It is the only writer of
+		// stream definitions (design 25 §3), so it alone reaches the JetStream API.
+		pub = []string{"mesh.control.>", "mesh.node.>", "$JS.API.>"}
+		sub = []string{"mesh.control.>", "$JS.API.>"}
+
+		// Work the mesh's own flows submit to a role, and the outcomes they wait on (ADR 0121). A
+		// build is the one today: the controller asks, and reads the answer from the seat's event
+		// like the catalogue does — which is why no holder needs to publish into anybody's inbox.
+		for _, seat := range meshSeatsTheControllerUses {
+			pub = append(pub, "mesh.seat."+seat+".accept.>")
+		}
+
+		// The two events it reacts to, and its ack subject on the stream they arrive from
+		// (streams.go). **Each named, not a pattern**: `mesh.mod.*.event.>` would make the
+		// controller a subscriber to every event in the mesh, and its permission list would stop
+		// saying what it is for. The ack grant below is scoped per stream because the controller's
+		// consumer name is the same on both and `$JS.ACK.CONTROL.controller.>` does not cover a
+		// delivery from EVENTS — a consumer that cannot ack has every message redelivered for
+		// ever, refused by the list it already has.
+		sub = append(sub, ControllerFollows...)
+		pub = append(pub, "$JS.ACK.EVENTS."+ControllerName+".>")
+
+		// **Where an enrolment's answer goes**, and `allow_responses` does not cover it. That
+		// permits one reply to the reply subject of a message the user received — and a message a
+		// JetStream consumer delivers has had that field claimed for the consumer's own ack address
+		// (design 25 §2), so the address the controller actually answers is the one the request
+		// carried in its payload, which is not a reply subject as the server understands it.
+		//
+		// Verified against a real server before this line existed: the answer was refused with
+		// "Permissions Violation for Publish to _INBOX.enrol.anchor…", and every enrolment on the
+		// mesh would have timed out while the controller logged success.
+		//
+		// **The enrolment inbox space, not a blanket `_INBOX.>`.** Design 25 §4 refuses that, and
+		// this is not it: nothing but an enrolling node ever subscribes under this prefix, each
+		// scoped to its own token's, so the controller publishing here is the mesh answering
+		// enrolments and can reach nothing else.
+		pub = append(pub, "_INBOX."+enrolmentPrefix+".>")
+
+	case KindPerson:
+		// Tools, and nothing else. Every subject a person may publish is a tool call; a person
+		// who could publish an event would be able to claim a module said something.
+		for _, t := range p.Invokes {
+			if t == "*" {
+				pub = append(pub, "mesh.mod.*.tool.>")
+				continue
+			}
+			module, tool, ok := strings.Cut(t, ".")
+			if !ok {
+				return Permissions{}, fmt.Errorf(
+					"%q does not name a tool: a person invokes <module>.<tool>, or * for every one", t)
+			}
+			pub = append(pub, "mesh.mod."+module+".tool."+tool)
+		}
+
+	case KindEnrolment:
+		// A leaked token is useless for anything but enrolling: it cannot read a declaration, hear
+		// an event, or subscribe any inbox but the one its own token derives (design 25 §6).
+		//
+		// **The inbox was missing and the handshake could not have completed without it.** An
+		// enrolling node publishes its request and waits on an address it states in the payload;
+		// with nothing to subscribe it waits out its timeout against a mesh that answered. Its own
+		// and no wider: `_INBOX.enrol.<node>.>`, so what is sealed to one machine cannot be read by
+		// another enrolling beside it.
+		if p.Node == "" {
+			// Refused rather than composed into `_INBOX.enrol..>`, which is a subject with an empty
+			// token in it — and worse, one every nameless enrolment user would share. A shared
+			// enrolment inbox is one machine able to read the credentials sealed to another.
+			return Permissions{}, errors.New(
+				"an enrolment user names no node, so its inbox would be shared with every other " +
+					"enrolment: a token is issued for a node record, and that record's name is " +
+					"the token's id")
+		}
+		pub = []string{"mesh.control.enrol"}
+		sub = []string{p.inbox()}
+
+	case KindNode:
+		// A host publishes its own node's control traffic and subscribes its own declaration —
+		// and nothing of any other node's.
+		pub = []string{"mesh.control." + p.Node + ".>"}
+		sub = []string{"mesh.node." + p.Node + ".declare"}
+
+	case KindModule:
+		// 1. Its own namespace: it publishes its events there and serves its tools there. Nothing
+		//    else may publish into it, so an event's source is a fact the server enforces rather
+		//    than a claim in the body (design 29 §2).
+		own := "mesh.mod." + p.Module
+		for _, e := range p.Emits {
+			pub = append(pub, own+".event."+e)
+		}
+		for _, t := range p.Serves {
+			sub = append(sub, own+".tool."+t)
+		}
+
+		// 2. What it consumes, by the emitter's own subject — an event is addressed to its
+		//    emitter, because the emitter's identity is the meaning (ADR 0118).
+		for _, c := range p.Consumes {
+			subject, err := consumedSubject(c)
+			if err != nil {
+				return Permissions{}, err
+			}
+			sub = append(sub, subject)
+		}
+
+		// 2b. Events of a role it watches, under the seat's own namespace. Subscribe only: watching a
+		//     role is hearing what it announced, not taking part in it.
+		for _, w := range p.Watches {
+			for _, e := range w.Emits {
+				sub = append(sub, seatSubject(w, "event", e))
+			}
+		}
+
+		// 3. Seats it holds: full participation.
+		for _, s := range p.Holds {
+			for _, a := range s.Accepts {
+				sub = append(sub, seatSubject(s, "accept", a))
+			}
+			for _, e := range s.Emits {
+				pub = append(pub, seatSubject(s, "event", e))
+			}
+			for _, t := range s.Serves {
+				sub = append(sub, seatSubject(s, "tool", t))
+			}
+		}
+
+		// 4. Seats it uses: publish only, and only the accepts half. A caller cannot subscribe a
+		//    seat's inbound subject and watch other modules' traffic, nor publish its outbound
+		//    events and lie about outcomes (design 29 §2).
+		for _, s := range p.Uses {
+			for _, a := range s.Accepts {
+				pub = append(pub, seatSubject(s, "accept", a))
+			}
+			for _, t := range s.Serves {
+				pub = append(pub, seatSubject(s, "tool", t))
+			}
+		}
+	}
+
+	if p.Kind == KindPerson {
+		// An inbox to hear answers in, and nothing else. No ack subject: a person has no durable
+		// consumer, because nothing is delivered to a person — they ask and are answered.
+		sub = append(sub, p.inbox())
+	}
+
+	if p.Kind == KindModule || p.Kind == KindNode || p.Kind == KindController {
+		// Its own reply space, and nothing wider.
+		sub = append(sub, p.inbox())
+
+		// Acking a JetStream delivery is a publish to that consumer's own ack address — a
+		// different subject from anything the consumer subscribes. Without it every message a
+		// module received would be redelivered forever, refused by the permission list it already
+		// has (design 25 §4). Scoped to this principal's own consumer name, so it can ack its own
+		// deliveries and no other's.
+		pub = append(pub, "$JS.ACK."+consumerStream(p)+"."+consumerDurable(p)+".>")
+	}
+
+	sort.Strings(pub)
+	sort.Strings(sub)
+	return Permissions{
+		Publish:   pub,
+		Subscribe: sub,
+		// Only something that serves is ever answering. A pure consumer is granted nothing here.
+		AllowResponses: p.Kind == KindModule && (len(p.Serves) > 0 || len(p.Holds) > 0) ||
+			p.Kind == KindController,
+	}, nil
+}
+
+// seatSubject places a seat's verb under the kind of traffic it is.
+//
+// **The kind token is load-bearing, not decoration.** A stream is defined by a subject filter, so
+// without it a stream over a seat or a module's namespace would capture that namespace's *tool*
+// traffic too — and a tool call must never be persisted (design 25 §3: tools stay on core NATS).
+// Found while defining the streams: the first draft of design 29 had one namespace per module
+// with no kind, which reads well and cannot be filtered.
+//
+// A seat serving more than its current protocol version carries the version as a token
+// (design 29 §8): the seat stays one role, and v1 and v2 run beside each other until nothing is
+// bound to the old one.
+func seatSubject(s Seat, kind, verb string) string {
+	return "mesh.seat." + s.Name + "." + kind + "." + verb
+}
+
+// consumerStream and consumerDurable are the two halves of a consumer's identity, and they are
+// two functions because conflating them was a real bug.
+//
+// **A durable name may not contain a dot; an ack subject is built from two names that do.** The
+// server acknowledges on `$JS.ACK.<stream>.<consumer>.…`, so a single string "EVENTS.one_audit"
+// reads correctly inside the permission and is rejected as a consumer name — *nats: invalid
+// consumer name*. Caught against a running server, and worth the comment because the shape of
+// the failure if it had not been is the one design 25 §4 warns about: a consumer that cannot ack
+// has every message redelivered forever, and its permission list looks right while it happens.
+//
+// They are derived here, beside the permission that must match them, because two places deriving
+// the same name is how a module ends up unable to ack its own deliveries.
+// consumedSubject is where a consumed event lands, from the local pattern a module declared.
+//
+// **The mesh's wildcards become this transport's** (design 29 §1): `*` is one name on both, and `**`
+// — the rest — is `>` here. A module writes neither transport's spelling, so a manifest stays correct
+// when the wire changes, which is the whole reason names are local.
+//
+// `**` on its own is every event from every module: the emitter is any, the event is anything. An
+// audit logger wants exactly that and says so in one token.
+func consumedSubject(pattern string) (string, error) {
+	if pattern == catalogueTheRest {
+		return "mesh.mod.*.event.>", nil
+	}
+	emitter, event, named := strings.Cut(pattern, ".")
+	if !named || emitter == "" || event == "" {
+		return "", fmt.Errorf(
+			"%q does not name an emitter and an event: a consumed event is <emitter>.<event>, or "+
+				"%q for every event", pattern, catalogueTheRest)
+	}
+	if emitter == catalogueTheRest {
+		return "", fmt.Errorf("%q stands for the rest of a name, so it cannot name the emitter", catalogueTheRest)
+	}
+	// Each name is checked before it becomes a subject: a name carrying a dot would add a token and
+	// silently widen the permission, which is the whole reason safeSubject exists.
+	var out []string
+	for _, part := range strings.Split(event, ".") {
+		switch part {
+		case catalogueTheRest:
+			out = append(out, ">")
+		case "*":
+			out = append(out, "*")
+		default:
+			if !safeSubject.MatchString(part) {
+				return "", fmt.Errorf("%q cannot be part of a subject: it would widen the permission", part)
+			}
+			out = append(out, part)
+		}
+	}
+	if emitter != "*" && !safeSubject.MatchString(emitter) {
+		return "", fmt.Errorf("%q cannot name an emitter: it would widen the permission", emitter)
+	}
+	return "mesh.mod." + emitter + ".event." + strings.Join(out, "."), nil
+}
+
+// catalogueTheRest is the mesh's wildcard for "the rest of a name", duplicated from the catalogue
+// package for the one direction of dependency the build queue's name is duplicated for.
+const catalogueTheRest = "**"
+
+func consumerStream(p Principal) string {
+	switch p.Kind {
+	case KindModule:
+		return "EVENTS"
+	case KindNode:
+		return "NODES"
+	case KindController:
+		return "CONTROL"
+	}
+	return ""
+}
+
+func consumerDurable(p Principal) string {
+	switch p.Kind {
+	case KindModule:
+		return p.Node + "_" + p.Module
+	case KindNode:
+		return p.Node
+	case KindController:
+		return "controller"
+	}
+	return ""
+}
+
+// Server is everything the composed file needs that is not a principal.
+type Server struct {
+	// ClientPort carries TLS itself. There is no plaintext port beside it: a bus reachable
+	// without TLS is one a module can reach without TLS by mistake.
+	ClientPort     int
+	MonitoringPort int
+	TLSCert        string
+	TLSKey         string
+	TLSCA          string
+	// StoreDir is a host directory bind, not a named volume — issue 115 is resolved and converted
+	// four modules away from named volumes; the bus's own data is not the place to bring one back.
+	StoreDir string
+}
+
+// Compose renders the server's whole configuration. The order is stable and the output is
+// deterministic, because the file's digest is what the module's entrypoint watches to decide
+// whether to reload: a composer that reordered a map on each run would signal a reload every time
+// the controller restarted, for a file that had not changed.
+func Compose(s Server, principals []Principal) (string, error) {
+	sorted := append([]Principal(nil), principals...)
+	sort.Slice(sorted, func(i, j int) bool { return sorted[i].Username() < sorted[j].Username() })
+
+	var b strings.Builder
+	b.WriteString("# Composed by the mesh controller. Do not edit: the next composition overwrites it.\n")
+	b.WriteString("# Accounts and permissions are derived from what each module declares and nothing\n")
+	b.WriteString("# else (novox/hq ADR 0043, design 29 §2).\n\n")
+
+	fmt.Fprintf(&b, "port: %d\n", s.ClientPort)
+	fmt.Fprintf(&b, "http: 127.0.0.1:%d\n\n", s.MonitoringPort)
+
+	// **No `verify`, and it said `verify: true` until this configuration was run.** That setting
+	// makes the server demand a *client* certificate, and nothing in the mesh presents one: a host
+	// pins this server's exact certificate and authenticates with the password the mesh minted
+	// (ADR 0004, design 25 §4), and so does a module's runtime. With it on, every connection in the
+	// mesh is refused at the TLS handshake before any password is looked at, and the error —
+	// "client didn't provide a certificate" — reads as a fault in the client.
+	//
+	// TLS is still required: the block is what requires it, and verify only decides whether client
+	// certificates are checked.
+	b.WriteString("tls {\n")
+	fmt.Fprintf(&b, "  cert_file: %q\n", s.TLSCert)
+	fmt.Fprintf(&b, "  key_file: %q\n", s.TLSKey)
+	fmt.Fprintf(&b, "  ca_file: %q\n", s.TLSCA)
+	b.WriteString("}\n\n")
+
+	b.WriteString("jetstream {\n")
+	fmt.Fprintf(&b, "  store_dir: %q\n", s.StoreDir)
+	b.WriteString("}\n\n")
+
+	accounts, err := ComposeAccounts(sorted)
+	if err != nil {
+		return "", err
+	}
+	b.WriteString(accounts)
+	return b.String(), nil
+}
+
+// ComposeAccounts is the accounts block alone — every user, and nothing about the server.
+//
+// **This is the only part of the configuration the mesh writes, and the split is deliberate.** A
+// server's ports, its TLS paths and its store directory are properties of the container the module
+// raises: they live in its image and its mounts, and they change when it does. The controller has no
+// business knowing them, and a controller that did would have to be kept in step with a Dockerfile
+// it never sees. What only the mesh knows is *who may connect*, so that is what it writes, and the
+// module's own configuration includes it.
+//
+// Four things checked against a running server before this shape was committed to: a user in an
+// included file authenticates; an unknown user is refused, so the include is the whole authority
+// rather than an addition to something; a publish outside a user's grant is refused; and rewriting
+// this file alone and signalling a reload makes a new user appear **without dropping the connection
+// the mesh already has** — which is what makes every later account, permission or person change cost
+// nothing (task 1.2's payoff).
+func ComposeAccounts(principals []Principal) (string, error) {
+	sorted := append([]Principal(nil), principals...)
+	sort.Slice(sorted, func(i, j int) bool { return sorted[i].Username() < sorted[j].Username() })
+
+	var b strings.Builder
+	b.WriteString("# The mesh's users, composed by the controller. Do not edit: the next\n")
+	b.WriteString("# composition overwrites it. Permissions are derived from what each module\n")
+	b.WriteString("# declares and nothing else (novox/hq ADR 0043, design 29 §2).\n\n")
+
+	// One account for the mesh: accounts in NATS isolate subject spaces entirely, and the mesh is
+	// one space (design 25 §4). The cost of that — that permissions are the only isolation — is
+	// paid in the scoping of every inbox and every ack subject.
+	b.WriteString("accounts {\n  MESH {\n    users = [\n")
+	for _, p := range sorted {
+		perms, err := PermissionsFor(p)
+		if err != nil {
+			return "", err
+		}
+		if p.PasswordHash == "" {
+			return "", fmt.Errorf("%s has no password hash: a user without one is a user anybody is", p.Username())
+		}
+		fmt.Fprintf(&b, "      { user: %q, password: %q, permissions: {\n", p.Username(), p.PasswordHash)
+		fmt.Fprintf(&b, "        publish: { allow: [%s] }\n", quoted(perms.Publish))
+		fmt.Fprintf(&b, "        subscribe: { allow: [%s] }\n", quoted(perms.Subscribe))
+		if perms.AllowResponses {
+			b.WriteString("        allow_responses: { max: 1, ttl: \"1m\" }\n")
+		}
+		b.WriteString("      } }\n")
+	}
+	b.WriteString("    ]\n  }\n}\n")
+	return b.String(), nil
+}
+
+func quoted(values []string) string {
+	if len(values) == 0 {
+		return ""
+	}
+	out := make([]string, len(values))
+	for i, v := range values {
+		out[i] = fmt.Sprintf("%q", v)
+	}
+	return strings.Join(out, ", ")
+}

internal/broker/nats_golden_test.go

@@ -0,0 +1,49 @@
+package broker
+
+import (
+	"flag"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+var update = flag.Bool("update", false, "rewrite the golden composition")
+
+// The composed file is the mesh's whole authority model, so a change to it should be visible in a
+// review rather than inferred from a diff of Go. The fixture is also the exact text checked
+// against the real server's parser (`nats-server -t`), which is what says this syntax is the
+// server's and not one we invented.
+func TestTheComposedConfigMatchesTheGolden(t *testing.T) {
+	seat := Seat{Name: "telegram-sender", Accepts: []string{"send"}, Emits: []string{"delivered", "failed"}}
+	got, err := Compose(
+		Server{ClientPort: 4222, MonitoringPort: 8222, StoreDir: "/data",
+			TLSCert: "/tls/tls.crt", TLSKey: "/tls/tls.key", TLSCA: "/tls/ca.crt"},
+		[]Principal{
+			{Kind: KindController, PasswordHash: "$2a$11$cccccccccccccccccccccc"},
+			{Kind: KindEnrolment, Node: "one", PasswordHash: "$2a$11$eeeeeeeeeeeeeeeeeeeeee"},
+			{Kind: KindNode, Node: "one", PasswordHash: "$2a$11$nnnnnnnnnnnnnnnnnnnnnn"},
+			{Kind: KindModule, Node: "one", Module: "telegram", Holds: []Seat{seat},
+				Serves: []string{"status"}, PasswordHash: "$2a$11$tttttttttttttttttttttt"},
+			{Kind: KindModule, Node: "two", Module: "shop", Uses: []Seat{seat},
+				Emits: []string{"order.placed"}, PasswordHash: "$2a$11$ssssssssssssssssssssss"},
+			{Kind: KindModule, Node: "two", Module: "audit",
+				Consumes: []string{"shop.order.placed"}, PasswordHash: "$2a$11$aaaaaaaaaaaaaaaaaaaaaa"},
+		})
+	if err != nil {
+		t.Fatal(err)
+	}
+	golden := filepath.Join("testdata", "composed.conf")
+	if *update {
+		if err := os.WriteFile(golden, []byte(got), 0o644); err != nil {
+			t.Fatal(err)
+		}
+		return
+	}
+	want, err := os.ReadFile(golden)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got != string(want) {
+		t.Errorf("composition changed; re-run with -update and read the diff:\n%s", got)
+	}
+}

internal/broker/nats_test.go

@@ -0,0 +1,326 @@
+package broker
+
+import (
+	"strings"
+	"testing"
+)
+
+func has(t *testing.T, subjects []string, want string) {
+	t.Helper()
+	for _, s := range subjects {
+		if s == want {
+			return
+		}
+	}
+	t.Fatalf("expected %q among %v", want, subjects)
+}
+
+func hasNot(t *testing.T, subjects []string, unwanted string) {
+	t.Helper()
+	for _, s := range subjects {
+		if s == unwanted {
+			t.Fatalf("did not expect %q among %v", unwanted, subjects)
+		}
+	}
+}
+
+// A module's authority comes from its declaration and nothing else (novox/hq ADR 0043).
+func TestAModulePublishesOnlyWhatItEmits(t *testing.T) {
+	p := Principal{Kind: KindModule, Node: "one", Module: "billing",
+		Emits: []string{"order.placed"}, PasswordHash: "x"}
+	perms, err := PermissionsFor(p)
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, perms.Publish, "mesh.mod.billing.event.order.placed")
+	hasNot(t, perms.Publish, "mesh.mod.billing.>")
+	hasNot(t, perms.Publish, "mesh.mod.shipping.event.order.placed")
+}
+
+// The gap AMQP left open — an emitter granted the events exchange whole — is closed by per-subject
+// permissions. A module cannot publish under another module's name.
+func TestAModuleCannotPublishUnderAnothersName(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "billing",
+		Emits: []string{"order.placed"}, PasswordHash: "x"})
+	for _, p := range perms.Publish {
+		if strings.HasPrefix(p, "mesh.mod.") && !strings.HasPrefix(p, "mesh.mod.billing.") {
+			t.Fatalf("billing may publish %q, which is not its own namespace", p)
+		}
+	}
+}
+
+// A caller of a seat may publish what the seat accepts, and nothing else of it: not its outbound
+// events, and not a subscription to its inbound queue (design 29 §2).
+func TestUsingASeatIsPublishOnlyAndInboundOnly(t *testing.T) {
+	seat := Seat{Name: "telegram-sender", Accepts: []string{"send"}, Emits: []string{"delivered", "failed"}}
+	perms, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "shop",
+		Uses: []Seat{seat}, PasswordHash: "x"})
+	has(t, perms.Publish, "mesh.seat.telegram-sender.accept.send")
+	hasNot(t, perms.Publish, "mesh.seat.telegram-sender.event.delivered")
+	hasNot(t, perms.Subscribe, "mesh.seat.telegram-sender.accept.send")
+}
+
+// The holder is the mirror image: it consumes what the seat accepts and publishes what it emits.
+func TestHoldingASeatIsTheMirrorOfUsingIt(t *testing.T) {
+	seat := Seat{Name: "telegram-sender", Accepts: []string{"send"}, Emits: []string{"delivered", "failed"}}
+	perms, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "telegram",
+		Holds: []Seat{seat}, PasswordHash: "x"})
+	has(t, perms.Subscribe, "mesh.seat.telegram-sender.accept.send")
+	has(t, perms.Publish, "mesh.seat.telegram-sender.event.delivered")
+	hasNot(t, perms.Publish, "mesh.seat.telegram-sender.accept.send")
+}
+
+// Without an ack permission a durable consumer never really consumes: every message it receives is
+// redelivered forever, refused by the permission list it already has (design 25 §4).
+func TestAModuleMayAckItsOwnDeliveriesAndNoOthers(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "billing",
+		Consumes: []string{"shop.order.placed"}, PasswordHash: "x"})
+	has(t, perms.Publish, "$JS.ACK.EVENTS.one_billing.>")
+	hasNot(t, perms.Publish, "$JS.ACK.>")
+	hasNot(t, perms.Publish, "$JS.ACK.EVENTS.one_shop.>")
+}
+
+// With one account, inbox privacy is the permission list or it is nothing.
+func TestAnInboxIsScopedToItsOwner(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "billing", PasswordHash: "x"})
+	has(t, perms.Subscribe, "_INBOX.one.billing.>")
+	hasNot(t, perms.Subscribe, "_INBOX.>")
+	hasNot(t, perms.Subscribe, "_INBOX.one.shop.>")
+}
+
+// A responder answers on the caller's inbox, which it has no permission for. allow_responses is
+// what makes a scoped inbox workable at all — the authority is bounded by having been asked.
+func TestOnlySomethingThatServesMayAnswer(t *testing.T) {
+	serving, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "billing",
+		Serves: []string{"status"}, PasswordHash: "x"})
+	if !serving.AllowResponses {
+		t.Fatal("a module serving a tool cannot answer the caller's inbox")
+	}
+	consumer, _ := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "audit",
+		Consumes: []string{"shop.order.placed"}, PasswordHash: "x"})
+	if consumer.AllowResponses {
+		t.Fatal("a pure consumer was granted the right to answer, which nothing asked it to do")
+	}
+}
+
+// A host reaches its own node's control traffic and its own declaration, and nothing of any
+// other node's.
+func TestAHostIsConfinedToItsOwnNode(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindNode, Node: "one", PasswordHash: "x"})
+	has(t, perms.Publish, "mesh.control.one.>")
+	has(t, perms.Subscribe, "mesh.node.one.declare")
+	hasNot(t, perms.Subscribe, "mesh.node.two.declare")
+	hasNot(t, perms.Subscribe, "mesh.node.>")
+}
+
+// A leaked enrolment token is useless for anything but enrolling (design 25 §6).
+func TestTheEnrolmentUserCanOnlyEnrol(t *testing.T) {
+	perms, err := PermissionsFor(Principal{Kind: KindEnrolment, Node: "anchor", PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(perms.Publish) != 1 || perms.Publish[0] != "mesh.control.enrol" {
+		t.Fatalf("enrolment may publish %v", perms.Publish)
+	}
+	// Its own inbox and nothing else. **Nothing else** is the point: no declaration, no event, and
+	// no other machine's answer — and the inbox itself is needed, because a node that cannot
+	// subscribe one waits out its timeout against a mesh that answered.
+	if len(perms.Subscribe) != 1 || perms.Subscribe[0] != "_INBOX.enrol.anchor.>" {
+		t.Fatalf("enrolment may subscribe %v, which is not its own inbox alone", perms.Subscribe)
+	}
+}
+
+// An enrolment user that names no node is refused: its inbox would be an empty subject token, and
+// one that every nameless enrolment user shared — which is one machine reading the credentials
+// sealed to another.
+func TestAnEnrolmentUserWithoutANodeIsRefused(t *testing.T) {
+	if _, err := PermissionsFor(Principal{Kind: KindEnrolment, PasswordHash: "x"}); err == nil {
+		t.Fatal("an enrolment user with no node was composed, so its inbox is shared")
+	}
+}
+
+// A name that would widen a permission is refused rather than quietly stretching one.
+func TestANameThatWouldWidenAPermissionIsRefused(t *testing.T) {
+	for _, bad := range []string{"bill.ing", "billing.>", "*", "bil>ling"} {
+		if _, err := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: bad, PasswordHash: "x"}); err == nil {
+			t.Fatalf("%q was accepted as part of a subject", bad)
+		}
+	}
+}
+
+// The entrypoint reloads on the file's digest changing, so an unchanged mesh must compose an
+// identical file — otherwise every controller restart signals a reload of the whole bus.
+func TestComposingTwiceGivesTheSameBytes(t *testing.T) {
+	s := Server{ClientPort: 4222, MonitoringPort: 8222, StoreDir: "/data",
+		TLSCert: "/tls/tls.crt", TLSKey: "/tls/tls.key", TLSCA: "/tls/ca.crt"}
+	ps := []Principal{
+		{Kind: KindModule, Node: "two", Module: "shop", Emits: []string{"order.placed"}, PasswordHash: "b"},
+		{Kind: KindController, PasswordHash: "c"},
+		{Kind: KindModule, Node: "one", Module: "billing", Consumes: []string{"shop.order.placed"}, PasswordHash: "a"},
+	}
+	first, err := Compose(s, ps)
+	if err != nil {
+		t.Fatal(err)
+	}
+	shuffled := []Principal{ps[2], ps[0], ps[1]}
+	second, err := Compose(s, shuffled)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if first != second {
+		t.Fatal("composition is order-dependent; every controller restart would reload the bus")
+	}
+}
+
+// A user without a password is a user anybody is.
+func TestAUserWithoutAPasswordIsRefused(t *testing.T) {
+	_, err := Compose(Server{ClientPort: 4222}, []Principal{{Kind: KindController}})
+	if err == nil {
+		t.Fatal("composed a user with no password hash")
+	}
+}
+
+// A person reaches the mesh's tools from a workstation (design 25 §7). Their authority is a list
+// of tools and nothing else.
+func TestAPersonMayAskOnlyTheToolsTheyWereGiven(t *testing.T) {
+	perms, err := PermissionsFor(Principal{Kind: KindPerson, Module: "jo",
+		Invokes: []string{"shop.price", "telegram.status"}, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, perms.Publish, "mesh.mod.shop.tool.price")
+	has(t, perms.Publish, "mesh.mod.telegram.tool.status")
+	hasNot(t, perms.Publish, "mesh.mod.shop.tool.refund")
+	hasNot(t, perms.Publish, "mesh.mod.*.tool.>")
+}
+
+// An administrator gets every tool, which is a different grant and looks like one.
+func TestAnAdministratorMayAskAnyTool(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindPerson, Module: "jo",
+		Invokes: []string{"*"}, PasswordHash: "x"})
+	has(t, perms.Publish, "mesh.mod.*.tool.>")
+}
+
+// **Nothing but tools.** A person who could publish an event would be able to claim a module
+// said something; one who could publish control traffic would be a second controller.
+func TestAPersonReachesNothingButTools(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindPerson, Module: "jo",
+		Invokes: []string{"*"}, PasswordHash: "x"})
+	for _, p := range perms.Publish {
+		if !strings.Contains(p, ".tool.") {
+			t.Errorf("a person may publish %q, which is not a tool call", p)
+		}
+	}
+	for _, s := range perms.Subscribe {
+		if !strings.HasPrefix(s, "_INBOX.person.") {
+			t.Errorf("a person may subscribe %q; only their own inbox should be reachable", s)
+		}
+	}
+}
+
+// A person has no durable consumer, because nothing is delivered to a person — so no ack
+// subject, and an ack permission would be authority over something that does not exist.
+func TestAPersonHasNoAckSubject(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindPerson, Module: "jo",
+		Invokes: []string{"*"}, PasswordHash: "x"})
+	for _, p := range perms.Publish {
+		if strings.HasPrefix(p, "$JS.ACK") {
+			t.Errorf("a person was granted %q, and has no consumer to acknowledge", p)
+		}
+	}
+}
+
+// A person asks and is answered; they never answer. allow_responses would let a person reply to
+// a request — which, on a bus where anyone may serve a tool, is somebody impersonating a module.
+func TestAPersonMayNotAnswer(t *testing.T) {
+	perms, _ := PermissionsFor(Principal{Kind: KindPerson, Module: "jo",
+		Invokes: []string{"*"}, PasswordHash: "x"})
+	if perms.AllowResponses {
+		t.Fatal("a person may answer a request, which is impersonating a module")
+	}
+}
+
+// Two people do not share an inbox, or one would read the other's answers.
+func TestTwoPeopleDoNotShareAnInbox(t *testing.T) {
+	a, _ := PermissionsFor(Principal{Kind: KindPerson, Module: "jo", Invokes: []string{"*"}, PasswordHash: "x"})
+	b, _ := PermissionsFor(Principal{Kind: KindPerson, Module: "sam", Invokes: []string{"*"}, PasswordHash: "x"})
+	if a.Subscribe[0] == b.Subscribe[0] {
+		t.Fatalf("both read %s", a.Subscribe[0])
+	}
+}
+
+// A malformed grant is refused rather than widened into something that happens to parse.
+func TestAToolGrantThatNamesNoToolIsRefused(t *testing.T) {
+	if _, err := PermissionsFor(Principal{Kind: KindPerson, Module: "jo",
+		Invokes: []string{"shop"}, PasswordHash: "x"}); err == nil {
+		t.Fatal("a grant naming a module but no tool was accepted")
+	}
+}
+
+// The controller can answer an enrolment, and reach no other inbox.
+//
+// **`allow_responses` does not cover this and that is the trap.** It permits one reply to the reply
+// subject of a message the user received — and a message a JetStream consumer delivers has had that
+// field claimed for the consumer's own ack address, so the address the controller actually answers is
+// the one the request carried in its payload, which the server does not recognise as a reply subject
+// at all.
+//
+// Found against a real server, after a live test on an *unpermissioned* one had passed: every
+// enrolment on the mesh would have timed out while the controller logged success.
+func TestTheControllerCanAnswerAnEnrolmentAndReachNoOtherInbox(t *testing.T) {
+	ctl, err := PermissionsFor(Principal{Kind: KindController, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	enrolling, err := PermissionsFor(Principal{Kind: KindEnrolment, Node: "anchor", PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// Whatever the enrolling node waits on, the controller must be able to publish to.
+	if len(enrolling.Subscribe) != 1 {
+		t.Fatalf("an enrolling node subscribes %v, and this test knows only how to check one",
+			enrolling.Subscribe)
+	}
+	waitsOn := enrolling.Subscribe[0]
+	if !covers(ctl.Publish, waitsOn) {
+		t.Fatalf("the controller may publish %v, none of which reaches %s — so every enrolment on "+
+			"the mesh times out while the controller logs success", ctl.Publish, waitsOn)
+	}
+
+	// And nothing wider. A node's own inbox and a module's are not the controller's to write into:
+	// that is the blanket grant design 25 §4 refuses.
+	for _, other := range []string{"_INBOX.node.anchor.x", "_INBOX.one.shop.x", "_INBOX.person.ada.x"} {
+		if covers(ctl.Publish, other) {
+			t.Errorf("the controller can publish to %s, which is an inbox privacy the permission "+
+				"list is the only thing protecting", other)
+		}
+	}
+}
+
+// covers says whether any granted subject pattern admits one concrete subject, with NATS's own
+// wildcard meanings: `*` is one token, `>` is the rest.
+func covers(granted []string, subject string) bool {
+	want := strings.Split(subject, ".")
+	for _, pattern := range granted {
+		if admits(strings.Split(pattern, "."), want) {
+			return true
+		}
+	}
+	return false
+}
+
+func admits(pattern, subject []string) bool {
+	for i, token := range pattern {
+		if token == ">" {
+			return i < len(subject)
+		}
+		if i >= len(subject) {
+			return false
+		}
+		if token != "*" && token != subject[i] {
+			return false
+		}
+	}
+	return len(pattern) == len(subject)
+}

internal/broker/onnats.go

@@ -0,0 +1,91 @@
+package broker
+
+import (
+	"fmt"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/envfile"
+)
+
+// Whether this mesh's own traffic is on the bus being built.
+//
+// **One switch, read in one place** (novox/hq ADR 0116 step 5). Every seam the bus change went
+// behind ships both implementations, and until the rollout every one of them chooses the bus the
+// mesh runs on today. This is what the rollout flips, and it is deliberately a single fact rather
+// than a fact per component: a controller whose outbound is on one bus and whose inbound is on the
+// other is a mesh that hears nothing, and no test of either half would catch it.
+
+// NATSVar is where the controller finds the bus being built. Unset is the ordinary case and means
+// the mesh runs on the bus it has always run on.
+const NATSVar = "MESH_BUS_NATS"
+
+// OnNATS is the address of the bus being built, and whether the mesh is on it.
+//
+// Read from the node's own settings rather than baked in, for the reason the broker's address is
+// (novox/hq 04-ISSUES/102): an address recorded once does not follow a node's ports.
+func OnNATS() (address string, on bool, err error) {
+	address, err = envfile.Placed(NATSVar)
+	if err != nil {
+		return "", false, err
+	}
+	address = strings.TrimSpace(address)
+	if address == "" {
+		return "", false, nil
+	}
+	return address, true, nil
+}
+
+// CredentialIn reads the user and password out of a bus address, and the address without them.
+//
+// The controller's own credential arrives in its address, the way the old bus's does. Split out so the
+// controller can record a hash of what it is actually using: its user is created by the installer at a
+// bootstrap password, before the controller exists to mint one, and a composition that left itself out
+// would produce a bus the writer cannot connect to.
+func CredentialIn(address string) (user, password, bare string) {
+	at := strings.LastIndex(address, "@")
+	if at < 0 {
+		return "", "", address
+	}
+	scheme := ""
+	rest := address[:at]
+	if i := strings.Index(rest, "://"); i >= 0 {
+		scheme, rest = rest[:i+3], rest[i+3:]
+	}
+	user, password, _ = strings.Cut(rest, ":")
+	return user, password, scheme + address[at+1:]
+}
+
+// BareAddress is a bus address with any credential stripped, for something that only needs to know
+// whether a server is answering there.
+func BareAddress(address string) string {
+	_, _, bare := CredentialIn(address)
+	if bare == "" {
+		return address
+	}
+	if strings.Contains(bare, "://") {
+		return bare
+	}
+	return "nats://" + bare
+}
+
+// MustBeOneBus refuses a configuration that names both buses for the mesh's own traffic.
+//
+// **Both clients ship and that is the point; both being live is not.** The rollout moves every node
+// at once (ADR 0116 step 5): a mesh half on each is one where a declaration goes out on one bus and
+// the report comes back on the other, and nothing anywhere says so — every component would log
+// success. Refused at start, where it can be said in one sentence.
+func MustBeOneBus(amqp, nats string) error {
+	if strings.TrimSpace(amqp) != "" && strings.TrimSpace(nats) != "" {
+		return fmt.Errorf(
+			"this control plane is told about both buses (%s and %s) and can only be on one. A mesh "+
+				"half on each is one where a declaration goes out on one and the report comes back "+
+				"on the other, and every component reports success while it happens. The rollout "+
+				"moves every node at once: unset %s to stay, or unset %s to move",
+			AMQPVarName, NATSVar, NATSVar, AMQPVarName)
+	}
+	return nil
+}
+
+// AMQPVarName is the variable naming the bus the mesh runs on today. Named here rather than
+// imported from the link package, for the one direction of dependency.
+const AMQPVarName = "MESH_BROKER_AMQP"

internal/broker/onnats_test.go

@@ -0,0 +1,60 @@
+package broker
+
+import (
+	"strings"
+	"testing"
+)
+
+// Which bus the mesh is on is one fact, and being told about both is refused.
+//
+// **Not a warning.** A mesh half on each bus is one where a declaration goes out on one and the
+// report comes back on the other, and every component reports success while it happens — which is
+// the exact failure ADR 0074 exists to catch, arriving through configuration instead of through code.
+func TestBeingToldAboutBothBusesIsRefused(t *testing.T) {
+	err := MustBeOneBus("amqps://broker:5671/", "nats://bus:4222")
+	if err == nil {
+		t.Fatal("a control plane told about both buses was allowed to start")
+	}
+	// The remedy is in the words, because whoever reads this has to choose one and the wrong choice
+	// is a rollout half done.
+	for _, want := range []string{AMQPVarName, NATSVar, "unset"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the refusal does not mention %s: %v", want, err)
+		}
+	}
+}
+
+// One bus, or none, is ordinary. None is a control plane that publishes nothing and holds records,
+// which several of its own commands are.
+func TestOneBusOrNeitherIsAllowed(t *testing.T) {
+	for _, c := range []struct{ what, amqp, nats string }{
+		{"the bus the mesh runs on today", "amqps://broker:5671/", ""},
+		{"the bus being built", "", "nats://bus:4222"},
+		{"neither", "", ""},
+		{"neither, with whitespace for an address", "  ", "\t"},
+	} {
+		if err := MustBeOneBus(c.amqp, c.nats); err != nil {
+			t.Errorf("%s was refused: %v", c.what, err)
+		}
+	}
+}
+
+// The controller's own credential arrives in its address, and has to be readable out of it — its user
+// is created by the installer at a bootstrap password, before the controller exists to mint one.
+func TestACredentialIsReadOutOfABusAddress(t *testing.T) {
+	for _, c := range []struct{ in, user, password, bare string }{
+		{"nats://controller:secret@127.0.0.1:4222", "controller", "secret", "nats://127.0.0.1:4222"},
+		{"controller:secret@127.0.0.1:4222", "controller", "secret", "127.0.0.1:4222"},
+		{"nats://127.0.0.1:4222", "", "", "nats://127.0.0.1:4222"},
+		{"127.0.0.1:4222", "", "", "127.0.0.1:4222"},
+		// A password containing an at-sign: split on the last one, or the address becomes part of the
+		// credential and the connection goes somewhere nobody named.
+		{"nats://controller:a@b@127.0.0.1:4222", "controller", "a@b", "nats://127.0.0.1:4222"},
+	} {
+		user, password, bare := CredentialIn(c.in)
+		if user != c.user || password != c.password || bare != c.bare {
+			t.Errorf("%q read as %q/%q at %q; wanted %q/%q at %q",
+				c.in, user, password, bare, c.user, c.password, c.bare)
+		}
+	}
+}

internal/broker/raise.go

@@ -0,0 +1,73 @@
+package broker
+
+import "fmt"
+
+// Bringing the bus's own objects into being, in the one order that works.
+//
+// **Asserted on every start rather than created once at genesis.** A stream somebody deleted, a mesh
+// raised from a restored backup, or a bus whose data directory was replaced all have records and no
+// objects — and a node whose consumer is missing hears nothing while everything else about it looks
+// correct. Idempotence is the whole requirement, and the parts are already idempotent; this is the
+// order they have to be asked in.
+
+// Raiser is everything asserting the bus's objects needs of a connection to it.
+type Raiser interface {
+	Asserter
+	Ensurer
+}
+
+// Raise asserts the mesh's streams, the controller's own consumers, and one consumer per node.
+//
+// **The order is not a preference.** A consumer on a stream that does not exist is refused, and the
+// refusal names the stream rather than the order — so somebody reading it goes looking for a deleted
+// stream instead of a reversed pair of lines. Nodes last, because the one a node reads lives on a
+// stream the mesh's own set defines.
+func Raise(r Raiser, nodes []string) error {
+	if err := AssertMeshStreams(r); err != nil {
+		return err
+	}
+	if err := AssertMeshConsumers(r); err != nil {
+		return err
+	}
+	if err := AssertNodeConsumers(r, nodes); err != nil {
+		return err
+	}
+	return nil
+}
+
+// RaiseSeats asserts one work queue per declared seat, and the worker of whoever holds it.
+//
+// Separate from Raise because it is answered by a different question: the mesh's own objects exist
+// because the mesh does, and a seat's exist because a module declaring one was registered. Kept
+// beside it so the order is visible — a holder's worker needs the seat's stream, and a seat's stream
+// needs nothing.
+func RaiseSeats(r Raiser, seats []DeclaredSeat, holders map[string]Holder) error {
+	for _, s := range SeatStreams(seats) {
+		if err := r.EnsureStream(s); err != nil {
+			return fmt.Errorf("asserting the work queue for %s: %w", s.Name, err)
+		}
+	}
+	for _, s := range seats {
+		h, held := holders[s.Name]
+		if !held {
+			// **The stream exists and the consumer does not, on purpose.** Work queues until a
+			// holder appears, so installing the module a week after something started sending to it
+			// flushes the backlog instead of having lost it.
+			continue
+		}
+		c, needed := HolderConsumerFor(h.Node, h.Module, s)
+		if !needed {
+			continue
+		}
+		if err := r.EnsureConsumer(c); err != nil {
+			return fmt.Errorf("asserting how %s on %s works %s: %w", h.Module, h.Node, s.Name, err)
+		}
+	}
+	return nil
+}
+
+// Holder is which module on which machine holds a seat.
+type Holder struct {
+	Node   string
+	Module string
+}

internal/broker/raise_live_test.go

@@ -0,0 +1,196 @@
+package broker
+
+import (
+	"os"
+	"testing"
+
+	"github.com/nats-io/nats.go"
+)
+
+// Raising the bus's objects against a real server.
+//
+// The pure tests above say what is asked for and in what order. Only a server can say whether it
+// accepts them — and two of these are claims about the server's own behaviour that nothing else
+// could answer: that asserting twice changes nothing, and that a consumer really is bound to the one
+// subject its node is allowed to read.
+//
+//	docker run -d --rm --name t -p 14227:4222 nats:2.10-alpine -js
+//	MESH_TEST_NATS=nats://127.0.0.1:14227 go test ./internal/broker/ -run TestRaising
+
+func aLiveBus(t *testing.T) *JetStream {
+	t.Helper()
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	js, err := Dial(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+	// **Nothing is deleted here, deliberately.** These objects are the mesh's own and every live
+	// test in every package shares one server: a test that deleted a stream to get a clean slate
+	// took it out from under whatever was running beside it, and the failure landed in the other
+	// test as "stream not found" — which reads as a bug in the code under test. Raise is idempotent
+	// by requirement, so asserting against whatever is already there is both safe and the realistic
+	// case.
+	return js
+}
+
+// Every object the mesh's own traffic needs, accepted by a real server, and asserting again changes
+// nothing — which is the whole requirement, because this runs on every start.
+func TestRaisingTheBusIsAcceptedAndIdempotent(t *testing.T) {
+	js := aLiveBus(t)
+
+	if err := Raise(js, []string{"anchor", "laptop"}); err != nil {
+		t.Fatalf("a real server refused the mesh's own objects: %v", err)
+	}
+	// Twice, with nothing in between. A start that failed the second time is a controller that
+	// cannot restart.
+	if err := Raise(js, []string{"anchor", "laptop"}); err != nil {
+		t.Fatalf("asserting the bus's objects a second time failed, so a restart would: %v", err)
+	}
+	// And again with a machine that was not there before, which is what enrolling one is.
+	if err := Raise(js, []string{"anchor", "laptop", "workstation"}); err != nil {
+		t.Fatalf("a machine joining an already-raised bus was refused: %v", err)
+	}
+
+	for _, s := range MeshStreams() {
+		if _, err := js.Context().StreamInfo(s.Name); err != nil {
+			t.Errorf("stream %s is not there: %v", s.Name, err)
+		}
+	}
+	for _, c := range MeshConsumers() {
+		if _, err := js.Context().ConsumerInfo(c.Stream, c.Name); err != nil {
+			t.Errorf("the controller's consumer on %s is not there: %v", c.Stream, err)
+		}
+	}
+	for _, node := range []string{"anchor", "laptop", "workstation"} {
+		info, err := js.Context().ConsumerInfo("NODES", node)
+		if err != nil {
+			t.Errorf("%s has no way to hear its declaration: %v", node, err)
+			continue
+		}
+		// **Its own subject and no other node's.** A consumer filtered on anything wider is a node
+		// reading another machine's declaration, and its own ack grant would not cover it either.
+		if info.Config.FilterSubject != "mesh.node."+node+".declare" {
+			t.Errorf("%s's consumer reads %q", node, info.Config.FilterSubject)
+		}
+		if info.Config.AckPolicy != nats.AckExplicitPolicy {
+			t.Errorf("%s's consumer acknowledges on delivery, so a declaration it died applying is "+
+				"never sent again", node)
+		}
+	}
+}
+
+// The store window needs unlimited redelivery on CONTROL: the bound belongs to the controller, and a
+// server that dead-lettered first would discard the push the stream exists to protect.
+func TestTheControlConsumerDoesNotDeadLetterBeforeTheControllerGivesUp(t *testing.T) {
+	js := aLiveBus(t)
+	if err := Raise(js, nil); err != nil {
+		t.Fatal(err)
+	}
+	info, err := js.Context().ConsumerInfo("CONTROL", ControllerName)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if info.Config.MaxDeliver > 0 {
+		t.Fatalf("max-deliver is %d: a push held through a store restart would be dead-lettered "+
+			"before the controller finished deciding about it", info.Config.MaxDeliver)
+	}
+}
+
+// A role's work queue exists before anybody holds it, against a real server.
+//
+// **The queue before the holder is the point** (novox/hq ADR 0121): work queues until somebody arrives
+// to do it, so assigning a build machine a week after something started asking for builds flushes the
+// backlog instead of having lost it. A stream created at assignment would make "the holder is not here
+// yet" mean "your requests are gone".
+func TestRaisingAMeshRolesWorkQueue(t *testing.T) {
+	js := aLiveBus(t)
+	seats := []DeclaredSeat{{Name: "mesh-build-machine", Accepts: []string{"build"},
+		Emits: []string{"built"}}}
+	t.Cleanup(func() { _ = js.Context().DeleteStream("SEAT_MESH_BUILD_MACHINE") })
+
+	if err := RaiseSeats(js, seats, nil); err != nil {
+		t.Fatalf("a real server refused a role's work queue: %v", err)
+	}
+	info, err := js.Context().StreamInfo("SEAT_MESH_BUILD_MACHINE")
+	if err != nil {
+		t.Fatalf("the role has no work queue: %v", err)
+	}
+	if info.Config.Retention != nats.WorkQueuePolicy {
+		t.Errorf("the queue retains as %v: work a holder took must leave it, or the next holder does "+
+			"it again", info.Config.Retention)
+	}
+	if len(info.Config.Subjects) != 1 || info.Config.Subjects[0] != "mesh.seat.mesh-build-machine.accept.>" {
+		t.Errorf("it carries %v rather than the role's own inbound subjects", info.Config.Subjects)
+	}
+	// Nobody holds it, so there is no worker — and asserting again changes nothing, because this runs
+	// on every start.
+	if err := RaiseSeats(js, seats, nil); err != nil {
+		t.Fatalf("asserting a role's queue a second time failed, so a restart would: %v", err)
+	}
+
+	// And once somebody holds it, the worker appears on that same queue.
+	if err := RaiseSeats(js, seats, map[string]Holder{
+		"mesh-build-machine": {Node: "anchor", Module: "builder"},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := js.Context().ConsumerInfo("SEAT_MESH_BUILD_MACHINE",
+		"SEAT_MESH_BUILD_MACHINE_worker"); err != nil {
+		t.Fatalf("the holder got no worker on the role's queue: %v", err)
+	}
+}
+
+// **A consumer created after the fact still sees what came before it**, which is why the mesh needs no
+// catch-up at all on this bus (novox/hq 04-ISSUES/050).
+//
+// On the bus the mesh runs on today a queue receives only what is published after it is bound, so
+// everything built before the catalogue existed was announced to nobody — and on a fresh mesh that is
+// always the foundation, because those are the things the catalogue needed in order to exist. A whole
+// mechanism was built for it: the catalogue asks, the controller re-publishes.
+//
+// A stream is a log and a consumer is a position in it. A consumer created later starts at the
+// beginning by default, so the builds are simply there. Asked of a real server rather than assumed,
+// because the whole decision about whether to keep that mechanism rests on it.
+func TestAConsumerCreatedAfterwardsStillSeesWhatCameBefore(t *testing.T) {
+	js := aLiveBus(t)
+	if err := AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	if err := js.Context().PurgeStream("EVENTS"); err != nil {
+		t.Fatal(err)
+	}
+
+	// Genesis: things are built before anything is listening.
+	built := []string{"base", "store", "mesh-catalog"}
+	for _, m := range built {
+		if _, err := js.Context().Publish("mesh.seat.mesh-build-machine.event.built",
+			[]byte(`{"module":"`+m+`"}`)); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	// Now the catalogue is installed and the controller creates its consumer.
+	c, ok := ConsumerFor(Principal{Kind: KindModule, Node: "one", Module: "mesh-catalog",
+		Watches: []Seat{{Name: "mesh-build-machine", Emits: []string{"built"}}}, PasswordHash: "x"})
+	if !ok {
+		t.Fatal("a module that watches a role got no consumer")
+	}
+	t.Cleanup(func() { _ = js.Context().DeleteConsumer(c.Stream, c.Name) })
+	if err := js.EnsureConsumer(c); err != nil {
+		t.Fatal(err)
+	}
+
+	info, err := js.Context().ConsumerInfo(c.Stream, c.Name)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if info.NumPending != uint64(len(built)) {
+		t.Fatalf("a consumer created after %d builds has %d waiting for it — if this is 0 the mesh "+
+			"does need a catch-up after all, and the reasoning for deleting it is wrong",
+			len(built), info.NumPending)
+	}
+}

internal/broker/readiness.go

@@ -0,0 +1,142 @@
+package broker
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// Whether a mesh could move its bus, and what is missing if not.
+//
+// **Asked before anything moves, and answerable from records alone.** The rollout moves every node at
+// once (novox/hq ADR 0116 step 5), so there is no partial state to inspect afterwards and no half to
+// roll back: either the mesh was ready or it was not. That makes a readiness question the most
+// valuable thing here — it costs nothing, it can be asked of a running mesh any number of times, and
+// every answer is a thing somebody can go and fix.
+//
+// Deliberately pure. It is handed what the mesh knows and returns sentences; nothing here connects to
+// anything, so it can be asked on a workstation about a mesh it has never reached.
+
+// Readiness is what the mesh knows about its own ability to move.
+type Readiness struct {
+	// TheBus is the address the mesh's own traffic would move to, empty when nothing names one.
+	TheBus string
+	// ServerStanding is whether a bus is reachable at that address, as somebody checked.
+	ServerStanding bool
+	// Holder is the node running the module that holds the bus seat, empty when nothing does.
+	Holder string
+	// AccountsComposed is whether that node has been sent the composed user list.
+	AccountsComposed bool
+	// Nodes is every machine the mesh knows.
+	Nodes []string
+	// Credentialled is which of them has a credential for the new bus.
+	Credentialled map[string]bool
+	// Modules is every assigned module, as `<node>/<module>`.
+	Modules []string
+	// ModuleCredentialled is which of those has one.
+	ModuleCredentialled map[string]bool
+	// StillOnTheOldBus is whether anything of the mesh's own still needs the bus it is leaving —
+	// which is not a reason to stop, because that broker stays as an ordinary provider of `amqp`
+	// (ADR 0119). Recorded so nobody reads the move as a retirement.
+	OldBusHasOtherClients bool
+}
+
+// NotReady is every reason this mesh cannot move its bus yet, in the order somebody would fix them.
+//
+// Empty means ready. **Each entry names one thing and what to do about it**, because a readiness check
+// that says "not ready" is a check nobody can act on — and this is read at the point where the next
+// step is irreversible.
+func NotReady(r Readiness) []string {
+	var why []string
+
+	if strings.TrimSpace(r.TheBus) == "" {
+		why = append(why, "nothing names the bus to move to: set "+NATSVar+" on the control node "+
+			"to the address the new server answers on")
+	}
+	if !r.ServerStanding {
+		why = append(why, "no bus is answering at that address. Step 2 of the change raises it beside "+
+			"the one the mesh is on, carrying nothing — assign the module that holds "+
+			"mesh-broker and push the machine that runs it")
+	}
+	if r.Holder == "" {
+		why = append(why, "no machine holds mesh-broker, so nothing would compose the bus's user "+
+			"list. Assign the module that claims it")
+	} else if !r.AccountsComposed {
+		why = append(why, fmt.Sprintf(
+			"%s holds mesh-broker and has not been sent the composed user list, so the bus would "+
+				"refuse every connection. `push %s`", r.Holder, r.Holder))
+	}
+
+	// A node with no credential cannot come back after the move, and a node that cannot come back is
+	// a machine the mesh has lost until somebody goes to it.
+	var missing []string
+	for _, n := range r.Nodes {
+		if !r.Credentialled[n] {
+			missing = append(missing, n)
+		}
+	}
+	sort.Strings(missing)
+	if len(missing) > 0 {
+		why = append(why, fmt.Sprintf(
+			"%d machine(s) have no credential for the new bus and would not come back: %s. Each needs "+
+				"one minted before the move, not after — after, there is no bus to ask over",
+			len(missing), strings.Join(missing, ", ")))
+	}
+
+	// A module without one keeps running and stops being reachable, which is a smaller fault and still
+	// one somebody should choose rather than discover.
+	var quiet []string
+	for _, m := range r.Modules {
+		if !r.ModuleCredentialled[m] {
+			quiet = append(quiet, m)
+		}
+	}
+	sort.Strings(quiet)
+	if len(quiet) > 0 {
+		why = append(why, fmt.Sprintf(
+			"%d module(s) have no credential for the new bus: %s. Each keeps serving and stops "+
+				"answering tools and hearing events until it is issued one",
+			len(quiet), strings.Join(quiet, ", ")))
+	}
+
+	return why
+}
+
+// WhatMoves is what the rollout would do, in order, for somebody reading before they commit.
+//
+// **Written out rather than summarised.** This is the one step with nothing to inspect afterwards, so
+// the last useful moment to disagree with it is while reading this.
+func WhatMoves(r Readiness) []string {
+	out := []string{
+		fmt.Sprintf("compose the bus's user list and send it to %s", holderOr(r.Holder)),
+		fmt.Sprintf("move this control plane to %s, and confirm it is heard", busOr(r.TheBus)),
+	}
+	nodes := append([]string(nil), r.Nodes...)
+	sort.Strings(nodes)
+	for _, n := range nodes {
+		out = append(out, fmt.Sprintf("move %s, and confirm it reports", n))
+	}
+	if len(r.Modules) > 0 {
+		out = append(out, fmt.Sprintf("move %d module runtime(s), and confirm each answers",
+			len(r.Modules)))
+	}
+	if r.OldBusHasOtherClients {
+		out = append(out, "leave the old broker running: it stays an ordinary provider of `amqp` for "+
+			"whatever else uses it (ADR 0119), and this move is not its retirement")
+	}
+	return out
+}
+
+func holderOr(node string) string {
+	if node == "" {
+		return "whichever machine holds mesh-broker"
+	}
+	return node
+}
+
+func busOr(address string) string {
+	if address == "" {
+		return "the new bus"
+	}
+	return address
+}

internal/broker/readiness_test.go

@@ -0,0 +1,97 @@
+package broker
+
+import (
+	"strings"
+	"testing"
+)
+
+// Whether a mesh could move its bus.
+//
+// Every case here is a way of moving that leaves something behind, and the one that matters most is a
+// machine with no credential: after the move there is no bus to ask it over, so it is lost until
+// somebody walks to it.
+
+func aMeshReadyToMove() Readiness {
+	return Readiness{
+		TheBus: "nats://127.0.0.1:5671", ServerStanding: true,
+		Holder: "anchor", AccountsComposed: true,
+		Nodes:               []string{"anchor", "laptop"},
+		Credentialled:       map[string]bool{"anchor": true, "laptop": true},
+		Modules:             []string{"anchor/gitea"},
+		ModuleCredentialled: map[string]bool{"anchor/gitea": true},
+	}
+}
+
+func TestAMeshWithEverythingInPlaceIsReady(t *testing.T) {
+	if why := NotReady(aMeshReadyToMove()); len(why) != 0 {
+		t.Fatalf("a mesh with everything in place was refused: %v", why)
+	}
+}
+
+// **A machine with no credential is the one that must stop this.** It keeps running and cannot come
+// back, and there is no bus left to tell it anything over — so the remedy has to happen before, and
+// the message says so.
+func TestAMachineWithNoCredentialStopsTheMove(t *testing.T) {
+	r := aMeshReadyToMove()
+	r.Credentialled = map[string]bool{"anchor": true}
+
+	why := NotReady(r)
+	if len(why) == 0 {
+		t.Fatal("a machine that could not come back did not stop the move")
+	}
+	said := strings.Join(why, "\n")
+	if !strings.Contains(said, "laptop") {
+		t.Errorf("the refusal does not name the machine: %s", said)
+	}
+	if !strings.Contains(said, "before the move") {
+		t.Errorf("the refusal does not say the remedy comes first: %s", said)
+	}
+}
+
+// A bus nobody has raised, a seat nobody holds, and a user list nobody has been sent: each stops it,
+// and each names its own next step, because "not ready" that cannot be acted on is not an answer.
+func TestEachThingMissingNamesItsOwnRemedy(t *testing.T) {
+	for _, c := range []struct {
+		what   string
+		break_ func(*Readiness)
+		says   string
+	}{
+		{"no address", func(r *Readiness) { r.TheBus = "" }, NATSVar},
+		{"no server", func(r *Readiness) { r.ServerStanding = false }, "carrying nothing"},
+		{"no holder", func(r *Readiness) { r.Holder = "" }, "mesh-broker"},
+		{"no user list", func(r *Readiness) { r.AccountsComposed = false }, "push anchor"},
+		{"a module with none", func(r *Readiness) {
+			r.ModuleCredentialled = map[string]bool{}
+		}, "anchor/gitea"},
+	} {
+		r := aMeshReadyToMove()
+		c.break_(&r)
+		why := NotReady(r)
+		if len(why) == 0 {
+			t.Errorf("%s did not stop the move", c.what)
+			continue
+		}
+		if !strings.Contains(strings.Join(why, "\n"), c.says) {
+			t.Errorf("%s: the refusal does not mention %q: %v", c.what, c.says, why)
+		}
+	}
+}
+
+// What the move would do is written out rather than summarised, because this is the one step with
+// nothing to inspect afterwards — so reading it is the last chance to disagree.
+func TestWhatMovesNamesEveryMachineAndSaysTheOldBrokerStays(t *testing.T) {
+	r := aMeshReadyToMove()
+	r.OldBusHasOtherClients = true
+	steps := strings.Join(WhatMoves(r), "\n")
+
+	for _, want := range []string{"anchor", "laptop", "user list", "module runtime"} {
+		if !strings.Contains(steps, want) {
+			t.Errorf("the plan does not mention %q:\n%s", want, steps)
+		}
+	}
+	// Said explicitly, so nobody reads the move as switching the old broker off — it stays serving
+	// whatever else uses it, and that is a decision already taken.
+	if !strings.Contains(steps, "not its retirement") {
+		t.Errorf("the plan does not say the old broker stays:\n%s", steps)
+	}
+}

internal/broker/streams.go

@@ -0,0 +1,238 @@
+package broker
+
+import (
+	"fmt"
+	"sort"
+)
+
+// The mesh's own streams.
+//
+// **These four and no more** (novox/hq ADR 0116 task 1.4, as revised by ADR 0118). An earlier
+// reading had the controller create *every* stream at genesis, from a fixed set. That is only the
+// mesh's own half: a seat's streams are created when the module declaring it is registered, and a
+// module's durable consumers when it is assigned — neither of which has happened at genesis. What
+// is here is the foundation, which exists before any module does.
+//
+// The controller is the only writer of stream definitions (design 25 §3). A module declares
+// nothing about them and cannot reach the JetStream API to make one.
+
+// Retention is how a stream decides what to keep, which is the whole of what distinguishes the
+// mesh's four relationships on the wire (design 29 §4).
+type Retention string
+
+const (
+	// RetentionWorkQueue: a message is removed once a consumer acknowledges it. Exactly one
+	// worker does the work, and a worker that dies has its message redelivered.
+	RetentionWorkQueue Retention = "workqueue"
+	// RetentionLastPerSubject: only the newest message on each subject survives. This is the
+	// state shape — a node that was away gets exactly the current declaration and nothing older.
+	RetentionLastPerSubject Retention = "last_per_subject"
+	// RetentionLimits: kept until it ages or the stream fills. Events, where a subscriber that
+	// was down catches up and nobody is obliged to act.
+	RetentionLimits Retention = "limits"
+)
+
+// A Stream is one of the mesh's own, as the controller asserts it.
+type Stream struct {
+	Name      string
+	Subjects  []string
+	Retention Retention
+	// MaxAge in seconds, zero for unbounded. Per stream — JetStream has no per-subject age,
+	// which is why differing retention between modules would mean a stream each.
+	MaxAge int
+	// MaxMsgsPerSubject caps each subject independently, so one noisy emitter cannot push
+	// another's events out of a shared stream. Verified: with a cap of 3, ten messages on one
+	// subject and one on another leave four in the stream, not three.
+	MaxMsgsPerSubject int
+	// Why is carried into the assertion so an operator reading the server's own state finds the
+	// reason there, rather than only in a repository they may not have.
+	Why string
+}
+
+// MeshStreams is the foundation set, in the order a person reads it.
+//
+// **CONTROL names its subjects rather than taking `mesh.control.>`**, because heartbeats live
+// under that prefix and must not be persisted: a lost heartbeat is the next heartbeat, and a
+// stream of them is a stream of the least valuable messages the mesh sends, competing for the
+// same retention as the ones that matter.
+//
+// **EVENTS filters on the `event` token**, which is the reason that token exists. A module's
+// namespace carries both its events and its tool calls; a filter of `mesh.mod.*.>` would persist
+// every tool invocation in the mesh, and a tool call must never be persisted (design 25 §3 keeps
+// tools on core NATS, where a lost call is a timeout the caller already handles).
+func MeshStreams() []Stream {
+	return []Stream{
+		{
+			Name: "CONTROL",
+			// A build's outcome is no longer here: it is the build-machine seat's own event, so one
+			// publish reaches whoever asked, the controller and the catalogue (novox/hq ADR 0121).
+			Subjects:  []string{"mesh.control.*.report", "mesh.control.enrol"},
+			Retention: RetentionWorkQueue,
+			Why: "the store-window guarantee (ADR 0083): the controller naks with a delay while its " +
+				"store is away and the message is redelivered; nothing is dropped",
+		},
+		{
+			Name:      "NODES",
+			Subjects:  []string{"mesh.node.*.declare"},
+			Retention: RetentionLastPerSubject,
+			Why: "one declaration per node, always the newest; a node that sees sequence n refuses " +
+				"n-1 by construction (issue 107)",
+		},
+		{
+			Name: "EVENTS",
+			// A seat's own events ride here too: they are 1:many like any event, and the
+			// `event` token keeps them clear of both the seat's work queue (`accept`) and its
+			// tools (`tool`), which must not be persisted.
+			Subjects:          []string{"mesh.mod.*.event.>", "mesh.seat.*.event.>"},
+			Retention:         RetentionLimits,
+			MaxAge:            7 * 24 * 60 * 60,
+			MaxMsgsPerSubject: 10000,
+			Why: "a subscriber that was down catches up; tool traffic under the same prefix is " +
+				"excluded by the event token; per-subject caps keep a noisy emitter from " +
+				"evicting a quiet one without splitting the stream",
+		},
+	}
+}
+
+// An Asserter is the part of a JetStream connection stream assertion needs. Narrow on purpose: it
+// keeps this testable without a server, and keeps the client library out of everything that only
+// wants to know what the streams are.
+type Asserter interface {
+	// EnsureStream creates the stream if absent and updates it to match if present. It must be
+	// idempotent: the controller asserts on every start, not only at genesis.
+	EnsureStream(s Stream) error
+}
+
+// AssertMeshStreams brings the foundation set into being, in order, and says which one failed
+// rather than that something did.
+//
+// Asserted on every start rather than created once at genesis, because a stream that was deleted,
+// or a mesh raised from a restored backup, must converge rather than run without the guarantee
+// its messages assume. Idempotence is the whole requirement.
+func AssertMeshStreams(a Asserter) error {
+	for _, s := range MeshStreams() {
+		if err := a.EnsureStream(s); err != nil {
+			return fmt.Errorf("asserting stream %s: %w", s.Name, err)
+		}
+	}
+	return nil
+}
+
+// Overlaps reports subject filters claimed by more than one stream.
+//
+// **Corrected against the server**: an earlier version of this comment said NATS accepts
+// overlapping streams and stores the message twice. It does not — it refuses the second stream
+// with "subjects overlap with an existing stream" (verified against nats-server 2.10). The check
+// still earns its place, for a different reason: the server's refusal arrives when the controller
+// is applying, naming one stream, at a moment when the mesh is half-configured. This one arrives
+// where the set is written, names both, and cannot reach a running mesh.
+//
+// It also decides a design question. Because overlap is refused rather than merged, a shared
+// EVENTS stream and a per-module stream cannot coexist — the module's would be refused — so
+// "one stream for most, its own for a module that wants different retention" is not an option
+// the server allows. It is all of one or all of the other.
+func Overlaps() []string {
+	seen := map[string]string{}
+	var clashes []string
+	for _, s := range MeshStreams() {
+		for _, subject := range s.Subjects {
+			if first, ok := seen[subject]; ok {
+				clashes = append(clashes, fmt.Sprintf("%s and %s both claim %s", first, s.Name, subject))
+				continue
+			}
+			seen[subject] = s.Name
+		}
+	}
+	sort.Strings(clashes)
+	return clashes
+}
+
+// The mesh's own consumers.
+//
+// A seat's streams and a module's consumers are derived from declarations (derived.go). These two
+// are not: **the controller is not a module and files no manifest**, so its authority and its
+// subscriptions cannot come from a declaration that does not exist. They are named here, where the
+// mesh's own streams are named, and narrowly — a controller subscribing `mesh.mod.*.event.>` would
+// hear every event in the mesh, which it has no business doing and which would make its permission
+// list stop explaining anything.
+
+// ControllerName is the controller's durable consumer on each stream it reads, and the name its
+// ack subject is derived from (nats.go: `$JS.ACK.<stream>.controller.>`).
+const ControllerName = "controller"
+
+// ControllerFollows are the events the controller reacts to: the catalogue saying a module's
+// current version moved, and a catalogue that has just started saying it may have missed builds.
+//
+// **Derived the same way a module's subscription is**, from the emitter and the bare local event
+// name, rather than written out. They were written out while the catalogue still spelled its events
+// as the old bus's routing keys, and the moment those were converted (novox/hq 04-ISSUES/127) a
+// hard-coded pair became a controller listening to a subject nothing publishes — the same fault, from
+// the other side. Deriving them means the conversion could not leave these behind.
+var ControllerFollows = []string{
+	moduleEventSubject("mesh-catalog", "upgraded"),
+	moduleEventSubject("mesh-catalog", "catching-up"),
+	// A build's outcome, which is the build-machine role's own event now (ADR 0121) rather than a
+	// message on the control branch. Same three audiences, one publish: whoever asked, this, and the
+	// catalogue.
+	seatEventSubject("mesh-build-machine", "built"),
+}
+
+// moduleEventSubject is where one module's event lands. The same derivation PermissionsFor uses, so
+// what the controller subscribes and what the emitter is permitted to publish cannot drift apart.
+func moduleEventSubject(module, event string) string {
+	return "mesh.mod." + module + ".event." + event
+}
+
+// seatEventSubject is where a role's own event lands, derived the same way a holder's permission is.
+func seatEventSubject(seat, verb string) string {
+	return "mesh.seat." + seat + ".event." + verb
+}
+
+// MeshConsumers is what the controller consumes, in the order a person reads it.
+//
+// **Unlimited redelivery on CONTROL, deliberately.** The store window's bound is the controller's,
+// not the server's (window.go): a message is held with a nak-and-delay until the controller either
+// takes it or gives up and says so. A max-deliver here would dead-letter a push that was being
+// held through a store restart — the exact message the stream exists to protect — some minutes
+// before the controller had finished deciding about it.
+func MeshConsumers() []Consumer {
+	return []Consumer{
+		{
+			Name:           ControllerName,
+			Stream:         "CONTROL",
+			Push:           true,
+			AckWaitSeconds: 30,
+			Why: "the controller is the single consumer of what nodes say; explicit ack and no " +
+				"max-deliver, because the store window's bound is the controller's own",
+		},
+		{
+			Name:           ControllerName,
+			Stream:         "EVENTS",
+			Filters:        ControllerFollows,
+			Push:           true,
+			AckWaitSeconds: 30,
+			MaxDeliver:     5,
+			Why: "the two events the mesh's own controller reacts to; after max-deliver it " +
+				"dead-letters, because an announcement it cannot act on will not become actionable",
+		},
+	}
+}
+
+// Ensurer is the part of a JetStream connection consumer assertion needs, narrow for the reason
+// Asserter is.
+type Ensurer interface {
+	EnsureConsumer(c Consumer) error
+}
+
+// AssertMeshConsumers brings the controller's own consumers into being, and says which one failed.
+//
+// After the streams, necessarily: a consumer on a stream that does not exist is refused, and the
+// refusal names the stream rather than the order.
+func AssertMeshConsumers(e Ensurer) error {
+	for _, c := range MeshConsumers() {
+		if err := e.EnsureConsumer(c); err != nil {
+			return fmt.Errorf("asserting consumer %s on %s: %w", c.Name, c.Stream, err)
+		}
+	}
+	return nil
+}

internal/broker/streams_test.go

@@ -0,0 +1,235 @@
+package broker
+
+import (
+	"errors"
+	"strings"
+	"testing"
+)
+
+type recorder struct {
+	seen []Stream
+	fail string
+}
+
+func (r *recorder) EnsureStream(s Stream) error {
+	if s.Name == r.fail {
+		return errors.New("refused")
+	}
+	r.seen = append(r.seen, s)
+	return nil
+}
+
+// The controller asserts on every start, not only at genesis: a stream that was deleted, or a mesh
+// raised from a backup, must converge rather than run without the guarantee its messages assume.
+func TestAssertingTwiceIsTheSameAsOnce(t *testing.T) {
+	a, b := &recorder{}, &recorder{}
+	if err := AssertMeshStreams(a); err != nil {
+		t.Fatal(err)
+	}
+	if err := AssertMeshStreams(a); err != nil {
+		t.Fatal(err)
+	}
+	if err := AssertMeshStreams(b); err != nil {
+		t.Fatal(err)
+	}
+	if len(a.seen) != 2*len(b.seen) {
+		t.Fatalf("asserted %d then %d; assertion is not repeatable", len(a.seen), len(b.seen))
+	}
+}
+
+func TestAFailedAssertionNamesItsStream(t *testing.T) {
+	err := AssertMeshStreams(&recorder{fail: "NODES"})
+	if err == nil || !strings.Contains(err.Error(), "NODES") {
+		t.Fatalf("got %v, which does not say which stream failed", err)
+	}
+}
+
+// Two streams matching one subject is accepted by NATS and stores the message twice under two
+// retentions. Nothing reports that, so it is refused where the set is written.
+func TestNoTwoStreamsClaimTheSameSubject(t *testing.T) {
+	if clashes := Overlaps(); len(clashes) != 0 {
+		t.Fatalf("overlapping subject filters: %v", clashes)
+	}
+}
+
+// A heartbeat under mesh.control.> must not be persisted: a lost one is the next one, and a
+// stream of them competes for retention with the messages that matter.
+func TestHeartbeatsAreNotInTheControlStream(t *testing.T) {
+	for _, s := range MeshStreams() {
+		for _, subject := range s.Subjects {
+			if subject == "mesh.control.>" || strings.Contains(subject, "alive") {
+				t.Fatalf("stream %s claims %q, which captures heartbeats", s.Name, subject)
+			}
+		}
+	}
+}
+
+// The reason the kind token exists: a filter over a module's whole namespace would persist every
+// tool call in the mesh.
+func TestTheEventsStreamDoesNotCaptureToolCalls(t *testing.T) {
+	var events Stream
+	for _, s := range MeshStreams() {
+		if s.Name == "EVENTS" {
+			events = s
+		}
+	}
+	// Nothing a tool call rides may match any of the filters — a module's or a seat's.
+	for _, tool := range []string{
+		"mesh.mod.billing.tool.status",
+		"mesh.seat.telegram-sender.tool.status",
+		"mesh.seat.telegram-sender.accept.send", // work, not an event: its own stream
+	} {
+		for _, f := range events.Subjects {
+			if subjectMatches(f, tool) {
+				t.Fatalf("%q matches the events filter %q, so it would be persisted here", tool, f)
+			}
+		}
+	}
+	// And both kinds of event do match.
+	for _, event := range []string{
+		"mesh.mod.billing.event.order.placed",
+		"mesh.seat.telegram-sender.event.delivered",
+	} {
+		matched := false
+		for _, f := range events.Subjects {
+			if subjectMatches(f, event) {
+				matched = true
+			}
+		}
+		if !matched {
+			t.Fatalf("%q matches no events filter, so nothing would keep it", event)
+		}
+	}
+}
+
+// subjectMatches is NATS subject matching, enough for these filters: `*` is one token, `>` is the
+// rest.
+func subjectMatches(filter, subject string) bool {
+	f, s := strings.Split(filter, "."), strings.Split(subject, ".")
+	for i, tok := range f {
+		if tok == ">" {
+			return i <= len(s)
+		}
+		if i >= len(s) {
+			return false
+		}
+		if tok != "*" && tok != s[i] {
+			return false
+		}
+	}
+	return len(f) == len(s)
+}
+
+// Each relationship's retention is the thing that makes it what it is (design 29 §4).
+func TestEachStreamCarriesTheRetentionItsShapeNeeds(t *testing.T) {
+	want := map[string]Retention{
+		"CONTROL": RetentionWorkQueue,
+		"NODES":   RetentionLastPerSubject,
+		"EVENTS":  RetentionLimits,
+	}
+	got := map[string]Retention{}
+	for _, s := range MeshStreams() {
+		got[s.Name] = s.Retention
+		if s.Why == "" {
+			t.Errorf("stream %s says no reason it exists", s.Name)
+		}
+	}
+	if len(got) != len(want) {
+		t.Fatalf("the foundation set is %v", got)
+	}
+	for name, r := range want {
+		if got[name] != r {
+			t.Errorf("%s retains as %q, expected %q", name, got[name], r)
+		}
+	}
+}
+
+// The order the bus's objects are asserted in, because getting it wrong is a refusal that names the
+// wrong thing: a consumer on a stream that does not exist is refused naming the *stream*, so
+// somebody reading it goes looking for a deletion instead of a reversed pair of lines.
+func TestTheBusesObjectsAreAssertedStreamsBeforeConsumers(t *testing.T) {
+	r := &recording{}
+	if err := Raise(r, []string{"anchor", "laptop"}); err != nil {
+		t.Fatal(err)
+	}
+
+	// Every stream before every consumer.
+	firstConsumer := -1
+	for i, step := range r.steps {
+		if strings.HasPrefix(step, "consumer ") && firstConsumer < 0 {
+			firstConsumer = i
+		}
+		if strings.HasPrefix(step, "stream ") && firstConsumer >= 0 {
+			t.Fatalf("a stream was asserted after a consumer: %v", r.steps)
+		}
+	}
+	if firstConsumer < 0 {
+		t.Fatalf("no consumer was asserted: %v", r.steps)
+	}
+
+	// And every node got one, named after it — without which that node hears nothing while
+	// everything else about it looks correct.
+	for _, node := range []string{"anchor", "laptop"} {
+		if !containsStep(r.steps, "consumer NODES/"+node) {
+			t.Errorf("%s was given no way to hear its declaration: %v", node, r.steps)
+		}
+	}
+	// And the controller its own, on both streams it reads.
+	for _, want := range []string{"consumer CONTROL/controller", "consumer EVENTS/controller"} {
+		if !containsStep(r.steps, want) {
+			t.Errorf("the controller is missing %s: %v", want, r.steps)
+		}
+	}
+}
+
+// A seat's work queue is asserted whether or not anybody holds it; the holder's worker only when
+// somebody does. **The stream without the consumer is the point**: work queues until a holder
+// appears, so installing the module later flushes the backlog instead of having lost it.
+func TestASeatsQueueExistsBeforeItsHolderDoes(t *testing.T) {
+	seats := []DeclaredSeat{{Name: "telegram-sender", Accepts: []string{"send"}}}
+
+	unheld := &recording{}
+	if err := RaiseSeats(unheld, seats, nil); err != nil {
+		t.Fatal(err)
+	}
+	if !containsStep(unheld.steps, "stream SEAT_TELEGRAM_SENDER") {
+		t.Fatalf("a declared seat got no work queue: %v", unheld.steps)
+	}
+	for _, step := range unheld.steps {
+		if strings.HasPrefix(step, "consumer ") {
+			t.Fatalf("a seat nobody holds got a worker: %v", unheld.steps)
+		}
+	}
+
+	held := &recording{}
+	if err := RaiseSeats(held, seats, map[string]Holder{
+		"telegram-sender": {Node: "anchor", Module: "telegram"},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if !containsStep(held.steps, "consumer SEAT_TELEGRAM_SENDER/SEAT_TELEGRAM_SENDER_worker") {
+		t.Fatalf("the seat's holder got no worker: %v", held.steps)
+	}
+}
+
+// recording is a connection to the bus that writes down what it was asked for.
+type recording struct{ steps []string }
+
+func (r *recording) EnsureStream(s Stream) error {
+	r.steps = append(r.steps, "stream "+s.Name)
+	return nil
+}
+
+func (r *recording) EnsureConsumer(c Consumer) error {
+	r.steps = append(r.steps, "consumer "+c.Stream+"/"+c.Name)
+	return nil
+}
+
+func containsStep(steps []string, want string) bool {
+	for _, s := range steps {
+		if s == want {
+			return true
+		}
+	}
+	return false
+}

internal/broker/testdata/composed.conf

@@ -0,0 +1,53 @@
+# Composed by the mesh controller. Do not edit: the next composition overwrites it.
+# Accounts and permissions are derived from what each module declares and nothing
+# else (novox/hq ADR 0043, design 29 §2).
+
+port: 4222
+http: 127.0.0.1:8222
+
+tls {
+  cert_file: "/tls/tls.crt"
+  key_file: "/tls/tls.key"
+  ca_file: "/tls/ca.crt"
+}
+
+jetstream {
+  store_dir: "/data"
+}
+
+# The mesh's users, composed by the controller. Do not edit: the next
+# composition overwrites it. Permissions are derived from what each module
+# declares and nothing else (novox/hq ADR 0043, design 29 §2).
+
+accounts {
+  MESH {
+    users = [
+      { user: "controller", password: "$2a$11$cccccccccccccccccccccc", permissions: {
+        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "_INBOX.enrol.>", "mesh.control.>", "mesh.node.>", "mesh.seat.mesh-build-machine.accept.>"] }
+        subscribe: { allow: ["$JS.API.>", "_INBOX.controller.>", "mesh.control.>", "mesh.mod.mesh-catalog.event.catching-up", "mesh.mod.mesh-catalog.event.upgraded", "mesh.seat.mesh-build-machine.event.built"] }
+        allow_responses: { max: 1, ttl: "1m" }
+      } }
+      { user: "enrol.one", password: "$2a$11$eeeeeeeeeeeeeeeeeeeeee", permissions: {
+        publish: { allow: ["mesh.control.enrol"] }
+        subscribe: { allow: ["_INBOX.enrol.one.>"] }
+      } }
+      { user: "node.one", password: "$2a$11$nnnnnnnnnnnnnnnnnnnnnn", permissions: {
+        publish: { allow: ["$JS.ACK.NODES.one.>", "mesh.control.one.>"] }
+        subscribe: { allow: ["_INBOX.node.one.>", "mesh.node.one.declare"] }
+      } }
+      { user: "one.telegram", password: "$2a$11$tttttttttttttttttttttt", permissions: {
+        publish: { allow: ["$JS.ACK.EVENTS.one_telegram.>", "mesh.seat.telegram-sender.event.delivered", "mesh.seat.telegram-sender.event.failed"] }
+        subscribe: { allow: ["_INBOX.one.telegram.>", "mesh.mod.telegram.tool.status", "mesh.seat.telegram-sender.accept.send"] }
+        allow_responses: { max: 1, ttl: "1m" }
+      } }
+      { user: "two.audit", password: "$2a$11$aaaaaaaaaaaaaaaaaaaaaa", permissions: {
+        publish: { allow: ["$JS.ACK.EVENTS.two_audit.>"] }
+        subscribe: { allow: ["_INBOX.two.audit.>", "mesh.mod.shop.event.order.placed"] }
+      } }
+      { user: "two.shop", password: "$2a$11$ssssssssssssssssssssss", permissions: {
+        publish: { allow: ["$JS.ACK.EVENTS.two_shop.>", "mesh.mod.shop.event.order.placed", "mesh.seat.telegram-sender.accept.send"] }
+        subscribe: { allow: ["_INBOX.two.shop.>"] }
+      } }
+    ]
+  }
+}

internal/broker/users.go

@@ -0,0 +1,127 @@
+package broker
+
+import (
+	"fmt"
+	"sort"
+)
+
+// Every user the composed file should contain, derived from what the mesh knows.
+//
+// **The list is derived, never kept.** A stored user list would be a second account of who may
+// reach the bus, able to disagree with the records it came from — and the disagreement would be
+// invisible, because both would look internally consistent. So this is a pure function of the
+// mesh's records, run again every time the file is written.
+//
+// Records are mirrored into this package's own types rather than imported from the catalogue, for
+// the reason DeclaredSeat is: composing authority is a different job from parsing a manifest, and
+// this package stays free of the other's types so a change to a manifest field cannot quietly widen
+// a permission.
+
+// Declared is one module on one node, as composing its authority needs it.
+type Declared struct {
+	Module   string
+	Emits    []string
+	Consumes []string
+	Serves   []string
+	// Holds are the seats this module claims, with the protocol each seat declares. A seat the
+	// mesh defines for itself declares no protocol, so holding one grants nothing on the bus —
+	// which is right: those seats are about who does a job, not about who may say what.
+	Holds []Seat
+	// Uses are the seats this module sends to.
+	Uses []Seat
+	// Watches are the seats whose events it consumes.
+	Watches []Seat
+}
+
+// Records is what composing a user list needs to know about the mesh, and nothing more.
+type Records struct {
+	// Nodes is every machine the mesh knows. Each gets a host user.
+	Nodes []string
+	// Assigned is the modules on each node, as they declare themselves.
+	Assigned map[string][]Declared
+	// Enrolling is every node with a live token — one enrolment user each, because the inbox an
+	// answer goes to is scoped to the token and a shared one is one machine reading another's
+	// sealed credentials (design 25 §6).
+	Enrolling []string
+	// People is each person's name against the tools they may invoke, `*` for an administrator.
+	People map[string][]string
+}
+
+// Users is every user the composed file should contain, in the order it will be written.
+//
+// The controller is always first and always present: a mesh whose own controller is not in the file
+// is a mesh that cannot be told anything, and there is no state of the records in which that is
+// correct.
+func Users(r Records) ([]Principal, error) {
+	out := []Principal{{Kind: KindController}}
+
+	for _, node := range sortedCopy(r.Nodes) {
+		out = append(out, Principal{Kind: KindNode, Node: node})
+		for _, d := range r.Assigned[node] {
+			out = append(out, Principal{
+				Kind: KindModule, Node: node, Module: d.Module,
+				Emits: d.Emits, Consumes: d.Consumes, Serves: d.Serves,
+				Holds: d.Holds, Uses: d.Uses, Watches: d.Watches,
+			})
+		}
+	}
+	for _, node := range sortedCopy(r.Enrolling) {
+		out = append(out, Principal{Kind: KindEnrolment, Node: node})
+	}
+	for _, person := range sortedNames(r.People) {
+		out = append(out, Principal{Kind: KindPerson, Module: person, Invokes: r.People[person]})
+	}
+
+	// Refused here rather than discovered by the server. Two users with one name is a file the
+	// server reads as one of them, and which one depends on the order — so a module assigned to a
+	// node twice, or a person named after nothing, is a composition that must not be written.
+	seen := map[string]string{}
+	for _, p := range out {
+		name := p.Username()
+		if name == "" || name == "." {
+			return nil, fmt.Errorf("a %s user has no name, so nothing could authenticate as it", p.Kind)
+		}
+		if first, already := seen[name]; already {
+			return nil, fmt.Errorf(
+				"two users would be called %q (a %s and a %s): the server would read the file as "+
+					"one of them, and which one depends on the order", name, first, p.Kind)
+		}
+		seen[name] = string(p.Kind)
+	}
+	return out, nil
+}
+
+// WithPasswords fills each user's hash from what the mesh minted, and says which users have none.
+//
+// **Separated from Users because they fail differently.** A user missing from the records is a bug
+// in deriving them; a user with no password is a step that has not happened yet — a module assigned
+// but never given a credential, a node enrolled before this existed. The second is ordinary and its
+// remedy is to mint one, so it is named rather than returned as an error, and the caller decides
+// whether a partial composition is worth writing.
+func WithPasswords(principals []Principal, hashes map[string]string) (filled []Principal, missing []string) {
+	for _, p := range principals {
+		hash, ok := hashes[p.Username()]
+		if !ok || hash == "" {
+			missing = append(missing, p.Username())
+			continue
+		}
+		p.PasswordHash = hash
+		filled = append(filled, p)
+	}
+	return filled, missing
+}
+
+func sortedCopy(in []string) []string {
+	out := append([]string(nil), in...)
+	sort.Strings(out)
+	return out
+}
+
+func sortedNames(in map[string][]string) []string {
+	out := make([]string, 0, len(in))
+	for k := range in {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}

internal/broker/users_test.go

@@ -0,0 +1,241 @@
+package broker
+
+import (
+	"strings"
+	"testing"
+)
+
+// Deriving the bus's user list from the mesh's records.
+//
+// Every test here is about a way the list could be wrong that the server would not tell anybody
+// about: a user missing, a user named twice, a user with authority it did not declare.
+
+func someRecords() Records {
+	return Records{
+		Nodes: []string{"two", "one"},
+		Assigned: map[string][]Declared{
+			"one": {{Module: "telegram", Serves: []string{"status"}}},
+			"two": {{Module: "shop", Emits: []string{"order.placed"}}},
+		},
+		Enrolling: []string{"three"},
+		People:    map[string][]string{"ada": {"mesh-catalog.catalog_tools"}},
+	}
+}
+
+func namesOf(t *testing.T, r Records) []string {
+	t.Helper()
+	users, err := Users(r)
+	if err != nil {
+		t.Fatal(err)
+	}
+	out := make([]string, 0, len(users))
+	for _, u := range users {
+		out = append(out, u.Username())
+	}
+	return out
+}
+
+// The controller is always there. A mesh whose own controller is not in the file is a mesh that
+// cannot be told anything, and there is no state of the records in which that is correct.
+func TestTheControllerIsAlwaysInTheList(t *testing.T) {
+	for _, r := range []Records{{}, someRecords()} {
+		names := namesOf(t, r)
+		if len(names) == 0 || names[0] != "controller" {
+			t.Fatalf("the controller is not first in %v", names)
+		}
+	}
+}
+
+// One user per node, one per module per node, one per live token and one per person — and nothing
+// else, because a user nobody derived is a user nobody can explain.
+func TestEveryRecordBecomesExactlyOneUser(t *testing.T) {
+	names := namesOf(t, someRecords())
+	want := []string{
+		"controller",
+		"node.one", "one.telegram",
+		"node.two", "two.shop",
+		"enrol.three",
+		"person.ada",
+	}
+	if strings.Join(names, ",") != strings.Join(want, ",") {
+		t.Fatalf("derived %v\n    want %v", names, want)
+	}
+}
+
+// Two users with one name is a file the server reads as one of them, and which one depends on the
+// order. Refused here, where both can be named, rather than left to be whichever the server picked.
+func TestTwoUsersWithOneNameAreRefused(t *testing.T) {
+	r := someRecords()
+	r.Assigned["one"] = append(r.Assigned["one"], Declared{Module: "telegram"})
+	_, err := Users(r)
+	if err == nil {
+		t.Fatal("a module assigned twice to one node composed two users with one name")
+	}
+	if !strings.Contains(err.Error(), "one.telegram") {
+		t.Fatalf("the refusal does not name the user: %v", err)
+	}
+}
+
+// A module's authority is what it declared and nothing more, carried through the derivation intact —
+// because this is the step where a mistake would grant something no manifest asked for.
+func TestAModulesAuthorityIsWhatItDeclared(t *testing.T) {
+	seat := Seat{Name: "telegram-sender", Accepts: []string{"send"}, Emits: []string{"delivered"}}
+	users, err := Users(Records{
+		Nodes: []string{"one"},
+		Assigned: map[string][]Declared{"one": {{
+			Module: "shop", Emits: []string{"order.placed"}, Uses: []Seat{seat},
+		}}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	perms, err := PermissionsFor(users[len(users)-1])
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, perms.Publish, "mesh.mod.shop.event.order.placed")
+	has(t, perms.Publish, "mesh.seat.telegram-sender.accept.send")
+	// A seat it uses, not one it holds: it may submit work and may not publish the seat's own
+	// events, or it could lie about outcomes on a role somebody else fills.
+	hasNot(t, perms.Publish, "mesh.seat.telegram-sender.event.delivered")
+	hasNot(t, perms.Subscribe, "mesh.seat.telegram-sender.accept.send")
+}
+
+// A user the mesh has never minted a password for is named rather than silently dropped or
+// composed as a user anybody is. It is an ordinary situation — a module assigned a moment ago — and
+// the remedy is to mint one, so the caller decides whether to write a partial file.
+func TestAUserWithNoPasswordIsNamedRatherThanWritten(t *testing.T) {
+	users, err := Users(someRecords())
+	if err != nil {
+		t.Fatal(err)
+	}
+	filled, missing := WithPasswords(users, map[string]string{
+		"controller": "$2a$hash", "node.one": "$2a$hash",
+	})
+	if len(filled) != 2 {
+		t.Fatalf("composed %d users from two hashes", len(filled))
+	}
+	if len(missing) != len(users)-2 {
+		t.Fatalf("%d users are missing a password, of %d: %v", len(missing), len(users), missing)
+	}
+	for _, p := range filled {
+		if p.PasswordHash == "" {
+			t.Fatalf("%s was kept with no password, which is a user anybody is", p.Username())
+		}
+	}
+}
+
+// And the whole thing composes: records in, a file the server would read out.
+func TestRecordsComposeIntoAFile(t *testing.T) {
+	users, err := Users(someRecords())
+	if err != nil {
+		t.Fatal(err)
+	}
+	hashes := map[string]string{}
+	for _, u := range users {
+		hashes[u.Username()] = "$2a$11$" + strings.Repeat("x", 22)
+	}
+	filled, missing := WithPasswords(users, hashes)
+	if len(missing) != 0 {
+		t.Fatalf("users with no password: %v", missing)
+	}
+	got, err := Compose(Server{ClientPort: 4222, MonitoringPort: 8222, StoreDir: "/data",
+		TLSCert: "/tls/tls.crt", TLSKey: "/tls/tls.key", TLSCA: "/tls/ca.crt"}, filled)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, want := range []string{
+		`user: "controller"`, `user: "node.one"`, `user: "one.telegram"`,
+		`user: "enrol.three"`, `user: "person.ada"`,
+		`"_INBOX.enrol.three.>"`, `"mesh.mod.mesh-catalog.tool.catalog_tools"`,
+	} {
+		if !strings.Contains(got, want) {
+			t.Errorf("the composed file does not contain %s", want)
+		}
+	}
+}
+
+// The accounts block alone is what the mesh writes, and it holds nothing about the server.
+//
+// **The split is the whole design decision** (ComposeAccounts): ports, TLS paths and a store
+// directory are properties of the container the module raises, and a controller that wrote them
+// would have to be kept in step with a Dockerfile it never sees. So this test says what must not be
+// in the file as plainly as what must.
+func TestWhatTheMeshWritesIsUsersAndNothingAboutTheServer(t *testing.T) {
+	users, err := Users(someRecords())
+	if err != nil {
+		t.Fatal(err)
+	}
+	hashes := map[string]string{}
+	for _, u := range users {
+		hashes[u.Username()] = "$2a$11$" + strings.Repeat("x", 22)
+	}
+	filled, missing := WithPasswords(users, hashes)
+	if len(missing) != 0 {
+		t.Fatalf("users with no password: %v", missing)
+	}
+	got, err := ComposeAccounts(filled)
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	for _, want := range []string{"accounts {", `user: "controller"`, `user: "one.telegram"`} {
+		if !strings.Contains(got, want) {
+			t.Errorf("the accounts file does not contain %s", want)
+		}
+	}
+	// None of the server's own settings. Each of these in the mesh's file is a value the controller
+	// would then own, and the module could no longer change its own image without the mesh agreeing.
+	for _, absent := range []string{"port:", "http:", "jetstream", "tls {", "store_dir", "cert_file"} {
+		if strings.Contains(got, absent) {
+			t.Errorf("the accounts file contains %q, which belongs to the module that raises the "+
+				"server, not to the mesh", absent)
+		}
+	}
+}
+
+// A user with no password is refused here too, not only by the whole-file composition: this is the
+// function the controller actually calls, and a user without a password is a user anybody is.
+func TestTheAccountsFileRefusesAUserWithNoPassword(t *testing.T) {
+	if _, err := ComposeAccounts([]Principal{{Kind: KindController}}); err == nil {
+		t.Fatal("a user with no password hash was written")
+	}
+}
+
+// **A user list is composed for a bus the mesh has not moved onto yet**, and that is the whole of
+// step 2 (novox/hq ADR 0116): the server stands in the mesh carrying nothing, on its own ports, while
+// every node is still on the bus it was on.
+//
+// Pinned because the first version of the composing step got it backwards — it wrote the list only
+// once the controller was already on the new bus, which is a step that cannot be taken: the module
+// comes up, finds no accounts file, and waits for one the controller had decided not to write.
+func TestAUserListIsComposedBeforeAnythingMovesOntoTheBus(t *testing.T) {
+	// Exactly the records of a mesh mid-change: everything running, nothing on the new bus.
+	users, err := Users(Records{
+		Nodes:    []string{"anchor"},
+		Assigned: map[string][]Declared{"anchor": {{Module: "nats"}}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	hashes := map[string]string{}
+	for _, u := range users {
+		hashes[u.Username()] = "$2a$11$" + strings.Repeat("x", 22)
+	}
+	filled, missing := WithPasswords(users, hashes)
+	if len(missing) != 0 {
+		t.Fatalf("users with no credential: %v", missing)
+	}
+	accounts, err := ComposeAccounts(filled)
+	if err != nil {
+		t.Fatal(err)
+	}
+	// The controller's own user above all: a file without it is a bus its writer cannot connect to,
+	// which is what the server would be left holding the moment it starts.
+	if !strings.Contains(accounts, `user: "controller"`) {
+		t.Fatalf("the composed list does not contain the controller:\n%s", accounts)
+	}
+	if !strings.Contains(accounts, `user: "node.anchor"`) {
+		t.Errorf("the composed list does not contain the machine running the bus")
+	}
+}

internal/catalogue/broker_seat_test.go

@@ -0,0 +1,101 @@
+package catalogue
+
+import (
+	"strings"
+	"testing"
+)
+
+// The mesh's bus is one per mesh, read from the catalogue beside this checkout.
+//
+// **This is step 2's claim, and it is checked here rather than in a bed** (novox/hq ADR 0116):
+// adoption puts the NATS server into the `mesh-broker` seat on a mesh that is already running,
+// and the property that matters is that a second one anywhere is refused *when it is assigned*,
+// not discovered later as two servers holding different halves of the mesh's traffic. A second
+// bus is not a degraded mesh; it is two meshes that both believe they are the one.
+func TestASecondMeshBusAnywhereIsRefusedByName(t *testing.T) {
+	nats := catalogueManifest(t, "nats")
+
+	if _, err := Resolve(shelf(nats), []string{"nats"}, workstation(), World{}); err != nil {
+		t.Fatalf("the bus alone does not resolve: %v", err)
+	}
+
+	elsewhere := World{Held: []Held{{Claim: "mesh-broker", Scope: ScopeMesh,
+		Node: "anchor", Module: "nats"}}}
+	other := workstation()
+	other.Name = "laptop"
+	_, err := Resolve(shelf(nats), []string{"nats"}, other, elsewhere)
+	if err == nil {
+		t.Fatal("a second bus was accepted on another machine")
+	}
+	if !strings.Contains(err.Error(), "mesh-broker") || !strings.Contains(err.Error(), "one per mesh") {
+		t.Fatalf("refused without naming the seat: %v", err)
+	}
+}
+
+// The seat is the server's role, not the product's name (novox/hq ADR 0079). A different
+// implementation of the bus claims the same seat, and the mesh refuses it for the same reason —
+// which is the property that lets the bus be replaced at all.
+func TestTheSeatRefusesADifferentBusToo(t *testing.T) {
+	nats := catalogueManifest(t, "nats")
+	held := World{Held: []Held{{Claim: "mesh-broker", Scope: ScopeMesh,
+		Node: "anchor", Module: "some-other-broker"}}}
+	other := workstation()
+	other.Name = "laptop"
+	if _, err := Resolve(shelf(nats), []string{"nats"}, other, held); err == nil {
+		t.Fatal("the seat admitted a second holder because the module's name differed")
+	}
+}
+
+// The old broker no longer claims the seat: it is an ordinary provider of `amqp`
+// (novox/hq ADR 0119), so it can sit on the same mesh as the bus without contending for it.
+func TestTheAmqpBrokerDoesNotContendForTheSeat(t *testing.T) {
+	lavinmq := catalogueManifest(t, "lavinmq")
+	for _, c := range lavinmq.Claims {
+		if c.Name == "mesh-broker" {
+			t.Fatal("the amqp broker still claims mesh-broker; it is a provider, not foundation")
+		}
+	}
+	busHeld := World{Held: []Held{{Claim: "mesh-broker", Scope: ScopeMesh,
+		Node: "anchor", Module: "nats"}}}
+	other := workstation()
+	other.Name = "laptop"
+	if _, err := Resolve(shelf(lavinmq), []string{"lavinmq"}, other, busHeld); err != nil {
+		t.Fatalf("the amqp broker was refused beside the mesh bus: %v", err)
+	}
+}
+
+// **A seat and the interface it delivers are different names, and renaming one must not rename
+// the other** (novox/hq ADR 0118). This nearly went wrong: the seats were renamed to the `mesh-*`
+// prefix, and a blanket search-and-replace also renamed `npm-package-registry` and `git` where
+// they are *provisions* — which a consumer requires and a provider offers. The tests failed with
+// "the package registry is served on <nil>", which does not say "you renamed an interface".
+func TestRenamingASeatDidNotRenameTheInterfaceItDelivers(t *testing.T) {
+	for _, pair := range []struct{ seat, delivers string }{
+		{"git", "git"},
+		{"npm-package-registry", "npm-package-registry"},
+		{"the-artifact-store", "artifact-store"},
+		{"mesh-store", "postgres-database"},
+		{"mesh-broker", "mesh-bus"},
+	} {
+		s, known := SeatNamed(pair.seat)
+		if !known {
+			t.Fatalf("%q is not a seat", pair.seat)
+		}
+		if s.Delivers != pair.delivers {
+			t.Errorf("the %s seat delivers %q, expected %q — renaming the seat moved the "+
+				"interface with it, and every consumer requiring it would stop resolving",
+				pair.seat, s.Delivers, pair.delivers)
+		}
+		// **Three of these deliberately share a name with what they deliver**, and that is not an
+		// incomplete rename. Renaming a seat that delivers a provision cascades to every consumer
+		// requiring it, with a mesh-wide window where a holder stops resolving mid-flight — so the
+		// trunk deferred exactly those three (novox/hq ADR 0121) while renaming the node-scoped ones.
+		// What this test is for is the other direction: that renaming a seat never moves the
+		// interface, which once produced "the package registry is served on <nil>".
+	}
+}
+
+// A manifest written against an old seat name is told what it became rather than refused as
+// unknown. **That map is the controller's store now, not this package** (novox/hq ADR 0122): a
+// rename is a row, so the courtesy survives a rename nobody recompiled for. Checked where the
+// table is read, not here, where there is no longer a hardcoded list to check against.

internal/catalogue/declaration.go

@@ -97,6 +97,13 @@ type Rendering struct {
 	// compose it a second time.
 	Suffix string
 
+	// BusUsers is the mesh's composed user list, for the module holding `mesh-broker`. Empty on
+	// every other node, and on this one until the controller has composed it.
+	//
+	// **Only the users, never the server's own settings**: those are the module's, in its image and
+	// its mounts (Manifest.BusUsers).
+	BusUsers string
+
 	// MeshRange is the private network's CIDR (the range node addresses are allocated from), for a
 	// module that must name the whole mesh rather than one machine — an intrusion filter that must
 	// never ban a tunnel peer, say. A per-mesh value the module cannot know, so it is carried here
@@ -376,6 +383,35 @@ func (r Resolution) compose(with Rendering, owner map[string]string) ([]map[stri
 				})
 			}
 		}
+		if m.BusUsers != "" {
+			// **The claim authorises it, not the field.** This file holds every user's password
+			// hash, so a module that could ask for it could read every credential on the bus.
+			// Checked from this manifest alone, which is the cheapest check there is: whether some
+			// other module also claims the seat is resolution's business elsewhere, and one holder
+			// mesh-wide is already guaranteed.
+			if !m.ClaimsSeat("mesh-broker") {
+				return nil, fmt.Errorf(
+					"%s asks for the mesh's user list and does not claim mesh-broker. That file "+
+						"holds every user's password hash, so the seat is what authorises it",
+					m.Module)
+			}
+			if with.BusUsers == "" {
+				// Asked for and not composed. Refused rather than skipped, for the reason a
+				// certificate is: a bus with no user list refuses every connection in the mesh, and
+				// an empty file would look like a configuration problem on the machine.
+				return nil, fmt.Errorf(
+					"%s holds mesh-broker and the mesh composed no user list, so the bus would "+
+						"refuse every connection", m.Module)
+			}
+			first = append(first, map[string]any{
+				"id": BusUsersID(), "type": "file", "path": m.BusUsers,
+				"content": with.BusUsers,
+				// Readable by the server and nothing else. Hashes rather than passwords, so this is
+				// not a set of working credentials — but a list of every user in the mesh is worth
+				// keeping to the one process that needs it.
+				"mode": "0600",
+			})
+		}
 		for _, name := range sortedKeys(m.OwnSecrets) {
 			sealed := with.Needed[m.Module][name]
 			if sealed == "" {

internal/catalogue/declaration_test.go

@@ -0,0 +1,80 @@
+package catalogue
+
+import "testing"
+
+// The mesh's user list reaches the module holding the bus, and nothing else.
+//
+// Three refusals and one delivery, because each of the refusals would be silent in a different way:
+// a module that asked and was given it could read every credential on the bus; a bus given an empty
+// file refuses every connection in the mesh and looks like a machine problem; and a bus that never
+// asked gets nothing rather than a file it does not read.
+func TestTheMeshsUserListGoesOnlyToTheModuleHoldingTheBus(t *testing.T) {
+	theBus := func() Manifest {
+		return Manifest{
+			Module: "nats", Version: "1",
+			Claims:    []Claim{{Name: "mesh-broker", Scope: ScopeMesh}},
+			BusUsers:  "/var/lib/nats-module/conf/accounts.conf",
+			Resources: []map[string]any{},
+		}
+	}
+
+	on := func(t *testing.T, m Manifest, with Rendering) ([]map[string]any, error) {
+		t.Helper()
+		return Resolution{Node: "anchor", Modules: []Manifest{m}}.Declaration(with)
+	}
+
+	t.Run("the holder is given it", func(t *testing.T) {
+		resources, err := on(t, theBus(), Rendering{BusUsers: "accounts { MESH { users = [] } }"})
+		if err != nil {
+			t.Fatal(err)
+		}
+		// Prefixed with the module it came from, like every resource: two modules may reasonably
+		// both call something "config", and without the prefix the second would silently replace
+		// the first.
+		var found map[string]any
+		for _, r := range resources {
+			if r["id"] == "nats."+BusUsersID() {
+				found = r
+			}
+		}
+		if found == nil {
+			t.Fatalf("the bus was given no user list: %+v", resources)
+		}
+		if found["path"] != "/var/lib/nats-module/conf/accounts.conf" {
+			t.Errorf("written to %v rather than where the module asked", found["path"])
+		}
+		if found["mode"] != "0600" {
+			t.Errorf("mode %v: a list of every user in the mesh belongs to the one process that "+
+				"needs it", found["mode"])
+		}
+	})
+
+	t.Run("a module that does not claim the seat is refused", func(t *testing.T) {
+		m := theBus()
+		m.Claims = nil
+		if _, err := on(t, m, Rendering{BusUsers: "accounts {}"}); err == nil {
+			t.Fatal("a module that claims nothing was handed every user's password hash")
+		}
+	})
+
+	t.Run("the holder with nothing composed is refused", func(t *testing.T) {
+		if _, err := on(t, theBus(), Rendering{}); err == nil {
+			t.Fatal("the bus was given an empty user list, so it would refuse every connection in " +
+				"the mesh and look like a machine problem")
+		}
+	})
+
+	t.Run("a module that did not ask gets nothing", func(t *testing.T) {
+		m := theBus()
+		m.BusUsers = ""
+		resources, err := on(t, m, Rendering{BusUsers: "accounts {}"})
+		if err != nil {
+			t.Fatal(err)
+		}
+		for _, r := range resources {
+			if r["id"] == "nats."+BusUsersID() {
+				t.Fatal("a module that asked for no user list was given one")
+			}
+		}
+	})
+}

internal/catalogue/events.go

@@ -0,0 +1,161 @@
+package catalogue
+
+import (
+	"fmt"
+	"regexp"
+	"strings"
+)
+
+// What a module may call an event, and what a consumer may ask for.
+//
+// A module names an event **locally**: `order.placed`, not a subject and not a routing key
+// (design 29 §1). A consumer names the emitter and the event: `billing.order.placed`. The mesh
+// derives the subject from those, so reorganising the subject space leaves every manifest correct.
+//
+// **Nothing checked this until every manifest in the catalogue was wrong the same way**
+// (novox/hq 04-ISSUES/127). All thirty-seven kept the old bus's routing key —
+// `module.<module>.<verb>` — which the derivation read as "a module called `module`", so every
+// cross-module subscription in the mesh pointed at a namespace nobody publishes to. Nothing failed:
+// the services started and none of them reacted. The documentation on these fields taught the old
+// form too, which is why the drift was uniform rather than scattered.
+
+// eventName is one name in a local event: lower-case, and no wildcard.
+var eventName = regexp.MustCompile(`^[a-z0-9][a-z0-9-]*$`)
+
+// The wildcards a consumer may use, spelled the mesh's way and derived to whatever the transport
+// spells them as.
+//
+// **A manifest holds no transport token**, which is the whole point of naming locally: the bus the
+// mesh runs on today spells these `*` and `#`, and the one being built spells them `*` and `>`. A
+// manifest that said either would be a manifest that stopped being true when the wire changed.
+const (
+	// OneName stands for exactly one name.
+	OneName = "*"
+	// TheRest stands for one or more names, and may only come last.
+	TheRest = "**"
+)
+
+// EventProblems is what is wrong with a manifest's events.
+//
+// Refused at registration, because the alternative is a module that installs, starts, connects and
+// reacts to nothing — and every log line says it is fine.
+func EventProblems(m Manifest) []string {
+	var problems []string
+
+	for _, e := range m.Emits {
+		if was, stale := staleEventForm(e, m.Module); stale {
+			problems = append(problems, fmt.Sprintf(
+				"%s emits %q, which is the old bus's routing key. An event is named locally now, so "+
+					"write %q — the mesh derives the subject (novox/hq design 29 §1)",
+				m.Module, was, strings.TrimPrefix(was, "module."+m.Module+".")))
+			continue
+		}
+		if strings.HasPrefix(e, "module.") {
+			problems = append(problems, fmt.Sprintf(
+				"%s emits %q: `module.` is reserved, because it is how the old bus spelled a "+
+					"routing key and an event named that way derives into a namespace nobody owns",
+				m.Module, e))
+			continue
+		}
+		if err := localName(e); err != nil {
+			problems = append(problems, fmt.Sprintf("%s emits %q: %v", m.Module, e, err))
+			continue
+		}
+		// **Its own name, never another's.** The bus enforces that a namespace belongs to the module
+		// it is named for, so an event named for somebody else cannot be published at all. If the
+		// event is about a role rather than about this module, it belongs on the seat: a name that
+		// is stable across whoever fills it (04-ISSUES/127).
+		if first, _, split := strings.Cut(e, "."); split && isAModuleNameOtherThan(first, m.Module) {
+			problems = append(problems, fmt.Sprintf(
+				"%s emits %q, which reads as another module's event. A module publishes under its "+
+					"own name only. If this is about a role rather than about %s, declare it on that "+
+					"seat, where the name survives the holder changing",
+				m.Module, e, m.Module))
+		}
+	}
+
+	for _, c := range m.Consumes {
+		if strings.HasPrefix(c, "module.") {
+			problems = append(problems, fmt.Sprintf(
+				"%s consumes %q, which is the old bus's pattern. A consumed event names its emitter "+
+					"and the event: write %q", m.Module, c, strings.TrimPrefix(c, "module.")))
+			continue
+		}
+		if c == "#" {
+			problems = append(problems, fmt.Sprintf(
+				"%s consumes %q, which is the old bus's wildcard for everything. Write %q",
+				m.Module, c, TheRest))
+			continue
+		}
+		if err := consumePattern(c); err != nil {
+			problems = append(problems, fmt.Sprintf("%s consumes %q: %v", m.Module, c, err))
+		}
+	}
+	return problems
+}
+
+// staleEventForm says an emitted name is this module's own old routing key, and what it was.
+func staleEventForm(event, module string) (string, bool) {
+	return event, module != "" && strings.HasPrefix(event, "module."+module+".")
+}
+
+// isAModuleNameOtherThan says a first token names some module of this mesh that is not this one.
+//
+// Only the mesh's own seats and the catalogue could answer this properly, and neither is reachable
+// from a parser given one manifest. So this catches the case that actually happened — a name that
+// is a *provision* the mesh defines, which is where "another module's event" comes from in practice
+// — and the whole-catalogue check catches the rest.
+func isAModuleNameOtherThan(first, module string) bool {
+	if first == module || first == "" {
+		return false
+	}
+	if _, isASeat := SeatNamed(first); isASeat {
+		return true
+	}
+	if _, isASeat := SeatDelivering(first); isASeat {
+		return true
+	}
+	return false
+}
+
+// localName checks one event name: dot-separated names, no wildcards, nothing else.
+func localName(event string) error {
+	if event == "" {
+		return fmt.Errorf("an event needs a name")
+	}
+	for _, part := range strings.Split(event, ".") {
+		if part == OneName || part == TheRest {
+			return fmt.Errorf("an emitted event names one event, so it carries no wildcard")
+		}
+		if !eventName.MatchString(part) {
+			return fmt.Errorf("%q is not a usable name: lower-case letters, digits and dashes", part)
+		}
+	}
+	return nil
+}
+
+// consumePattern checks a consumed pattern: the emitter, then the event, with wildcards.
+func consumePattern(pattern string) error {
+	if pattern == "" {
+		return fmt.Errorf("a consumed event needs an emitter and an event")
+	}
+	parts := strings.Split(pattern, ".")
+	for i, part := range parts {
+		switch {
+		case part == TheRest:
+			if i != len(parts)-1 {
+				return fmt.Errorf("%q stands for the rest of a name, so nothing may follow it", TheRest)
+			}
+		case part == OneName:
+		case !eventName.MatchString(part):
+			return fmt.Errorf("%q is not a usable name: lower-case letters, digits and dashes", part)
+		}
+	}
+	// `**` alone is every event from every module, which the audit logger wants and says plainly.
+	if len(parts) == 1 && parts[0] != TheRest {
+		return fmt.Errorf(
+			"%q names an emitter and no event. Write <emitter>.<event>, or %q for every event",
+			pattern, TheRest)
+	}
+	return nil
+}

internal/catalogue/manifest.go

@@ -176,15 +176,29 @@ type Manifest struct {
 	// Requires are names that must be provided by something assigned to the same node.
 	Requires []string `json:"requires,omitempty"`
 
-	// Emits are the event types this module publishes onto the broker — dotted topic keys, e.g.
-	// "module.umami.site.created". Declared so the mesh knows the event graph; events are
-	// provisioning's lighter sibling — 1:many and broadcast, no credential (novox/hq ADR 0041).
+	// Emits are the events this module publishes, named **locally**: `order.placed`, not a subject
+	// and not a routing key. The mesh derives where it lands (design 29 §1), so reorganising the
+	// subject space leaves this manifest correct. Events are provisioning's lighter sibling — 1:many
+	// and broadcast, no credential (novox/hq ADR 0041).
+	//
+	// A module publishes under its own name only. If the event is about a *role* rather than about
+	// this module, it belongs on that seat, where the name outlives whoever holds it.
+	//
+	// **This said "dotted topic keys, e.g. module.umami.site.created" until 04-ISSUES/127**, which
+	// is the old bus's routing key, and is why every manifest in the catalogue had the same mistake:
+	// nobody was guessing, everybody followed this comment.
 	Emits []string `json:"emits,omitempty"`
 
-	// Consumes are the event patterns this module subscribes to — topic patterns over module,
-	// mesh and node events alike, e.g. "node.*.joined" or "#" (the audit logger). The runtime
-	// wires the subscription; the module ships the handler. A Consumes for an event nothing on
-	// the mesh Emits is a dangling edge.
+	// Consumes are the events this module reacts to, each naming its emitter and the event:
+	// `billing.order.placed`. `*` stands for one name and `**` for the rest, so `*.download.completed`
+	// is that event from any module and `**` is every event in the mesh.
+	//
+	// Spelled the mesh's way rather than the wire's, for the reason Emits is: the bus the mesh runs
+	// on today spells these `*` and `#`, the one being built spells them `*` and `>`, and a manifest
+	// naming either would stop being true when the wire changed.
+	//
+	// The runtime wires the subscription; the module ships the handler. A Consumes for an event
+	// nothing on the mesh Emits is a dangling edge.
 	Consumes []string `json:"consumes,omitempty"`
 
 	// Claims are singular resources. Two modules claiming one thing within a scope cannot both
@@ -192,13 +206,33 @@ type Manifest struct {
 	// that every new module would force its predecessors to update.
 	Claims []Claim `json:"claims,omitempty"`
 
-	// DefinesSeats are the seats this module defines for itself (novox/hq ADR 0121). The control
-	// plane defines the system seats — `mesh-*` and `node-*` — and a module may define its own,
-	// named outside that namespace, to coordinate its own instances: the mesh enforces
-	// one-holder-per-scope for it without knowing what it means. A module's declared seat is the
-	// only non-system name it may then claim; a claim to a name neither the mesh nor the module
-	// defines is refused.
-	DefinesSeats []Claim `json:"seats,omitempty"`
+	// DefinesSeats are the seats this module defines for itself, with their protocols
+	// (novox/hq ADR 0121, ADR 0129). The control plane defines the system seats — `mesh-*` and
+	// `node-*` — and a module may define its own, named outside that namespace, to coordinate its
+	// own instances: the mesh enforces one-holder-per-scope for it without knowing what it means. A
+	// module's declared seat is the only non-system name it may then claim; a claim to a name
+	// neither the mesh nor the module defines is refused.
+	//
+	// **Two lines of work built this at once**, one calling it `Seats` with a protocol and one
+	// `DefinesSeats` without. Same key in the file, so no manifest is affected: this is the trunk's
+	// name with the richer type, because what a role accepts, emits and serves is what lets the mesh
+	// check that a holder answers what its seat promises.
+	DefinesSeats []SeatDeclaration `json:"seats,omitempty"`
+
+	// Uses are seats this module sends to. It names the *seat*, never the module holding it, so
+	// the implementation can be replaced under it and no caller changes. A caller gets publish
+	// on that seat's inbound subjects and nothing else — not its outbound events, and not a
+	// subscription to the queue it writes to (design 29 §2).
+	Uses []string `json:"uses,omitempty"`
+
+	// Tools are the tools this module answers — request and reply, awaited.
+	//
+	// **New, and not `serves`**, which this manifest already uses for the facts a consumer needs
+	// in order to reach a provision. Two meanings under one key would be a footgun in the one
+	// file a module author reads most. Until now a module's tools were known only at runtime,
+	// from MESH_TOOL_MODULES in its image; declaring them is what lets the mesh check that a
+	// module claiming a seat answers what that seat's protocol promises (novox/hq ADR 0118).
+	Tools []string `json:"tools,omitempty"`
 
 	// Capabilities the machine must have. A different field from Requires because the remedy
 	// differs: a missing module can be assigned, and a missing capability means the wrong
@@ -424,6 +458,20 @@ type Manifest struct {
 	// A directory rather than one document for the same reason as above: each value is sealed
 	// separately and the mesh cannot open any of them to build a list.
 	Grants map[string]string `json:"grants,omitempty"`
+
+	// BusUsers is where this module wants the mesh's user list written, and it is only ever
+	// answered for the module holding `mesh-broker`.
+	//
+	// **The mesh writes who may connect; the module owns everything else about its server**
+	// (novox/hq design 25 §4, task 1.7). Ports, TLS paths and a store directory live in this
+	// module's image and its mounts and change when it does, so the module's own configuration
+	// carries them and includes this file. A controller that wrote the whole configuration would
+	// have to be kept in step with a Dockerfile it never sees.
+	//
+	// **Asking for it is not enough to receive it.** This file holds every user's password hash, so
+	// a module that could ask for it could read every credential on the bus — and the claim on
+	// `mesh-broker` is what authorises it, checked from this manifest alone.
+	BusUsers string `json:"bus-users,omitempty"`
 }
 
 // Build says how to produce this module's artifacts from its source.
@@ -672,7 +720,22 @@ type Certificate struct {
 
 // CertificateID and AuthorityID are the resource identities of what the mesh issued.
 func CertificateID() string { return "certificate" }
-func AuthorityID() string   { return "certificate-authority" }
+
+// ClaimsSeat says whether this manifest claims one named seat.
+func (m Manifest) ClaimsSeat(seat string) bool {
+	for _, c := range m.Claims {
+		if c.Name == seat {
+			return true
+		}
+	}
+	return false
+}
+
+// BusUsersID names the mesh's composed user list, so it is the same resource across every
+// declaration and a change to it is an update rather than a second file beside the old one — which
+// on a bus reading a directory would be two account lists, and the server would take both.
+func BusUsersID() string  { return "bus-users" }
+func AuthorityID() string { return "certificate-authority" }
 
 // FilteringID names the computed rule set, so it is the same resource across every declaration
 // and a change to it is an update rather than an addition beside the old one.
@@ -1001,6 +1064,10 @@ func ParseManifest(raw []byte) (Manifest, error) {
 			problems = append(problems, fmt.Sprintf("%s requires itself", m.Module))
 		}
 	}
+	// What it may call an event, and what it may ask to hear (events.go). Checked here because a
+	// module whose event names are wrong installs, starts, connects and reacts to nothing, with
+	// every log line saying it is fine (novox/hq 04-ISSUES/127).
+	problems = append(problems, EventProblems(m)...)
 	wellFormed := true
 	for _, c := range m.Claims {
 		if !name.MatchString(c.Name) {
@@ -1021,6 +1088,10 @@ func ParseManifest(raw []byte) (Manifest, error) {
 	if wellFormed {
 		problems = append(problems, claimProblems(m)...)
 	}
+	// What one manifest can be judged on: a declaration's shape, its scope, and the reserved
+	// prefix. Whether a seat anybody names exists, and whether a holder answers for it, are
+	// facts about the catalogue and are checked at registration (CatalogueProblems).
+	problems = append(problems, declaredSeatProblems(m)...)
 	if m.Computed != "" && len(m.Resources) > 0 {
 		// One or the other. A module that both ships files and has them computed would leave
 		// nobody able to say where a given file came from.

internal/catalogue/no_subjects_test.go

@@ -0,0 +1,121 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"regexp"
+	"strings"
+	"testing"
+)
+
+// **A manifest holds no subject** (novox/hq design 29 §1).
+//
+// A module names its events, tools and seats locally, and the mesh derives where they land. The
+// property that buys: reorganise the subject space and every manifest in the catalogue is still
+// correct. It holds today by construction — nothing reads a subject from a manifest — and a rule
+// held by construction is one a later field breaks quietly, with the symptom appearing as a
+// permission that does not match a subject rather than as a manifest that was wrong.
+func TestNoManifestContainsASubject(t *testing.T) {
+	root := filepath.Join("..", "..", "..", "mesh-catalog", "modules")
+	entries, err := os.ReadDir(root)
+	if err != nil {
+		t.Skipf("catalogue sibling not present: %v", err)
+	}
+
+	// Anything in the mesh's own subject space, and anything shaped like a wire address.
+	subject := regexp.MustCompile(`^(mesh|\$JS)\.[a-zA-Z0-9_*>.-]+$`)
+
+	var found []string
+	var walk func(module string, path string, v any)
+	walk = func(module, path string, v any) {
+		switch t := v.(type) {
+		case string:
+			if subject.MatchString(t) {
+				found = append(found, module+" "+path+" = "+t)
+			}
+		case map[string]any:
+			for k, inner := range t {
+				walk(module, path+"."+k, inner)
+			}
+		case []any:
+			for _, inner := range t {
+				walk(module, path+"[]", inner)
+			}
+		}
+	}
+
+	checked := 0
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		raw, err := os.ReadFile(filepath.Join(root, e.Name(), "module.json"))
+		if err != nil {
+			continue
+		}
+		var m any
+		if err := json.Unmarshal(raw, &m); err != nil {
+			t.Errorf("%s: %v", e.Name(), err)
+			continue
+		}
+		checked++
+		walk(e.Name(), "", m)
+	}
+	if checked == 0 {
+		t.Skip("no manifests read")
+	}
+	if len(found) > 0 {
+		t.Errorf("a manifest names a subject, so reorganising the subject space would mean "+
+			"editing the catalogue:\n  %s", strings.Join(found, "\n  "))
+	}
+	t.Logf("%d manifests hold no subject", checked)
+}
+
+// **Every module's event names are what design 29 says, across the whole catalogue.**
+//
+// The rule above holds by construction and turned out to be weaker than it reads: a manifest holds
+// no subject, and every manifest in the catalogue still held the old bus's routing key, which
+// derives into a namespace nobody owns (novox/hq 04-ISSUES/127). Nothing failed — the services
+// started and none of them reacted. This is the check that was missing.
+func TestEveryManifestsEventNamesAreLocal(t *testing.T) {
+	manifests := theCatalogue(t)
+
+	var problems []string
+	for _, m := range manifests {
+		problems = append(problems, EventProblems(m)...)
+	}
+	if len(problems) > 0 {
+		t.Fatalf("the catalogue holds %d event name(s) the mesh would derive wrongly:\n  %s",
+			len(problems), strings.Join(problems, "\n  "))
+	}
+}
+
+// theCatalogue is every manifest beside this checkout, parsed the way registration parses one.
+func theCatalogue(t *testing.T) []Manifest {
+	t.Helper()
+	root := filepath.Join("..", "..", "..", "mesh-catalog", "modules")
+	entries, err := os.ReadDir(root)
+	if err != nil {
+		t.Skipf("catalogue sibling not present: %v", err)
+	}
+	var out []Manifest
+	for _, e := range entries {
+		if !e.IsDir() {
+			continue
+		}
+		raw, err := os.ReadFile(filepath.Join(root, e.Name(), "module.json"))
+		if err != nil {
+			continue
+		}
+		var m Manifest
+		if err := json.Unmarshal(raw, &m); err != nil {
+			t.Fatalf("%s: %v", e.Name(), err)
+		}
+		out = append(out, m)
+	}
+	if len(out) == 0 {
+		t.Skip("no manifests found beside this checkout")
+	}
+	return out
+}

internal/catalogue/seats.go

@@ -27,6 +27,17 @@ type Seat struct {
 	// provision may only be held by a module providing it at the seat's scope, and its holder is
 	// what a requirement for that provision resolves to when several modules provide it.
 	Delivers string
+	// Accepts, Emits and Serves are the protocol of the role, as local verbs — the same three a
+	// module declares for a seat of its own (novox/hq ADR 0118), and empty for most of these: a seat
+	// is usually about who does a job and not about what may be said to them.
+	//
+	// **Named here so the mesh has no role it cannot describe** (ADR 0121). Without them a build
+	// machine had three audiences for one outcome and nothing derived a grant for any of them, and an
+	// event about a role had nowhere to live but the namespace of whichever module held that role
+	// today — which the bus refuses, because a namespace belongs to who it is named for.
+	Accepts []string
+	Emits   []string
+	Serves  []string
 	// Decision is the record that made it a seat.
 	Decision string
 }
@@ -39,7 +50,12 @@ type Seat struct {
 var defaultSeats = []Seat{
 	{Name: "mesh-controller", Scope: ScopeMesh, Decision: "novox/hq ADR 0079"},
 	{Name: "mesh-store", Scope: ScopeMesh, Delivers: "postgres-database", Decision: "novox/hq ADR 0079"},
-	{Name: "mesh-broker", Scope: ScopeMesh, Delivers: "amqp", Decision: "novox/hq ADR 0079"},
+	// **Delivers the mesh's own bus, not `amqp`.** Those were the same word until
+	// ADR 0127 separated them: `amqp` is a backing service a module may require, and this seat is
+	// the mesh's own transport. ADR 0128 then made that connection something a module requires
+	// rather than receives ambiently — 23 of the catalogue's modules never speak, and an ambient
+	// connection would mint a credential for each.
+	{Name: "mesh-broker", Scope: ScopeMesh, Delivers: "mesh-bus", Decision: "novox/hq ADR 0079"},
 	{Name: "the-artifact-store", Scope: ScopeMesh, Delivers: "artifact-store", Decision: "novox/hq ADR 0075"},
 	{Name: "mesh-catalog", Scope: ScopeMesh, Decision: "novox/hq ADR 0121"},
 	// Deferred renames (novox/hq ADR 0121): these deliver a provision, so renaming them is a
@@ -47,7 +63,11 @@ var defaultSeats = []Seat{
 	// They keep their names until that migration is done deliberately, apart from the node-* pass.
 	{Name: "npm-package-registry", Scope: ScopeMesh, Delivers: "npm-package-registry", Decision: "novox/hq ADR 0109"},
 	{Name: "git", Scope: ScopeMesh, Delivers: "git", Decision: "novox/hq ADR 0111"},
-	{Name: "mesh-build-machine", Scope: ScopeMesh, Decision: "novox/hq ADR 0121"},
+	// A build is work submitted to this role and its outcome is the role's own event (ADR 0129).
+	// One publish reaches whoever asked, the controller that records it, and the catalogue that
+	// places it in the graph — what the old bus's shared exchange did for free.
+	{Name: "mesh-build-machine", Scope: ScopeMesh,
+		Accepts: []string{"build"}, Emits: []string{"built"}, Decision: "novox/hq ADR 0121"},
 	{Name: "node-dns-resolver", Scope: ScopeNode, Decision: "novox/hq ADR 0121"},
 	{Name: "node-intrusion-prevention", Scope: ScopeNode, Decision: "novox/hq ADR 0121"},
 	{Name: "node-packet-filter", Scope: ScopeNode, Decision: "novox/hq ADR 0121"},
@@ -152,7 +172,7 @@ func SeatDelivering(provision string) (Seat, bool) {
 func claimProblems(m Manifest) []string {
 	var problems []string
 
-	defined := map[string]Claim{}
+	defined := map[string]SeatDeclaration{}
 	for _, d := range m.DefinesSeats {
 		if _, isSystem := SeatNamed(d.Name); isSystem || isSystemSeatName(d.Name) {
 			problems = append(problems, fmt.Sprintf(
@@ -185,9 +205,12 @@ func claimProblems(m Manifest) []string {
 		}
 		d, ours := defined[c.Name]
 		if !ours {
-			problems = append(problems, fmt.Sprintf(
-				"%s claims %q, which is not a seat this mesh defines and not one %s declares itself "+
-					"(novox/hq ADR 0121) — the seats are: %s", m.Module, c.Name, m.Module, seatNames()))
+			// **A claim on a seat this manifest does not declare is not the parser's to judge.**
+			// A module may hold a seat another module declared — that is why ADR 0126 has callers
+			// name the seat and not its provider, so an implementation can be replaced without
+			// touching a caller. Whether the seat exists is a fact about the whole catalogue, so
+			// the refusal is at registration, where every declaration is in view
+			// (`CatalogueProblems`: "which no module declares and the mesh does not define").
 			continue
 		}
 		if c.At() != d.At() {
@@ -243,3 +266,18 @@ func HolderAmong(provision string, providers []Provider, held []Held) (Provider,
 	}
 	return Provider{}, false
 }
+
+// SeatsWithAProtocol are the mesh's own seats that say something about what may be said to them or by
+// them, which is the set the bus derives streams, consumers and permissions from.
+//
+// Most of the set is not here, and that is the ordinary case: a seat saying only who does a job grants
+// nothing on the bus and needs no queue.
+func SeatsWithAProtocol() []Seat {
+	var out []Seat
+	for _, s := range seats {
+		if len(s.Accepts) > 0 || len(s.Emits) > 0 || len(s.Serves) > 0 {
+			out = append(out, s)
+		}
+	}
+	return out
+}

internal/catalogue/seats_declared.go

@@ -0,0 +1,239 @@
+package catalogue
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// Seats a module declares of its own (novox/hq ADR 0118).
+//
+// The set of seats a mesh has is **derived**: the mesh's own, in seats.go, plus those declared by
+// every module it has registered. Still closed — a seat named nowhere is refused — but computed
+// from the catalogue rather than written in the controller, which is what ADR 0110 actually
+// needed and a hand-maintained table could not keep. Its own evidence: the enumeration done by
+// hand while that record was written reported eleven claims where there were thirteen.
+//
+// **What can be checked from one manifest and what cannot.** A declaration's shape, its scope,
+// and the reserved prefix are facts about the manifest in front of you. Whether a seat anybody
+// names actually exists, whether two modules declared the same one, and whether a holder
+// satisfies the protocol are facts about the *catalogue* — so they are checked at registration,
+// by CatalogueProblems, which is the last moment the mesh can still say no.
+
+// meshSeatPrefix is reserved to the mesh. The prefix *is* the reservation rule: no list of
+// reserved names to maintain, no way for the mesh's own namespace to be colonised by a manifest,
+// and nothing to keep in step when a mesh seat is added.
+const meshSeatPrefix = "mesh-"
+
+// A SeatDeclaration is a role a module offers on the bus: what may be sent to it, what it says,
+// and what it answers. A caller declares that it uses the *seat*, never the module, so the
+// implementation can be replaced under it.
+type SeatDeclaration struct {
+	Name  string `json:"name"`
+	Scope string `json:"scope,omitempty"`
+
+	// Accepts are the verbs others may submit work on. Each becomes a work-queue subject, and
+	// the holder is the only consumer — so exactly one worker does the job, by construction
+	// rather than by how carefully somebody wrote a subscribe call.
+	Accepts []string `json:"accepts,omitempty"`
+	// Emits are the verbs the holder publishes: 1:many, nobody obliged to act.
+	Emits []string `json:"emits,omitempty"`
+	// Serves are the verbs the holder answers: request and reply, awaited.
+	Serves []string `json:"serves,omitempty"`
+
+	// RetainSeconds is how long the inbound backlog survives with no holder, zero for the
+	// mesh's default. Retention belongs to whoever owns the namespace (design 29 §3) — a seat
+	// owns its own, which is why a seat is also the answer for a module that needs retention
+	// its events cannot have.
+	RetainSeconds int `json:"retain-seconds,omitempty"`
+}
+
+// At is this declaration's scope, with the default applied. Mesh by default, because a seat
+// declared by a module is nearly always "there is one of these in the mesh" — a per-node worker
+// is the deliberate case, and says so.
+func (s SeatDeclaration) At() string {
+	if s.Scope == "" {
+		return ScopeMesh
+	}
+	return s.Scope
+}
+
+// verbs is everything the protocol names, for the checks that do not care which half.
+func (s SeatDeclaration) verbs() []string {
+	out := append([]string{}, s.Accepts...)
+	out = append(out, s.Emits...)
+	return append(out, s.Serves...)
+}
+
+// declaredSeatProblems is what one manifest can be judged on alone.
+func declaredSeatProblems(m Manifest) []string {
+	var problems []string
+	seen := map[string]bool{}
+
+	for _, s := range m.DefinesSeats {
+		switch {
+		case s.Name == "":
+			problems = append(problems, fmt.Sprintf("%s declares a seat with no name", m.Module))
+			continue
+		case !name.MatchString(s.Name):
+			problems = append(problems, fmt.Sprintf(
+				"%s declares a seat named %q, which is not a usable name", m.Module, s.Name))
+			continue
+		case strings.HasPrefix(s.Name, meshSeatPrefix):
+			// The mesh's own code dereferences its seats by name — the resolver *is* the thing
+			// that finds the store — so the prefix is not a convention, it is a namespace.
+			problems = append(problems, fmt.Sprintf(
+				"%s declares a seat named %q; %q is reserved to the mesh, which defines its own "+
+					"seats (novox/hq ADR 0118)", m.Module, s.Name, meshSeatPrefix+"*"))
+			continue
+		}
+		if seen[s.Name] {
+			problems = append(problems, fmt.Sprintf(
+				"%s declares the seat %q twice", m.Module, s.Name))
+			continue
+		}
+		seen[s.Name] = true
+
+		if _, isMesh := SeatNamed(s.Name); isMesh {
+			problems = append(problems, fmt.Sprintf(
+				"%s declares %q, which is a seat the mesh already defines", m.Module, s.Name))
+		}
+		switch s.At() {
+		case ScopeNode, ScopeSite, ScopeMesh:
+		default:
+			problems = append(problems, fmt.Sprintf(
+				"%s declares seat %s at scope %q; a seat is held per node, per site or per mesh",
+				m.Module, s.Name, s.Scope))
+		}
+		// A seat with an empty protocol is allowed, and is the mesh saying what a machine is:
+		// which module is this node's packet filter, or its showcase. Design 26 calls it a seat
+		// that delivers nothing, and that is most of the node-scoped ones. ADR 0126's "a declared
+		// seat carries a protocol" governs what a holder must satisfy, not that every seat offers
+		// something — a marker seat's protocol is satisfied by holding it. Nothing can reach this
+		// state by accident: a mistyped field name is refused by the parser above, so an empty
+		// protocol was written as one.
+		for _, v := range s.verbs() {
+			if !name.MatchString(v) {
+				problems = append(problems, fmt.Sprintf(
+					"%s declares %s.%s, which is not a usable verb", m.Module, s.Name, v))
+			}
+		}
+	}
+
+	for _, u := range m.Uses {
+		if !name.MatchString(u) {
+			problems = append(problems, fmt.Sprintf("%s uses %q, which is not a usable seat name", m.Module, u))
+		}
+	}
+	return problems
+}
+
+// A Shelf is every manifest the mesh has registered, by module name.
+type Shelf map[string]Manifest
+
+// CatalogueProblems are the rules no single manifest can be judged against.
+//
+// Run at registration, which is the last moment the mesh can still refuse: after it, a caller is
+// bound to a seat and a refusal is an outage rather than a conversation.
+func CatalogueProblems(shelf Shelf) []string {
+	var problems []string
+
+	// Who declares what, and who declared it first.
+	declaredBy := map[string]string{}
+	declared := map[string]SeatDeclaration{}
+	for _, module := range shelfOrder(shelf) {
+		for _, s := range shelf[module].DefinesSeats {
+			if s.Name == "" {
+				continue
+			}
+			if first, taken := declaredBy[s.Name]; taken {
+				// The second loses. A seat name meaning two different protocols is the failure
+				// nobody could diagnose afterwards — a caller would bind to whichever happened
+				// to register first, and the symptom would appear in the other module.
+				problems = append(problems, fmt.Sprintf(
+					"%s declares the seat %q, which %s already declares; a seat name means one "+
+						"protocol", module, s.Name, first))
+				continue
+			}
+			declaredBy[s.Name] = module
+			declared[s.Name] = s
+		}
+	}
+
+	exists := func(seat string) bool {
+		if _, isMesh := SeatNamed(seat); isMesh {
+			return true
+		}
+		_, ok := declaredBy[seat]
+		return ok
+	}
+
+	for _, module := range shelfOrder(shelf) {
+		m := shelf[module]
+
+		// A `uses` naming nothing is where ADR 0110's guarantee lands under a derived set: the
+		// same refusal, at the same moment, from a set nobody maintains by hand.
+		for _, u := range m.Uses {
+			if !exists(u) {
+				problems = append(problems, fmt.Sprintf(
+					"%s uses the seat %q, which no module declares and the mesh does not define",
+					module, u))
+			}
+		}
+
+		for _, c := range m.Claims {
+			if !exists(c.Name) {
+				problems = append(problems, fmt.Sprintf(
+					"%s claims the seat %q, which no module declares and the mesh does not define",
+					module, c.Name))
+				continue
+			}
+			s, isModuleSeat := declared[c.Name]
+			if !isModuleSeat {
+				continue // a mesh seat: already judged by claimProblems
+			}
+			if c.At() != s.At() {
+				problems = append(problems, fmt.Sprintf(
+					"%s claims %s at scope %q, and %s declares it at %s",
+					module, c.Name, c.At(), declaredBy[c.Name], s.At()))
+			}
+			// A holder that does not answer what the seat promises is a caller's timeout, found
+			// at assignment instead.
+			if missing := unserved(m, s); len(missing) > 0 {
+				problems = append(problems, fmt.Sprintf(
+					"%s claims %s but does not serve %s, which that seat's protocol promises",
+					module, c.Name, strings.Join(missing, ", ")))
+			}
+		}
+	}
+	sort.Strings(problems)
+	return problems
+}
+
+// unserved is what a seat's protocol promises and the claimant does not answer. Only the tools
+// are checked: `accepts` and `emits` are wired by the runtime from the declaration, while a tool
+// is code the module either has or has not written.
+func unserved(m Manifest, s SeatDeclaration) []string {
+	has := map[string]bool{}
+	for _, t := range m.Tools {
+		has[t] = true
+	}
+	var missing []string
+	for _, t := range s.Serves {
+		if !has[t] {
+			missing = append(missing, t)
+		}
+	}
+	return missing
+}
+
+// shelfOrder is the catalogue in a stable order, so two runs report the same problems in the same
+// sequence — a refusal that reorders itself is a refusal nobody can diff.
+func shelfOrder(shelf Shelf) []string {
+	out := make([]string, 0, len(shelf))
+	for k := range shelf {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}

internal/catalogue/seats_declared_test.go

@@ -0,0 +1,108 @@
+package catalogue
+
+import (
+	"strings"
+	"testing"
+)
+
+func problemsFor(t *testing.T, shelf Shelf) string {
+	t.Helper()
+	return strings.Join(CatalogueProblems(shelf), "; ")
+}
+
+func telegram() Manifest {
+	return Manifest{Module: "telegram", Tools: []string{"status"}, DefinesSeats: []SeatDeclaration{{
+		Name: "telegram-sender", Scope: ScopeMesh,
+		Accepts: []string{"send"}, Emits: []string{"delivered", "failed"}, Serves: []string{"status"},
+	}}, Claims: []Claim{{Name: "telegram-sender", Scope: ScopeMesh}}}
+}
+
+// The whole point: a module contributes a capability without the mesh being changed.
+func TestAModuleDeclaresItsOwnSeatAndHoldsIt(t *testing.T) {
+	shop := Manifest{Module: "shop", Uses: []string{"telegram-sender"}}
+	if got := problemsFor(t, Shelf{"telegram": telegram(), "shop": shop}); got != "" {
+		t.Fatalf("a declared seat and its caller were refused: %s", got)
+	}
+}
+
+// The prefix is the reservation rule, so there is no list to maintain and none to drift.
+func TestAModuleCannotDeclareAMeshSeat(t *testing.T) {
+	for _, n := range []string{"mesh-broker", "mesh-anything", "mesh-store"} {
+		m := Manifest{Module: "impostor", DefinesSeats: []SeatDeclaration{{Name: n, Accepts: []string{"x"}}}}
+		got := strings.Join(declaredSeatProblems(m), "; ")
+		if !strings.Contains(got, "reserved to the mesh") {
+			t.Fatalf("%q was accepted as a module's seat: %q", n, got)
+		}
+	}
+}
+
+// A seat name meaning two protocols is the failure nobody could diagnose afterwards.
+func TestTwoModulesCannotDeclareTheSameSeat(t *testing.T) {
+	other := Manifest{Module: "aardvark", DefinesSeats: []SeatDeclaration{{
+		Name: "telegram-sender", Scope: ScopeMesh, Accepts: []string{"something-else"}}}}
+	got := problemsFor(t, Shelf{"telegram": telegram(), "aardvark": other})
+	if !strings.Contains(got, "already declares") {
+		t.Fatalf("both declarations stood: %s", got)
+	}
+	// The first declarer keeps it; only the second is refused.
+	if strings.Count(got, "already declares") != 1 {
+		t.Fatalf("expected exactly one refusal: %s", got)
+	}
+}
+
+// Where ADR 0110's guarantee lands under a derived set: a typo is refused, not resolved to
+// nothing at runtime.
+func TestUsingASeatNobodyDeclaresIsRefused(t *testing.T) {
+	shop := Manifest{Module: "shop", Uses: []string{"telegram-sendr"}}
+	got := problemsFor(t, Shelf{"telegram": telegram(), "shop": shop})
+	if !strings.Contains(got, "telegram-sendr") || !strings.Contains(got, "no module declares") {
+		t.Fatalf("a misspelled seat was accepted: %s", got)
+	}
+}
+
+// A holder that does not answer what the seat promises is a caller's timeout, found here instead.
+func TestAHolderMustServeWhatItsSeatPromises(t *testing.T) {
+	m := telegram()
+	m.Tools = nil // declares the seat, serves none of it
+	got := problemsFor(t, Shelf{"telegram": m})
+	if !strings.Contains(got, "does not serve status") {
+		t.Fatalf("a holder was accepted that answers nothing its seat promises: %s", got)
+	}
+}
+
+// A seat with no protocol is a marker: which module is this node's showcase, or its packet filter.
+// Most node-scoped seats are markers, so refusing one would refuse the majority of the set.
+func TestASeatWithoutAProtocolIsAMarkerNotAMistake(t *testing.T) {
+	m := Manifest{Module: "vague", DefinesSeats: []SeatDeclaration{{Name: "something", Scope: ScopeNode}}}
+	if got := strings.Join(declaredSeatProblems(m), "; "); got != "" {
+		t.Fatalf("a marker seat was refused: %s", got)
+	}
+}
+
+// A claim at the wrong scope is a different seat than the one declared.
+func TestAClaimMustMatchTheDeclaredScope(t *testing.T) {
+	m := telegram()
+	m.Claims = []Claim{{Name: "telegram-sender", Scope: ScopeNode}}
+	got := problemsFor(t, Shelf{"telegram": m})
+	if !strings.Contains(got, "scope") {
+		t.Fatalf("a claim at the wrong scope was accepted: %s", got)
+	}
+}
+
+// The mesh's own seats still work, and are not shadowed by the derived half.
+func TestTheMeshsOwnSeatsAreStillClaimable(t *testing.T) {
+	m := Manifest{Module: "nats", Claims: []Claim{{Name: "mesh-broker", Scope: ScopeMesh}}}
+	if got := problemsFor(t, Shelf{"nats": m}); got != "" {
+		t.Fatalf("a mesh seat was refused by the derived check: %s", got)
+	}
+}
+
+// A refusal that reorders itself between runs is a refusal nobody can diff.
+func TestTheProblemsAreStable(t *testing.T) {
+	shelf := Shelf{"telegram": telegram(), "shop": {Module: "shop", Uses: []string{"nope"}},
+		"other": {Module: "other", Uses: []string{"also-nope"}}}
+	first, second := problemsFor(t, shelf), problemsFor(t, shelf)
+	if first != second {
+		t.Fatalf("unstable:\n%s\n%s", first, second)
+	}
+}

internal/catalogue/seats_test.go

@@ -74,17 +74,40 @@ func claimed(claims string) []byte {
 	return []byte(`{"module":"thing","version":"1","provides":[{"name":"npm-package-registry","scope":"mesh"}],"claims":` + claims + `}`)
 }
 
-func TestAClaimOnASeatTheMeshDoesNotDefineIsRefused(t *testing.T) {
-	_, err := ParseManifest(claimed(`[{"name":"the-anything","scope":"node"}]`))
-	if err == nil {
-		t.Fatal("a module invented a seat by claiming it")
+// **The refusal moved, it did not go** (novox/hq ADR 0118, superseding 0110). A module may now
+// declare its own seats, so whether a claimed seat exists is a fact about the *catalogue* and not
+// about the manifest in front of the parser: a claim on a seat another registered module declares
+// is perfectly good, and the parser cannot tell the two cases apart. So the parser accepts it and
+// registration refuses it — the same guarantee, at the same moment work would otherwise start,
+// from a set nobody maintains by hand.
+func TestAClaimOnASeatNobodyDeclaresIsRefusedAtRegistration(t *testing.T) {
+	m, err := ParseManifest(claimed(`[{"name":"the-anything","scope":"node"}]`))
+	if err != nil {
+		t.Fatalf("the parser judged a claim it cannot judge alone: %v", err)
 	}
-	if !strings.Contains(err.Error(), "the-anything") || !strings.Contains(err.Error(), "not a seat") {
-		t.Fatalf("the refusal does not say the seat is unknown: %v", err)
+
+	problems := CatalogueProblems(Shelf{m.Module: m})
+	if len(problems) == 0 {
+		t.Fatal("a module invented a seat by claiming it, and registration allowed it")
 	}
-	// And it says what the seats are, because "no" without the list sends somebody reading code.
-	if !strings.Contains(err.Error(), "node-packet-filter") {
-		t.Fatalf("the refusal does not list the seats: %v", err)
+	joined := strings.Join(problems, "; ")
+	if !strings.Contains(joined, "the-anything") || !strings.Contains(joined, "no module declares") {
+		t.Fatalf("the refusal does not say the seat is nobody's: %v", problems)
+	}
+}
+
+// And the same claim is fine once something declares that seat, which is the case the parser
+// could not have distinguished.
+func TestAClaimOnASeatAnotherModuleDeclaresIsAccepted(t *testing.T) {
+	claimant, err := ParseManifest(claimed(`[{"name":"the-anything","scope":"node"}]`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	declarer := Manifest{Module: "someone", DefinesSeats: []SeatDeclaration{
+		{Name: "the-anything", Scope: ScopeNode, Accepts: []string{"work"}},
+	}}
+	if problems := CatalogueProblems(Shelf{claimant.Module: claimant, "someone": declarer}); len(problems) != 0 {
+		t.Fatalf("a claim on a declared seat was refused: %v", problems)
 	}
 }
 
@@ -96,8 +119,15 @@ func TestAModuleDefinesAndClaimsItsOwnSeat(t *testing.T) {
 	if _, err := ParseManifest(ok); err != nil {
 		t.Fatalf("a module could not define and claim its own seat: %v", err)
 	}
-	// Claiming a name it neither the mesh nor the module defines is still refused.
-	if _, err := ParseManifest(claimed(`[{"name":"the-anything","scope":"node"}]`)); err == nil {
+	// Claiming a name nobody defines is still refused — but **at registration, not here**: with
+	// seats declared by modules, a claim on a seat *another* module declares is good, and the
+	// parser cannot tell that from an invented name. See
+	// TestAClaimOnASeatNobodyDeclaresIsRefusedAtRegistration.
+	claimant, err := ParseManifest(claimed(`[{"name":"the-anything","scope":"node"}]`))
+	if err != nil {
+		t.Fatalf("the parser judged a claim it cannot judge alone: %v", err)
+	}
+	if len(CatalogueProblems(Shelf{claimant.Module: claimant})) == 0 {
 		t.Fatal("a module claimed a seat nobody defines")
 	}
 	// A module may not carve its seat out of the mesh's own namespace.
@@ -271,11 +301,11 @@ func TestUseSeatsReplacesTheSetButNeverEmptiesIt(t *testing.T) {
 // and a held record naming the old name break nothing after a rename.
 func TestAFormerNameResolvesAfterARename(t *testing.T) {
 	defer func() { UseSeats(DefaultSeats()); UseAliases(nil) }()
-	UseSeats([]Seat{{Name: "mesh-git", Scope: ScopeMesh, Delivers: "git", Decision: "novox/hq ADR 0121"}})
-	UseAliases(map[string]string{"git": "mesh-git"})
+	UseSeats([]Seat{{Name: "git", Scope: ScopeMesh, Delivers: "git", Decision: "novox/hq ADR 0121"}})
+	UseAliases(map[string]string{"git": "git"})
 
 	// The old name resolves to the renamed seat.
-	if s, ok := SeatNamed("git"); !ok || s.Name != "mesh-git" {
+	if s, ok := SeatNamed("git"); !ok || s.Name != "git" {
 		t.Fatalf("the former name did not resolve to the renamed seat: %+v ok=%v", s, ok)
 	}
 	// And a holder recorded under the old name is still found for the provision the seat delivers.

internal/inventory/busrecords.go

@@ -0,0 +1,178 @@
+package inventory
+
+import (
+	"context"
+	"fmt"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// What the bus's user list is derived from, read out of the mesh's records.
+//
+// The deriving itself is pure and lives in the broker package; this is the reading, and it is kept
+// apart for the reason that package keeps its own types: a permission must be a function of what a
+// module declared, and a query that decided anything would be a second place authority came from.
+
+// BusRecords is every fact the composer needs about who may reach the bus.
+//
+// **A module's authority comes from the manifest, not from the assignment.** The assignment says
+// *where* it runs; what it may say is in what it declared, so the two are read together and the
+// manifest is the one that decides.
+func (i *Inventory) BusRecords(ctx context.Context) (broker.Records, error) {
+	nodes, err := i.Nodes(ctx)
+	if err != nil {
+		return broker.Records{}, fmt.Errorf("cannot read the mesh's machines: %w", err)
+	}
+	declared, err := i.Catalogue(ctx)
+	if err != nil {
+		return broker.Records{}, fmt.Errorf("cannot read the catalogue: %w", err)
+	}
+
+	// Every seat any module declares, by name, so a module's claim can be resolved to the protocol
+	// that seat promises. **Across the whole catalogue, not one manifest**: a seat is declared by
+	// one module and held by another, which is the whole reason a seat exists (ADR 0118).
+	seats := map[string]catalogue.SeatDeclaration{}
+	for _, m := range declared {
+		for _, s := range m.DefinesSeats {
+			seats[s.Name] = s
+		}
+	}
+	// And the mesh's own, which carry protocol too (novox/hq ADR 0121). Added after the modules'
+	// rather than before, because a `mesh-*` name is the mesh's and registration refuses a module
+	// declaring one — so this cannot be shadowed, and if it ever were, the mesh's own would win.
+	for _, own := range catalogue.SeatsWithAProtocol() {
+		seats[own.Name] = catalogue.SeatDeclaration{
+			Name: own.Name, Scope: own.Scope,
+			Accepts: own.Accepts, Emits: own.Emits, Serves: own.Serves,
+		}
+	}
+
+	out := broker.Records{Assigned: map[string][]broker.Declared{}, People: map[string][]string{}}
+	for _, n := range nodes {
+		out.Nodes = append(out.Nodes, n.Name)
+		modules, err := i.Assigned(ctx, n.Name)
+		if err != nil {
+			return broker.Records{}, fmt.Errorf("cannot read what %s runs: %w", n.Name, err)
+		}
+		for _, module := range modules {
+			m, known := declared[module]
+			if !known {
+				// Assigned and not in the catalogue. Said rather than composed with no authority:
+				// a user with an empty permission list is a module that starts, connects, and is
+				// refused by the server on its first publish — an authorisation error that says
+				// nothing about a missing manifest.
+				//
+				// **The catalogue refuses to forget an assigned module, so this is the second line
+				// and not the first.** It earns its place there anyway: relying on another
+				// package's invariant is how a rule ends up enforced by nothing.
+				return broker.Records{}, fmt.Errorf(
+					"%s is assigned to %s and is not in the catalogue, so what it may say cannot "+
+						"be derived", module, n.Name)
+			}
+			out.Assigned[n.Name] = append(out.Assigned[n.Name], declaredFor(m, seats))
+		}
+	}
+
+	enrolling, err := i.NodesWithALiveToken(ctx)
+	if err != nil {
+		return broker.Records{}, err
+	}
+	out.Enrolling = enrolling
+
+	people, err := i.People(ctx)
+	if err != nil {
+		return broker.Records{}, err
+	}
+	for _, p := range people {
+		out.People[p.Name] = p.Invokes
+	}
+	return out, nil
+}
+
+// declaredFor is one module's manifest as the composer needs it: what it says about itself, and the
+// protocol of every seat it holds or uses.
+func declaredFor(m catalogue.Manifest, seats map[string]catalogue.SeatDeclaration) broker.Declared {
+	// A consumed name is a module's event unless it names a seat, and only somebody holding the seat
+	// set can tell (novox/hq ADR 0121). Split here, because the composer cannot look at a name and
+	// know — and a role's event read as a module's is a subscription to a namespace nobody owns.
+	var fromModules []string
+	var watches []broker.Seat
+	for _, c := range m.Consumes {
+		emitter, event, named := strings.Cut(c, ".")
+		if named {
+			if s, isASeat := seats[emitter]; isASeat {
+				watches = append(watches, broker.Seat{Name: s.Name, Emits: []string{event}})
+				continue
+			}
+		}
+		fromModules = append(fromModules, c)
+	}
+
+	d := broker.Declared{
+		Module:   m.Module,
+		Emits:    m.Emits,
+		Consumes: fromModules,
+		Watches:  watches,
+		// The tools it answers, which is `tools` and not `serves`: the manifest's `serves` is the
+		// facts a consumer needs to reach a provision, a different meaning under a similar word.
+		Serves: m.Tools,
+	}
+	for _, c := range m.Claims {
+		// Every seat with a protocol, the mesh's own included. One that says only who does a job is
+		// not here and grants nothing, which is most of them.
+		if s, hasAProtocol := seats[c.Name]; hasAProtocol {
+			d.Holds = append(d.Holds, asSeat(s))
+		}
+	}
+	for _, name := range m.Uses {
+		if s, declaredSomewhere := seats[name]; declaredSomewhere {
+			d.Uses = append(d.Uses, asSeat(s))
+		}
+	}
+	return d
+}
+
+func asSeat(s catalogue.SeatDeclaration) broker.Seat {
+	return broker.Seat{Name: s.Name, Accepts: s.Accepts, Emits: s.Emits, Serves: s.Serves}
+}
+
+// MeshSeats are the mesh's own seats that carry a protocol, as the bus needs them: what to make a work
+// queue for, and whose holder gets a worker on it (novox/hq ADR 0121).
+func MeshSeats() []broker.DeclaredSeat {
+	var out []broker.DeclaredSeat
+	for _, s := range catalogue.SeatsWithAProtocol() {
+		out = append(out, broker.DeclaredSeat{
+			Name: s.Name, Accepts: s.Accepts, Emits: s.Emits, Serves: s.Serves,
+		})
+	}
+	return out
+}
+
+// NodesWithALiveToken is every machine holding a token that could still be presented — issued, not
+// expired, not redeemed.
+//
+// **One enrolment user per such token** (design 25 §6): the inbox an answer goes to is scoped to the
+// token, because an answer carries that machine's credentials sealed to it and a shared inbox is one
+// machine able to read another's.
+func (i *Inventory) NodesWithALiveToken(ctx context.Context) ([]string, error) {
+	rows, err := i.store.Pool().Query(ctx,
+		`select distinct n.name
+		   from enrolment_token t join node n on n.id = t.node
+		  where t.redeemed is null and t.expires > now()
+		  order by n.name`)
+	if err != nil {
+		return nil, fmt.Errorf("cannot read which machines hold a live token: %w", err)
+	}
+	defer rows.Close()
+	var out []string
+	for rows.Next() {
+		var name string
+		if err := rows.Scan(&name); err != nil {
+			return nil, err
+		}
+		out = append(out, name)
+	}
+	return out, rows.Err()
+}

internal/inventory/busrecords_test.go

@@ -0,0 +1,164 @@
+package inventory
+
+import (
+	"context"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// Reading the bus's user list out of the mesh's records, against a real store.
+//
+// What each of these is about is a user that would be **missing or wrong in a way nothing reports**:
+// the server reads whatever file it is given, and a module whose user is absent fails on its first
+// publish with an authorisation error that says nothing about a missing assignment.
+
+func aMeshWith(t *testing.T, manifests ...catalogue.Manifest) (*Inventory, context.Context) {
+	t.Helper()
+	inv := ForTest(t)
+	ctx := context.Background()
+	for _, m := range manifests {
+		if err := inv.RegisterModule(ctx, m, Source{Repository: "/r"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return inv, ctx
+}
+
+func theSeatDeclarer() catalogue.Manifest {
+	return catalogue.Manifest{
+		Module: "telegram", Version: "1",
+		DefinesSeats: []catalogue.SeatDeclaration{{
+			Name: "telegram-sender", Accepts: []string{"send"}, Emits: []string{"delivered"},
+		}},
+		Claims: []catalogue.Claim{{Name: "telegram-sender", Scope: catalogue.ScopeMesh}},
+	}
+}
+
+// A module assigned to a machine becomes a user with the authority its manifest declared — and the
+// protocol of a seat declared by a *different* module, which is the whole reason a seat exists.
+func TestAnAssignedModuleBecomesAUserWithWhatItDeclared(t *testing.T) {
+	shop := catalogue.Manifest{
+		Module: "shop", Version: "1",
+		Emits: []string{"order.placed"}, Tools: []string{"price"},
+		Uses: []string{"telegram-sender"},
+	}
+	inv, ctx := aMeshWith(t, theSeatDeclarer(), shop)
+	if _, err := inv.AddNode(ctx, "one"); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := inv.Assign(ctx, "one", "shop"); err != nil {
+		t.Fatal(err)
+	}
+
+	records, err := inv.BusRecords(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	on := records.Assigned["one"]
+	if len(on) != 1 || on[0].Module != "shop" {
+		t.Fatalf("the machine's modules read as %+v", on)
+	}
+	if len(on[0].Uses) != 1 || on[0].Uses[0].Accepts[0] != "send" {
+		t.Fatalf("the seat it uses carries no protocol: %+v — so it would be granted nothing on a "+
+			"seat it was assigned to send to", on[0].Uses)
+	}
+	if len(on[0].Serves) != 1 || on[0].Serves[0] != "price" {
+		t.Fatalf("its tools read as %v, and a module that cannot subscribe its own tool subject "+
+			"serves nothing", on[0].Serves)
+	}
+
+	// And it derives into a user the server would accept.
+	users, err := broker.Users(records)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var found bool
+	for _, u := range users {
+		if u.Username() != "one.shop" {
+			continue
+		}
+		found = true
+		perms, err := broker.PermissionsFor(u)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if !granted(perms.Publish, "mesh.mod.shop.event.order.placed") ||
+			!granted(perms.Publish, "mesh.seat.telegram-sender.accept.send") ||
+			!granted(perms.Subscribe, "mesh.mod.shop.tool.price") {
+			t.Fatalf("one.shop's authority is not what it declared: %+v", perms)
+		}
+	}
+	if !found {
+		t.Fatal("no user was derived for the assigned module")
+	}
+}
+
+// A machine holding a live token gets an enrolment user; one whose token is spent or expired does
+// not. **An enrolment user outliving its token is a right to join that nobody issued.**
+func TestOnlyAMachineWithALiveTokenHasAnEnrolmentUser(t *testing.T) {
+	inv, ctx := aMeshWith(t)
+	for _, name := range []string{"live", "expired", "none"} {
+		if _, err := inv.AddNode(ctx, name); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if _, err := inv.IssueToken(ctx, "live", time.Hour); err != nil {
+		t.Fatal(err)
+	}
+	// Briefly, then waited out: a token with no lifetime is refused at issue, which is the right
+	// refusal and leaves this as the way to have an expired one.
+	if _, err := inv.IssueToken(ctx, "expired", 10*time.Millisecond); err != nil {
+		t.Fatal(err)
+	}
+	time.Sleep(50 * time.Millisecond)
+
+	records, err := inv.BusRecords(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Join(records.Enrolling, ",") != "live" {
+		t.Fatalf("machines with a live token read as %v", records.Enrolling)
+	}
+}
+
+// A module assigned and absent from the catalogue is refused rather than composed with no authority.
+//
+// **The catalogue refuses to forget an assigned module, so this is the second line and not the
+// first** — and it earns its place there: relying on another package's invariant is how a rule ends
+// up enforced by nothing. Checked against the derivation directly, because the situation cannot be
+// reached through the store.
+func TestAnAssignmentWithNoManifestDerivesNoAuthority(t *testing.T) {
+	// What BusRecords would have produced had it composed a ghost: a module with nothing declared.
+	users, err := broker.Users(broker.Records{
+		Nodes:    []string{"one"},
+		Assigned: map[string][]broker.Declared{"one": {{Module: "ghost"}}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	perms, err := broker.PermissionsFor(users[len(users)-1])
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Its inbox and its ack subject, and nothing it could say. That is a module which starts,
+	// connects, and is refused by the server on its first publish — an authorisation error that
+	// says nothing about a missing manifest, which is why BusRecords names it instead.
+	for _, p := range perms.Publish {
+		if strings.HasPrefix(p, "mesh.mod.ghost.event.") {
+			t.Fatalf("a module with no manifest was granted %s", p)
+		}
+	}
+}
+
+func granted(all []string, one string) bool {
+	for _, s := range all {
+		if s == one {
+			return true
+		}
+	}
+	return false
+}

internal/inventory/bususers.go

@@ -0,0 +1,232 @@
+package inventory
+
+import (
+	"context"
+	"crypto/rand"
+	"encoding/base64"
+	"errors"
+	"fmt"
+
+	"github.com/jackc/pgx/v5"
+	"golang.org/x/crypto/bcrypt"
+)
+
+// The bus's own users, as records.
+//
+// **Only the credential is kept here.** A user's *authority* is derived from what its module
+// declares, every time the file is written (novox/hq ADR 0043) — a stored copy of a permission list
+// would be a second account of a user's authority, able to disagree with the first, and the
+// disagreement would be invisible until somebody compared a composed file with a manifest.
+//
+// What cannot be derived is the password, and on the bus being built it has to outlive its own
+// minting: the whole user list is one file, rewritten whenever any of it changes, so a person's
+// access change would blank every module's password if the mesh kept nothing (design 25 §4, and the
+// migration beside this).
+
+// BusUser is one user of the bus, as the mesh records it.
+type BusUser struct {
+	Username string
+	Kind     string
+	Node     string
+	Module   string
+	// PasswordHash is what the composed file carries. The plaintext is returned once, by Mint, and
+	// then exists only where it was sealed.
+	PasswordHash string
+}
+
+// The kinds of bus user the mesh records. The same words the composer uses, so a row and a
+// principal do not need a translation table between them.
+const (
+	BusController = "controller"
+	BusNode       = "node"
+	BusModule     = "module"
+	BusEnrolment  = "enrolment"
+	BusPerson     = "person"
+)
+
+// MintBusPassword makes a bus password and records its hash under a username, replacing whatever was
+// there, and returns the plaintext **once**.
+//
+// **Once is the whole contract.** The caller seals it to whoever will use it — into an enrolment
+// reply, into a module's sealed environment — and the mesh keeps only the hash, so a credential is
+// never recoverable from the store. A caller that loses it must mint again, which is a rotation and
+// is meant to feel like one.
+func (i *Inventory) MintBusPassword(ctx context.Context, u BusUser) (string, error) {
+	if u.Username == "" || u.Kind == "" {
+		return "", errors.New("a bus user needs a username and a kind")
+	}
+	raw := make([]byte, 32)
+	if _, err := rand.Read(raw); err != nil {
+		return "", fmt.Errorf("cannot generate a bus password: %w", err)
+	}
+	password := base64.RawURLEncoding.EncodeToString(raw)
+
+	// The cost the server will pay on every connection. Left at the library's default rather than
+	// raised: a node reconnecting after a network blip pays it, and the mesh's own links reconnect
+	// far more often than a person logs in anywhere.
+	hash, err := bcrypt.GenerateFromPassword([]byte(password), bcrypt.DefaultCost)
+	if err != nil {
+		return "", fmt.Errorf("cannot hash a bus password: %w", err)
+	}
+
+	if _, err := i.store.Pool().Exec(ctx,
+		`insert into bus_user (username, kind, node, module, password_hash)
+		 values ($1, $2, $3, $4, $5)
+		 on conflict (username) do update
+		   set kind = excluded.kind, node = excluded.node, module = excluded.module,
+		       password_hash = excluded.password_hash, minted_at = now()`,
+		u.Username, u.Kind, u.Node, u.Module, string(hash)); err != nil {
+		return "", fmt.Errorf("cannot record the bus user %s: %w", u.Username, err)
+	}
+	return password, nil
+}
+
+// BusUsers is every user the composed file should contain, by username.
+//
+// Returned as a map because the composer asks by username: the principals are derived from records
+// elsewhere, and this is only what each one's password is. A principal with no row here has no
+// password, and the composer refuses it rather than writing a user anybody is.
+func (i *Inventory) BusUsers(ctx context.Context) (map[string]BusUser, error) {
+	rows, err := i.store.Pool().Query(ctx,
+		`select username, kind, node, module, password_hash from bus_user order by username`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := map[string]BusUser{}
+	for rows.Next() {
+		var u BusUser
+		if err := rows.Scan(&u.Username, &u.Kind, &u.Node, &u.Module, &u.PasswordHash); err != nil {
+			return nil, err
+		}
+		out[u.Username] = u
+	}
+	return out, rows.Err()
+}
+
+// BusUserHash is one user's hash, or false when the mesh has never minted one for it.
+func (i *Inventory) BusUserHash(ctx context.Context, username string) (string, bool, error) {
+	var hash string
+	err := i.store.Pool().QueryRow(ctx,
+		`select password_hash from bus_user where username = $1`, username).Scan(&hash)
+	if errors.Is(err, pgx.ErrNoRows) {
+		return "", false, nil
+	}
+	return hash, err == nil, err
+}
+
+// ForgetBusUser removes one user, so the next composition does not contain it.
+//
+// **Removal is what makes revocation real here.** On a bus with a management call, deleting an
+// account ends its connections; here the credential stops working when the file no longer names it,
+// which is the next composition — so forgetting the row and composing are one act, and a caller
+// that does the first without the second has revoked nothing.
+func (i *Inventory) ForgetBusUser(ctx context.Context, username string) error {
+	_, err := i.store.Pool().Exec(ctx, `delete from bus_user where username = $1`, username)
+	return err
+}
+
+// ForgetBusUsersOf removes every user belonging to one node — its host's, and every module assigned
+// to it. What a forgotten node leaves behind on the bus is otherwise a set of credentials for a
+// machine the mesh no longer knows.
+func (i *Inventory) ForgetBusUsersOf(ctx context.Context, node string) error {
+	if node == "" {
+		return errors.New("forgetting the bus users of no node would forget every user that has none")
+	}
+	_, err := i.store.Pool().Exec(ctx, `delete from bus_user where node = $1`, node)
+	return err
+}
+
+// SeedBusUser records a hash of a credential the mesh did not mint, so a composition contains it.
+//
+// **Genesis is the reason this exists.** The controller's own user is created before the controller
+// runs — by the installer, at a well-known bootstrap password, the way the store's and the old bus's
+// are (`postgres:bootstrap`, `guest:guest`). Nothing minted it, so nothing recorded a hash for it, and
+// the controller's first composition would leave itself out of the very file it was writing: a bus
+// nothing can connect to, produced by the thing connected to it.
+//
+// Idempotent, and it does not overwrite. A credential the mesh *did* mint is the one that counts, so
+// once there is a row this does nothing — otherwise a restart would put the bootstrap password back
+// over a rotated one.
+func (i *Inventory) SeedBusUser(ctx context.Context, u BusUser, password string) error {
+	if u.Username == "" || u.Kind == "" || password == "" {
+		return errors.New("a bus user needs a username, a kind and the credential it is using")
+	}
+	hash, err := bcrypt.GenerateFromPassword([]byte(password), bcrypt.DefaultCost)
+	if err != nil {
+		return fmt.Errorf("cannot hash a bus password: %w", err)
+	}
+	_, err = i.store.Pool().Exec(ctx,
+		`insert into bus_user (username, kind, node, module, password_hash)
+		 values ($1, $2, $3, $4, $5)
+		 on conflict (username) do nothing`,
+		u.Username, u.Kind, u.Node, u.Module, string(hash))
+	return err
+}
+
+// A person who may call the mesh's tools (novox/hq design 25 §7).
+//
+// **Their authority is a list of tools and nothing else.** Not a module: they hold no seat, nothing is
+// addressed to them, nothing is delivered to them, and they have no consumer to acknowledge. What
+// they have is permission to ask.
+
+// Person is somebody who may reach the mesh's tools.
+type Person struct {
+	Name string
+	// Invokes are the tools they may call, each `<module>.<tool>`, or the single entry `*` for an
+	// administrator.
+	Invokes []string
+}
+
+// RecordPerson adds somebody, or changes what they may call.
+//
+// Replacing rather than merging: what a person may call is stated in full, so a change that meant to
+// remove a tool does remove it. A list that could only grow is a permission nobody can take back.
+func (i *Inventory) RecordPerson(ctx context.Context, p Person) error {
+	if p.Name == "" {
+		return errors.New("a person needs a name: it becomes their user on the bus")
+	}
+	if len(p.Invokes) == 0 {
+		return fmt.Errorf(
+			"%s may call nothing, so there is no reason for them to reach the mesh. Name the tools, "+
+				"or `*` for an administrator", p.Name)
+	}
+	_, err := i.store.Pool().Exec(ctx,
+		`insert into person (name, invokes) values ($1, $2)
+		 on conflict (name) do update set invokes = excluded.invokes`,
+		p.Name, p.Invokes)
+	return err
+}
+
+// People is everybody who may reach the mesh's tools.
+func (i *Inventory) People(ctx context.Context) ([]Person, error) {
+	rows, err := i.store.Pool().Query(ctx, `select name, invokes from person order by name`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var out []Person
+	for rows.Next() {
+		var p Person
+		if err := rows.Scan(&p.Name, &p.Invokes); err != nil {
+			return nil, err
+		}
+		out = append(out, p)
+	}
+	return out, rows.Err()
+}
+
+// ForgetPerson removes somebody and the credential they were given.
+//
+// **Both, or neither is a revocation.** A person's row gone and their bus user left behind is a
+// credential that still works and that nothing derives, which is the worst of both: it keeps working
+// and nobody can explain why.
+func (i *Inventory) ForgetPerson(ctx context.Context, name string) error {
+	if name == "" {
+		return errors.New("forgetting nobody would forget everybody")
+	}
+	if _, err := i.store.Pool().Exec(ctx, `delete from person where name = $1`, name); err != nil {
+		return err
+	}
+	return i.ForgetBusUser(ctx, "person."+name)
+}

internal/inventory/bususers_test.go

@@ -0,0 +1,160 @@
+package inventory
+
+import (
+	"context"
+	"testing"
+
+	"golang.org/x/crypto/bcrypt"
+)
+
+// The bus's users as records — against a real store, because what is being checked is that the
+// column exists, the upsert behaves, and a plaintext is returned exactly once.
+
+func aBusUser(module string) BusUser {
+	return BusUser{Username: "one." + module, Kind: BusModule, Node: "one", Module: module}
+}
+
+// The plaintext comes back once and the store keeps only a hash that verifies against it. **A
+// credential recoverable from the mesh's store is one whose blast radius is the store's**, so what
+// is asserted is that the password is not in there.
+func TestABusPasswordIsReturnedOnceAndOnlyItsHashIsKept(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+
+	password, err := inv.MintBusPassword(ctx, aBusUser("shop"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if password == "" {
+		t.Fatal("no password came back, so nothing can be sealed to the module")
+	}
+
+	hash, known, err := inv.BusUserHash(ctx, "one.shop")
+	if err != nil || !known {
+		t.Fatalf("the user was not recorded: %v %v", known, err)
+	}
+	if hash == password {
+		t.Fatal("the store holds the password itself")
+	}
+	if err := bcrypt.CompareHashAndPassword([]byte(hash), []byte(password)); err != nil {
+		t.Fatalf("the recorded hash does not verify the password it was made from: %v", err)
+	}
+}
+
+// Minting again replaces what was there rather than failing or adding a second row: that is a
+// rotation, and the old credential stops working at the next composition.
+func TestMintingAgainRotatesRatherThanAddsAUser(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+
+	first, err := inv.MintBusPassword(ctx, aBusUser("shop"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	second, err := inv.MintBusPassword(ctx, aBusUser("shop"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if first == second {
+		t.Fatal("minting twice produced the same password")
+	}
+	users, err := inv.BusUsers(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(users) != 1 {
+		t.Fatalf("%d users after two mints for one name", len(users))
+	}
+	hash := users["one.shop"].PasswordHash
+	if err := bcrypt.CompareHashAndPassword([]byte(hash), []byte(second)); err != nil {
+		t.Fatal("the kept hash is not the newest password's")
+	}
+	if bcrypt.CompareHashAndPassword([]byte(hash), []byte(first)) == nil {
+		t.Fatal("the previous password still verifies, so a rotation revoked nothing")
+	}
+}
+
+// Forgetting a node takes every credential that belonged to it — its host's and every module
+// assigned to it. What a forgotten node leaves behind otherwise is a working set of credentials for
+// a machine the mesh no longer knows.
+func TestForgettingANodeTakesItsBusUsersWithIt(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+
+	for _, u := range []BusUser{
+		{Username: "node.one", Kind: BusNode, Node: "one"},
+		aBusUser("shop"),
+		{Username: "node.two", Kind: BusNode, Node: "two"},
+		{Username: "controller", Kind: BusController},
+	} {
+		if _, err := inv.MintBusPassword(ctx, u); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := inv.ForgetBusUsersOf(ctx, "one"); err != nil {
+		t.Fatal(err)
+	}
+	users, err := inv.BusUsers(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, still := users["node.one"]; still {
+		t.Fatal("a forgotten node's host credential still works")
+	}
+	if _, still := users["one.shop"]; still {
+		t.Fatal("a module on a forgotten node still has a credential")
+	}
+	// And nothing else went with it: the controller has no node, and another machine's user is
+	// another machine's.
+	for _, kept := range []string{"node.two", "controller"} {
+		if _, ok := users[kept]; !ok {
+			t.Fatalf("%s was removed with another node's users", kept)
+		}
+	}
+}
+
+// Forgetting the users of no node would forget every user that has none — the controller and every
+// person — so it is refused rather than run.
+func TestForgettingTheUsersOfNoNodeIsRefused(t *testing.T) {
+	inv := ForTest(t)
+	if err := inv.ForgetBusUsersOf(context.Background(), ""); err == nil {
+		t.Fatal("forgetting the bus users of no node was allowed")
+	}
+}
+
+// The controller's own user is created by the installer, so the mesh has to be able to record a
+// credential it did not mint — or the first composition leaves the writer out of the file it writes.
+func TestACredentialTheMeshDidNotMintIsRecordedOnceAndNotOverwritten(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+
+	if err := inv.SeedBusUser(ctx, BusUser{Username: "controller", Kind: BusController},
+		"bootstrap"); err != nil {
+		t.Fatal(err)
+	}
+	hash, known, err := inv.BusUserHash(ctx, "controller")
+	if err != nil || !known {
+		t.Fatalf("the credential was not recorded: %v %v", known, err)
+	}
+	if err := bcrypt.CompareHashAndPassword([]byte(hash), []byte("bootstrap")); err != nil {
+		t.Fatalf("what was recorded does not verify the credential given: %v", err)
+	}
+
+	// Minted since, then seeded again — which is what a restart does. The rotation must stand, or
+	// every restart would put the bootstrap password back over it.
+	minted, err := inv.MintBusPassword(ctx, BusUser{Username: "controller", Kind: BusController})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := inv.SeedBusUser(ctx, BusUser{Username: "controller", Kind: BusController},
+		"bootstrap"); err != nil {
+		t.Fatal(err)
+	}
+	hash, _, err = inv.BusUserHash(ctx, "controller")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := bcrypt.CompareHashAndPassword([]byte(hash), []byte(minted)); err != nil {
+		t.Fatal("a restart put the bootstrap credential back over a rotated one")
+	}
+}

internal/inventory/migrations/0037-the-bus-keeps-its-users-hashes.sql

@@ -0,0 +1,36 @@
+-- Every bus user's password hash, because the file has to be written again.
+--
+-- novox/hq design 25 §4, task 1.7. On the bus the mesh runs on today an account is created by a
+-- management call: the mesh mints a password, hands it over, seals the plaintext to whoever will
+-- use it, and keeps nothing. That works because the broker remembers.
+--
+-- The bus being built has no management call — its users are a file the controller composes, and
+-- **the whole file is written every time any of it changes**. So the first person's access change
+-- would silently blank every module's password. The hash has to outlive its own minting, which is
+-- state the mesh did not need before and does now.
+--
+-- Keyed by username, because the username is exactly what the composed file needs and what a
+-- principal derives from its own identity. Nothing else about the user is here: **permissions are
+-- not stored.** They are derived from what each module declares, every time the file is written
+-- (ADR 0043) — a stored copy would be a second account of a user's authority, able to disagree
+-- with the first, and the disagreement would be invisible until somebody compared a file with a
+-- manifest.
+--
+-- The hash and not the password. A file on a node's disk holds the hash, and so does this: a
+-- credential recoverable from the mesh's store is one whose blast radius is the store's.
+create table bus_user (
+    username      text primary key,
+    -- kind and what it names, so a user whose subject is gone can be found and removed: a module
+    -- unassigned, a node forgotten, a token spent. Recorded rather than parsed back out of the
+    -- username, because a name is for the server and a parser over it would be a second grammar.
+    kind          text        not null,
+    node          text        not null default '',
+    module        text        not null default '',
+    password_hash text        not null,
+    minted_at     timestamptz not null default now()
+);
+
+-- Finding every user of one kind, and every user belonging to one node — which is what removing a
+-- node, or composing after an assignment, asks.
+create index bus_user_kind on bus_user (kind);
+create index bus_user_node on bus_user (node) where node <> '';

internal/inventory/migrations/0038-a-person-may-reach-the-mesh.sql

@@ -0,0 +1,19 @@
+-- A person who may call the mesh's tools from a workstation.
+--
+-- novox/hq design 25 §7. Everything else that reaches the bus is a machine or a module running on
+-- one; this is the exception the mesh has always had informally — somebody at a terminal — and never
+-- recorded. Until now "the operator" meant whoever held the keys, which is a role and not a record,
+-- so nothing could say who may call what.
+--
+-- **The authority is a list of tools and nothing else.** A person is not a module: they hold no seat,
+-- nothing is addressed to them, nothing is delivered to them, and they have no consumer to
+-- acknowledge. What they have is permission to ask. That is why there is no scope column and no node
+-- column — a person is not on a machine.
+create table person (
+    name    text        primary key,
+    -- The tools this person may invoke, each `<module>.<tool>`, or the single entry `*` for an
+    -- administrator. Stored as given: the permission is derived from it at every composition, so a
+    -- normalised form here would be a second opinion about authority (novox/hq ADR 0043).
+    invokes text[]      not null default '{}',
+    created timestamptz not null default now()
+);

internal/inventory/nodes.go

@@ -821,6 +821,25 @@ func (i *Inventory) RecordSent(ctx context.Context, node, digest string) error {
 	return err
 }
 
+// Outstanding is the digest of the declaration a machine was last sent, by its name, and empty
+// for one that has never been sent anything.
+//
+// **By name rather than by id**, because the caller is the serving loop and what a node puts in a
+// report is its name. Asked of one machine rather than read from Waiting's sweep, because it is
+// asked per message: a report names the declaration it is about, and a report about one the mesh
+// has already moved past is not acted on (design 25 §3).
+func (i *Inventory) Outstanding(ctx context.Context, name string) (string, error) {
+	var sent string
+	err := i.store.Pool().QueryRow(ctx,
+		`select coalesce(sent, '') from node where name = $1`, name).Scan(&sent)
+	if errors.Is(err, pgx.ErrNoRows) {
+		// Not an error worth carrying up: a report from a machine the mesh has no record of has
+		// nothing to be stale against, and whatever is wrong with it is the listener's to say.
+		return "", nil
+	}
+	return sent, err
+}
+
 // Waiting is every machine whose declaration has changed since it was last sent one.
 //
 // The caller works out what each machine should be now, because only it can — resolution is the

internal/inventory/people_test.go

@@ -0,0 +1,120 @@
+package inventory
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// Somebody who may call the mesh's tools, and what the bus makes of them.
+
+// A person's authority is a list of tools, and it becomes exactly that on the bus — nothing on
+// control, nothing on nodes, nothing they could publish as a module.
+func TestAPersonMayCallToolsAndNothingElse(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+	if err := inv.RecordPerson(ctx, Person{Name: "ada",
+		Invokes: []string{"mesh-catalog.catalog_tools"}}); err != nil {
+		t.Fatal(err)
+	}
+
+	records, err := inv.BusRecords(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := records.People["ada"]; len(got) != 1 || got[0] != "mesh-catalog.catalog_tools" {
+		t.Fatalf("ada may call %v", got)
+	}
+
+	users, err := broker.Users(records)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var found bool
+	for _, u := range users {
+		if u.Username() != "person.ada" {
+			continue
+		}
+		found = true
+		perms, err := broker.PermissionsFor(u)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if len(perms.Publish) != 1 || perms.Publish[0] != "mesh.mod.mesh-catalog.tool.catalog_tools" {
+			t.Errorf("ada may publish %v, which should be the one tool and nothing else", perms.Publish)
+		}
+		for _, s := range perms.Publish {
+			if strings.HasPrefix(s, "mesh.control") || strings.HasPrefix(s, "mesh.node") ||
+				strings.Contains(s, ".event.") {
+				t.Errorf("a person may publish %s — an event would let them claim a module said "+
+					"something, and control is not theirs", s)
+			}
+		}
+		if perms.AllowResponses {
+			t.Error("a person may answer a request, which is impersonating a module on a bus where " +
+				"anyone may serve a tool")
+		}
+	}
+	if !found {
+		t.Fatal("no bus user was derived for a recorded person")
+	}
+}
+
+// Stating what somebody may call replaces what was there. A list that could only grow is a permission
+// nobody can take back.
+func TestChangingWhatAPersonMayCallRemovesWhatIsNotNamed(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+	if err := inv.RecordPerson(ctx, Person{Name: "ada", Invokes: []string{"a.one", "b.two"}}); err != nil {
+		t.Fatal(err)
+	}
+	if err := inv.RecordPerson(ctx, Person{Name: "ada", Invokes: []string{"a.one"}}); err != nil {
+		t.Fatal(err)
+	}
+	people, err := inv.People(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(people) != 1 || len(people[0].Invokes) != 1 || people[0].Invokes[0] != "a.one" {
+		t.Fatalf("ada may call %v; the removed tool is still there", people)
+	}
+}
+
+// Somebody who may call nothing is refused: there is no reason for them to reach the mesh, and an
+// empty list is more likely a mistake than an intention.
+func TestSomebodyWhoMayCallNothingIsRefused(t *testing.T) {
+	inv := ForTest(t)
+	if err := inv.RecordPerson(context.Background(), Person{Name: "ada"}); err == nil {
+		t.Fatal("somebody who may call nothing was recorded")
+	}
+}
+
+// Forgetting somebody takes their credential with them. **Both, or it is not a revocation**: a
+// person's row gone and their bus user left behind is a credential that still works and that nothing
+// derives.
+func TestForgettingAPersonTakesTheirCredential(t *testing.T) {
+	inv := ForTest(t)
+	ctx := context.Background()
+	if err := inv.RecordPerson(ctx, Person{Name: "ada", Invokes: []string{"a.one"}}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := inv.MintBusPassword(ctx, BusUser{Username: "person.ada", Kind: BusPerson}); err != nil {
+		t.Fatal(err)
+	}
+
+	if err := inv.ForgetPerson(ctx, "ada"); err != nil {
+		t.Fatal(err)
+	}
+	if _, known, err := inv.BusUserHash(ctx, "person.ada"); err != nil || known {
+		t.Fatalf("a forgotten person's credential still works: %v %v", known, err)
+	}
+	records, err := inv.BusRecords(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, still := records.People["ada"]; still {
+		t.Fatal("a forgotten person is still composed into the bus")
+	}
+}

internal/link/build.go

@@ -82,6 +82,16 @@ type BuildResult struct {
 	// about the source.
 	On string `json:"on"`
 
+	// Module is what was built, read out of the manifest — the only place it is authoritative, since
+	// a request names a repository and a path.
+	//
+	// **Here because one message now reaches three audiences** (novox/hq ADR 0121). On the bus the
+	// mesh runs on today the answer and the announcement were two publishes to two topologies, so a
+	// result needed no module name and the announcement carried one. On the bus being built the
+	// outcome is the role's own event, and the catalogue reading it needs to know what was built.
+	// Empty on a failed build, which produced no module version.
+	Module string `json:"module,omitempty"`
+
 	// Commit is what was actually built. The mesh records it, which is what makes "is this
 	// current?" answerable without building again.
 	Commit string `json:"commit,omitempty"`

internal/link/builds.go

@@ -0,0 +1,108 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"time"
+)
+
+// Asking a role to build something, and being told what came of it.
+//
+// A build is work submitted to a role, not a message to a machine (novox/hq ADR 0121). The
+// build-machine seat accepts a build and emits an outcome, so the same publish that answers whoever
+// asked also reaches the controller that records it and the catalogue that places it in the module
+// graph — and no build machine needs permission to publish into anybody's inbox.
+//
+// **This is the one flow whose shape differs from every other**, which is why it has its own seam
+// rather than living in `Bus`. Everything else the controller sends is either an event nobody must
+// act on or a declaration a node reconciles toward; a build is a request that takes minutes and has
+// exactly one answer. Too long for request/reply, too particular to be an event.
+
+// TheBuildMachine is the role a build is submitted to.
+const TheBuildMachine = "mesh-build-machine"
+
+// BuildWork is where a build request lands, and BuildOutcome is where its result does. Derived from
+// the seat, so both sides name the role and neither names the other.
+func BuildWork() string    { return "mesh.seat." + TheBuildMachine + ".accept.build" }
+func BuildOutcome() string { return "mesh.seat." + TheBuildMachine + ".event.built" }
+
+// KeyRoleBuilt is the build outcome under the role's name, on the bus the mesh runs on today.
+//
+// The same event as KeyModuleBuilt and published beside it, because a catalogue installed before this
+// change listens for the module's name and one installed after listens for the role's. Both, until
+// this bus retires: a rename needs publisher and subscriber to change together, and a deployment
+// cannot promise which arrives first.
+const KeyRoleBuilt = "built"
+
+// Builders is how work reaches a build machine and how the outcome comes back.
+type Builders interface {
+	// Submit asks for one build and waits for its outcome.
+	//
+	// The wait is long by nature. A build clones, pulls a base image and runs a container build, so
+	// a timeout here says "nothing is doing builds" rather than "this build is slow" — and the two
+	// need different remedies, which is why the message distinguishes them.
+	Submit(ctx context.Context, request BuildRequest, wait time.Duration) (BuildResult, error)
+
+	// Close lets go of whatever was dialled.
+	Close()
+}
+
+// BuildMachine is a machine taking work from the role it holds.
+type BuildMachine interface {
+	// Take hands each request to do until the context ends, and says why it stopped.
+	Take(ctx context.Context, do func(context.Context, Build)) error
+	Close()
+}
+
+// Build is one request a machine has been handed.
+type Build interface {
+	// Request is what to build.
+	Request() BuildRequest
+
+	// Announce publishes the outcome as the role's own event.
+	//
+	// One publish, three audiences: whoever asked matches it by the id their request carried, the
+	// controller records it, and the catalogue places it. On the bus the mesh runs on today that
+	// fan-out came from a shared exchange; here the mesh derived the subject.
+	Announce(ctx context.Context, result BuildResult) error
+
+	// Done settles the request. Called only after the outcome is away, so a machine that dies
+	// before announcing leaves the work for another rather than losing it.
+	Done() error
+
+	// Hold hands the work back for another attempt after the delay.
+	Hold(after time.Duration) error
+}
+
+// waitingFor is the message a caller gets when nothing answered. Its own function because both
+// transports say it, and saying it differently in two places is how one of them ends up vague.
+func waitingFor(wait time.Duration) error {
+	return fmt.Errorf(
+		"no build machine answered within %s. Either nothing holds %s — in which case the work is "+
+			"queued and will be done when something does — or a build is taking longer than this",
+		wait, TheBuildMachine)
+}
+
+// theOutcomeOf reads a result and says whether it is the answer to this request.
+func theOutcomeOf(body []byte, id string) (BuildResult, bool, error) {
+	var result BuildResult
+	if err := json.Unmarshal(body, &result); err != nil {
+		return BuildResult{}, false, fmt.Errorf("a build machine answered with something unreadable: %w", err)
+	}
+	// Somebody else's build. Skipped rather than returned, because returning it would attribute one
+	// build's outcome to another's.
+	return result, result.ID == id, nil
+}
+
+// ModuleOf reads the module's name out of a manifest a build produced, which is the only place it is
+// authoritative — a request named a repository and a path, not a module.
+func ModuleOf(manifest json.RawMessage) string {
+	var named struct {
+		Module string `json:"module"`
+	}
+	if err := json.Unmarshal(manifest, &named); err != nil {
+		return ""
+	}
+	return named.Module
+}

internal/link/builds_current.go

@@ -0,0 +1,202 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+// The build flow on the bus the mesh runs on today.
+//
+// Moved behind the seam rather than changed. The queue, the reply binding and the correlation are
+// what they were, because the mesh is running on this.
+
+// currentBuilds asks for builds over a channel.
+type currentBuilds struct{ channel *amqp.Channel }
+
+// BuildsOverCurrent is the asking side on the bus the mesh has.
+func BuildsOverCurrent(channel *amqp.Channel) Builders { return currentBuilds{channel: channel} }
+
+func (b currentBuilds) Close() {}
+
+func (b currentBuilds) Submit(ctx context.Context, request BuildRequest,
+	wait time.Duration) (BuildResult, error) {
+
+	// Its own queue for the answer, declared before the ask. Consuming from the shared exchange
+	// would mean competing with the controller's own consumer for a message meant for this caller.
+	replies, err := b.channel.QueueDeclare(ReplyQueue(request.ID), false, true, true, false, nil)
+	if err != nil {
+		return BuildResult{}, err
+	}
+	// **A builder never publishes to the default exchange**, because permission there is per
+	// exchange and not per queue — a builder allowed to use it could publish into any node's queue,
+	// which is the privilege a build machine most obviously should not have. The cost is that every
+	// asker sees every result, which is why the correlation is checked below rather than assumed.
+	if err := b.channel.QueueBind(replies.Name, KeyBuilt, Exchange, false, nil); err != nil {
+		return BuildResult{}, err
+	}
+	answers, err := b.channel.ConsumeWithContext(ctx, replies.Name, "", true, true, false, false, nil)
+	if err != nil {
+		return BuildResult{}, err
+	}
+
+	body, err := json.Marshal(request)
+	if err != nil {
+		return BuildResult{}, err
+	}
+	if err := b.channel.PublishWithContext(ctx, "", BuildQueue, false, false, amqp.Publishing{
+		ContentType:   "application/json",
+		DeliveryMode:  amqp.Persistent,
+		CorrelationId: request.ID,
+		ReplyTo:       replies.Name,
+		Body:          body,
+	}); err != nil {
+		return BuildResult{}, err
+	}
+
+	waiting, cancel := context.WithTimeout(ctx, wait)
+	defer cancel()
+	for {
+		select {
+		case <-waiting.Done():
+			return BuildResult{}, waitingFor(wait)
+		case delivery, ok := <-answers:
+			if !ok {
+				return BuildResult{}, errors.New("the connection closed while waiting for a build")
+			}
+			result, mine, err := theOutcomeOf(delivery.Body, request.ID)
+			if err != nil {
+				return BuildResult{}, err
+			}
+			if mine {
+				return result, nil
+			}
+		}
+	}
+}
+
+// --- the machine's side ---------------------------------------------------------------------
+
+type currentMachine struct {
+	conn    *amqp.Connection
+	channel *amqp.Channel
+	on      string
+}
+
+// MachineOverCurrent takes build work over a channel.
+func MachineOverCurrent(conn *amqp.Connection, channel *amqp.Channel, on string) BuildMachine {
+	return &currentMachine{conn: conn, channel: channel, on: on}
+}
+
+func (m *currentMachine) Close() {}
+
+func (m *currentMachine) Take(ctx context.Context, do func(context.Context, Build)) error {
+	if _, err := m.channel.QueueDeclare(BuildQueue, true, false, false, false, nil); err != nil {
+		return err
+	}
+	// One at a time. A machine that took five requests at once would run five container builds
+	// against one runtime and finish all of them slower than it would have finished the first — and
+	// the queue is what shares work between machines, so nothing is lost by it.
+	if err := m.channel.Qos(1, 0, false); err != nil {
+		return err
+	}
+	// Not auto-acknowledged: a request acknowledged on arrival is a build that vanishes if this
+	// process dies mid-way, with nobody waiting on it ever hearing why.
+	requests, err := m.channel.ConsumeWithContext(ctx, BuildQueue, "mesh-builder",
+		false, false, false, false, nil)
+	if err != nil {
+		return err
+	}
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case delivery, ok := <-requests:
+			if !ok {
+				return errors.New("the broker closed the connection")
+			}
+			var request BuildRequest
+			if err := json.Unmarshal(delivery.Body, &request); err != nil {
+				// Unreadable: rejected rather than retried, because the next attempt reads the same
+				// bytes. Nobody waiting hears an answer, which is correct — there was no request.
+				_ = delivery.Reject(false)
+				continue
+			}
+			do(ctx, &currentBuild{request: request, delivery: delivery, on: m.on, channel: m.channel})
+		}
+	}
+}
+
+type currentBuild struct {
+	request  BuildRequest
+	delivery amqp.Delivery
+	on       string
+	channel  *amqp.Channel
+}
+
+func (b *currentBuild) Request() BuildRequest { return b.request }
+
+// Announce answers and announces, which on this bus are two publishes to two exchanges.
+//
+// The reply goes to whoever asked, correlated to their request; the announcement says to the whole
+// mesh that a module now exists at a commit (novox/hq ADR 0072). Only a successful build is
+// announced: a failed one produced no module version, and announcing one would put something in the
+// graph that was never made.
+func (b *currentBuild) Announce(ctx context.Context, result BuildResult) error {
+	body, err := json.Marshal(result)
+	if err != nil {
+		return err
+	}
+	if err := b.channel.PublishWithContext(ctx, Exchange, KeyBuilt, false, false, amqp.Publishing{
+		ContentType:   "application/json",
+		CorrelationId: result.ID,
+		Body:          body,
+	}); err != nil {
+		return fmt.Errorf("cannot answer a build request: %w", err)
+	}
+	if result.Failed != "" || result.Commit == "" {
+		return nil
+	}
+
+	// **Announced under both names on this bus, for exactly as long as this bus lives.**
+	//
+	// A build's outcome belongs to the role now (novox/hq ADR 0121), so a catalogue built from the
+	// current manifests listens for the role's name. A catalogue that is *already running* listens for
+	// the module's, because that is what it was told when it was installed. A rename on a live bus
+	// needs the publisher and the subscriber to change together, and a merge cannot promise that: one
+	// of them is deployed first, and in that window the graph silently stops being updated — which is
+	// the failure this whole change was cleaning up after.
+	//
+	// So both, and the order stops mattering. The module's own name goes with the bus, in step 5's
+	// retirement list; nothing has ever run on the bus being built, so there is no legacy name there
+	// and this doubling has no counterpart.
+	announced := announcementOf(result)
+	if err := EmitEvent(ctx, OverCurrent{Channel: b.channel}, KeyModuleBuilt, "builder", b.on,
+		announced); err != nil {
+		return err
+	}
+	return EmitEvent(ctx, OverCurrent{Channel: b.channel}, KeyRoleBuilt, TheBuildMachine, b.on,
+		announced)
+}
+
+func (b *currentBuild) Done() error { return b.delivery.Ack(false) }
+
+func (b *currentBuild) Hold(time.Duration) error {
+	// No delayed redelivery on this bus: handed back at once, which is what it has always done.
+	return b.delivery.Nack(false, true)
+}
+
+// announcementOf is what the mesh is told about a finished build. One function, so the two
+// transports cannot describe the same build differently.
+func announcementOf(result BuildResult) map[string]any {
+	return map[string]any{
+		"module": ModuleOf(result.Manifest), "commit": result.Commit,
+		"repository": result.Repository, "path": result.Path, "ref": result.Ref,
+		"manifest": json.RawMessage(result.Manifest), "against": result.Against,
+		"made": result.Made,
+	}
+}

internal/link/builds_current_test.go

@@ -0,0 +1,36 @@
+package link
+
+import (
+	"strings"
+	"testing"
+)
+
+// **The old bus announces a build under both names, and that is not belt-and-braces.**
+//
+// A build's outcome belongs to the role now, so a catalogue built from the current manifests listens
+// for the role's name — and a catalogue already running listens for the module's, because that is what
+// it was told when it was installed. A rename on a live bus needs publisher and subscriber to change
+// together, which a deployment cannot promise: one arrives first, and in that window the module graph
+// silently stops being updated.
+//
+// Caught by asking what merging this would do to the mesh that is actually running, which is the only
+// place the question could have been asked — the tests were green and both buses were self-consistent.
+func TestTheOldBusAnnouncesABuildUnderBothNames(t *testing.T) {
+	// The routing key a catalogue installed before the change is bound to.
+	if KeyModuleBuilt != "module.builder.built" {
+		t.Fatalf("the module's own name is %q; a catalogue already running is bound to the old one",
+			KeyModuleBuilt)
+	}
+	// And the local name a catalogue built from the current manifests declares, which the old bus's
+	// client turns into `module.<role>.built`.
+	if KeyRoleBuilt != "built" {
+		t.Fatalf("the role's event is %q, and a holder emits its verbs bare", KeyRoleBuilt)
+	}
+	if TheBuildMachine != "mesh-build-machine" {
+		t.Fatalf("the role is %q", TheBuildMachine)
+	}
+	// The two must differ, or one publish would serve both and this doubling would be pointless.
+	if strings.HasSuffix(KeyModuleBuilt, "."+TheBuildMachine+"."+KeyRoleBuilt) {
+		t.Fatal("the two names are the same, so nothing was renamed and this is dead weight")
+	}
+}

internal/link/builds_nats.go

@@ -0,0 +1,206 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// The build flow on the bus being built.
+//
+// **One publish where the old bus needed two** (novox/hq ADR 0121). There, the answer went to a
+// reply queue and the announcement to an events exchange, because the two audiences were reached by
+// two topologies. Here the outcome is the role's own event: whoever asked matches it by the id their
+// request carried, the controller records it, the catalogue places it in the graph. So a build
+// machine publishes once and needs permission for nothing but its own role's subjects — no reply
+// queue to declare, and no grant over anybody's inbox.
+
+// natsBuilds asks for builds over a connection.
+type natsBuilds struct {
+	js    *broker.JetStream
+	owned bool
+}
+
+// BuildsOverNATS is the asking side on the bus being built. It dials, because the command that asks
+// for a build is a one-shot and holds nothing else.
+func BuildsOverNATS(address string) (Builders, error) {
+	js, err := broker.Dial(address)
+	if err != nil {
+		return nil, fmt.Errorf("cannot reach the bus at %s to ask for a build: %w", address, err)
+	}
+	return &natsBuilds{js: js, owned: true}, nil
+}
+
+func (b *natsBuilds) Close() {
+	if b.owned && b.js != nil {
+		b.js.Close()
+	}
+}
+
+func (b *natsBuilds) Submit(ctx context.Context, request BuildRequest,
+	wait time.Duration) (BuildResult, error) {
+
+	// Subscribed before the ask, so an outcome cannot arrive before there is anywhere for it to
+	// land. Core, not the stream: the asker is waiting now, and the durable copy of this outcome is
+	// the same event on EVENTS, which the controller records.
+	outcomes, err := b.js.Conn().SubscribeSync(BuildOutcome())
+	if err != nil {
+		return BuildResult{}, fmt.Errorf("cannot listen for a build's outcome: %w", err)
+	}
+	defer func() { _ = outcomes.Unsubscribe() }()
+	if err := b.js.Conn().Flush(); err != nil {
+		return BuildResult{}, err
+	}
+
+	body, err := json.Marshal(request)
+	if err != nil {
+		return BuildResult{}, err
+	}
+	// Into the role's work queue and awaited: work the bus never accepted must fail here rather than
+	// be assumed, because nothing else will ever say so.
+	publish, cancel := context.WithTimeout(ctx, 30*time.Second)
+	defer cancel()
+	if _, err := b.js.Context().Publish(BuildWork(), body, nats.Context(publish)); err != nil {
+		return BuildResult{}, fmt.Errorf("cannot submit a build: %w", err)
+	}
+
+	waiting, cancelWait := context.WithTimeout(ctx, wait)
+	defer cancelWait()
+	for {
+		msg, err := outcomes.NextMsgWithContext(waiting)
+		switch {
+		case errors.Is(err, context.DeadlineExceeded):
+			return BuildResult{}, waitingFor(wait)
+		case errors.Is(err, context.Canceled):
+			return BuildResult{}, ctx.Err()
+		case err != nil:
+			return BuildResult{}, fmt.Errorf("waiting for a build's outcome: %w", err)
+		}
+		result, mine, err := theOutcomeOf(msg.Data, request.ID)
+		if err != nil {
+			return BuildResult{}, err
+		}
+		if mine {
+			return result, nil
+		}
+	}
+}
+
+// --- the machine's side ---------------------------------------------------------------------
+
+type natsMachine struct {
+	js   *broker.JetStream
+	on   string
+	seat string
+	sub  *nats.Subscription
+}
+
+// MachineOverNATS takes build work from the role this machine holds.
+func MachineOverNATS(js *broker.JetStream, on string) BuildMachine {
+	return &natsMachine{js: js, on: on, seat: TheBuildMachine}
+}
+
+func (m *natsMachine) Close() {
+	if m.sub != nil {
+		_ = m.sub.Unsubscribe()
+	}
+}
+
+// Take binds to the role's worker and hands each request over, one at a time.
+//
+// **Bound, never created.** The work queue and the worker on it are the controller's to define
+// (design 25 §3), and a build machine reaches no part of the JetStream API — so a missing one is said
+// as the mesh's to answer rather than quietly created with whatever this client defaults to.
+func (m *natsMachine) Take(ctx context.Context, do func(context.Context, Build)) error {
+	worker, found := broker.HolderConsumerFor(m.on, "builder",
+		broker.DeclaredSeat{Name: m.seat, Accepts: []string{"build"}})
+	if !found {
+		return fmt.Errorf("%s accepts no work, so there is nothing for this machine to take", m.seat)
+	}
+
+	// One at a time, which the consumer's own ack-pending limit enforces rather than a prefetch
+	// setting: a machine that took five requests at once would run five container builds against one
+	// runtime and finish all of them slower than the first.
+	work := make(chan *nats.Msg, 1)
+	// **The consumer's own filter, not the one subject this machine cares about.** The client checks
+	// what is asked for against the consumer's filter and refuses anything that is not the same —
+	// "subject does not match consumer" — so subscribing `…accept.build` against a consumer filtered
+	// on `…accept.>` is rejected even though it is narrower. Learned twice now, on two different
+	// consumers, which is why it is written down here.
+	filter := worker.Filters[0]
+	sub, err := m.js.Context().ChanQueueSubscribe(filter, worker.Queue, work,
+		nats.Bind(worker.Stream, worker.Name), nats.ManualAck())
+	if err != nil {
+		return fmt.Errorf(
+			"this machine cannot take work from %s: %w. The mesh creates that queue and this "+
+				"machine's worker on it, and a build machine may not create one itself — so this is "+
+				"the mesh's to answer, not this machine's", m.seat, err)
+	}
+	m.sub = sub
+
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case msg, ok := <-work:
+			if !ok {
+				return errors.New("the bus stopped delivering build work")
+			}
+			var request BuildRequest
+			if err := json.Unmarshal(msg.Data, &request); err != nil {
+				// Unreadable: terminated rather than retried, because the next attempt reads the same
+				// bytes. Nobody waiting hears an answer, which is right — there was no request.
+				_ = msg.Term()
+				continue
+			}
+			do(ctx, &natsBuild{request: request, msg: msg, on: m.on, js: m.js})
+		}
+	}
+}
+
+type natsBuild struct {
+	request BuildRequest
+	msg     *nats.Msg
+	on      string
+	js      *broker.JetStream
+}
+
+func (b *natsBuild) Request() BuildRequest { return b.request }
+
+// Announce publishes the outcome as the role's own event, once, for all three audiences.
+//
+// Into the stream, so a controller that was restarting still records it and a catalogue that was
+// down still catches up. The asker is listening on core for the same subject and gets it either way:
+// a stream delivers to its durable consumers and the plain subscribers both.
+func (b *natsBuild) Announce(ctx context.Context, result BuildResult) error {
+	// One body, three readers. Whoever asked matches the id; the controller records it; the catalogue
+	// needs to know what was built, which only the manifest says — so the result carries it rather
+	// than a second message carrying a second shape.
+	//
+	// **A failed build names no module**, because it produced no module version and the catalogue
+	// would otherwise put something in the graph that was never made. The asker still gets its
+	// answer: a failure is the answer.
+	if result.Failed == "" && result.Commit != "" {
+		result.Module = ModuleOf(result.Manifest)
+	}
+	body, err := json.Marshal(result)
+	if err != nil {
+		return err
+	}
+	publish, cancel := context.WithTimeout(ctx, 30*time.Second)
+	defer cancel()
+	if _, err := b.js.Context().Publish(BuildOutcome(), body, nats.Context(publish)); err != nil {
+		return fmt.Errorf("cannot announce a build's outcome: %w", err)
+	}
+	return nil
+}
+
+func (b *natsBuild) Done() error { return b.msg.Ack() }
+
+func (b *natsBuild) Hold(after time.Duration) error { return b.msg.NakWithDelay(after) }

internal/link/builds_nats_test.go

@@ -0,0 +1,215 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"log"
+	"os"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// A build, end to end, against a real server.
+//
+// The claim worth checking is the one ADR 0121 rests on: **one publish reaches three audiences**.
+// Whoever asked matches the outcome by the id their request carried; the controller records it; the
+// catalogue places it. On the old bus that fan-out came from a shared exchange, and it would be easy
+// to write a version where only the asker hears it and nobody notices for weeks.
+//
+//	docker run -d --rm --name t -p 14230:4222 nats:2.10-alpine -js
+//	MESH_TEST_NATS=nats://127.0.0.1:14230 go test ./internal/link/ -run TestNatsABuild
+
+func aBusWithTheBuildRole(t *testing.T) *broker.JetStream {
+	t.Helper()
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	js, err := broker.Dial(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+
+	seats := []broker.DeclaredSeat{{Name: TheBuildMachine, Accepts: []string{"build"},
+		Emits: []string{"built"}}}
+	if err := broker.AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	if err := broker.RaiseSeats(js, seats, map[string]broker.Holder{
+		TheBuildMachine: {Node: "anchor", Module: "builder"},
+	}); err != nil {
+		t.Fatal(err)
+	}
+	clean := func() {
+		_ = js.Context().DeleteStream("SEAT_MESH_BUILD_MACHINE")
+		for _, s := range broker.MeshStreams() {
+			_ = js.Context().PurgeStream(s.Name)
+		}
+	}
+	t.Cleanup(clean)
+	return js
+}
+
+// The whole round trip: asked, taken, built, and the outcome heard by the asker and by a consumer of
+// the role's event who never asked for anything.
+func TestNatsABuildIsTakenAndItsOutcomeReachesEverybody(t *testing.T) {
+	js := aBusWithTheBuildRole(t)
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+
+	// A third party on the role's event — what the catalogue is. Subscribed first, so nothing is
+	// missed.
+	heard := make(chan BuildResult, 4)
+	watching, err := js.Conn().Subscribe(BuildOutcome(), func(msg *nats.Msg) {
+		var r BuildResult
+		if json.Unmarshal(msg.Data, &r) == nil {
+			heard <- r
+		}
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer func() { _ = watching.Unsubscribe() }()
+	_ = js.Conn().Flush()
+
+	// A build machine holding the role.
+	machine := MachineOverNATS(js, "anchor")
+	defer machine.Close()
+	failed := make(chan error, 1)
+	go func() {
+		failed <- machine.Take(ctx, func(ctx context.Context, work Build) {
+			r := work.Request()
+			_ = work.Announce(ctx, BuildResult{
+				ID: r.ID, Repository: r.Repository, On: "anchor", Commit: "abc1234",
+				Manifest: json.RawMessage(`{"module":"shop"}`),
+			})
+			_ = work.Done()
+		})
+	}()
+
+	ask := &natsBuilds{js: js}
+	result, err := ask.Submit(ctx, BuildRequest{ID: "b-1", Repository: "/r"}, 15*time.Second)
+	if err != nil {
+		select {
+		case why := <-failed:
+			t.Fatalf("the machine could not take work: %v", why)
+		default:
+		}
+		t.Fatalf("the asker never got an outcome: %v", err)
+	}
+	if result.ID != "b-1" || result.Commit != "abc1234" {
+		t.Fatalf("the asker got %+v", result)
+	}
+	// Named in the outcome, because only the manifest says what was built and the catalogue reading
+	// this event needs to know.
+	if result.Module != "shop" {
+		t.Errorf("the outcome names module %q, so a catalogue reading it cannot place the build",
+			result.Module)
+	}
+
+	select {
+	case also := <-heard:
+		if also.ID != "b-1" {
+			t.Fatalf("a third party heard %+v", also)
+		}
+	case <-time.After(5 * time.Second):
+		t.Fatal("nobody but the asker heard the outcome, so the catalogue would never place the build")
+	}
+
+	// And the work left the queue: a request a machine took and settled must not be given to another.
+	deadline := time.Now().Add(5 * time.Second)
+	for time.Now().Before(deadline) {
+		info, err := js.Context().StreamInfo("SEAT_MESH_BUILD_MACHINE")
+		if err == nil && info.State.Msgs == 0 {
+			return
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatal("the request is still queued after being settled, so another machine would build it again")
+}
+
+// Work submitted with no machine holding the role waits rather than failing, and is done when one
+// arrives. **That is what a queue is for**, and the alternative — refusing because nobody is there
+// yet — would make installing a build machine an ordering problem.
+func TestNatsABuildWaitsForAMachineRatherThanFailing(t *testing.T) {
+	js := aBusWithTheBuildRole(t)
+
+	body, _ := json.Marshal(BuildRequest{ID: "b-2", Repository: "/r"})
+	if _, err := js.Context().Publish(BuildWork(), body); err != nil {
+		t.Fatal(err)
+	}
+	info, err := js.Context().StreamInfo("SEAT_MESH_BUILD_MACHINE")
+	if err != nil || info.State.Msgs != 1 {
+		t.Fatalf("the work did not queue: %+v %v", info, err)
+	}
+
+	// Now a machine arrives and finds it waiting.
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	took := make(chan string, 1)
+	machine := MachineOverNATS(js, "anchor")
+	defer machine.Close()
+	go func() {
+		_ = machine.Take(ctx, func(ctx context.Context, work Build) {
+			took <- work.Request().ID
+			_ = work.Announce(ctx, BuildResult{ID: work.Request().ID, On: "anchor", Failed: "no"})
+			_ = work.Done()
+		})
+	}()
+	select {
+	case id := <-took:
+		if id != "b-2" {
+			t.Fatalf("the machine took %q", id)
+		}
+	case <-time.After(10 * time.Second):
+		t.Fatal("a machine that arrived after the work did never got it, so the backlog was lost")
+	}
+}
+
+// A machine that dies before saying anything leaves the work for another.
+func TestNatsWorkAMachineDidNotAnswerGoesBackToTheQueue(t *testing.T) {
+	js := aBusWithTheBuildRole(t)
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+
+	body, _ := json.Marshal(BuildRequest{ID: "b-3", Repository: "/r"})
+	if _, err := js.Context().Publish(BuildWork(), body); err != nil {
+		t.Fatal(err)
+	}
+
+	var once sync.Once
+	handed := make(chan struct{}, 2)
+	machine := MachineOverNATS(js, "anchor")
+	defer machine.Close()
+	go func() {
+		_ = machine.Take(ctx, func(ctx context.Context, work Build) {
+			handed <- struct{}{}
+			// The first time, hand it straight back — a machine that stopped mid-build.
+			var settled bool
+			once.Do(func() { _ = work.Hold(200 * time.Millisecond); settled = true })
+			if !settled {
+				_ = work.Announce(ctx, BuildResult{ID: work.Request().ID, On: "anchor"})
+				_ = work.Done()
+			}
+		})
+	}()
+
+	for i := 0; i < 2; i++ {
+		select {
+		case <-handed:
+		case <-time.After(10 * time.Second):
+			t.Fatalf("the work was handed over %d time(s); unanswered work must come back", i)
+		}
+	}
+}
+
+func quietLog() *log.Logger { return log.New(io.Discard, "", 0) }
+
+var _ = quietLog

internal/link/bus.go

@@ -0,0 +1,194 @@
+package link
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/nats-io/nats.go"
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+// Bus is what the controller needs of the mesh's bus, **in the mesh's own words rather than a
+// transport's** (novox/hq ADR 0116 step 3).
+//
+// Until now every one of these functions took the transport's own channel type, so the transport
+// reached every caller and changing it meant touching all of them. The seam is small — the
+// controller sends exactly two kinds of message that expect no answer, and asks two kinds of
+// question — which is why the bus can be replaced at all.
+//
+// Two implementations live below, and both ship until the rollout (ADR 0116: nothing moves a
+// node's bus before step 5). Both shipping is what makes them comparable — the same caller, the
+// same arguments, and one conformance fixture holding them to one envelope.
+type Bus interface {
+	// PublishEvent announces something that happened, under the emitter's own name. 1:many, and
+	// nobody is obliged to act (ADR 0041).
+	PublishEvent(ctx context.Context, key, source, node string, body []byte) error
+
+	// PublishDeclaration delivers one node what it should be. Addressed to that node alone: a
+	// declaration is not an event, and replaying yesterday's is actively harmful
+	// (design 29 §4, the *state* shape).
+	PublishDeclaration(ctx context.Context, node string, body []byte) error
+
+	// AskTool sends one question to a module's tool and awaits one answer. A tool nobody serves
+	// must say so **at once** rather than after the whole wait: the difference between "that
+	// module is down" and "that tool is slow" is the first thing a person asking wants.
+	AskTool(ctx context.Context, module, tool string, args []byte, timeout time.Duration) ([]byte, error)
+}
+
+// --- The bus the mesh runs on today -----------------------------------------------------
+
+// OverCurrent is the bus the mesh runs on today, until the rollout.
+type OverCurrent struct{ Channel *amqp.Channel }
+
+func (b OverCurrent) PublishEvent(ctx context.Context, key, source, node string, body []byte) error {
+	id, err := eventID()
+	if err != nil {
+		return err
+	}
+	return b.Channel.PublishWithContext(ctx, EventsExchange, key, false, false, amqp.Publishing{
+		ContentType:  "application/json",
+		DeliveryMode: amqp.Persistent,
+		MessageId:    id,
+		Timestamp:    time.Now().UTC(),
+		Body:         body,
+		Headers: amqp.Table{
+			"x-event-id":   id,
+			"x-source":     source,
+			"x-node":       node,
+			"x-time":       time.Now().UTC().Format(time.RFC3339),
+			"content-type": "application/json",
+		},
+	})
+}
+
+// AskTool is implemented over the existing reply-queue machinery in ask.go; this seam does not
+// change how it works today.
+func (b OverCurrent) AskTool(ctx context.Context, module, tool string, args []byte, timeout time.Duration) ([]byte, error) {
+	answer, err := Ask(ctx, b.Channel, module, tool, args, timeout)
+	if err != nil {
+		return nil, err
+	}
+	return answer.Result, nil
+}
+
+func (b OverCurrent) PublishDeclaration(ctx context.Context, node string, body []byte) error {
+	// To the queue directly rather than through an exchange: a declaration is for one node, and
+	// routing it by name through a shared exchange would mean a binding per node that nothing
+	// removes when a node is retired.
+	return b.Channel.PublishWithContext(ctx, "", QueueFor(node), false, false, amqp.Publishing{
+		ContentType:  "application/json",
+		DeliveryMode: amqp.Persistent,
+		Body:         body,
+	})
+}
+
+// --- NATS, the bus being built ----------------------------------------------------------------
+
+// OverNATS is the bus as a JetStream context.
+type OverNATS struct {
+	Conn *nats.Conn
+	JS   nats.JetStreamContext
+}
+
+// The subjects a node publishes on, and the controller listens to.
+//
+// One tree, and each name says who it is about: `mesh.control.<node>.…` is a node's own, which is
+// what lets a node's account be granted exactly its own prefix and nothing of any other node's
+// (design 25 §2, §4). The two that belong to no node — an enrolment, because a machine enrolling
+// has no name the mesh has agreed to yet, and a build's outcome, because a builder is not
+// reporting about itself — are named directly.
+const (
+	// EnrolSubject is where a joining machine asks. Its enrolment user may publish here and
+	// nowhere else, so a leaked token buys nothing but the chance to enrol.
+	EnrolSubject = "mesh.control.enrol"
+
+	// BuiltSubject is where a build's outcome lands, for results nobody was waiting for.
+	BuiltSubject = "mesh.control.built"
+
+	// AliveSubjects is every node's heartbeat. Core NATS, never a stream: a lost heartbeat is the
+	// next heartbeat, and a stream of them is the mesh's least valuable message competing for
+	// retention with its most valuable (design 25 §3).
+	AliveSubjects = "mesh.control.*.alive"
+)
+
+// ReportSubject is where one node says what it did. On the CONTROL stream, because it is the
+// message the store-window guarantee is about (ADR 0083).
+func ReportSubject(node string) string { return "mesh.control." + node + ".report" }
+
+// AliveSubject is one node's heartbeat.
+func AliveSubject(node string) string { return "mesh.control." + node + ".alive" }
+
+// EventSubject is where a module's event lands. Derived from the emitter, never taken from the
+// caller: a source that could differ from the subject is an envelope that can lie about its
+// origin, and on NATS the account's permissions make the subject the authority (design 29 §2).
+func EventSubject(source, key string) string {
+	return "mesh.mod." + source + ".event." + key
+}
+
+// DeclareSubject is where one node's declaration lands. Last-per-subject on the NODES stream, so
+// a node that was away gets exactly the current one and a replayed older one is refused by
+// sequence — the wire-level answer to novox/hq issue 107.
+func DeclareSubject(node string) string { return "mesh.node." + node + ".declare" }
+
+func (b OverNATS) PublishEvent(ctx context.Context, key, source, node string, body []byte) error {
+	id, err := eventID()
+	if err != nil {
+		return err
+	}
+	h := nats.Header{}
+	h.Set("x-event-id", id)
+	h.Set("x-source", source)
+	h.Set("x-node", node)
+	h.Set("x-time", time.Now().UTC().Format(time.RFC3339))
+	h.Set("content-type", "application/json")
+
+	// The id is also the publish's message id, so the server refuses a duplicate inside its
+	// window. That narrows the window a consumer must deduplicate in; it does not remove the
+	// requirement, because the window is finite (design 19, delivery).
+	_, err = b.JS.PublishMsg(&nats.Msg{
+		Subject: EventSubject(source, key),
+		Header:  h,
+		Data:    body,
+	}, nats.MsgId(id), nats.Context(ctx))
+	if err != nil {
+		return fmt.Errorf("emitting %s: %w", key, err)
+	}
+	return nil
+}
+
+func (b OverNATS) PublishDeclaration(ctx context.Context, node string, body []byte) error {
+	_, err := b.JS.Publish(DeclareSubject(node), body, nats.Context(ctx))
+	if err != nil {
+		return fmt.Errorf("declaring to %s: %w", node, err)
+	}
+	return nil
+}
+
+// ToolSubject is where a module answers. Derived from the module and the tool, so a caller names
+// what it wants rather than where it lives.
+func ToolSubject(module, tool string) string { return "mesh.mod." + module + ".tool." + tool }
+
+func (b OverNATS) AskTool(ctx context.Context, module, tool string, args []byte, timeout time.Duration) ([]byte, error) {
+	if len(args) == 0 {
+		args = []byte(`{}`)
+	}
+	ask, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	// **No reply queue, and no correlation to check.** The caller's inbox is its own — each
+	// account is granted one prefix and no other (design 25 §4) — so an answer cannot reach the
+	// wrong asker and there is nothing to correlate against. That also settles a cost recorded
+	// in build.go: on a shared reply exchange every asker saw every result.
+	msg, err := b.Conn.RequestWithContext(ask, ToolSubject(module, tool), args)
+	if err != nil {
+		if errors.Is(err, nats.ErrNoResponders) {
+			// Said at once rather than after the whole wait: nothing is subscribed to that
+			// subject, which is a different fact from a tool being slow.
+			return nil, fmt.Errorf("nothing serves %s.%s", module, tool)
+		}
+		return nil, fmt.Errorf("asking %s.%s: %w", module, tool, err)
+	}
+	return msg.Data, nil
+}

internal/link/bus_test.go

@@ -0,0 +1,154 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+)
+
+// **Both implementations, one fixture.** The point of the seam is not that the transport can be
+// swapped — it is that the two can be held to the same envelope while both ship, so the day the
+// bus moves is a configuration change rather than a discovery.
+//
+// Against a real server, because what the fixture pins is what reaches the wire:
+//
+//	docker run -d --rm --name t -p 14222:4222 nats:2.10-alpine -js
+//	MESH_TEST_NATS=nats://127.0.0.1:14222 go test ./internal/link/ -run TestTheNatsBus
+func TestTheNatsBusEmitsTheEnvelopeTheFixturePins(t *testing.T) {
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	f := loadFixture(t, "events/module-event.json")
+
+	conn, err := nats.Connect(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer conn.Close()
+	js, err := conn.JetStream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := js.AddStream(&nats.StreamConfig{
+		Name: "EVENTS", Subjects: []string{"mesh.mod.*.event.>"},
+	}); err != nil && err != nats.ErrStreamNameAlreadyInUse {
+		t.Fatal(err)
+	}
+
+	bus := OverNATS{JS: js}
+	body, _ := json.Marshal(f.Given.Body)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	if err := EmitEvent(ctx, bus, f.Given.Key, f.Given.Module, f.Given.Node, f.Given.Body); err != nil {
+		t.Fatal(err)
+	}
+
+	// Read back from the stream, not from the thing that wrote it.
+	raw, err := js.GetLastMsg("EVENTS", f.Wire.Subject)
+	if err != nil {
+		t.Fatalf("nothing landed on %s, which the fixture names: %v", f.Wire.Subject, err)
+	}
+	for _, h := range f.Wire.RequiredHeaders {
+		if raw.Header.Get(h) == "" {
+			t.Errorf("%s is not set on the wire, and the fixture requires it", h)
+		}
+	}
+	if got := raw.Header.Get("x-source"); got != f.Given.Module {
+		t.Errorf("x-source is %q; the subject says %q", got, f.Given.Module)
+	}
+	if string(raw.Data) != string(body) {
+		t.Errorf("the payload is %s, expected the body alone: %s", raw.Data, body)
+	}
+	// The envelope must not also be nested inside the payload.
+	var nested map[string]any
+	if json.Unmarshal(raw.Data, &nested) == nil {
+		if _, has := nested["key"]; has {
+			t.Error("the payload carries the envelope's own fields, which the fixture refuses")
+		}
+	}
+}
+
+// The subject a declaration lands on is one node's, and nothing else's — the state shape.
+func TestADeclarationIsAddressedToOneNode(t *testing.T) {
+	if got := DeclareSubject("anchor"); got != "mesh.node.anchor.declare" {
+		t.Fatalf("a declaration would go to %q", got)
+	}
+	if DeclareSubject("anchor") == DeclareSubject("laptop") {
+		t.Fatal("two nodes share a declaration subject, so each would apply the other's")
+	}
+}
+
+// A module cannot emit under another's name: the subject is derived from the source, and the
+// server's permissions make that subject the authority.
+func TestAnEventsSubjectIsDerivedFromItsSource(t *testing.T) {
+	if got := EventSubject("shop", "order.placed"); got != "mesh.mod.shop.event.order.placed" {
+		t.Fatalf("an event would land on %q", got)
+	}
+	if EventSubject("shop", "x") == EventSubject("billing", "x") {
+		t.Fatal("two modules share an event subject, so neither owns its own name")
+	}
+}
+
+// A tool nobody serves says so at once. The difference between "that module is down" and "that
+// tool is slow" is the first thing a person asking wants, and waiting out the timeout to say it
+// is how a fast answer becomes a slow non-answer.
+func TestAskingAToolNobodyServesFailsAtOnce(t *testing.T) {
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	conn, err := nats.Connect(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer conn.Close()
+
+	bus := OverNATS{Conn: conn}
+	began := time.Now()
+	_, err = bus.AskTool(context.Background(), "nobody", "home", nil, 30*time.Second)
+	if err == nil {
+		t.Fatal("asking a tool nothing serves succeeded")
+	}
+	if took := time.Since(began); took > 2*time.Second {
+		t.Errorf("took %v to say nothing serves it; the caller waited out the timeout", took)
+	}
+	if !strings.Contains(err.Error(), "nothing serves") {
+		t.Errorf("the refusal does not say nobody is there: %v", err)
+	}
+}
+
+// And a served tool answers, with no reply queue to declare and no correlation to check.
+func TestAskingAServedToolAnswers(t *testing.T) {
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	conn, err := nats.Connect(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer conn.Close()
+
+	sub, err := conn.Subscribe(ToolSubject("shop", "price"), func(m *nats.Msg) {
+		_ = m.Respond([]byte(`{"total":12}`))
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer sub.Unsubscribe()
+
+	got, err := OverNATS{Conn: conn}.AskTool(context.Background(), "shop", "price",
+		[]byte(`{"qty":4}`), 5*time.Second)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if string(got) != `{"total":12}` {
+		t.Fatalf("the answer came back as %s", got)
+	}
+}

internal/link/conformance_test.go

@@ -0,0 +1,107 @@
+package link
+
+import (
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"regexp"
+	"testing"
+	"time"
+)
+
+// The Go implementation, held to the shared fixtures (novox/hq ADR 0074, design 19).
+//
+// **Read from the sdk's conformance directory by sibling path**, the way the lab finds its
+// siblings — deliberately not copied here. A fixture copied into each implementation is two
+// fixtures, and two fixtures drift, which is the exact failure the suite exists to prevent.
+type fixture struct {
+	Name  string `json:"name"`
+	Given struct {
+		Module  string            `json:"module"`
+		Node    string            `json:"node"`
+		Key     string            `json:"key"`
+		Body    map[string]any    `json:"body"`
+		Headers map[string]string `json:"headers"`
+	} `json:"given"`
+	Wire struct {
+		Subject         string            `json:"subject"`
+		RequiredHeaders []string          `json:"requiredHeaders"`
+		HeaderFormats   map[string]string `json:"headerFormats"`
+	} `json:"wire"`
+}
+
+func loadFixture(t *testing.T, name string) fixture {
+	t.Helper()
+	path := filepath.Join("..", "..", "..", "mesh-sdk", "conformance", name)
+	raw, err := os.ReadFile(path)
+	if err != nil {
+		t.Skipf("the sdk's conformance fixtures are not beside this checkout: %v", err)
+	}
+	var f fixture
+	if err := json.Unmarshal(raw, &f); err != nil {
+		t.Fatalf("%s: %v", name, err)
+	}
+	return f
+}
+
+// Every header the fixture requires is one this implementation actually sets.
+func TestTheGoEmitterSetsEveryRequiredHeader(t *testing.T) {
+	f := loadFixture(t, "events/module-event.json")
+	sent := goEventHeaders(f.Given.Key, f.Given.Module, f.Given.Node)
+	for _, want := range f.Wire.RequiredHeaders {
+		if _, ok := sent[want]; !ok {
+			t.Errorf("the Go emitter does not set %q, which the fixture requires — an event it "+
+				"emits is one a conforming consumer refuses", want)
+		}
+	}
+}
+
+// And each value is in the shape the fixture pins, because a header present but differently
+// formatted is the disagreement that does not announce itself.
+func TestTheGoEmittersHeaderFormatsMatch(t *testing.T) {
+	f := loadFixture(t, "events/module-event.json")
+	sent := goEventHeaders(f.Given.Key, f.Given.Module, f.Given.Node)
+
+	if got := sent["content-type"]; got != f.Wire.HeaderFormats["content-type"] {
+		t.Errorf("content-type is %q, the fixture says %q", got, f.Wire.HeaderFormats["content-type"])
+	}
+	if _, err := time.Parse(time.RFC3339, sent["x-time"]); err != nil {
+		t.Errorf("x-time %q is not RFC3339, which the fixture requires: %v", sent["x-time"], err)
+	}
+	if pattern := f.Wire.HeaderFormats["x-event-id"]; pattern != "" {
+		if !regexp.MustCompile(pattern).MatchString(sent["x-event-id"]) {
+			t.Errorf("x-event-id %q does not match %q", sent["x-event-id"], pattern)
+		}
+	}
+	// The origin the envelope claims is the one the bus enforces by namespace. A disagreement
+	// here means the envelope is lying about where it came from.
+	if sent["x-source"] != f.Given.Module {
+		t.Errorf("x-source is %q for module %q", sent["x-source"], f.Given.Module)
+	}
+}
+
+// The subject a module's event lands on is derived, not carried — so this implementation must
+// derive the same one the fixture names.
+func TestTheGoSubjectMatchesTheFixture(t *testing.T) {
+	f := loadFixture(t, "events/module-event.json")
+	got := "mesh.mod." + f.Given.Module + ".event." + f.Given.Key
+	if got != f.Wire.Subject {
+		t.Errorf("this implementation would publish on %q; the fixture says %q", got, f.Wire.Subject)
+	}
+}
+
+// goEventHeaders is the header set EmitEvent produces, factored so conformance can see it
+// without a broker. Kept beside the emitter so the two cannot drift apart silently.
+func goEventHeaders(eventType, source, node string) map[string]string {
+	id, err := eventID()
+	if err != nil {
+		panic(err)
+	}
+	return map[string]string{
+		"x-event-id":   id,
+		"x-source":     source,
+		"x-node":       node,
+		"x-time":       time.Now().UTC().Format(time.RFC3339),
+		"content-type": "application/json",
+	}
+}

internal/link/declare.go

@@ -5,8 +5,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"time"
-
-	amqp "github.com/rabbitmq/amqp091-go"
 )
 
 // Signer is whatever holds the control plane's signing key.
@@ -21,7 +19,7 @@ type Signer interface {
 // something else, and the node would refuse a declaration that was genuinely the mesh's.
 //
 // Published to the node's own queue, which its account alone may read.
-func Declare(ctx context.Context, channel *amqp.Channel, signer Signer, node string,
+func Declare(ctx context.Context, bus Bus, signer Signer, node string,
 	declaration []byte, timeout time.Duration) error {
 
 	if !json.Valid(declaration) {
@@ -41,13 +39,5 @@ func Declare(ctx context.Context, channel *amqp.Channel, signer Signer, node str
 	publish, cancel := context.WithTimeout(ctx, timeout)
 	defer cancel()
 
-	// Published to the queue directly rather than through the exchange: a declaration is for one
-	// node, and routing it by name through a shared exchange would mean a binding per node that
-	// nothing removes when a node is retired.
-	return channel.PublishWithContext(publish, "", QueueFor(node), false, false,
-		amqp.Publishing{
-			ContentType:  "application/json",
-			DeliveryMode: amqp.Persistent,
-			Body:         body,
-		})
+	return bus.PublishDeclaration(publish, node, body)
 }

internal/link/enrol_bus_credential_test.go

@@ -0,0 +1,107 @@
+package link_test
+
+import (
+	"crypto/ed25519"
+	"crypto/rand"
+	"testing"
+	"time"
+
+	"golang.org/x/crypto/bcrypt"
+
+	"github.com/novox/mesh-controller/internal/identity"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// What a node is given to come back with, on the bus being built.
+//
+// **The credential becomes usable at the next composition, not when it is made**, which is the one
+// real difference from the bus the mesh runs on today: there a management call makes it live at once.
+// So what has to be true here is that the mesh recorded it and told the node, and the rest is a push.
+
+// aMeshReadyToEnrol is both stores with a signing key established, which a control plane does at
+// start: one that cannot sign is one whose declarations every node correctly refuses.
+func aMeshReadyToEnrol(t *testing.T) (*inventory.Inventory, *identity.Identity) {
+	t.Helper()
+	inv := inventory.ForTest(t)
+	ident := identity.ForTest(t)
+	if _, err := ident.Establish(t.Context()); err != nil {
+		t.Fatal(err)
+	}
+	return inv, ident
+}
+
+func aTokenFor(t *testing.T, inv *inventory.Inventory, node string) (string, ed25519.PublicKey) {
+	t.Helper()
+	ctx := t.Context()
+	if _, err := inv.AddNode(ctx, node); err != nil {
+		t.Fatal(err)
+	}
+	issued, err := inv.IssueToken(ctx, node, time.Hour)
+	if err != nil {
+		t.Fatal(err)
+	}
+	public, _, err := ed25519.GenerateKey(rand.Reader)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return issued.Secret, public
+}
+
+// A node enrolling onto the bus being built is told a password of its own, and the mesh keeps only
+// its hash — which is what the next composition writes into the bus's user list.
+func TestANodeEnrollingOnTheNewBusIsMintedACredentialTheMeshOnlyHashes(t *testing.T) {
+	inv, ident := aMeshReadyToEnrol(t)
+	ctx := t.Context()
+	secret, public := aTokenFor(t, inv, "anchor")
+
+	reply, err := link.Enrolment{Inventory: inv, Identity: ident, OnNATS: true}.Enrol(ctx, link.EnrolRequest{
+		Node: "anchor", Secret: secret, PublicKey: public})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if reply.Password == "" {
+		t.Fatal("the node was told no password, so it keeps a one-time secret as a credential")
+	}
+	if reply.Password == secret {
+		t.Fatal("the node was handed the token's own secret back: a credential that lives for years " +
+			"must not be the string that was pasted into a terminal")
+	}
+
+	// Recorded under the name the composed file will use, and as a hash: a credential recoverable
+	// from the mesh's store is one whose blast radius is the store's.
+	hash, known, err := inv.BusUserHash(ctx, "node.anchor")
+	if err != nil || !known {
+		t.Fatalf("the mesh kept no credential for the node it enrolled: %v %v", known, err)
+	}
+	if hash == reply.Password {
+		t.Fatal("the store holds the password itself")
+	}
+	if err := bcrypt.CompareHashAndPassword([]byte(hash), []byte(reply.Password)); err != nil {
+		t.Fatalf("what the mesh kept does not verify what it told the node: %v", err)
+	}
+}
+
+// On the bus the mesh runs on today, with no management configured, nothing is minted and the node is
+// told so by being given no password — it keeps the token's secret, which it says out loud.
+//
+// **This is the check that the switch is a switch.** A node enrolling on one bus must not come away
+// with a credential for the other: it would be half-moved, and nothing anywhere would say which half.
+func TestANodeEnrollingOnTheOldBusIsMintedNoCredentialForTheNewOne(t *testing.T) {
+	inv, ident := aMeshReadyToEnrol(t)
+	ctx := t.Context()
+	secret, public := aTokenFor(t, inv, "anchor")
+
+	reply, err := link.Enrolment{Inventory: inv, Identity: ident}.Enrol(ctx, link.EnrolRequest{
+		Node: "anchor", Secret: secret, PublicKey: public})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if reply.Password != "" {
+		t.Fatalf("a node on the old bus was given a password from nowhere: %q", reply.Password)
+	}
+	if _, known, err := inv.BusUserHash(ctx, "node.anchor"); err != nil || known {
+		t.Fatalf("a node enrolling on the old bus was given a credential for the new one: %v %v",
+			known, err)
+	}
+}

internal/link/enrol_reply_probe_test.go

@@ -0,0 +1,81 @@
+package link
+
+import (
+	"os"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+)
+
+// **Verified 2026-09-27**: the caller asked for a reply to `_INBOX.LCr3M83q…` and the consumer
+// saw `$JS.ACK.PROBE.probe_consumer.1.1.1…`. The design was right, and enrolment's payload-borne
+// reply subject is necessary rather than defensive.
+//
+// Design 25 §2 asserts that a reply address is **eaten** when a message crosses a stream: core
+// request/reply puts the requester's inbox in the message's Reply field, but a JetStream consumer
+// has already claimed that field for its own ack address by the time a handler sees it. The whole
+// enrolment design rests on it — the reply subject travels in the payload instead — so it is
+// checked rather than believed.
+//
+//	docker run -d --rm --name t -p 14222:4222 nats:2.10-alpine -js
+//	MESH_TEST_NATS=nats://127.0.0.1:14222 go test ./internal/link/ -run TestAReplyAddress
+func TestAReplyAddressDoesNotSurviveAStream(t *testing.T) {
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	conn, err := nats.Connect(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer conn.Close()
+	js, err := conn.JetStream()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := js.AddStream(&nats.StreamConfig{
+		Name: "PROBE", Subjects: []string{"probe.>"}, Retention: nats.WorkQueuePolicy,
+	}); err != nil && err != nats.ErrStreamNameAlreadyInUse {
+		t.Fatal(err)
+	}
+	defer js.DeleteStream("PROBE")
+
+	seen := make(chan *nats.Msg, 1)
+	sub, err := js.Subscribe("probe.enrol", func(m *nats.Msg) { seen <- m },
+		nats.Durable("probe_consumer"), nats.ManualAck())
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer sub.Unsubscribe()
+
+	// A caller doing what core request/reply does: publish with its own inbox as the reply.
+	inbox := nats.NewInbox()
+	if err := conn.PublishMsg(&nats.Msg{Subject: "probe.enrol", Reply: inbox, Data: []byte("{}")}); err != nil {
+		t.Fatal(err)
+	}
+
+	select {
+	case m := <-seen:
+		t.Logf("the caller asked for a reply to   %s", inbox)
+		t.Logf("the consumer sees a Reply field of %s", m.Reply)
+		if m.Reply == inbox {
+			t.Fatalf("the reply address SURVIVED the stream. Design 25 §2 says it does not, and " +
+				"builds enrolment around carrying the reply subject in the payload to work " +
+				"around it. If this holds generally, that work is unnecessary and the design " +
+				"should say so.")
+		}
+		if m.Reply == "" {
+			t.Fatal("the Reply field is empty rather than claimed; the design says it carries " +
+				"the consumer's ack address, which is a different fact")
+		}
+		// Answering it would send the enrolling node's credentials to an ack subject.
+		if len(m.Reply) < 7 || m.Reply[:7] != "$JS.ACK" {
+			t.Errorf("the Reply field is %q, which is neither the caller's inbox nor an ack "+
+				"address — the design's reasoning assumes one of the two", m.Reply)
+		}
+		m.Ack()
+	case <-time.After(5 * time.Second):
+		t.Fatal("nothing was delivered")
+	}
+}

internal/link/enrolment.go

@@ -28,6 +28,15 @@ type Enrolment struct {
 	Identity   *identity.Identity
 	Management *broker.Management
 	Broker     broker.Broker
+
+	// OnNATS says the mesh's own traffic is on the bus being built, so a node's credential is
+	// minted into the mesh's records and composed into the bus's user list rather than pushed
+	// through a management call (novox/hq design 25 §4).
+	//
+	// **One bus, and a node gets a credential for exactly one** — refused at start if the
+	// controller is told about both (broker.MustBeOneBus), because a node holding a credential for
+	// each is one that could be half-moved, and nothing would say which half.
+	OnNATS bool
 }
 
 // Enrol records what the node presented and spends the token.
@@ -159,7 +168,33 @@ func (e Enrolment) Enrol(ctx context.Context, request EnrolRequest) (reply Enrol
 	// the spend and not before: a replaced password on an attempt that failed would be held by
 	// nobody. If the broker will not take it now, the enrolment still stands — the node keeps
 	// the token's secret as its password, which it is told, and which is said here.
-	if e.Management != nil {
+	switch {
+	case e.OnNATS:
+		// **Minted into the mesh's records, not pushed to a server.** The bus's users are a file
+		// the controller composes, so a credential becomes usable at the next composition rather
+		// than at the moment it is made — and the plaintext is returned once, here, and then exists
+		// only on the machine it was sealed to.
+		//
+		// The node reconnects as itself and may be refused until that composition reaches the
+		// machine running the bus. That is what the host's reconnect backoff is for and it is
+		// survivable by design (ADR 0004: disconnection is an ordinary situation); waiting for the
+		// push here would hold an enrolment open for as long as a declaration takes to apply.
+		password, err := e.Inventory.MintBusPassword(ctx, inventory.BusUser{
+			Username: broker.Principal{Kind: broker.KindNode, Node: node.Name}.Username(),
+			Kind:     inventory.BusNode,
+			Node:     node.Name,
+		})
+		if err != nil {
+			// Not fatal to the enrolment: the node is recorded and the token is spent, and a node
+			// that keeps the token's secret is told so. Said loudly, because until this is minted
+			// the machine has no credential of its own.
+			log.Printf("%s is enrolled and the mesh could not mint its bus credential, so it keeps "+
+				"the token's secret as its password: %v", node.Name, err)
+		} else {
+			reply.Password = password
+		}
+
+	case e.Management != nil:
 		password, err := freshPassword()
 		if err != nil {
 			return EnrolReply{}, err
@@ -234,6 +269,17 @@ var ErrNoBrokerManagement = errors.New("no broker management configured")
 // A node states; the owning context writes (novox/hq ADR 0006). What a node says it applied is
 // its own account of its own machine, kept as a copy for recovery — so this writes it down and
 // decides nothing from it.
+// Outstanding is the declaration the mesh last sent a node, so a report about an older one is not
+// acted on (design 25 §3, window.go).
+//
+// **Here rather than on Listener.** A report is recorded by whatever keeps records, and a great
+// many things that record reports have no idea what was sent — every test in this package among
+// them. So the serving loop asks for this when the listener happens to be able to answer, and
+// where it cannot, a report has nothing to be stale against and is simply acted on.
+func (e Enrolment) Outstanding(ctx context.Context, node string) (string, error) {
+	return e.Inventory.Outstanding(ctx, node)
+}
+
 func (e Enrolment) Heard(ctx context.Context, report Report) (err error) {
 	// A store that could not be asked right now is said as such, so the report is kept for
 	// another attempt rather than acknowledged and lost (novox/hq issue 082).

internal/link/events.go

@@ -6,9 +6,6 @@ import (
 	"encoding/hex"
 	"encoding/json"
 	"fmt"
-	"time"
-
-	amqp "github.com/rabbitmq/amqp091-go"
 )
 
 // Emitting a module event from Go.
@@ -29,32 +26,17 @@ const (
 
 // EmitEvent publishes one module event, in the envelope the sdk's consumers expect.
 //
-// Persistent, because an event that a broker restart loses is not an announcement. The publish is
-// not confirmed here: the caller has already done the work the event describes, and a build that
-// succeeded must not be reported as failed because saying so failed.
-func EmitEvent(ctx context.Context, channel *amqp.Channel, eventType, source, node string, body any) error {
+// The envelope is the transport's to write (bus.go) and this is only what goes in it, which is
+// what lets one conformance fixture hold both implementations to the same headers.
+func EmitEvent(ctx context.Context, bus Bus, eventType, source, node string, body any) error {
 	payload, err := json.Marshal(body)
 	if err != nil {
 		return fmt.Errorf("cannot serialise a %s event: %w", eventType, err)
 	}
-	id, err := eventID()
-	if err != nil {
-		return err
-	}
-	return channel.PublishWithContext(ctx, EventsExchange, eventType, false, false, amqp.Publishing{
-		ContentType:  "application/json",
-		DeliveryMode: amqp.Persistent,
-		MessageId:    id,
-		Timestamp:    time.Now().UTC(),
-		Body:         payload,
-		Headers: amqp.Table{
-			"x-event-id":   id,
-			"x-source":     source,
-			"x-node":       node,
-			"x-time":       time.Now().UTC().Format(time.RFC3339),
-			"content-type": "application/json",
-		},
-	})
+	// The publish is not confirmed by the caller: it has already done the work the event
+	// describes, and a build that succeeded must not be reported as failed because saying so
+	// failed. Each transport decides what "published" means for it.
+	return bus.PublishEvent(ctx, eventType, source, node, payload)
 }
 
 // eventID is what a consumer deduplicates on: delivery is at-least-once, so a handler must be able

internal/link/protocol.go

@@ -80,6 +80,20 @@ type EnrolRequest struct {
 	// it. Nil from a node that found none, which is every converged one.
 	Tunnel *Tunnel `json:"tunnel,omitempty"`
 
+	// ReplyTo is where the answer goes, as a field of the request rather than the transport's own
+	// reply address.
+	//
+	// **Because a stream eats the transport's field** (design 25 §2, verified against a running
+	// server): a message a JetStream consumer delivers has had its reply field claimed for that
+	// consumer's own ack address, so by the time the controller sees an enrolment, the field names
+	// where the *controller* must acknowledge, not where the node is waiting. An enrolment is the
+	// case that matters — a caller waiting on an ephemeral inbox, over a subject the store window
+	// may legitimately delay by several nak cycles.
+	//
+	// Empty on the bus the mesh runs on today, where the delivery carries the reply queue and the
+	// field means what it has always meant.
+	ReplyTo string `json:"reply_to,omitempty"`
+
 	// Redelivered is set by the control plane, never sent: the broker handed this request over a
 	// second time. Such a request does not finish an enrolment already spent — the first time may
 	// have answered, and the node holds what it was told.

internal/link/receive.go

@@ -0,0 +1,125 @@
+package link
+
+import (
+	"context"
+	"time"
+)
+
+// The consume side of the bus, in the mesh's own words.
+//
+// The outbound half went behind `Bus` (bus.go) and the transport stopped reaching its callers.
+// This is the other half, and it is the larger one: everything a node or a module says arrives
+// here, and until now every handler took the transport's own delivery type — so the serving loop
+// could not be moved to another bus without moving enrolment, reports, builds, upgrades and
+// catch-up with it in one breath.
+//
+// **Two implementations, both shipping** (novox/hq ADR 0116: nothing moves a node's bus before
+// step 5). Both shipping is what makes them comparable, and it is what lets the store-window
+// guarantee (ADR 0083) be stated once — in window.go, pure — rather than twice, once per
+// transport, where the two would eventually disagree about the thing that matters most.
+
+// The kinds of message the controller acts on.
+//
+// A transport maps its own addressing onto these — a routing key on the bus the mesh has, a
+// subject on the one being built — and nothing past this point knows which it was. They are never
+// on the wire: the wire is the transport's business, and a kind that travelled would be a third
+// name for the same thing.
+const (
+	KindEnrolment   = "enrolment"
+	KindReport      = "report"
+	KindHeartbeat   = "heartbeat"
+	KindBuilt       = "built"
+	KindModuleMoved = "module-moved"
+	KindCatchUp     = "catch-up"
+)
+
+// Control is one thing a node or a module said, as the controller must act on it.
+//
+// **Settling is stated as what the mesh means, not as the transport's verbs.** The two buses
+// spell them differently — an ack and a reject against a delivery tag, an ack and a term against
+// a stream sequence — and the guarantee is the same either way: `Took` is done with, `Drop` is
+// understood and not worth another attempt, and `Hold` is the store window, where the message is
+// kept and comes back.
+//
+// A handler that returns without calling any of the three leaves the message unsettled on
+// purpose. That is the right answer while shutting down: a cancelled context is not an answer
+// about a message, and the bus should hand it to whatever consumes next (novox/hq issue 083).
+type Control interface {
+	// Kind is which of the constants above this is.
+	Kind() string
+
+	// Body is the message itself — the payload alone, never the envelope.
+	Body() []byte
+
+	// Redelivered says the bus has handed this message over before. An enrolment cares and
+	// nothing else does: one already spent is not finished a second time.
+	Redelivered() bool
+
+	// HeldFor is how long this message has been waiting to be taken. Zero on a first delivery.
+	//
+	// **Read from the message rather than remembered by the controller.** On the bus being built
+	// it is the age of the publish, which a controller that restarted mid-window still reads
+	// correctly — the whole reason the holding moves into the server. On the bus the mesh has it
+	// is how long this process has held it, which is the most that transport can say.
+	HeldFor() time.Duration
+
+	// Answer replies to whoever is waiting on this message; only an enrolment expects one.
+	//
+	// Each transport knows where its own answer goes, and they do not agree about it: one carries
+	// a reply queue in the delivery, and on the other the field that would have carried it has
+	// been claimed by the consumer's own ack subject, so the address travels in the payload
+	// (design 25 §2, verified). That difference is exactly what this seam exists to keep out of
+	// the handler.
+	Answer(ctx context.Context, body []byte) error
+
+	// Took settles the message: acted on, or understood and needing no action.
+	Took() error
+
+	// About names what this message is about — a node's report, one module's move, one build's
+	// outcome — and is said before the store is asked.
+	//
+	// A transport that holds messages **in memory** uses it to set aside anything older it is
+	// holding about the same thing: the older is the past, and letting it come back after the
+	// newer was acted on would undo the newer.
+	//
+	// **This is the one thing holding-in-memory can do that holding-in-the-server cannot**, and
+	// naming it here rather than hiding it is deliberate. On the bus being built the message
+	// belongs to the server and comes back whatever happened meanwhile, so this is ignored and the
+	// digest a report carries answers the same question instead (window.go, design 25 §3).
+	About(what string)
+
+	// Hold keeps the message and asks for it again after the delay — the store window.
+	Hold(after time.Duration) error
+
+	// Drop settles the message without acting on it: refused, stale, or given up on. It is not
+	// delivered again.
+	Drop() error
+}
+
+// Inbound is where control messages come from.
+type Inbound interface {
+	// Also asks for one more kind to be delivered.
+	//
+	// **Nothing is subscribed unless something is listening for it.** A durable queue or a
+	// durable stream consumer that nobody reads fills quietly, and the first symptom is a bus out
+	// of disk rather than anything about modules.
+	Also(kind string) error
+
+	// Receive delivers every message to act until the context ends, and says why it stopped.
+	Receive(ctx context.Context, act func(context.Context, Control)) error
+
+	// Close lets go of whatever the implementation holds.
+	Close()
+}
+
+// Outstanding answers which declaration the mesh last sent a node — the digest, not the
+// declaration.
+//
+// **Asked before the store is waited on** (design 25 §3): a report about a declaration the mesh
+// has already moved past is not worth holding a slot in the window that a current message needs.
+// It is a separate interface from Listener rather than a method on it, because a controller that
+// only publishes needs neither and something that records reports need not also be able to say
+// what was sent.
+type Outstanding interface {
+	Outstanding(ctx context.Context, node string) (string, error)
+}

internal/link/receive_current.go

@@ -0,0 +1,317 @@
+package link
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+// The consume side on the bus the mesh runs on today.
+//
+// Everything here was the serving loop's until the seam went in: the queues, the binds, the
+// prefetch, and the list of messages the store could not take yet. It moved rather than changed —
+// the behaviour this transport has is the behaviour it had, because the mesh is running on it and
+// a bus nothing speaks yet is no reason to alter the one every node is on (ADR 0116).
+
+// Prefetch is how many messages the bus hands the controller before it has settled them.
+//
+// More than one because a message the store could not take is held, unsettled, while the loop
+// goes on answering others — an enrolment above all, which a host is waiting on (novox/hq issue
+// 083). Bounded, because what is held is also what the bus has not kept on its own disk as
+// pending.
+const Prefetch = 64
+
+// PrefetchHeadroom is how much of the prefetch is never held, so the loop always has messages to
+// answer — an enrolment above all — while others wait for the store.
+const PrefetchHeadroom = 8
+
+// TryAgainAfter is how often the held are looked at. A store comes back in seconds, and a report a
+// few seconds late is still current.
+const TryAgainAfter = 2 * time.Second
+
+// currentInbound consumes what nodes say over the bus the mesh has.
+type currentInbound struct {
+	conn    *amqp.Connection
+	channel *amqp.Channel
+	// upgrades and catchups are bound only when something is listening (Also).
+	upgrades bool
+	catchups bool
+	// held is every message the store could not take, by the bus's own delivery tag. Kept here
+	// rather than in the serving loop because holding a delivery unacknowledged is this
+	// transport's way of keeping it, and the other's is to hand it back to the server.
+	held map[uint64]*holding
+	// again is how often the held are looked at; zero means TryAgainAfter. Set by tests.
+	again time.Duration
+}
+
+// holding is one message kept for the store, and when to try it again.
+type holding struct {
+	message *currentControl
+	due     time.Time
+	about   string
+}
+
+// Current is the consume side of the bus the mesh runs on today.
+func Current(conn *amqp.Connection, channel *amqp.Channel) Inbound {
+	return &currentInbound{conn: conn, channel: channel, held: map[uint64]*holding{}}
+}
+
+// Also binds the queue one more kind arrives on.
+//
+// The kinds nodes publish all share one queue and are bound at Connect, because a node may
+// publish any of them and binding one while forgetting another is a message the bus accepts, finds
+// no queue for, and drops — the publisher sees success and the consumer sees nothing. The two that
+// are events get their own queue each, and only when something is listening.
+func (c *currentInbound) Also(kind string) error {
+	switch kind {
+	case KindModuleMoved:
+		if err := c.bindEvent(UpgradeQueue, KeyModuleUpgraded); err != nil {
+			return err
+		}
+		c.upgrades = true
+	case KindCatchUp:
+		if err := c.bindEvent(CatchUpQueue, KeyCatchingUp); err != nil {
+			return err
+		}
+		c.catchups = true
+	default:
+		return fmt.Errorf("nothing binds a queue for %s on this bus", kind)
+	}
+	return nil
+}
+
+func (c *currentInbound) bindEvent(queue, key string) error {
+	if _, err := c.channel.QueueDeclare(queue, true, false, false, false, nil); err != nil {
+		return fmt.Errorf("cannot declare the %s queue: %w", queue, err)
+	}
+	if err := c.channel.QueueBind(queue, key, EventsExchange, false, nil); err != nil {
+		return fmt.Errorf("cannot bind %s to %s/%s: %w", queue, EventsExchange, key, err)
+	}
+	return nil
+}
+
+func (c *currentInbound) Close() {}
+
+// Receive consumes until the context ends.
+//
+// One consumer per queue, deliberately: with two on one queue the bus would round-robin between
+// them and each would receive half of what it expects — a fault this project has already had,
+// between a module's daemon and its capability server.
+func (c *currentInbound) Receive(ctx context.Context, act func(context.Context, Control)) error {
+	// A bounded prefetch rather than one. The loop still takes messages one at a time; what the
+	// prefetch buys is that a message the store could not take can be held while the loop goes on
+	// to the next, instead of every enrolment waiting behind it (novox/hq issue 083). Anything
+	// held goes back to the bus if the controller stops, because nothing held is acknowledged.
+	if err := c.channel.Qos(Prefetch, 0, false); err != nil {
+		return err
+	}
+
+	deliveries, err := c.channel.ConsumeWithContext(ctx, ControlQueue, "control-plane",
+		false, false, false, false, nil)
+	if err != nil {
+		return err
+	}
+
+	// Its own queue and its own consumer for each event, for the reason above: two consumers on
+	// one queue split its messages between them, and an upgrade or a catch-up request going to
+	// whichever half was not listening is a gap that looks like a working mesh.
+	var upgrades, catchups <-chan amqp.Delivery
+	if c.upgrades {
+		upgrades, err = c.channel.ConsumeWithContext(ctx, UpgradeQueue, "control-plane-upgrades",
+			false, false, false, false, nil)
+		if err != nil {
+			return err
+		}
+	}
+	if c.catchups {
+		catchups, err = c.channel.ConsumeWithContext(ctx, CatchUpQueue, "control-plane-catchup",
+			false, false, false, false, nil)
+		if err != nil {
+			return err
+		}
+	}
+
+	closed := c.conn.NotifyClose(make(chan *amqp.Error, 1))
+
+	again := c.again
+	if again == 0 {
+		again = TryAgainAfter
+	}
+	ticker := time.NewTicker(again)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case <-ticker.C:
+			if ctx.Err() != nil {
+				return nil
+			}
+			c.retryHeld(ctx, act)
+		case delivery, ok := <-catchups:
+			if !ok {
+				if catchups != nil {
+					return errors.New("the bus stopped delivering catch-up requests")
+				}
+				continue
+			}
+			act(ctx, c.wrap(KindCatchUp, delivery))
+		case delivery, ok := <-upgrades:
+			// A nil channel blocks for ever, so this case simply never fires when nothing is
+			// listening for upgrades. Closed is different, and means the bus stopped.
+			if !ok {
+				if upgrades != nil {
+					return errors.New("the bus stopped delivering upgrades")
+				}
+				continue
+			}
+			act(ctx, c.wrap(KindModuleMoved, delivery))
+		case reason := <-closed:
+			// Said rather than returned quietly. A controller whose bus connection dropped is a
+			// mesh where nothing can be told anything, and the reason is the first thing anybody
+			// will want.
+			return fmt.Errorf("the bus connection closed: %v", reason)
+		case delivery, ok := <-deliveries:
+			if !ok {
+				return errors.New("the bus stopped delivering")
+			}
+			kind, known := kindOfKey[delivery.RoutingKey]
+			if !known {
+				// Rejected without requeue: a message nothing understands will not be understood
+				// on the next attempt either, and requeuing it would spin.
+				_ = delivery.Reject(false)
+				continue
+			}
+			act(ctx, c.wrap(kind, delivery))
+		}
+	}
+}
+
+// kindOfKey is how this transport's addressing becomes what the mesh calls a message.
+var kindOfKey = map[string]string{
+	KeyEnrol:          KindEnrolment,
+	KeyReport:         KindReport,
+	KeyAlive:          KindHeartbeat,
+	KeyBuilt:          KindBuilt,
+	KeyModuleUpgraded: KindModuleMoved,
+	KeyCatchingUp:     KindCatchUp,
+}
+
+func (c *currentInbound) wrap(kind string, delivery amqp.Delivery) *currentControl {
+	return &currentControl{kind: kind, delivery: delivery, on: c}
+}
+
+// retryHeld hands every message whose delay has passed back to the loop. Each handler holds it
+// again, settles it, or lets it go past the bound.
+func (c *currentInbound) retryHeld(ctx context.Context, act func(context.Context, Control)) {
+	now := time.Now()
+	due := make([]*currentControl, 0, len(c.held))
+	for _, h := range c.held {
+		if !h.due.After(now) {
+			due = append(due, h.message)
+		}
+	}
+	for _, m := range due {
+		if ctx.Err() != nil {
+			return
+		}
+		act(ctx, m)
+	}
+}
+
+// currentControl is one delivery from the bus the mesh has, as the controller reads it.
+type currentControl struct {
+	kind     string
+	delivery amqp.Delivery
+	on       *currentInbound
+	// about is what this message is about, as the handler named it; empty until it does.
+	about string
+	// first is when this message was first held for the store; zero while it has not been.
+	first time.Time
+}
+
+func (m *currentControl) Kind() string      { return m.kind }
+func (m *currentControl) Body() []byte      { return m.delivery.Body }
+func (m *currentControl) Redelivered() bool { return m.delivery.Redelivered }
+
+func (m *currentControl) HeldFor() time.Duration {
+	if m.first.IsZero() {
+		return 0
+	}
+	return time.Since(m.first)
+}
+
+// Answer publishes to the reply queue the request named.
+func (m *currentControl) Answer(ctx context.Context, body []byte) error {
+	if m.delivery.ReplyTo == "" {
+		return errors.New("that request named no reply queue, so nothing can be told the answer")
+	}
+	return m.on.channel.PublishWithContext(ctx, "", m.delivery.ReplyTo, false, false,
+		amqp.Publishing{
+			ContentType:   "application/json",
+			CorrelationId: m.delivery.CorrelationId,
+			Body:          body,
+		})
+}
+
+func (m *currentControl) Took() error {
+	m.forget()
+	return m.delivery.Ack(false)
+}
+
+// Drop rejects without requeue: on this bus that is what "understood, and not worth another
+// attempt" is spelled as, and it is what feeds a dead-letter queue where one is configured.
+func (m *currentControl) Drop() error {
+	m.forget()
+	return m.delivery.Reject(false)
+}
+
+// About names what this message is about, and lets go of whatever is held about the same thing:
+// the held one is the past, and acting on it after this one would undo this one. Acknowledged
+// rather than left to come back, because a held message nothing will act on is a place in the
+// prefetch nothing gets back.
+func (m *currentControl) About(what string) {
+	m.about = what
+	if what == "" {
+		return
+	}
+	for tag, h := range m.on.held {
+		if h.about != what || tag == m.delivery.DeliveryTag {
+			continue
+		}
+		delete(m.on.held, tag)
+		_ = h.message.delivery.Ack(false)
+	}
+}
+
+// Hold keeps the message unacknowledged and sets it aside to be handed back after the delay.
+//
+// Held no further than the prefetch leaves room: past that the bus would hand the loop nothing
+// new — enrolments included — until something held was let go. A message that cannot be held says
+// so, and the handler settles it its own way.
+func (m *currentControl) Hold(after time.Duration) error {
+	if m.on.held == nil {
+		m.on.held = map[uint64]*holding{}
+	}
+	if _, already := m.on.held[m.delivery.DeliveryTag]; !already {
+		if len(m.on.held) >= Prefetch-PrefetchHeadroom {
+			return fmt.Errorf("%d messages are already held for the store, and holding more "+
+				"would stop the queue", len(m.on.held))
+		}
+		m.first = time.Now()
+	}
+	m.on.held[m.delivery.DeliveryTag] = &holding{
+		message: m, due: time.Now().Add(after), about: m.about,
+	}
+	return nil
+}
+
+func (m *currentControl) forget() {
+	if m.on != nil {
+		delete(m.on.held, m.delivery.DeliveryTag)
+	}
+}

internal/link/receive_current_test.go

@@ -0,0 +1,71 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"io"
+	"log"
+	"testing"
+	"time"
+
+	amqp "github.com/rabbitmq/amqp091-go"
+)
+
+// The harness for the consume side on the bus the mesh runs on today.
+//
+// Messages arrive through the seam, so what these tests exercise is the controller's decision
+// about a message and this transport's way of keeping one — which is what the seam separated. A
+// fake acknowledger stands in for the bus, because what is asserted is how a message was settled
+// and that needs no server.
+
+// settled is how the bus was told to settle one message.
+type settled struct{ acked, nacked, requeued, rejected bool }
+
+func (a *settled) Ack(uint64, bool) error { a.acked = true; return nil }
+func (a *settled) Nack(_ uint64, _ bool, requeue bool) error {
+	a.nacked, a.requeued = true, requeue
+	return nil
+}
+func (a *settled) Reject(uint64, bool) error { a.rejected = true; return nil }
+
+// unsettled is a message the controller has neither taken nor let go: it is held, and the bus will
+// hand it to whatever consumes next if the controller stops.
+func (a *settled) unsettled() bool { return !a.acked && !a.nacked && !a.rejected }
+
+var tag uint64
+
+func quiet() *log.Logger { return log.New(io.Discard, "", 0) }
+
+// serving is a controller with nothing but a way of receiving, ready for a listener, a recorder,
+// an upgrader or a replayer to be set on it.
+func serving() (*Server, *currentInbound) {
+	in := &currentInbound{held: map[uint64]*holding{}}
+	return &Server{inbound: in, bus: OverCurrent{}, log: quiet()}, in
+}
+
+// sends is one message arriving over this transport, as the controller reads it.
+func (c *currentInbound) sends(t *testing.T, to *settled, kind string, v any) Control {
+	t.Helper()
+	body, err := json.Marshal(v)
+	if err != nil {
+		t.Fatal(err)
+	}
+	tag++
+	return &currentControl{kind: kind, on: c, delivery: amqp.Delivery{
+		Acknowledger: to, Body: body, DeliveryTag: tag,
+	}}
+}
+
+// dueNow brings every held message forward, so a test need not wait out the backoff a real store
+// restart would be given (RedeliverAfter).
+func (c *currentInbound) dueNow() {
+	for _, h := range c.held {
+		h.due = time.Now().Add(-time.Second)
+	}
+}
+
+// retries hands every held message back to the controller, the way the ticker does.
+func (c *currentInbound) retries(ctx context.Context, s *Server) {
+	c.dueNow()
+	c.retryHeld(ctx, s.act)
+}

internal/link/receive_nats.go

@@ -0,0 +1,290 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// The consume side on the bus being built.
+//
+// The shape is the AMQP one's, because the seam made them comparable: one loop, one message at a
+// time, and the same window deciding. What differs is where a held message lives — and that is the
+// whole point of the move. On the bus the mesh has, holding one means keeping an unacknowledged
+// delivery in this process, bounded by the prefetch and lost if the controller stops. Here it is a
+// `nak` with a delay: the message stays the server's, the controller keeps nothing but the moment
+// it first could not take it, and a controller that restarts mid-window has nothing to lose.
+
+// natsInbound consumes what nodes and modules say over NATS.
+type natsInbound struct {
+	js *broker.JetStream
+	// follows is the kinds asked for beyond what nodes say (Also). The events those are are the
+	// only ones the controller subscribes, and only when something is listening.
+	follows map[string]bool
+	// since is when the controller first could not take a message, by that message's place in its
+	// stream.
+	//
+	// **A timestamp, not a message.** This is the whole difference the move buys: the AMQP side
+	// keeps the delivery, and this keeps eight bytes saying when the window opened. A controller
+	// that restarts loses these and starts the window again, which is correct — it is holding
+	// nothing, and the messages are all still on the server.
+	since map[uint64]time.Time
+}
+
+// Nats is the consume side of the bus being built.
+func Nats(js *broker.JetStream) Inbound {
+	return &natsInbound{js: js, follows: map[string]bool{}, since: map[uint64]time.Time{}}
+}
+
+// Also records one more kind to subscribe. Nothing is subscribed here: the controller's consumer on
+// the events stream carries both of these as filters, so it is created once, in Receive, with
+// whatever was asked for — and not at all when nothing was.
+func (n *natsInbound) Also(kind string) error {
+	switch kind {
+	case KindModuleMoved, KindCatchUp:
+		n.follows[kind] = true
+		return nil
+	default:
+		return fmt.Errorf("nothing subscribes %s separately on this bus", kind)
+	}
+}
+
+func (n *natsInbound) Close() {}
+
+// Receive consumes until the context ends.
+//
+// Three subscriptions, and each is a channel the one loop selects on. **Channels rather than
+// callbacks**: the library would run a handler on its own goroutine, and the window's bookkeeping —
+// which message is held, and since when — is read and written without a lock because the AMQP loop
+// never had two. A second goroutine would make that wrong in a way no test would catch.
+func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Control)) error {
+	if err := broker.AssertMeshConsumers(n.js); err != nil {
+		return err
+	}
+	js, conn := n.js.Context(), n.js.Conn()
+
+	// What nodes say, off the CONTROL stream. Bound to the durable the controller asserted rather
+	// than creating one here: the consumer is an object with a configuration — ack policy, ack
+	// wait, redelivery — and a client that creates its own would be a second opinion about it.
+	control := make(chan *nats.Msg, Prefetch)
+	said, err := js.ChanSubscribe("", control, nats.Bind("CONTROL", broker.ControllerName))
+	if err != nil {
+		return fmt.Errorf("subscribing to what nodes say: %w", err)
+	}
+	defer func() { _ = said.Unsubscribe() }()
+
+	// Heartbeats, on core NATS and off any stream (design 25 §3). Their own subscription because
+	// they are their own guarantee: a lost one is the next one.
+	beats := make(chan *nats.Msg, Prefetch)
+	alive, err := conn.ChanSubscribe(AliveSubjects, beats)
+	if err != nil {
+		return fmt.Errorf("subscribing to heartbeats: %w", err)
+	}
+	defer func() { _ = alive.Unsubscribe() }()
+
+	// The events the controller follows, when something is listening for them.
+	var events chan *nats.Msg
+	if len(n.follows) > 0 {
+		events = make(chan *nats.Msg, Prefetch)
+		followed, err := js.ChanSubscribe("", events, nats.Bind("EVENTS", broker.ControllerName))
+		if err != nil {
+			return fmt.Errorf("subscribing to what the catalogue says: %w", err)
+		}
+		defer func() { _ = followed.Unsubscribe() }()
+	}
+
+	// A connection that dropped is said, not discovered. A controller whose bus connection is gone
+	// is a mesh where nothing can be told anything.
+	gone := make(chan error, 1)
+	conn.SetDisconnectErrHandler(func(_ *nats.Conn, err error) {
+		select {
+		case gone <- err:
+		default:
+		}
+	})
+
+	for {
+		select {
+		case <-ctx.Done():
+			return nil
+		case err := <-gone:
+			return fmt.Errorf("the bus connection dropped: %w", err)
+		case msg := <-beats:
+			n.deliver(ctx, act, msg, false)
+		case msg := <-events:
+			n.deliver(ctx, act, msg, true)
+		case msg, ok := <-control:
+			if !ok {
+				return errors.New("the bus stopped delivering")
+			}
+			n.deliver(ctx, act, msg, true)
+		}
+	}
+}
+
+// deliver names one message and hands it to the loop, or drops it where the mesh has no name for
+// its subject — which cannot happen through a filter the controller wrote, and is said rather than
+// ignored for exactly that reason.
+func (n *natsInbound) deliver(ctx context.Context, act func(context.Context, Control),
+	msg *nats.Msg, streamed bool) {
+
+	kind, known := kindOfSubject(msg.Subject)
+	if !known {
+		if streamed {
+			_ = msg.Term()
+		}
+		return
+	}
+	m := &natsControl{kind: kind, msg: msg, on: n}
+	if streamed {
+		// A message with no metadata is not from a stream, whatever it was delivered on, and the
+		// window has nothing to hold it by. Said by leaving the sequence at zero.
+		if meta, err := msg.Metadata(); err == nil {
+			m.seq = meta.Sequence.Stream
+			m.delivered = meta.NumDelivered
+		}
+	}
+	act(ctx, m)
+}
+
+// kindOfSubject is how this transport's addressing becomes what the mesh calls a message.
+//
+// By subject, which is the only thing the server enforces: a body claiming to be a report does not
+// make it one, and on this bus the subject an account may publish *is* its authority (design 29
+// §2). The mirror of the routing-key table on the bus the mesh has.
+func kindOfSubject(subject string) (string, bool) {
+	switch subject {
+	case EnrolSubject:
+		return KindEnrolment, true
+	case BuiltSubject:
+		return KindBuilt, true
+	}
+	if node, rest, ok := strings.Cut(strings.TrimPrefix(subject, "mesh.control."), "."); ok &&
+		node != "" && !strings.Contains(node, ".") {
+		switch rest {
+		case "report":
+			return KindReport, true
+		case "alive":
+			return KindHeartbeat, true
+		}
+	}
+	switch subject {
+	case broker.ControllerFollows[0]:
+		return KindModuleMoved, true
+	case broker.ControllerFollows[1]:
+		return KindCatchUp, true
+	case BuildOutcome():
+		// A build's outcome is the role's event now, so it arrives on the events stream rather than
+		// the control branch — and is acted on by the same handler, because what the controller does
+		// with it did not change (novox/hq ADR 0121).
+		return KindBuilt, true
+	}
+	return "", false
+}
+
+// natsControl is one message from the bus being built, as the controller reads it.
+type natsControl struct {
+	kind string
+	msg  *nats.Msg
+	on   *natsInbound
+	// seq is this message's place in its stream; zero for a core message, which has none and
+	// cannot be held.
+	seq uint64
+	// delivered is how many times the server has handed this message over, this time included.
+	delivered uint64
+}
+
+func (m *natsControl) Kind() string { return m.kind }
+func (m *natsControl) Body() []byte { return m.msg.Data }
+
+// Redelivered is what the server counted, not what the controller remembers. Which is the answer to
+// a question the AMQP side could only guess at across a restart: an enrolment redelivered because
+// the controller stopped mid-answer reads as redelivered to the controller that comes back.
+func (m *natsControl) Redelivered() bool { return m.delivered > 1 }
+
+func (m *natsControl) HeldFor() time.Duration {
+	if m.seq == 0 {
+		return 0
+	}
+	first, held := m.on.since[m.seq]
+	if !held {
+		return 0
+	}
+	return time.Since(first)
+}
+
+// About is nothing here, and that is the point.
+//
+// Setting a held message aside when a newer one about the same thing arrives is what a controller
+// holding deliveries in memory can do. A naked message belongs to the server and comes back
+// whatever happened meanwhile, so the question "is this the past?" is answered by what the message
+// says instead — the digest of the declaration a report is about (window.go, design 25 §3).
+func (m *natsControl) About(string) {}
+
+// Answer publishes to the reply subject the request carries **in its payload**.
+//
+// Not `Respond`, and not the message's reply field: a message a JetStream consumer delivers has had
+// that field claimed for the consumer's own ack address, so answering it would send the reply to
+// `$JS.ACK.CONTROL.controller.…` and the enrolling node would wait out its timeout. Verified
+// against a running server (design 25 §2), which is why it is a field of the request and this reads
+// it from there.
+func (m *natsControl) Answer(ctx context.Context, body []byte) error {
+	var addressed replyAddressed
+	if err := json.Unmarshal(m.msg.Data, &addressed); err != nil {
+		return fmt.Errorf("that request cannot be read, so its reply address cannot be: %w", err)
+	}
+	if addressed.ReplyTo == "" {
+		return errors.New("that request named no reply subject in its payload, so nothing can be " +
+			"told the answer")
+	}
+	return m.on.js.Conn().PublishMsg(&nats.Msg{Subject: addressed.ReplyTo, Data: body})
+}
+
+func (m *natsControl) Took() error {
+	m.forget()
+	if m.seq == 0 {
+		// Core NATS: nothing is keeping it, so there is nothing to settle.
+		return nil
+	}
+	return m.msg.Ack(nats.Context(context.Background()))
+}
+
+// Drop terminates the delivery: understood, and the server is told not to send it again. Different
+// from an ack only in the server's own accounting, which is where somebody asking "what happened to
+// that message" will look.
+func (m *natsControl) Drop() error {
+	m.forget()
+	if m.seq == 0 {
+		return nil
+	}
+	return m.msg.Term()
+}
+
+// Hold hands the message back with a delay, and remembers when the window opened.
+func (m *natsControl) Hold(after time.Duration) error {
+	if m.seq == 0 {
+		return errors.New("a message that is not in a stream cannot be held: nothing is keeping it")
+	}
+	if _, already := m.on.since[m.seq]; !already {
+		m.on.since[m.seq] = time.Now()
+	}
+	return m.msg.NakWithDelay(after)
+}
+
+func (m *natsControl) forget() {
+	if m.on != nil && m.seq != 0 {
+		delete(m.on.since, m.seq)
+	}
+}
+
+// replyAddressed is the one field every message that expects an answer carries.
+type replyAddressed struct {
+	ReplyTo string `json:"reply_to,omitempty"`
+}

internal/link/receive_nats_test.go

@@ -0,0 +1,376 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"os"
+	"sync"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// The consume side against a real server, because what is being checked is what the server does.
+//
+// Reasoning cannot answer any of these: whether a nak-with-delay really comes back, whether the
+// delay is honoured, whether terminating a delivery really stops it, or whether an answer published
+// to an address carried in the payload reaches a caller waiting on its own inbox. Each is a claim
+// about a server, so each is asked of one:
+//
+//	docker run -d --rm --name t -p 14222:4222 nats:2.10-alpine -js
+//	MESH_TEST_NATS=nats://127.0.0.1:14222 go test ./internal/link/ -run TestNats
+
+func aBus(t *testing.T) *broker.JetStream {
+	t.Helper()
+	url := os.Getenv("MESH_TEST_NATS")
+	if url == "" {
+		t.Skip("MESH_TEST_NATS unset")
+	}
+	js, err := broker.Dial(url)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+
+	// **Streams purged, consumers removed.** Both halves, and each was learned by getting it wrong.
+	//
+	// The streams are emptied rather than deleted and recreated, because delete-then-add is not a
+	// reset: the server's teardown races the creation, and a test then inherits the previous one's
+	// messages — which reads as a redelivery bug in the code under test.
+	//
+	// The consumers are removed, because deleting a stream used to take them with it and purging
+	// does not. A durable *push* consumer that survives between tests keeps pushing to a delivery
+	// subject the previous test's subscription has gone from: the messages count as delivered, go
+	// nowhere, and the next test waits out its timeout for an announcement the server believes it
+	// already sent. The controller recreates what it needs on start, so leaving none is correct.
+	if err := broker.AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	clean := func() {
+		for _, c := range broker.MeshConsumers() {
+			_ = js.Context().DeleteConsumer(c.Stream, c.Name)
+		}
+		for _, s := range broker.MeshStreams() {
+			_ = js.Context().PurgeStream(s.Name)
+		}
+	}
+	clean()
+	t.Cleanup(clean)
+	return js
+}
+
+// serving1 is a controller reading from a real bus, and a way to stop it.
+func servingOn(t *testing.T, js *broker.JetStream, l Listener) (*Server, func()) {
+	t.Helper()
+	s := &Server{inbound: Nats(js), bus: OverNATS{Conn: js.Conn(), JS: js.Context()},
+		listener: l, log: quiet()}
+	ctx, stop := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { defer close(done); _ = s.Serve(ctx) }()
+	return s, func() {
+		stop()
+		<-done
+	}
+}
+
+// counted records reports and can be told to refuse them, from another goroutine.
+type counted struct {
+	mu    sync.Mutex
+	err   error
+	heard []Report
+}
+
+func (c *counted) Heard(_ context.Context, r Report) error {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	if c.err != nil {
+		return c.err
+	}
+	c.heard = append(c.heard, r)
+	return nil
+}
+
+func (c *counted) refusing(err error) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.err = err
+}
+
+func (c *counted) count() int {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return len(c.heard)
+}
+
+func eventually(t *testing.T, what string, is func() bool) {
+	t.Helper()
+	deadline := time.Now().Add(8 * time.Second)
+	for time.Now().Before(deadline) {
+		if is() {
+			return
+		}
+		time.Sleep(20 * time.Millisecond)
+	}
+	t.Fatalf("%s did not happen within the wait", what)
+}
+
+// A report published by a node reaches the controller, is recorded, and is acknowledged — so the
+// stream does not hold it. A work queue is the check: what is acknowledged leaves it.
+func TestNatsAReportIsHeardAndLeavesTheStream(t *testing.T) {
+	js := aBus(t)
+	store := &counted{}
+	_, stop := servingOn(t, js, store)
+	defer stop()
+
+	body, _ := json.Marshal(Report{Node: "anchor", Declared: "d1", Applied: []string{"store"}})
+	if _, err := js.Context().Publish(ReportSubject("anchor"), body); err != nil {
+		t.Fatal(err)
+	}
+	eventually(t, "a report being recorded", func() bool { return store.count() == 1 })
+	eventually(t, "the report leaving the work queue", func() bool {
+		info, err := js.Context().StreamInfo("CONTROL")
+		return err == nil && info.State.Msgs == 0
+	})
+}
+
+// **The store window, in the server.** A report the store cannot take is naked with a delay and
+// comes back; once the store is there it is recorded and leaves the stream. The controller holds
+// nothing in the meantime — which is what the sequence check below is for: the message is still on
+// the server while it waits.
+func TestNatsAReportTheStoreCannotTakeIsHeldByTheServerAndComesBack(t *testing.T) {
+	js := aBus(t)
+	store := &counted{}
+	store.refusing(errors.Join(ErrTryAgain, errors.New("the database system is starting up")))
+	_, stop := servingOn(t, js, store)
+	defer stop()
+
+	body, _ := json.Marshal(Report{Node: "anchor", Declared: "d1", Applied: []string{"store"}})
+	if _, err := js.Context().Publish(ReportSubject("anchor"), body); err != nil {
+		t.Fatal(err)
+	}
+
+	// Held: the message is the server's, unacknowledged, and still in the stream.
+	eventually(t, "the report being redelivered at least once", func() bool {
+		info, err := js.Context().ConsumerInfo("CONTROL", broker.ControllerName)
+		return err == nil && info.NumRedelivered >= 1
+	})
+	info, err := js.Context().StreamInfo("CONTROL")
+	if err != nil || info.State.Msgs != 1 {
+		t.Fatalf("a held report did not stay on the server: %+v, %v", info, err)
+	}
+	if store.count() != 0 {
+		t.Fatalf("a report was recorded by a store that was refusing it")
+	}
+
+	store.refusing(nil)
+	eventually(t, "the report being recorded once the store was back",
+		func() bool { return store.count() == 1 })
+	eventually(t, "the recorded report leaving the work queue", func() bool {
+		info, err := js.Context().StreamInfo("CONTROL")
+		return err == nil && info.State.Msgs == 0
+	})
+}
+
+// A report about a declaration the mesh has moved past is settled without being acted on, and
+// leaves the stream rather than coming back for ever.
+func TestNatsASupersededReportIsSettledAndNotActedOn(t *testing.T) {
+	js := aBus(t)
+	store := &sentAndHeardSafely{sent: "d2"}
+	_, stop := servingOn(t, js, store)
+	defer stop()
+
+	body, _ := json.Marshal(Report{Node: "anchor", Declared: "d1", Applied: []string{"store"}})
+	if _, err := js.Context().Publish(ReportSubject("anchor"), body); err != nil {
+		t.Fatal(err)
+	}
+	eventually(t, "the superseded report leaving the stream", func() bool {
+		info, err := js.Context().StreamInfo("CONTROL")
+		return err == nil && info.State.Msgs == 0
+	})
+	if store.count() != 0 {
+		t.Fatalf("a report about a superseded declaration was acted on")
+	}
+}
+
+// sentAndHeardSafely is sentAndHeard, read from two goroutines.
+type sentAndHeardSafely struct {
+	mu    sync.Mutex
+	sent  string
+	heard []Report
+}
+
+func (s *sentAndHeardSafely) Heard(_ context.Context, r Report) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.heard = append(s.heard, r)
+	return nil
+}
+
+func (s *sentAndHeardSafely) Outstanding(context.Context, string) (string, error) {
+	return s.sent, nil
+}
+
+func (s *sentAndHeardSafely) count() int {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return len(s.heard)
+}
+
+// **An enrolment answered through a reply address the stream would have eaten.**
+//
+// The caller waits on its own inbox and states that address in the request's payload. The check is
+// that the answer arrives there — which is the whole reason the address is a field rather than the
+// transport's reply, and this is the test design 25 §2 asks for so the reason cannot quietly become
+// folklore.
+func TestNatsAnEnrolmentIsAnsweredOnTheAddressInItsPayload(t *testing.T) {
+	js := aBus(t)
+	s := &Server{inbound: Nats(js), bus: OverNATS{Conn: js.Conn(), JS: js.Context()},
+		enroller: enrolsAs{reply: EnrolReply{Accepted: true, Node: "anchor"}}, log: quiet()}
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	go func() { _ = s.Serve(ctx) }()
+
+	inbox := nats.NewInbox()
+	answers, err := js.Conn().SubscribeSync(inbox)
+	if err != nil {
+		t.Fatal(err)
+	}
+	body, _ := json.Marshal(EnrolRequest{Node: "anchor", Secret: "t", ReplyTo: inbox})
+	if _, err := js.Context().Publish(EnrolSubject, body); err != nil {
+		t.Fatal(err)
+	}
+
+	msg, err := answers.NextMsg(8 * time.Second)
+	if err != nil {
+		t.Fatalf("no answer reached the address the request named: %v", err)
+	}
+	var reply EnrolReply
+	if err := json.Unmarshal(msg.Data, &reply); err != nil {
+		t.Fatal(err)
+	}
+	if !reply.Accepted || reply.Node != "anchor" {
+		t.Fatalf("the answer was not the mesh's: %+v", reply)
+	}
+	// And the address really is not the one the transport carried: what the consumer saw was its
+	// own ack subject, which is why this had to travel in the payload.
+	if msg.Subject != inbox {
+		t.Fatalf("the answer arrived on %s, not the address the request named", msg.Subject)
+	}
+}
+
+// enrolsAs answers every request the same way.
+type enrolsAs struct{ reply EnrolReply }
+
+func (e enrolsAs) Enrol(context.Context, EnrolRequest) (EnrolReply, error) { return e.reply, nil }
+
+// A heartbeat is core NATS: it reaches the controller and nothing is persisted, so the stream the
+// reports live in stays empty.
+func TestNatsAHeartbeatIsHeardAndNothingIsKept(t *testing.T) {
+	js := aBus(t)
+	store := &counted{}
+	_, stop := servingOn(t, js, store)
+	defer stop()
+
+	// Given time to subscribe: a core subscription that is not yet up misses what is published,
+	// which is the guarantee a heartbeat has and not a fault.
+	eventually(t, "the heartbeat subscription coming up", func() bool {
+		body, _ := json.Marshal(Alive{Node: "anchor"})
+		_ = js.Conn().Publish(AliveSubject("anchor"), body)
+		_ = js.Conn().Flush()
+		return store.count() >= 1
+	})
+	info, err := js.Context().StreamInfo("CONTROL")
+	if err != nil || info.State.Msgs != 0 {
+		t.Fatalf("a heartbeat was persisted, and the mesh's least valuable message now competes "+
+			"for retention with its most valuable: %+v, %v", info, err)
+	}
+}
+
+// The two events the controller follows arrive over one durable consumer with two filters, and it
+// can acknowledge them.
+//
+// **Both halves are the point.** A consumer with several filter subjects is a 2.10 feature and this
+// is the first thing in the mesh to use one; and a delivery from the events stream is acknowledged
+// on a different ack subject from a delivery from the control stream, which the controller's own
+// permission list has to cover or every announcement is redelivered for ever.
+func TestNatsTheEventsTheControllerFollowsArriveAndAreAcknowledged(t *testing.T) {
+	js := aBus(t)
+	told := &toldAbout{}
+	s := &Server{inbound: Nats(js), bus: OverNATS{Conn: js.Conn(), JS: js.Context()}, log: quiet()}
+	if err := s.Follows(told); err != nil {
+		t.Fatal(err)
+	}
+	if err := s.Answers(replaysWith{}); err != nil {
+		t.Fatal(err)
+	}
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	go func() { _ = s.Serve(ctx) }()
+
+	moved, _ := json.Marshal(Upgraded{Module: "gitea", Commit: "abcdef0123"})
+	if _, err := js.Context().Publish(broker.ControllerFollows[0], moved); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := js.Context().Publish(broker.ControllerFollows[1], []byte(`{}`)); err != nil {
+		t.Fatal(err)
+	}
+
+	eventually(t, "the catalogue's upgrade reaching the controller",
+		func() bool { return told.count() == 1 })
+	eventually(t, "both announcements being acknowledged", func() bool {
+		info, err := js.Context().ConsumerInfo("EVENTS", broker.ControllerName)
+		return err == nil && info.NumAckPending == 0 && info.Delivered.Consumer == 2
+	})
+}
+
+type toldAbout struct {
+	mu   sync.Mutex
+	saw  []Upgraded
+	fail error
+}
+
+func (u *toldAbout) Upgraded(_ context.Context, m Upgraded) error {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	if u.fail != nil {
+		return u.fail
+	}
+	u.saw = append(u.saw, m)
+	return nil
+}
+
+func (u *toldAbout) count() int {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	return len(u.saw)
+}
+
+// A store that never comes back: the report is let go once the bound passes, and it leaves the
+// stream rather than being held for ever. The bound is the controller's, not the server's — nothing
+// here sets max-deliver, and that is deliberate (streams.go).
+func TestNatsAReportIsLetGoOnceTheStoreHasBeenGoneTooLong(t *testing.T) {
+	js := aBus(t)
+	store := &counted{}
+	store.refusing(errors.Join(ErrTryAgain, errors.New("connection refused")))
+	s := &Server{inbound: Nats(js), bus: OverNATS{Conn: js.Conn(), JS: js.Context()},
+		listener: store, log: quiet(), giveUp: 1500 * time.Millisecond}
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	go func() { _ = s.Serve(ctx) }()
+
+	body, _ := json.Marshal(Report{Node: "anchor", Declared: "d1", Applied: []string{"store"}})
+	if _, err := js.Context().Publish(ReportSubject("anchor"), body); err != nil {
+		t.Fatal(err)
+	}
+	eventually(t, "the report being let go once the bound passed", func() bool {
+		info, err := js.Context().StreamInfo("CONTROL")
+		return err == nil && info.State.Msgs == 0
+	})
+	if store.count() != 0 {
+		t.Fatalf("a report was recorded by a store that never came back")
+	}
+}

internal/link/report_retry_test.go

@@ -2,26 +2,11 @@ package link
 
 import (
 	"context"
-	"encoding/json"
 	"errors"
-	"io"
-	"log"
 	"testing"
 	"time"
-
-	amqp "github.com/rabbitmq/amqp091-go"
 )
 
-type saidTo struct{ acked, nacked, requeued bool }
-
-func (a *saidTo) Ack(uint64, bool) error { a.acked = true; return nil }
-func (a *saidTo) Nack(_ uint64, _ bool, requeue bool) error {
-	a.nacked, a.requeued = true, requeue
-	return nil
-}
-func (a *saidTo) Reject(uint64, bool) error { return nil }
-func (a *saidTo) unsettled() bool           { return !a.acked && !a.nacked }
-
 type heardWith struct{ err error }
 
 func (h heardWith) Heard(context.Context, Report) error { return h.err }
@@ -31,39 +16,31 @@ type switchable struct{ err error }
 
 func (h *switchable) Heard(context.Context, Report) error { return h.err }
 
-var tag uint64
-
-func aReport(t *testing.T, to *saidTo, node, declared string) amqp.Delivery {
-	t.Helper()
-	body, err := json.Marshal(Report{Node: node, Declared: declared, Applied: []string{"store"}})
-	if err != nil {
-		t.Fatal(err)
-	}
-	tag++
-	return amqp.Delivery{Acknowledger: to, RoutingKey: KeyReport, Body: body, DeliveryTag: tag}
+func aReport(node, declared string) Report {
+	return Report{Node: node, Declared: declared, Applied: []string{"store"}}
 }
 
-func quiet() *log.Logger { return log.New(io.Discard, "", 0) }
-
 // A report the store could not take right now is held, unsettled, and recorded when the store is
 // back; one the store answered no to is acknowledged; one recorded is acknowledged (issue 082, 083).
 func TestAReportTheStoreCouldNotTakeIsHeldAndOneItRefusedIsNot(t *testing.T) {
 	store := &switchable{err: errors.Join(ErrTryAgain, errors.New("starting up"))}
-	s := &Server{listener: store, log: quiet()}
-	held := &saidTo{}
-	s.handleReport(context.Background(), aReport(t, held, "anchor", "d1"))
-	if !held.unsettled() || len(s.parked) != 1 {
-		t.Fatalf("a report the store could not take was not held: %+v, %d held", held, len(s.parked))
+	s, in := serving()
+	s.listener = store
+	held := &settled{}
+	s.act(context.Background(), in.sends(t, held, KindReport, aReport("anchor", "d1")))
+	if !held.unsettled() || len(in.held) != 1 {
+		t.Fatalf("a report the store could not take was not held: %+v, %d held", held, len(in.held))
 	}
 	store.err = nil
-	s.retryHeld(context.Background())
-	if !held.acked || len(s.parked) != 0 {
-		t.Fatalf("a held report was not recorded once the store was back: %+v, %d held", held, len(s.parked))
+	in.retries(context.Background(), s)
+	if !held.acked || len(in.held) != 0 {
+		t.Fatalf("a held report was not recorded once the store was back: %+v, %d held", held, len(in.held))
 	}
 
-	refused := &saidTo{}
-	s = &Server{listener: heardWith{err: errors.New("a report named no node")}, log: quiet()}
-	s.handleReport(context.Background(), aReport(t, refused, "anchor", "d1"))
+	refused := &settled{}
+	s, in = serving()
+	s.listener = heardWith{err: errors.New("a report named no node")}
+	s.act(context.Background(), in.sends(t, refused, KindReport, aReport("anchor", "d1")))
 	if !refused.acked || refused.nacked {
 		t.Fatalf("a report the store answered no to was not acknowledged: %+v", refused)
 	}
@@ -72,33 +49,35 @@ func TestAReportTheStoreCouldNotTakeIsHeldAndOneItRefusedIsNot(t *testing.T) {
 // A newer report from the same node supersedes one of its reports still held: recorded after the
 // newer, the older would overwrite what the node is doing now.
 func TestANewerReportSupersedesAHeldOneFromTheSameNode(t *testing.T) {
-	s := &Server{listener: heardWith{err: errors.Join(ErrTryAgain, errors.New("starting up"))}, log: quiet()}
-	older, newer, other := &saidTo{}, &saidTo{}, &saidTo{}
-	s.handleReport(context.Background(), aReport(t, older, "anchor", "d1"))
-	s.handleReport(context.Background(), aReport(t, other, "laptop", "d7"))
-	s.handleReport(context.Background(), aReport(t, newer, "anchor", "d2"))
+	s, in := serving()
+	s.listener = heardWith{err: errors.Join(ErrTryAgain, errors.New("starting up"))}
+	older, newer, other := &settled{}, &settled{}, &settled{}
+	s.act(context.Background(), in.sends(t, older, KindReport, aReport("anchor", "d1")))
+	s.act(context.Background(), in.sends(t, other, KindReport, aReport("laptop", "d7")))
+	s.act(context.Background(), in.sends(t, newer, KindReport, aReport("anchor", "d2")))
 	if !older.acked {
 		t.Fatalf("the older report was not set aside by the newer: %+v", older)
 	}
-	if !newer.unsettled() || !other.unsettled() || len(s.parked) != 2 {
+	if !newer.unsettled() || !other.unsettled() || len(in.held) != 2 {
 		t.Fatalf("the newer report and another node's were not both held: newer %+v other %+v, %d held",
-			newer, other, len(s.parked))
+			newer, other, len(in.held))
 	}
 }
 
 // A store that has not come back within the bound is not restarting: the report is let go, loudly,
 // rather than held for ever.
 func TestAReportIsLetGoOnceTheStoreHasBeenGoneTooLong(t *testing.T) {
-	s := &Server{listener: heardWith{err: errors.Join(ErrTryAgain, errors.New("connection refused"))},
-		log: quiet(), giveUp: time.Millisecond}
-	held := &saidTo{}
-	s.handleReport(context.Background(), aReport(t, held, "anchor", "d1"))
+	s, in := serving()
+	s.listener = heardWith{err: errors.Join(ErrTryAgain, errors.New("connection refused"))}
+	s.giveUp = time.Millisecond
+	held := &settled{}
+	s.act(context.Background(), in.sends(t, held, KindReport, aReport("anchor", "d1")))
 	if !held.unsettled() {
 		t.Fatalf("the first failure was not held: %+v", held)
 	}
 	time.Sleep(5 * time.Millisecond)
-	s.retryHeld(context.Background())
-	if !held.acked || len(s.parked) != 0 {
-		t.Fatalf("a report past the bound was not let go: %+v, %d held", held, len(s.parked))
+	in.retries(context.Background(), s)
+	if !held.acked || len(in.held) != 0 {
+		t.Fatalf("a report past the bound was not let go: %+v, %d held", held, len(in.held))
 	}
 }

internal/link/serve.go

@@ -7,31 +7,30 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
-	"github.com/novox/mesh-controller/internal/envfile"
-	"github.com/novox/mesh-controller/internal/inventory"
 	"log"
 	"os"
 	"time"
 
 	amqp "github.com/rabbitmq/amqp091-go"
+
+	"github.com/novox/mesh-controller/internal/envfile"
 )
 
-// AMQPVar is the control plane's own connection to the broker.
+// AMQPVar is the controller's own connection to the bus the mesh runs on today.
 const AMQPVar = "MESH_BROKER_AMQP"
 
-// Enroller is what the control plane does with an enrolment request.
+// Enroller is what the controller does with an enrolment request.
 //
-// An interface so the serving loop can be tested against a real broker without a database, and
-// so the two concerns — moving messages, and deciding — stay apart.
+// An interface so the serving loop can be tested against a real bus without a database, and so the
+// two concerns — moving messages, and deciding — stay apart.
 type Enroller interface {
 	// Enrol spends the token, records the key, and reports the node's name. The error is
 	// returned to the node as a refusal; it must be the same for every reason a token can fail.
 	Enrol(ctx context.Context, request EnrolRequest) (EnrolReply, error)
 }
 
-// Server consumes what nodes say.
-// Listener is what the control plane does with a report. Separate from Enroller so the two can
-// be given independently, and so a server that only sends declarations needs neither.
+// Listener is what the controller does with a report. Separate from Enroller so the two can be
+// given independently, and so a server that only sends declarations needs neither.
 type Listener interface {
 	Heard(ctx context.Context, report Report) error
 }
@@ -46,106 +45,77 @@ type Recorder interface {
 	Built(ctx context.Context, result BuildResult) error
 }
 
-type Server struct {
-	conn     *amqp.Connection
-	channel  *amqp.Channel
-	enroller Enroller
-	listener Listener
-	recorder Recorder
-	log      *log.Logger
-	upgrader Upgrader
-	replayer Replayer
-	// Messages the store could not take right now, held unacknowledged and tried again on a
-	// ticker, by subject (novox/hq issues 082, 083). again is the ticker's interval, zero meaning
-	// TryAgainAfter; giveUp is how long one is kept, zero meaning GiveUpAfter.
-	again  time.Duration
-	giveUp time.Duration
-	parked map[string]*held
-}
-
-// held is one message the store could not take, kept to be tried again.
-type held struct {
-	delivery amqp.Delivery
-	retry    func(context.Context, amqp.Delivery)
-	what     string
-	first    time.Time
-}
-
-// ErrTryAgain marks a listener's failure as "not now": what it was given is worth keeping and
-// asking again, as when the store is restarting (novox/hq issue 082).
-var ErrTryAgain = errors.New("not now, try again")
-
-// TryAgainAfter is how often messages the store could not take are tried again. A store comes
-// back in seconds, and a report a few seconds late is still current.
-const TryAgainAfter = 2 * time.Second
-
-// GiveUpAfter bounds how long one message is kept trying. A store that has not come back in this
-// long is not restarting, and the message is let go with a line saying it was lost.
-const GiveUpAfter = 2 * time.Minute
-
-// Prefetch is how many messages the broker hands the control plane before it has settled them.
-// More than one because a message the store could not take is held, unsettled, while the loop goes
-// on answering others — an enrolment above all, which a host is waiting on (novox/hq issue 083).
-// Bounded, because what is held is also what the broker has not kept on its own disk as pending.
-const Prefetch = 64
-
-// PrefetchHeadroom is how much of the prefetch is never held, so the loop always has messages to
-// answer — an enrolment above all — while others wait for the store.
-const PrefetchHeadroom = 8
-
-// Records tells the server where to keep build results.
-//
-// Set after Connect rather than passed to it, because a control plane that only publishes — the
-// `build` command, which waits for its own answer — needs a connection and no recorder, and
-// making it supply one would have it construct something it never uses.
-func (s *Server) Records(r Recorder) { s.recorder = r }
-
-// Upgrader is what the control plane does when the catalogue says a module moved.
+// Upgrader is what the controller does when the catalogue says a module moved.
 //
 // An interface for the same reason Enroller is one: deciding what an upgrade means for the
 // machines running it is a different concern from noticing that one was announced, and only the
 // first needs a database.
 type Upgrader interface {
 	// Upgraded is told which module moved and between which commits. An error is logged and the
-	// message is not requeued: an upgrade the control plane could not act on is not one it will
-	// act on by being handed the same message again, and a poison message on a durable queue
-	// would stop every upgrade behind it — except the store unreachable for the moment, which is
-	// asked again for a bounded time (novox/hq issue 083).
+	// message is not handed back: an upgrade the controller could not act on is not one it will
+	// act on by being given the same message again, and a poison message on a durable queue would
+	// stop every upgrade behind it — except the store unreachable for the moment, which is asked
+	// again for a bounded time (novox/hq issue 083).
 	Upgraded(ctx context.Context, u Upgraded) error
 }
 
-// Follows says what to do about upgrades, and binds the queue they arrive on.
+// Server acts on what nodes and modules say.
 //
-// **Not bound unless something is listening.** A durable queue bound to every upgrade with no
-// consumer fills up quietly, and the first symptom is a broker out of disk rather than anything
-// about modules.
+// **It holds no transport.** What arrives comes through Inbound and what it publishes goes through
+// Bus, so this file is the controller's *decisions* about messages and nothing about wires. The
+// connection below is the bus the mesh runs on today, kept because the command line publishes over
+// the same one until the rollout (ADR 0116).
+type Server struct {
+	inbound Inbound
+	bus     Bus
+	conn    *amqp.Connection
+	channel *amqp.Channel
+
+	enroller Enroller
+	listener Listener
+	recorder Recorder
+	upgrader Upgrader
+	replayer Replayer
+
+	log *log.Logger
+	// giveUp is how long one message is held for the store; zero means GiveUpAfter.
+	giveUp time.Duration
+}
+
+// ErrTryAgain marks a listener's failure as "not now": what it was given is worth keeping and
+// asking again, as when the store is restarting (novox/hq issue 082).
+var ErrTryAgain = errors.New("not now, try again")
+
+// GiveUpAfter bounds how long one message is kept trying. A store that has not come back in this
+// long is not restarting, and the message is let go with a line saying it was lost.
+const GiveUpAfter = 2 * time.Minute
+
+// Records tells the server where to keep build results.
+//
+// Set after Connect rather than passed to it, because a controller that only publishes — the
+// `build` command, which waits for its own answer — needs a connection and no recorder, and making
+// it supply one would have it construct something it never uses.
+func (s *Server) Records(r Recorder) { s.recorder = r }
+
+// Follows says what to do about upgrades, and asks for them to be delivered.
 func (s *Server) Follows(u Upgrader) error {
-	if _, err := s.channel.QueueDeclare(UpgradeQueue, true, false, false, false, nil); err != nil {
-		return fmt.Errorf("cannot declare the %s queue: %w", UpgradeQueue, err)
-	}
-	if err := s.channel.QueueBind(UpgradeQueue, KeyModuleUpgraded, EventsExchange, false, nil); err != nil {
-		return fmt.Errorf("cannot bind %s to %s/%s: %w", UpgradeQueue, EventsExchange, KeyModuleUpgraded, err)
+	if err := s.inbound.Also(KindModuleMoved); err != nil {
+		return err
 	}
 	s.upgrader = u
 	return nil
 }
 
-// Answers binds the queue a catalogue's catch-up request arrives on.
-//
-// **Not bound unless something is listening**, for the same reason upgrades are not: a durable
-// queue with no consumer fills quietly and the first symptom is a broker out of disk.
+// Answers says what to do about a catalogue's catch-up request, and asks for them to be delivered.
 func (s *Server) Answers(r Replayer) error {
-	if _, err := s.channel.QueueDeclare(CatchUpQueue, true, false, false, false, nil); err != nil {
-		return fmt.Errorf("cannot declare the %s queue: %w", CatchUpQueue, err)
-	}
-	if err := s.channel.QueueBind(CatchUpQueue, KeyCatchingUp, EventsExchange, false, nil); err != nil {
-		return fmt.Errorf("cannot bind %s to %s/%s: %w", CatchUpQueue, EventsExchange, KeyCatchingUp, err)
+	if err := s.inbound.Also(KindCatchUp); err != nil {
+		return err
 	}
 	s.replayer = r
 	return nil
 }
 
-// Connect opens the control plane's own connection to the broker.
+// Connect opens the controller's own connection to the bus.
 //
 // On the port MESH_BROKER_AMQP_PORT names when the node's settings moved the broker (novox/hq
 // 04-ISSUES/102) — the URL is genesis's, sealed, and its port is the one thing in it the node may
@@ -164,7 +134,7 @@ func Connect(enroller Enroller, listener Listener) (*Server, error) {
 
 	conn, err := amqp.Dial(url)
 	if err != nil {
-		// Not quoted back: the URL carries the control plane's own broker password.
+		// Not quoted back: the URL carries the controller's own bus password.
 		return nil, fmt.Errorf("cannot reach the broker named in %s: %w", AMQPVar, err)
 	}
 	channel, err := conn.Channel()
@@ -173,17 +143,17 @@ func Connect(enroller Enroller, listener Listener) (*Server, error) {
 		return nil, err
 	}
 
-	// Declared here rather than assumed. The control plane is the only thing that may create
-	// them — a node's account can write to this exchange and read its own queue, and configure
-	// nothing else, so a node arriving before the control plane has ever run finds nothing and
-	// says so, rather than quietly creating a topology nobody designed.
+	// Declared here rather than assumed. The controller is the only thing that may create them — a
+	// node's account can write to this exchange and read its own queue, and configure nothing
+	// else, so a node arriving before the controller has ever run finds nothing and says so,
+	// rather than quietly creating a topology nobody designed.
 	if err := channel.ExchangeDeclare(Exchange, "direct", true, false, false, false, nil); err != nil {
 		conn.Close()
 		return nil, fmt.Errorf("cannot declare the %s exchange: %w", Exchange, err)
 	}
-	// The events exchange too. The control plane is not the only publisher on it — modules
-	// announce onto it with their own accounts — but it is the only thing permitted to create it,
-	// for the same reason it is the only thing permitted to create the direct one.
+	// The events exchange too. The controller is not the only publisher on it — modules announce
+	// onto it with their own accounts — but it is the only thing permitted to create it, for the
+	// same reason it is the only thing permitted to create the direct one.
 	if err := channel.ExchangeDeclare(EventsExchange, "topic", true, false, false, false, nil); err != nil {
 		conn.Close()
 		return nil, fmt.Errorf("cannot declare the %s exchange: %w", EventsExchange, err)
@@ -192,7 +162,7 @@ func Connect(enroller Enroller, listener Listener) (*Server, error) {
 		conn.Close()
 		return nil, fmt.Errorf("cannot declare the %s queue: %w", ControlQueue, err)
 	}
-	// Every key a node may publish. Binding one and forgetting another is a message the broker
+	// Every key a node may publish. Binding one and forgetting another is a message the bus
 	// accepts, finds no queue for, and drops — the publisher sees success and the consumer sees
 	// nothing. That is exactly what happened to reports: `report` was left unbound while `enrol`
 	// worked, so nodes announced what they had applied into a void for an afternoon.
@@ -203,14 +173,24 @@ func Connect(enroller Enroller, listener Listener) (*Server, error) {
 		}
 	}
 
-	return &Server{conn: conn, channel: channel, enroller: enroller, listener: listener,
-		log: log.New(os.Stdout, "", log.LstdFlags)}, nil
+	return &Server{
+		inbound:  Current(conn, channel),
+		bus:      OverCurrent{Channel: channel},
+		conn:     conn,
+		channel:  channel,
+		enroller: enroller,
+		listener: listener,
+		log:      log.New(os.Stdout, "", log.LstdFlags),
+	}, nil
 }
 
-// Channel is the control plane's channel, for sending declarations.
+// Channel is the controller's channel, for the command line's own publishing.
 func (s *Server) Channel() *amqp.Channel { return s.channel }
 
 func (s *Server) Close() {
+	if s.inbound != nil {
+		s.inbound.Close()
+	}
 	if s.channel != nil {
 		_ = s.channel.Close()
 	}
@@ -219,131 +199,120 @@ func (s *Server) Close() {
 	}
 }
 
-// Serve consumes until the context ends.
-//
-// One consumer, deliberately: with two, the broker would round-robin between them and each would
-// receive half of what it expects — a fault this project has already had, between a module's
-// daemon and its capability server.
+// Serve acts on what arrives until the context ends.
 func (s *Server) Serve(ctx context.Context) error {
-	// A bounded prefetch rather than one. The loop still takes messages one at a time; what the
-	// prefetch buys is that a message the store could not take can be held while the loop goes on
-	// to the next, instead of every enrolment waiting behind it (novox/hq issue 083). Anything held
-	// goes back to the broker if the control plane stops, because nothing held is acknowledged.
-	if err := s.channel.Qos(Prefetch, 0, false); err != nil {
-		return err
-	}
-
-	deliveries, err := s.channel.ConsumeWithContext(ctx, ControlQueue, "control-plane",
-		false, false, false, false, nil)
-	if err != nil {
-		return err
-	}
-
-	// The upgrade queue, when something is listening for them. A second queue rather than a
-	// second consumer on the first: two consumers on one queue split its messages between them,
-	// which is the fault the comment above exists about. Two queues share nothing.
-	var upgrades <-chan amqp.Delivery
+	s.log.Printf("consuming what nodes say: %s, %s, %s, %s",
+		KindEnrolment, KindReport, KindHeartbeat, KindBuilt)
 	if s.upgrader != nil {
-		upgrades, err = s.channel.ConsumeWithContext(ctx, UpgradeQueue, "control-plane-upgrades",
-			false, false, false, false, nil)
-		if err != nil {
-			return err
-		}
+		s.log.Printf("following %s", KindModuleMoved)
 	}
-
-	// Its own queue and its own consumer, for the reason above: two consumers on one queue split
-	// its messages, and a catch-up request going to whichever half was not listening is a gap that
-	// looks like a working mesh.
-	var catchups <-chan amqp.Delivery
 	if s.replayer != nil {
-		catchups, err = s.channel.ConsumeWithContext(ctx, CatchUpQueue, "control-plane-catchup",
-			false, false, false, false, nil)
-		if err != nil {
-			return err
-		}
-	}
-
-	closed := s.conn.NotifyClose(make(chan *amqp.Error, 1))
-	s.log.Printf("consuming %s, bound to %s/{%s,%s,%s,%s}",
-		ControlQueue, Exchange, KeyEnrol, KeyReport, KeyAlive, KeyBuilt)
-	if s.upgrader != nil {
-		s.log.Printf("consuming %s, bound to %s/%s", UpgradeQueue, EventsExchange, KeyModuleUpgraded)
-	}
-
-	again := s.again
-	if again == 0 {
-		again = TryAgainAfter
-	}
-	ticker := time.NewTicker(again)
-	defer ticker.Stop()
-
-	for {
-		select {
-		case <-ctx.Done():
-			return nil
-		case <-ticker.C:
-			if ctx.Err() != nil {
-				return nil
-			}
-			s.retryHeld(ctx)
-		case delivery, ok := <-catchups:
-			if !ok {
-				if catchups != nil {
-					return errors.New("the broker stopped delivering catch-up requests")
-				}
-				continue
-			}
-			s.catchingUp(ctx, delivery)
-		case delivery, ok := <-upgrades:
-			// A nil channel blocks for ever, so this case simply never fires when nothing is
-			// listening for upgrades. Closed is different, and means the broker stopped.
-			if !ok {
-				if upgrades != nil {
-					return errors.New("the broker stopped delivering upgrades")
-				}
-				continue
-			}
-			s.upgraded(ctx, delivery)
-		case reason := <-closed:
-			// Said rather than returned quietly. A control plane whose broker connection dropped
-			// is a mesh where nothing can be told anything, and the reason is the first thing
-			// anybody will want.
-			return fmt.Errorf("the broker connection closed: %v", reason)
-		case delivery, ok := <-deliveries:
-			if !ok {
-				return errors.New("the broker stopped delivering")
-			}
-			s.handle(ctx, delivery)
-		}
+		s.log.Printf("answering %s", KindCatchUp)
 	}
+	return s.inbound.Receive(ctx, s.act)
 }
 
-func (s *Server) handle(ctx context.Context, delivery amqp.Delivery) {
-	switch delivery.RoutingKey {
-	case KeyEnrol:
-		s.handleEnrol(ctx, delivery)
-	case KeyReport:
-		s.handleReport(ctx, delivery)
-	case KeyAlive:
-		s.handleAlive(delivery)
-	case KeyBuilt:
-		s.handleBuilt(ctx, delivery)
+// act is one message, whichever bus it came over.
+func (s *Server) act(ctx context.Context, m Control) {
+	switch m.Kind() {
+	case KindEnrolment:
+		s.enrolling(ctx, m)
+	case KindReport:
+		s.reported(ctx, m)
+	case KindHeartbeat:
+		s.heartbeat(m)
+	case KindBuilt:
+		s.wasBuilt(ctx, m)
+	case KindModuleMoved:
+		s.moved(ctx, m)
+	case KindCatchUp:
+		s.catchingUp(ctx, m)
 	default:
-		// Rejected without requeue: a message nothing understands will not be understood on the
-		// next attempt either, and requeuing it would spin.
-		s.log.Printf("refusing a message with routing key %q", delivery.RoutingKey)
-		_ = delivery.Reject(false)
+		// Dropped: a message nothing understands will not be understood on the next attempt
+		// either, and asking for it again would spin.
+		s.log.Printf("refusing a message the mesh has no name for: %q", m.Kind())
+		_ = m.Drop()
 	}
 }
 
-// handleAlive records that a node was heard from, and nothing else.
+// decide asks the window what to do about one message, and holds it when that is the answer —
+// because holding is the one verdict that means the same thing everywhere.
 //
-// Deliberately silent: a node saying it is there every minute would fill the log with the
-// ordinary case, and a log where the ordinary case is loud is a log nobody reads.
-func (s *Server) handleAlive(delivery amqp.Delivery) {
+// The other three come back, because "settled without acting" is a build result dropped and an
+// upgrade taken, and only the handler knows which its message is. Hold means the message is the
+// bus's problem now and **the caller must not settle it**; that is also what a cancelled context
+// gets, because shutting down is not an answer about a message (novox/hq issue 083, on review).
+//
+// declaredIn is the declaration this message is about, empty when it is about none; outstanding is
+// what the mesh last sent that node, empty when it has sent none or could not be asked.
+func (s *Server) decide(ctx context.Context, m Control, what, declaredIn, outstanding string,
+	err error) Verdict {
+
+	if ctx.Err() != nil {
+		return Hold
+	}
+	window := StoreWindow{GiveUpAfter: s.giveUpAfter()}
+	switch v := window.Decide(err, declaredIn, outstanding, m.HeldFor()); v {
+	case Hold:
+		// Said once, on the first hold. Every redelivery saying it again would fill the log with
+		// one store restart.
+		first := m.HeldFor() == 0
+		if held := m.Hold(RedeliverAfter(m.HeldFor())); held != nil {
+			s.log.Printf("LOST %s: it could not be held for the store (%v): %v", what, held, err)
+			return GiveUp
+		}
+		if first {
+			s.log.Printf("could not keep %s yet; holding it to try again: %v", what, err)
+		}
+		return Hold
+	case Stale:
+		s.log.Printf("set aside %s: the mesh has moved past that declaration", what)
+		return Stale
+	case GiveUp:
+		s.log.Printf("LOST %s: the store has not come back in %s: %v", what, s.giveUpAfter(), err)
+		return GiveUp
+	default:
+		return v
+	}
+}
+
+func (s *Server) giveUpAfter() time.Duration {
+	if s.giveUp == 0 {
+		return GiveUpAfter
+	}
+	return s.giveUp
+}
+
+// outstanding is the digest of the declaration the mesh last sent a node.
+//
+// Nothing is guessed when it cannot be answered: a listener that keeps no record of what was sent
+// gives a report nothing to be stale against, and a store that cannot be read will hold the report
+// anyway, so there is nothing for the check to decide.
+func (s *Server) outstanding(ctx context.Context, node string) string {
+	if node == "" {
+		return ""
+	}
+	asks, ok := s.listener.(Outstanding)
+	if !ok {
+		return ""
+	}
+	digest, err := asks.Outstanding(ctx, node)
+	if err != nil {
+		return ""
+	}
+	return digest
+}
+
+// heartbeat records that a node was heard from, and nothing else.
+//
+// Deliberately silent: a node saying it is there every minute would fill the log with the ordinary
+// case, and a log where the ordinary case is loud is a log nobody reads. Not held for the store
+// either — the next heartbeat is a minute away, and a heartbeat kept for two minutes to be written
+// late says nothing the one after it will not say better.
+func (s *Server) heartbeat(m Control) {
 	var alive Alive
-	if err := json.Unmarshal(delivery.Body, &alive); err != nil || alive.Node == "" {
-		_ = delivery.Reject(false)
+	if err := json.Unmarshal(m.Body(), &alive); err != nil || alive.Node == "" {
+		_ = m.Drop()
 		return
 	}
 	if s.listener != nil {
@@ -351,34 +320,49 @@ func (s *Server) handleAlive(delivery amqp.Delivery) {
 			s.log.Printf("could not record that %s is here: %v", alive.Node, err)
 		}
 	}
-	_ = delivery.Ack(false)
+	_ = m.Took()
 }
 
-// handleReport records what a node says it did.
+// reported records what a node says it did.
 //
 // A node states; nothing here writes anything the node claimed about itself beyond that it was
 // heard from. What it applied is its own account of its own machine, and the mesh keeps the last
 // one as a copy for recovery rather than as a source (novox/hq 09-the-node-lifecycle).
-func (s *Server) handleReport(ctx context.Context, delivery amqp.Delivery) {
+func (s *Server) reported(ctx context.Context, m Control) {
 	var report Report
-	if err := json.Unmarshal(delivery.Body, &report); err != nil {
+	if err := json.Unmarshal(m.Body(), &report); err != nil {
 		s.log.Printf("a report could not be read: %v", err)
-		_ = delivery.Reject(false)
+		_ = m.Drop()
 		return
 	}
-	// A node's newer report supersedes one of its older reports still held: the older is its
-	// past, and recorded after the newer it would overwrite what the node is doing now.
-	subject := "report " + report.Node
-	s.supersede(subject, delivery)
+	m.About("report " + report.Node)
+	what := fmt.Sprintf("%s's report of declaration %s", report.Node, report.Declared)
+
 	if s.listener != nil {
-		if err := s.listener.Heard(context.Background(), report); err != nil {
-			// Held, not acknowledged, while the store cannot take it: the node reports an apply
-			// once, and a report lost here is a node the mesh never hears from again — the store
+		// **Asked before the store, not after** (design 25 §3). A report about a declaration the
+		// mesh has moved past would otherwise wait out a restarting store to be written and then
+		// overwrite what the node is doing now — and on the bus being built, where the holding is
+		// the server's, it comes back after the newer was applied whatever the controller does.
+		outstanding := s.outstanding(ctx, report.Node)
+		declaredIn := staleAgainst(report)
+		if Superseded(declaredIn, outstanding) {
+			s.log.Printf("set aside %s: the mesh has moved past that declaration", what)
+			_ = m.Took()
+			return
+		}
+
+		err := s.listener.Heard(context.Background(), report)
+		switch s.decide(ctx, m, what, declaredIn, outstanding, err) {
+		case Hold:
+			// Held, not settled, while the store cannot take it: the node reports an apply once,
+			// and a report lost here is a node the mesh never hears from again — the store
 			// restarting under the adoption that node just applied lost exactly that (issue 082).
-			what := fmt.Sprintf("%s's report of declaration %s", report.Node, report.Declared)
-			if s.tryLater(ctx, delivery, subject, what, err, s.handle) {
-				return
-			}
+			return
+		case Stale, GiveUp:
+			_ = m.Took()
+			return
+		}
+		if err != nil {
 			// Said rather than swallowed. A report the mesh heard and failed to write down is a
 			// node whose recovery copy is silently older than it looks.
 			s.log.Printf("could not record %s's report: %v", report.Node, err)
@@ -393,124 +377,49 @@ func (s *Server) handleReport(ctx context.Context, delivery amqp.Delivery) {
 	default:
 		s.log.Printf("%s applied %d resource(s)", report.Node, len(report.Applied))
 	}
-	s.settled(subject, delivery)
-	_ = delivery.Ack(false)
+	_ = m.Took()
 }
 
-// tryLater holds a message the store could not take right now, to be tried again on the ticker,
-// and says whether it did (novox/hq issues 082, 083).
+// staleAgainst is the declaration a report may be judged stale against, and it is empty for a
+// report that is not only an account of an apply.
 //
-// "Right now" is the store unreachable or restarting — ErrTryAgain from a listener, or an error the
-// inventory reads as an outage. Anything else is an answer, and is left to the caller to settle.
-// Held means unacknowledged and set aside: the loop goes on to the next message, so an enrolment a
-// host is waiting on is answered while a report waits for the store. One message is held at most
-// giveUpAfter; past it, it is let go with a line saying it was lost, and the caller settles it.
-func (s *Server) tryLater(ctx context.Context, delivery amqp.Delivery, subject, what string, err error,
-	retry func(context.Context, amqp.Delivery)) bool {
-	// Shutting down: nothing is settled. Unsettled, the broker hands the message to whatever
-	// consumes next — a cancelled context is not an answer about the message (issue 083, review).
-	if ctx.Err() != nil {
-		return true
+// **A report carries two different things, and only one of them is about a declaration.** What the
+// node applied is; what the machine *is* — the tunnel it took over, the ports its own bundle holds,
+// what an adopted node found and is keeping, a node moving its overlay key — is not. Those reach
+// the mesh on a report because a report is the message a node sends, and nowhere else: a rekey
+// dropped as stale is a node whose overlay key never moves, and no retry is coming, because the
+// node said it once.
+//
+// So staleness is asked only of a report that is purely an apply's account. The rest is acted on
+// whenever it arrives, which is the behaviour the mesh has had all along.
+func staleAgainst(report Report) string {
+	if report.Rekey != nil || report.Tunnel != nil || len(report.Held) > 0 ||
+		report.Firewall != "" || len(report.Reachable) > 0 || len(report.Carried) > 0 {
+		return ""
 	}
-	if !errors.Is(err, ErrTryAgain) && !inventory.Unreachable(err) {
-		return false
-	}
-	if s.parked == nil {
-		s.parked = map[string]*held{}
-	}
-	h, ok := s.parked[subject]
-	if !ok || h.delivery.DeliveryTag != delivery.DeliveryTag {
-		// Held no further than the prefetch leaves room: past it, the broker would hand the loop
-		// nothing new — enrolments included — until something held was let go.
-		if !ok && len(s.parked) >= Prefetch-PrefetchHeadroom {
-			s.log.Printf("LOST %s: %d messages are already held for the store, and holding more "+
-				"would stop the queue: %v", what, len(s.parked), err)
-			return false
-		}
-		h = &held{delivery: delivery, retry: retry, what: what, first: time.Now()}
-		s.parked[subject] = h
-		s.log.Printf("could not keep %s yet; holding it to try again: %v", what, err)
-		return true
-	}
-	if time.Since(h.first) >= s.giveUpAfter() {
-		delete(s.parked, subject)
-		s.log.Printf("LOST %s: the store has not come back in %s: %v", what, s.giveUpAfter(), err)
-		return false
-	}
-	return true
+	return report.Declared
 }
 
-// supersede drops a message held for a subject when a newer one for it arrives: the older is
-// acknowledged, because acting on it after the newer would undo the newer.
-func (s *Server) supersede(subject string, newer amqp.Delivery) {
-	h, ok := s.parked[subject]
-	if !ok || h.delivery.DeliveryTag == newer.DeliveryTag {
-		return
-	}
-	delete(s.parked, subject)
-	s.log.Printf("set aside %s: a newer one arrived", h.what)
-	_ = h.delivery.Ack(false)
-}
-
-// settled forgets a message once it has been handled either way.
-func (s *Server) settled(subject string, delivery amqp.Delivery) {
-	if h, ok := s.parked[subject]; ok && h.delivery.DeliveryTag == delivery.DeliveryTag {
-		delete(s.parked, subject)
-	}
-}
-
-// digest names a message by its content.
-func digest(body []byte) string {
-	sum := sha256.Sum256(body)
-	return hex.EncodeToString(sum[:])
-}
-
-// retryHeld tries every held message again. Each handler holds it again, settles it, or lets it
-// go past the bound.
-func (s *Server) retryHeld(ctx context.Context) {
-	for _, h := range s.snapshot() {
-		if ctx.Err() != nil {
-			return
-		}
-		h.retry(ctx, h.delivery)
-	}
-}
-
-func (s *Server) snapshot() []*held {
-	out := make([]*held, 0, len(s.parked))
-	for _, h := range s.parked {
-		out = append(out, h)
-	}
-	return out
-}
-
-func (s *Server) giveUpAfter() time.Duration {
-	if s.giveUp == 0 {
-		return GiveUpAfter
-	}
-	return s.giveUp
-}
-
-func (s *Server) handleEnrol(ctx context.Context, delivery amqp.Delivery) {
+func (s *Server) enrolling(ctx context.Context, m Control) {
 	reply := EnrolReply{Refusal: "that token cannot be used"}
 
 	var request EnrolRequest
-	if err := json.Unmarshal(delivery.Body, &request); err != nil {
+	if err := json.Unmarshal(m.Body(), &request); err != nil {
 		s.log.Printf("an enrolment request could not be read: %v", err)
 	} else {
-		request.Redelivered = delivery.Redelivered
+		request.Redelivered = m.Redelivered()
 		accepted, err := s.enroller.Enrol(ctx, request)
 		switch {
 		case errors.Is(err, ErrTryAgain):
 			// Not a refusal: nothing was spent, and the same request asked again will be
-			// answered. Replied at once rather than held, so the node — which is waiting on
-			// this answer — decides when to ask, and the queue behind it moves (issue 083).
+			// answered. Replied at once rather than held, so the node — which is waiting on this
+			// answer — decides when to ask, and the queue behind it moves (issue 083).
 			reply = EnrolReply{TryAgain: true, Refusal: "the mesh cannot answer right now; ask again"}
 			s.log.Printf("asked %q to enrol again shortly: %v", request.Node, err)
 		case err != nil && request.Redelivered:
-			// Said as what it most likely is: the broker handed this request over again after
-			// the control plane stopped mid-answer, and an enrolment already spent is not
-			// finished a second time. The node may need a new token.
+			// Said as what it most likely is: the bus handed this request over again after the
+			// controller stopped mid-answer, and an enrolment already spent is not finished a
+			// second time. The node may need a new token.
 			s.log.Printf("refusing a redelivered enrolment for %q — it may have finished before "+
 				"the control plane stopped, and if the node did not get its answer it needs a new "+
 				"token: %v", request.Node, err)
@@ -524,169 +433,160 @@ func (s *Server) handleEnrol(ctx context.Context, delivery amqp.Delivery) {
 		}
 	}
 
-	s.reply(ctx, delivery, reply)
-
-	// Acknowledged after the reply is sent, so a control plane that dies mid-answer leaves the
-	// request on the broker rather than having consumed it silently. Asked again by the same
-	// presenter, an enrolment finishes: the token is held for its key and spent last (issue 083).
-	_ = delivery.Ack(false)
-}
-
-func (s *Server) reply(ctx context.Context, delivery amqp.Delivery, reply EnrolReply) {
-	if delivery.ReplyTo == "" {
-		s.log.Print("an enrolment request named no reply queue, so nothing can be told the answer")
-		return
-	}
-	body, err := json.Marshal(reply)
-	if err != nil {
+	if body, err := json.Marshal(reply); err != nil {
 		s.log.Printf("cannot encode a reply: %v", err)
-		return
+	} else {
+		answer, cancel := context.WithTimeout(ctx, 10*time.Second)
+		if err := m.Answer(answer, body); err != nil {
+			s.log.Printf("cannot answer an enrolment: %v", err)
+		}
+		cancel()
 	}
-	timeout, cancel := context.WithTimeout(ctx, 10*time.Second)
-	defer cancel()
 
-	if err := s.channel.PublishWithContext(timeout, "", delivery.ReplyTo, false, false,
-		amqp.Publishing{
-			ContentType:   "application/json",
-			CorrelationId: delivery.CorrelationId,
-			Body:          body,
-		}); err != nil {
-		s.log.Printf("cannot reply to %s: %v", delivery.ReplyTo, err)
-	}
+	// Settled after the answer is sent, so a controller that dies mid-answer leaves the request on
+	// the bus rather than having consumed it silently. Asked again by the same presenter, an
+	// enrolment finishes: the token is held for its key and spent last (issue 083).
+	_ = m.Took()
 }
 
-// handleBuilt keeps what a builder said, whichever way it went.
+// wasBuilt keeps what a builder said, whichever way it went.
 //
 // This is for results nobody was waiting for. A build asked for with `build` is answered directly
 // to the asker; one triggered any other way is published here, and without this it would be
 // reported into the void — which is the same as not reporting it.
-func (s *Server) handleBuilt(ctx context.Context, delivery amqp.Delivery) {
+func (s *Server) wasBuilt(ctx context.Context, m Control) {
 	var result BuildResult
-	if err := json.Unmarshal(delivery.Body, &result); err != nil {
+	if err := json.Unmarshal(m.Body(), &result); err != nil {
 		s.log.Printf("a build result could not be read: %v", err)
-		_ = delivery.Reject(false)
+		_ = m.Drop()
 		return
 	}
 	if s.recorder == nil {
-		// Nothing to keep it in. Rejected rather than dropped silently, so the broker's own
-		// counters show something arriving that nothing handles.
+		// Nothing to keep it in. Dropped rather than swallowed, so the bus's own counters show
+		// something arriving that nothing handles.
 		s.log.Printf("a build result arrived and this control plane keeps none")
-		_ = delivery.Reject(false)
+		_ = m.Drop()
 		return
 	}
 	// Each build result its own subject: none supersedes another, and recording one twice is
 	// harmless — the build is kept by its id.
-	subject := "build " + digest(delivery.Body)
-	s.supersede(subject, delivery)
-	if err := s.recorder.Built(ctx, result); err != nil {
+	m.About("build " + digest(m.Body()))
+	err := s.recorder.Built(ctx, result)
+	switch s.decide(ctx, m, fmt.Sprintf("a build result from %s", result.On), "", "", err) {
+	case Hold:
 		// Held while the store cannot take it: a build result lost here is never announced (083).
-		if s.tryLater(ctx, delivery, subject, fmt.Sprintf("a build result from %s", result.On), err, s.handle) {
-			return
-		}
+		return
+	case Stale, GiveUp:
+		_ = m.Drop()
+		return
+	}
+	if err != nil {
 		s.log.Printf("cannot keep a build result from %s: %v", result.On, err)
-		s.settled(subject, delivery)
-		_ = delivery.Reject(false)
+		_ = m.Drop()
 		return
 	}
-	s.settled(subject, delivery)
 	switch {
 	case result.Failed != "":
 		s.log.Printf("%s could not build %s", result.On, result.Repository)
 	default:
 		s.log.Printf("%s built %s from %s", result.On, result.Repository, result.Commit)
 	}
-	_ = delivery.Ack(false)
+	_ = m.Took()
 }
 
 // catchingUp answers a catalogue that has just started and may have missed builds.
 //
-// Acknowledged after the work. A replay that fails for a reason other than the store is not one
-// that succeeds by being handed the same request again, so that is acknowledged and said; but a
-// store that could not be read right now is held and asked again, bounded, rather than the
-// request lost until the catalogue next restarts (issue 083). One request stands for all: a newer
-// one supersedes one still held.
-func (s *Server) catchingUp(ctx context.Context, delivery amqp.Delivery) {
-	const subject = "catch-up"
-	s.supersede(subject, delivery)
-	holding := false
-	defer func() {
-		if !holding {
-			s.settled(subject, delivery)
-			_ = delivery.Ack(false)
-		}
-	}()
+// Settled after the work. A replay that fails for a reason other than the store is not one that
+// succeeds by being handed the same request again, so that is settled and said; but a store that
+// could not be read right now is held and asked again, bounded, rather than the request lost until
+// the catalogue next restarts (issue 083). One request stands for all: a newer one supersedes one
+// still held.
+func (s *Server) catchingUp(ctx context.Context, m Control) {
+	m.About("catch-up")
 	if s.replayer == nil {
 		s.log.Printf("a catalogue asked to catch up and this control plane has nothing to replay")
+		_ = m.Took()
 		return
 	}
 	announcements, err := s.replayer.Announceable(ctx)
+	switch s.decide(ctx, m, "a catalogue's request to catch up", "", "", err) {
+	case Hold:
+		return
+	case Stale, GiveUp:
+		_ = m.Took()
+		return
+	}
 	if err != nil {
-		if s.tryLater(ctx, delivery, subject, "a catalogue's request to catch up", err, s.catchingUp) {
-			holding = true
-			return
-		}
 		s.log.Printf("a catalogue asked to catch up and the mesh could not read its builds: %v", err)
+		_ = m.Took()
 		return
 	}
 	sent := 0
 	for _, a := range announcements {
 		a.Replay = true
-		if err := EmitEvent(ctx, s.channel, KeyModuleBuilt, "control-plane", "", a); err != nil {
+		if err := EmitEvent(ctx, s.bus, KeyModuleBuilt, "control-plane", "", a); err != nil {
 			// Said and abandoned rather than retried: the catalogue asks again every time it
 			// starts, and half a graph delivered twice is no better than half delivered once.
 			s.log.Printf("replaying %s at %s failed, and the rest is abandoned: %v",
 				a.Module, short(a.Commit), err)
+			_ = m.Took()
 			return
 		}
 		sent++
 	}
 	s.log.Printf("a catalogue asked to catch up; re-announced %d build(s)", sent)
+	_ = m.Took()
 }
 
-// upgraded hands one announcement to whatever is following them.
+// moved hands one announcement to whatever is following them.
 //
-// **Acknowledged whatever happens, but one thing.** A failure here is usually the control plane
-// being unable to act on an upgrade — a machine that cannot be resolved, a broker that will not
-// take a declaration — and none of those get better by being handed the same message again. The
-// one exception is the upgrader saying the store could not be read for the moment (ErrTryAgain):
-// that is held and asked again, bounded (novox/hq issue 083). Only the upgrader's word counts
-// here, not an error that merely looks like an outage — a push that timed out on the second
-// machine is not asked again, or the first would be pushed every few seconds for two minutes.
-// A newer move of the same module supersedes one still held.
-func (s *Server) upgraded(ctx context.Context, delivery amqp.Delivery) {
+// **Settled whatever happens, but one thing.** A failure here is usually the controller being
+// unable to act on an upgrade — a machine that cannot be resolved, a bus that will not take a
+// declaration — and none of those get better by being handed the same message again. The one
+// exception is the upgrader saying the store could not be read for the moment (ErrTryAgain): that
+// is held and asked again, bounded (novox/hq issue 083). Only the upgrader's word counts here, not
+// an error that merely looks like an outage — a push that timed out on the second machine is not
+// asked again, or the first would be pushed every few seconds for two minutes. A newer move of the
+// same module supersedes one still held.
+func (s *Server) moved(ctx context.Context, m Control) {
 	var u Upgraded
-	_ = json.Unmarshal(delivery.Body, &u)
-	subject := "upgrade " + u.Module
-	s.supersede(subject, delivery)
-	holding := false
-	defer func() {
-		if !holding {
-			s.settled(subject, delivery)
-			_ = delivery.Ack(false)
-		}
-	}()
-	if err := json.Unmarshal(delivery.Body, &u); err != nil {
+	if err := json.Unmarshal(m.Body(), &u); err != nil {
 		s.log.Printf("an upgrade announcement could not be read: %v", err)
+		_ = m.Took()
 		return
 	}
+	m.About("upgrade " + u.Module)
 	if u.Module == "" {
 		s.log.Printf("an upgrade announcement named no module; ignored")
+		_ = m.Took()
 		return
 	}
-	if err := s.upgrader.Upgraded(ctx, u); err != nil {
-		// Shutting down is not an answer about the announcement: left for the broker.
-		if ctx.Err() != nil {
-			holding = true
-			return
-		}
-		if errors.Is(err, ErrTryAgain) &&
-			s.tryLater(ctx, delivery, subject, fmt.Sprintf("%s's move to %s", u.Module, short(u.Commit)), err, s.upgraded) {
-			holding = true
-			return
-		}
+	err := s.upgrader.Upgraded(ctx, u)
+	// Only "not now" is worth holding. Anything else is an answer, and the window would read a
+	// timed-out push as an outage and ask for it again.
+	holdable := err
+	if !errors.Is(err, ErrTryAgain) {
+		holdable = nil
+	}
+	what := fmt.Sprintf("%s's move to %s", u.Module, short(u.Commit))
+	switch s.decide(ctx, m, what, "", "", holdable) {
+	case Hold:
+		return
+	case Stale, GiveUp:
+		_ = m.Took()
+		return
+	}
+	if err != nil {
 		s.log.Printf("%s moved to %s and the mesh could not act on it: %v",
 			u.Module, short(u.Commit), err)
 	}
+	_ = m.Took()
+}
+
+// digest names a message by its content.
+func digest(body []byte) string {
+	sum := sha256.Sum256(body)
+	return hex.EncodeToString(sum[:])
 }
 
 // short is a commit as people read it.

internal/link/stale_report_test.go

@@ -0,0 +1,135 @@
+package link
+
+import (
+	"context"
+	"errors"
+	"testing"
+)
+
+// Supersession as a check rather than a memory (design 25 §3).
+//
+// Holding a message in memory let the controller drop an older report when a newer one for the
+// same node arrived. On the bus being built the message belongs to the server and comes back
+// whatever happened meanwhile — so the older report is redelivered *after* the newer was applied,
+// and acting on it would undo the newer.
+//
+// The answer was already in the message: a report carries the digest of the declaration it is
+// about, so "is this the past?" is a question the message answers.
+
+// sentAndHeard records reports and knows what was last sent, which is the pair the check needs.
+type sentAndHeard struct {
+	sent  string
+	heard []Report
+	err   error
+}
+
+func (s *sentAndHeard) Heard(_ context.Context, r Report) error {
+	if s.err != nil {
+		return s.err
+	}
+	s.heard = append(s.heard, r)
+	return nil
+}
+
+func (s *sentAndHeard) Outstanding(context.Context, string) (string, error) { return s.sent, nil }
+
+// A report about a declaration the mesh has moved past is settled and not acted on. Settled rather
+// than dropped, because there is nothing wrong with the message — it is simply the past, and
+// redelivering it for ever is worse than letting it go.
+func TestAReportAboutASupersededDeclarationIsNotActedOn(t *testing.T) {
+	store := &sentAndHeard{sent: "d2"}
+	s, in := serving()
+	s.listener = store
+	to := &settled{}
+	s.act(context.Background(), in.sends(t, to, KindReport, aReport("anchor", "d1")))
+
+	if len(store.heard) != 0 {
+		t.Fatalf("a report about a superseded declaration was acted on: %+v", store.heard)
+	}
+	if !to.acked {
+		t.Fatalf("a superseded report was not settled, so it comes back for ever: %+v", *to)
+	}
+}
+
+// The report about the declaration that *is* outstanding is acted on, and so is one from a node
+// the mesh has no digest for — an older host that says nothing about which declaration it applied
+// has nothing to be judged against, and refusing it would silence every node built before reports
+// carried the digest.
+func TestAReportAboutTheOutstandingDeclarationIsActedOn(t *testing.T) {
+	for _, c := range []struct{ what, sent, declared string }{
+		{"the one outstanding", "d2", "d2"},
+		{"a report that says nothing about which", "d2", ""},
+		{"a node nothing was ever sent", "", "d1"},
+	} {
+		store := &sentAndHeard{sent: c.sent}
+		s, in := serving()
+		s.listener = store
+		to := &settled{}
+		s.act(context.Background(), in.sends(t, to, KindReport, aReport("anchor", c.declared)))
+		if len(store.heard) != 1 || !to.acked {
+			t.Errorf("%s: was not acted on and acknowledged: heard %+v, settled %+v",
+				c.what, store.heard, *to)
+		}
+	}
+}
+
+// **Staleness is decided before the store is waited on**, not after: a redelivery that lost its
+// race is not worth holding a slot in the window that a current message needs.
+func TestASupersededReportIsNotHeldForTheStore(t *testing.T) {
+	store := &sentAndHeard{sent: "d2", err: errors.Join(ErrTryAgain, errors.New("starting up"))}
+	s, in := serving()
+	s.listener = store
+	to := &settled{}
+	s.act(context.Background(), in.sends(t, to, KindReport, aReport("anchor", "d1")))
+	if !to.acked || len(in.held) != 0 {
+		t.Fatalf("a superseded report waited for the store: %+v, %d held", *to, len(in.held))
+	}
+}
+
+// **Half of a report is not about a declaration, and that half is never stale.**
+//
+// What the machine *is* — the tunnel it took over, the ports its own bundle holds, what an adopted
+// node found and is keeping, a node moving its overlay key — reaches the mesh on a report and
+// nowhere else. A rekey set aside as stale is a node whose overlay key never moves, and no retry is
+// coming, because the node said it once. So a report carrying any of these is acted on whenever it
+// arrives, however far the mesh has moved on.
+func TestAReportCarryingWhatOnlyTheNodeKnowsIsActedOnHoweverOldItIs(t *testing.T) {
+	for _, c := range []struct {
+		what   string
+		report Report
+	}{
+		{"a rekey", Report{Node: "anchor", Declared: "d1",
+			Rekey: &Rekey{Previous: "k1", OverlayKey: "k2"}}},
+		{"the tunnel it carried", Report{Node: "anchor", Declared: "d1",
+			Tunnel: &CarriedTunnel{Interface: "wg0", State: "taken"}}},
+		{"what an adopted node holds", Report{Node: "anchor", Declared: "d1",
+			Held: []Held{{ID: "conf", Module: "web", Kind: "file"}}}},
+		{"the firewall it found", Report{Node: "anchor", Declared: "d1", Firewall: "ufw"}},
+		{"what is reachable on it", Report{Node: "anchor", Declared: "d1",
+			Reachable: []Reach{{Protocol: "tcp", Port: 443}}}},
+		{"the ports its own bundle holds", Report{Node: "anchor", Declared: "d1",
+			Carried: []int{5432}}},
+	} {
+		store := &sentAndHeard{sent: "d9"}
+		s, in := serving()
+		s.listener = store
+		to := &settled{}
+		s.act(context.Background(), in.sends(t, to, KindReport, c.report))
+		if len(store.heard) != 1 {
+			t.Errorf("%s was set aside as stale, and the mesh will never hear it again: %+v",
+				c.what, *to)
+		}
+	}
+}
+
+// A heartbeat is not held for the store: the next one is a minute away, and one kept for two
+// minutes to be written late says nothing the one after it will not say better.
+func TestAHeartbeatIsNotHeldForTheStore(t *testing.T) {
+	s, in := serving()
+	s.listener = heardWith{err: errors.Join(ErrTryAgain, errors.New("starting up"))}
+	to := &settled{}
+	s.act(context.Background(), in.sends(t, to, KindHeartbeat, Alive{Node: "anchor"}))
+	if !to.acked || len(in.held) != 0 {
+		t.Fatalf("a heartbeat was held for the store: %+v, %d held", *to, len(in.held))
+	}
+}

internal/link/store_window_test.go

@@ -2,15 +2,11 @@ package link
 
 import (
 	"context"
-	"encoding/json"
 	"errors"
 	"fmt"
-	"io"
-	"log"
 	"testing"
 
 	"github.com/jackc/pgx/v5/pgconn"
-	amqp "github.com/rabbitmq/amqp091-go"
 )
 
 // What a store restarting under an adoption answers with (novox/hq issues 082, 083).
@@ -28,29 +24,6 @@ type replaysWith struct{ err error }
 
 func (r replaysWith) Announceable(context.Context) ([]Announcement, error) { return nil, r.err }
 
-type settledAs struct{ acked, nacked, requeued, rejected bool }
-
-func (a *settledAs) Ack(uint64, bool) error { a.acked = true; return nil }
-func (a *settledAs) Nack(_ uint64, _ bool, requeue bool) error {
-	a.nacked, a.requeued = true, requeue
-	return nil
-}
-func (a *settledAs) Reject(uint64, bool) error { a.rejected = true; return nil }
-
-func a(t *testing.T, to *settledAs, key string, v any) amqp.Delivery {
-	t.Helper()
-	body, err := json.Marshal(v)
-	if err != nil {
-		t.Fatal(err)
-	}
-	tag++
-	return amqp.Delivery{Acknowledger: to, RoutingKey: key, Body: body, DeliveryTag: tag}
-}
-
-func quietServer() *Server { return &Server{log: log.New(io.Discard, "", 0)} }
-
-func (a *settledAs) held() bool { return !a.acked && !a.nacked && !a.rejected }
-
 // A build result the store could not take right now is handed back; one it refused is rejected,
 // as before; one it kept is acknowledged.
 func TestABuildResultWaitsOutARestartingStore(t *testing.T) {
@@ -58,16 +31,16 @@ func TestABuildResultWaitsOutARestartingStore(t *testing.T) {
 	for _, c := range []struct {
 		what string
 		err  error
-		want func(*settledAs) bool
+		want func(*settled) bool
 	}{
-		{"restarting", restarting, func(s *settledAs) bool { return s.held() }},
-		{"refused", errors.New("no such module"), func(s *settledAs) bool { return s.rejected && !s.nacked }},
-		{"kept", nil, func(s *settledAs) bool { return s.acked && !s.nacked }},
+		{"restarting", restarting, func(s *settled) bool { return s.unsettled() }},
+		{"refused", errors.New("no such module"), func(s *settled) bool { return s.rejected && !s.nacked }},
+		{"kept", nil, func(s *settled) bool { return s.acked && !s.nacked }},
 	} {
-		s := quietServer()
+		s, in := serving()
 		s.recorder = recordsWith{err: c.err}
-		to := &settledAs{}
-		s.handleBuilt(context.Background(), a(t, to, KeyBuilt, built))
+		to := &settled{}
+		s.act(context.Background(), in.sends(t, to, KindBuilt, built))
 		if !c.want(to) {
 			t.Errorf("%s: a build result was settled as %+v", c.what, *to)
 		}
@@ -81,17 +54,17 @@ func TestAnUpgradeWaitsOutARestartingStoreAndNothingElse(t *testing.T) {
 	for _, c := range []struct {
 		what string
 		err  error
-		want func(*settledAs) bool
+		want func(*settled) bool
 	}{
-		{"the store away, said by the upgrader", errors.Join(ErrTryAgain, restarting), func(s *settledAs) bool { return s.held() }},
-		{"a push that timed out", context.DeadlineExceeded, func(s *settledAs) bool { return s.acked && !s.nacked }},
-		{"cannot act", errors.New("anchor cannot be resolved"), func(s *settledAs) bool { return s.acked && !s.nacked }},
-		{"acted", nil, func(s *settledAs) bool { return s.acked && !s.nacked }},
+		{"the store away, said by the upgrader", errors.Join(ErrTryAgain, restarting), func(s *settled) bool { return s.unsettled() }},
+		{"a push that timed out", context.DeadlineExceeded, func(s *settled) bool { return s.acked && !s.nacked }},
+		{"cannot act", errors.New("anchor cannot be resolved"), func(s *settled) bool { return s.acked && !s.nacked }},
+		{"acted", nil, func(s *settled) bool { return s.acked && !s.nacked }},
 	} {
-		s := quietServer()
+		s, in := serving()
 		s.upgrader = upgradesWith{err: c.err}
-		to := &settledAs{}
-		s.upgraded(context.Background(), a(t, to, "upgraded", moved))
+		to := &settled{}
+		s.act(context.Background(), in.sends(t, to, KindModuleMoved, moved))
 		if !c.want(to) {
 			t.Errorf("%s: an upgrade was settled as %+v", c.what, *to)
 		}
@@ -104,16 +77,16 @@ func TestACatchUpWaitsOutARestartingStore(t *testing.T) {
 	for _, c := range []struct {
 		what string
 		err  error
-		want func(*settledAs) bool
+		want func(*settled) bool
 	}{
-		{"restarting", restarting, func(s *settledAs) bool { return s.held() }},
-		{"unreadable", errors.New("a build row is malformed"), func(s *settledAs) bool { return s.acked && !s.nacked }},
-		{"nothing to replay", nil, func(s *settledAs) bool { return s.acked && !s.nacked }},
+		{"restarting", restarting, func(s *settled) bool { return s.unsettled() }},
+		{"unreadable", errors.New("a build row is malformed"), func(s *settled) bool { return s.acked && !s.nacked }},
+		{"nothing to replay", nil, func(s *settled) bool { return s.acked && !s.nacked }},
 	} {
-		s := quietServer()
+		s, in := serving()
 		s.replayer = replaysWith{err: c.err}
-		to := &settledAs{}
-		s.catchingUp(context.Background(), a(t, to, "catch-up", map[string]string{}))
+		to := &settled{}
+		s.act(context.Background(), in.sends(t, to, KindCatchUp, map[string]string{}))
 		if !c.want(to) {
 			t.Errorf("%s: a catch-up request was settled as %+v", c.what, *to)
 		}
@@ -121,15 +94,15 @@ func TestACatchUpWaitsOutARestartingStore(t *testing.T) {
 }
 
 // Shutting down is not an answer about a message: one handled with a cancelled context is left
-// unsettled, for the broker to hand to whatever consumes next (issue 083, review).
-func TestAMessageHandledDuringShutdownIsLeftForTheBroker(t *testing.T) {
+// unsettled, for the bus to hand to whatever consumes next (issue 083, review).
+func TestAMessageHandledDuringShutdownIsLeftForTheBus(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	cancel()
-	s := quietServer()
+	s, in := serving()
 	s.recorder = recordsWith{err: context.Canceled}
-	to := &settledAs{}
-	s.handleBuilt(ctx, a(t, to, KeyBuilt, BuildResult{On: "anchor", Repository: "/r", Commit: "abc"}))
-	if !to.held() {
+	to := &settled{}
+	s.act(ctx, in.sends(t, to, KindBuilt, BuildResult{On: "anchor", Repository: "/r", Commit: "abc"}))
+	if !to.unsettled() {
 		t.Fatalf("a build result handled during shutdown was settled, and so lost: %+v", *to)
 	}
 }
@@ -137,45 +110,45 @@ func TestAMessageHandledDuringShutdownIsLeftForTheBroker(t *testing.T) {
 // Two identical build results: the newer sets the older aside rather than leaving it unsettled
 // for ever, holding a place in the prefetch.
 func TestAnIdenticalBuildResultSetsTheHeldOneAside(t *testing.T) {
-	s := quietServer()
+	s, in := serving()
 	s.recorder = recordsWith{err: restarting}
 	built := BuildResult{On: "anchor", Repository: "/r", Commit: "abc"}
-	first, second := &settledAs{}, &settledAs{}
-	s.handleBuilt(context.Background(), a(t, first, KeyBuilt, built))
-	s.handleBuilt(context.Background(), a(t, second, KeyBuilt, built))
-	if !first.acked || !second.held() || len(s.parked) != 1 {
+	first, second := &settled{}, &settled{}
+	s.act(context.Background(), in.sends(t, first, KindBuilt, built))
+	s.act(context.Background(), in.sends(t, second, KindBuilt, built))
+	if !first.acked || !second.unsettled() || len(in.held) != 1 {
 		t.Fatalf("an identical build result did not set the held one aside: first %+v second %+v, %d held",
-			*first, *second, len(s.parked))
+			*first, *second, len(in.held))
 	}
 }
 
 // What is held stops short of the prefetch, so the loop always has room to answer an enrolment.
 func TestWhatIsHeldLeavesRoomInThePrefetch(t *testing.T) {
-	s := quietServer()
+	s, in := serving()
 	s.recorder = recordsWith{err: restarting}
-	var last *settledAs
+	var last *settled
 	for i := 0; i < Prefetch; i++ {
-		last = &settledAs{}
-		s.handleBuilt(context.Background(), a(t, last, KeyBuilt, BuildResult{On: "anchor", Commit: fmt.Sprint(i)}))
+		last = &settled{}
+		s.act(context.Background(), in.sends(t, last, KindBuilt, BuildResult{On: "anchor", Commit: fmt.Sprint(i)}))
 	}
-	if len(s.parked) != Prefetch-PrefetchHeadroom {
-		t.Fatalf("%d messages were held; the ceiling is %d", len(s.parked), Prefetch-PrefetchHeadroom)
+	if len(in.held) != Prefetch-PrefetchHeadroom {
+		t.Fatalf("%d messages were held; the ceiling is %d", len(in.held), Prefetch-PrefetchHeadroom)
 	}
-	if last.held() {
+	if last.unsettled() {
 		t.Fatalf("a message past the ceiling was held: %+v", *last)
 	}
 }
 
-// An upgrade handled during shutdown is left for the broker too — the upgrader's error is the
+// An upgrade handled during shutdown is left for the bus too — the upgrader's error is the
 // cancelled context, which is no answer about the announcement.
-func TestAnUpgradeHandledDuringShutdownIsLeftForTheBroker(t *testing.T) {
+func TestAnUpgradeHandledDuringShutdownIsLeftForTheBus(t *testing.T) {
 	ctx, cancel := context.WithCancel(context.Background())
 	cancel()
-	s := quietServer()
+	s, in := serving()
 	s.upgrader = upgradesWith{err: context.Canceled}
-	to := &settledAs{}
-	s.upgraded(ctx, a(t, to, "upgraded", Upgraded{Module: "gitea", Commit: "abcdef0123"}))
-	if !to.held() {
-		t.Fatalf("an upgrade handled during shutdown was settled, and so lost: %+v", *to)
+	to := &settled{}
+	s.act(ctx, in.sends(t, to, KindModuleMoved, Upgraded{Module: "gitea", Commit: "abcdef0123"}))
+	if !to.unsettled() {
+		t.Fatalf("an upgrade was settled during shutdown, and so lost: %+v", *to)
 	}
 }

internal/link/window.go

@@ -0,0 +1,120 @@
+package link
+
+import (
+	"errors"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// The store window, as a decision rather than a mechanism.
+//
+// The guarantee (novox/hq ADR 0083): a push the controller cannot record because its store is
+// restarting is **held and retried**, not dropped and not falsely acknowledged. On the bus the
+// mesh runs on today that is done by keeping the delivery unacknowledged in memory and settling
+// it later. On the bus being built it is a `nak` with a delay: the server holds it and redelivers,
+// so the controller keeps no list of parked messages and a controller that restarts mid-window
+// loses nothing it was holding.
+//
+// **The decision is the same either way, and the mechanism is not the interesting part.** What is
+// interesting is that moving the holding into the server introduces a problem the in-memory
+// version did not have, and the answer was already in the message.
+
+// Verdict is what to do with one control message.
+type Verdict int
+
+const (
+	// Take it: apply, then acknowledge.
+	Take Verdict = iota
+	// Hold it: the store cannot record this yet. Nak with a delay and let the server redeliver.
+	Hold
+	// Stale: this is about a declaration the node has already moved past, and applying it would
+	// undo what came after. Acknowledge without acting — redelivering forever is worse.
+	Stale
+	// GiveUp: the store has not come back in time. Settle it and say so, loudly.
+	GiveUp
+)
+
+// StoreWindow decides. Pure, so the guarantee is testable without a bus, a store or a clock.
+type StoreWindow struct {
+	// GiveUpAfter is how long one message may be held before it is let go with a line saying so.
+	GiveUpAfter time.Duration
+}
+
+// Decide answers for one delivery.
+//
+//   - err is what the store said, or nil.
+//   - declaredIn is the digest of the declaration this message is about, empty when it is not
+//     about one (an enrolment, a build result).
+//   - outstanding is the digest the mesh last sent that node, empty when it has sent none.
+//   - heldFor is how long this message has already been held; zero on first delivery.
+func (w StoreWindow) Decide(err error, declaredIn, outstanding string, heldFor time.Duration) Verdict {
+	// **Staleness is checked before the store, not after.** A redelivery that lost its race is
+	// not worth waiting on a store for, and asking the store first would mean a message about a
+	// superseded declaration holding a slot in the window that a current one needs.
+	if Superseded(declaredIn, outstanding) {
+		return Stale
+	}
+	if err == nil {
+		return Take
+	}
+	if !errors.Is(err, ErrTryAgain) && !inventory.Unreachable(err) {
+		// Not the store being away: a refusal is an answer, and holding it would turn a message
+		// the mesh understood into one it retries forever.
+		return Take
+	}
+	if heldFor >= w.GiveUpAfter {
+		return GiveUp
+	}
+	return Hold
+}
+
+// Superseded says a message is about a declaration the mesh has already moved past.
+//
+// Stated on its own because it is asked in two places for one reason: here, so the window's whole
+// decision is in one pure function, and by the serving loop *before* it asks the store, because
+// that is the point — a message about the past must not wait on a store, or it holds a slot in the
+// window that a current message needs.
+//
+// Unanswerable is not stale. A message that names no declaration, and a node the mesh has never
+// sent one, both give nothing to compare: the mesh acts on the message rather than guessing, which
+// is also what keeps a host built before reports carried the digest from going silent.
+func Superseded(declaredIn, outstanding string) bool {
+	return declaredIn != "" && outstanding != "" && declaredIn != outstanding
+}
+
+// RedeliverAfter is how long the server should hold a naked message before trying again.
+//
+// Backed off, and bounded. A store restarting is back in seconds; a store that is gone is not
+// helped by being asked every second, and the delay is what keeps a window of held messages from
+// becoming a spin.
+func RedeliverAfter(heldFor time.Duration) time.Duration {
+	switch {
+	case heldFor < 5*time.Second:
+		return time.Second
+	case heldFor < 30*time.Second:
+		return 5 * time.Second
+	default:
+		return 15 * time.Second
+	}
+}
+
+// The problem holding-in-the-server introduces, and why the answer was already in the message.
+//
+// Holding a delivery in memory let the controller do something a server cannot: when a newer
+// report for the same node arrived, it dropped the older one, "because acting on it after the
+// newer would undo the newer". A `nak`ed message is the server's, and the server will redeliver
+// it whatever else has happened in the meantime — so the older report comes back *after* the
+// newer was applied, and applying it would undo exactly what that comment describes.
+//
+// **A report already says which declaration it is about.** `Declared` is the digest of the exact
+// bytes the mesh sent, and it exists because an earlier attempt to order reports by time lost the
+// race it invited — an apply that started under the previous declaration finishes after the next
+// is sent, and the report reads as newer than the send. Clocks cannot answer *which*; the digest
+// is the answer itself.
+//
+// So supersession stops being a thing the controller remembers and becomes a thing it checks: a
+// report whose digest is not the one outstanding for that node is stale, and is acknowledged
+// without being acted on. Which is the same shape as a node refusing a superseded declaration by
+// sequence (novox/hq issue 107) — ordering settled by what the message says, not by when it
+// happened to arrive.

internal/link/window_test.go

@@ -0,0 +1,84 @@
+package link
+
+import (
+	"errors"
+	"testing"
+	"time"
+)
+
+func window() StoreWindow { return StoreWindow{GiveUpAfter: 2 * time.Minute} }
+
+// The guarantee itself (novox/hq ADR 0083): a push the store cannot record is held, not dropped
+// and not falsely acknowledged.
+func TestAMessageTheStoreCannotTakeYetIsHeld(t *testing.T) {
+	if got := window().Decide(ErrTryAgain, "", "", 0); got != Hold {
+		t.Fatalf("got %v; a push the store could not record was not held", got)
+	}
+}
+
+// A refusal is an answer. Holding it would turn a message the mesh understood into one it
+// retries forever.
+func TestARefusalIsNotHeld(t *testing.T) {
+	if got := window().Decide(errors.New("that node does not exist"), "", "", 0); got != Take {
+		t.Fatalf("got %v; a refusal was mistaken for the store being away", got)
+	}
+}
+
+// Held has a limit, and past it the message is settled rather than held for ever.
+func TestAStoreThatNeverComesBackEndsTheHold(t *testing.T) {
+	if got := window().Decide(ErrTryAgain, "", "", 3*time.Minute); got != GiveUp {
+		t.Fatalf("got %v; the window has no end", got)
+	}
+	if got := window().Decide(ErrTryAgain, "", "", time.Minute); got != Hold {
+		t.Fatalf("got %v; the window ended early", got)
+	}
+}
+
+// **The problem that holding in the server introduces.** A naked message is redelivered whatever
+// else happened meanwhile, so a report about a superseded declaration comes back after the newer
+// one was applied — and applying it would undo the newer.
+func TestAReportAboutASupersededDeclarationIsNotApplied(t *testing.T) {
+	got := window().Decide(nil, "digest-of-the-old-one", "digest-of-the-current-one", 0)
+	if got != Stale {
+		t.Fatalf("got %v; a redelivery that lost its race would have undone what came after", got)
+	}
+}
+
+func TestAReportAboutTheOutstandingDeclarationIsApplied(t *testing.T) {
+	if got := window().Decide(nil, "same", "same", 0); got != Take {
+		t.Fatalf("got %v; a current report was discarded", got)
+	}
+}
+
+// A message that is about no declaration — an enrolment, a build result — is never stale: there
+// is nothing for it to be out of date with.
+func TestAMessageAboutNoDeclarationIsNeverStale(t *testing.T) {
+	if got := window().Decide(nil, "", "whatever-is-outstanding", 0); got != Take {
+		t.Fatalf("got %v; an enrolment was treated as a stale report", got)
+	}
+	if got := window().Decide(nil, "a-digest", "", 0); got != Take {
+		t.Fatalf("got %v; a report was called stale against a node that was sent nothing", got)
+	}
+}
+
+// Staleness is decided before the store is waited on: a redelivery that lost its race must not
+// hold a slot in the window that a current message needs.
+func TestAStaleMessageIsNotHeldForTheStore(t *testing.T) {
+	if got := window().Decide(ErrTryAgain, "old", "current", 0); got != Stale {
+		t.Fatalf("got %v; a superseded message was held for a store it would never be applied to", got)
+	}
+}
+
+// The delay backs off: a store that is gone is not helped by being asked every second, and the
+// delay is what keeps a window of held messages from becoming a spin.
+func TestRedeliveryBacksOff(t *testing.T) {
+	first := RedeliverAfter(0)
+	later := RedeliverAfter(10 * time.Second)
+	last := RedeliverAfter(time.Minute)
+	if !(first < later && later < last) {
+		t.Fatalf("delays do not back off: %v %v %v", first, later, last)
+	}
+	if last > 30*time.Second {
+		t.Fatalf("a held message waits %v between attempts, which is longer than a store restart", last)
+	}
+}