Merge pull request 'Assert every declared state's bucket on each push (hq ADR 0201)' (#262) from fix/buckets-on-push into main

The raise at start was the only place buckets were asserted, so a module
registered and assigned since had none until the control plane restarted —
found on the first module to declare state.

Dockerfile

@@ -1,5 +1,11 @@
-ARG GO_BASE=golang:1.25-alpine
+# The Go it builds with, pinned here because genesis builds this file with no arguments (novox/hq
-# The control plane's image.
+# issue 223) — the Makefile passes the same digest. A tag older than go.mod asks for is how
+# `make image` broke once before (issue 146).
+ARG GO_BASE=golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c
+# The control plane's image — for genesis and the lab only. The mesh runs the controller as a Go
+# bundle the host starts as a process (module.json; novox/hq issue 213), and builds no image of it.
+# Genesis builds this file and raises it as the container the process replaces on the first push
+# (mesh-host internal/bootstrap, novox/hq issue 223).
 #
 # novox/hq ADR 0006: this image is pinned by digest in the bundle the host carries, fetched on a
 # machine where no mesh exists yet, and run before there is anything to check it against. So it

Makefile

@@ -27,17 +27,21 @@ build:
 IMAGE ?= mesh-controller:$(VERSION)
 DEV_TAG ?= mesh-controller:development
 
-# The base the module declares, read from the manifest rather than written here twice.
+# The Go base the image is built on.
 #
 # **`make image` was broken and stayed broken**, because the Dockerfile's fallback base was a Go
 # older than go.mod asks for: every build died at `go mod download` with "go.mod requires go >=
 # 1.26.0", and the pipeline never saw it because the pipeline passes the declared base in. Anybody
 # building the image by hand hit it and had to find the digest themselves (novox/hq 04-ISSUES/146,
 # what it cost).
-GO_BASE ?= $(shell python3 -c "import json;print(next(o['image'] for o in json.load(open('module.json'))['build']['on'] if o['arg']=='GO_BASE'))" 2>/dev/null)
+#
+# **Pinned here since the manifest stopped building an image** (novox/hq issue 213): the mesh builds
+# the controller as a Go bundle with its own toolchain, and only `make image` — genesis and the lab —
+# still needs a Go base. The digest is the one the manifest declared until then.
+GO_BASE ?= golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c
 
 image:
-	@test -n "$(GO_BASE)" || { echo "module.json declares no GO_BASE; pass GO_BASE=<image> or fix the manifest"; exit 1; }
+	@test -n "$(GO_BASE)" || { echo "no GO_BASE; pass GO_BASE=<image>"; exit 1; }
 	docker build --build-arg GO_BASE=$(GO_BASE) --build-arg VERSION=$(VERSION) -t $(IMAGE) -t $(DEV_TAG) .
 	@echo
 	@docker image inspect $(IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'
@@ -48,7 +52,7 @@ BUILDER_IMAGE ?= mesh-builder:$(VERSION)
 BUILDER_DEV_TAG ?= mesh-builder:development
 
 builder-image:
-	@test -n "$(GO_BASE)" || { echo "module.json declares no GO_BASE; pass GO_BASE=<image> or fix the manifest"; exit 1; }
+	@test -n "$(GO_BASE)" || { echo "no GO_BASE; pass GO_BASE=<image>"; exit 1; }
 	docker build --build-arg GO_BASE=$(GO_BASE) -f cmd/mesh-builder/Dockerfile -t $(BUILDER_IMAGE) -t $(BUILDER_DEV_TAG) .
 	@echo
 	@docker image inspect $(BUILDER_IMAGE) --format 'built {{.RepoTags}}  {{.Size}} bytes'

README.md

@@ -193,6 +193,13 @@ passes every check that only looks at the message.
 
 ## The image
 
+**The mesh no longer runs the controller from it** (novox/hq issue 213). The module declares a Go
+bundle, `controller`, which the host on the controller's machine unpacks and runs as the process
+`mesh-controller` under the account of the same name (ADR 0188 §1, §3). The image stays for what
+still runs a container of the controller: genesis, which raises the first controller from it and
+installs the module from its manifest (mesh-host `internal/bootstrap`), and the lab. Neither is the
+mesh's own build any more — `make image` builds it.
+
 `FROM scratch`, holding one statically linked binary and nothing else — no shell, no package
 manager, no libc, no CA certificates.
 

cmd/mesh-controller/addresses_test.go

@@ -145,7 +145,7 @@ func TestTheRegistryTrustAndEveryImageFollowThePortTheNodeGaveTheStore(t *testin
 //
 // Composed from the control plane's own manifest against a real inventory: the store's module is
 // given 6852 on this node the way genesis or an operator gives it, and the control plane's
-// container is told so beside the sealed connection genesis wrote.
+// process is told so beside the sealed connection genesis wrote.
 func TestTheControlPlaneIsToldWhereTheNodePutTheStoreAndTheBroker(t *testing.T) {
 	open := aMesh(t)
 	ctx := t.Context()
@@ -157,8 +157,8 @@ func TestTheControlPlaneIsToldWhereTheNodePutTheStoreAndTheBroker(t *testing.T)
 	if err != nil {
 		t.Fatal(err)
 	}
-	control, err := withSeatPorts(m).Resolve([]catalogue.Built{{Name: "server", Kind: catalogue.ArtifactImage,
+	control, err := withSeatPorts(m).Resolve([]catalogue.Built{{Name: "controller", Kind: catalogue.ArtifactBundle,
-		Reference: "registry.example/control@" + aDigest}})
+		Reference: "https://registry.example/mesh-controller/controller.tar.gz", Digest: aDigest}})
 	if err != nil {
 		t.Fatal(err)
 	}
@@ -200,12 +200,12 @@ func TestTheControlPlaneIsToldWhereTheNodePutTheStoreAndTheBroker(t *testing.T)
 
 	var env map[string]any
 	for _, r := range composed(t, open, "anchor").Resources {
-		if r["id"] == "mesh-controller.server" {
+		if r["id"] == "mesh-controller.controller" {
 			env, _ = r["env"].(map[string]any)
 		}
 	}
 	if env == nil {
-		t.Fatal("the control plane's container is not in its own node's declaration")
+		t.Fatal("the control plane's process is not in its own node's declaration")
 	}
 	for key, want := range map[string]string{
 		"MESH_STORE_INVENTORY_PORT": "6852",
@@ -238,7 +238,7 @@ func withSeatPorts(m catalogue.Manifest) catalogue.Manifest {
 	out := m
 	out.Resources = nil
 	for _, r := range m.Resources {
-		if r["type"] != "container" {
+		if r["type"] != "container" && r["type"] != "process" {
 			out.Resources = append(out.Resources, r)
 			continue
 		}

cmd/mesh-controller/build.go

@@ -558,6 +558,10 @@ func takeIn(ctx context.Context, inv *inventory.Inventory, result link.BuildResu
 		}
 		return manifest, kept, err
 	}
+	// The keep set just moved, and new bytes just landed (novox/hq ADR 0189). Asked here rather
+	// than on a timer of its own: this is the only moment either is true. Never fatal — the build
+	// worked and the module is registered.
+	collect(ctx, inv)
 	return manifest, kept, nil
 }
 

cmd/mesh-controller/check.go

@@ -7,6 +7,7 @@ import (
 	"os"
 	"path/filepath"
 	"sort"
+	"strings"
 
 	"github.com/novox/mesh-controller/internal/catalogue"
 )
@@ -90,6 +91,17 @@ func moduleCheck(paths []string, out io.Writer) error {
 		if len(m.Invokes) > 0 {
 			fmt.Fprintf(out, ", invokes %s", joinInvokes(m.Invokes))
 		}
+		// The state it keeps and reads (novox/hq ADR 0201), so a reviewer sees what lands on the bus.
+		if len(m.State) > 0 {
+			kept := make([]string, 0, len(m.State))
+			for _, s := range m.State {
+				kept = append(kept, s.Name)
+			}
+			fmt.Fprintf(out, ", keeps state %s", strings.Join(kept, ", "))
+		}
+		if len(m.Reads) > 0 {
+			fmt.Fprintf(out, ", reads %s", strings.Join(m.Reads, ", "))
+		}
 		fmt.Fprintln(out)
 	}
 	if failed > 0 {

cmd/mesh-controller/collect.go

@@ -0,0 +1,108 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"os"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/artifacts"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Letting the artifact store go of what the mesh no longer keeps (novox/hq ADR 0189, issue 108).
+//
+// **Run where the records change.** A build is the moment new bytes landed in the store and the
+// moment the keep set moved, so it is the moment to say what may go — and it needs no timer of
+// its own. Reclaiming the bytes is the store's own nightly step; this only decides.
+//
+// Never fatal to a build. The build succeeded, the module is registered, and a store that could
+// not be reached is a thing to say rather than a reason to undo any of that. The next build asks
+// again, and the references it could not collect are still uncollected, so nothing is lost by
+// having failed.
+
+// collect asks the store to let go of everything the mesh made and no longer keeps, and records
+// what it let go of. Says what it did and what it could not; returns nothing, because nothing
+// upstream should branch on it.
+func collect(ctx context.Context, inv *inventory.Inventory) {
+	references, err := inv.ToCollect(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not work out what the artifact store may let go of: %v\n", err)
+		return
+	}
+	if len(references) == 0 {
+		return
+	}
+	shelf, err := inv.Catalogue(ctx)
+	if err != nil {
+		fmt.Fprintf(os.Stderr, "could not read the catalogue to find the artifact store: %v\n", err)
+		return
+	}
+	// As the mesh reaches it from the network. Empty means the store is not on the network — on a
+	// mesh being raised it is not yet, and there the store holds one build of anything and has
+	// nothing to collect.
+	address, err := artifactStoreAddress(ctx, inv, shelf, "")
+	if err != nil || address == "" {
+		if err != nil {
+			fmt.Fprintf(os.Stderr, "could not find the artifact store to collect from: %v\n", err)
+		}
+		return
+	}
+
+	// **Bounded, because this runs inside somebody's build.** The first sweep of a mesh that has
+	// never collected has the whole history to get through, and a person waiting on `build` should
+	// not pay for it. Two bounds, and what is left over is simply offered again next time —
+	// builds are frequent, and the point is that the store stops growing, not that it empties
+	// tonight.
+	within, stop := context.WithTimeout(ctx, sweepBudget)
+	defer stop()
+	store := artifacts.Store{Address: address}
+
+	var done []string
+	var left int
+	for i, reference := range references {
+		if i >= mostPerSweep || within.Err() != nil {
+			left = len(references) - i
+			break
+		}
+		err := store.LetGo(within, reference)
+		if err == nil || errors.Is(err, artifacts.Gone) {
+			// Gone is the outcome wanted, already true. Recorded so the next sweep does not ask
+			// again for ever.
+			done = append(done, reference)
+			continue
+		}
+		// **Stopped at the first refusal, not pushed through.** A store that refuses one refuses
+		// all of them — deletion disabled, the store down, the network gone — so going on would
+		// be a hundred identical failures and a hundred identical log lines in front of whoever
+		// was building something.
+		fmt.Fprintf(os.Stderr, "the artifact store kept %s, so nothing more was asked of it: %v\n",
+			reference, err)
+		left = len(references) - i
+		break
+	}
+
+	if len(done) > 0 {
+		// Recorded outside `within`: the deletions happened, and losing the record of them because
+		// the sweep ran out of budget would mean asking about them again for ever.
+		if err := inv.MarkCollected(ctx, done); err != nil {
+			fmt.Fprintf(os.Stderr, "the store let go of %d artifact(s) and the record of it did not keep: %v\n",
+				len(done), err)
+			return
+		}
+		fmt.Fprintf(os.Stderr, "the artifact store let go of %d artifact(s) the mesh no longer keeps\n",
+			len(done))
+	}
+	if left > 0 {
+		fmt.Fprintf(os.Stderr, "%d more to collect; the next build asks again\n", left)
+	}
+}
+
+// mostPerSweep is how many artifacts one sweep will ask about. Enough that a mesh building
+// several times a day converges within days of this landing; small enough that no single build
+// waits on the whole backlog.
+const mostPerSweep = 200
+
+// sweepBudget is the longest a sweep will keep a build waiting.
+const sweepBudget = 60 * time.Second

cmd/mesh-controller/push.go

@@ -747,6 +747,16 @@ func issueMemberships(ctx context.Context, open *stores, server *link.Server, se
 	if !ok {
 		return nil
 	}
+	// **Every declared state's bucket, before the memberships that name it** (novox/hq ADR 0201). The
+	// raise at start asserts them too, but a module registered and assigned since would otherwise have
+	// its bucket only after the control plane next restarts — found the first time a module declared
+	// state: its bundle asked for a bucket that did not exist. Idempotent and cheap; a failure is said
+	// and the push stands, as a membership's is.
+	if buckets, err := open.inventory.DeclaredBuckets(ctx); err != nil {
+		fmt.Printf("  the modules' state could not be read, so no bucket was asserted: %v\n", err)
+	} else if _, err := broker.RaiseBuckets(broker.OnConn(bus.Conn), buckets); err != nil {
+		fmt.Printf("  the modules' state could not be asserted on the bus: %v — the next push tries again\n", err)
+	}
 	// The declarations are sent and recorded by now; a membership that cannot be issued is said
 	// and does not unsay them. Every runtime without one serves the shape it derives (ADR 0160), so
 	// the push stands, the first failure is named once, and the next push tries again.
@@ -898,6 +908,21 @@ func raiseTheBus(ctx context.Context, inv *inventory.Inventory, address string)
 	if err := broker.RaiseSeats(js, inventory.MeshSeats(), holders); err != nil {
 		return err
 	}
+	// Every module's state (novox/hq ADR 0201), from the catalogue: a bucket exists from
+	// registration, so a module reading one may watch it before its owner runs anywhere. One that
+	// nothing declares any more is said and kept — what it holds is data.
+	buckets, err := inv.DeclaredBuckets(ctx)
+	if err != nil {
+		return err
+	}
+	undeclared, err := broker.RaiseBuckets(js, buckets)
+	if err != nil {
+		return err
+	}
+	if len(undeclared) > 0 {
+		fmt.Printf("the bus holds state nothing declares any more, kept because it is data: %s — "+
+			"removing it is a person's act\n", strings.Join(undeclared, ", "))
+	}
 	// And how every module hears what it consumes. Derived from the same records the user list is
 	// composed from, so a module the mesh grants a consumer's subjects has that consumer waiting.
 	// Done on every raise, not only when a credential is issued: every module moved onto this bus
@@ -921,8 +946,8 @@ func raiseTheBus(ctx context.Context, inv *inventory.Inventory, address string)
 		}
 		hearing++
 	}
-	fmt.Printf("the bus at %s has its streams, %d machine(s) can hear a declaration, and %d module(s) "+
+	fmt.Printf("the bus at %s has its streams, %d machine(s) can hear a declaration, %d module(s) "+
-		"can hear what they consume\n", broker.BareAddress(address), len(names), hearing)
+		"can hear what they consume, and %d bucket(s) of state\n", broker.BareAddress(address), len(names), hearing, len(buckets))
 	return nil
 }
 

internal/artifacts/store.go

@@ -0,0 +1,99 @@
+// Package artifacts speaks to the mesh's artifact store over its own door.
+//
+// Only what the mesh needs that nothing else does: letting go of something it put there
+// (novox/hq ADR 0189, issue 108). Pushing is the builder's, through the container runtime; reading
+// is every machine's, through its runtime. This is the one operation that belongs to the thing
+// holding the records, because it is the only one that is a decision rather than a transfer.
+package artifacts
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// Store is the artifact store at an address, as this machine reaches it.
+type Store struct {
+	// Address is `host:port` — the store as the caller reaches it now, composed and never
+	// recorded (novox/hq 04-ISSUES/102).
+	Address string
+	// HTTP is the client used; nil is a client with a modest timeout.
+	HTTP *http.Client
+}
+
+// Gone is the answer when the store does not hold it: the outcome wanted, already true.
+var Gone = fmt.Errorf("the store does not hold it")
+
+// LetGo asks the store to drop one artifact the mesh recorded making.
+//
+// Takes a reference as the mesh records it — `artifact-store://<module>/<artifact>@sha256:…` for
+// an image, `…/blobs/sha256:…` for an archive — because that is the identity every record uses,
+// and composes the address here at the moment of use.
+//
+// Returns Gone when the store answers that it does not have it. That is not a failure: the sweep
+// wants the artifact absent, and it is. It is distinguished from success only so a caller can say
+// which of the two happened.
+func (s Store) LetGo(ctx context.Context, reference string) error {
+	path, kept := catalogue.InArtifactStore(reference)
+	if !kept {
+		// Nothing the mesh put in its own store. Refused rather than attempted: composing a
+		// delete for a reference of unknown shape is how a sweep reaches something that is not
+		// the mesh's.
+		return fmt.Errorf("%s is not a reference into the mesh's artifact store", reference)
+	}
+	if s.Address == "" {
+		return fmt.Errorf("this mesh has no artifact store on its network to ask about %s", reference)
+	}
+	repository, kind, digest, err := split(path)
+	if err != nil {
+		return err
+	}
+	url := "http://" + s.Address + "/v2/" + repository + "/" + kind + "/" + digest
+
+	request, err := http.NewRequestWithContext(ctx, http.MethodDelete, url, nil)
+	if err != nil {
+		return err
+	}
+	client := s.HTTP
+	if client == nil {
+		client = &http.Client{Timeout: 30 * time.Second}
+	}
+	response, err := client.Do(request)
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	switch response.StatusCode {
+	case http.StatusAccepted, http.StatusOK, http.StatusNoContent:
+		return nil
+	case http.StatusNotFound:
+		return Gone
+	case http.StatusMethodNotAllowed:
+		// The registry was started without deletion enabled. Said plainly, because the remedy is
+		// a setting on the store's module and not anything about this artifact.
+		return fmt.Errorf(
+			"the artifact store refuses deletion: its server was started without it enabled "+
+				"(REGISTRY_STORAGE_DELETE_ENABLED), so nothing can be collected until the store "+
+				"module is applied again (novox/hq ADR 0189). Asking about %s", reference)
+	default:
+		return fmt.Errorf("the artifact store answered %s for %s", response.Status, reference)
+	}
+}
+
+// split reads a recorded path into the repository, which endpoint names the thing, and the digest.
+//
+// Two shapes, which are the two the mesh records: `<repository>@sha256:<hex>` is a manifest, and
+// `<repository>/blobs/sha256:<hex>` is a blob.
+func split(path string) (repository, kind, digest string, err error) {
+	if before, after, ok := strings.Cut(path, "@sha256:"); ok {
+		return before, "manifests", "sha256:" + after, nil
+	}
+	if before, after, ok := strings.Cut(path, "/blobs/sha256:"); ok {
+		return before, "blobs", "sha256:" + after, nil
+	}
+	return "", "", "", fmt.Errorf("%q names nothing the store holds by digest", path)
+}

internal/artifacts/store_test.go

@@ -0,0 +1,94 @@
+package artifacts
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// Asking the store to let go of what the mesh no longer keeps (novox/hq ADR 0189, issue 108).
+//
+// A fake store records what it was asked to delete, so what is asserted is the mesh's decision
+// and the shape of the request — not the registry's behaviour, which is the registry's to test.
+
+func fakeStore(t *testing.T, answer int) (Store, *[]string) {
+	t.Helper()
+	var asked []string
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodDelete {
+			t.Errorf("the store was asked %s %s; collecting is a delete", r.Method, r.URL.Path)
+		}
+		asked = append(asked, r.URL.Path)
+		w.WriteHeader(answer)
+	}))
+	t.Cleanup(server.Close)
+	return Store{Address: strings.TrimPrefix(server.URL, "http://")}, &asked
+}
+
+func TestAnImageAndAnArchiveAreAskedForAtTheirOwnEndpoints(t *testing.T) {
+	// The two shapes the mesh records: a manifest by digest, and a blob by digest. They are
+	// different endpoints, and asking at the wrong one answers 404 — which this would then
+	// record as collected, leaving the bytes on disk for ever while the record says otherwise.
+	store, asked := fakeStore(t, http.StatusAccepted)
+	ctx := context.Background()
+
+	image := catalogue.ArtifactStoreScheme + "web/app@sha256:abc123"
+	archive := catalogue.ArtifactStoreScheme + "web/config/blobs/sha256:def456"
+	if err := store.LetGo(ctx, image); err != nil {
+		t.Fatal(err)
+	}
+	if err := store.LetGo(ctx, archive); err != nil {
+		t.Fatal(err)
+	}
+	want := []string{"/v2/web/app/manifests/sha256:abc123", "/v2/web/config/blobs/sha256:def456"}
+	if len(*asked) != 2 || (*asked)[0] != want[0] || (*asked)[1] != want[1] {
+		t.Fatalf("the store was asked %v; want %v", *asked, want)
+	}
+}
+
+func TestAStoreThatDoesNotHaveItAnswersGone(t *testing.T) {
+	// The outcome wanted, already true. Told apart from success only so the sweep can say which
+	// happened; both are recorded, because retrying for ever is the thing to avoid.
+	store, _ := fakeStore(t, http.StatusNotFound)
+	err := store.LetGo(context.Background(), catalogue.ArtifactStoreScheme+"web/app@sha256:abc123")
+	if !errors.Is(err, Gone) {
+		t.Fatalf("a store that does not hold it answered %v, want Gone", err)
+	}
+}
+
+func TestAStoreWithDeletionOffSaysSoAndNamesTheRemedy(t *testing.T) {
+	// The registry answers 405 when it was started without deletion enabled. The remedy is a
+	// setting on the store's module, and saying "405" would send somebody to the wrong place.
+	store, _ := fakeStore(t, http.StatusMethodNotAllowed)
+	err := store.LetGo(context.Background(), catalogue.ArtifactStoreScheme+"web/app@sha256:abc123")
+	if err == nil {
+		t.Fatal("a store that refuses deletion was read as success")
+	}
+	if !strings.Contains(err.Error(), "REGISTRY_STORAGE_DELETE_ENABLED") {
+		t.Fatalf("the refusal does not name the remedy: %v", err)
+	}
+}
+
+func TestAReferenceThatIsNotTheMeshsOwnIsNeverAsked(t *testing.T) {
+	// The whole safety of the sweep is that it names only what the mesh recorded putting there.
+	// A reference of another shape — a vendor's image, a package version — is refused rather
+	// than composed into a delete somewhere that is not the mesh's store.
+	store, asked := fakeStore(t, http.StatusAccepted)
+	for _, reference := range []string{
+		"docker.io/library/registry@sha256:abc123",
+		"registry@sha256:abc123",
+		"1.4.2",
+	} {
+		if err := store.LetGo(context.Background(), reference); err == nil {
+			t.Errorf("%s was asked about; it is not a reference into the mesh's store", reference)
+		}
+	}
+	if len(*asked) != 0 {
+		t.Fatalf("the store was asked about %v", *asked)
+	}
+}

internal/broker/jetstream.go

@@ -1,6 +1,7 @@
 package broker
 
 import (
+	"context"
 	"crypto/sha256"
 	"crypto/tls"
 	"crypto/x509"
@@ -12,6 +13,7 @@ import (
 	"time"
 
 	"github.com/nats-io/nats.go"
+	"github.com/nats-io/nats.go/jetstream"
 )
 
 // The JetStream side of the controller: the one place the mesh's streams and consumers are
@@ -84,6 +86,13 @@ func pinnedTo(path string) (*tls.Config, error) {
 	return PinnedToFingerprint(want), nil
 }
 
+// OnConn is the JetStream handle over a connection the caller already holds — the control plane's
+// link — for asserting what the bus holds without dialling a second time.
+func OnConn(conn *nats.Conn) *JetStream {
+	js, _ := conn.JetStream()
+	return &JetStream{conn: conn, js: js}
+}
+
 // DialPinned is Dial with the server's certificate pinned by a fingerprint the caller already holds
 // — a module or a build machine that was handed one beside its credential, and has no file.
 func DialPinned(url, fingerprint string, opts ...nats.Option) (*JetStream, error) {
@@ -316,3 +325,50 @@ func retentionOf(r Retention) nats.RetentionPolicy {
 		return nats.LimitsPolicy
 	}
 }
+
+// EnsureBucket creates a module's bucket if it is absent and brings its options to match if it is
+// present (novox/hq ADR 0201).
+//
+// **An update, never a delete and recreate**, for the reason a stream is updated: recreating
+// discards what the bucket holds, and what a module's state holds is data. The mesh's caps are
+// asserted with the owner's options, so a bucket made by hand converges to them.
+func (j *JetStream) EnsureBucket(b Bucket) error {
+	history := b.History
+	if history == 0 {
+		history = 1
+	}
+	js, err := jetstream.New(j.conn)
+	if err != nil {
+		return err
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	if _, err := js.CreateOrUpdateKeyValue(ctx, jetstream.KeyValueConfig{
+		Bucket:       b.Bucket(),
+		Description:  b.Why(),
+		History:      uint8(history),
+		TTL:          time.Duration(b.TTLSeconds) * time.Second,
+		MaxValueSize: StateMaxValueBytes,
+		MaxBytes:     StateMaxBytes,
+		Storage:      jetstream.FileStorage,
+	}); err != nil {
+		return fmt.Errorf("asserting bucket %s: %w", b.Bucket(), err)
+	}
+	return nil
+}
+
+// BucketNames is every key-value bucket on the server, the mesh's and anybody else's.
+func (j *JetStream) BucketNames() ([]string, error) {
+	js, err := jetstream.New(j.conn)
+	if err != nil {
+		return nil, err
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+	lister := js.KeyValueStoreNames(ctx)
+	var out []string
+	for name := range lister.Name() {
+		out = append(out, name)
+	}
+	return out, lister.Error()
+}

internal/broker/membership.go

@@ -45,6 +45,11 @@ type Membership struct {
 	// module that must tell the mesh from the world, the route proxy serving an internal name, reads
 	// it here rather than keeping a definition of its own.
 	Mesh []string `json:"mesh,omitempty"`
+	// State is every bucket this module's code may reach, by the name it uses for each, and whether
+	// it may write it (novox/hq ADR 0201): the runtime answers a bundle's state verbs from this list
+	// and refuses, with the reason, what is not on it — the bus enforces only the union over every
+	// module on the machine.
+	State []StateIssued `json:"state,omitempty"`
 }
 
 // Served is one address a tool is answered on.
@@ -103,6 +108,7 @@ func MembershipFor(node string, d Declared, where Placements) Membership {
 			m.Seats = append(m.Seats, SeatServed{Seat: s.Name, Verb: verb, Subject: seatToolSubject(s, verb, node)})
 		}
 	}
+	m.State = stateIssuedFor(d)
 	if len(d.Invokes) > 0 {
 		m.Reaches = map[string][]string{}
 		for _, t := range d.Invokes {

internal/broker/nats.go

@@ -103,6 +103,12 @@ type Principal struct {
 	// permission and nothing beside it.
 	Invokes []string
 
+	// State is the local names of the state this principal's module keeps, and Reads the state of
+	// others it reads as `<module>.<name>` (novox/hq ADR 0201): a bucket each, kept by the owner's
+	// instances and read by whoever declares it.
+	State []string
+	Reads []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
@@ -421,6 +427,10 @@ func PermissionsFor(p Principal) (Permissions, error) {
 			}
 		}
 
+		// 5. Its state, and the state of others it reads (novox/hq ADR 0201): every one read and
+		//    watched, its own written too.
+		pub = append(pub, stateGrants(p.Module, p.State, p.Reads)...)
+
 	case KindNodeTools:
 		// **One process serves what every module on the machine would have served for itself**
 		// (novox/hq ADR 0175). Each carried module's whole tool namespace — the same grant that
@@ -487,6 +497,13 @@ func PermissionsFor(p Principal) (Permissions, error) {
 				"$JS.API.CONSUMER.MSG.NEXT."+stream+"."+durable,
 				"$JS.ACK."+stream+"."+durable+".>")
 		}
+		// **And it keeps and reads state for the modules it carries** (novox/hq ADR 0201): the union
+		// of what each may do with a bucket — an owner's write, a reader's read. That one module's code
+		// does not write another's bucket through it is the runtime's to keep, from the membership
+		// each assignment is issued, as it keeps each module's events under that module's own name.
+		for _, d := range p.Carries {
+			pub = append(pub, stateGrants(d.Module, stateNames(d.State), d.Reads)...)
+		}
 		sub = unique(sub)
 		pub = unique(pub)
 	}

internal/broker/state.go

@@ -0,0 +1,169 @@
+package broker
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// A module's state on the bus (novox/hq ADR 0201, design 32 §4, design 25 §3).
+//
+// A module names the state it keeps (`state`) and the state of others it reads (`reads`), and each
+// is a key-value bucket: the server's own last-per-subject stream with direct reads, delete markers
+// and watches, which is the state relationship the mesh already uses for declarations, opened to
+// modules. The controller creates every bucket from the catalogue — from registration, like a
+// seat's stream, so a reader may watch before the owner runs anywhere — and no module can.
+//
+// Pure, like everything else in this package that decides what the bus holds; jetstream.go is the
+// part that asks a server.
+
+// The mesh's caps on a bucket, the same for every module: a value is a piece of state, not a file,
+// and a bucket that grew without bound would be one module filling the bus's disk for everyone.
+const (
+	StateMaxValueBytes = 256 * 1024
+	StateMaxBytes      = 64 * 1024 * 1024
+)
+
+// A Bucket is one module's declared state as the bus holds it.
+type Bucket struct {
+	Module string
+	Name   string
+	// History is how many values a key keeps; zero is one.
+	History int
+	// TTLSeconds is how long a value lives; zero is until replaced or deleted.
+	TTLSeconds int
+}
+
+// BucketName is the bucket a module's state lives in: the module and the local name joined by an
+// underscore, which neither may contain, so two modules can never derive one bucket.
+func BucketName(module, name string) string { return module + "_" + name }
+
+// Bucket is this bucket's name on the bus.
+func (b Bucket) Bucket() string { return BucketName(b.Module, b.Name) }
+
+// Why is carried into the server's description of the bucket, so somebody reading the server's
+// own state finds whose it is and why it is kept.
+func (b Bucket) Why() string {
+	return fmt.Sprintf("%s's state %q (novox/hq ADR 0201): its current value per key, written by %s, "+
+		"read by whatever declares it reads it; kept when %s is unassigned, because it is data",
+		b.Module, b.Name, b.Module, b.Module)
+}
+
+// bucketOfRead is the bucket a read names, `<module>.<name>`, or false when it names none.
+func bucketOfRead(read string) (string, bool) {
+	at := strings.LastIndex(read, ".")
+	if at <= 0 || at == len(read)-1 {
+		return "", false
+	}
+	module, name := read[:at], read[at+1:]
+	if !safeSubject.MatchString(module) || !safeSubject.MatchString(name) {
+		return "", false
+	}
+	return BucketName(module, name), true
+}
+
+// stateGrants is what a principal publishes to reach the state its modules keep and read: for every
+// bucket, binding to it, reading a key directly, and an ordered consumer for listing and watching,
+// created and deleted on the bucket's own stream, with its flow control answered; for a bucket an
+// owner keeps, writing under the bucket's own subjects too.
+//
+// **Measured against a running server, 2026-10-04** (novox/hq research 024), and each one is there
+// because leaving it out failed: without STREAM.INFO nothing binds; without DIRECT.GET nothing is
+// read; without CONSUMER.CREATE no key is listed and nothing is watched; without CONSUMER.DELETE a
+// watch cannot be stopped and lingers on the server. A write outside these is refused by the server
+// — and reaches the writer as a timeout, not a refusal, which is why the runtime refuses first.
+func stateGrants(module string, keeps []string, reads []string) []string {
+	var out []string
+	read := func(bucket string) {
+		stream := "KV_" + bucket
+		out = append(out,
+			"$JS.API.STREAM.INFO."+stream,
+			"$JS.API.DIRECT.GET."+stream+".>",
+			"$JS.API.CONSUMER.CREATE."+stream+".>",
+			"$JS.API.CONSUMER.DELETE."+stream+".>",
+			"$JS.FC."+stream+".>")
+	}
+	for _, name := range keeps {
+		if !safeSubject.MatchString(name) {
+			continue
+		}
+		bucket := BucketName(module, name)
+		read(bucket)
+		out = append(out, "$KV."+bucket+".>")
+	}
+	for _, r := range reads {
+		if bucket, ok := bucketOfRead(r); ok {
+			read(bucket)
+		}
+	}
+	return out
+}
+
+// StateIssued is one bucket an assignment may reach, by the name its module uses for it: its own
+// state by the local name, another's as `<module>.<name>` (novox/hq ADR 0201).
+type StateIssued struct {
+	Name   string `json:"name"`
+	Bucket string `json:"bucket"`
+	Writes bool   `json:"writes,omitempty"`
+}
+
+// stateIssuedFor is every bucket a module's code may reach, as its membership lists them.
+func stateIssuedFor(d Declared) []StateIssued {
+	var out []StateIssued
+	for _, b := range d.State {
+		out = append(out, StateIssued{Name: b.Name, Bucket: BucketName(d.Module, b.Name), Writes: true})
+	}
+	for _, r := range d.Reads {
+		if bucket, ok := bucketOfRead(r); ok {
+			out = append(out, StateIssued{Name: r, Bucket: bucket})
+		}
+	}
+	return out
+}
+
+// stateNames is the local names of a module's own buckets.
+func stateNames(buckets []Bucket) []string {
+	out := make([]string, 0, len(buckets))
+	for _, b := range buckets {
+		out = append(out, b.Name)
+	}
+	return out
+}
+
+// A BucketAsserter is the part of a JetStream connection bucket assertion needs.
+type BucketAsserter interface {
+	// EnsureBucket creates the bucket if absent and brings its options to match if present, never
+	// discarding what it holds.
+	EnsureBucket(b Bucket) error
+	// BucketNames is every key-value bucket on the server.
+	BucketNames() ([]string, error)
+}
+
+// RaiseBuckets asserts every declared bucket and answers the buckets on the server that nothing
+// declares any more.
+//
+// **Those are reported, never removed** (novox/hq ADR 0201, ADR 0030): what a module stored is
+// data, and a manifest edited, a module renamed or a catalogue entry dropped is an ordinary day's
+// work that must not take data with it. Removing one is a person's act.
+func RaiseBuckets(a BucketAsserter, buckets []Bucket) (undeclared []string, err error) {
+	sorted := append([]Bucket(nil), buckets...)
+	sort.Slice(sorted, func(i, j int) bool { return sorted[i].Bucket() < sorted[j].Bucket() })
+	declared := map[string]bool{}
+	for _, b := range sorted {
+		if err := a.EnsureBucket(b); err != nil {
+			return nil, fmt.Errorf("asserting %s's state %q: %w", b.Module, b.Name, err)
+		}
+		declared[b.Bucket()] = true
+	}
+	names, err := a.BucketNames()
+	if err != nil {
+		return nil, fmt.Errorf("listing the bus's state: %w", err)
+	}
+	for _, n := range names {
+		if !declared[n] {
+			undeclared = append(undeclared, n)
+		}
+	}
+	sort.Strings(undeclared)
+	return undeclared, nil
+}

internal/broker/state_test.go

@@ -0,0 +1,180 @@
+package broker
+
+import (
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/nats-io/nats.go"
+)
+
+// The grants measured against a running server (novox/hq research 024): an owner reads and writes
+// its bucket, a reader only reads, and neither reaches any other bucket.
+func TestAnOwnerWritesItsStateAndAReaderOnlyReads(t *testing.T) {
+	owner, err := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "claude-code",
+		State: []string{"servers"}, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, s := range []string{
+		"$KV.claude-code_servers.>",
+		"$JS.API.STREAM.INFO.KV_claude-code_servers",
+		"$JS.API.DIRECT.GET.KV_claude-code_servers.>",
+		"$JS.API.CONSUMER.CREATE.KV_claude-code_servers.>",
+		"$JS.API.CONSUMER.DELETE.KV_claude-code_servers.>",
+		"$JS.FC.KV_claude-code_servers.>",
+	} {
+		has(t, owner.Publish, s)
+	}
+	hasNot(t, owner.Publish, "$KV.>")
+	hasNot(t, owner.Publish, "$JS.API.>")
+
+	reader, err := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "console",
+		Reads: []string{"claude-code.servers"}, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, reader.Publish, "$JS.API.DIRECT.GET.KV_claude-code_servers.>")
+	has(t, reader.Publish, "$JS.API.CONSUMER.CREATE.KV_claude-code_servers.>")
+	hasNot(t, reader.Publish, "$KV.claude-code_servers.>")
+	for _, s := range reader.Subscribe {
+		if s == "$KV.claude-code_servers.>" {
+			t.Fatalf("a reader subscribes the bucket's subjects directly: %v", reader.Subscribe)
+		}
+	}
+}
+
+// One runtime carries every module on its machine, so its grant is the union: the owner's write
+// where an owner is carried, a read where only a reader is.
+func TestTheRuntimeKeepsAndReadsStateForItsModules(t *testing.T) {
+	perms, err := PermissionsFor(Principal{Kind: KindNodeTools, Node: "one", Module: RuntimeModule,
+		Carries: []Declared{
+			{Module: "claude-code", State: []Bucket{{Module: "claude-code", Name: "servers"}},
+				Reads: []string{"licence-manager.bindings"}},
+			{Module: "audit"},
+		}, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	has(t, perms.Publish, "$KV.claude-code_servers.>")
+	has(t, perms.Publish, "$JS.API.DIRECT.GET.KV_licence-manager_bindings.>")
+	hasNot(t, perms.Publish, "$KV.licence-manager_bindings.>")
+}
+
+// A module with no state is granted nothing of any bucket — the composition of every module that
+// existed before this is unchanged.
+func TestAModuleWithNoStateReachesNoBucket(t *testing.T) {
+	perms, err := PermissionsFor(Principal{Kind: KindModule, Node: "one", Module: "billing",
+		Emits: []string{"order.placed"}, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, s := range perms.Publish {
+		if strings.HasPrefix(s, "$KV.") || strings.HasPrefix(s, "$JS.FC.") || strings.Contains(s, ".KV_") {
+			t.Fatalf("granted %q without declaring state", s)
+		}
+	}
+}
+
+// A read that names no bucket grants nothing rather than something that happens to parse.
+func TestAReadThatNamesNoBucketGrantsNothing(t *testing.T) {
+	if got := stateGrants("a", nil, []string{"nodot", "x.", ".y", "a.b>"}); len(got) != 0 {
+		t.Fatalf("granted %v for reads that name no bucket", got)
+	}
+}
+
+// The membership lists every bucket the module's code may reach, by the name the module uses for
+// it, and whether it may write it — the list the runtime refuses from.
+func TestAMembershipListsTheStateItsModuleMayReach(t *testing.T) {
+	m := MembershipFor("one", Declared{Module: "claude-code",
+		State: []Bucket{{Module: "claude-code", Name: "servers"}},
+		Reads: []string{"licence-manager.bindings"}}, Placements{})
+	want := []StateIssued{
+		{Name: "servers", Bucket: "claude-code_servers", Writes: true},
+		{Name: "licence-manager.bindings", Bucket: "licence-manager_bindings"},
+	}
+	if !slices.Equal(m.State, want) {
+		t.Fatalf("issued %+v, want %+v", m.State, want)
+	}
+	if none := MembershipFor("one", Declared{Module: "audit"}, Placements{}); none.State != nil {
+		t.Fatalf("a module with no state was issued %+v", none.State)
+	}
+}
+
+type buckets struct {
+	ensured []string
+	on      []string
+}
+
+func (b *buckets) EnsureBucket(x Bucket) error {
+	b.ensured = append(b.ensured, x.Bucket())
+	return nil
+}
+func (b *buckets) BucketNames() ([]string, error) { return b.on, nil }
+
+// Every declared bucket is asserted; one on the server that nothing declares is said, not removed.
+func TestRaisingStateReportsWhatNothingDeclares(t *testing.T) {
+	b := &buckets{on: []string{"claude-code_servers", "gone_old", "ours_by_hand"}}
+	undeclared, err := RaiseBuckets(b, []Bucket{{Module: "claude-code", Name: "servers"}, {Module: "a", Name: "b"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !slices.Equal(b.ensured, []string{"a_b", "claude-code_servers"}) {
+		t.Fatalf("asserted %v", b.ensured)
+	}
+	if !slices.Equal(undeclared, []string{"gone_old", "ours_by_hand"}) {
+		t.Fatalf("reported %v", undeclared)
+	}
+}
+
+// Against a real server: a bucket is created with the owner's options and the mesh's caps,
+// asserting it again changes nothing and keeps what it holds, and a changed option is brought to
+// match in place.
+func TestABucketIsAssertedInPlace(t *testing.T) {
+	js := aLiveBus(t)
+	b := Bucket{Module: "statetest", Name: "servers"}
+	if _, err := RaiseBuckets(js, []Bucket{b}); err != nil {
+		t.Fatalf("a real server refused a module's bucket: %v", err)
+	}
+	kv, err := js.Context().KeyValue(b.Bucket())
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := kv.Put("all.one", []byte(`{"kept":true}`)); err != nil {
+		t.Fatal(err)
+	}
+	b.History = 3
+	if _, err := RaiseBuckets(js, []Bucket{b}); err != nil {
+		t.Fatalf("asserting the bucket again failed, so a restart would: %v", err)
+	}
+	got, err := kv.Get("all.one")
+	if err != nil || string(got.Value()) != `{"kept":true}` {
+		t.Fatalf("asserting again lost what the bucket held: %v %v", got, err)
+	}
+	status, err := kv.Status()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if status.History() != 3 {
+		t.Fatalf("history is %d, the owner declared 3", status.History())
+	}
+	if s, ok := status.(*nats.KeyValueBucketStatus); ok {
+		if c := s.StreamInfo().Config; c.MaxMsgSize != StateMaxValueBytes || c.MaxBytes != StateMaxBytes {
+			t.Fatalf("the mesh's caps are not on the bucket: value %d, bucket %d", c.MaxMsgSize, c.MaxBytes)
+		}
+	}
+}
+
+// Against a real server: the handle over a connection the control plane already holds asserts a
+// bucket as Dial's does — what a push uses, so a module registered since the last start has its
+// bucket before its membership names it.
+func TestABucketIsAssertedOverAHeldConnection(t *testing.T) {
+	js := aLiveBus(t)
+	held := OnConn(js.Conn())
+	if _, err := RaiseBuckets(held, []Bucket{{Module: "statetest", Name: "held"}}); err != nil {
+		t.Fatalf("asserting over a held connection failed: %v", err)
+	}
+	if _, err := js.Context().KeyValue("statetest_held"); err != nil {
+		t.Fatalf("the bucket is not there: %v", err)
+	}
+}

internal/broker/users.go

@@ -33,6 +33,10 @@ type Declared struct {
 	Watches []Seat
 	// Invokes are the tools it calls, `<module>.<tool>` or `*` (novox/hq ADR 0152).
 	Invokes []string
+	// State is the state it keeps, each a bucket its instances write (novox/hq ADR 0201).
+	State []Bucket
+	// Reads are other modules' state it reads, each `<module>.<name>` (novox/hq ADR 0201).
+	Reads []string
 }
 
 // Records is what composing a user list needs to know about the mesh, and nothing more.
@@ -81,6 +85,7 @@ func Users(r Records) ([]Principal, error) {
 				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, Invokes: d.Invokes,
+				State: stateNames(d.State), Reads: d.Reads,
 			})
 		}
 		if runtimeHere {

internal/builder/builder.go

@@ -215,7 +215,7 @@ func Build(ctx context.Context, run Runner, publish Publisher,
 		sort.Slice(artifacts, func(i, j int) bool { return artifacts[i].Name < artifacts[j].Name })
 		for _, a := range artifacts {
 			say("artifact", "%s (%s%s) — starting", a.Name, a.Kind, langSuffix(a))
-			made, err := one(ctx, run, publish, manifest.Module, within, workspace, commit, credentials, a, args, held, npmrcPath, seatBases, say)
+			made, err := one(ctx, run, publish, manifest.Module, within, workspace, commit, credentials, a, args, held, npmrcPath, npmrc, seatBases, say)
 			if err != nil {
 				say("artifact", "%s FAILED: %v", a.Name, err)
 				return Result{}, err
@@ -443,7 +443,7 @@ func wantsPackages(manifest catalogue.Manifest, within string) bool {
 
 func one(ctx context.Context, run Runner, publish Publisher,
 	module, tree, workspace, commit, credentials string, a catalogue.Artifact, args []string,
-	held map[string]string, npmrc string, seats map[string]string,
+	held map[string]string, npmrc string, registry Npmrc, seats map[string]string,
 	say func(step, format string, args ...any)) (catalogue.Built, error) {
 
 	switch a.Kind {
@@ -582,6 +582,11 @@ func one(ctx context.Context, run Runner, publish Publisher,
 					"holds no copy of it. Build %s first",
 				module, a.Name, chain.Language, chain.Base, chain.Artifact, chain.Base)
 		}
+		// The module's own packages first, where the compiler and the bundler resolve them from
+		// (dependencies.go); nothing at all for a module whose package.json names only the SDK.
+		if err := installOwn(ctx, run, tree, chain, base, registry, say); err != nil {
+			return catalogue.Built{}, fmt.Errorf("%s: %s: %w", module, a.Name, err)
+		}
 		say("bundle", "compiling %s in %s's toolchain", a.Language, chain.Base)
 		compiled, err := compile(ctx, run, tree, chain, base, a)
 		if err != nil {

internal/builder/dependencies.go

@@ -0,0 +1,147 @@
+package builder
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"sort"
+	"strings"
+)
+
+// A module's own packages, installed before its bundle is compiled, so the bundler inlines them.
+//
+// **A bundle could only import what the toolchain happened to carry.** The compiler and the bundler
+// resolve an import by walking up from the module's source: the module's own directory first, then
+// the toolchain image's node_modules. Nothing ever put anything in the first, so a module needing a
+// database driver (`pg`, `mongodb`, `mssql`) could not be a bundle at all, and kept a container whose
+// recipe installed it by hand (novox/hq ADR 0198 §4: "the backend's own driver inside the bundle").
+// Now the module's `package.json` says what it depends on, as any Node package does, and the build
+// installs exactly that into the module's own directory before compiling.
+//
+// **The SDK the toolchain carries is the one a bundle is built with, whatever the module says**
+// (novox/hq issue 212: the toolchain is rebuilt on every SDK release and every bundle after it). A
+// module's `package.json` names `@novox/mesh-sdk` with a range — it has to, to type-check on a
+// workstation — and installing that range would shadow the toolchain's copy for this module alone:
+// one module compiled against an older SDK than its neighbours, chosen by a caret nobody re-reads.
+// So the SDK is taken out of what is installed (and never fetched), and any copy something else
+// pulls in is removed afterwards; every import of it resolves past the module's node_modules to the
+// toolchain's. A module therefore cannot pin a different SDK, by design: the toolchain is the pin.
+//
+// **Correctness before speed.** Every build installs afresh into a fresh clone, from the lockfile
+// when the module has one (`npm ci`, exact) and from its ranges otherwise; nothing installed is kept
+// between builds. What is shared is npm's own download cache, a named volume, which is
+// content-addressed and verified by integrity on every read — it saves the network, never the
+// install. Install scripts do not run: the build node runs nobody's postinstall, and what a script
+// would build natively could not be inlined into one file anyway.
+
+// sdkPackage is the package a TypeScript bundle's launcher serves through, and the one package a
+// module's own dependencies never supply (above).
+const sdkPackage = "@novox/mesh-sdk"
+
+// npmCache is the named volume npm's download cache lives in across builds on one build node.
+const npmCache = "mesh-builder-npm-cache"
+
+// ownDependencies is what a module's package.json depends on beyond the SDK, sorted; nothing when
+// the module has no package.json or depends on nothing else — which builds exactly as before.
+func ownDependencies(tree string) ([]string, error) {
+	raw, err := os.ReadFile(filepath.Join(tree, "package.json"))
+	if errors.Is(err, os.ErrNotExist) {
+		return nil, nil
+	}
+	if err != nil {
+		return nil, err
+	}
+	var p struct {
+		Dependencies map[string]string `json:"dependencies"`
+	}
+	if err := json.Unmarshal(raw, &p); err != nil {
+		return nil, fmt.Errorf("the module's package.json is not JSON: %w", err)
+	}
+	var names []string
+	for name := range p.Dependencies {
+		if name != sdkPackage {
+			names = append(names, name)
+		}
+	}
+	sort.Strings(names)
+	return names, nil
+}
+
+// installSteps is the script run inside the toolchain image, from the module's own directory ($0).
+// It works in a scratch copy so the module's package.json and lockfile are never rewritten, takes
+// the SDK out of what is installed, installs production dependencies only, removes any copy of the
+// SDK something pulled in, and puts the result at the module's node_modules.
+const installSteps = `set -e
+work="$(mktemp -d)"
+cp "$0/package.json" "$work/"
+if [ -f "$0/package-lock.json" ]; then cp "$0/package-lock.json" "$work/"; fi
+cd "$work"
+node -e '
+const fs = require("fs"), sdk = process.argv[1];
+const p = JSON.parse(fs.readFileSync("package.json", "utf8"));
+for (const k of ["dependencies", "peerDependencies", "optionalDependencies"]) if (p[k]) delete p[k][sdk];
+delete p.devDependencies; delete p.scripts;
+fs.writeFileSync("package.json", JSON.stringify(p));
+' "$1"
+shift
+if [ -f package-lock.json ]; then
+  npm ci --omit=dev --omit=peer --ignore-scripts --no-audit --no-fund "$@"
+else
+  npm install --omit=dev --omit=peer --ignore-scripts --no-audit --no-fund --no-package-lock "$@"
+fi
+find node_modules -depth -type d -path "*/node_modules/@novox/mesh-sdk" -exec rm -rf {} +
+rm -rf "$0/node_modules"
+cp -a node_modules "$0/node_modules"
+`
+
+// installOwn installs a TypeScript module's own production dependencies into its directory, in the
+// toolchain image, before the compile — or does nothing at all for a module that has none.
+func installOwn(ctx context.Context, run Runner, tree string, chain Toolchain, base string,
+	registry Npmrc, say func(step, format string, args ...any)) error {
+	if chain.Language != "typescript" {
+		return nil
+	}
+	deps, err := ownDependencies(tree)
+	if err != nil || len(deps) == 0 {
+		return err
+	}
+	scoped := strings.TrimSpace(registry.Scope)
+	if !registry.Enabled() {
+		// **No registry, no scoped package.** Without the mesh's registry a scoped name resolves on
+		// the public one, where anybody may have published it: a dependency that installs is not
+		// the dependency the module meant.
+		for _, d := range deps {
+			if strings.HasPrefix(d, "@novox/") {
+				return fmt.Errorf("the module depends on %s, and this build knows no package registry "+
+					"for its scope; it would resolve from the public registry, which is not where the "+
+					"mesh publishes it", d)
+			}
+		}
+	}
+	const within = "/app/modules/module"
+	invocation := []string{"run", "--rm",
+		"--volume", tree + ":" + within,
+		"--volume", npmCache + ":/root/.npm",
+		"--workdir", within}
+	var flags []string
+	if registry.Enabled() {
+		// The registry is reached where the binding says it is, which may be this machine's own
+		// loopback — the reason an image build that resolves packages runs on the host network too.
+		invocation = append(invocation, "--network", "host")
+		reg := strings.TrimSpace(registry.Registry)
+		if !strings.HasSuffix(reg, "/") {
+			reg += "/"
+		}
+		flags = append(flags, "--"+scoped+":registry="+reg)
+	}
+	invocation = append(invocation, base, "sh", "-c", installSteps, within, sdkPackage)
+	invocation = append(invocation, flags...)
+	say("bundle", "installing the module's own packages: %s", strings.Join(deps, ", "))
+	if _, err := run(ctx, tree, "docker", invocation...); err != nil {
+		return fmt.Errorf("installing the module's own packages (%s): %w", strings.Join(deps, ", "), err)
+	}
+	return nil
+}

internal/builder/dependencies_test.go

@@ -0,0 +1,131 @@
+package builder
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+// A module's own packages (dependencies.go): installed into its own directory, in the toolchain,
+// before the compile, so the bundler inlines them — the SDK always the toolchain's.
+
+func buildWithPackageJSON(t *testing.T, pkg string, extra map[string]string, registry Npmrc) (*recorded, error) {
+	t.Helper()
+	files := map[string]string{"index.ts": "console.log(1)"}
+	if pkg != "" {
+		files["package.json"] = pkg
+	}
+	for k, v := range extra {
+		files[k] = v
+	}
+	r, workspace := aRepository(t, aBundle, files)
+	held := map[string]string{"mesh-tools/build": "registry.invalid/mesh-tools/build@sha256:" + strings.Repeat("b", 64)}
+	_, err := Build(context.Background(), compiling{r}.run, r,
+		"https://forge.invalid/greeter.git", "", "", workspace, held, registry, GitCredential{}, nil)
+	return r, err
+}
+
+func installs(r *recorded) []string {
+	var out []string
+	for _, line := range r.ran {
+		if strings.HasPrefix(line, "docker run") && strings.Contains(line, "npm ci") {
+			out = append(out, line)
+		}
+	}
+	return out
+}
+
+func compileIndex(r *recorded) int {
+	for i, line := range r.ran {
+		if strings.Contains(line, "--outDir") {
+			return i
+		}
+	}
+	return -1
+}
+
+func TestAModulesOwnPackagesAreInstalledInTheToolchainBeforeTheCompile(t *testing.T) {
+	r, err := buildWithPackageJSON(t, `{"type":"module","dependencies":{"@novox/mesh-sdk":"^0.1.0","pg":"^8"},"devDependencies":{"typescript":"^5"}}`,
+		nil, Npmrc{Scope: "@novox", Registry: "https://forge.invalid/api/packages/novox/npm"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	got := installs(r)
+	if len(got) != 1 {
+		t.Fatalf("want one install of the module's own packages:\n%s", strings.Join(r.ran, "\n"))
+	}
+	line := got[0]
+	for _, want := range []string{
+		"mesh-tools/build@sha256:",       // in the toolchain image
+		":/app/modules/module",           // into the module's own directory
+		"--workdir /app/modules/module",  //
+		npmCache + ":/root/.npm",         // npm's verified download cache, and only that
+		"--omit=dev", "--ignore-scripts", // production packages, no build-node scripts
+		"npm ci", "npm install", "--no-package-lock", // the lockfile when there is one, else the ranges
+		"--@novox:registry=https://forge.invalid/api/packages/novox/npm/", // the scope from the mesh's registry
+		"--network host",
+		"@novox/mesh-sdk", // named, to be taken out of what is installed
+	} {
+		if !strings.Contains(line, want) {
+			t.Errorf("the install lacks %q:\n%s", want, line)
+		}
+	}
+	// The SDK is the toolchain's: never installed from the module's range, and any copy removed.
+	if !strings.Contains(line, `delete p[k][sdk]`) || !strings.Contains(line, `-path "*/node_modules/@novox/mesh-sdk" -exec rm -rf`) {
+		t.Errorf("the module's own SDK range could shadow the toolchain's SDK:\n%s", line)
+	}
+	if i, c := strings.Index(strings.Join(r.ran, "\n"), "npm ci"), compileIndex(r); c < 0 ||
+		i > strings.Index(strings.Join(r.ran, "\n"), "--outDir") {
+		t.Fatalf("the install did not run before the compile:\n%s", strings.Join(r.ran, "\n"))
+	}
+}
+
+// **A module with nothing beyond the SDK builds exactly as before**: the same commands, no install.
+func TestAModuleDependingOnlyOnTheSDKBuildsExactlyAsBefore(t *testing.T) {
+	without, err := buildWithPackageJSON(t, "", nil, Npmrc{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, pkg := range []string{
+		`{"type":"module","dependencies":{"@novox/mesh-sdk":"^0.1.0"},"devDependencies":{"typescript":"^5"}}`,
+		`{"type":"module"}`,
+	} {
+		with, err := buildWithPackageJSON(t, pkg, map[string]string{"package-lock.json": "{}"}, Npmrc{Scope: "@novox", Registry: "https://forge.invalid/npm/"})
+		if err != nil {
+			t.Fatal(err)
+		}
+		if strings.Contains(strings.Join(with.ran, "\n"), "npm ") {
+			t.Fatalf("a module depending on nothing but the SDK ran npm:\n%s", strings.Join(with.ran, "\n"))
+		}
+		if len(with.ran) != len(without.ran) {
+			t.Fatalf("a module depending only on the SDK built differently from one with no package.json:\n%s\n---\n%s",
+				strings.Join(with.ran, "\n"), strings.Join(without.ran, "\n"))
+		}
+	}
+}
+
+// Without the mesh's registry a scoped package would resolve on the public one: refused by name.
+func TestAScopedPackageWithNoRegistryIsRefused(t *testing.T) {
+	r, err := buildWithPackageJSON(t, `{"dependencies":{"@novox/mesh-sdk":"^0.1.0","@novox/other":"^1"}}`, nil, Npmrc{})
+	if err == nil || !strings.Contains(err.Error(), "@novox/other") {
+		t.Fatalf("a scoped package was installed with no registry for its scope: %v", err)
+	}
+	if strings.Contains(strings.Join(r.ran, "\n"), "--outDir") {
+		t.Fatal("the compile ran after the refusal")
+	}
+	// A public package installs without one, from the public registry and nothing else.
+	r, err = buildWithPackageJSON(t, `{"dependencies":{"mssql":"^11"}}`, nil, Npmrc{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := installs(r); len(got) != 1 || strings.Contains(got[0], ":registry=") || strings.Contains(got[0], "--network host") {
+		t.Fatalf("a public package's install: %v", got)
+	}
+}
+
+func TestAnUnreadablePackageJSONIsRefusedByName(t *testing.T) {
+	_, err := buildWithPackageJSON(t, `{"dependencies":`, nil, Npmrc{})
+	if err == nil || !strings.Contains(err.Error(), "package.json") {
+		t.Fatalf("a broken package.json was not refused by name: %v", err)
+	}
+}

internal/builder/toolchain.go

@@ -67,8 +67,9 @@ type Toolchain struct {
 	// `package.json` saying `"type": "module"` — Node reads a bare `.js` as CommonJS otherwise, so a
 	// bundle with its dependencies and without that line still fails to start — and the pruned,
 	// production-only node_modules the runtime itself ships with: the SDK's and the runtime's
-	// dependencies, and nothing module-specific yet (novox/hq ADR 0188 §5: a skeleton; a module's
+	// dependencies, and nothing module-specific (a module's own npm dependencies are installed into
-	// own npm dependencies are a later step). Empty for a language whose bundle carries its own —
+	// its own directory before the compile and inlined by the bundler: dependencies.go). Empty for a
+	// language whose bundle carries its own —
 	// a Go binary is static, a Python bundle is installed with its dependencies.
 	//
 	// A toolchain image without the directory fails the build by name rather than packing a bundle

internal/catalogue/bound_into_files.go

@@ -46,7 +46,7 @@ func boundUsed(content string) [][2]string {
 // Three facts the mesh states about any provision, plus whatever the provider said it serves. A
 // module may not reach a binding it does not have — the same boundary as a secret, for the same
 // reason.
-func knownFor(m Manifest, needs []Needed, node string) map[string]map[string]string {
+func knownFor(m Manifest, needs []Needed, node string) (map[string]map[string]string, error) {
 	out := map[string]map[string]string{}
 	for _, want := range m.Wants() {
 		for i := range needs {
@@ -54,12 +54,20 @@ func knownFor(m Manifest, needs []Needed, node string) map[string]map[string]str
 			if n.Name != want || n.For != m.Module {
 				continue
 			}
+			as := ConsumerIdentity(node, IdentitySource(m.Slug, m.Module))
 			values := map[string]string{
 				"at":   n.At,
 				"from": n.From,
-				"as":   ConsumerIdentity(node, IdentitySource(m.Slug, m.Module)),
+				"as":   as,
 			}
-			for key, value := range n.Serves {
+			// What the provider derives for this consumer rather than for all of them
+			// (novox/hq ADR 0201). Filled here, the one place a provision and the module
+			// requiring it are both in hand.
+			served, err := ServedTo(n.Serves, as)
+			if err != nil {
+				return nil, fmt.Errorf("%s requires %s: %w", m.Module, want, err)
+			}
+			for key, value := range served {
 				// The provider's own vocabulary. Rendered plainly: a port is 5432, not 5432.000000,
 				// which is what a float would write and what a connection string would refuse.
 				values[key] = plainly(value)
@@ -67,7 +75,7 @@ func knownFor(m Manifest, needs []Needed, node string) map[string]map[string]str
 			out[want] = values
 		}
 	}
-	return out
+	return out, nil
 }
 
 // withOwnNames adds a module's own composed names to what it may name from one binding:

internal/catalogue/consumer_into_serves.go

@@ -0,0 +1,311 @@
+package catalogue
+
+import (
+	"fmt"
+	"regexp"
+	"sort"
+	"strings"
+)
+
+// What a provider derives for one consumer, said once in the provider's definition and delivered
+// to both ends (novox/hq ADR 0201, issue 124).
+//
+// A `serves` block is otherwise literal: the same values for every consumer. Where the provider
+// *names the resource* — a bucket, a database, a vhost — the name is derived from who is asking,
+// and before this the mesh had no channel for it. The provider recomputed it in its own code and
+// every consumer transcribed it into its own definition by hand, which is a copy of somebody
+// else's rule kept in agreement by nobody. One of three transcriptions was wrong for months.
+//
+// **The mesh learns no protocol here; it spells its own name in an alphabet it already knows.**
+// The only fact a served value may name is the identity the mesh itself minted for the consumer,
+// in one of two alphabets: as it was minted, and as a DNS label. Everything a provider wants
+// around it — a prefix, a suffix, a separator — it writes around the placeholder, because a
+// served value is a string.
+
+// consumerFact is `${consumer:<fact>}` or `${consumer:<fact>:<alphabet>}`.
+var consumerFact = regexp.MustCompile(`\$\{consumer:([a-z][a-z0-9-]*)(?::([a-z][a-z0-9-]*))?\}`)
+
+// consumerFacts are what a served value may name about the consumer it is being derived for.
+// One entry, deliberately: the identity is the one thing about a consumer the mesh itself chose,
+// so it is the one thing the mesh can hand to a provider without either end guessing.
+var consumerFacts = []string{"as"}
+
+// consumerAlphabets are the ways the mesh will write that identity. `dns` is the mesh's own
+// identifier with its separator written `-` instead of `_` — the whole of the difference between
+// the alphabet the mesh mints in and the one buckets, vhosts and hostnames accept.
+var consumerAlphabets = []string{"dns"}
+
+// ServedTo fills a provider's served values for one consumer.
+//
+// `as` is the identity the mesh minted for that consumer — the same string it is told to present
+// as a login. Values with no placeholder are returned exactly as they were, and a block with no
+// placeholder at all is returned unchanged, so this costs nothing for the providers that derive
+// nothing.
+//
+// Only strings carry placeholders. A number, a boolean or a nested object is a value the provider
+// stated outright, and is left alone.
+func ServedTo(serves map[string]any, as string) (map[string]any, error) {
+	if len(serves) == 0 {
+		return serves, nil
+	}
+	var out map[string]any
+	for _, key := range sortedAnyKeys(serves) {
+		text, ok := serves[key].(string)
+		if !ok || !strings.Contains(text, "${consumer:") {
+			continue
+		}
+		filled, err := consumerInto(text, as)
+		if err != nil {
+			return nil, fmt.Errorf("the value served as %q: %w", key, err)
+		}
+		if out == nil {
+			// Copied only once something actually changes: the caller's map is the manifest's,
+			// and a provider that derives nothing must not have it rewritten underneath it.
+			out = make(map[string]any, len(serves))
+			for k, v := range serves {
+				out[k] = v
+			}
+		}
+		out[key] = filled
+	}
+	if out == nil {
+		return serves, nil
+	}
+	return out, nil
+}
+
+// consumerInto replaces every `${consumer:…}` in one value.
+//
+// **A fact or an alphabet the mesh does not have is refused, not left standing.** Written through,
+// the literal `${consumer:as}` would reach a configuration file and be read as a bucket name,
+// failing somewhere that names neither the module nor the mesh — the same reasoning `${bound:…}`
+// is refused by (boundInto).
+func consumerInto(value, as string) (string, error) {
+	var failed error
+	out := consumerFact.ReplaceAllStringFunc(value, func(match string) string {
+		parts := consumerFact.FindStringSubmatch(match)
+		fact, alphabet := parts[1], parts[2]
+		if fact != "as" {
+			if failed == nil {
+				failed = fmt.Errorf(
+					"says %s, and the mesh states %s about a consumer", match, orNothing(consumerFacts))
+			}
+			return match
+		}
+		switch alphabet {
+		case "":
+			return as
+		case "dns":
+			return asDNSLabel(as)
+		default:
+			if failed == nil {
+				failed = fmt.Errorf(
+					"says %s, and the mesh writes an identity as %s", match, orNothing(consumerAlphabets))
+			}
+			return match
+		}
+	})
+	if failed != nil {
+		return "", failed
+	}
+	return out, nil
+}
+
+// asDNSLabel writes a minted identity as a DNS label.
+//
+// The mesh's identities are already lower-case letters, digits and `_` (ConsumerIdentity), and
+// already short enough for the tightest backend they reach (CheckIdentity, twenty characters). So
+// this is the separator and nothing else — no lower-casing of what is already lower case, no
+// truncation to a limit the identity is already inside, no padding of a name that is already long
+// enough. Each of those would be the mesh guessing at a rule it has not been given.
+func asDNSLabel(as string) string {
+	return strings.ReplaceAll(as, "_", "-")
+}
+
+// CheckServes refuses a `serves` block that names a consumer fact or an alphabet the mesh does not
+// have, when the definition is parsed rather than when a consumer is resolved.
+//
+// A provision nobody consumes yet still has its rule read: a definition that would be refused the
+// first time somebody required it is a definition that is wrong now.
+func CheckServes(m Manifest) []string {
+	var problems []string
+	for _, provision := range sortedServes(m.Serves) {
+		for _, key := range sortedAnyKeys(m.Serves[provision]) {
+			text, ok := m.Serves[provision][key].(string)
+			if !ok {
+				continue
+			}
+			// A probe identity, because what is checked is the shape of the statement and not
+			// what any consumer is called.
+			if _, err := consumerInto(text, "mesh_node_module"); err != nil {
+				problems = append(problems, fmt.Sprintf(
+					"%s serves %s, and the value it serves as %q %s", m.Module, provision, key, err))
+			}
+		}
+	}
+	return problems
+}
+
+func sortedServes(serves map[string]map[string]any) []string {
+	out := make([]string, 0, len(serves))
+	for k := range serves {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
+
+func sortedAnyKeys(values map[string]any) []string {
+	out := make([]string, 0, len(values))
+	for k := range values {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
+
+// derivedFor is what the provider on this machine derives for one consumer of one provision
+// (novox/hq ADR 0201).
+//
+// Settled first, then derived: an operator may set a prefix on what the provider serves and the
+// mesh still fills the consumer's half of it ([ADR 0174]). Only the keys that actually name the
+// consumer are returned — the rest of a `serves` block is the same for every consumer and is
+// already in the provider's own definition, so repeating it here would be a second copy to go
+// stale.
+//
+// The first module in the resolved order that says it serves the provision answers, which is the
+// choice servedOnThisMachine makes for the consumer's half. Nothing serving it on this machine is
+// not an error: a contribution can reach a machine whose provider is a record or an adapter, and
+// then there is nothing derived to tell.
+func (r Resolution) derivedFor(provision, as, consumer, local string, settings SettingsBy) (map[string]any, error) {
+	for _, m := range r.Modules {
+		serves, said := m.Serves[provision]
+		if !said {
+			continue
+		}
+		var names map[string]any
+		for key, value := range serves {
+			if text, ok := value.(string); ok && strings.Contains(text, "${consumer:") {
+				if names == nil {
+					names = map[string]any{}
+				}
+				names[key] = value
+			}
+		}
+		if names == nil {
+			return nil, nil
+		}
+		// **A consumer that keeps several holders of this provision is refused** — this is issue
+		// 124's own failure one case to the side, and it would be just as quiet.
+		//
+		// Each holder gets its own login, `…_<local>` (ADR 0094), and a provider derives from the
+		// login, so it would make one resource per holder. The consumer's side has no such
+		// dimension: one binding file per provision, one `${bound:<provision>:<key>}`, both
+		// derived from the un-suffixed identity. So the provider would create the holder's
+		// resource and the consumer would be configured against a name nothing made — it would
+		// authenticate successfully and be refused on every object, which reads like a credential
+		// fault and is not one.
+		//
+		// Lifting this means giving the consumer's side a local dimension. That is a decision,
+		// not an omission, and until it is taken the mesh says so rather than guessing.
+		if local != "" {
+			return nil, fmt.Errorf(
+				"%s keeps several holders of %s (this one is %q), and %s derives %s for each "+
+					"consumer from the login the mesh minted. Each holder has its own login, and a "+
+					"consumer is told one value per requirement — so the two ends would name "+
+					"different things and nothing would compare them (novox/hq ADR 0201)",
+				consumer, local, provision, m.Module, orNothing(sortedAnyKeys(names)))
+		}
+		settled, err := Settle(names, settings[m.Module])
+		if err != nil {
+			return nil, fmt.Errorf("%s serving %s: %w", m.Module, provision, err)
+		}
+		derived, err := ServedTo(settled, as)
+		if err != nil {
+			return nil, fmt.Errorf("%s serving %s to %s: %w", m.Module, provision, as, err)
+		}
+		return derived, nil
+	}
+	return nil, nil
+}
+
+// notTranscribed refuses a consumer's file that writes out the value its provider derives for it,
+// instead of asking for it (novox/hq ADR 0201, issue 124).
+//
+// **What would have caught the one wrong instance.** The object store's three consumers each wrote
+// their bucket into their own configuration by hand. One of them named a predecessor's bucket, and
+// nothing compared it to what the provider would actually create: the module would have
+// authenticated successfully and been refused on every object, which reads like a credential fault
+// and is not one. It looked authoritative for months.
+//
+// The test is exact and costs one string search: a definition whose file already contains the
+// value the mesh is about to derive for it has written down somebody else's rule. It cannot be a
+// coincidence — a derived value carries the identity the mesh minted for this very consumer on
+// this very machine, which nothing else would spell out — and it cannot be checked afterwards,
+// because after substitution every consumer's file contains it legitimately.
+//
+// Only values that actually name the consumer are judged. A provider that serves a constant under
+// the same key serves the same constant to everyone, and a consumer repeating it is redundant
+// rather than wrong.
+func notTranscribed(resource map[string]any, known map[string]map[string]string, module string) error {
+	if fmt.Sprint(resource["type"]) != "file" {
+		return nil
+	}
+	content, ok := resource["content"].(string)
+	if !ok || content == "" {
+		return nil
+	}
+	for _, provision := range sortedKnown(known) {
+		values := known[provision]
+		identity := values["as"]
+		if identity == "" {
+			continue
+		}
+		for _, key := range sortedStringKeys(values) {
+			if key == "as" {
+				// The login is not derived from itself, and a consumer that must present it in a
+				// connection string legitimately has it from `${bound:…}` — which is what it will
+				// be after substitution, so this would judge the substitution, not the module.
+				continue
+			}
+			value := values[key]
+			if value == "" || !namesTheConsumer(value, identity) {
+				continue
+			}
+			if !strings.Contains(content, value) {
+				continue
+			}
+			return fmt.Errorf(
+				"%s writes %q into %v, and that is exactly what %s derives for it — a definition "+
+					"keeping its own copy of somebody else's naming rule is one that can disagree "+
+					"with it, silently. Say ${bound:%s:%s} and be told",
+				module, value, resource["id"], provision, provision, key)
+		}
+	}
+	return nil
+}
+
+// namesTheConsumer is whether a derived value was built from this consumer's identity — in the
+// alphabet it was minted in, or as a DNS label. A value that does not contain it was not derived
+// from it, whatever else it may be.
+func namesTheConsumer(value, identity string) bool {
+	return strings.Contains(value, identity) || strings.Contains(value, asDNSLabel(identity))
+}
+
+func sortedKnown(known map[string]map[string]string) []string {
+	out := make([]string, 0, len(known))
+	for k := range known {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}
+
+func sortedStringKeys(values map[string]string) []string {
+	out := make([]string, 0, len(values))
+	for k := range values {
+		out = append(out, k)
+	}
+	sort.Strings(out)
+	return out
+}

internal/catalogue/controller_is_a_process_test.go

@@ -0,0 +1,131 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"strings"
+	"testing"
+)
+
+// novox/hq issue 213: the controller is a Go program and was the one piece of the mesh's own Go
+// code still shipped as an image (ADR 0188 §1). Its own manifest, composed for the machine that runs
+// it, is a Go bundle run by the host as a process — and no container.
+func TestTheControllerIsAProcessAndNoContainer(t *testing.T) {
+	raw, err := os.ReadFile("../../module.json")
+	if err != nil {
+		t.Fatal(err)
+	}
+	m, err := ParseManifest(raw)
+	if err != nil {
+		t.Fatalf("the controller's own manifest does not parse:\n%v", err)
+	}
+	if m.Build == nil || len(m.Build.Artifacts) != 1 {
+		t.Fatalf("the controller builds %+v; it is one bundle", m.Build)
+	}
+	a := m.Build.Artifacts[0]
+	if a.Kind != ArtifactBundle || a.Language != "go" || a.System == "" || BinaryOf(a) != "mesh-controller" {
+		t.Fatalf("the controller's artifact is %+v, not a Go bundle naming its system and binary", a)
+	}
+	for _, c := range m.Capabilities {
+		if c == "container-runtime" {
+			t.Error("the controller still requires a container runtime on its machine")
+		}
+	}
+
+	digest := "sha256:" + strings.Repeat("c", 64)
+	control, err := m.Resolve([]Built{{Name: a.Name, Kind: ArtifactBundle,
+		Reference: ArtifactStoreScheme + "mesh-controller/" + a.Name + "@" + digest, Digest: digest}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// The node's runtime does not launch it: it serves its seat's verbs itself.
+	if loads := control.Bundles[0].Loads; len(loads) != 0 {
+		t.Errorf("the node's runtime would launch the controller as a tools bundle: %v", loads)
+	}
+
+	needed := map[string]map[string]string{"mesh-controller": {}}
+	for name := range m.OwnSecrets {
+		needed["mesh-controller"][name] = "sealed-" + name
+	}
+	out, err := Resolution{Node: "anchor", Modules: []Manifest{control}}.Declaration(Rendering{
+		Needed: needed, ArtifactStore: "anchor.internal:5100",
+		Seats: map[string]map[int]int{"mesh-store": {5432: 6852}},
+	})
+	if err != nil {
+		t.Fatalf("the controller does not compose: %v", err)
+	}
+
+	var process, step map[string]any
+	account, firstSecret := -1, -1
+	for i, r := range out {
+		switch {
+		case r["type"] == "container":
+			t.Errorf("the controller's declaration still runs a container: %v", r)
+		case r["id"] == "mesh-controller.controller":
+			process = r
+		case r["id"] == "mesh-controller.controller-prepare":
+			step = r
+			if process != nil {
+				t.Error("the controller's preparation is placed after the process it prepares for")
+			}
+		case r["type"] == "user" && r["name"] == "mesh-controller":
+			account = i
+		case strings.HasPrefix(fmt.Sprint(r["id"]), "mesh-controller.needs-") && firstSecret < 0:
+			firstSecret = i
+		}
+	}
+	if process == nil {
+		t.Fatalf("the controller's process is not in its declaration: %v", out)
+	}
+	if run, _ := json.Marshal(process["run"]); string(run) != `["./mesh-controller","serve"]` {
+		t.Errorf("the controller is run as %s, not its own bundle's binary", run)
+	}
+	if process["source"] != "anchor.internal:5100/mesh-controller/"+a.Name+"@"+digest || process["digest"] != digest {
+		t.Errorf("the controller's bundle is fetched from %v (%v)", process["source"], process["digest"])
+	}
+	// The user: an account the host declares, which owns what the process reads.
+	if process["user"] != "mesh-controller" || account < 0 {
+		t.Errorf("the controller runs as %v, and the account declared is at %d", process["user"], account)
+	}
+	if firstSecret >= 0 && account > firstSecret {
+		t.Error("the controller's secrets are written before the account they belong to exists")
+	}
+	for _, r := range out {
+		if strings.HasPrefix(fmt.Sprint(r["id"]), "mesh-controller.needs-") && r["owner"] != "mesh-controller" {
+			t.Errorf("%v belongs to %v, which the controller's process cannot read", r["id"], r["owner"])
+		}
+	}
+	if dir := fileNamed(out, "mesh-controller.mesh-state"); dir == nil || dir["owner"] != "mesh-controller" {
+		t.Errorf("the controller's state directory is not its account's to enter: %v", dir)
+	}
+	// Each mount became a path the process reads: nothing it is told is a path inside a container.
+	state := fmt.Sprint(fileNamed(out, "mesh-controller.mesh-state")["path"])
+	env, _ := process["env"].(map[string]any)
+	for key, value := range env {
+		v := fmt.Sprint(value)
+		if strings.HasPrefix(v, "/run/secrets") || strings.HasPrefix(v, "/broker-tls") {
+			t.Errorf("%s=%s is a path inside the container the controller no longer runs in", key, v)
+		}
+		if strings.HasSuffix(key, "_FILE") && !strings.HasPrefix(v, state+"/") {
+			t.Errorf("%s=%s is not one of the files the mesh places for it", key, v)
+		}
+	}
+	if env["MESH_BROKER_CERTIFICATE"] != "/var/lib/mesh-broker-tls/tls.crt" {
+		t.Errorf("the controller reads the broker's certificate from %v", env["MESH_BROKER_CERTIFICATE"])
+	}
+	if env["MESH_STORE_INVENTORY_PORT"] != "6852" {
+		t.Errorf("the controller is told the store is on %v; the node put it on 6852", env["MESH_STORE_INVENTORY_PORT"])
+	}
+	// The handover: the container it ran as goes only once this is running.
+	if got, _ := json.Marshal(process["replaces"]); string(got) != `["mesh-controller.server"]` {
+		t.Errorf("the controller's process replaces %s, not the container it ran as", got)
+	}
+	// And its state is prepared first, by the same program as the same account.
+	if step == nil || step["run-once"] != true || step["user"] != "mesh-controller" {
+		t.Fatalf("the controller's preparation is %v", step)
+	}
+	if run, _ := json.Marshal(step["run"]); string(run) != `["./mesh-controller","prepare"]` {
+		t.Errorf("the controller's preparation runs %s", run)
+	}
+}

internal/catalogue/declaration.go

@@ -415,6 +415,13 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 	filtering := AsNftables(rules, with.Mesh, r.PublicDomain != "", with.Foundation,
 		with.OutwardLinks, with.TunnelInterface)
 
+	// **A variable two modules set is refused whether or not anything places it** (novox/hq ADR
+	// 0203 §5): the account has one environment, and a machine whose holder arrives later should not
+	// be the moment two modules are found to disagree about it.
+	if err := variablesSetOnce(r.Modules); err != nil {
+		return nil, err
+	}
+
 	var out []map[string]any
 	for _, m := range r.Modules {
 		if with.Adopted && m.Filtering != nil {
@@ -645,7 +652,16 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 			if err != nil {
 				return nil, err
 			}
-			file, err := boundFile(*found, m.Binds[to], ConsumerIdentity(r.Node, IdentitySource(m.Slug, m.Module)), own)
+			as := ConsumerIdentity(r.Node, IdentitySource(m.Slug, m.Module))
+			// What the provider derives for THIS consumer, filled here where the consumer is
+			// known (novox/hq ADR 0201). The same fill knownFor does below, so the binding file
+			// and the module's `${bound:…}` substitutions cannot say different things.
+			told := *found
+			told.Serves, err = ServedTo(told.Serves, as)
+			if err != nil {
+				return nil, fmt.Errorf("%s is told about %s: %w", m.Module, to, err)
+			}
+			file, err := boundFile(told, m.Binds[to], as, own)
 			if err != nil {
 				return nil, err
 			}
@@ -719,7 +735,10 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 			return nil, err
 		}
 		// And what its bindings say, for the half of a connection that is not secret.
-		known := knownFor(m, r.Needs, r.Node)
+		known, err := knownFor(m, r.Needs, r.Node)
+		if err != nil {
+			return nil, err
+		}
 		// A requirement answered on this same machine is not in r.Needs — its binding file is
 		// written from `here` (above) — and so `${bound:…}` could not name it, though the file
 		// beside it said the same facts. Filled from the same answer, so the two cannot disagree.
@@ -736,7 +755,11 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 			}
 			local := *answered
 			local.For = m.Module
-			for provision, values := range knownFor(m, []Needed{local}, r.Node) {
+			here, err := knownFor(m, []Needed{local}, r.Node)
+			if err != nil {
+				return nil, err
+			}
+			for provision, values := range here {
 				known[provision] = values
 			}
 		}
@@ -761,6 +784,17 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 		// And the machine underneath, which no binding of its own can tell it.
 		thisMachine := machineFacts(r, with.Names, with.MeshRange)
 
+		// **A definition that already holds the answer transcribed it** (novox/hq ADR 0201).
+		// Judged over what the module itself declares, and before anything is substituted: the
+		// mesh's own generated files — the binding, the contributions — legitimately carry the
+		// derived value, and after substitution so does every consumer's file, so this is the one
+		// moment the two can be told apart.
+		for _, own := range m.Resources {
+			if err := notTranscribed(own, known, m.Module); err != nil {
+				return nil, err
+			}
+		}
+
 		// Which of this module's files carry a secret, for the rule that a container may not read
 		// one of them as its environment without saying so (ADR 0086, issue 041).
 		secretFiles := secretFilesOf(resources)
@@ -862,6 +896,13 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 				return nil, err
 			}
 			publishedOn(copied, m.Module, with)
+			// The account's environment and every module's shell code, where this module holds the
+			// seat that places them (novox/hq ADR 0203, ADR 0204). Gathered from every module on
+			// the node, as the jails are, and **last of every placeholder pass**: shell code is a
+			// shell's own syntax, full of `${…}` no pass above should ever be shown.
+			if err := contributionsInto(copied, m, r.Modules, thisMachine); err != nil {
+				return nil, err
+			}
 			copied["id"] = m.Module + "." + fmt.Sprint(resource["id"])
 			// A service saying what it reflects names resources within its own module, so those
 			// are prefixed too or they would point at nothing.
@@ -880,6 +921,15 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 			if renamed := reflectsRenamed(m.Module, resource["reload-on"]); renamed != nil {
 				copied["reload-on"] = renamed
 			}
+			// And which of its module's containers a scheduled step holds still (novox/hq ADR 0189).
+			// **The loudest of the three when it is missed.** An unprefixed `restart-on` matches
+			// nothing and a service quietly never restarts; an unprefixed `while-stopped` names a
+			// container the declaration does not contain, and the host refuses the whole
+			// declaration — so the machine takes nothing at all, for every push, until this is
+			// right. That is what it did on the control node (2026-10-04).
+			if renamed := reflectsRenamed(m.Module, resource[WhileStopped]); renamed != nil {
+				copied[WhileStopped] = renamed
+			}
 			// And what a process replaces (novox/hq issue 213): a resource of this module's that it
 			// no longer declares, named as the host recorded it, or the host hands nothing over and
 			// removes it first.
@@ -1175,6 +1225,19 @@ type Contribution struct {
 	// requirement's name — everything providing `reverse-proxy` understands the same shape, which
 	// is what makes swapping one for another cost nothing.
 	Values map[string]any `json:"values"`
+	// Derived is what this provider's own definition said it derives for this consumer, already
+	// derived (novox/hq ADR 0201).
+	//
+	// **The provider is told, rather than recomputing it.** A served value may name the consumer's
+	// identity — a bucket named for who is asking, a database prefixed with it — and before this
+	// the rule lived twice: once in the provisioner's code, once transcribed into every consumer's
+	// definition. The mesh fills the provider's own statement here and delivers the same filled
+	// value to the consumer, so the two cannot disagree: there is no second computation to
+	// disagree with.
+	//
+	// Only the keys that are per-consumer. The rest of what the provider serves is the same for
+	// everyone and is in its own definition, where it already is.
+	Derived map[string]any `json:"derived,omitempty"`
 }
 
 // grantPath is where one consumer's sealed credential lands on the providing machine.
@@ -1266,12 +1329,17 @@ func (r Resolution) contributions(settings SettingsBy, grants []Grant,
 			// told about it and withdraws the login on its next pass.
 			continue
 		}
+		as := holderAs(ConsumerIdentity(g.Consumer, IdentitySource(g.Slug, g.From)), g.Local)
+		derived, err := r.derivedFor(g.Provision, as, g.From, g.Local, settings)
+		if err != nil {
+			return nil, err
+		}
 		out[g.Provision] = append(out[g.Provision], Contribution{
-			From: g.From, Node: g.Consumer, At: g.At, Values: g.Values,
+			From: g.From, Node: g.Consumer, At: g.At, Values: g.Values, Derived: derived,
 			// One holder per local name: the identity the consumer is known by, and the local name
 			// after it where the module keeps several (ADR 0094). Not a login any backend checks —
 			// a secret is not a login — so the identity limit does not apply to the suffix.
-			As:     holderAs(ConsumerIdentity(g.Consumer, IdentitySource(g.Slug, g.From)), g.Local),
+			As:     as,
 			Secret: grantPath(directories[g.Provision], g.Consumer, holderAs(g.From, g.Local)),
 		})
 		if granted[g.Provision] == nil {

internal/catalogue/derived_for_consumer_test.go

@@ -0,0 +1,343 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// What a provider derives for each consumer, said once and delivered to both ends
+// (novox/hq ADR 0201, issue 124).
+//
+// The failure these are written against: the object store's provisioner derived each consumer's
+// bucket from the login the mesh minted, in its own code, and the mesh had no channel to tell the
+// consumer which bucket that was — so all three consumers wrote the answer into their own
+// definitions by hand. Two were right. One named a predecessor's bucket and would have
+// authenticated successfully and been refused on every object. Each of them also named the
+// machine the module happens to run on, which a definition may not do.
+
+// store is an object store in the shape minio has: it serves a region and a port to everyone, and
+// a bucket named for whoever is asking.
+func store() Manifest {
+	return Manifest{
+		Module: "store", Version: "1",
+		Provides: FromAnywhere("s3-bucket"),
+		Listens:  []Listening{{Port: 9000, Protocol: "tcp", From: FromMesh}},
+		Serves: map[string]map[string]any{"s3-bucket": {
+			"region": "eu-west",
+			"bucket": "${consumer:as:dns}",
+		}},
+		Receives: map[string]string{"s3-bucket": "/var/lib/store/grants/mesh.json"},
+		Grants:   map[string]string{"s3-bucket": "/var/lib/store/grants"},
+		Resources: []map[string]any{{
+			"id": "server", "type": "container", "name": "store", "ports": []any{"9000"},
+		}},
+	}
+}
+
+// files is a consumer that writes the bucket into its own configuration — which is the thing it
+// could not do before, and had to transcribe.
+func files() Manifest {
+	return Manifest{
+		Module: "files", Version: "1", Slug: "files",
+		Requires: []string{"s3-bucket"},
+		Binds:    map[string]string{"s3-bucket": "/var/lib/files/store.json"},
+		Secrets:  map[string]string{"s3-bucket": "/var/lib/files/store.secret"},
+		Resources: []map[string]any{{
+			"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
+			"content": "BUCKET=${bound:s3-bucket:bucket}\nREGION=${bound:s3-bucket:region}\n",
+		}},
+	}
+}
+
+// pics is a second consumer of the same provider on the same machine: two derivations, neither
+// the other's.
+func pics() Manifest {
+	return Manifest{
+		Module: "pics", Version: "1", Slug: "pics",
+		Requires: []string{"s3-bucket"},
+		Binds:    map[string]string{"s3-bucket": "/var/lib/pics/store.json"},
+		Secrets:  map[string]string{"s3-bucket": "/var/lib/pics/store.secret"},
+		Resources: []map[string]any{{
+			"id": "env", "type": "file", "path": "/var/lib/pics/env", "mode": "0600",
+			"content": "BUCKET=${bound:s3-bucket:bucket}\n",
+		}},
+	}
+}
+
+// The three places the derived value lands must agree, because agreeing is the whole point: the
+// consumer's own file, the binding it reads as JSON, and the provider's contributions entry.
+func TestADerivedValueReachesBothEndsAndAgrees(t *testing.T) {
+	r, err := Resolve(shelf(store(), files()), []string{"store", "files"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	out, err := r.Declaration(Rendering{Grants: []Grant{{
+		Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+		Values: map[string]any{}, Sealed: "c2VhbGVk",
+	}}})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	// The mesh minted this identity for the consumer; the bucket is that identity as a DNS label.
+	// Derived here with the mesh's own function, so the test cannot agree with a wrong rule.
+	as := ConsumerIdentity("workstation", IdentitySource("files", "files"))
+	want := strings.ReplaceAll(as, "_", "-")
+	if want == as || !strings.Contains(as, "_") {
+		t.Fatalf("the mesh's identity %q has no separator to rewrite; this test proves nothing", as)
+	}
+
+	env := fileNamed(out, "files.env")
+	if env == nil {
+		t.Fatalf("the consumer was given no file: %v", out)
+	}
+	if got := env["content"].(string); !strings.Contains(got, "BUCKET="+want+"\n") {
+		t.Errorf("the consumer's own file was not told the bucket:\n%s\nwant BUCKET=%s", got, want)
+	}
+
+	binding := fileNamed(out, "files.bound-s3-bucket")
+	if binding == nil {
+		t.Fatalf("the consumer was given no binding: %v", out)
+	}
+	var said struct {
+		Serves map[string]any `json:"serves"`
+	}
+	if err := json.Unmarshal([]byte(binding["content"].(string)), &said); err != nil {
+		t.Fatal(err)
+	}
+	if said.Serves["bucket"] != want {
+		t.Errorf("the binding says the bucket is %q, want %q", said.Serves["bucket"], want)
+	}
+	// And what is the same for everybody is still the same for everybody.
+	if said.Serves["region"] != "eu-west" {
+		t.Errorf("the binding lost what the provider serves to all: %v", said.Serves)
+	}
+
+	given := storeGrants(t, out)
+	if len(given) != 1 {
+		t.Fatalf("the provider was told about %d consumer(s): %v", len(given), given)
+	}
+	if given[0].Derived["bucket"] != want {
+		t.Errorf("the provider was told the bucket is %v, and the consumer was told %q — "+
+			"the two ends disagree, which is the whole failure", given[0].Derived["bucket"], want)
+	}
+	// Only the per-consumer half. The region is the same for everyone and is already in the
+	// provider's own definition; repeating it here would be a copy to go stale.
+	if _, carried := given[0].Derived["region"]; carried {
+		t.Errorf("the provider was handed back what it already says for everyone: %v", given[0].Derived)
+	}
+}
+
+// Two consumers of one provider on one machine get two buckets, and neither gets the other's.
+func TestTwoConsumersOfOneProviderGetTheirOwnDerivation(t *testing.T) {
+	r, err := Resolve(shelf(store(), files(), pics()),
+		[]string{"store", "files", "pics"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	out, err := r.Declaration(Rendering{Grants: []Grant{
+		{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+			Values: map[string]any{}, Sealed: "c2VhbGVk"},
+		{Provision: "s3-bucket", Consumer: "workstation", From: "pics", Slug: "pics",
+			Values: map[string]any{}, Sealed: "c2VhbGVk"},
+	}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	forFiles := strings.ReplaceAll(ConsumerIdentity("workstation", IdentitySource("files", "files")), "_", "-")
+	forPics := strings.ReplaceAll(ConsumerIdentity("workstation", IdentitySource("pics", "pics")), "_", "-")
+	if forFiles == forPics {
+		t.Fatal("the two consumers were given the same identity; this test proves nothing")
+	}
+	if got := fileNamed(out, "files.env")["content"].(string); !strings.Contains(got, "BUCKET="+forFiles+"\n") {
+		t.Errorf("files was not given its own bucket:\n%s", got)
+	}
+	if got := fileNamed(out, "pics.env")["content"].(string); !strings.Contains(got, "BUCKET="+forPics+"\n") {
+		t.Errorf("pics was not given its own bucket:\n%s", got)
+	}
+	var buckets []any
+	for _, g := range storeGrants(t, out) {
+		buckets = append(buckets, g.Derived["bucket"])
+	}
+	if len(buckets) != 2 || buckets[0] == buckets[1] {
+		t.Errorf("the provider was told %v; it must be told one bucket per consumer", buckets)
+	}
+}
+
+// An operator may still set what the provider serves, and the mesh still derives the rest: the
+// setting is laid on first, then the consumer's half is filled.
+func TestASettingComposesWithADerivedValue(t *testing.T) {
+	r, err := Resolve(shelf(store(), files()), []string{"store", "files"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	out, err := r.Declaration(Rendering{
+		Settings: SettingsBy{"store": {{From: "the operator",
+			Values: map[string]any{"bucket": "team-${consumer:as:dns}"}}}},
+		Grants: []Grant{{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+			Values: map[string]any{}, Sealed: "c2VhbGVk"}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := "team-" + strings.ReplaceAll(ConsumerIdentity("workstation", IdentitySource("files", "files")), "_", "-")
+	if got := fileNamed(out, "files.env")["content"].(string); !strings.Contains(got, "BUCKET="+want+"\n") {
+		t.Errorf("the operator's prefix did not survive the derivation:\n%s\nwant BUCKET=%s", got, want)
+	}
+	if given := storeGrants(t, out); given[0].Derived["bucket"] != want {
+		t.Errorf("the provider was told %v, the consumer %q", given[0].Derived["bucket"], want)
+	}
+}
+
+// A fact or an alphabet the mesh does not have is refused where the definition is, not where a
+// consumer happens to be resolved — and the refusal says what may be said instead.
+func TestAServedValueNamingSomethingTheMeshDoesNotHaveIsRefused(t *testing.T) {
+	for _, c := range []struct{ value, says string }{
+		{"${consumer:node}", "as"},
+		{"${consumer:as:punycode}", "dns"},
+	} {
+		m := store()
+		m.Serves["s3-bucket"]["bucket"] = c.value
+		raw, err := json.Marshal(m)
+		if err != nil {
+			t.Fatal(err)
+		}
+		_, err = ParseManifest(raw)
+		if err == nil {
+			t.Fatalf("%s was accepted", c.value)
+		}
+		if !strings.Contains(err.Error(), c.value) {
+			t.Errorf("the refusal of %s does not quote it: %v", c.value, err)
+		}
+		if !strings.Contains(err.Error(), c.says) {
+			t.Errorf("the refusal of %s does not say what may be said (%q): %v", c.value, c.says, err)
+		}
+	}
+}
+
+// `dns` is checked against an identity the mesh actually mints, not an invented string.
+func TestTheDNSAlphabetIsTheMintedIdentityWithItsSeparatorRewritten(t *testing.T) {
+	as := ConsumerIdentity("anchor", IdentitySource("ncloud", "nextcloud"))
+	if err := CheckIdentity("anchor", IdentitySource("ncloud", "nextcloud")); err != nil {
+		t.Fatalf("the mesh would not mint this identity at all: %v", err)
+	}
+	label := asDNSLabel(as)
+	if strings.Contains(label, "_") {
+		t.Errorf("%q is not a DNS label", label)
+	}
+	if strings.ReplaceAll(label, "-", "_") != as {
+		t.Errorf("%q is not %q with its separator rewritten", label, as)
+	}
+}
+
+// The check that would have caught the one wrong instance: a consumer that writes the derived
+// value into its own definition instead of asking for it is refused, whether it transcribed the
+// right answer or a predecessor's.
+func TestAConsumerThatTranscribesWhatItsProviderDerivesIsRefused(t *testing.T) {
+	as := ConsumerIdentity("workstation", IdentitySource("files", "files"))
+	transcribed := strings.ReplaceAll(as, "_", "-")
+
+	m := files()
+	m.Resources = []map[string]any{{
+		"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
+		// Exactly what the provider will create — correct today, and a copy of a rule that is
+		// not this module's.
+		"content": "BUCKET=" + transcribed + "\n",
+	}}
+	r, err := Resolve(shelf(store(), m), []string{"store", "files"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = r.Declaration(Rendering{Grants: []Grant{{
+		Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+		Values: map[string]any{}, Sealed: "c2VhbGVk",
+	}}})
+	if err == nil {
+		t.Fatal("a definition holding its own copy of the provider's naming rule was accepted")
+	}
+	if !strings.Contains(err.Error(), "${bound:s3-bucket:bucket}") {
+		t.Errorf("the refusal does not say what to write instead: %v", err)
+	}
+
+	// And a constant the provider serves to everyone is not a transcription: repeating it is
+	// redundant, not wrong, and refusing it would be the mesh policing style.
+	m.Resources = []map[string]any{{
+		"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
+		"content": "REGION=eu-west\n",
+	}}
+	r, err = Resolve(shelf(store(), m), []string{"store", "files"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := r.Declaration(Rendering{Grants: []Grant{{
+		Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+		Values: map[string]any{}, Sealed: "c2VhbGVk",
+	}}}); err != nil {
+		t.Errorf("a value the provider serves to everyone was judged a transcription: %v", err)
+	}
+}
+
+func storeGrants(t *testing.T, out []map[string]any) []Contribution {
+	t.Helper()
+	for _, r := range out {
+		if r["path"] != "/var/lib/store/grants/mesh.json" {
+			continue
+		}
+		var parsed struct {
+			Given []Contribution `json:"given"`
+		}
+		if err := json.Unmarshal([]byte(r["content"].(string)), &parsed); err != nil {
+			t.Fatal(err)
+		}
+		return parsed.Given
+	}
+	t.Fatalf("the provider was given no contributions file: %v", out)
+	return nil
+}
+
+// A consumer that keeps SEVERAL holders of one provision is refused, rather than told one thing
+// while its provider is told another.
+//
+// **This is issue 124's own failure, one case to the side.** The mesh gives each holder its own
+// login — `mesh_node_mod_<local>` (ADR 0094) — and the provider derives from the login, so it
+// would make one resource per holder. The consumer's side has no such dimension: there is one
+// binding file per provision and one `${bound:<provision>:<key>}`, both derived from the
+// un-suffixed identity. So the provider would create `…-mod-cold` and the consumer would be
+// configured against `…-mod`: it would authenticate successfully and be refused on every object,
+// which is exactly the fault this whole record exists to end.
+//
+// Refused, loudly, at the one place that can see both halves. Lifting it means giving the
+// consumer's side a local dimension, which is a decision and not an omission.
+func TestAConsumerWithSeveralHoldersOfADerivingProviderIsRefused(t *testing.T) {
+	m := files()
+	// Two holders of the one provision, the shape ADR 0094 gives a module that keeps several.
+	m.Secrets = nil
+	m.SecretsMany = map[string]map[string]string{"s3-bucket": {
+		"hot":  "/var/lib/files/hot.secret",
+		"cold": "/var/lib/files/cold.secret",
+	}}
+	m.Resources = []map[string]any{{
+		"id": "env", "type": "file", "path": "/var/lib/files/env", "mode": "0600",
+		"content": "BUCKET=${bound:s3-bucket:bucket}\n",
+	}}
+	r, err := Resolve(shelf(store(), m), []string{"store", "files"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = r.Declaration(Rendering{Grants: []Grant{
+		{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+			Local: "hot", Values: map[string]any{}, Sealed: "c2VhbGVk"},
+		{Provision: "s3-bucket", Consumer: "workstation", From: "files", Slug: "files",
+			Local: "cold", Values: map[string]any{}, Sealed: "c2VhbGVk"},
+	}})
+	if err == nil {
+		t.Fatal("a consumer with several holders of a deriving provider was accepted; " +
+			"its two ends would have disagreed in silence")
+	}
+	for _, want := range []string{"files", "s3-bucket", "bucket"} {
+		if !strings.Contains(err.Error(), want) {
+			t.Errorf("the refusal does not name %q: %v", want, err)
+		}
+	}
+}

internal/catalogue/environment_into.go

@@ -0,0 +1,523 @@
+package catalogue
+
+import (
+	"fmt"
+	"regexp"
+	"sort"
+	"strings"
+)
+
+// The account's environment and the login shell's code, composed from the modules a node runs
+// (novox/hq ADR 0203, ADR 0204).
+//
+// **The same shape as the jails.** Every module may contribute — a toolchain its directory on PATH,
+// a version manager a variable naming its home, a prompt the code that loads it — naming no node, no
+// path and no file of the shell's (ADR 0112). The one module holding the matching seat places the
+// result with a placeholder in its own file, and the controller fills it from every module on the
+// node. A node not running a module has none of its contribution, and unassigning one takes its
+// lines away at the next composition.
+//
+// **Two kinds of contribution, kept apart on purpose.** The environment is facts, which the
+// controller writes in two standard formats — POSIX assignment and the service manager's
+// environment.d — so a terminal, a script, the login shell's `execute` and a graphical session all
+// read the same values (ADR 0203). Shell code is not a fact: it is text in one shell's syntax, which
+// the controller sorts into a slot and pastes without reading, as it pastes a jail's stanza (ADR
+// 0204).
+
+// EnvironmentSeat and LoginShellSeat are the seats whose holders may place what the modules
+// contributed: the account's environment, and the login shell's code.
+const (
+	EnvironmentSeat = "node-environment"
+	LoginShellSeat  = "node-login-shell"
+)
+
+// Where an environment entry on PATH goes: before the account's existing PATH, or after it.
+const (
+	PathAtStart = "start"
+	PathAtEnd   = "end"
+)
+
+// Environment is what one module adds to the account's environment (novox/hq ADR 0203).
+type Environment struct {
+	// Variables are names and literal values. A value may name the machine's own facts with
+	// ${machine:…}, resolved before anything is written, and nothing else that expands.
+	Variables map[string]string `json:"variables,omitempty"`
+	// Path is entries on the account's PATH, each at its start or its end, in the order declared.
+	Path []PathEntry `json:"path,omitempty"`
+}
+
+// PathEntry is one directory a module puts on the account's PATH.
+type PathEntry struct {
+	Entry string `json:"entry"`
+	At    string `json:"at"`
+}
+
+// ShellCode is one piece of code a module adds to a shell's startup (novox/hq ADR 0204).
+type ShellCode struct {
+	// For is the shell the code is written in.
+	For string `json:"for"`
+	// Slot is where it runs among the other modules' code: first, normal or last. Named rather
+	// than numbered, because every contributor would guess a number and a collision says nothing.
+	Slot string `json:"slot"`
+	// Code is never interpreted — it is the shell's syntax, and only the shell reads it.
+	Code string `json:"code"`
+}
+
+// The shells and slots a contribution may name (novox/hq ADR 0204). Closed, so a typo is a refusal
+// at the check rather than code that silently lands in no placeholder.
+var (
+	knownShells = []string{"zsh", "bash", "fish"}
+	knownSlots  = []string{"first", "normal", "last"}
+)
+
+// The two renderings of the environment a holder may place (novox/hq ADR 0203, decision 3).
+const (
+	EnvironmentPOSIX   = "posix"
+	EnvironmentSystemd = "systemd"
+)
+
+// ofEnvironment and ofShell are where a holder places what was contributed: ${environment:posix},
+// ${environment:systemd} and ${shell:<shell>:<slot>}. Loose inside the braces on purpose, so a
+// misspelt key is found and refused rather than left in a file as a literal nobody reads.
+var (
+	ofEnvironment = regexp.MustCompile(`\$\{environment:([^}]*)\}`)
+	ofShell       = regexp.MustCompile(`\$\{shell:([^}]*)\}`)
+)
+
+// variableName is a POSIX shell variable name, which is also what environment.d accepts.
+var variableName = regexp.MustCompile(`^[A-Za-z_][A-Za-z0-9_]*$`)
+
+// environmentProblems is what is wrong with this module's environment contribution, from the
+// manifest alone.
+func (m Manifest) environmentProblems() []string {
+	if m.Environment == nil {
+		return nil
+	}
+	var problems []string
+	for _, n := range sortedKeys(m.Environment.Variables) {
+		switch {
+		case !variableName.MatchString(n):
+			problems = append(problems, fmt.Sprintf(
+				"%s sets the variable %q, which is not a name a shell accepts: a letter or an "+
+					"underscore, then letters, digits and underscores", m.Module, n))
+			continue
+		case n == "PATH":
+			// PATH is the one variable every module shares, so no module may set it whole: a second
+			// setter would replace the first's entries, and the account's own PATH with them.
+			problems = append(problems, fmt.Sprintf(
+				"%s sets PATH as a variable; a module adds an entry under environment.path, at the "+
+					"start or the end, and PATH is composed from every module's (novox/hq ADR 0203)", m.Module))
+			continue
+		}
+		if why := literalProblem(m.Environment.Variables[n]); why != "" {
+			problems = append(problems, fmt.Sprintf(
+				"%s sets %s to %q, which %s — %s", m.Module, n, m.Environment.Variables[n], why, literalRule))
+		}
+	}
+	seen := map[string]bool{}
+	for i, p := range m.Environment.Path {
+		switch {
+		case p.Entry == "":
+			problems = append(problems, fmt.Sprintf("%s's PATH entry %d names no directory", m.Module, i+1))
+		case strings.Contains(ofMachine.ReplaceAllString(p.Entry, ""), ":"):
+			// A colon is PATH's own separator, so an entry holding one is two entries, and the
+			// check that it is already present would look for the wrong thing.
+			problems = append(problems, fmt.Sprintf(
+				"%s puts %q on PATH, which holds a colon, PATH's own separator", m.Module, p.Entry))
+		case seen[p.Entry]:
+			problems = append(problems, fmt.Sprintf("%s puts %q on PATH twice", m.Module, p.Entry))
+		default:
+			if why := literalProblem(p.Entry); why != "" {
+				problems = append(problems, fmt.Sprintf(
+					"%s puts %q on PATH, which %s — %s", m.Module, p.Entry, why, literalRule))
+			}
+		}
+		seen[p.Entry] = true
+		if p.At != PathAtStart && p.At != PathAtEnd {
+			problems = append(problems, fmt.Sprintf(
+				"%s puts %q on PATH at %q; an entry goes at %q or %q of the account's PATH",
+				m.Module, p.Entry, p.At, PathAtStart, PathAtEnd))
+		}
+	}
+	return problems
+}
+
+// literalRule is why a value must be literal, said with every refusal of one.
+const literalRule = "a value is literal, so a POSIX shell and the service manager read it alike, and " +
+	"names the machine only through the mesh's own ${machine:…} facts (novox/hq ADR 0203)"
+
+// literalProblem is why a value cannot be written, unquoted by either reader, as the same string in
+// both formats — or nothing. A `$` would expand differently in each; a quote or a backslash is
+// quoting in one and a character in the other; a line break ends the line in both.
+func literalProblem(v string) string {
+	switch {
+	case strings.ContainsAny(v, `'"`):
+		return "holds a quote"
+	case strings.Contains(v, `\`):
+		return "holds a backslash"
+	case strings.ContainsAny(v, "\n\r"):
+		return "holds a line break"
+	case strings.ContainsRune(v, 0):
+		return "holds a NUL"
+	case strings.Contains(ofMachine.ReplaceAllString(v, ""), "$"):
+		return "holds a $ that is not one of the machine's ${machine:…} facts"
+	}
+	return ""
+}
+
+// shellProblems is what is wrong with this module's shell code, from the manifest alone. The code
+// itself is not judged: it is the shell's syntax, which the controller does not read.
+func (m Manifest) shellProblems() []string {
+	var problems []string
+	for i, c := range m.Shell {
+		if !oneOf(knownShells, c.For) {
+			problems = append(problems, fmt.Sprintf(
+				"%s's shell code %d is for %q; the shells are %s", m.Module, i+1, c.For,
+				strings.Join(knownShells, ", ")))
+		}
+		if !oneOf(knownSlots, c.Slot) {
+			problems = append(problems, fmt.Sprintf(
+				"%s's shell code %d goes in the slot %q; the slots are %s", m.Module, i+1, c.Slot,
+				strings.Join(knownSlots, ", ")))
+		}
+		if strings.TrimSpace(c.Code) == "" {
+			problems = append(problems, fmt.Sprintf("%s's shell code %d has no code", m.Module, i+1))
+		}
+	}
+	return problems
+}
+
+// contributionPlaceholderProblems is every place this module's resources name the environment or
+// the shell's code and may not — judged from the manifest, so the catalogue check refuses it before
+// a mesh does, and again at composition in the same words.
+func (m Manifest) contributionPlaceholderProblems() []string {
+	var problems []string
+	for _, r := range m.Resources {
+		problems = append(problems, placeholderProblems(m, r)...)
+	}
+	return problems
+}
+
+// placeholderProblems is what is wrong with one resource's ${environment:…} and ${shell:…}.
+//
+// **The seat authorises it, not the placeholder** (novox/hq ADR 0203 §5, ADR 0204 §3), as the seat
+// authorises the bus's user list: a module that does not hold the account's environment writing it
+// would be a second writer of a file there is one of, and a module that does not hold the login
+// shell writing every module's shell code would be a second shell.
+func placeholderProblems(m Manifest, r map[string]any) []string {
+	var problems []string
+	for _, field := range sortedKeys(r) {
+		s, ok := r[field].(string)
+		if !ok {
+			continue
+		}
+		env := ofEnvironment.FindAllStringSubmatch(s, -1)
+		code := ofShell.FindAllStringSubmatch(s, -1)
+		if len(env)+len(code) == 0 {
+			continue
+		}
+		if field != "content" {
+			// Placed only where a file's bytes are, which is where every one of them is meant to go:
+			// a path or an owner holding several lines of shell is nothing the host could act on.
+			problems = append(problems, fmt.Sprintf(
+				"%s's resource %v names %s in its %s; the environment and the shell's code are placed "+
+					"only in a file's content", m.Module, r["id"], placeholderOf(env, code), field))
+			continue
+		}
+		for _, e := range env {
+			if e[1] != EnvironmentPOSIX && e[1] != EnvironmentSystemd {
+				problems = append(problems, fmt.Sprintf(
+					"%s's resource %v names %s; the environment is ${environment:%s} or ${environment:%s}",
+					m.Module, r["id"], e[0], EnvironmentPOSIX, EnvironmentSystemd))
+			}
+		}
+		if len(env) > 0 && !m.ClaimsSeat(EnvironmentSeat) {
+			problems = append(problems, fmt.Sprintf(
+				"%s's resource %v names %s and %s does not claim %s; the account's environment is "+
+					"written by that seat's holder alone (novox/hq ADR 0203)",
+				m.Module, r["id"], env[0][0], m.Module, EnvironmentSeat))
+		}
+		for _, c := range code {
+			shell, slot, two := strings.Cut(c[1], ":")
+			if !two || !oneOf(knownShells, shell) || !oneOf(knownSlots, slot) {
+				problems = append(problems, fmt.Sprintf(
+					"%s's resource %v names %s; shell code is ${shell:<shell>:<slot>}, the shell one of "+
+						"%s and the slot one of %s", m.Module, r["id"], c[0],
+					strings.Join(knownShells, ", "), strings.Join(knownSlots, ", ")))
+			}
+		}
+		if len(code) > 0 && !m.ClaimsSeat(LoginShellSeat) {
+			problems = append(problems, fmt.Sprintf(
+				"%s's resource %v names %s and %s does not claim %s; every module's shell code is "+
+					"placed by the login shell's holder alone (novox/hq ADR 0204)",
+				m.Module, r["id"], code[0][0], m.Module, LoginShellSeat))
+		}
+	}
+	return problems
+}
+
+func placeholderOf(env, code [][]string) string {
+	if len(env) > 0 {
+		return env[0][0]
+	}
+	return code[0][0]
+}
+
+// contributedEnvironment is one node's environment, gathered and in the order it is written.
+type contributedEnvironment struct {
+	// variables is by module in name order, each module's sorted by name.
+	variables []setBy
+	// start and end are PATH's entries in their final order, each once.
+	start, end []placedOn
+}
+
+type setBy struct {
+	module string
+	names  []string
+	values map[string]string
+}
+
+type placedOn struct {
+	module, entry string
+}
+
+// inModuleOrder is the modules sorted by name — the order contributions are written in (novox/hq
+// ADR 0203, ADR 0204), so the same set composes byte for byte whatever order they were assigned in.
+func inModuleOrder(modules []Manifest) []Manifest {
+	out := append([]Manifest(nil), modules...)
+	sort.SliceStable(out, func(a, b int) bool { return out[a].Module < out[b].Module })
+	return out
+}
+
+// variablesSetOnce refuses a variable two modules on one node both set (novox/hq ADR 0203 §5),
+// naming both. Neither is chosen: whichever was written last would win in one reader and not
+// necessarily in the other, and the module that lost would not be told.
+func variablesSetOnce(modules []Manifest) error {
+	setter := map[string]string{}
+	for _, m := range inModuleOrder(modules) {
+		if m.Environment == nil {
+			continue
+		}
+		for _, n := range sortedKeys(m.Environment.Variables) {
+			if first, taken := setter[n]; taken {
+				return fmt.Errorf(
+					"%s and %s both set %s on this machine; the account has one environment, so one "+
+						"of them must stop setting it (novox/hq ADR 0203)", first, m.Module, n)
+			}
+			setter[n] = m.Module
+		}
+	}
+	return nil
+}
+
+// environmentOn gathers every module's environment on a node, with the machine's facts in place.
+//
+// A PATH entry two modules both add is written once, where the first puts it: two toolchains
+// sharing ~/.local/bin is ordinary, and nothing about it is in conflict.
+func environmentOn(modules []Manifest, facts map[string]string) (contributedEnvironment, error) {
+	var env contributedEnvironment
+	if err := variablesSetOnce(modules); err != nil {
+		return env, err
+	}
+	placed := map[string]bool{}
+	for _, m := range inModuleOrder(modules) {
+		if m.Environment == nil {
+			continue
+		}
+		if len(m.Environment.Variables) > 0 {
+			set := setBy{module: m.Module, values: map[string]string{}}
+			for _, n := range sortedKeys(m.Environment.Variables) {
+				v, err := factsIn(m.Environment.Variables[n], facts, m.Module, n)
+				if err != nil {
+					return env, err
+				}
+				set.names = append(set.names, n)
+				set.values[n] = v
+			}
+			env.variables = append(env.variables, set)
+		}
+		for _, p := range m.Environment.Path {
+			entry, err := factsIn(p.Entry, facts, m.Module, "a PATH entry")
+			if err != nil {
+				return env, err
+			}
+			if strings.Contains(entry, ":") {
+				return env, fmt.Errorf("%s puts %q on PATH on this machine, which holds a colon, PATH's own separator",
+					m.Module, entry)
+			}
+			if placed[entry] {
+				continue
+			}
+			placed[entry] = true
+			if p.At == PathAtEnd {
+				env.end = append(env.end, placedOn{m.Module, entry})
+			} else {
+				env.start = append(env.start, placedOn{m.Module, entry})
+			}
+		}
+	}
+	return env, nil
+}
+
+// factsIn resolves a contributed value's ${machine:…} facts with this machine's — first, before
+// either format is written, so both say the same thing (novox/hq ADR 0203).
+func factsIn(v string, facts map[string]string, module, what string) (string, error) {
+	for _, key := range machineUsed(v) {
+		value, has := facts[key]
+		if !has {
+			return "", fmt.Errorf("%s sets %s to a value that says ${machine:%s}, and this machine says %s",
+				module, what, key, orNothing(namesOfFacts(facts)))
+		}
+		v = strings.ReplaceAll(v, fmt.Sprintf("${machine:%s}", key), value)
+	}
+	// Judged again once filled: a fact is the mesh's, and still has to be a literal both readers
+	// take alike.
+	if why := literalProblem(v); why != "" {
+		return "", fmt.Errorf("%s sets %s to %q on this machine, which %s — %s", module, what, v, why, literalRule)
+	}
+	return v, nil
+}
+
+// posix is the environment as lines a POSIX shell sources (novox/hq ADR 0203 §3): every variable
+// exported, every PATH entry added only when it is missing, so sourcing the file twice — a login
+// shell that starts another — changes nothing. POSIX sh only, because sh, bash and zsh all read it.
+//
+// The start entries are written last-first: each is put in front of PATH, so the last written ends
+// up first, and the result reads in module order, then the order each module declared.
+func (e contributedEnvironment) posix() string {
+	var b strings.Builder
+	for _, set := range e.variables {
+		fmt.Fprintf(&b, "# %s\n", set.module)
+		for _, n := range set.names {
+			fmt.Fprintf(&b, "export %s='%s'\n", n, set.values[n])
+		}
+	}
+	named := ""
+	for i := len(e.start) - 1; i >= 0; i-- {
+		p := e.start[i]
+		if p.module != named {
+			fmt.Fprintf(&b, "# %s\n", p.module)
+			named = p.module
+		}
+		fmt.Fprintf(&b, "case \":${PATH}:\" in *':%s:'*) ;; *) PATH='%s'\"${PATH:+:${PATH}}\" ;; esac\n",
+			p.entry, p.entry)
+	}
+	named = ""
+	for _, p := range e.end {
+		if p.module != named {
+			fmt.Fprintf(&b, "# %s\n", p.module)
+			named = p.module
+		}
+		fmt.Fprintf(&b, "case \":${PATH}:\" in *':%s:'*) ;; *) PATH=\"${PATH:+${PATH}:}\"'%s' ;; esac\n",
+			p.entry, p.entry)
+	}
+	if len(e.start)+len(e.end) > 0 {
+		b.WriteString("export PATH\n")
+	}
+	return b.String()
+}
+
+// systemd is the same environment as the service manager's environment.d reads it (novox/hq ADR
+// 0203 §3), for the account's user manager and so for everything a graphical session starts. Read
+// once per manager start, so it needs no guard against running twice; the account's existing PATH
+// sits between the start and the end entries.
+func (e contributedEnvironment) systemd() string {
+	var b strings.Builder
+	for _, set := range e.variables {
+		fmt.Fprintf(&b, "# %s\n", set.module)
+		for _, n := range set.names {
+			fmt.Fprintf(&b, "%s=%s\n", n, set.values[n])
+		}
+	}
+	if len(e.start) > 0 {
+		fmt.Fprintf(&b, "# %s\nPATH=%s${PATH:+:$PATH}\n", modulesOf(e.start), entriesOf(e.start))
+	}
+	if len(e.end) > 0 {
+		fmt.Fprintf(&b, "# %s\nPATH=${PATH:+$PATH:}%s\n", modulesOf(e.end), entriesOf(e.end))
+	}
+	return b.String()
+}
+
+// modulesOf names who contributed a line holding several modules' entries, in the order they appear.
+func modulesOf(entries []placedOn) string {
+	var names []string
+	seen := map[string]bool{}
+	for _, p := range entries {
+		if !seen[p.module] {
+			seen[p.module] = true
+			names = append(names, p.module)
+		}
+	}
+	return strings.Join(names, ", ")
+}
+
+func entriesOf(entries []placedOn) string {
+	out := make([]string, len(entries))
+	for i, p := range entries {
+		out[i] = p.entry
+	}
+	return strings.Join(out, ":")
+}
+
+// shellCode is every module's code for one shell and one slot (novox/hq ADR 0204 §3): in module
+// order, each module's pieces in the order it declared them, each preceded by a line naming the
+// module, and empty when nothing is contributed.
+func shellCode(modules []Manifest, shell, slot string) string {
+	var b strings.Builder
+	for _, m := range inModuleOrder(modules) {
+		named := false
+		for _, c := range m.Shell {
+			if c.For != shell || c.Slot != slot {
+				continue
+			}
+			if !named {
+				fmt.Fprintf(&b, "# %s\n", m.Module)
+				named = true
+			}
+			b.WriteString(c.Code)
+			if !strings.HasSuffix(c.Code, "\n") {
+				b.WriteString("\n")
+			}
+		}
+	}
+	return b.String()
+}
+
+// contributionsInto fills a holder's file with the node's environment and its shell code.
+//
+// **Last, after every other placeholder pass, and in one pass each.** Shell code is contributed text
+// in a shell's own syntax — `${XDG_CACHE_HOME:-$HOME/.cache}`, `${(%):-%n}` — and the rendered
+// environment holds `${PATH:+…}`: a scanner for the mesh's own placeholders that ran after these
+// were in place would read the shell's expansions as the mesh's and refuse them, or fill a
+// `${machine:…}` some module wrote for its shell to see. So nothing runs after them, the environment
+// is filled before the shell's code is, and each is replaced in a single pass over what the holder
+// wrote, so a contributed piece is never scanned again.
+func contributionsInto(resource map[string]any, m Manifest, modules []Manifest, facts map[string]string) error {
+	if problems := placeholderProblems(m, resource); len(problems) > 0 {
+		return fmt.Errorf("%s", problems[0])
+	}
+	content, ok := resource["content"].(string)
+	if !ok {
+		return nil
+	}
+	if ofEnvironment.MatchString(content) {
+		env, err := environmentOn(modules, facts)
+		if err != nil {
+			return err
+		}
+		content = ofEnvironment.ReplaceAllStringFunc(content, func(placeholder string) string {
+			if ofEnvironment.FindStringSubmatch(placeholder)[1] == EnvironmentSystemd {
+				return env.systemd()
+			}
+			return env.posix()
+		})
+	}
+	if ofShell.MatchString(content) {
+		content = ofShell.ReplaceAllStringFunc(content, func(placeholder string) string {
+			shell, slot, _ := strings.Cut(ofShell.FindStringSubmatch(placeholder)[1], ":")
+			return shellCode(modules, shell, slot)
+		})
+	}
+	resource["content"] = content
+	return nil
+}

internal/catalogue/environment_into_test.go

@@ -0,0 +1,471 @@
+package catalogue
+
+import (
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// Defends novox/hq ADR 0203 (the account's environment is one module's, and every module
+// contributes to it) and ADR 0204 (shell code in named slots, placed by the login shell's holder).
+
+// contributors is a fixed set of contributions, in no particular order: what the renderings are
+// asserted against byte for byte. go-toolchain and zsh both put ~/.local/bin on PATH, which is the
+// ordinary case of two modules sharing a directory, and is written once.
+func contributors() []Manifest {
+	return []Manifest{
+		{Module: "zsh", Environment: &Environment{
+			Variables: map[string]string{"XDG_CONFIG_HOME": "${machine:account-home}/.config", "EDITOR": "vim"},
+			Path: []PathEntry{
+				{Entry: "${machine:account-home}/.local/bin", At: PathAtStart},
+				{Entry: "${machine:account-home}/bin", At: PathAtStart},
+				{Entry: "/opt/scripts", At: PathAtEnd},
+			},
+		}},
+		{Module: "go-toolchain", Environment: &Environment{
+			Variables: map[string]string{"GOPATH": "${machine:account-home}/go"},
+			Path: []PathEntry{
+				{Entry: "${machine:account-home}/go/bin", At: PathAtStart},
+				{Entry: "/usr/local/go/bin", At: PathAtStart},
+				{Entry: "${machine:account-home}/.local/bin", At: PathAtStart},
+			},
+		}},
+		{Module: "agent", Environment: &Environment{
+			Variables: map[string]string{"DISABLE_AUTOUPDATER": "1"},
+			Path:      []PathEntry{{Entry: "/opt/agent/bin", At: PathAtEnd}},
+		}},
+		// A module contributing nothing is in the set and writes nothing.
+		{Module: "postgres"},
+	}
+}
+
+var operatorFacts = map[string]string{"name": "workstation", "account": "op", "account-home": "/home/op"}
+
+// The final PATH this set composes, around whatever the account had: the start entries in module
+// order and then declared order, the account's own, then the end entries.
+const composedPOSIX = `# agent
+export DISABLE_AUTOUPDATER='1'
+# go-toolchain
+export GOPATH='/home/op/go'
+# zsh
+export EDITOR='vim'
+export XDG_CONFIG_HOME='/home/op/.config'
+# zsh
+case ":${PATH}:" in *':/home/op/bin:'*) ;; *) PATH='/home/op/bin'"${PATH:+:${PATH}}" ;; esac
+# go-toolchain
+case ":${PATH}:" in *':/home/op/.local/bin:'*) ;; *) PATH='/home/op/.local/bin'"${PATH:+:${PATH}}" ;; esac
+case ":${PATH}:" in *':/usr/local/go/bin:'*) ;; *) PATH='/usr/local/go/bin'"${PATH:+:${PATH}}" ;; esac
+case ":${PATH}:" in *':/home/op/go/bin:'*) ;; *) PATH='/home/op/go/bin'"${PATH:+:${PATH}}" ;; esac
+# agent
+case ":${PATH}:" in *':/opt/agent/bin:'*) ;; *) PATH="${PATH:+${PATH}:}"'/opt/agent/bin' ;; esac
+# zsh
+case ":${PATH}:" in *':/opt/scripts:'*) ;; *) PATH="${PATH:+${PATH}:}"'/opt/scripts' ;; esac
+export PATH
+`
+
+const composedSystemd = `# agent
+DISABLE_AUTOUPDATER=1
+# go-toolchain
+GOPATH=/home/op/go
+# zsh
+EDITOR=vim
+XDG_CONFIG_HOME=/home/op/.config
+# go-toolchain, zsh
+PATH=/home/op/go/bin:/usr/local/go/bin:/home/op/.local/bin:/home/op/bin${PATH:+:$PATH}
+# agent, zsh
+PATH=${PATH:+$PATH:}/opt/agent/bin:/opt/scripts
+`
+
+func TestTheEnvironmentRendersForAPOSIXShellByteForByte(t *testing.T) {
+	env, err := environmentOn(contributors(), operatorFacts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := env.posix(); got != composedPOSIX {
+		t.Fatalf("the POSIX rendering is\n%s\nnot\n%s", got, composedPOSIX)
+	}
+}
+
+func TestTheEnvironmentRendersForTheServiceManagerByteForByte(t *testing.T) {
+	env, err := environmentOn(contributors(), operatorFacts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := env.systemd(); got != composedSystemd {
+		t.Fatalf("the environment.d rendering is\n%s\nnot\n%s", got, composedSystemd)
+	}
+}
+
+// Sourcing twice changes nothing (ADR 0203 §3): a login shell that starts another reads the file
+// again, and a PATH that grew each time would be the symptom. Run by a real `sh`, because the claim
+// is about what a shell does with the file, not about what the file looks like.
+func TestThePOSIXEnvironmentSourcedTwiceLeavesPATHAsOnce(t *testing.T) {
+	sh, err := exec.LookPath("sh")
+	if err != nil {
+		t.Skip("no sh on this machine")
+	}
+	script := "PATH=/usr/bin:/bin\n" + composedPOSIX + "once=$PATH\n" + composedPOSIX +
+		`[ "$PATH" = "$once" ] || { echo "changed: $once -> $PATH"; exit 1; }` + "\n" +
+		`echo "$PATH"; echo "$GOPATH"`
+	out, err := exec.Command(sh, "-c", script).CombinedOutput()
+	if err != nil {
+		t.Fatalf("sourcing twice: %v\n%s", err, out)
+	}
+	lines := strings.Split(strings.TrimSpace(string(out)), "\n")
+	want := "/home/op/go/bin:/usr/local/go/bin:/home/op/.local/bin:/home/op/bin:/usr/bin:/bin:/opt/agent/bin:/opt/scripts"
+	if lines[0] != want {
+		t.Fatalf("PATH is %s, not %s", lines[0], want)
+	}
+	if lines[1] != "/home/op/go" {
+		t.Fatalf("GOPATH was not exported: %q", lines[1])
+	}
+	// And an entry the account already has stays where it is, and once.
+	out, err = exec.Command(sh, "-c", "PATH=/opt/scripts:/usr/bin\n"+composedPOSIX+`echo "$PATH"`).CombinedOutput()
+	if err != nil {
+		t.Fatalf("%v\n%s", err, out)
+	}
+	if got := strings.TrimSpace(string(out)); got !=
+		"/home/op/go/bin:/usr/local/go/bin:/home/op/.local/bin:/home/op/bin:/opt/scripts:/usr/bin:/opt/agent/bin" {
+		t.Fatalf("an entry already on PATH was added again or moved: %s", got)
+	}
+}
+
+// The environment.d rendering, read by the service manager's own generator where this machine has
+// one — the same reader an account's user manager runs, so the PATH it composes is the one asserted.
+func TestTheServiceManagerReadsTheSystemdRenderingAsMeant(t *testing.T) {
+	generator := "/usr/lib/systemd/user-environment-generators/30-systemd-environment-d-generator"
+	if _, err := os.Stat(generator); err != nil {
+		t.Skip("no environment.d generator on this machine")
+	}
+	config := t.TempDir()
+	if err := os.MkdirAll(filepath.Join(config, "environment.d"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(config, "environment.d", "50-mesh.conf"), []byte(composedSystemd), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	cmd := exec.Command(generator)
+	cmd.Env = []string{"PATH=/usr/bin:/bin", "HOME=" + config, "XDG_CONFIG_HOME=" + config}
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("%v\n%s", err, out)
+	}
+	want := "PATH=/home/op/go/bin:/usr/local/go/bin:/home/op/.local/bin:/home/op/bin:/usr/bin:/bin:/opt/agent/bin:/opt/scripts"
+	if !strings.Contains(string(out), want+"\n") || !strings.Contains(string(out), "GOPATH=/home/op/go\n") {
+		t.Fatalf("the service manager read\n%s", out)
+	}
+}
+
+// Nothing contributed renders nothing, in both formats — not an empty `export PATH`.
+func TestNoContributionsRenderNothing(t *testing.T) {
+	env, err := environmentOn([]Manifest{{Module: "postgres"}}, operatorFacts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if env.posix() != "" || env.systemd() != "" {
+		t.Fatalf("an empty environment rendered %q and %q", env.posix(), env.systemd())
+	}
+}
+
+// A ${machine:…} fact the machine does not have is refused naming the module, as a file's is.
+func TestAContributedFactTheMachineLacksIsRefused(t *testing.T) {
+	_, err := environmentOn(contributors(), map[string]string{"name": "server"})
+	if err == nil || !strings.Contains(err.Error(), "go-toolchain sets GOPATH") ||
+		!strings.Contains(err.Error(), "${machine:account-home}") {
+		t.Fatalf("a missing account home was not refused by name: %v", err)
+	}
+}
+
+// ADR 0203 §5: two modules setting one variable are refused, both named — neither silently wins.
+func TestAVariableTwoModulesSetIsRefusedNamingBoth(t *testing.T) {
+	modules := append(contributors(), Manifest{Module: "neovim", Environment: &Environment{
+		Variables: map[string]string{"EDITOR": "nvim"}}})
+	_, err := environmentOn(modules, operatorFacts)
+	if err == nil || err.Error() != "neovim and zsh both set EDITOR on this machine; the account has one "+
+		"environment, so one of them must stop setting it (novox/hq ADR 0203)" {
+		t.Fatalf("a variable set twice was not refused naming both: %v", err)
+	}
+	// And at composition, whether or not the node holds the environment.
+	r := Resolution{Node: "workstation", Account: "op", Modules: modules}
+	if _, err := r.Declaration(Rendering{}); err == nil || !strings.Contains(err.Error(), "neovim and zsh both set EDITOR") {
+		t.Fatalf("composition accepted a variable set twice: %v", err)
+	}
+}
+
+// shells contributes code for several shells and slots, in no order.
+func shells() []Manifest {
+	return []Manifest{
+		{Module: "zsh-syntax-highlighting", Shell: []ShellCode{
+			{For: "zsh", Slot: "last", Code: "source /usr/share/zsh/plugins/zsh-syntax-highlighting/zsh-syntax-highlighting.zsh"},
+		}},
+		{Module: "powerlevel10k", Shell: []ShellCode{
+			{For: "zsh", Slot: "first", Code: "if [[ -r \"${XDG_CACHE_HOME:-$HOME/.cache}/p10k-instant-prompt-${(%):-%n}.zsh\" ]]; then\n" +
+				"  source \"${XDG_CACHE_HOME:-$HOME/.cache}/p10k-instant-prompt-${(%):-%n}.zsh\"\nfi\n"},
+			{For: "zsh", Slot: "normal", Code: "source ~/.local/share/powerlevel10k/powerlevel10k.zsh-theme"},
+			{For: "zsh", Slot: "normal", Code: "[[ -f ~/.local/share/powerlevel10k/p10k.zsh ]] && source ~/.local/share/powerlevel10k/p10k.zsh"},
+		}},
+		{Module: "zsh-autosuggestions", Shell: []ShellCode{
+			{For: "zsh", Slot: "normal", Code: "source /usr/share/zsh/plugins/zsh-autosuggestions/zsh-autosuggestions.zsh"},
+			{For: "bash", Slot: "normal", Code: "echo not for zsh"},
+		}},
+		{Module: "direnv", Shell: []ShellCode{
+			{For: "fish", Slot: "last", Code: "direnv hook fish | source"},
+			{For: "bash", Slot: "last", Code: "eval \"$(direnv hook bash)\""},
+		}},
+	}
+}
+
+// ADR 0204 §3: a slot holds that shell's code only, in module order, each module's pieces in the
+// order it declared them under a line naming it; empty when nothing is contributed.
+func TestShellCodeLandsInItsSlotInModuleOrderForItsShellOnly(t *testing.T) {
+	if got, want := shellCode(shells(), "zsh", "normal"), "# powerlevel10k\n"+
+		"source ~/.local/share/powerlevel10k/powerlevel10k.zsh-theme\n"+
+		"[[ -f ~/.local/share/powerlevel10k/p10k.zsh ]] && source ~/.local/share/powerlevel10k/p10k.zsh\n"+
+		"# zsh-autosuggestions\n"+
+		"source /usr/share/zsh/plugins/zsh-autosuggestions/zsh-autosuggestions.zsh\n"; got != want {
+		t.Fatalf("zsh's normal slot is\n%s\nnot\n%s", got, want)
+	}
+	if got, want := shellCode(shells(), "zsh", "last"), "# zsh-syntax-highlighting\n"+
+		"source /usr/share/zsh/plugins/zsh-syntax-highlighting/zsh-syntax-highlighting.zsh\n"; got != want {
+		t.Fatalf("zsh's last slot is\n%s\nnot\n%s", got, want)
+	}
+	if got, want := shellCode(shells(), "bash", "last"), "# direnv\neval \"$(direnv hook bash)\"\n"; got != want {
+		t.Fatalf("bash's last slot is %q, not %q", got, want)
+	}
+	if got := shellCode(shells(), "fish", "first"); got != "" {
+		t.Fatalf("a slot nobody contributed to holds %q", got)
+	}
+}
+
+// The holder of node-login-shell, as WP3's zsh module writes its block, with its own zsh around the
+// slots — which holds `${…}` of the shell's own that no mesh pass may touch either.
+func zshHolder() Manifest {
+	return Manifest{Module: "zsh", Claims: []Claim{{Name: LoginShellSeat, Scope: ScopeNode}},
+		Resources: []map[string]any{
+			{"id": "zshrc", "type": "file", "path": "${machine:account-home}/.zshrc", "content": "" +
+				"${shell:zsh:first}" +
+				"PROMPT='%n@%m ${PWD/#$HOME/~} '\n" +
+				"${shell:zsh:normal}" +
+				"alias ll='ls -l'\n" +
+				"${shell:zsh:last}"},
+		}}
+}
+
+// The case the ordering exists for: contributed zsh code full of `${…}` reaches the file byte for
+// byte, because the shell's code is placed after every other placeholder pass and in one pass — a
+// scanner for the mesh's placeholders that ran after it would read `${XDG_CACHE_HOME:-…}` and
+// `${(%):-%n}` as the mesh's, or fill a `${machine:…}` some module wrote for its shell to see.
+func TestShellCodeReachesTheHoldersFileByteForByte(t *testing.T) {
+	modules := append(shells(), zshHolder(), Manifest{Module: "sly", Shell: []ShellCode{
+		{For: "zsh", Slot: "last", Code: "echo ${machine:account-home} ${secret:x} ${shell:zsh:first} ${environment:posix}"},
+	}})
+	r := Resolution{Node: "workstation", Account: "op", Modules: modules}
+	out, err := r.Declaration(Rendering{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var zshrc map[string]any
+	for _, res := range out {
+		if res["id"] == "zsh.zshrc" {
+			zshrc = res
+		}
+	}
+	if zshrc == nil {
+		t.Fatalf("the holder's file was not composed: %v", out)
+	}
+	if zshrc["path"] != "/home/op/.zshrc" {
+		t.Fatalf("the holder's own placeholders were not filled first: %v", zshrc["path"])
+	}
+	want := "# powerlevel10k\n" +
+		"if [[ -r \"${XDG_CACHE_HOME:-$HOME/.cache}/p10k-instant-prompt-${(%):-%n}.zsh\" ]]; then\n" +
+		"  source \"${XDG_CACHE_HOME:-$HOME/.cache}/p10k-instant-prompt-${(%):-%n}.zsh\"\nfi\n" +
+		"PROMPT='%n@%m ${PWD/#$HOME/~} '\n" +
+		"# powerlevel10k\n" +
+		"source ~/.local/share/powerlevel10k/powerlevel10k.zsh-theme\n" +
+		"[[ -f ~/.local/share/powerlevel10k/p10k.zsh ]] && source ~/.local/share/powerlevel10k/p10k.zsh\n" +
+		"# zsh-autosuggestions\n" +
+		"source /usr/share/zsh/plugins/zsh-autosuggestions/zsh-autosuggestions.zsh\n" +
+		"alias ll='ls -l'\n" +
+		"# sly\n" +
+		"echo ${machine:account-home} ${secret:x} ${shell:zsh:first} ${environment:posix}\n" +
+		"# zsh-syntax-highlighting\n" +
+		"source /usr/share/zsh/plugins/zsh-syntax-highlighting/zsh-syntax-highlighting.zsh\n"
+	if got := zshrc["content"]; got != want {
+		t.Fatalf("the holder's .zshrc is\n%s\nnot\n%s", got, want)
+	}
+}
+
+// The holder of node-environment places both renderings, and they are the same as rendered alone.
+func TestTheEnvironmentHolderPlacesBothRenderings(t *testing.T) {
+	holder := Manifest{Module: "node-env", Claims: []Claim{{Name: EnvironmentSeat, Scope: ScopeNode}},
+		Resources: []map[string]any{
+			{"id": "posix", "type": "file", "path": "${machine:account-home}/.config/mesh/environment.sh",
+				"content": "# The mesh's environment.\n${environment:posix}"},
+			{"id": "systemd", "type": "file", "path": "${machine:account-home}/.config/environment.d/50-mesh.conf",
+				"content": "${environment:systemd}"},
+		}}
+	r := Resolution{Node: "workstation", Account: "op", Modules: append(contributors(), holder)}
+	out, err := r.Declaration(Rendering{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	by := map[string]any{}
+	for _, res := range out {
+		by[res["id"].(string)] = res["content"]
+	}
+	if by["node-env.posix"] != "# The mesh's environment.\n"+composedPOSIX {
+		t.Fatalf("the POSIX file is\n%v", by["node-env.posix"])
+	}
+	if by["node-env.systemd"] != composedSystemd {
+		t.Fatalf("the environment.d file is\n%v", by["node-env.systemd"])
+	}
+}
+
+// ADR 0203 §5 and ADR 0204 §3: a placeholder outside the seat's holder is refused — by the parser,
+// which is what the catalogue check and registration run, and again at composition, in the same words.
+func TestAPlaceholderOutsideTheHolderIsRefused(t *testing.T) {
+	for _, c := range []struct{ content, want string }{
+		{"${environment:posix}", "toolchain's resource rc names ${environment:posix} and toolchain does not claim node-environment"},
+		{"${shell:zsh:normal}", "toolchain's resource rc names ${shell:zsh:normal} and toolchain does not claim node-login-shell"},
+	} {
+		raw := `{"module":"toolchain","resources":[{"id":"rc","type":"file","path":"/etc/rc","content":"` + c.content + `"}]}`
+		if _, err := ParseManifest([]byte(raw)); err == nil || !strings.Contains(err.Error(), c.want) {
+			t.Errorf("the catalogue check accepted %s outside its holder: %v", c.content, err)
+		}
+		m := Manifest{Module: "toolchain", Resources: []map[string]any{
+			{"id": "rc", "type": "file", "path": "/etc/rc", "content": c.content}}}
+		r := Resolution{Node: "workstation", Account: "op", Modules: []Manifest{m}}
+		if _, err := r.Declaration(Rendering{}); err == nil || !strings.Contains(err.Error(), c.want) {
+			t.Errorf("composition accepted %s outside its holder: %v", c.content, err)
+		}
+	}
+}
+
+// A key nobody renders is refused, not left in the file as a literal.
+func TestAnUnknownPlaceholderKeyIsRefused(t *testing.T) {
+	for _, c := range []struct{ content, want string }{
+		{"${environment:foo}", "names ${environment:foo}; the environment is ${environment:posix} or ${environment:systemd}"},
+		{"${shell:zsh:middle}", "names ${shell:zsh:middle}; shell code is ${shell:<shell>:<slot>}"},
+		{"${shell:tcsh:first}", "names ${shell:tcsh:first}; shell code is ${shell:<shell>:<slot>}"},
+		{"${shell:zsh}", "names ${shell:zsh}; shell code is ${shell:<shell>:<slot>}"},
+	} {
+		raw := `{"module":"holder","claims":[{"name":"node-environment","scope":"node"},{"name":"node-login-shell","scope":"node"}],` +
+			`"resources":[{"id":"rc","type":"file","path":"/etc/rc","content":"` + c.content + `"}]}`
+		if _, err := ParseManifest([]byte(raw)); err == nil || !strings.Contains(err.Error(), c.want) {
+			t.Errorf("%s was accepted: %v", c.content, err)
+		}
+	}
+	// And outside a file's content, where nothing could be placed.
+	raw := `{"module":"holder","claims":[{"name":"node-environment","scope":"node"}],` +
+		`"resources":[{"id":"rc","type":"file","path":"/etc/${environment:posix}","content":"x"}]}`
+	if _, err := ParseManifest([]byte(raw)); err == nil ||
+		!strings.Contains(err.Error(), "names ${environment:posix} in its path; the environment and the shell's code are placed only in a file's content") {
+		t.Errorf("a placeholder in a path was accepted: %v", err)
+	}
+}
+
+// What ADR 0203 §2 allows a contribution to say, refused at parse when it says anything else.
+func TestAMalformedEnvironmentIsRefusedAtParse(t *testing.T) {
+	for _, c := range []struct{ environment, want string }{
+		{`{"variables":{"1X":"a"}}`, `tool sets the variable "1X", which is not a name a shell accepts`},
+		{`{"variables":{"MY-VAR":"a"}}`, `tool sets the variable "MY-VAR", which is not a name a shell accepts`},
+		{`{"variables":{"PATH":"/bin"}}`, `tool sets PATH as a variable; a module adds an entry under environment.path`},
+		{`{"variables":{"A":"$HOME/x"}}`, `tool sets A to "$HOME/x", which holds a $ that is not one of the machine's ${machine:…} facts`},
+		{`{"variables":{"A":"${HOME}/x"}}`, `tool sets A to "${HOME}/x", which holds a $`},
+		{`{"variables":{"A":"it's"}}`, `tool sets A to "it's", which holds a quote`},
+		{`{"variables":{"A":"say \"hi\""}}`, `which holds a quote`},
+		{`{"variables":{"A":"a\\b"}}`, `which holds a backslash`},
+		{`{"variables":{"A":"a\nb"}}`, `which holds a line break`},
+		{`{"variables":{"A":"a\u0000b"}}`, `which holds a NUL`},
+		{`{"path":[{"entry":"","at":"start"}]}`, `tool's PATH entry 1 names no directory`},
+		{`{"path":[{"entry":"/a:/b","at":"start"}]}`, `tool puts "/a:/b" on PATH, which holds a colon`},
+		{`{"path":[{"entry":"$HOME/bin","at":"start"}]}`, `tool puts "$HOME/bin" on PATH, which holds a $`},
+		{`{"path":[{"entry":"/a","at":"middle"}]}`, `tool puts "/a" on PATH at "middle"; an entry goes at "start" or "end"`},
+		{`{"path":[{"entry":"/a"}]}`, `tool puts "/a" on PATH at ""`},
+		{`{"path":[{"entry":"/a","at":"start"},{"entry":"/a","at":"end"}]}`, `tool puts "/a" on PATH twice`},
+		{`{"variables":{"A":"x"},"paths":[]}`, `unknown field "paths"`},
+	} {
+		_, err := ParseManifest([]byte(`{"module":"tool","environment":` + c.environment + `}`))
+		if err == nil || !strings.Contains(err.Error(), c.want) {
+			t.Errorf("%s: want %q, got %v", c.environment, c.want, err)
+		}
+	}
+	// What is allowed: a literal, and the machine's own facts.
+	if _, err := ParseManifest([]byte(`{"module":"tool","environment":{` +
+		`"variables":{"GOPATH":"${machine:account-home}/go","DISABLE_X":"1","ANSWER":"a b+c=d"},` +
+		`"path":[{"entry":"${machine:account-home}/go/bin","at":"start"},{"entry":"/opt/x","at":"end"}]}}`)); err != nil {
+		t.Fatalf("a well-formed environment was refused: %v", err)
+	}
+}
+
+func TestMalformedShellCodeIsRefusedAtParse(t *testing.T) {
+	for _, c := range []struct{ shell, want string }{
+		{`[{"for":"tcsh","slot":"normal","code":"x"}]`, `tool's shell code 1 is for "tcsh"; the shells are zsh, bash, fish`},
+		{`[{"for":"zsh","slot":"middle","code":"x"}]`, `tool's shell code 1 goes in the slot "middle"; the slots are first, normal, last`},
+		{`[{"for":"zsh","slot":"last","code":"x"},{"for":"zsh","slot":"last","code":" \n"}]`, `tool's shell code 2 has no code`},
+		{`[{"for":"zsh","slot":"last","code":"x","order":1}]`, `unknown field "order"`},
+	} {
+		_, err := ParseManifest([]byte(`{"module":"tool","shell":` + c.shell + `}`))
+		if err == nil || !strings.Contains(err.Error(), c.want) {
+			t.Errorf("%s: want %q, got %v", c.shell, c.want, err)
+		}
+	}
+	// The code itself is never judged: a shell's own `${…}` is not the mesh's.
+	if _, err := ParseManifest([]byte(`{"module":"tool","shell":[{"for":"zsh","slot":"first",` +
+		`"code":"source \"${XDG_CACHE_HOME:-$HOME/.cache}/p10k-instant-prompt-${(%):-%n}.zsh\""}]}`)); err != nil {
+		t.Fatalf("shell code was judged as if it were the mesh's: %v", err)
+	}
+}
+
+// ADR 0203 §1 and ADR 0204 §1: both seats are the mesh's own, held once per machine; the login
+// shell's contract is `execute`, described and with a schema an agent can call.
+func TestTheSeatTableCarriesTheEnvironmentAndTheLoginShell(t *testing.T) {
+	env, ok := SeatNamed("node-environment")
+	if !ok || env.Scope != ScopeNode || env.Decision != "novox/hq ADR 0203" ||
+		len(env.Serves)+len(env.Accepts)+len(env.Emits) != 0 || env.Delivers != "" {
+		t.Fatalf("node-environment is not a node seat with no protocol: %+v (defined %v)", env, ok)
+	}
+	shell, ok := SeatNamed("node-login-shell")
+	if !ok || shell.Scope != ScopeNode || shell.Decision != "novox/hq ADR 0204" {
+		t.Fatalf("node-login-shell is not a node seat: %+v (defined %v)", shell, ok)
+	}
+	if len(shell.Serves) != 1 || shell.Serves[0].Name != "execute" || shell.Serves[0].Description == "" {
+		t.Fatalf("the login shell serves %+v, not execute alone", shell.Serves)
+	}
+	props, _ := shell.Serves[0].Input["properties"].(map[string]any)
+	required, _ := shell.Serves[0].Input["required"].([]string)
+	if _, has := props["command"]; !has || len(required) != 1 || required[0] != "command" {
+		t.Fatalf("execute does not require a command: %v", shell.Serves[0].Input)
+	}
+	if _, has := props["timeout_seconds"]; !has {
+		t.Fatalf("execute takes no timeout: %v", props)
+	}
+}
+
+// ADR 0204 §1: the login shell is the mesh's, so no module declares it — neither under the mesh's
+// name nor under the name a module gave it before.
+func TestNoModuleMayDeclareTheLoginShell(t *testing.T) {
+	for _, n := range []string{"login-shell", "node-login-shell", "node-environment"} {
+		raw := `{"module":"zsh","seats":[{"name":"` + n + `","scope":"node","serves":["execute"]}],"tools":["execute"]}`
+		_, err := ParseManifest([]byte(raw))
+		if err == nil {
+			t.Errorf("a module declaring %q was accepted", n)
+		}
+	}
+	got := strings.Join(declaredSeatProblems(Manifest{Module: "zsh",
+		DefinesSeats: []SeatDeclaration{{Name: "login-shell", Scope: ScopeNode}}}), "; ")
+	if !strings.Contains(got, `zsh declares a seat named "login-shell"; the login shell is the mesh's own seat node-login-shell`) {
+		t.Fatalf("declaring login-shell was not refused by name: %q", got)
+	}
+	// And a shell module claiming the mesh's seat, serving execute, is what the seat is for.
+	m, err := ParseManifest([]byte(`{"module":"zsh","tools":["execute"],` +
+		`"claims":[{"name":"node-login-shell","scope":"node"}]}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := CanHold(m, Seat{Name: LoginShellSeat, Scope: ScopeNode, Serves: loginShellVerbs()}); err != nil {
+		t.Fatalf("a shell module claiming the seat cannot hold it: %v", err)
+	}
+}

internal/catalogue/genesis_image_test.go

@@ -0,0 +1,29 @@
+package catalogue
+
+import (
+	"os"
+	"regexp"
+	"testing"
+)
+
+// novox/hq issue 223: genesis raises the controller as a container built from this repository's own
+// Dockerfile, with no build arguments — the manifest no longer builds an image, so nothing passes a
+// base in. The Dockerfile's own default must therefore be a Go that builds this module, pinned by
+// digest, and the replacement the manifest's process names must be the container genesis raises.
+func TestGenesisCanBuildTheControllersImageAsItStands(t *testing.T) {
+	raw, err := os.ReadFile("../../Dockerfile")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !regexp.MustCompile(`(?m)^ARG GO_BASE=golang@sha256:[0-9a-f]{64}$`).Match(raw) {
+		t.Fatal("the Dockerfile's default Go base is not pinned by digest; genesis builds it with no arguments")
+	}
+	makefile, err := os.ReadFile("../../Makefile")
+	if err != nil {
+		t.Fatal(err)
+	}
+	pin := regexp.MustCompile(`golang@sha256:[0-9a-f]{64}`)
+	if string(pin.Find(raw)) != string(pin.Find(makefile)) {
+		t.Errorf("the Dockerfile and the Makefile build on different Go: %s, %s", pin.Find(raw), pin.Find(makefile))
+	}
+}

internal/catalogue/manifest.go

@@ -325,6 +325,15 @@ type Manifest struct {
 	// person's account (design 25 §7) already had the same shape.
 	Invokes []string `json:"invokes,omitempty"`
 
+	// State is the current state this module keeps on the bus, by local name: each a key-value
+	// bucket the controller creates, which every instance of the module writes and reads
+	// (novox/hq ADR 0201). Not history — that is an event — and never a secret, sealed or not.
+	State []StateDeclaration `json:"state,omitempty"`
+
+	// Reads are other modules' state this module reads and watches, each `<module>.<name>`
+	// (novox/hq ADR 0201). Read-only: only the owner's instances write.
+	Reads []string `json:"reads,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
 	// machine.
@@ -508,6 +517,17 @@ type Manifest struct {
 	// holder. Like Filtering: one module per node gathers what every module declared and writes it.
 	Jailing *Jailing `json:"jailing,omitempty"`
 
+	// Environment is what this module adds to the operator account's environment: variables, and
+	// entries on PATH (novox/hq ADR 0203). Facts, not lines of one shell's syntax — the holder of
+	// node-environment places them, and the controller writes them in each reader's format. Like
+	// Jails: any module contributes, gathered from every module on the node, written by the holder.
+	Environment *Environment `json:"environment,omitempty"`
+
+	// Shell is code this module adds to the login shell's startup, for a named shell in a named
+	// slot (novox/hq ADR 0204). The controller never reads it: it is placed, in module order, where
+	// the holder of node-login-shell put the slot's placeholder.
+	Shell []ShellCode `json:"shell,omitempty"`
+
 	// Guards are ports of this module's the mesh refuses on an adopted node except from the
 	// private network and from the machine itself (novox/hq ADR 0100) — the store's port and the
 	// broker's management port. The ports the software uses; the mesh guards where the machine
@@ -1316,6 +1336,8 @@ func ParseManifest(raw []byte) (Manifest, error) {
 	// 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)...)
+	// And what it may call its state, and whose it may read (state.go, novox/hq ADR 0201).
+	problems = append(problems, StateProblems(m)...)
 	wellFormed := true
 	for _, c := range m.Claims {
 		if !name.MatchString(c.Name) {
@@ -1428,6 +1450,10 @@ func ParseManifest(raw []byte) (Manifest, error) {
 				"%s serves %q to whoever requires it, and does not provide it", m.Module, to))
 		}
 	}
+	// A served value may be derived for the consumer it is served to (novox/hq ADR 0201). Read
+	// here, where the definition is, rather than when somebody first requires it: a rule that
+	// would be refused at the first consumer is wrong from the moment it is written.
+	problems = append(problems, CheckServes(m)...)
 	for to, where := range m.Binds {
 		if !placedOrAbsolute(where) {
 			problems = append(problems, fmt.Sprintf(
@@ -1622,6 +1648,13 @@ func ParseManifest(raw []byte) (Manifest, error) {
 				}
 			}
 		}
+		// **A scheduled step may hold this module's own containers still while it runs**
+		// (novox/hq ADR 0189). What the host judges is the declaration it receives — whether each
+		// id is a container placed on that machine; what belongs here is what only the definition
+		// shows: that the ids are this module's, that they are containers, and that the step is
+		// scheduled. A module naming a neighbour's container would be a module that can stop the
+		// mesh, and the manifest is where that is visible.
+		problems = append(problems, whileStoppedProblems(m, r, hasSchedule(r))...)
 	}
 	for name, own := range m.OwnSecrets {
 		if !placedOrAbsolute(own.Path) {
@@ -1783,6 +1816,12 @@ func ParseManifest(raw []byte) (Manifest, error) {
 	problems = append(problems, m.unknownDirRefs()...)
 	problems = append(problems, m.unknownAccessRefs()...)
 	problems = append(problems, m.jailProblems()...)
+	// What a module adds to the account's environment and to the login shell, and the holder's
+	// placeholders for them (novox/hq ADR 0203, ADR 0204) — here, so the catalogue check refuses
+	// them in the words registration does.
+	problems = append(problems, m.environmentProblems()...)
+	problems = append(problems, m.shellProblems()...)
+	problems = append(problems, m.contributionPlaceholderProblems()...)
 
 	for i, r := range m.Resources {
 		id, _ := r["id"].(string)
@@ -2153,3 +2192,71 @@ func (o OwnSecrets) Paths() map[string]string {
 // InstancesInterchangeable is the one value of a definition's `instances`: the module is the same
 // on every machine, so any instance may answer for the module.
 const InstancesInterchangeable = "interchangeable"
+
+// WhileStopped is the resource key naming the containers a scheduled step holds still while it
+// runs (novox/hq ADR 0189). Carried to the host unchanged, like `schedule`.
+const WhileStopped = "while-stopped"
+
+// hasSchedule is whether a resource declares a cadence, as a string.
+func hasSchedule(r map[string]any) bool {
+	s, _ := r["schedule"].(string)
+	return s != ""
+}
+
+// whileStoppedProblems judges one container's maintenance window against its own definition
+// (novox/hq ADR 0189).
+//
+// Three things the manifest is the only place to see: that the step is scheduled (a one-time
+// offline job says *before* rather than *instead of* — at apply the host already has a window,
+// because the declaration is applied in order and a run-once step gates what follows); that every
+// id it names is **this module's own** container; and that it does not name itself.
+//
+// The host checks the fourth — that the container is actually placed on that machine — because
+// that is a fact about the declaration and not about the definition.
+func whileStoppedProblems(m Manifest, r map[string]any, scheduled bool) []string {
+	raw, present := r[WhileStopped]
+	if !present {
+		return nil
+	}
+	ids, ok := raw.([]any)
+	if !ok {
+		return []string{fmt.Sprintf(
+			"%s declares %s on %v as a %T; it is a list of this module's container ids",
+			m.Module, WhileStopped, r["id"], raw)}
+	}
+	var problems []string
+	if len(ids) > 0 && !scheduled {
+		problems = append(problems, fmt.Sprintf(
+			"%s declares %s on %v, which has no schedule. A maintenance window is for a recurring "+
+				"step: at apply the mesh already has one, because a run-once step gates what is "+
+				"declared after it (novox/hq ADR 0189)", m.Module, WhileStopped, r["id"]))
+	}
+	containers := map[string]bool{}
+	for _, own := range m.Resources {
+		if fmt.Sprint(own["type"]) == "container" {
+			containers[fmt.Sprint(own["id"])] = true
+		}
+	}
+	for _, each := range ids {
+		id, ok := each.(string)
+		if !ok {
+			problems = append(problems, fmt.Sprintf(
+				"%s declares %s on %v naming a %T; each entry is a container's id",
+				m.Module, WhileStopped, r["id"], each))
+			continue
+		}
+		if id == fmt.Sprint(r["id"]) {
+			problems = append(problems, fmt.Sprintf(
+				"%s declares %s on %v naming itself", m.Module, WhileStopped, r["id"]))
+			continue
+		}
+		if !containers[id] {
+			problems = append(problems, fmt.Sprintf(
+				"%s declares %s on %v naming %q, which is not a container this module declares. "+
+					"A step may hold still its own module's containers and nobody else's — one "+
+					"that could quiesce a neighbour could stop the mesh",
+				m.Module, WhileStopped, r["id"], id))
+		}
+	}
+	return problems
+}

internal/catalogue/seat_into_test.go

@@ -101,7 +101,7 @@ func TestTheControlPlanesOwnAddressesFollowTheNodesPorts(t *testing.T) {
 	}
 	// The manifest itself names them now; withSeatPorts is a no-op on it, and this holds it so.
 	for _, r := range m.Resources {
-		if r["type"] != "container" {
+		if r["type"] != "process" {
 			continue
 		}
 		env, _ := r["env"].(map[string]any)
@@ -113,8 +113,9 @@ func TestTheControlPlanesOwnAddressesFollowTheNodesPorts(t *testing.T) {
 	}
 	m = withSeatPorts(m)
 	control, err := m.Resolve([]Built{{
-		Name: "server", Kind: ArtifactImage,
+		Name: "controller", Kind: ArtifactBundle,
-		Reference: ArtifactStoreScheme + "mesh-controller/server@sha256:" + strings.Repeat("c", 64),
+		Reference: ArtifactStoreScheme + "mesh-controller/controller@sha256:" + strings.Repeat("c", 64),
+		Digest:    "sha256:" + strings.Repeat("c", 64),
 	}})
 	if err != nil {
 		t.Fatal(err)
@@ -135,9 +136,9 @@ func TestTheControlPlanesOwnAddressesFollowTheNodesPorts(t *testing.T) {
 	if err != nil {
 		t.Fatalf("the control plane does not compose: %v", err)
 	}
-	server := fileNamed(out, "mesh-controller.server")
+	server := fileNamed(out, "mesh-controller.controller")
 	if server == nil {
-		t.Fatalf("the control plane's container is not in the declaration: %v", out)
+		t.Fatalf("the control plane's process is not in the declaration: %v", out)
 	}
 	env, _ := server["env"].(map[string]any)
 	for key, want := range map[string]string{
@@ -151,8 +152,8 @@ func TestTheControlPlanesOwnAddressesFollowTheNodesPorts(t *testing.T) {
 			t.Errorf("the control plane is told %s=%v; the node put it on %s", key, env[key], want)
 		}
 	}
-	if got := server["image"]; got != "anchor.internal:5100/mesh-controller/server@sha256:"+strings.Repeat("c", 64) {
+	if got := server["source"]; got != "anchor.internal:5100/mesh-controller/controller@sha256:"+strings.Repeat("c", 64) {
-		t.Errorf("the control plane's own image is %v, not routed through the store", got)
+		t.Errorf("the control plane's own bundle is fetched from %v, not routed through the store", got)
 	}
 
 	// And on a mesh where the foundation is where genesis raised it, nothing is added.
@@ -160,7 +161,7 @@ func TestTheControlPlanesOwnAddressesFollowTheNodesPorts(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	env, _ = fileNamed(out, "mesh-controller.server")["env"].(map[string]any)
+	env, _ = fileNamed(out, "mesh-controller.controller")["env"].(map[string]any)
 	if env["MESH_STORE_INVENTORY_PORT"] != "" {
 		t.Errorf("with no settings, the control plane is told %v", env)
 	}
@@ -186,7 +187,7 @@ func withSeatPorts(m Manifest) Manifest {
 	out := m
 	out.Resources = nil
 	for _, r := range m.Resources {
-		if r["type"] != "container" {
+		if r["type"] != "container" && r["type"] != "process" {
 			out.Resources = append(out.Resources, r)
 			continue
 		}

internal/catalogue/seats.go

@@ -150,6 +150,17 @@ var defaultSeats = []Seat{
 	// served by the node tools runtime (ADR 0175).
 	{Name: "node-service-manager", Scope: ScopeNode, Decision: "novox/hq ADR 0177",
 		Serves: serviceManagerVerbs()},
+	// The operator account's environment (novox/hq ADR 0203): one module per machine writes it, and
+	// every module contributes to it. No verbs — the seat says who places the environment's files,
+	// and their path is its protocol: a shell sources ~/.config/mesh/environment.sh without knowing
+	// which module wrote it.
+	{Name: EnvironmentSeat, Scope: ScopeNode, Decision: "novox/hq ADR 0203"},
+	// The login shell (novox/hq ADR 0204, replacing the module-declared `login-shell` of ADR 0176):
+	// the mesh's, so a second shell module claims the seat rather than declaring a second one, and
+	// the seat exists whether or not zsh's definition is registered. `execute` is the contract any
+	// node may call; the holder places every module's shell code in its slots.
+	{Name: LoginShellSeat, Scope: ScopeNode, Decision: "novox/hq ADR 0204",
+		Serves: loginShellVerbs()},
 	// Deferred (novox/hq ADR 0121): renaming to mesh-private-network is a scope + server/client
 	// model change, not a rename, so it stays until that is built.
 	{Name: "the-private-network", Scope: ScopeNode, Decision: "novox/hq ADR 0110"},
@@ -456,3 +467,17 @@ func serviceManagerVerbs() []Verb {
 			Input: scoped(map[string]string{"unit": unit["unit"], "lines": "how many lines from the end (default 100)"}, []string{"unit"})},
 	}
 }
+
+// loginShellVerbs is the contract every holder of node-login-shell serves (novox/hq ADR 0176, ADR
+// 0204): one command, run the way the operator's own terminal would run it, bounded below the
+// runtime's thirty-second call limit so a hung command answers rather than times the caller out.
+func loginShellVerbs() []Verb {
+	return []Verb{
+		{Name: "execute", Description: "Run one command on this machine as the operator account, in a " +
+			"non-interactive login shell in its home; answers with what it printed and how it exited.",
+			Input: schema(map[string]string{
+				"command":         "the command line, as you would type it",
+				"timeout_seconds": "give up after this long, at most 25 (default 20)",
+			}, []string{"command"})},
+	}
+}

internal/catalogue/seats_declared.go

@@ -25,6 +25,10 @@ import (
 // and nothing to keep in step when a mesh seat is added.
 const meshSeatPrefix = "mesh-"
 
+// retiredLoginShell is the one name outside the prefix a module may not declare: the login shell's,
+// from when a module declared it (novox/hq ADR 0176), before it became the mesh's (ADR 0204).
+const retiredLoginShell = "login-shell"
+
 // 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.
@@ -88,6 +92,15 @@ func declaredSeatProblems(m Manifest) []string {
 					"seats (novox/hq ADR 0118)", m.Module, s.Name, meshSeatPrefix+"*"))
 			continue
 		}
+		if s.Name == retiredLoginShell {
+			// The name ADR 0176 gave the login shell when the zsh module declared it. The seat is
+			// the mesh's now, so a module declaring the old name would be a second login shell
+			// beside it, with a protocol of its own (novox/hq ADR 0204).
+			problems = append(problems, fmt.Sprintf(
+				"%s declares a seat named %q; the login shell is the mesh's own seat %s, which a shell "+
+					"module claims and none declares (novox/hq ADR 0204)", m.Module, s.Name, LoginShellSeat))
+			continue
+		}
 		if seen[s.Name] {
 			problems = append(problems, fmt.Sprintf(
 				"%s declares the seat %q twice", m.Module, s.Name))
@@ -215,6 +228,13 @@ func CatalogueProblems(shelf Shelf) []string {
 			}
 		}
 	}
+	// A read of a module's state that module does not keep (novox/hq ADR 0201) — said only where the
+	// owner is on the shelf, as a consumer may be installed before its emitter.
+	var manifests []Manifest
+	for _, module := range shelfOrder(shelf) {
+		manifests = append(manifests, shelf[module])
+	}
+	problems = append(problems, StateReadsNothingDeclares(manifests)...)
 	sort.Strings(problems)
 	return problems
 }

internal/catalogue/seats_test.go

@@ -44,10 +44,11 @@ func TestTheSeatsAreAClosedSetAndEachNamesItsDecision(t *testing.T) {
 			delivered[s.Delivers] = s.Name
 		}
 	}
-	// Seventeen since node-build-agent (novox/hq ADR 0190) — sixteen once the retired
+	// Nineteen since node-environment and node-login-shell (novox/hq ADR 0203, ADR 0204), after
-	// mesh-build-machine row goes, when no registered manifest claims it any more.
+	// node-build-agent made seventeen (ADR 0190) — eighteen once the retired mesh-build-machine row
-	if len(Seats()) != 17 {
+	// goes, when no registered manifest claims it any more.
-		t.Errorf("the mesh defines %d seats rather than 17; the set is closed, so a change here is "+
+	if len(Seats()) != 19 {
+		t.Errorf("the mesh defines %d seats rather than 19; the set is closed, so a change here is "+
 			"a decision (novox/hq ADR 0110): %s", len(Seats()), seatNames())
 	}
 }

internal/catalogue/state.go

@@ -0,0 +1,150 @@
+package catalogue
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"regexp"
+	"strings"
+)
+
+// What a module may call its state, and whose state it may ask to read (novox/hq ADR 0201).
+//
+// A module names its state **locally** — `servers`, never a bucket or a subject — and another
+// module's as `<module>.<name>`, the way a consumed event names its emitter (design 32 §1). The
+// mesh derives the bucket from the two names, so the module and the local name must each be one
+// token: the bucket joins them with an underscore, which neither may contain, so two modules can
+// never derive one bucket.
+
+// stateName is one local name of a module's state: lower-case, no dot, no underscore.
+var stateName = regexp.MustCompile(`^[a-z0-9][a-z0-9-]*$`)
+
+// The mesh's caps on what a module may ask of a bucket's history.
+const (
+	// StateMostHistory is the most past values a key may keep. The server's own limit.
+	StateMostHistory = 64
+)
+
+// StateDeclaration is one bucket a module owns: its local name, and the options that are the
+// owner's to choose, as a seat chooses how long its backlog survives (design 32 §3).
+type StateDeclaration struct {
+	Name string `json:"name"`
+	// History is how many values a key keeps, the current one included; zero is one.
+	History int `json:"history,omitempty"`
+	// TTLSeconds is how long a value lives once written; zero is until it is replaced or deleted.
+	TTLSeconds int `json:"ttl-seconds,omitempty"`
+}
+
+// UnmarshalJSON reads a bucket as its bare name, or as {name, history, ttl-seconds}.
+func (s *StateDeclaration) UnmarshalJSON(raw []byte) error {
+	trimmed := bytes.TrimSpace(raw)
+	if len(trimmed) > 0 && trimmed[0] == '"' {
+		return json.Unmarshal(trimmed, &s.Name)
+	}
+	type plain StateDeclaration
+	var full plain
+	dec := json.NewDecoder(bytes.NewReader(trimmed))
+	dec.DisallowUnknownFields()
+	if err := dec.Decode(&full); err != nil {
+		return fmt.Errorf("a state is either a name or {name, history, ttl-seconds}: %w", err)
+	}
+	*s = StateDeclaration(full)
+	return nil
+}
+
+// MarshalJSON writes back the short form when there is nothing else to say.
+func (s StateDeclaration) MarshalJSON() ([]byte, error) {
+	if s.History == 0 && s.TTLSeconds == 0 {
+		return json.Marshal(s.Name)
+	}
+	type plain StateDeclaration
+	return json.Marshal(plain(s))
+}
+
+// ReadState splits a read into the owning module and the local name, or says why it is not one.
+func ReadState(read string) (module, local string, err error) {
+	at := strings.LastIndex(read, ".")
+	if at <= 0 || at == len(read)-1 {
+		return "", "", fmt.Errorf("%q does not name a module and its state: a read is <module>.<name>", read)
+	}
+	module, local = read[:at], read[at+1:]
+	if !stateName.MatchString(module) {
+		return "", "", fmt.Errorf("%q cannot own state: a module whose state is read is one plain name", module)
+	}
+	if !stateName.MatchString(local) {
+		return "", "", fmt.Errorf("%q is not a state name: lower-case letters, digits and hyphens", local)
+	}
+	return module, local, nil
+}
+
+// StateProblems is what is wrong with a manifest's state and reads.
+//
+// Refused at registration, because a bucket name the bus cannot hold is a module that installs,
+// starts, and is refused on its first write with a reason about a bucket nobody named.
+func StateProblems(m Manifest) []string {
+	var problems []string
+	if len(m.State) > 0 && !stateName.MatchString(m.Module) {
+		problems = append(problems, fmt.Sprintf(
+			"%s keeps state, and a module's name is part of its buckets' names, which take one plain "+
+				"name — no dot (novox/hq ADR 0201)", m.Module))
+	}
+	seen := map[string]bool{}
+	for _, s := range m.State {
+		switch {
+		case !stateName.MatchString(s.Name):
+			problems = append(problems, fmt.Sprintf(
+				"%s keeps state %q: a state is named locally — lower-case letters, digits and hyphens, "+
+					"no dot and no underscore; the mesh derives the bucket (novox/hq ADR 0201)", m.Module, s.Name))
+		case seen[s.Name]:
+			problems = append(problems, fmt.Sprintf("%s keeps state %q twice", m.Module, s.Name))
+		}
+		seen[s.Name] = true
+		if s.History < 0 || s.History > StateMostHistory {
+			problems = append(problems, fmt.Sprintf(
+				"%s keeps %d values of %q; a key keeps between 1 and %d", m.Module, s.History, s.Name, StateMostHistory))
+		}
+		if s.TTLSeconds < 0 {
+			problems = append(problems, fmt.Sprintf("%s gives %q a negative lifetime", m.Module, s.Name))
+		}
+	}
+	for _, r := range m.Reads {
+		module, _, err := ReadState(r)
+		if err != nil {
+			problems = append(problems, fmt.Sprintf("%s reads %v", m.Module, err))
+			continue
+		}
+		if module == m.Module {
+			problems = append(problems, fmt.Sprintf(
+				"%s reads %q, which is its own state: a module reads and writes what it keeps already", m.Module, r))
+		}
+	}
+	return problems
+}
+
+// StateReadsNothingDeclares is every read across a catalogue whose owner is present and declares no
+// such state. An absent owner says nothing — a module may be installed long before the one whose
+// state it reads, as a consumer may before its emitter (design 32 §1).
+func StateReadsNothingDeclares(manifests []Manifest) []string {
+	declared := map[string]map[string]bool{}
+	for _, m := range manifests {
+		own := map[string]bool{}
+		for _, s := range m.State {
+			own[s.Name] = true
+		}
+		declared[m.Module] = own
+	}
+	var problems []string
+	for _, m := range manifests {
+		for _, r := range m.Reads {
+			module, local, err := ReadState(r)
+			if err != nil {
+				continue
+			}
+			if own, present := declared[module]; present && !own[local] {
+				problems = append(problems, fmt.Sprintf(
+					"%s reads %q, and %s keeps no state called %q", m.Module, r, module, local))
+			}
+		}
+	}
+	return problems
+}

internal/catalogue/state_test.go

@@ -0,0 +1,94 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// A module declares the state it keeps and the state it reads (novox/hq ADR 0201), a bucket by its
+// bare name or with the owner's options.
+func TestAManifestMaySayWhatStateItKeepsAndReads(t *testing.T) {
+	m, err := ParseManifest([]byte(`{"module":"claude-code","version":"1",` +
+		`"state":["servers",{"name":"seen","history":5,"ttl-seconds":3600}],` +
+		`"reads":["licence-manager.bindings"]}`))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(m.State) != 2 || m.State[0].Name != "servers" || m.State[1].History != 5 || m.State[1].TTLSeconds != 3600 {
+		t.Fatalf("state not read: %+v", m.State)
+	}
+	if len(m.Reads) != 1 || m.Reads[0] != "licence-manager.bindings" {
+		t.Fatalf("reads not read: %v", m.Reads)
+	}
+	// Written back as it came in: the short form where nothing else is said.
+	out, _ := json.Marshal(m.State)
+	if string(out) != `["servers",{"name":"seen","history":5,"ttl-seconds":3600}]` {
+		t.Fatalf("written back as %s", out)
+	}
+}
+
+// A name the bus could not hold, or that would let two modules derive one bucket, is refused at
+// registration in the manifest's words.
+func TestAStateNameIsLocalAndOneToken(t *testing.T) {
+	for _, c := range []struct{ manifest, says string }{
+		{`{"module":"a","version":"1","state":["mesh.servers"]}`, `keeps state "mesh.servers": a state is named locally`},
+		{`{"module":"a","version":"1","state":["my_servers"]}`, `keeps state "my_servers"`},
+		{`{"module":"a","version":"1","state":["s","s"]}`, `keeps state "s" twice`},
+		{`{"module":"a","version":"1","state":[{"name":"s","history":65}]}`, `a key keeps between 1 and 64`},
+		{`{"module":"a.b","version":"1","state":["s"]}`, `no dot`},
+		{`{"module":"a","version":"1","reads":["bindings"]}`, `a read is <module>.<name>`},
+		{`{"module":"a","version":"1","reads":["a.s"]}`, `which is its own state`},
+		{`{"module":"a","version":"1","state":[{"name":"s","shared":true}]}`, `{name, history, ttl-seconds}`},
+	} {
+		_, err := ParseManifest([]byte(c.manifest))
+		if err == nil {
+			t.Errorf("%s was accepted", c.manifest)
+			continue
+		}
+		if !strings.Contains(err.Error(), c.says) {
+			t.Errorf("%s refused for the wrong reason: %v", c.manifest, err)
+		}
+	}
+}
+
+// A read whose owner is present must name a state that owner keeps; an absent owner says nothing,
+// because a module may be installed before the one whose state it reads.
+func TestAReadNamesStateItsOwnerKeeps(t *testing.T) {
+	owner := Manifest{Module: "licence-manager", State: []StateDeclaration{{Name: "bindings"}}}
+	good := Manifest{Module: "claude-code", Reads: []string{"licence-manager.bindings", "absent.anything"}}
+	bad := Manifest{Module: "other", Reads: []string{"licence-manager.tokens"}}
+	if p := StateReadsNothingDeclares([]Manifest{owner, good}); len(p) != 0 {
+		t.Fatalf("a read of declared state was refused: %v", p)
+	}
+	p := StateReadsNothingDeclares([]Manifest{owner, bad})
+	if len(p) != 1 || !strings.Contains(p[0], `licence-manager keeps no state called "tokens"`) {
+		t.Fatalf("a read of state nobody keeps was not named: %v", p)
+	}
+}
+
+// **Across the whole catalogue**: every state name is local, and every read whose owner is present
+// names state that owner keeps.
+func TestEveryManifestsStateIsLocalAndEveryReadIsKept(t *testing.T) {
+	manifests := theCatalogue(t)
+	var problems []string
+	for _, m := range manifests {
+		problems = append(problems, StateProblems(m)...)
+	}
+	problems = append(problems, StateReadsNothingDeclares(manifests)...)
+	if len(problems) > 0 {
+		t.Fatalf("the catalogue's state is not what ADR 0201 says:\n  %s", strings.Join(problems, "\n  "))
+	}
+}
+
+// `module check` says it too: the cross-catalogue pass names a read nothing on the shelf keeps.
+func TestTheCataloguePassNamesAReadItsOwnerDoesNotKeep(t *testing.T) {
+	shelf := Shelf{
+		"licence-manager": {Module: "licence-manager", State: []StateDeclaration{{Name: "bindings"}}},
+		"claude-code":     {Module: "claude-code", Reads: []string{"licence-manager.tokens"}},
+	}
+	problems := CatalogueProblems(shelf)
+	if len(problems) != 1 || !strings.Contains(problems[0], `keeps no state called "tokens"`) {
+		t.Fatalf("the catalogue pass said %v", problems)
+	}
+}

internal/catalogue/while_stopped_test.go

@@ -0,0 +1,145 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"fmt"
+	"strings"
+	"testing"
+)
+
+// A scheduled step may hold its module's own containers still while it runs (novox/hq ADR 0189).
+//
+// The host judges what it receives — whether each id is a container on that machine. What the
+// definition is the only place to see is judged here, near whoever wrote it.
+
+func aStoreManifest(step map[string]any) []byte {
+	m := map[string]any{
+		"module": "distribution", "version": "1",
+		"resources": []any{
+			map[string]any{"id": "store", "type": "container", "name": "mesh-registry",
+				"image": "registry@sha256:" + strings.Repeat("a", 64)},
+			step,
+		},
+	}
+	raw, _ := json.Marshal(m)
+	return raw
+}
+
+func TestAMaintenanceWindowOnItsOwnModulesContainerIsAccepted(t *testing.T) {
+	raw := aStoreManifest(map[string]any{
+		"id": "collect", "type": "container", "name": "mesh-registry-collect",
+		"image":    "registry@sha256:" + strings.Repeat("a", 64),
+		"schedule": "30 3 * * *", "while-stopped": []any{"store"},
+	})
+	if _, err := ParseManifest(raw); err != nil {
+		t.Fatalf("a step holding its own module's container still was refused: %v", err)
+	}
+}
+
+func TestAMaintenanceWindowIsRefusedWhereTheDefinitionShowsItCannotMean(t *testing.T) {
+	for _, c := range []struct {
+		name string
+		step map[string]any
+		says string
+	}{
+		{
+			"on a step with no schedule",
+			map[string]any{"id": "collect", "type": "container", "name": "c",
+				"image":         "registry@sha256:" + strings.Repeat("a", 64),
+				"while-stopped": []any{"store"}},
+			"gates what is declared after it",
+		},
+		{
+			"on a run-once step, which already has order",
+			map[string]any{"id": "collect", "type": "container", "name": "c",
+				"image":    "registry@sha256:" + strings.Repeat("a", 64),
+				"run-once": true, "while-stopped": []any{"store"}},
+			"A maintenance window is for a recurring step",
+		},
+		{
+			"naming a container this module does not declare",
+			map[string]any{"id": "collect", "type": "container", "name": "c",
+				"image":    "registry@sha256:" + strings.Repeat("a", 64),
+				"schedule": "30 3 * * *", "while-stopped": []any{"the-broker"}},
+			"could quiesce a neighbour could stop the mesh",
+		},
+		{
+			"naming itself",
+			map[string]any{"id": "collect", "type": "container", "name": "c",
+				"image":    "registry@sha256:" + strings.Repeat("a", 64),
+				"schedule": "30 3 * * *", "while-stopped": []any{"collect"}},
+			"naming itself",
+		},
+		{
+			"written as something that is not a list",
+			map[string]any{"id": "collect", "type": "container", "name": "c",
+				"image":    "registry@sha256:" + strings.Repeat("a", 64),
+				"schedule": "30 3 * * *", "while-stopped": "store"},
+			"a list of this module's container ids",
+		},
+	} {
+		_, err := ParseManifest(aStoreManifest(c.step))
+		if err == nil {
+			t.Errorf("%s was accepted", c.name)
+			continue
+		}
+		if !strings.Contains(err.Error(), c.says) {
+			t.Errorf("%s: the refusal does not say %q:\n%v", c.name, c.says, err)
+		}
+	}
+}
+
+// A composed declaration names the step's held containers the way the machine knows them.
+//
+// **The gap that let a bug through to the control node.** The manifest says `while-stopped:
+// ["store"]`, because a module names its own resources locally; the declaration a machine
+// receives calls that container `distribution.store`, because every resource is composed under
+// its module. `restart-on` and `reload-on` are rewritten for exactly this reason, and
+// `while-stopped` was not — so the host found no container by that id and refused the whole
+// declaration, every push, until it was fixed.
+//
+// It passed every test on both sides: the controller's tests read manifests, the host's read
+// hand-written declarations with bare ids. Only composing one and judging the result catches it.
+func TestAComposedWindowNamesTheContainerAsTheMachineKnowsIt(t *testing.T) {
+	store := Manifest{
+		Module: "distribution", Version: "1",
+		Provides: FromAnywhere("artifact-store"),
+		Listens:  []Listening{{Port: 5000, Protocol: "tcp", From: FromMesh}},
+		Serves:   map[string]map[string]any{"artifact-store": {"port": 5000}},
+		Resources: []map[string]any{
+			{"id": "store", "type": "container", "name": "mesh-registry",
+				"image": "registry@sha256:" + strings.Repeat("a", 64), "ports": []any{"5000"}},
+			{"id": "collect", "type": "container", "name": "mesh-registry-collect",
+				"image":    "registry@sha256:" + strings.Repeat("a", 64),
+				"schedule": "30 3 * * *", WhileStopped: []any{"store"}},
+		},
+	}
+	r, err := Resolve(shelf(store), []string{"distribution"}, reachable(), World{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	out, err := r.Declaration(Rendering{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	collect := fileNamed(out, "distribution.collect")
+	if collect == nil {
+		for _, res := range out {
+			if res["id"] == "distribution.collect" {
+				collect = res
+			}
+		}
+	}
+	if collect == nil {
+		t.Fatalf("the step was not composed at all: %v", out)
+	}
+	held, _ := collect[WhileStopped].([]any)
+	if len(held) != 1 {
+		t.Fatalf("the composed step holds %v still; want one container", collect[WhileStopped])
+	}
+	if got := fmt.Sprint(held[0]); got != "distribution.store" {
+		t.Fatalf("the composed step says it holds %q still, and the machine's container is "+
+			"called %q — the host refuses a declaration naming a container it does not have, "+
+			"whole, so the machine would take nothing at all", got, "distribution.store")
+	}
+}

internal/inventory/busrecords.go

@@ -132,6 +132,9 @@ func declaredFor(m catalogue.Manifest, seats map[string]catalogue.SeatDeclaratio
 		Serves: m.Tools,
 		// And what it calls (novox/hq ADR 0152) — the console's `*`, nothing else's.
 		Invokes: m.Invokes,
+		// And the state it keeps and reads (novox/hq ADR 0201).
+		State: bucketsOf(m),
+		Reads: m.Reads,
 	}
 	for _, c := range m.Claims {
 		// Every seat with a protocol, the mesh's own included. One that says only who does a job is
@@ -148,6 +151,30 @@ func declaredFor(m catalogue.Manifest, seats map[string]catalogue.SeatDeclaratio
 	return d
 }
 
+// bucketsOf is the state a module keeps, as the bus holds it.
+func bucketsOf(m catalogue.Manifest) []broker.Bucket {
+	var out []broker.Bucket
+	for _, s := range m.State {
+		out = append(out, broker.Bucket{Module: m.Module, Name: s.Name, History: s.History, TTLSeconds: s.TTLSeconds})
+	}
+	return out
+}
+
+// DeclaredBuckets is every bucket the catalogue declares, registered modules assigned or not: a
+// bucket exists from registration, like a seat's stream, so a module reading it may watch before its
+// owner runs anywhere (novox/hq ADR 0201).
+func (i *Inventory) DeclaredBuckets(ctx context.Context) ([]broker.Bucket, error) {
+	declared, err := i.Catalogue(ctx)
+	if err != nil {
+		return nil, fmt.Errorf("cannot read the catalogue: %w", err)
+	}
+	var out []broker.Bucket
+	for _, m := range declared {
+		out = append(out, bucketsOf(m)...)
+	}
+	return out, nil
+}
+
 func asSeat(s catalogue.SeatDeclaration) broker.Seat {
 	return broker.Seat{Name: s.Name, Scope: s.Scope, Accepts: s.Accepts, Emits: s.Emits,
 		Serves: catalogue.VerbNames(s.Serves)}

internal/inventory/catalogue.go

@@ -78,11 +78,12 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
 	// (novox/hq ADR 0175, to-be 38 WP2.4): a module serving its tools from a container built on the
 	// runtime's image. Refused at registration, by name, for a module that is new to the catalogue
 	// or that was registered in another shape — the mechanism that keeps the old pattern from
-	// returning by habit. **Not refused for a module already registered in that shape**: the
+	// returning by habit. Before the runtime exists the pattern is accepted as it always was.
-	// catalogue holds some thirty of them the day the runtime arrives, each moves to a bundle in
+	//
-	// its own change (to-be 38 WP4 onward), and a gate that refused every rebuild of every unmoved
+	// *Since 2026-10-04 (to-be 38 WP4b's last step):* refused for **every** module. While some
-	// module in the meantime would stop the whole pipeline to make a point the record already makes.
+	// thirty modules still stood in that shape, one already registered so was rebuilt without
-	// Before the runtime exists the pattern is accepted as it always was.
+	// complaint, so the pipeline kept running while each moved; every module has moved since, and
+	// the exception would only let one move back.
 	if m.Module != catalogue.RuntimeModule {
 		if why := catalogue.ToolContainerOnTheRuntime(m, from.Against); why != "" {
 			runtime, err := i.hasModule(ctx, catalogue.RuntimeModule)
@@ -90,16 +91,10 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
 				return err
 			}
 			if runtime {
-				already, err := i.registeredInThatShape(ctx, m.Module)
-				if err != nil {
-					return err
-				}
-				if !already {
 				return fmt.Errorf("%s is not registered: %s", m.Module, why)
 			}
 		}
 	}
-	}
 	raw, err := json.Marshal(m)
 	if err != nil {
 		return err
@@ -152,26 +147,6 @@ func (i *Inventory) RegisterModule(ctx context.Context, m catalogue.Manifest, fr
 	return nil
 }
 
-// registeredInThatShape is whether the catalogue already holds this module as a tools container on
-// the runtime's image — judged from the manifest it holds and what that module's newest build stood
-// on, the same two things the gate judges a new registration by. False for a module the catalogue
-// does not hold.
-func (i *Inventory) registeredInThatShape(ctx context.Context, name string) (bool, error) {
-	held, err := i.Catalogue(ctx)
-	if err != nil {
-		return false, err
-	}
-	stored, has := held[name]
-	if !has {
-		return false, nil
-	}
-	against, err := i.BuiltAgainst(ctx)
-	if err != nil {
-		return false, err
-	}
-	return catalogue.ToolContainerOnTheRuntime(stored, against[name]) != "", nil
-}
-
 // hasModule is whether the catalogue holds a module of that name.
 func (i *Inventory) hasModule(ctx context.Context, name string) (bool, error) {
 	var one int

internal/inventory/catalogue_test.go

@@ -688,9 +688,9 @@ func TestRegisteringWithoutProvenanceKeepsTheSeat(t *testing.T) {
 
 // Once the node's tool runtime is in the catalogue, a module serving its tools from a container
 // built on the runtime's image is refused at registration, naming the record (novox/hq ADR 0175,
-// to-be 38 WP2.4) — for a module new to the catalogue or one that had moved away from it; a module
+// to-be 38 WP2.4) — for every module, since every module has moved (WP4b's last step; WP3's
-// already standing in that shape is rebuilt as before, so the catalogue's pipeline keeps running
+// amendment let one already standing in that shape be rebuilt while each moved). Before the
-// while each moves (WP3's amendment). Before the runtime, it is accepted as it always was — so a
+// runtime, it is accepted as it always was — so a
 // mesh converts in the order the design says and nothing is refused before there is anything to
 // move to.
 func TestAToolContainerIsRefusedOnceTheRuntimeIsRegistered(t *testing.T) {
@@ -715,11 +715,11 @@ func TestAToolContainerIsRefusedOnceTheRuntimeIsRegistered(t *testing.T) {
 		t.Fatal(err)
 	}
 
-	// **A module already registered in that shape is rebuilt without complaint** (to-be 38 WP2.4 as
+	// **A module already registered in that shape is refused too** (WP4b's last step): every module
-	// amended by WP3): some thirty of them stand the day the runtime arrives, and each moves in its
+	// has moved, and a rebuild in the old shape is one moving back.
-	// own change. The gate is against the pattern spreading, not against the pipeline running.
+	if err := inv.RegisterModule(ctx, filter, Source{Repository: "/r", Against: stoodOn}); err == nil ||
-	if err := inv.RegisterModule(ctx, filter, Source{Repository: "/r", Against: stoodOn}); err != nil {
+		!strings.Contains(err.Error(), "ADR 0175") {
-		t.Fatalf("a rebuild of a module that already had the pattern was refused: %v", err)
+		t.Fatalf("a rebuild of a module in the old pattern was registered beside the runtime: %v", err)
 	}
 	// A module new to the catalogue in that shape is refused, naming the record.
 	newcomer := filter

internal/inventory/collection.go

@@ -0,0 +1,242 @@
+package inventory
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+)
+
+// What the artifact store keeps, and what it may let go (novox/hq ADR 0189, issue 108).
+//
+// The store has never collected anything: every build pushes another layer set and nothing has
+// ever removed one. The registry's own answer — collect what no tag names — is wrong here, because
+// the mesh pushes each artifact under one moving tag and pins machines by digest, so every build
+// but the newest is untagged and some machine may still be running it.
+//
+// **So the mesh decides, from its own records, and it never has to look in the store to do it.**
+// It has never put anything there it did not record, which means every digest it could remove is
+// already in a build row. A digest the mesh did not record making is therefore never named here —
+// not as a safety margin but as the rule restated, and it is what keeps the sweep away from the
+// images genesis pushed before any record existed (04-ISSUES/102, F4).
+
+// KeptBuilds is how many successful builds of each module keep their artifacts, counting the
+// newest. The newest is what the mesh hands a machine now; the four behind it are how far back a
+// release that turns out wrong can be taken.
+const KeptBuilds = 5
+
+// ToCollect is every artifact the mesh made, no longer keeps, and has not already collected.
+//
+// Three reasons an artifact stays, and nothing else is a reason:
+//
+//   - **a definition names it** — the reference appears in a module's recorded manifest, which is
+//     what the mesh would hand a machine now. No age limit: this is the floor;
+//   - **the mesh can still go back to it** — it is an artifact of one of the KeptBuilds most
+//     recent successful builds of its module;
+//   - it was already collected, in which case there is nothing left to do.
+//
+// Returned in a stated order so two runs over the same records ask for the same things in the
+// same sequence, which is what makes a failed sweep safe to simply run again.
+func (i *Inventory) ToCollect(ctx context.Context) ([]string, error) {
+	keep, err := i.keptReferences(ctx)
+	if err != nil {
+		return nil, err
+	}
+	rows, err := i.store.Pool().Query(ctx,
+		// Every artifact of every successful build, oldest first, minus what has already been
+		// collected. A failed build published nothing, so it names nothing to remove.
+		`select b.made
+		   from build b
+		  where b.failed = '' and b.module is not null and b.module <> ''
+		  order by b.at asc, b.id asc`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+
+	collected, err := i.alreadyCollected(ctx)
+	if err != nil {
+		return nil, err
+	}
+	seen := map[string]bool{}
+	var out []string
+	for rows.Next() {
+		var raw []byte
+		if err := rows.Scan(&raw); err != nil {
+			return nil, err
+		}
+		var made []Artifact
+		if err := json.Unmarshal(raw, &made); err != nil {
+			// One unreadable record must not stop the rest being collected — and an artifact this
+			// row named is simply not offered, which errs toward keeping.
+			continue
+		}
+		for _, a := range made {
+			if a.Reference == "" || keep[a.Reference] || collected[a.Reference] || seen[a.Reference] {
+				continue
+			}
+			seen[a.Reference] = true
+			out = append(out, a.Reference)
+		}
+	}
+	return out, rows.Err()
+}
+
+// keptReferences is every artifact reference the mesh still keeps, for either of the two reasons.
+func (i *Inventory) keptReferences(ctx context.Context) (map[string]bool, error) {
+	keep := map[string]bool{}
+
+	// **Whatever a definition the mesh holds names.** Read as text rather than by walking the
+	// resource shapes: a reference may be a container's image, a bundle's source, or a field some
+	// later kind of resource grows, and what matters is only whether the mesh could hand this
+	// string to a machine. A manifest that mentions it is a manifest that might.
+	manifests, err := i.store.Pool().Query(ctx, `select manifest::text from module where manifest is not null`)
+	if err != nil {
+		return nil, err
+	}
+	defer manifests.Close()
+	var named []string
+	for manifests.Next() {
+		var text string
+		if err := manifests.Scan(&text); err != nil {
+			return nil, err
+		}
+		named = append(named, text)
+	}
+	if err := manifests.Err(); err != nil {
+		return nil, err
+	}
+
+	// The KeptBuilds most recent successful builds of each module, whole.
+	recent, err := i.store.Pool().Query(ctx,
+		`select made from (
+		    select made, row_number() over (partition by module order by at desc, id desc) as back
+		      from build
+		     where failed = '' and module is not null and module <> ''
+		 ) ranked where back <= $1`, KeptBuilds)
+	if err != nil {
+		return nil, err
+	}
+	defer recent.Close()
+	for recent.Next() {
+		var raw []byte
+		if err := recent.Scan(&raw); err != nil {
+			return nil, err
+		}
+		var made []Artifact
+		if err := json.Unmarshal(raw, &made); err != nil {
+			continue
+		}
+		for _, a := range made {
+			if a.Reference != "" {
+				keep[a.Reference] = true
+			}
+		}
+	}
+	if err := recent.Err(); err != nil {
+		return nil, err
+	}
+
+	// And anything a manifest mentions. Done after the recent set so the scan runs over the
+	// candidates rather than over every reference ever recorded: a manifest holds a reference
+	// composed with the store's address or kept bare, so the search is for the digest within it.
+	if len(named) > 0 {
+		all, err := i.everyReferenceMade(ctx)
+		if err != nil {
+			return nil, err
+		}
+		for _, reference := range all {
+			if keep[reference] {
+				continue
+			}
+			digest := digestIn(reference)
+			if digest == "" {
+				// Not something the store holds by digest; nothing here can speak for it, so it
+				// is kept rather than guessed about.
+				keep[reference] = true
+				continue
+			}
+			for _, text := range named {
+				if strings.Contains(text, digest) {
+					keep[reference] = true
+					break
+				}
+			}
+		}
+	}
+	return keep, nil
+}
+
+// everyReferenceMade is every artifact reference any successful build recorded.
+func (i *Inventory) everyReferenceMade(ctx context.Context) ([]string, error) {
+	rows, err := i.store.Pool().Query(ctx,
+		`select made from build where failed = '' and module is not null and module <> ''`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	seen := map[string]bool{}
+	var out []string
+	for rows.Next() {
+		var raw []byte
+		if err := rows.Scan(&raw); err != nil {
+			return nil, err
+		}
+		var made []Artifact
+		if err := json.Unmarshal(raw, &made); err != nil {
+			continue
+		}
+		for _, a := range made {
+			if a.Reference == "" || seen[a.Reference] {
+				continue
+			}
+			seen[a.Reference] = true
+			out = append(out, a.Reference)
+		}
+	}
+	return out, rows.Err()
+}
+
+// digestIn is the `sha256:<hex>` a reference names, empty when it names none.
+func digestIn(reference string) string {
+	for _, marker := range []string{"@sha256:", "/sha256:"} {
+		if _, after, ok := strings.Cut(reference, marker); ok {
+			return "sha256:" + after
+		}
+	}
+	return ""
+}
+
+// alreadyCollected is what the store has already been asked to let go.
+func (i *Inventory) alreadyCollected(ctx context.Context) (map[string]bool, error) {
+	rows, err := i.store.Pool().Query(ctx, `select reference from artifact_collected`)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := map[string]bool{}
+	for rows.Next() {
+		var reference string
+		if err := rows.Scan(&reference); err != nil {
+			return nil, err
+		}
+		out[reference] = true
+	}
+	return out, rows.Err()
+}
+
+// MarkCollected records that the store no longer holds these.
+//
+// **A store that answered "not found" is recorded too.** The outcome wanted is that the artifact
+// is gone, and it is; retrying it every sweep for ever is the failure this table exists to
+// prevent. Only a store that could not be reached, or refused, leaves a reference unmarked — and
+// then the next sweep asks again, which is what should happen.
+func (i *Inventory) MarkCollected(ctx context.Context, references []string) error {
+	for _, reference := range references {
+		if _, err := i.store.Pool().Exec(ctx,
+			`insert into artifact_collected (reference) values ($1) on conflict (reference) do nothing`,
+			reference); err != nil {
+			return err
+		}
+	}
+	return nil
+}

internal/inventory/collection_test.go

@@ -0,0 +1,147 @@
+package inventory
+
+import (
+	"context"
+	"fmt"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// What the store keeps, and what it may let go (novox/hq ADR 0189, issue 108).
+//
+// The store has collected nothing since it was raised, and the registry's own answer — collect
+// what no tag names — would delete images machines are running, because the mesh pushes under one
+// moving tag and pins by digest. So the rule is the mesh's, read from its own records, and these
+// are the three reasons an artifact stays and the one reason it goes.
+
+// ref is an artifact reference as the mesh records one.
+func ref(module, artifact string, n int) string {
+	return fmt.Sprintf("%s%s/%s@sha256:%064x", catalogue.ArtifactStoreScheme, module, artifact, n)
+}
+
+// built records one successful build of a module publishing one image.
+func built(t *testing.T, inv *Inventory, id, module string, n int) string {
+	t.Helper()
+	reference := ref(module, "app", n)
+	b := aBuild(id, module, "")
+	b.Made = []Artifact{{Name: "app", Kind: "image", Reference: reference}}
+	if err := inv.RecordBuild(context.Background(), b); err != nil {
+		t.Fatal(err)
+	}
+	return reference
+}
+
+func TestTheStoreKeepsTheRecentBuildsAndLetsGoOfTheRest(t *testing.T) {
+	inv := fresh(t)
+	ctx := context.Background()
+
+	// Eight builds of one module, oldest first. Five are kept — the newest, and the four a
+	// release that turns out wrong can be taken back to.
+	var made []string
+	for i := 1; i <= 8; i++ {
+		made = append(made, built(t, inv, fmt.Sprintf("b%02d", i), "web", i))
+	}
+
+	go_, err := inv.ToCollect(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	want := made[:3] // the three oldest
+	if len(go_) != len(want) {
+		t.Fatalf("offered %v to collect; want the %d oldest of %d", go_, len(want), len(made))
+	}
+	for i := range want {
+		if go_[i] != want[i] {
+			t.Fatalf("offered %v; want %v — and in that order, so a failed sweep is safe to run again",
+				go_, want)
+		}
+	}
+}
+
+func TestADefinitionNamingAnArtifactKeepsItHoweverOldItIs(t *testing.T) {
+	// The floor: no age limit. A module recorded at an older commit still names what the mesh
+	// would hand a machine now, and that is what must not be collected out from under it.
+	inv := fresh(t)
+	ctx := context.Background()
+
+	var made []string
+	for i := 1; i <= 8; i++ {
+		made = append(made, built(t, inv, fmt.Sprintf("b%02d", i), "web", i))
+	}
+	oldest := made[0]
+
+	// A definition the mesh holds, whose container runs that oldest image.
+	m := catalogue.Manifest{Module: "web", Version: "1", Resources: []map[string]any{{
+		"id": "app", "type": "container", "name": "web", "image": oldest,
+	}}}
+	if err := inv.RegisterModule(ctx, m, Source{Repository: "https://forge.invalid/web.git"}); err != nil {
+		t.Fatal(err)
+	}
+
+	go_, err := inv.ToCollect(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, reference := range go_ {
+		if reference == oldest {
+			t.Fatalf("the mesh offered to collect %s, which a definition it holds names", oldest)
+		}
+	}
+	if len(go_) != 2 {
+		t.Fatalf("offered %v; want the two oldest that nothing names", go_)
+	}
+}
+
+func TestWhatHasBeenCollectedIsNotOfferedAgain(t *testing.T) {
+	// Without this the sweep reissues a delete for every artifact it has ever collected, every
+	// time it runs, for ever — a number of requests that grows with the mesh's whole history.
+	inv := fresh(t)
+	ctx := context.Background()
+	for i := 1; i <= 7; i++ {
+		built(t, inv, fmt.Sprintf("b%02d", i), "web", i)
+	}
+	first, err := inv.ToCollect(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(first) != 2 {
+		t.Fatalf("offered %v, want two", first)
+	}
+	if err := inv.MarkCollected(ctx, first); err != nil {
+		t.Fatal(err)
+	}
+	again, err := inv.ToCollect(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(again) != 0 {
+		t.Fatalf("offered %v again after collecting it", again)
+	}
+}
+
+func TestAFailedBuildNamesNothingToCollectAndEachModuleIsCountedOnItsOwn(t *testing.T) {
+	inv := fresh(t)
+	ctx := context.Background()
+
+	// A failed build published nothing, so it is neither kept nor collected — and it must not
+	// count against the module's five.
+	for i := 1; i <= 6; i++ {
+		built(t, inv, fmt.Sprintf("w%02d", i), "web", i)
+	}
+	if err := inv.RecordBuild(ctx, aBuild("w99", "web", "the recipe would not build")); err != nil {
+		t.Fatal(err)
+	}
+	// And a second module with three builds keeps all three: five each, not five between them.
+	for i := 1; i <= 3; i++ {
+		built(t, inv, fmt.Sprintf("d%02d", i), "db", 100+i)
+	}
+
+	go_, err := inv.ToCollect(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(go_) != 1 || go_[0] != ref("web", "app", 1) {
+		t.Fatalf("offered %v; want only web's oldest — db's three are all within its five", go_)
+	}
+}

internal/inventory/migrations/0056-what-the-store-no-longer-keeps.sql

@@ -0,0 +1,23 @@
+-- What the artifact store no longer keeps (novox/hq ADR 0189, issue 108).
+--
+-- The mesh removes from its store only what it put there and can account for: every digest it
+-- could remove is already in a build record, so the sweep reads its own records rather than
+-- enumerating the store. What it does not get from those records is whether it has already
+-- removed something -- `build.made` says what that build published, for ever, which is history
+-- and not an index of what is on disk.
+--
+-- Without this the sweep would reissue a delete for every artifact it has ever collected, every
+-- time it runs, and each one would answer 404 -- a number of requests that grows with the mesh's
+-- whole history and never shrinks.
+--
+-- Keyed by the reference as the mesh records it (`artifact-store://<module>/<artifact>@sha256:…`),
+-- because that is the identity the record uses everywhere else. Not a foreign key to build: two
+-- builds can publish the same digest (the same source built twice produces the same bytes), and
+-- what is collected is the artifact, not the attempt that made it.
+create table artifact_collected (
+    reference text        primary key,
+
+    -- When the store answered. Kept so a reader of an old build record can tell "this artifact is
+    -- gone" from "this artifact was never there", which are different kinds of surprise.
+    at        timestamptz not null default now()
+);

internal/store/manifest_test.go

@@ -30,7 +30,7 @@ func TestTheManifestsOwnPlaceholderUnfilledLeavesTheStoreWhereTheFileSays(t *tes
 	}
 	var written string
 	for _, r := range m.Resources {
-		if r.Type == "container" {
+		if r.Type == "process" {
 			written = r.Env["MESH_STORE_INVENTORY_PORT"]
 		}
 	}

module.json

@@ -2,9 +2,6 @@
   "module": "mesh-controller",
   "version": "1",
   "slug": "control",
-  "capabilities": [
-    "container-runtime"
-  ],
   "claims": [
     {
       "name": "mesh-controller",
@@ -26,7 +23,7 @@
     "broker-address": "${dir:mesh-state}/broker-address",
     "bus": "${dir:mesh-state}/bus"
   },
-  "secrets-owner": "65534:65534",
+  "secrets-owner": "mesh-controller",
   "prepares": true,
   "tools": [
     "tools",
@@ -43,62 +40,58 @@
     "build"
   ],
   "resources": [
+    {
+      "id": "account",
+      "type": "user",
+      "name": "mesh-controller",
+      "shell": "/usr/bin/nologin",
+      "home": "/var/lib/mesh-controller"
+    },
     {
       "id": "mesh-state",
       "type": "directory",
       "mode": "0700",
-      "place": "mesh"
+      "place": "mesh",
+      "owner": "mesh-controller"
     },
     {
-      "id": "server",
+      "id": "controller",
-      "type": "container",
+      "type": "process",
       "name": "mesh-controller",
-      "network": "host",
+      "artifact": "controller",
-      "args": [
+      "run": [
+        "./mesh-controller",
         "serve"
       ],
+      "user": "mesh-controller",
       "env": {
-        "MESH_BROKER_CERTIFICATE": "/broker-tls/tls.crt",
+        "MESH_BROKER_CERTIFICATE": "/var/lib/mesh-broker-tls/tls.crt",
-        "MESH_STORE_INVENTORY_FILE": "/run/secrets/inventory",
+        "MESH_STORE_INVENTORY_FILE": "${dir:mesh-state}/inventory",
-        "MESH_STORE_IDENTITY_FILE": "/run/secrets/identity",
+        "MESH_STORE_IDENTITY_FILE": "${dir:mesh-state}/identity",
-        "MESH_STORE_LICENCES_FILE": "/run/secrets/licences",
+        "MESH_STORE_LICENCES_FILE": "${dir:mesh-state}/licences",
-        "MESH_BROKER_MANAGEMENT_FILE": "/run/secrets/broker-management",
+        "MESH_BROKER_MANAGEMENT_FILE": "${dir:mesh-state}/broker-management",
-        "MESH_BROKER_ADDRESS_FILE": "/run/secrets/broker-address",
+        "MESH_BROKER_ADDRESS_FILE": "${dir:mesh-state}/broker-address",
         "MESH_STORE_INVENTORY_PORT": "${seat:mesh-store:5432}",
         "MESH_STORE_IDENTITY_PORT": "${seat:mesh-store:5432}",
         "MESH_STORE_LICENCES_PORT": "${seat:mesh-store:5432}",
         "MESH_BROKER_MANAGEMENT_PORT": "${seat:mesh-broker:15672}",
         "MESH_BROKER_ADDRESS_PORT": "${seat:mesh-broker:5671}",
-        "MESH_BUS_NATS_FILE": "/run/secrets/bus"
+        "MESH_BUS_NATS_FILE": "${dir:mesh-state}/bus"
       },
-      "volumes": [
+      "replaces": [
-        "/var/lib/mesh-broker-tls:/broker-tls:ro",
+        "server"
-        "${dir:mesh-state}/inventory:/run/secrets/inventory:ro",
-        "${dir:mesh-state}/identity:/run/secrets/identity:ro",
-        "${dir:mesh-state}/licences:/run/secrets/licences:ro",
-        "${dir:mesh-state}/broker:/run/secrets/broker:ro",
-        "${dir:mesh-state}/bus:/run/secrets/bus:ro",
-        "${dir:mesh-state}/broker-management:/run/secrets/broker-management:ro",
-        "${dir:mesh-state}/broker-address:/run/secrets/broker-address:ro"
-      ],
-      "artifact": "server",
-      "restart-on": [
-        "control-env"
       ]
     }
   ],
   "build": {
     "artifacts": [
       {
-        "name": "server",
+        "name": "controller",
-        "kind": "image",
+        "kind": "bundle",
-        "from": "Dockerfile"
+        "language": "go",
-      }
+        "system": "arch",
-    ],
+        "from": "cmd/mesh-controller",
-    "on": [
+        "binary": "mesh-controller"
-      {
-        "arg": "GO_BASE",
-        "image": "golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c"
       }
     ]
   }