Review before merge: refuse a silent disagreement, and bound the sweep

Three things found reading this back, each of which would have been quiet.

A consumer that keeps several holders of one provision (ADR 0094) gets a
login per holder, and a provider derives from the login — so it would make a
resource per holder while the consumer is told one value for the requirement.
That is issue 124's own failure one case to the side: authenticate, then be
refused on every object. Refused now, naming both ends.

The sweep runs inside somebody's build and was unbounded. At most two hundred
artifacts and sixty seconds, stopping at the first refusal because a store
that refuses one refuses all; the rest is offered again next build.

The citation and migration renumbers are in the commit before this one.

Dockerfile

@@ -1,5 +1,11 @@
-ARG GO_BASE=golang:1.25-alpine
-# The control plane's image.
+# The Go it builds with, pinned here because genesis builds this file with no arguments (novox/hq
+# 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,
-		Reference: "registry.example/control@" + aDigest}})
+	control, err := withSeatPorts(m).Resolve([]catalogue.Built{{Name: "controller", Kind: catalogue.ArtifactBundle,
+		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/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/plans_one_at_a_time_test.go

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// novox/hq issue 213: for the moment a machine hands its controller over, the container and the
+// process both run the plan timer on one store. Only the one holding the plans moves them; the other
+// leaves them alone, and moves them once they are let go.
+func TestAControllerLeavesThePlansToTheOneHoldingThem(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	now := time.Now().UTC()
+	// Every tier done: the next step is the plan's last, and needs nothing but the store.
+	plan := inventory.Plan{ID: "plan-213", Repository: "r", Commit: "abc", Created: now, Updated: now,
+		State: inventory.PlanRolling, Tier: 1, Tiers: [][]string{{"app"}},
+		Modules: map[string]*inventory.PlanModule{"app": {State: "built"}}}
+	if err := open.inventory.SavePlan(ctx, plan); err != nil {
+		t.Fatal(err)
+	}
+
+	// The other controller: its own connections to the same store, holding the plans.
+	other, err := inventory.Open(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(other.Close)
+	release, err := other.HoldPlans(ctx, false)
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(release) // before the close above: a pool waits for a connection still held
+
+	advancePlans(ctx, open)
+	if p, err := open.inventory.PlanByID(ctx, "plan-213"); err != nil || !p.Open() {
+		t.Fatalf("a controller moved a plan another held: %+v %v", p, err)
+	}
+
+	release()
+	advancePlans(ctx, open)
+	if p, err := open.inventory.PlanByID(ctx, "plan-213"); err != nil || p.State != inventory.PlanDone {
+		t.Fatalf("the plan did not move once it was let go: %+v %v", p, err)
+	}
+}

cmd/mesh-controller/release_plan.go

@@ -2,6 +2,7 @@ package main
 
 import (
 	"context"
+	"errors"
 	"flag"
 	"fmt"
 	"sort"
@@ -296,6 +297,15 @@ func askTier(ctx context.Context, inv *inventory.Inventory, p *inventory.Plan) e
 // build's request time is not known, and such an outcome is taken as before.
 func planBuilt(ctx context.Context, open *stores, module, commit, failed string, asked time.Time) {
 	inv := open.inventory
+	// One controller works the plans at a time (novox/hq issue 213); an outcome waits its turn rather
+	// than write over what the holder is about to save. Not taken, it is still in the build records,
+	// which the holder settles the plan from (issue 214).
+	release, err := inv.HoldPlans(ctx, true)
+	if err != nil {
+		fmt.Printf("plans: %s's outcome is left to the build records: %v\n", module, err)
+		return
+	}
+	defer release()
 	plans, err := inv.OpenPlans(ctx)
 	if err != nil {
 		fmt.Printf("plans: cannot read them: %v\n", err)
@@ -342,13 +352,30 @@ func planBuilt(ctx context.Context, open *stores, module, commit, failed string,
 			fmt.Printf("%s: %s; the tiers after it are not asked\n", p.ID, p.Note)
 		}
 	}
-	advancePlans(ctx, open)
+	advanceHeld(ctx, open)
 }
 
 // advancePlans moves every open plan as far as the facts allow: a tier whose modules are all built
 // and whose gates are applied gives way to the next; the last tier done is the plan done. Called
 // after every outcome and on a timer, so a plan waiting on a machine's report moves when it comes.
+//
+// **One controller at a time** (novox/hq issue 213). A plan is read, changed and saved whole; two
+// controllers — the old and the new while a machine hands its controller over — would each ask a
+// tier the other had just asked. Taken without waiting: whoever holds the plans is moving them.
 func advancePlans(ctx context.Context, open *stores) {
+	release, err := open.inventory.HoldPlans(ctx, false)
+	if err != nil {
+		if !errors.Is(err, inventory.ErrPlansBusy) {
+			fmt.Printf("plans: cannot hold them: %v\n", err)
+		}
+		return
+	}
+	defer release()
+	advanceHeld(ctx, open)
+}
+
+// advanceHeld is advancePlans for a caller already holding the plans.
+func advanceHeld(ctx context.Context, open *stores) {
 	inv := open.inventory
 	plans, err := inv.OpenPlans(ctx)
 	if err != nil {
@@ -676,6 +703,19 @@ func plansCommand(ctx context.Context, args []string) error {
 		}
 		p.State = inventory.PlanFailed
 		p.Note = "stopped by hand at tier " + fmt.Sprint(p.Tier)
+		release, err := inv.HoldPlans(ctx, true)
+		if err != nil {
+			return err
+		}
+		defer release()
+		if p, err = inv.PlanByID(ctx, positionals[1]); err != nil {
+			return err
+		}
+		if !p.Open() {
+			return fmt.Errorf("%s is already %s", p.ID, p.State)
+		}
+		p.State = inventory.PlanFailed
+		p.Note = "stopped by hand at tier " + fmt.Sprint(p.Tier)
 		if err := inv.SavePlan(ctx, p); err != nil {
 			return err
 		}

cmd/mesh-controller/upgrades.go

@@ -315,6 +315,13 @@ func (f following) SourceMoved(ctx context.Context, m link.SourceMoved) error {
 	for _, e := range moved {
 		movedNames = append(movedNames, e.Manifest.Module)
 	}
+	// Written and its first tier asked as one act on the plans (novox/hq issue 213): a timer on
+	// another controller reading it between the two would ask the tier again.
+	release, err := inv.HoldPlans(ctx, true)
+	if err != nil {
+		return notNow(err)
+	}
+	defer release()
 	plan := planOfMerge(m, movedNames, edges)
 	if hasCycle(plan.Tiers, edges) {
 		fmt.Printf("  the last tier depends on itself: %s — built together, in no order\n",

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/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
-	// own npm dependencies are a later step). Empty for a language whose bundle carries its own —
+	// dependencies, and nothing module-specific (a module's own npm dependencies are installed into
+	// 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/a_service_process_test.go

@@ -0,0 +1,166 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+)
+
+// novox/hq issue 213 (ADR 0188 §1, §3): a module's own Go service is a bundle the host runs as a
+// process, not an image. Each test holds one thing that had to change in the composer for the
+// controller to be declared that way.
+
+const aServiceDigest = "sha256:dddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddddd"
+
+// aServiceModule is the controller's shape in miniature: it answers tools of its own, its code is a
+// Go bundle a process runs as an account it declares, its secrets belong to that account, it
+// prepares its state, and its process replaces the container it used to run as.
+func aServiceModule(t *testing.T) Manifest {
+	t.Helper()
+	raw := `{
+	  "module": "svc", "version": "1", "tools": ["status"], "prepares": true,
+	  "own-secrets": {"store": "${dir:state}/store"},
+	  "secrets-owner": "svc",
+	  "resources": [
+	    {"id": "state", "type": "directory", "mode": "0700", "place": "mesh", "owner": "svc"},
+	    {"id": "service", "type": "process", "name": "svc", "artifact": "code",
+	     "run": ["./svc", "serve"], "user": "svc", "replaces": ["server"],
+	     "env": {"SVC_STORE_FILE": "${dir:state}/store", "SVC_STORE_PORT": "${seat:mesh-store:5432}"}},
+	    {"id": "account", "type": "user", "name": "svc", "shell": "/usr/bin/nologin", "home": "/var/lib/svc"}
+	  ],
+	  "build": {"artifacts": [{"name": "code", "kind": "bundle", "language": "go", "system": "arch",
+	    "from": "cmd/svc", "binary": "svc"}]}
+	}`
+	m, err := ParseManifest([]byte(raw))
+	if err != nil {
+		t.Fatalf("the service's manifest is refused: %v", err)
+	}
+	resolved, err := m.Resolve([]Built{{Name: "code", Kind: ArtifactBundle,
+		Reference: ArtifactStoreScheme + "svc/code@" + aServiceDigest, Digest: aServiceDigest}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	return resolved
+}
+
+func composeTheService(t *testing.T, with Rendering) []map[string]any {
+	t.Helper()
+	with.Needed = map[string]map[string]string{"svc": {"store": "sealed-store"}}
+	with.ArtifactStore = "anchor.internal:5100"
+	out, err := Resolution{Node: "anchor", Modules: []Manifest{aServiceModule(t)}}.Declaration(with)
+	if err != nil {
+		t.Fatalf("the service does not compose: %v", err)
+	}
+	return out
+}
+
+func indexOf(out []map[string]any, id string) int {
+	for i, r := range out {
+		if r["id"] == id {
+			return i
+		}
+	}
+	return -1
+}
+
+// A module that declares tools has every bundle served by the node's runtime unless it says
+// otherwise — and the controller declares the verbs it answers as tools. Its service bundle is run
+// by its own process; launched a second time by the runtime it would be a second controller
+// pretending to be an MCP server.
+func TestABundleItsOwnProcessRunsIsNotServedByTheRuntime(t *testing.T) {
+	m := aServiceModule(t)
+	if len(m.Bundles) != 1 {
+		t.Fatalf("the service's bundle was not kept: %+v", m.Bundles)
+	}
+	if loads := m.Bundles[0].Loads; len(loads) != 0 {
+		t.Fatalf("the runtime would launch the service's own bundle as tools: %v", loads)
+	}
+	// And a bundle no resource runs still is served, as a module declaring tools always had it.
+	tools := Manifest{Module: "t", Version: "1", Tools: []string{"x"},
+		Build: &Build{Artifacts: []Artifact{{Name: "tools", Kind: ArtifactBundle, Language: "go",
+			System: "arch", From: "cmd/t"}}}}
+	resolved, err := tools.Resolve([]Built{{Name: "tools", Kind: ArtifactBundle,
+		Reference: ArtifactStoreScheme + "t/tools@" + aServiceDigest, Digest: aServiceDigest}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if loads := resolved.Bundles[0].Loads; len(loads) != 1 || loads[0] != "t" {
+		t.Fatalf("a tools bundle nothing runs is no longer served: %v", loads)
+	}
+}
+
+// The account is created before anything is given to it. Its secrets are mesh-computed and so
+// placed before the module's own resources; given to a user the machine did not have yet, they were
+// refused on the first apply and the process started without them.
+func TestAModulesAccountComesBeforeWhatBelongsToIt(t *testing.T) {
+	out := composeTheService(t, Rendering{})
+	account, secret := indexOf(out, "svc.account"), indexOf(out, "svc."+NeedID("store"))
+	if account < 0 || secret < 0 {
+		t.Fatalf("the account or the secret is missing: %v", out)
+	}
+	if account > secret {
+		t.Fatalf("the secret owned by svc is written before svc exists: account at %d, secret at %d",
+			account, secret)
+	}
+	if owner := out[secret]["owner"]; owner != "svc" {
+		t.Errorf("the secret belongs to %v, not the account its process runs as", owner)
+	}
+}
+
+// The process is the module's program; its preparation is the same program asked to prepare, as a
+// step before it — with the same account and environment, and handing nothing over.
+func TestAProcessIsPreparedByItsOwnProgram(t *testing.T) {
+	out := composeTheService(t, Rendering{})
+	step, process := indexOf(out, "svc.service-prepare"), indexOf(out, "svc.service")
+	if step < 0 || process < 0 || step > process {
+		t.Fatalf("the preparation is not a step before the process (%d, %d): %v", step, process, out)
+	}
+	s := out[step]
+	if s["type"] != "process" || s["run-once"] != true || s["name"] != "svc-prepare" {
+		t.Errorf("the preparation is not a run-once process: %v", s)
+	}
+	if run, _ := json.Marshal(s["run"]); string(run) != `["./svc","prepare"]` {
+		t.Errorf("the preparation runs %s", run)
+	}
+	if s["user"] != "svc" || s["source"] != out[process]["source"] {
+		t.Errorf("the preparation does not run the same bundle as the same account: %v", s)
+	}
+	if env, _ := s["env"].(map[string]any); env["SVC_STORE_FILE"] == nil {
+		t.Errorf("the preparation is not given the process's environment: %v", s["env"])
+	}
+	if _, has := s["replaces"]; has {
+		t.Errorf("the preparation would hand over what the process replaces: %v", s)
+	}
+	if _, has := s["args"]; has {
+		t.Errorf("the preparation carries a container's args: %v", s)
+	}
+}
+
+// What the process replaces is named as the host recorded it, `<module>.<id>`; unprefixed, the host
+// matches nothing and removes the container first, as before.
+func TestWhatAProcessReplacesIsNamedAsTheHostRecordedIt(t *testing.T) {
+	out := composeTheService(t, Rendering{})
+	p := out[indexOf(out, "svc.service")]
+	if got, _ := json.Marshal(p["replaces"]); string(got) != `["svc.server"]` {
+		t.Fatalf("the process replaces %s", got)
+	}
+}
+
+func TestWhatReplacesMayNameIsRefusedNearItsAuthor(t *testing.T) {
+	for what, resource := range map[string]string{
+		"a container":        `{"id":"c","type":"container","name":"c","image":"x@` + aServiceDigest + `","replaces":["old"]}`,
+		"a step":             `{"id":"p","type":"process","name":"p","run":["./p"],"run-once":true,"replaces":["old"]}`,
+		"something declared": `{"id":"p","type":"process","name":"p","run":["./p"],"replaces":["p"]}`,
+		"another module's":   `{"id":"p","type":"process","name":"p","run":["./p"],"replaces":["other.old"]}`,
+		"not a list":         `{"id":"p","type":"process","name":"p","run":["./p"],"replaces":"old"}`,
+	} {
+		raw := `{"module":"m","version":"1","resources":[` + resource + `]}`
+		if _, err := ParseManifest([]byte(raw)); err == nil || !strings.Contains(err.Error(), "replace") {
+			t.Errorf("replaces on %s was accepted: %v", what, err)
+		}
+	}
+	ok := `{"module":"m","version":"1","resources":[{"id":"p","type":"process","name":"p","run":["./p"],"replaces":["old"]}]}`
+	if _, err := ParseManifest([]byte(ok)); err != nil {
+		t.Errorf("a process replacing what its module no longer declares was refused: %v", err)
+	}
+}

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/build.go

@@ -77,6 +77,7 @@ func (m Manifest) Resolve(built []Built) (Manifest, error) {
 	// build compare equal.
 	out.Bundles = nil
 	if m.Build != nil {
+		run := runByAResource(m)
 		for _, a := range m.Build.Artifacts {
 			if a.Kind != ArtifactBundle {
 				continue
@@ -84,8 +85,13 @@ func (m Manifest) Resolve(built []Built) (Manifest, error) {
 			made := by[a.Name]
 			// What the runtime loads: what the artifact said, else every entrypoint of a module
 			// that declares tools, else nothing (the field's own rule; see Artifact.Loads).
+			//
+			// **Never, unasked, a bundle one of the module's own resources runs** (novox/hq issue 213).
+			// A process the host runs is the module's service, not its tools: the controller declares
+			// the verbs it answers as `tools` and serves them itself, and its bundle would otherwise
+			// have been launched a second time by the node's runtime, as an MCP child it is not.
 			loads := append([]string(nil), a.Loads...)
-			if a.Loads == nil && len(m.Tools) > 0 {
+			if a.Loads == nil && len(m.Tools) > 0 && !run[a.Name] {
 				loads = append([]string(nil), a.Entrypoints...)
 				// A bundle compiled to a binary has no entrypoints: the binary is what it is, and what
 				// the runtime starts to serve it (novox/hq ADR 0193). So a Go tools bundle is served
@@ -453,6 +459,18 @@ func BinaryOf(a Artifact) string {
 	return a.Name
 }
 
+// runByAResource is the artifacts one of a module's own resources names — a process that runs it,
+// a step, an archive that unpacks it — by name.
+func runByAResource(m Manifest) map[string]bool {
+	named := map[string]bool{}
+	for _, r := range m.Resources {
+		if a, ok := r["artifact"].(string); ok && a != "" {
+			named[a] = true
+		}
+	}
+	return named
+}
+
 // undeliveredBundles says which of a module's bundles nothing would ever put on a machine (novox/hq
 // 04-ISSUES/216). A bundle reaches a machine three ways: the node's runtime serves it (it says
 // `loads`, or its module declares `tools`), a resource names it (a process, a step, an archive), or
@@ -463,12 +481,7 @@ func undeliveredBundles(m Manifest) []string {
 	if m.Build == nil || m.Module == RuntimeModule {
 		return nil
 	}
-	named := map[string]bool{}
-	for _, r := range m.Resources {
-		if a, ok := r["artifact"].(string); ok && a != "" {
-			named[a] = true
-		}
-	}
+	named := runByAResource(m)
 	var problems []string
 	for _, a := range m.Build.Artifacts {
 		if a.Kind != ArtifactBundle || named[a.Name] || len(a.Loads) > 0 || len(m.Tools) > 0 {

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

@@ -645,7 +645,16 @@ 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
 			}
@@ -693,7 +702,15 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 
 		// Now, and not before: a module whose resources are computed replaces them wholesale, and
 		// merging earlier would throw away the files it still needs.
-		resources = append(append([]map[string]any{}, first...), resources...)
+		//
+		// **Except the module's own accounts, which go before even those** (novox/hq issue 213). What
+		// the mesh computes may belong to one: a module whose code runs as an account it declares has
+		// its secrets written owned by that account, and a file given to a user the machine does not
+		// have yet fails — so on the first apply the secrets were refused, the process started without
+		// them, and the second apply healed it, which is the fault the paragraph above describes.
+		// An account depends on nothing the mesh computes.
+		accounts, rest := accountsFirst(resources)
+		resources = append(append(accounts, first...), rest...)
 
 		// No container is given the mesh's names (novox/hq ADR 0148). It used to be: every
 		// container got the whole roster as `--add-host` entries at creation, and a name that
@@ -711,7 +728,10 @@ 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.
@@ -728,7 +748,11 @@ 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
 			}
 		}
@@ -753,6 +777,17 @@ 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)
@@ -872,6 +907,12 @@ 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 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.
+			if renamed := reflectsRenamed(m.Module, resource["replaces"]); renamed != nil {
+				copied["replaces"] = renamed
+			}
 			// **What reads one of this module's own secrets is restarted when it changes** (novox/hq
 			// issue 203, issue 206). A credential is re-issued by the mesh, and a container that
 			// mounted the old file keeps the old one open: the build machine ran for an hour on a
@@ -1161,6 +1202,19 @@ 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.
@@ -1252,12 +1306,17 @@ 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 {
@@ -2010,12 +2069,18 @@ func preparationTarget(m Manifest) string {
 		return ""
 	}
 	for _, r := range m.Resources {
-		if fmt.Sprint(r["type"]) != "container" || !ownArtifact(r, m.Module) {
+		// A container, or a process the host runs from a bundle the module built (novox/hq issue
+		// 213): the same program in the same context, hosted as a unit rather than a container.
+		kind := fmt.Sprint(r["type"])
+		if (kind != "container" && kind != "process") || !ownArtifact(r, m.Module) {
 			continue
 		}
 		if once, _ := r["run-once"].(bool); once {
 			continue
 		}
+		if r["schedule"] != nil {
+			continue
+		}
 		return fmt.Sprint(r["id"])
 	}
 	return ""
@@ -2029,7 +2094,10 @@ func ownArtifact(resource map[string]any, module string) bool {
 		return true
 	}
 	image, _ := resource["image"].(string)
-	return strings.HasPrefix(image, ArtifactStoreScheme+module+"/")
+	// A process or an archive carries what was built as its source (novox/hq issue 213).
+	source, _ := resource["source"].(string)
+	return strings.HasPrefix(image, ArtifactStoreScheme+module+"/") ||
+		strings.HasPrefix(source, ArtifactStoreScheme+module+"/")
 }
 
 // prepared is the module's own resource as the step that prepares its state: the same image, the same
@@ -2051,7 +2119,19 @@ func prepared(from map[string]any) map[string]any {
 	step["id"] = fmt.Sprint(from["id"]) + "-prepare"
 	step["name"] = fmt.Sprint(from["name"]) + "-prepare"
 	step["run-once"] = true
-	step["args"] = []any{PreparationArgument}
+	if fmt.Sprint(from["type"]) == "process" {
+		// A process says its whole command: the program, then its arguments. The step is the same
+		// program asked to prepare (novox/hq issue 213). It replaces nothing — what the process
+		// replaces is handed over to the process, never to the step that runs before it — and a
+		// step is not restarted, it runs again when what it reads changed, which `restart-on` says.
+		run := stringsIn(from["run"])
+		if len(run) > 0 {
+			step["run"] = []any{run[0], PreparationArgument}
+		}
+		delete(step, "replaces")
+	} else {
+		step["args"] = []any{PreparationArgument}
+	}
 	delete(step, "ports")
 	delete(step, "ip")
 	delete(step, "schedule")
@@ -2196,3 +2276,16 @@ func withRestartOn(have any, add []string) []any {
 	}
 	return out
 }
+
+// accountsFirst splits a module's resources into its accounts and everything else, each in the order
+// written.
+func accountsFirst(resources []map[string]any) (accounts, rest []map[string]any) {
+	for _, r := range resources {
+		if fmt.Sprint(r["type"]) == "user" {
+			accounts = append(accounts, r)
+			continue
+		}
+		rest = append(rest, r)
+	}
+	return accounts, rest
+}

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/foundation_manifests_test.go

@@ -1,6 +1,7 @@
 package catalogue
 
 import (
+	"encoding/json"
 	"fmt"
 	"os"
 	"reflect"
@@ -170,7 +171,7 @@ func TestTheForgeHoldsTheNpmAndGitSeats(t *testing.T) {
 // **And the forge's own address follows it**, composed from the manifest in the catalogue beside
 // this checkout (novox/hq 04-ISSUES/088).
 //
-// The forge is reached a third way that neither test above covers: by its own sidecar, over the
+// The forge is reached a third way that neither test above covers: by its own code, over the
 // machine's loopback, told where to go in its environment. The `2999:3000` mapping that lets the
 // forge go on binding 3000 does nothing for a caller dialling the machine — so a literal there is
 // wrong on every node whose assignment differs, and wrong for a second reason on a node given the
@@ -178,13 +179,13 @@ func TestTheForgeHoldsTheNpmAndGitSeats(t *testing.T) {
 // in an `env` at all is a declaration, not a manifest.
 func TestTheForgesOwnAddressFollowsThePortTheNodeGaveIt(t *testing.T) {
 	forge, err := catalogueManifest(t, "gitea").Resolve([]Built{{
-		Name: "runtime", Kind: ArtifactImage,
-		Reference: "registry.example/gitea-runtime@sha256:" + strings.Repeat("a", 64),
+		Name: "code", Kind: ArtifactBundle,
+		Reference: ArtifactStoreScheme + "gitea/code/blobs/" + bundleDigest, Digest: bundleDigest,
 	}})
 	if err != nil {
 		t.Fatalf("the forge's manifest does not resolve against its own build: %v", err)
 	}
-	r := Resolution{Node: "anchor", Modules: []Manifest{forge}, Needs: []Needed{
+	r := Resolution{Node: "anchor", Modules: []Manifest{forge, theRuntime(t)}, Needs: []Needed{
 		{Name: "postgres-database", For: "gitea", From: "anchor", At: "127.0.0.1",
 			Serves: map[string]any{"port": float64(5432)}, Sealed: "sealed-db"},
 		{Name: "route", For: "gitea", From: "anchor"},
@@ -194,8 +195,8 @@ func TestTheForgesOwnAddressFollowsThePortTheNodeGaveIt(t *testing.T) {
 
 	// The number this node was given for the forge — the one the machine it is about to run on
 	// already publishes.
-	out, err := r.Declaration(Rendering{
-		Needed: map[string]map[string]string{"gitea": {"broker": "sealed-broker"}},
+	out, err := r.Declaration(Rendering{ArtifactStore: "anchor.internal:5101",
+		Needed: map[string]map[string]string{RuntimeModule: {"broker": "sealed-broker"}},
 		Given:  map[string]map[int]int{"gitea": {3000: 2999}},
 	})
 	if err != nil {
@@ -210,14 +211,19 @@ func TestTheForgesOwnAddressFollowsThePortTheNodeGaveIt(t *testing.T) {
 	if published := fmt.Sprint(server["ports"]); !strings.Contains(published, "2999:3000") {
 		t.Fatalf("the forge is not published on the port this node gave it: %v", server["ports"])
 	}
-	runtime := fileNamed(out, "gitea.runtime")
+	// The forge's own code runs in the node's runtime (novox/hq ADR 0198), given its words there.
+	runtime := fileNamed(out, RuntimeModule+"."+RuntimeProcessID())
 	if runtime == nil {
-		t.Fatalf("the forge's sidecar is not in the declaration: %v", out)
+		t.Fatalf("the node's runtime is not in the declaration: %v", ids(out))
 	}
-	env, _ := runtime["env"].(map[string]any)
-	if env["MESH_GITEA_URL"] != "http://127.0.0.1:2999" {
-		t.Fatalf("the forge's sidecar dials %v while the machine publishes the forge on 2999 — "+
-			"whatever reads it dials a dead port", env["MESH_GITEA_URL"])
+	env, _ := runtime["env"].(map[string]string)
+	var given map[string]map[string]string
+	if err := json.Unmarshal([]byte(env[RuntimeToolEnv]), &given); err != nil {
+		t.Fatalf("the runtime's %s is not JSON: %q", RuntimeToolEnv, env[RuntimeToolEnv])
+	}
+	if given["gitea"]["MESH_GITEA_URL"] != "http://127.0.0.1:2999" {
+		t.Fatalf("the forge's code dials %v while the machine publishes the forge on 2999 — "+
+			"whatever reads it dials a dead port", given["gitea"]["MESH_GITEA_URL"])
 	}
 }
 
@@ -229,13 +235,13 @@ func declaredGiteaSsh(t *testing.T, given map[int]int) map[string]any {
 	t.Helper()
 	forge := catalogueManifest(t, "gitea")
 	resolved, err := forge.Resolve([]Built{{
-		Name: "runtime", Kind: ArtifactImage,
-		Reference: "registry.example/gitea-runtime@sha256:" + strings.Repeat("a", 64),
+		Name: "code", Kind: ArtifactBundle,
+		Reference: ArtifactStoreScheme + "gitea/code/blobs/" + bundleDigest, Digest: bundleDigest,
 	}})
 	if err != nil {
 		t.Fatalf("the forge's manifest does not resolve against its own build: %v", err)
 	}
-	r := Resolution{Node: "anchor", Modules: []Manifest{resolved}, Needs: []Needed{
+	r := Resolution{Node: "anchor", Modules: []Manifest{resolved, theRuntime(t)}, Needs: []Needed{
 		{Name: "postgres-database", For: "gitea", From: "anchor", At: "127.0.0.1",
 			Serves: map[string]any{"port": float64(5432)}, Sealed: "sealed-db"},
 		{Name: "route", For: "gitea", From: "anchor"},
@@ -246,8 +252,8 @@ func declaredGiteaSsh(t *testing.T, given map[int]int) map[string]any {
 	for k, v := range given {
 		givenPorts[k] = v
 	}
-	out, err := r.Declaration(Rendering{
-		Needed: map[string]map[string]string{"gitea": {"broker": "sealed-broker"}},
+	out, err := r.Declaration(Rendering{ArtifactStore: "anchor.internal:5101",
+		Needed: map[string]map[string]string{RuntimeModule: {"broker": "sealed-broker"}},
 		Ports:  map[string]map[int]int{"gitea": givenPorts},
 		Given:  map[string]map[int]int{"gitea": given},
 	})

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

@@ -1428,6 +1428,10 @@ 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(
@@ -1515,6 +1519,53 @@ func ParseManifest(raw []byte) (Manifest, error) {
 				"program that reads what the mesh delivered and reconciles",
 			m.Module, r["id"]))
 	}
+	// **What a process replaces is something the module no longer declares** (novox/hq issue 213).
+	// The host keeps it running until the process is, then removes it: so it is named by the id the
+	// module used to give it, it is never a resource the module still declares — that would be
+	// applied and removed by one declaration — and only a process that stays up has anything to
+	// hand over to. Said here, near the author, as the host would refuse it far away.
+	ids := map[string]bool{}
+	for _, r := range m.Resources {
+		ids[fmt.Sprint(r["id"])] = true
+	}
+	for _, r := range m.Resources {
+		raw, present := r["replaces"]
+		if !present {
+			continue
+		}
+		if fmt.Sprint(r["type"]) != "process" {
+			problems = append(problems, fmt.Sprintf(
+				"%s: %v says what it replaces, and only a process does", m.Module, r["id"]))
+			continue
+		}
+		if once, _ := r["run-once"].(bool); once || r["schedule"] != nil {
+			problems = append(problems, fmt.Sprintf(
+				"%s: %v replaces something and runs once or on a schedule — only a process that stays "+
+					"up is there a moment later to hand over to", m.Module, r["id"]))
+		}
+		list, ok := raw.([]any)
+		if !ok {
+			problems = append(problems, fmt.Sprintf(
+				"%s: %v replaces %v; replaces is a list of the ids this module no longer declares",
+				m.Module, r["id"], raw))
+			continue
+		}
+		for _, item := range list {
+			id, ok := item.(string)
+			switch {
+			case !ok || strings.TrimSpace(id) == "":
+				problems = append(problems, fmt.Sprintf(
+					"%s: %v replaces %v, which is not an id", m.Module, r["id"], item))
+			case strings.Contains(id, "."):
+				problems = append(problems, fmt.Sprintf(
+					"%s: %v replaces %q; a process replaces only a resource of its own module, named "+
+						"by its own id", m.Module, r["id"], id))
+			case ids[id]:
+				problems = append(problems, fmt.Sprintf(
+					"%s: %v replaces %q, which this module still declares", m.Module, r["id"], id))
+			}
+		}
+	}
 	// **A module that prepares its state must have code the mesh can run** (novox/hq ADR 0135). The
 	// preparation is the module's own program in its preparation mode, so it is derived from the
 	// resource that runs that program — and a module declaring none has asked for something the mesh
@@ -1522,7 +1573,7 @@ func ParseManifest(raw []byte) (Manifest, error) {
 	// quietly prepares nothing.
 	if m.Prepares && preparationTarget(m) == "" {
 		problems = append(problems, fmt.Sprintf(
-			"%s says it prepares its state, and declares no container running an artifact it built — "+
+			"%s says it prepares its state, and declares no container or process running an artifact it built — "+
 				"the preparation is this module's own program, so there has to be one for the mesh to "+
 				"run it in", m.Module))
 	}
@@ -1575,6 +1626,13 @@ 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) {
@@ -2106,3 +2164,71 @@ 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/port_into.go

@@ -31,7 +31,7 @@ import (
 //
 // So a module asks. `${port:8080}` is "the machine-side port you gave me for the 8080 I said I
 // listen on", and the module writes that where it would otherwise have written a literal — in a
-// file's content, or in a value of a container's `env`.
+// file's content, or in a value of a container's or a process's `env`.
 //
 // **The environment is filled by the control plane, exactly as a bound value is.** A port is not
 // secret — the mesh holds it in the clear — so there is nothing for the host to be the only
@@ -64,7 +64,12 @@ func portsUsed(content string) []int {
 }
 
 // portInto replaces a resource's ${port:…} placeholders with what this machine assigned — in a
-// file's content, and in a value of a container's environment.
+// file's content, and in a value of a container's or a process's environment.
+//
+// **A process's environment is a container's** (novox/hq to-be 38 WP4c). A module's code moving out
+// of its container becomes a process on the machine and still has to be told what the container
+// was told; filled for one kind and not the other, the literal reached the process and was read as
+// a port, and the modules that moved first wrote their run-once steps a 0600 env file instead.
 //
 // A port the module did not say it listens on is refused, for the same reason a binding's unknown
 // key is: the module is asking about something it never declared, and the answer would be a guess.
@@ -84,7 +89,7 @@ func portInto(resource map[string]any, module string, listens []Listening, with
 		}
 		resource["content"] = filled
 
-	case "container":
+	case "container", "process":
 		env, ok := resource["env"].(map[string]any)
 		if !ok {
 			return nil
@@ -106,8 +111,8 @@ func portInto(resource map[string]any, module string, listens []Listening, with
 				continue
 			}
 			value, err := portsFilledInto(written,
-				fmt.Sprintf("%s's container %s sets %s to something that",
-					module, resource["name"], key), module, listens, with)
+				fmt.Sprintf("%s's %s %s sets %s to something that",
+					module, resource["type"], resource["name"], key), module, listens, with)
 			if err != nil {
 				return err
 			}

internal/catalogue/process_env_test.go

@@ -0,0 +1,80 @@
+package catalogue
+
+import (
+	"strings"
+	"testing"
+)
+
+// **A process's environment is composed as a container's is** (novox/hq to-be 38 WP4c).
+//
+// A module's code moving out of its container becomes a process on the machine, and what its
+// container's environment asked for — the port this machine gave the module, the place it put the
+// module's directory — it still has to be told. Filled for a container and not for a process, the
+// literal `${port:8080}` reached the process as its environment and was read as a port; the modules
+// that moved first wrote their run-once steps an env file instead.
+func processModule(env map[string]any) Manifest {
+	return Manifest{
+		Module:  "showcase",
+		Listens: []Listening{{Port: 8080, From: FromMesh}},
+		Resources: []map[string]any{
+			{"id": "data", "type": "directory", "mode": "0700"},
+			{"id": "setup", "type": "process", "name": "showcase-setup", "run-once": true,
+				"run": []any{"/usr/bin/showcase", "setup"}, "env": env},
+		},
+	}
+}
+
+func TestAProcessIsToldItsPortAndItsPlaceInItsEnvironment(t *testing.T) {
+	env := map[string]any{
+		"SHOWCASE_URL":  "http://127.0.0.1:${port:8080}",
+		"SHOWCASE_DATA": "${dir:data}/objects",
+		"SHOWCASE_DB":   "127.0.0.1:${seat:mesh-store:5432}",
+		"GREETING":      "hello",
+	}
+	out, err := Resolution{Node: "anchor", Modules: []Manifest{processModule(env)}}.Declaration(Rendering{
+		Ports: map[string]map[int]int{"showcase": {8080: 21000}},
+		Seats: map[string]map[int]int{"mesh-store": {5432: 6852}},
+	})
+	if err != nil {
+		t.Fatalf("a process asking for its port and its place does not compose: %v", err)
+	}
+	setup := fileNamed(out, "showcase.setup")
+	if setup == nil {
+		t.Fatalf("the process is not in the declaration: %v", out)
+	}
+	got, _ := setup["env"].(map[string]any)
+	for key, want := range map[string]string{
+		"SHOWCASE_URL":  "http://127.0.0.1:21000",
+		"SHOWCASE_DATA": "/var/lib/showcase/data/objects",
+		"SHOWCASE_DB":   "127.0.0.1:6852",
+		"GREETING":      "hello",
+	} {
+		if got[key] != want {
+			t.Errorf("the process is told %s=%v, want %q", key, got[key], want)
+		}
+	}
+	if env["SHOWCASE_URL"] != "http://127.0.0.1:${port:8080}" {
+		t.Fatalf("composing for one machine edited the module's own manifest: %v", env)
+	}
+}
+
+// An unknown reference in a process's environment is refused as a container's is, naming the
+// process and the variable — left alone, it would reach the machine as a literal.
+func TestAProcessAskingAboutAnUndeclaredPortIsRefused(t *testing.T) {
+	env := map[string]any{"SHOWCASE_URL": "http://127.0.0.1:${port:9999}"}
+	_, err := Resolution{Node: "anchor", Modules: []Manifest{processModule(env)}}.Declaration(Rendering{})
+	if err == nil {
+		t.Fatal("a process was told a port its module never said it listens on")
+	}
+	for _, said := range []string{"showcase-setup", "SHOWCASE_URL", "${port:9999}", "8080"} {
+		if !strings.Contains(err.Error(), said) {
+			t.Errorf("the refusal does not say %q: %v", said, err)
+		}
+	}
+
+	env = map[string]any{"SHOWCASE_DATA": "${dir:date}/objects"}
+	if _, err := (Resolution{Node: "anchor", Modules: []Manifest{processModule(env)}}).Declaration(Rendering{}); err == nil ||
+		!strings.Contains(err.Error(), "${dir:date}") {
+		t.Fatalf("a process naming no directory of its module was not refused: %v", err)
+	}
+}

internal/catalogue/seat_into.go

@@ -43,8 +43,8 @@ import (
 var ofSeat = regexp.MustCompile(`\$\{seat:([a-z0-9][a-z0-9-]*):([0-9]+)\}`)
 
 // seatInto replaces a resource's ${seat:…} placeholders with where this machine put each seat's
-// holder — in a file's content, and in a value of a container's environment. The same two places
-// portInto fills, for the same reason: they are where a process reads a number from.
+// holder — in a file's content, and in a value of a container's or a process's environment. The
+// same places portInto fills, for the same reason: they are where a program reads a number from.
 func seatInto(resource map[string]any, module string, with Rendering) error {
 	switch fmt.Sprint(resource["type"]) {
 	case "file":
@@ -58,7 +58,7 @@ func seatInto(resource map[string]any, module string, with Rendering) error {
 		}
 		resource["content"] = filled
 
-	case "container":
+	case "container", "process":
 		env, ok := resource["env"].(map[string]any)
 		if !ok {
 			return nil
@@ -78,8 +78,8 @@ func seatInto(resource map[string]any, module string, with Rendering) error {
 				continue
 			}
 			value, err := seatsFilledInto(written,
-				fmt.Sprintf("%s's container %s sets %s to something that",
-					module, resource["name"], key), with)
+				fmt.Sprintf("%s's %s %s sets %s to something that",
+					module, resource["type"], resource["name"], key), with)
 			if err != nil {
 				return err
 			}

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,
-		Reference: ArtifactStoreScheme + "mesh-controller/server@sha256:" + strings.Repeat("c", 64),
+		Name: "controller", Kind: ArtifactBundle,
+		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) {
-		t.Errorf("the control plane's own image is %v, not routed through the store", got)
+	if got := server["source"]; got != "anchor.internal:5100/mesh-controller/controller@sha256:"+strings.Repeat("c", 64) {
+		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/while_stopped_test.go

@@ -0,0 +1,89 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"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)
+		}
+	}
+}

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
-	// 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
-	// module in the meantime would stop the whole pipeline to make a point the record already makes.
-	// Before the runtime exists the pattern is accepted as it always was.
+	// returning by habit. Before the runtime exists the pattern is accepted as it always was.
+	//
+	// *Since 2026-10-04 (to-be 38 WP4b's last step):* refused for **every** module. While some
+	// thirty modules still stood in that shape, one already registered so was rebuilt without
+	// 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)
@@ -90,13 +91,7 @@ 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)
-				}
+				return fmt.Errorf("%s is not registered: %s", m.Module, why)
 			}
 		}
 	}
@@ -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
-// already standing in that shape is rebuilt as before, so the catalogue's pipeline keeps running
-// while each moves (WP3's amendment). Before the runtime, it is accepted as it always was — so a
+// to-be 38 WP2.4) — for every module, since every module has moved (WP4b's last step; WP3's
+// amendment let one already standing in that shape be rebuilt while each moved). Before the
+// 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
-	// amended by WP3): some thirty of them stand the day the runtime arrives, and each moves in its
-	// 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 {
-		t.Fatalf("a rebuild of a module that already had the pattern was refused: %v", err)
+	// **A module already registered in that shape is refused too** (WP4b's last step): every module
+	// has moved, and a rebuild in the old shape is one moving back.
+	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 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/hold.go

@@ -87,3 +87,62 @@ func (i *Inventory) tryHold(ctx context.Context, sorted []string) (func(), strin
 	}
 	return release, "", nil
 }
+
+// ErrPlansBusy is the plans held by another act — on a machine replacing its controller, the other
+// controller — for longer than a caller waits, or at all for one that does not wait.
+var ErrPlansBusy = errors.New("another controller is working the plans")
+
+// HoldPlans makes working the plans one act at a time, across every controller on the store
+// (novox/hq issue 213). A plan is read, changed and written whole; two controllers doing that at
+// once — the old and the new for the moment a machine hands its controller over, or a controller
+// and a person's `plans stop` — each act on what the other has not saved yet: a tier asked twice,
+// an outcome written over. A session-level advisory lock on one connection, released by the
+// returned function and by the session ending, so a controller that dies holding it holds nothing.
+//
+// wait false gives ErrPlansBusy at once when another holds them — the timer's way: the holder is
+// moving the plans already. wait true looks again every HoldPoll for up to HoldWaitFor — an
+// outcome's or a merge's way, which must be written.
+func (i *Inventory) HoldPlans(ctx context.Context, wait bool) (func(), error) {
+	deadline := time.Now().Add(HoldWaitFor)
+	for {
+		release, took, err := i.tryLock(ctx, "mesh-plans")
+		if err != nil || took {
+			return release, err
+		}
+		if !wait || time.Now().After(deadline) {
+			return nil, ErrPlansBusy
+		}
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		case <-time.After(HoldPoll):
+		}
+	}
+}
+
+// tryLock takes one named advisory lock on a connection of its own, or gives the connection back.
+func (i *Inventory) tryLock(ctx context.Context, key string) (func(), bool, error) {
+	conn, err := i.store.Pool().Acquire(ctx)
+	if err != nil {
+		return nil, false, err
+	}
+	var once sync.Once
+	release := func() {
+		once.Do(func() {
+			if _, err := conn.Exec(context.WithoutCancel(ctx), `select pg_advisory_unlock_all()`); err != nil {
+				_ = conn.Conn().Close(context.WithoutCancel(ctx))
+			}
+			conn.Release()
+		})
+	}
+	var took bool
+	if err := conn.QueryRow(ctx, `select pg_try_advisory_lock(hashtext($1)::bigint)`, key).Scan(&took); err != nil {
+		release()
+		return nil, false, err
+	}
+	if !took {
+		release()
+		return nil, false, nil
+	}
+	return release, true, nil
+}

internal/inventory/hold_plans_test.go

@@ -0,0 +1,38 @@
+package inventory
+
+import (
+	"errors"
+	"testing"
+	"time"
+)
+
+// novox/hq issue 213: while a machine hands its controller over from the container to the process,
+// two controllers run on one store for a moment. Working the plans is one act at a time across them.
+func TestThePlansAreWorkedByOneControllerAtATime(t *testing.T) {
+	first := ForTest(t)
+	// A second controller: its own connections to the same store.
+	second, err := Open(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(second.Close)
+
+	release, err := first.HoldPlans(t.Context(), false)
+	if err != nil {
+		t.Fatalf("the plans could not be held when nobody held them: %v", err)
+	}
+	t.Cleanup(release) // a pool closing waits for a connection still held; release is idempotent
+	if _, err := second.HoldPlans(t.Context(), false); !errors.Is(err, ErrPlansBusy) {
+		t.Fatalf("a second controller held the plans while the first did: %v", err)
+	}
+	// A waiter gets them once they are let go.
+	was := HoldPoll
+	HoldPoll = 10 * time.Millisecond
+	defer func() { HoldPoll = was }()
+	go func() { time.Sleep(50 * time.Millisecond); release() }()
+	again, err := second.HoldPlans(t.Context(), true)
+	if err != nil {
+		t.Fatalf("a waiting controller never got the plans once they were let go: %v", err)
+	}
+	again()
+}

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/link/receive_nats.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"log"
 	"strings"
 	"time"
 
@@ -78,10 +79,15 @@ func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Con
 	// than creating one here: the consumer is an object with a configuration — ack policy, ack
 	// wait, redelivery — and a client that creates its own would be a second opinion about it.
 	control := make(chan *nats.Msg, Prefetch)
-	said, err := js.ChanSubscribe("", control, nats.Bind("CONTROL", broker.ControllerName))
+	said, err := standingBy(ctx, log.Default(), "CONTROL", func() (*nats.Subscription, error) {
+		return js.ChanSubscribe("", control, nats.Bind("CONTROL", broker.ControllerName))
+	})
 	if err != nil {
 		return fmt.Errorf("subscribing to what nodes say: %w", err)
 	}
+	if said == nil {
+		return nil // stopped while standing by
+	}
 	defer func() { _ = said.Unsubscribe() }()
 
 	// Heartbeats, on core NATS and off any stream (design 25 §3). Their own subscription because
@@ -97,10 +103,15 @@ func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Con
 	var events chan *nats.Msg
 	if len(n.follows) > 0 {
 		events = make(chan *nats.Msg, Prefetch)
-		followed, err := js.ChanSubscribe("", events, nats.Bind("EVENTS", broker.ControllerName))
+		followed, err := standingBy(ctx, log.Default(), "EVENTS", func() (*nats.Subscription, error) {
+			return js.ChanSubscribe("", events, nats.Bind("EVENTS", broker.ControllerName))
+		})
 		if err != nil {
 			return fmt.Errorf("subscribing to what the catalogue says: %w", err)
 		}
+		if followed == nil {
+			return nil
+		}
 		defer func() { _ = followed.Unsubscribe() }()
 	}
 
@@ -324,3 +335,44 @@ func (m *natsControl) forget() {
 type replyAddressed struct {
 	ReplyTo string `json:"reply_to,omitempty"`
 }
+
+// StandbyPoll is how often a controller standing by looks again for its consumers. A variable so a
+// test need not wait.
+var StandbyPoll = 2 * time.Second
+
+// standingBy binds one of the controller's consumers, waiting while another controller holds it.
+//
+// **Two controllers, one consumer** (novox/hq issue 213). The controller's consumers are push
+// consumers with no delivery group, so the server lets one subscription bind each — on purpose:
+// two would each act on every message (issue 146). When a machine hands its controller over from
+// the container to the process, the host starts the process first and removes the container only
+// once the process is up; the process then finds the consumers bound. Exiting on that would never
+// be up, so the container would never go. It stands by instead — the seat's verbs are already
+// served from a queue group, and the plans wait on their lock — and binds as soon as the other lets
+// go. Nil and no error is ctx ending while it waited.
+func standingBy(ctx context.Context, logger interface{ Printf(string, ...any) }, stream string,
+	bind func() (*nats.Subscription, error)) (*nats.Subscription, error) {
+	said := false
+	for {
+		sub, err := bind()
+		if err == nil {
+			if said {
+				logger.Printf("took the controller's consumer on %s: the controller that held it let go", stream)
+			}
+			return sub, nil
+		}
+		if !strings.Contains(err.Error(), "already bound") {
+			return nil, err
+		}
+		if !said {
+			logger.Printf("another controller holds the controller's consumer on %s; standing by "+
+				"until it lets go", stream)
+			said = true
+		}
+		select {
+		case <-ctx.Done():
+			return nil, nil
+		case <-time.After(StandbyPoll):
+		}
+	}
+}

internal/link/standby_nats_test.go

@@ -0,0 +1,69 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+)
+
+// novox/hq issue 213: while a machine hands its controller over, the new controller (the process)
+// is started while the old one (the container) still holds the controller's consumers. It must not
+// exit — the host would read that as a replacement that did not come up and never remove the
+// container — and must not act on what the old one is handed. It stands by, and takes the consumers
+// when the old one lets go.
+func TestNatsASecondControllerStandsByAndTakesOverWhenTheFirstLetsGo(t *testing.T) {
+	js := aBus(t)
+	was := StandbyPoll
+	StandbyPoll = 50 * time.Millisecond
+	defer func() { StandbyPoll = was }()
+
+	old := &counted{}
+	_, stopOld := servingOn(t, js, old)
+	eventually(t, "the first controller binding its consumer", func() bool {
+		info, err := js.Context().ConsumerInfo("CONTROL", broker.ControllerName)
+		return err == nil && info.PushBound
+	})
+
+	// The new one, on a connection of its own as the process would have.
+	second, err := broker.Dial(os.Getenv("MESH_TEST_NATS"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(second.Close)
+	fresh := &counted{}
+	s := &Server{inbound: Nats(second), bus: OverNATS{Conn: second.Conn(), JS: second.Context()},
+		listener: fresh, log: quiet()}
+	ctx, stopNew := context.WithCancel(context.Background())
+	defer stopNew()
+	ended := make(chan error, 1)
+	go func() { ended <- s.Serve(ctx) }()
+
+	select {
+	case err := <-ended:
+		t.Fatalf("the second controller stopped instead of standing by: %v", err)
+	case <-time.After(500 * time.Millisecond):
+	}
+
+	report := func(declared string) {
+		body, _ := json.Marshal(Report{Node: "anchor", Declared: declared, Applied: []string{"store"}})
+		if _, err := js.Context().Publish(ReportSubject("anchor"), body); err != nil {
+			t.Fatal(err)
+		}
+	}
+	report("d1")
+	eventually(t, "the holding controller hearing the report", func() bool { return old.count() == 1 })
+	if fresh.count() != 0 {
+		t.Fatal("the controller standing by acted on a report the holder was handed")
+	}
+
+	stopOld()
+	report("d2")
+	eventually(t, "the second controller taking over once the first let go", func() bool { return fresh.count() == 1 })
+	if old.count() != 1 {
+		t.Errorf("the first controller heard %d reports", old.count())
+	}
+}

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",
-      "type": "container",
+      "id": "controller",
+      "type": "process",
       "name": "mesh-controller",
-      "network": "host",
-      "args": [
+      "artifact": "controller",
+      "run": [
+        "./mesh-controller",
         "serve"
       ],
+      "user": "mesh-controller",
       "env": {
-        "MESH_BROKER_CERTIFICATE": "/broker-tls/tls.crt",
-        "MESH_STORE_INVENTORY_FILE": "/run/secrets/inventory",
-        "MESH_STORE_IDENTITY_FILE": "/run/secrets/identity",
-        "MESH_STORE_LICENCES_FILE": "/run/secrets/licences",
-        "MESH_BROKER_MANAGEMENT_FILE": "/run/secrets/broker-management",
-        "MESH_BROKER_ADDRESS_FILE": "/run/secrets/broker-address",
+        "MESH_BROKER_CERTIFICATE": "/var/lib/mesh-broker-tls/tls.crt",
+        "MESH_STORE_INVENTORY_FILE": "${dir:mesh-state}/inventory",
+        "MESH_STORE_IDENTITY_FILE": "${dir:mesh-state}/identity",
+        "MESH_STORE_LICENCES_FILE": "${dir:mesh-state}/licences",
+        "MESH_BROKER_MANAGEMENT_FILE": "${dir:mesh-state}/broker-management",
+        "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": "${dir:mesh-state}/bus"
       },
-      "volumes": [
-        "/var/lib/mesh-broker-tls:/broker-tls:ro",
-        "${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"
+      "replaces": [
+        "server"
       ]
     }
   ],
   "build": {
     "artifacts": [
       {
-        "name": "server",
-        "kind": "image",
-        "from": "Dockerfile"
-      }
-    ],
-    "on": [
-      {
-        "arg": "GO_BASE",
-        "image": "golang@sha256:8ac98ca534ac3f51e1f420a1dd2c15e74c75cfa0f23f3ad27eb5d7236c349a0c"
+        "name": "controller",
+        "kind": "bundle",
+        "language": "go",
+        "system": "arch",
+        "from": "cmd/mesh-controller",
+        "binary": "mesh-controller"
       }
     ]
   }