Back off before standing for the controller lease after a loss, and while the bus does not answer (issue 906)

A controller that lost its lease to a stalled bus exited and was started
again five seconds later, asserting every bucket and consumer on the bus
that had just stalled; a start that could not assert the lease bucket
exited at once, into the same loop. A candidate now waits one key's age
after the newest epoch lost in the last ten minutes, doubled per further
loss, at most two minutes, before it touches the bus; and retries its
stand in-process from 5 s doubling to 60 s instead of exiting.

cmd/mesh-controller/acting.go

@@ -30,7 +30,9 @@ import (
 //     closes at once and the process exits, so its service manager restarts it as a candidate (serve,
 //     in push.go). A renewal the bus does not answer in time is not a loss (novox/hq issue 477): the
 //     lease is kept, no new act starts once half the key's age passed unrenewed, and the next renewal
-//     made opens the gate again.
+//     made opens the gate again. A candidate backs off before it stands (novox/hq issue 906, lease_backoff.go):
+//     after an epoch lost lately, and while its bus does not answer it, in its own process rather than in a
+//     restart every five seconds.
 //   - **A command run at a shell** — `push` in the installer, the lab, a person repairing a mesh whose
 //     controller is down (issue 201) — acts **under the holder's epoch** when a controller holds the
 //     lease: it is the same mesh's word, composed and sent under the store's hold of each machine like
@@ -72,6 +74,9 @@ type actor struct {
 	// lease over in a second rather than in a renewal of five (novox/hq issue 591). A renewal is given
 	// the lease's own RenewEvery to answer however often it is made, so a slow bus does not lose it.
 	renewEvery, poll time.Duration
+	// lossBase, when set, is the backoff after a lost epoch in place of the key's age: a test's (novox/hq
+	// issue 906).
+	lossBase time.Duration
 }
 
 // theLease is this process's standing to the lease.
@@ -211,48 +216,73 @@ func (a *actor) serveUnderTheLease(ctx context.Context, inv *inventory.Inventory
 	a.mu.Lock()
 	a.serving = true
 	a.mu.Unlock()
-	// Its own connection, held as long as the lease, and named so (novox/hq issue 327).
-	js, err := broker.Dial(address, nats.Name(broker.ConnectionName+" lease"))
-	if err != nil {
-		return nil, fmt.Errorf("the mesh is on the bus at %s and this control plane cannot reach it to take the "+
-			"lease: %w", broker.BareAddress(address), err)
-	}
-	api, err := jetstream.New(js.Conn())
-	if err != nil {
-		js.Close()
-		return nil, err
-	}
-	asserting, cancel := context.WithTimeout(ctx, 10*time.Second)
-	err = broker.EnsureLeaseBucket(asserting, api)
-	cancel()
-	if err != nil {
-		js.Close()
-		return nil, err
-	}
 	say := func(format string, args ...any) { fmt.Printf(format+"\n", args...) }
-	l, err := lease.Open(ctx, api, broker.LeaseBucket, lease.Options{Holder: holderOf(instance),
-		RenewEvery: a.renewEvery, RenewWithin: lease.RenewEvery, Poll: a.poll, Floor: inv.HighestEpoch, Say: say, Health: controllerHealth, Moved: func(was, floor uint64) {
-			a.mu.Lock()
-			defer a.mu.Unlock()
-			a.reset = time.Now()
-			a.resetSaid = fmt.Sprintf("the lease bucket was at revision %d with epoch %d already issued: it was "+
-				"raised again from nothing (a bus whose data was replaced), and its revisions were moved past %d so "+
-				"no machine refuses the next epoch", was, floor, floor)
-		}})
+	// **Backing off before standing** (novox/hq issue 906, lease_backoff.go): after an epoch lost lately, before
+	// this process touches the bus; and while the bus does not answer it, here rather than in a restart.
+	if wait := a.backoffAfterLosses(ctx, inv, say); wait > 0 {
+		if err := sleepOrDone(ctx, wait); err != nil {
+			return nil, err
+		}
+	}
+	var (
+		js         *broker.JetStream
+		l          *lease.Lease
+		epoch      uint64
+		unwritable error
+	)
+	err = standUp(ctx, sleepOrDone, say, func(ctx context.Context) error {
+		unwritable = nil
+		if js == nil {
+			// Its own connection, held as long as the lease, and named so (novox/hq issue 327).
+			dialled, err := broker.Dial(address, nats.Name(broker.ConnectionName+" lease"))
+			if err != nil {
+				return fmt.Errorf("the mesh is on the bus at %s and this control plane cannot reach it to take the "+
+					"lease: %w", broker.BareAddress(address), err)
+			}
+			js = dialled
+		}
+		api, err := jetstream.New(js.Conn())
+		if err != nil {
+			return err
+		}
+		asserting, cancel := context.WithTimeout(ctx, 10*time.Second)
+		err = broker.EnsureLeaseBucket(asserting, api)
+		cancel()
+		if err != nil {
+			return err
+		}
+		l, err = lease.Open(ctx, api, broker.LeaseBucket, lease.Options{Holder: holderOf(instance),
+			RenewEvery: a.renewEvery, RenewWithin: lease.RenewEvery, Poll: a.poll, Floor: inv.HighestEpoch, Say: say, Health: controllerHealth, Moved: func(was, floor uint64) {
+				a.mu.Lock()
+				defer a.mu.Unlock()
+				a.reset = time.Now()
+				a.resetSaid = fmt.Sprintf("the lease bucket was at revision %d with epoch %d already issued: it was "+
+					"raised again from nothing (a bus whose data was replaced), and its revisions were moved past %d so "+
+					"no machine refuses the next epoch", was, floor, floor)
+			}})
+		if err != nil {
+			return err
+		}
+		epoch, err = l.Take(ctx)
+		if errors.Is(err, lease.ErrUnwritable) {
+			unwritable = err
+			return nil
+		}
+		// Any other error: whether another controller acts cannot be told, so this one does not act, and tries
+		// again.
+		return err
+	})
 	if err != nil {
-		js.Close()
+		if js != nil {
+			js.Close()
+		}
 		return nil, err
 	}
 	gone := make(chan struct{})
-	epoch, err := l.Take(ctx)
-	switch {
+	switch err := unwritable; {
 	case ctx.Err() != nil:
 		js.Close()
 		return nil, ctx.Err()
-	case err != nil && !errors.Is(err, lease.ErrUnwritable):
-		// Whether another controller acts cannot be told: this one does not act, and exits to try again.
-		js.Close()
-		return nil, err
 	case err != nil:
 		// Nobody holds it and the bus will not let it be written: unleased, said, tried again (see above).
 		a.mu.Lock()

cmd/mesh-controller/builder_says_default_test.go

@@ -0,0 +1,9 @@
+package main
+
+import "context"
+
+// A test's builder says it writes references, as R4e's does, unless the test says otherwise (issue 1886): no test reads
+// the bus for it.
+func init() {
+	builderSaysReferences = func(context.Context) (bool, error) { return true, nil }
+}

cmd/mesh-controller/builder_says_test.go

@@ -0,0 +1,243 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	sdkdelivery "git.novox.be/novox/mesh-sdk/go/delivery"
+
+	"github.com/novox/mesh-controller/internal/builderkeys"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// theBuilderSays sets what the running builder says for one test.
+func theBuilderSays(t *testing.T, says bool, err error) {
+	t.Helper()
+	was := builderSaysReferences
+	builderSaysReferences = func(context.Context) (bool, error) { return says, err }
+	t.Cleanup(func() { builderSaysReferences = was })
+}
+
+// askingBuildAgent records the modules build-agent is asked for, by path.
+func askingBuildAgent(t *testing.T) *[]string {
+	t.Helper()
+	var asked []string
+	was := askABuild
+	askABuild = func(_ context.Context, _ buildSource, path, _ string) (string, error) {
+		asked = append(asked, path)
+		return "build-9", nil
+	}
+	t.Cleanup(func() { askABuild = was })
+	return &asked
+}
+
+// 1886/AC2: while no running builder says it writes references, a walk asks build-agent for a module of its tier, as
+// before R4e, and takes nothing from the builder's key.
+func TestAWalkAsksBuildAgentWhileTheBuilderDoesNotSayItWritesReferences(t *testing.T) {
+	theBuilderSays(t, false, nil)
+	asked := askingBuildAgent(t)
+	p, byName := aWalkOfOne("sudo")
+	askModule(t.Context(), p, "sudo", byName)
+	s := p.Modules["sudo"]
+	if s.Key || s.Build != "build-9" || s.State != "asked" || len(*asked) != 1 {
+		t.Fatalf("the module is not asked of build-agent: %+v, asked %v", s, *asked)
+	}
+}
+
+// 1886/AC3: once a running builder says it writes references, a walk takes the module from the builder's key (ADR
+// 0498 §8) and asks build-agent for nothing.
+func TestAWalkTakesTheKeyOnceTheBuilderSaysItWritesReferences(t *testing.T) {
+	theBuilderSays(t, true, nil)
+	asked := askingBuildAgent(t)
+	p, byName := aWalkOfOne("sudo")
+	askModule(t.Context(), p, "sudo", byName)
+	if s := p.Modules["sudo"]; !s.Key || s.Build != "" || len(*asked) != 0 {
+		t.Fatalf("the module is not taken from its key: %+v, asked %v", s, *asked)
+	}
+}
+
+// 1886/AC2: a say that cannot be read leaves the module unasked and says why, so the next look asks it: neither path is
+// guessed.
+func TestASayNotReadLeavesTheModuleUnasked(t *testing.T) {
+	theBuilderSays(t, false, errors.New("the bus did not answer"))
+	asked := askingBuildAgent(t)
+	p, byName := aWalkOfOne("sudo")
+	askModule(t.Context(), p, "sudo", byName)
+	s := p.Modules["sudo"]
+	if s.State != "" || s.Key || len(*asked) != 0 || !strings.Contains(s.Why, "the bus did not answer") {
+		t.Fatalf("a say not read: %+v, asked %v", s, *asked)
+	}
+	if got := unaskedOf(*p); len(got) != 1 {
+		t.Fatalf("the module is not left for the next look to ask: %v", got)
+	}
+}
+
+// 1886/AC2: the one function every ask goes through lets build-agent build any module while the builder does not say
+// it writes references, and refuses it again once it does.
+func TestBuildAgentIsAskedForAnyModuleWhileTheBuilderDoesNotSay(t *testing.T) {
+	was := askedModuleOf
+	t.Cleanup(func() { askedModuleOf = was })
+	askedModuleOf = func(_ context.Context, _ buildSource, path string) (string, error) {
+		return strings.TrimPrefix(path, "modules/"), nil
+	}
+	theBuilderSays(t, false, nil)
+	if err := onlyTheBuildersOwn(t.Context(), buildSource{Repository: "novox/mesh-catalog", Seat: "git"}, "modules/sudo"); err != nil {
+		t.Fatalf("an ask while the builder does not say is refused: %v", err)
+	}
+	theBuilderSays(t, true, nil)
+	if err := onlyTheBuildersOwn(t.Context(), buildSource{Repository: "novox/mesh-catalog", Seat: "git"}, "modules/sudo"); !errors.Is(err, errNotTheBuildersOwn) {
+		t.Fatalf("an ask once the builder says is not refused: %v", err)
+	}
+}
+
+// 1886/AC4: an open walk that waits on a key the running builder cannot reference (done without a reference, as R4e's
+// own walk did in plan-1792707174627079192) is judged again at the next look: the module is handed back unasked, and
+// that look's ask goes to build-agent. With the builder saying it writes references, the same walk keeps waiting.
+func TestAnOpenWalkWaitingOnAnUnreferencedKeyIsAskedOfBuildAgent(t *testing.T) {
+	t.Cleanup(func() {
+		keyTakes.Lock()
+		keyTakes.running, keyTakes.waits, keyTakes.taken = map[string]bool{}, map[string]string{}, map[string]bool{}
+		keyTakes.Unlock()
+	})
+	fp := "src2:" + strings.Repeat("e", 64)
+	unreferenced := builderkeys.Key{Module: "sudo", Fingerprint: fp, Found: true, Revision: 1,
+		Build: sdkdelivery.BuilderBuild{State: sdkdelivery.BuildDone, Manifest: strings.Repeat("b", 64),
+			Artifacts: []sdkdelivery.BuiltArtifact{{Name: "tools", Kind: "bundle", Digest: "sha256:1"}}}}
+	walk := func() (*inventory.Plan, map[string]inventory.Entry, planKeysOf) {
+		theBuilderSays(t, true, nil)
+		p, byName := aWalkOfOne("sudo")
+		askModule(t.Context(), p, "sudo", byName)
+		keys := planKeysOf{values: map[string]sdkdelivery.Module{"sudo@c0ffee0123456789": {FP2: fp}},
+			keys: map[string]builderkeys.Key{"sudo": unreferenced}}
+		return p, byName, keys
+	}
+	take := func(string, sdkdelivery.Module, string) { t.Fatal("an unreferenced key was taken in") }
+	records := recordsOf{builds: map[string]inventory.Build{}}
+
+	// Said: it waits for the reference, as ADR 0498 §8 has it.
+	p, _, keys := walk()
+	if _, err := settleKeys(t.Context(), records, keys, p, p.Tiers[0], take); err != nil {
+		t.Fatal(err)
+	}
+	if s := p.Modules["sudo"]; s.State != "asked" || !s.Key || !strings.Contains(s.Why, "waits for the reference") {
+		t.Fatalf("with the builder saying, the walk does not wait for the reference: %+v", s)
+	}
+
+	// Not said: handed back, and the next look asks build-agent.
+	p, byName, keys := walk()
+	theBuilderSays(t, false, nil)
+	asked := askingBuildAgent(t)
+	changed, err := settleKeys(t.Context(), records, keys, p, p.Tiers[0], take)
+	if err != nil || !changed {
+		t.Fatalf("the walk was not judged again: %v %v", changed, err)
+	}
+	s := p.Modules["sudo"]
+	if s.Key || s.State != "" || p.State != inventory.PlanBuilding || !strings.Contains(s.Why, "does not say") {
+		t.Fatalf("the module is not handed back to be asked of build-agent: %+v %s", s, p.State)
+	}
+	for _, m := range unaskedOf(*p) {
+		askModule(t.Context(), p, m, byName)
+	}
+	if s := p.Modules["sudo"]; s.Key || s.Build != "build-9" || len(*asked) != 1 {
+		t.Fatalf("the next look did not ask build-agent: %+v, asked %v", s, *asked)
+	}
+}
+
+// 1886/AC4: a key the running builder failed while it does not say it writes references fails no walk: the module is
+// asked of build-agent, whose build is the one delivered.
+func TestAFailedKeyOfABuilderThatDoesNotSayFailsNoWalk(t *testing.T) {
+	theBuilderSays(t, true, nil)
+	p, byName := aWalkOfOne("sudo")
+	askModule(t.Context(), p, "sudo", byName)
+	_ = byName
+	fp := "src2:" + strings.Repeat("f", 64)
+	theBuilderSays(t, false, nil)
+	keys := planKeysOf{values: map[string]sdkdelivery.Module{"sudo@c0ffee0123456789": {FP2: fp}},
+		keys: map[string]builderkeys.Key{"sudo": {Module: "sudo", Fingerprint: fp, Found: true,
+			Build: sdkdelivery.BuilderBuild{State: sdkdelivery.BuildFailed, Error: "not stored: tools: no reference"}}}}
+	if _, err := settleKeys(t.Context(), recordsOf{}, keys, p, p.Tiers[0], func(string, sdkdelivery.Module, string) {}); err != nil {
+		t.Fatal(err)
+	}
+	if s := p.Modules["sudo"]; p.State != inventory.PlanBuilding || s.State != "" || s.Key {
+		t.Fatalf("a failed key of a builder that does not say failed the walk: %s %q %+v", p.State, p.Note, s)
+	}
+}
+
+// 1886/AC6 (ADR 0522 §6): the parts that write references, the builder and the build modules, are build-agent's own until
+// R4g: asked of build-agent and never taken from keys they write themselves, whatever the builder says.
+func TestTheReferenceWritersAreAskedOfBuildAgentWhateverTheBuilderSays(t *testing.T) {
+	theBuilderSays(t, true, nil)
+	asked := askingBuildAgent(t)
+	for _, m := range []string{"mesh-build-go", "mesh-build-node", "mesh-build-image", "mesh-delivery-builder"} {
+		p, byName := aWalkOfOne(m)
+		askModule(t.Context(), p, m, byName)
+		if s := p.Modules[m]; s.Key || s.Build != "build-9" {
+			t.Errorf("%s is not asked of build-agent: %+v", m, s)
+		}
+	}
+	if len(*asked) != 4 {
+		t.Fatalf("build-agent was asked %v", *asked)
+	}
+}
+
+// 1886 review round 1 (ADR 0522 §2): the guard of the one ask function judges by the say askModule read, never a second
+// read: a second read that fails, or that says otherwise, fails no plan.
+func TestTheAskGuardUsesTheSayTheWalkRead(t *testing.T) {
+	was := askedModuleOf
+	t.Cleanup(func() { askedModuleOf = was })
+	askedModuleOf = func(_ context.Context, _ buildSource, path string) (string, error) {
+		return strings.TrimPrefix(path, "modules/"), nil
+	}
+	reads := 0
+	wasSays := builderSaysReferences
+	t.Cleanup(func() { builderSaysReferences = wasSays })
+	for _, second := range []struct {
+		name string
+		says bool
+		err  error
+	}{{"a second read that fails", false, errors.New("the bus did not answer")}, {"a say that turned true", true, nil}} {
+		reads = 0
+		builderSaysReferences = func(context.Context) (bool, error) {
+			reads++
+			if reads == 1 {
+				return false, nil
+			}
+			return second.says, second.err
+		}
+		wasAsk := askABuild
+		askABuild = func(ctx context.Context, source buildSource, path, _ string) (string, error) {
+			if err := onlyTheBuildersOwn(ctx, source, path); err != nil {
+				return "", err
+			}
+			return "build-9", nil
+		}
+		p, byName := aWalkOfOne("sudo")
+		askModule(t.Context(), p, "sudo", byName)
+		askABuild = wasAsk
+		if s := p.Modules["sudo"]; p.State == inventory.PlanFailed || s.Build != "build-9" || s.Key {
+			t.Errorf("%s: the walk %s, the module %+v", second.name, p.State, s)
+		}
+	}
+}
+
+// 1886 merge check hang: the look must judge an ask and a settle by one say. A settle that read the say from another
+// source than the ask (the bus's bucket, against the walk's read) handed a module back, the next ask took it from its key
+// again, and the look's loop never ended: walk_from_keys_integration_test hung.
+func TestTheSettleJudgesByTheSameSayAsTheAsk(t *testing.T) {
+	theBuilderSays(t, true, nil)
+	p, _ := aWalkOfOne("sudo")
+	p.Modules["sudo"] = &inventory.PlanModule{}
+	askModule(t.Context(), p, "sudo", map[string]inventory.Entry{"sudo": {Source: inventory.Source{Repository: "novox/mesh-catalog", Path: "modules/sudo", Ref: "main"}}})
+	fp := "src2:" + strings.Repeat("1", 64)
+	keys := planKeysOf{values: map[string]sdkdelivery.Module{"sudo@c0ffee0123456789": {FP2: fp}},
+		keys: map[string]builderkeys.Key{"sudo": {Module: "sudo", Fingerprint: fp, Found: true,
+			Build: sdkdelivery.BuilderBuild{State: sdkdelivery.BuildBuilding, By: "shanks"}}}}
+	if _, err := settleKeys(t.Context(), recordsOf{}, keys, p, p.Tiers[0], func(string, sdkdelivery.Module, string) {}); err != nil {
+		t.Fatal(err)
+	}
+	if s := p.Modules["sudo"]; !s.Key || s.State != "asked" {
+		t.Fatalf("the settle judged by another say than the ask's: %+v", s)
+	}
+}

cmd/mesh-controller/lease_backoff.go

@@ -0,0 +1,117 @@
+package main
+
+import (
+	"context"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// Backing off before standing (novox/hq issue 906).
+//
+// A controller loses its lease when its bus stops answering for longer than the key's age: on the control-node,
+// a bus whose disks are saturated has taken 30 to 42 seconds to answer (2026-10-23 05:52). The process then
+// exits, as ADR 0229 rule 4 says, and its service manager starts it again five seconds later. **A start is the
+// heaviest thing a controller asks of the bus** — every bucket, stream and consumer asserted again — and a start
+// that cannot assert the lease's bucket used to exit at once and be started again five seconds later, on the
+// same overloaded bus, nine times in twelve minutes (2026-10-16). So a candidate backs off twice:
+//
+//   - **After a loss**, before it touches the bus: from the newest epoch lost in the last ten minutes (the
+//     controller's store says it, rule 7), it waits one key's age, doubled for every further loss in that
+//     window, at most two minutes. A single loss costs a key's age; a controller losing it again and again
+//     stands less and less often.
+//   - **While the bus does not answer it**, in its own process: asserting the lease's bucket, opening it or
+//     reading who holds it is tried again after 5 s, doubling to at most 60 s, instead of exiting into a restart
+//     every five seconds. Nothing acts meanwhile: the process holds no lease, and the gate refuses every act.
+
+const (
+	// lossWindow is how far back lost epochs count towards the backoff.
+	lossWindow = 10 * time.Minute
+	// lossMost is the longest a candidate waits after losses.
+	lossMost = 2 * time.Minute
+	// standFirst and standMost are the waits between tries of a candidate the bus does not answer.
+	standFirst = 5 * time.Second
+	standMost  = time.Minute
+)
+
+// lossBackoff is how long a candidate waits now before it stands, after the epochs it read: one key's age from
+// the newest loss in the window, doubled for every further loss in it, at most lossMost.
+func lossBackoff(epochs []inventory.Epoch, now time.Time) time.Duration {
+	return lossBackoffFrom(epochs, now, broker.LeaseTTL)
+}
+
+func lossBackoffFrom(epochs []inventory.Epoch, now time.Time, base time.Duration) time.Duration {
+	lost, newest := lostSince(epochs, now)
+	if lost == 0 {
+		return 0
+	}
+	delay := base
+	for i := 1; i < lost && delay < lossMost; i++ {
+		delay *= 2
+	}
+	delay = min(delay, lossMost)
+	return max(newest.Add(delay).Sub(now), 0)
+}
+
+// standUp runs try until it answers nil, waiting standFirst between tries and doubling to standMost; it answers
+// ctx's error once ctx ends.
+func standUp(ctx context.Context, sleep func(context.Context, time.Duration) error, say func(string, ...any),
+	try func(context.Context) error) error {
+	wait := standFirst
+	for {
+		err := try(ctx)
+		if err == nil {
+			return nil
+		}
+		if ctx.Err() != nil {
+			return ctx.Err()
+		}
+		say("this controller cannot stand for the lease yet: %v — it tries again in %s, in this process, rather "+
+			"than start again on a bus that does not answer (novox/hq issue 906)", err, wait)
+		if err := sleep(ctx, wait); err != nil {
+			return err
+		}
+		wait = min(wait*2, standMost)
+	}
+}
+
+// backoffAfterLosses is how long this candidate waits before it stands, read from the epochs its store kept,
+// and said. A store that cannot be read makes it wait nothing: the backoff spares the bus, and never keeps a
+// controller from standing.
+func (a *actor) backoffAfterLosses(ctx context.Context, inv *inventory.Inventory, say func(string, ...any)) time.Duration {
+	now := time.Now()
+	reading, cancel := context.WithTimeout(ctx, 10*time.Second)
+	epochs, err := inv.EpochsSince(reading, now.Add(-lossWindow))
+	cancel()
+	if err != nil {
+		say("whether the lease was lost lately cannot be read, so this controller stands without backing off: %v", err)
+		return 0
+	}
+	base := a.lossBase
+	if base <= 0 {
+		base = broker.LeaseTTL
+	}
+	wait := lossBackoffFrom(epochs, now, base)
+	if wait > 0 {
+		lost, _ := lostSince(epochs, now)
+		say("the controller lease was lost lately (%d time(s) in the last %s): this controller backs off %s "+
+			"before it stands, so its start does not add to the load that lost it (novox/hq issue 906)",
+			lost, lossWindow, wait.Round(time.Second))
+	}
+	return wait
+}
+
+// lostSince is how many epochs were lost in the window before now, and when the newest of them was.
+func lostSince(epochs []inventory.Epoch, now time.Time) (lost int, newest time.Time) {
+	for _, e := range epochs {
+		if e.Ended == nil || e.How != inventory.EpochLost || now.Sub(*e.Ended) > lossWindow {
+			continue
+		}
+		lost++
+		if e.Ended.After(newest) {
+			newest = *e.Ended
+		}
+	}
+	return lost, newest
+}

cmd/mesh-controller/lease_backoff_test.go

@@ -0,0 +1,97 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"slices"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A controller that lost its lease to a slow bus backs off before it stands again (novox/hq issue 906): it
+// does not add its start — every bucket and consumer asserted again — to the load that lost it the lease.
+
+func lostAt(epoch uint64, ended time.Time, how string) inventory.Epoch {
+	return inventory.Epoch{Epoch: epoch, Instance: "controller@anchor", Taken: ended.Add(-time.Hour), Ended: &ended,
+		How: how}
+}
+
+// 906/AC1: given epochs lost in the last ten minutes, when a candidate starts, then it waits from the newest
+// loss one key's age, doubled for each further loss, at most two minutes; nothing else makes it wait.
+func TestACandidateAfterALostLeaseBacksOff(t *testing.T) {
+	now := time.Date(2026, 10, 23, 5, 52, 40, 0, time.UTC)
+	for _, c := range []struct {
+		name   string
+		epochs []inventory.Epoch
+		want   time.Duration
+	}{
+		{"nothing lost", nil, 0},
+		{"one held, one given back", []inventory.Epoch{
+			{Epoch: 3, Taken: now.Add(-time.Minute)},
+			lostAt(2, now.Add(-2*time.Second), inventory.EpochReleased),
+		}, 0},
+		{"one lost five seconds ago", []inventory.Epoch{lostAt(7, now.Add(-5*time.Second), inventory.EpochLost)},
+			10 * time.Second},
+		{"one lost long ago enough", []inventory.Epoch{lostAt(7, now.Add(-20*time.Second), inventory.EpochLost)}, 0},
+		{"three lost, the newest a second ago", []inventory.Epoch{
+			lostAt(9, now.Add(-time.Second), inventory.EpochLost),
+			lostAt(8, now.Add(-3*time.Minute), inventory.EpochLost),
+			lostAt(7, now.Add(-6*time.Minute), inventory.EpochLost),
+		}, 59 * time.Second},
+		{"six lost: at most two minutes", []inventory.Epoch{
+			lostAt(12, now, inventory.EpochLost), lostAt(11, now.Add(-time.Minute), inventory.EpochLost),
+			lostAt(10, now.Add(-2*time.Minute), inventory.EpochLost), lostAt(9, now.Add(-3*time.Minute), inventory.EpochLost),
+			lostAt(8, now.Add(-4*time.Minute), inventory.EpochLost), lostAt(7, now.Add(-5*time.Minute), inventory.EpochLost),
+		}, 2 * time.Minute},
+		{"a loss older than ten minutes does not count", []inventory.Epoch{
+			lostAt(9, now.Add(-time.Second), inventory.EpochLost),
+			lostAt(8, now.Add(-11*time.Minute), inventory.EpochLost),
+		}, 14 * time.Second},
+		{"expired is not lost", []inventory.Epoch{lostAt(9, now.Add(-time.Second), inventory.EpochExpired)}, 0},
+	} {
+		t.Run(c.name, func(t *testing.T) {
+			if got := lossBackoff(c.epochs, now); got != c.want {
+				t.Fatalf("a candidate after %d epoch(s) waits %s, want %s", len(c.epochs), got, c.want)
+			}
+		})
+	}
+}
+
+// 906/AC2: given a bus that does not answer a candidate, when it asserts the lease's bucket or reads who holds
+// it, then the candidate tries again in its own process after 5 s, doubling to at most 60 s, instead of exiting
+// to be restarted by its service manager every five seconds.
+func TestACandidateTheBusDoesNotAnswerTriesAgainWithBackoff(t *testing.T) {
+	var waited []time.Duration
+	sleep := func(ctx context.Context, d time.Duration) error { waited = append(waited, d); return nil }
+	tries := 0
+	slow := errors.New("asserting bucket mesh-controller_lease: context deadline exceeded")
+	err := standUp(t.Context(), sleep, func(string, ...any) {}, func(context.Context) error {
+		tries++
+		if tries < 7 {
+			return slow
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("a candidate whose bus answered at the seventh try gave up: %v", err)
+	}
+	want := []time.Duration{5 * time.Second, 10 * time.Second, 20 * time.Second, 40 * time.Second, time.Minute,
+		time.Minute}
+	if tries != 7 || !slices.Equal(waited, want) {
+		t.Fatalf("a candidate tried %d time(s), waiting %v between; want 7, waiting %v", tries, waited, want)
+	}
+}
+
+// 906/AC2: and a candidate stopping while it waits stops, with why.
+func TestACandidateStoppingWhileItBacksOffStops(t *testing.T) {
+	ctx, cancel := context.WithCancel(t.Context())
+	cancel()
+	err := standUp(ctx, sleepOrDone, func(string, ...any) {}, func(context.Context) error {
+		return errors.New("who holds the lease cannot be read: context deadline exceeded")
+	})
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("a candidate stopped while backing off answered %v", err)
+	}
+}

cmd/mesh-controller/lease_integration_test.go

@@ -152,3 +152,33 @@ func TestTwoControllersOneActs(t *testing.T) {
 		t.Fatalf("B's epoch does not say it was lost: %+v", ended)
 	}
 }
+
+// 906/AC1, on a real bus and store: a controller starting right after an epoch was lost backs off before it
+// stands — it takes the lease no sooner than the backoff after the loss, and then holds it as usual.
+func TestAControllerAfterALostLeaseBacksOffBeforeItStands(t *testing.T) {
+	url, _ := aBusForTheLease(t)
+	inv := inventory.ForTest(t)
+	ctx := t.Context()
+	// The controller before it lost epoch 5, and said so in the store, as `took` does on a loss.
+	if _, err := inv.TookEpoch(ctx, inventory.Epoch{Epoch: 5, Instance: "controller@anchor pid 1", Host: "anchor",
+		Taken: time.Now()}); err != nil {
+		t.Fatal(err)
+	}
+	if err := inv.EndEpoch(ctx, 5, inventory.EpochLost); err != nil {
+		t.Fatal(err)
+	}
+	// The key's age stands in as a backoff of two seconds, so the case runs in seconds (issue 591).
+	const backoff = 2 * time.Second
+	a := &actor{renewEvery: lease.RenewEvery / 5, poll: lease.Poll / 10, lossBase: backoff}
+	start := time.Now()
+	if _, err := a.serveUnderTheLease(ctx, inv, url); err != nil {
+		t.Fatal(err)
+	}
+	defer a.release()
+	if waited := time.Since(start); waited < backoff-300*time.Millisecond {
+		t.Fatalf("a controller took the lease %s after an epoch was lost, without backing off %s", waited, backoff)
+	}
+	if epoch, err := a.epoch(ctx); err != nil || epoch <= 5 {
+		t.Fatalf("after backing off the controller holds epoch %d (%v), want one above the lost 5", epoch, err)
+	}
+}

cmd/mesh-controller/new_modules.go

@@ -11,6 +11,7 @@ import (
 	"time"
 
 	sdkdelivery "git.novox.be/novox/mesh-sdk/go/delivery"
+	"github.com/nats-io/nats.go/jetstream"
 	"github.com/novox/mesh-controller/internal/builderkeys"
 	"github.com/novox/mesh-controller/internal/conditions"
 	"github.com/novox/mesh-controller/internal/inventory"
@@ -209,7 +210,12 @@ func watchNewModules(ctx context.Context, open *stores) {
 		fmt.Printf("new modules: the builder's keys are not watched: %v\n", err)
 		return
 	}
-	if err := builderkeys.Watch(ctx, keys.JS, func(module string) { wakeNewModule(ctx, open, module) }); err != nil {
+	says := watchingSays{js: keys.JS}
+	says.watch(ctx)
+	if err := builderkeys.Watch(ctx, keys.JS, func(module string) {
+		says.watch(ctx)
+		wakeNewModule(ctx, open, module)
+	}); err != nil {
 		closeKeys()
 		fmt.Printf("new modules: the builder's keys are not watched: %v\n", err)
 		return
@@ -272,3 +278,32 @@ func newModulesOf(ctx context.Context, repository string) ([]string, error) {
 	sort.Strings(out)
 	return out, nil
 }
+
+// watchingSays watches the builder's bucket of instances once it is there (issue 1886): a change of what the builder says
+// asks the look, which judges every open walk again. A builder from before R4e makes no such bucket; a write of the
+// builder's keys asks again whether it is there now, so nothing waits on a timer.
+type watchingSays struct {
+	js jetstream.JetStream
+	mu *sync.Mutex
+	on *bool
+}
+
+func (w *watchingSays) watch(ctx context.Context) {
+	if w.mu == nil {
+		w.mu, w.on = &sync.Mutex{}, new(bool)
+	}
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if *w.on {
+		return
+	}
+	on, err := builderkeys.WatchSays(ctx, w.js, askToLook)
+	if err != nil {
+		fmt.Printf("plans: what the builder says it writes is not watched now: %v\n", err)
+		return
+	}
+	if on {
+		*w.on = true
+		askToLook() // what it says now may differ from what the last look read
+	}
+}

cmd/mesh-controller/plan_keys.go

@@ -32,7 +32,12 @@ func isBuildersOwn(module string) bool { return buildersOwn(module) }
 // repository's; and mesh-tools, the toolchain image build-agent's builds stand on. The builder writes no key of any of them,
 // and a walk waiting for one would hold its tier for ever. mesh-tools-go is not among them: it builds from a Dockerfile in
 // its own directory on a pinned base, and has its key from mesh-build-image.
-var buildAgentsOwn = map[string]bool{sdkdelivery.BuilderModule: true, builder.AgentModule: true, "mesh-tools": true}
+//
+// **And the parts that write references** (ADR 0522 §6, issue 1886): the build modules push and answer each reference,
+// and the builder writes it into the key, so none of them is ever taken from a key they write themselves. Until R4g
+// build-agent is the way they are delivered whatever the builder says.
+var buildAgentsOwn = map[string]bool{sdkdelivery.BuilderModule: true, builder.AgentModule: true, "mesh-tools": true,
+	"mesh-build-go": true, "mesh-build-node": true, "mesh-build-image": true}
 
 // buildersOwn is isBuildersOwn's rule: a variable only so a test of the asks left (their bus waits, their plan key) may
 // name modules of its own as build-agent's.
@@ -44,6 +49,32 @@ type planKeyReads interface {
 	Build(ctx context.Context, module, fp string) (builderkeys.Key, error)
 }
 
+// builderSayKey carries, on an ask's context, the builder's say the walk read before asking (withBuilderSay).
+type builderSayKey struct{}
+
+// withBuilderSay hands the say a walk read to the ask's guard, so the guard never reads it a second time.
+func withBuilderSay(ctx context.Context, says bool) context.Context {
+	return context.WithValue(ctx, builderSayKey{}, says)
+}
+
+// builderSayOf is the say a walk handed the ask, and whether it handed one.
+func builderSayOf(ctx context.Context) (says, handed bool) {
+	says, handed = ctx.Value(builderSayKey{}).(bool)
+	return says, handed
+}
+
+// builderSaysReferences is whether a running builder instance says, in its own bucket of instances, that the keys it
+// writes name references (issue 1886). Until one does, a walk asks build-agent for every module, as before R4e: the
+// keys the running builder writes could never be delivered from. A variable so a test needs no bus.
+var builderSaysReferences = func(ctx context.Context) (bool, error) {
+	keys, closeKeys, err := openBuilderKeys()
+	if err != nil {
+		return false, err
+	}
+	defer closeKeys()
+	return keys.SaysReferences(ctx)
+}
+
 // openPlanKeys opens what a walk reads of the indexer and the builder. A variable so a test needs no bus.
 var openPlanKeys = func() (planKeyReads, func(), error) {
 	r, closeKeys, err := openBuilderKeys()
@@ -102,6 +133,20 @@ func settleKeys(ctx context.Context, inv keyRecords, keys planKeyReads, p *inven
 			changed = true
 		}
 	}
+	// One say for the ask and the settle (builderSaysReferences): two sources that differ hand a module back and take it
+	// from its key again at every turn of the look, which then never ends.
+	says, err := builderSaysReferences(ctx)
+	if err != nil {
+		return changed, fmt.Errorf("whether the builder writes references cannot be read: %w", err)
+	}
+	// **A walk never waits on a key the running builder cannot reference** (issue 1886): while no builder says it writes
+	// references, a module that waits on its key, or whose key failed, is handed back unasked, and the look asks it of
+	// build-agent. A done key with every reference is still taken.
+	toAgent := func(s *inventory.PlanModule, why string) {
+		s.State, s.Key, s.AskedAt, s.KeyAt, s.Build, s.Fingerprint = "", false, nil, "", "", ""
+		s.Why = "the builder does not say it writes references, so build-agent is asked (" + why + ")"
+		changed = true
+	}
 	for _, m := range tier {
 		s := p.Modules[m]
 		if s == nil || !s.Key || s.State != "asked" {
@@ -109,6 +154,10 @@ func settleKeys(ctx context.Context, inv keyRecords, keys planKeyReads, p *inven
 		}
 		trunk, err := keys.ModuleAt(ctx, m, s.KeyAt)
 		if errors.Is(err, builderkeys.ErrNoValue) {
+			if !says {
+				toAgent(s, "the indexer holds no value of it")
+				continue
+			}
 			set(s, fmt.Sprintf("waits for the indexer's value of %s at %s", m, orNow(s.KeyAt)))
 			continue
 		}
@@ -120,6 +169,10 @@ func settleKeys(ctx context.Context, inv keyRecords, keys planKeyReads, p *inven
 			if why == "" {
 				why = "the indexer wrote no fingerprint of the new form"
 			}
+			if !says {
+				toAgent(s, why)
+				continue
+			}
 			set(s, fmt.Sprintf("waits: %s at %s has no fingerprint to take a build by: %s", m, orNow(s.KeyAt), why))
 			continue
 		}
@@ -131,6 +184,9 @@ func settleKeys(ctx context.Context, inv keyRecords, keys planKeyReads, p *inven
 			return changed, err
 		}
 		switch {
+		case !says && (k.Failed() || !k.Done() || len(k.Unreferenced()) > 0):
+			toAgent(s, k.Says())
+			continue
 		case k.Failed():
 			// **A failed key fails the walk in the builder's own words** (`digests differ:`, `not stored:`, `not pinned:`,
 			// `version taken:`), and a rebuild of it is a new revision the next walk takes.

cmd/mesh-controller/plan_retry.go

@@ -203,6 +203,21 @@ func retryRefusal(p inventory.Plan, plans []inventory.Plan, onPath map[string]st
 		return oneWalkAtATime(p, plans, onPath)
 	}
 	stopped := stoppedRollouts(p)
+	// **A walk stopped by hand while its tier waited is resumed** (issue 1886): nothing in it failed, so nothing is
+	// asked again; it goes on from where it stopped, unless a newer walk holds what it has yet to send.
+	if len(stopped) == 0 && stoppedByHandWaiting(p) {
+		if q, found := newerOpenPlan(p, plans); found {
+			return fmt.Errorf("%s supersedes it: a newer merge of %s (%s at %s) is open, and resuming %s would build "+
+				"what that one replaced", q.ID, q.Repository, q.ID, short(q.Commit), p.ID)
+		}
+		for _, m := range holdingABuild(p) {
+			if q, found := newerPlanFor(m, p, plans); found {
+				return fmt.Errorf("%s has a newer plan, %s (%s, %s at %s): resuming %s would send the older build of it "+
+					"over the newer one", m, q.ID, q.State, q.Repository, short(q.Commit), p.ID)
+			}
+		}
+		return oneWalkAtATime(p, plans, onPath)
+	}
 	if len(stopped) == 0 {
 		return fmt.Errorf("nothing in tier %d of %s failed to build or stopped rolling out — it stopped at: %s",
 			p.Tier, p.ID, p.Note)
@@ -333,6 +348,8 @@ func plansForRetry(ctx context.Context, inv planReader, p inventory.Plan) ([]inv
 
 // retryPlan asks a failed plan's failed modules again, under new ids, and sets it building again at
 // that tier: what follows is the plan going on as if they had built the first time.
+// A walk stopped by hand while its tier waited, with nothing failed, is resumed instead: nothing asked again
+// (resumeStopped, issue 1886).
 func retryPlan(ctx context.Context, open *stores, id string) (string, error) {
 	inv := open.inventory
 	release, err := inv.HoldPlans(ctx, true)
@@ -388,6 +405,13 @@ func retryPlan(ctx context.Context, open *stores, id string) (string, error) {
 	if again := unjudgedAtGate(p); len(failed) == 0 && len(again) > 0 {
 		return retryTierWhole(ctx, open, &p, again)
 	}
+	if len(failed) == 0 && len(stoppedRollouts(p)) == 0 && stoppedByHandWaiting(p) {
+		said := resumeStopped(&p)
+		if err := inv.SavePlan(ctx, &p); err != nil {
+			return "", err
+		}
+		return said, nil
+	}
 	if len(failed) == 0 {
 		return retryRollouts(ctx, open, &p)
 	}
@@ -583,3 +607,89 @@ func toRetry(p *inventory.Plan, recorded map[string][]inventory.Build, byID map[
 	}
 	return failedIn(*p)
 }
+
+// stoppedByHandWaiting says a walk was stopped by hand while its current tier still waited (issue 1886): stopped by a
+// person, nothing of the tier failed, and a module of it asked, built and not sent, or not yet asked. A stop that left
+// merges out is not one: retryPlan refuses it before this is read (novox/hq ADR 0299).
+func stoppedByHandWaiting(p inventory.Plan) bool {
+	if p.State != inventory.PlanFailed || p.Delivery == nil || p.Delivery.Stopped == "" ||
+		p.Tier >= len(p.Tiers) || len(failedIn(p)) > 0 {
+		return false
+	}
+	asked, unsent, unasked := waitingIn(p)
+	return asked+unsent+unasked > 0
+}
+
+// waitingIn counts what a plan's current tier waits on: modules asked, built and not sent, not yet asked.
+func waitingIn(p inventory.Plan) (asked, unsent, unasked int) {
+	if p.Tier >= len(p.Tiers) {
+		return 0, 0, 0
+	}
+	for _, m := range p.Tiers[p.Tier] {
+		s := p.Modules[m]
+		switch {
+		case s == nil || s.State == "":
+			unasked++
+		case s.State == "asked":
+			asked++
+		case s.State == "built" && s.SentAt == nil:
+			unsent++
+		}
+	}
+	return asked, unsent, unasked
+}
+
+// holdingABuild is the modules of a stopped walk's current tier that hold a build of its commit, not yet sent: built and
+// not sent, or asked under a build id. Only these could put an older build over a newer plan's (review of issue 1886): a
+// module not yet asked is asked at its followed branch's head when the walk reaches it, and one asked on the builder's key
+// is settled by the look from the key of the commit it reads then. Sorted.
+func holdingABuild(p inventory.Plan) []string {
+	if p.Tier >= len(p.Tiers) {
+		return nil
+	}
+	var out []string
+	for _, m := range p.Tiers[p.Tier] {
+		s := p.Modules[m]
+		if s == nil || s.SentAt != nil || s.Key {
+			continue
+		}
+		if s.State == "built" || (s.State == "asked" && s.Build != "") {
+			out = append(out, m)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// resumeStopped sets a walk stopped by hand building again at the tier it stopped at, its stop cleared and its asks kept but
+// build-agent's own (issue 1886): the next look carries it on as any walk — a build done waits for its reference, and one done on the
+// builder's key is handed on (novox/hq ADR 0522 §3). Its note keeps who stopped it and that a person resumed it.
+func resumeStopped(p *inventory.Plan) string {
+	// **Build-agent's own are asked of build-agent** (novox/hq ADR 0522 §6): one marked taken from a key it writes itself is
+	// set unasked, and the look asks the rest of a partly asked tier (unaskedOf), of build-agent, whatever the builder says.
+	var again []string
+	for _, m := range p.Tiers[p.Tier] {
+		if s := p.Modules[m]; s != nil && s.State == "asked" && s.Key && isBuildersOwn(m) {
+			s.State, s.Key, s.AskedAt, s.Build, s.Why = "", false, nil, "", ""
+			again = append(again, m)
+		}
+	}
+	sort.Strings(again)
+	asked, unsent, unasked := waitingIn(*p)
+	later := 0
+	for i := p.Tier + 1; i < len(p.Tiers); i++ {
+		later += len(p.Tiers[i])
+	}
+	by, why := p.Delivery.Stopped, p.Delivery.StoppedWhy
+	waits := fmt.Sprintf("%d asked, %d built and not sent, %d not yet asked in it; %d not yet asked in the tiers after",
+		asked, unsent, unasked, later)
+	if len(again) > 0 {
+		waits += "; " + strings.Join(again, ", ") + " set unasked, for the look to ask of build-agent (its own, ADR 0522 §6)"
+	}
+	p.State = inventory.PlanBuilding
+	goesOnAgain(p)
+	p.Delivery.Kept = nil
+	p.Note = fmt.Sprintf("tier %d resumed by hand after the stop by %s (%s): %s", p.Tier, by, why, waits)
+	return fmt.Sprintf("%s resumed at tier %d of %d, stopped by %s: %s; this verb asks nothing — the next look carries it "+
+		"on as any walk", p.ID, p.Tier+1, len(p.Tiers), by, waits)
+}

cmd/mesh-controller/plan_retry_stopped_test.go

@@ -0,0 +1,155 @@
+package main
+
+import (
+	"slices"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// A walk stopped by hand while its tier waited — modules asked on the builder's keys, two built and never sent, the next
+// tier not yet asked — is resumed by `plans retry` (issue 1886): set building again, its stop cleared, the asks it holds
+// kept, so the next look carries it on as any walk. Nothing in it failed, so the retry of failed builds had nothing to do
+// and refused it, leaving a stopped R4e with no verb to carry its 48 modules on.
+func TestRetryResumesAWalkStoppedByHand(t *testing.T) {
+	at := time.Date(2026, 10, 23, 0, 1, 53, 0, time.UTC)
+	sent := at.Add(-time.Hour)
+	p := inventory.Plan{ID: "plan-r4e", Repository: "novox/mesh-sdk", Branch: "main", Commit: "7aacbc16",
+		Created: at.Add(-2 * time.Hour), State: inventory.PlanFailed, Tier: 1,
+		Note: "stopped by a person at tier 1: issue 1886",
+		Delivery: &inventory.PlanDelivery{Stopped: "a person (fix-1886@shanks)", StoppedWhy: "issue 1886",
+			Kept: []inventory.PlanKept{{Module: "mesh-controller"}}},
+		Modules: map[string]*inventory.PlanModule{}}
+	tier0 := []string{"mesh-controller", "mesh-sdk", "pacman", "route-proxy"}
+	for _, m := range tier0 {
+		p.Modules[m] = &inventory.PlanModule{State: "built", BuiltAt: &sent, SentAt: &sent}
+	}
+	var tier1 []string
+	for _, m := range []string{"build-agent", "mesh-tools"} {
+		tier1 = append(tier1, m)
+		p.Modules[m] = &inventory.PlanModule{State: "built", AskedAt: &sent, BuiltAt: &sent, Build: "build-" + m}
+	}
+	onKeys := []string{"mesh-build-go", "mesh-build-image", "mesh-build-node", "mesh-delivery-checker",
+		"mesh-delivery-forge-bridge", "mesh-delivery-gate", "mesh-delivery-indexer", "mesh-delivery-view"}
+	for _, m := range onKeys {
+		tier1 = append(tier1, m)
+		p.Modules[m] = &inventory.PlanModule{State: "asked", AskedAt: &sent, Key: true, Build: "build-" + m}
+	}
+	tier2 := []string{"gitea", "mesh-catalog", "mongodb"}
+	for _, m := range tier2 {
+		p.Modules[m] = &inventory.PlanModule{}
+	}
+	p.Tiers = [][]string{tier0, tier1, tier2}
+
+	if err := retryRefusal(p, []inventory.Plan{p}, ownPathOf(nil)); err != nil {
+		t.Fatalf("a walk stopped by hand with its tier waiting was refused a retry: %v", err)
+	}
+	said := resumeStopped(&p)
+	if p.State != inventory.PlanBuilding {
+		t.Errorf("resumed walk is %s, wanted building", p.State)
+	}
+	if p.Delivery.Stopped != "" || p.Delivery.StoppedWhy != "" || len(p.Delivery.Kept) > 0 {
+		t.Errorf("resumed walk still says it stopped: %+v", p.Delivery)
+	}
+	if !strings.Contains(p.Note, "resumed by hand") || !strings.Contains(p.Note, "a person (fix-1886@shanks)") {
+		t.Errorf("the note does not record the hand resume and the stop it undid: %q", p.Note)
+	}
+	for _, m := range onKeys {
+		s := p.Modules[m]
+		if isBuildersOwn(m) {
+			// Build-agent's own are asked of build-agent whatever the builder says (ADR 0522 §6): set unasked, for the look.
+			if s.State != "" || s.Key || s.Build != "" || s.AskedAt != nil {
+				t.Errorf("%s, build-agent's own, was left on the builder's key: %+v", m, s)
+			}
+			if !slices.Contains(unaskedOf(p), m) {
+				t.Errorf("%s is not among what the look asks next: %v", m, unaskedOf(p))
+			}
+			continue
+		}
+		if s.State != "asked" || !s.Key || s.Build != "build-"+m {
+			t.Errorf("%s's ask was not kept for the look to hand on: %+v", m, s)
+		}
+	}
+	for _, w := range []string{"plan-r4e", "tier 2 of 3", "5 asked", "2 built and not sent", "3 not yet asked in it",
+		"3 not yet asked in the tiers after", "mesh-build-go, mesh-build-image, mesh-build-node set unasked"} {
+		if !strings.Contains(said, w) {
+			t.Errorf("the answer %q does not say %q", said, w)
+		}
+	}
+
+	// What it still refuses: a walk stopped with nothing left in its tier, and one a related open walk runs beside.
+	done := p
+	done.State, done.Delivery = inventory.PlanFailed, &inventory.PlanDelivery{Stopped: "a person"}
+	done.Modules = map[string]*inventory.PlanModule{}
+	for _, tier := range done.Tiers {
+		for _, m := range tier {
+			done.Modules[m] = &inventory.PlanModule{State: "built", BuiltAt: &sent, SentAt: &sent}
+		}
+	}
+	if err := retryRefusal(done, nil, ownPathOf(nil)); err == nil || !strings.Contains(err.Error(), "nothing in tier 1") {
+		t.Errorf("a stopped walk with nothing left in its tier was retried: %v", err)
+	}
+	stopped := p
+	stopped.State, stopped.Delivery = inventory.PlanFailed, &inventory.PlanDelivery{Stopped: "a person"}
+	beside := inventory.Plan{ID: "plan-442", Repository: "novox/mesh-sdk", Branch: "release", Commit: "bbadbb2a",
+		Created: at, State: inventory.PlanBuilding, Tiers: [][]string{{"mesh-sdk"}},
+		Modules: map[string]*inventory.PlanModule{"mesh-sdk": {State: "asked"}}}
+	if err := retryRefusal(stopped, []inventory.Plan{beside}, ownPathOf(nil)); err == nil || !strings.Contains(err.Error(), "one walk at a time") {
+		t.Errorf("a stopped walk was resumed beside a related open walk: %v", err)
+	}
+	// A newer open plan of its repository and branch supersedes it: resuming would build what that one replaced.
+	newerOpen := inventory.Plan{ID: "plan-newer", Repository: "novox/mesh-sdk", Branch: "main", Commit: "d00dfeed",
+		Created: at, State: inventory.PlanBuilding, Tiers: [][]string{{"zsh"}},
+		Modules: map[string]*inventory.PlanModule{"zsh": {State: "asked"}}}
+	if err := retryRefusal(stopped, []inventory.Plan{stopped, newerOpen}, ownPathOf(nil)); err == nil ||
+		!strings.Contains(err.Error(), "plan-newer supersedes it") {
+		t.Errorf("a stopped walk was resumed under a newer open plan of its branch: %v", err)
+	}
+	// A newer plan holding a build of its tier not yet sent refuses it: resuming would send the older build over the newer.
+	newerBuilt := inventory.Plan{ID: "plan-agent", Repository: "novox/mesh-controller", Branch: "main", Commit: "d00dfeed",
+		Created: at, State: inventory.PlanDone, Tiers: [][]string{{"build-agent"}},
+		Modules: map[string]*inventory.PlanModule{"build-agent": {State: "built"}}}
+	if err := retryRefusal(stopped, []inventory.Plan{stopped, newerBuilt}, ownPathOf(nil)); err == nil ||
+		!strings.Contains(err.Error(), "older build") {
+		t.Errorf("a stopped walk was resumed over a newer plan of a build it has not sent: %v", err)
+	}
+}
+
+// A newer plan holding a module the stopped walk has not yet asked does not refuse its resume (review of issue 1886): that
+// module is asked at its followed branch's head when the walk reaches it, so no older build goes over the newer. Open and
+// related, the newer walk refuses it under one walk at a time; done, it does not. One holding a build the walk has not sent
+// still refuses it.
+func TestResumeIsNotRefusedByANewerPlanOfAModuleNotYetAsked(t *testing.T) {
+	at := time.Date(2026, 10, 23, 0, 1, 53, 0, time.UTC)
+	sent := at.Add(-time.Hour)
+	p := inventory.Plan{ID: "plan-r4e", Repository: "novox/mesh-sdk", Branch: "main", Commit: "7aacbc16",
+		Created: at.Add(-2 * time.Hour), State: inventory.PlanFailed, Tier: 1,
+		Delivery: &inventory.PlanDelivery{Stopped: "a person", StoppedWhy: "issue 1886"},
+		Tiers:    [][]string{{"mesh-sdk"}, {"build-agent", "mesh-delivery-gate"}, {"mesh-delivery-builder", "gitea"}},
+		Modules: map[string]*inventory.PlanModule{
+			"mesh-sdk":              {State: "built", BuiltAt: &sent, SentAt: &sent},
+			"build-agent":           {State: "built", BuiltAt: &sent},
+			"mesh-delivery-gate":    {State: "asked", AskedAt: &sent, Key: true},
+			"mesh-delivery-builder": {}, "gitea": {}}}
+	newer := func(state string, holds string) inventory.Plan {
+		return inventory.Plan{ID: "plan-1792715180015327197", Repository: "novox/mesh-delivery", Branch: "main",
+			Commit: "d00dfeed", Created: at, State: state, Tiers: [][]string{{holds}},
+			Modules: map[string]*inventory.PlanModule{holds: {State: "asked"}}}
+	}
+	if err := retryRefusal(p, []inventory.Plan{p, newer(inventory.PlanDone, "mesh-delivery-builder")}, ownPathOf(nil)); err != nil {
+		t.Errorf("a done newer plan of a module not yet asked refused the resume: %v", err)
+	}
+	if err := retryRefusal(p, []inventory.Plan{p, newer(inventory.PlanDone, "mesh-delivery-gate")}, ownPathOf(nil)); err != nil {
+		t.Errorf("a done newer plan of a module on the builder's key refused the resume: %v", err)
+	}
+	if err := retryRefusal(p, []inventory.Plan{p, newer(inventory.PlanBuilding, "mesh-delivery-builder")}, ownPathOf(nil)); err == nil ||
+		!strings.Contains(err.Error(), "one walk at a time") {
+		t.Errorf("an open newer walk related to it did not refuse the resume under one walk at a time: %v", err)
+	}
+	if err := retryRefusal(p, []inventory.Plan{p, newer(inventory.PlanDone, "build-agent")}, ownPathOf(nil)); err == nil ||
+		!strings.Contains(err.Error(), "older build") {
+		t.Errorf("a newer plan of a built module not yet sent did not refuse the resume: %v", err)
+	}
+}

cmd/mesh-controller/release_plan.go

@@ -457,6 +457,18 @@ func onlyTheBuildersOwn(ctx context.Context, source buildSource, path string) er
 	if isBuildersOwn(module) {
 		return nil
 	}
+	// **While no running builder says it writes references, build-agent builds every module** (issue 1886): judged by the
+	// say the walk read when it has one, read here only for an ask that carries none.
+	says, read := builderSayOf(ctx)
+	if !read {
+		var err error
+		if says, err = builderSaysReferences(ctx); err != nil {
+			says = true // not known: refused below, as before the say existed
+		}
+	}
+	if !says {
+		return nil
+	}
 	if module == "" {
 		module = fmt.Sprintf("%s at %q", source, orRoot(path))
 	}
@@ -529,10 +541,22 @@ func askModule(ctx context.Context, p *inventory.Plan, name string, byName map[s
 	}
 	// **Every module but the builder's own is taken from the builder's done key, asking no build** (novox/hq ADR 0503
 	// §6, ADR 0501, ADR 0453 §5): the look settles it (settleFromKeys).
+	// **Only while a running builder says it writes references** (issue 1886): until then the keys it writes could never
+	// be delivered from, and the module is asked of build-agent as before R4e. A say not read leaves it unasked, said,
+	// for the next look.
 	if !isBuildersOwn(name) {
-		sayBuilding(ctx, p.ID, now)
-		takeFromKey(p, state, e, now)
-		return
+		says, err := builderSaysReferences(ctx)
+		if err != nil {
+			state.Why = "waits: whether the builder writes references cannot be read: " + firstLine(err.Error())
+			return
+		}
+		if says {
+			sayBuilding(ctx, p.ID, now)
+			takeFromKey(p, state, e, now)
+			return
+		}
+		// The ask's guard judges by this say, never by a second read that could fail or differ (review of issue 1886).
+		ctx = withBuilderSay(ctx, says)
 	}
 	source := buildSource{Repository: e.Source.Repository, Seat: e.Source.Seat}
 	fmt.Printf("  tier %d: ", p.Tier)

internal/builderkeys/builderkeys.go

@@ -9,6 +9,7 @@ package builderkeys
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
 	"strings"
@@ -309,3 +310,99 @@ func firstLine(s string) string {
 	line, _, _ := strings.Cut(strings.TrimSpace(s), "\n")
 	return line
 }
+
+// The builder's say of what its keys hold (novox/hq issue 1886): each running builder instance keeps a key, named by its
+// node, in the builder's own announced bucket of instances, saying what the keys it writes hold. An instance that writes
+// references (ADR 0498 §1) says WritesReferences there; a builder from before R4e keeps no such key, and has no bucket.
+// The bucket is the builder's and the shape is read by the controller alone, so neither is in mesh-sdk.
+const (
+	// BuilderInstancesBucket is the builder's bucket of its running instances, `<owner module>_<name>`.
+	BuilderInstancesBucket = delivery.BuilderModule + "_instances"
+	// WritesReferences is the word an instance says when every done key it writes names where its artifacts are.
+	WritesReferences = "references"
+)
+
+// instanceSays is an instance's value in BuilderInstancesBucket: what the keys it writes hold.
+type instanceSays struct {
+	Writes []string `json:"writes"`
+}
+
+// saysReferences says whether any instance's value says it writes references. A value that does not decode says
+// nothing.
+func saysReferences(values [][]byte) bool {
+	for _, v := range values {
+		var says instanceSays
+		if json.Unmarshal(v, &says) != nil {
+			continue
+		}
+		for _, w := range says.Writes {
+			if w == WritesReferences {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// SaysReferences says whether a running builder instance says, in its bucket of instances, that it writes references. No
+// bucket (the builder from before R4e, which never made it) and no key say no; a bucket that cannot be read is an error.
+func (r Reader) SaysReferences(ctx context.Context) (bool, error) {
+	kv, err := r.bucket(ctx, BuilderInstancesBucket)
+	if err != nil || kv == nil {
+		return false, err
+	}
+	keys, err := kv.Keys(ctx)
+	if errors.Is(err, jetstream.ErrNoKeysFound) {
+		return false, nil
+	}
+	if err != nil {
+		return false, fmt.Errorf("listing the builder's instances: %w", err)
+	}
+	var values [][]byte
+	for _, key := range keys {
+		e, err := kv.Get(ctx, key)
+		if errors.Is(err, jetstream.ErrKeyNotFound) {
+			continue
+		}
+		if err != nil {
+			return false, fmt.Errorf("the builder's instance %s: %w", key, err)
+		}
+		values = append(values, e.Value())
+	}
+	return saysReferences(values), nil
+}
+
+// WatchSays calls wake for every change of the builder's bucket of instances from now on, until ctx ends: what makes the
+// look judge an open walk again when the builder starts or stops saying it writes references (issue 1886), with no
+// timer. It says false, and watches nothing, while the bucket is not there; the caller asks again when the builder
+// writes a key, which a builder that makes the bucket does.
+func WatchSays(ctx context.Context, js jetstream.JetStream, wake func()) (bool, error) {
+	kv, err := js.KeyValue(ctx, BuilderInstancesBucket)
+	if errors.Is(err, jetstream.ErrBucketNotFound) {
+		return false, nil
+	}
+	if err != nil {
+		return false, fmt.Errorf("the bucket %s: %w", BuilderInstancesBucket, err)
+	}
+	w, err := kv.WatchAll(ctx, jetstream.UpdatesOnly())
+	if err != nil {
+		return false, fmt.Errorf("watching %s: %w", BuilderInstancesBucket, err)
+	}
+	go func() {
+		defer w.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case e, ok := <-w.Updates():
+				if !ok {
+					return
+				}
+				if e != nil {
+					wake() // a put and a delete both change what the builder says
+				}
+			}
+		}
+	}()
+	return true, nil
+}

internal/builderkeys/says_integration_test.go

@@ -0,0 +1,54 @@
+//go:build integration
+
+package builderkeys
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go/jetstream"
+)
+
+// 1886/AC1 and AC4 on a real bus: no bucket of instances says no and watches nothing; an instance's key saying
+// references says yes, and its put and its delete each wake the watcher, with no timer.
+func TestTheBuildersSayIsReadAndWatchedOnTheBus(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+	js := aBus(t)
+	r := Reader{JS: js}
+	if says, err := r.SaysReferences(ctx); err != nil || says {
+		t.Fatalf("no bucket: %v %v", says, err)
+	}
+	if on, err := WatchSays(ctx, js, func() {}); err != nil || on {
+		t.Fatalf("no bucket is watched: %v %v", on, err)
+	}
+	kv, err := js.CreateKeyValue(ctx, jetstream.KeyValueConfig{Bucket: BuilderInstancesBucket, History: 1})
+	if err != nil {
+		t.Fatal(err)
+	}
+	woken := make(chan struct{}, 4)
+	if on, err := WatchSays(ctx, js, func() { woken <- struct{}{} }); err != nil || !on {
+		t.Fatalf("the bucket is not watched: %v %v", on, err)
+	}
+	wake := func(what string) {
+		t.Helper()
+		select {
+		case <-woken:
+		case <-ctx.Done():
+			t.Fatalf("%s woke nothing", what)
+		}
+	}
+	put(t, kv, "shanks", []byte(`{"writes":["references"]}`))
+	wake("a put")
+	if says, err := r.SaysReferences(ctx); err != nil || !says {
+		t.Fatalf("an instance saying references: %v %v", says, err)
+	}
+	if err := kv.Delete(ctx, "shanks"); err != nil {
+		t.Fatal(err)
+	}
+	wake("a delete")
+	if says, err := r.SaysReferences(ctx); err != nil || says {
+		t.Fatalf("no instance left: %v %v", says, err)
+	}
+}

internal/builderkeys/says_test.go

@@ -0,0 +1,24 @@
+package builderkeys
+
+import "testing"
+
+// 1886/AC1: a builder that keeps no say, or a say without the word, is a builder that does not write references; one
+// instance that says it is enough, since it builds every key done without a reference again (ADR 0498 §8).
+func TestTheBuilderSaysReferencesOnlyWhenAnInstanceSaysTheWord(t *testing.T) {
+	cases := []struct {
+		name   string
+		values [][]byte
+		want   bool
+	}{
+		{"no instance", nil, false},
+		{"an instance saying nothing", [][]byte{[]byte(`{"writes":[]}`)}, false},
+		{"a value not read", [][]byte{[]byte(`{not json`)}, false},
+		{"another word", [][]byte{[]byte(`{"writes":["digests"]}`)}, false},
+		{"one instance says it", [][]byte{[]byte(`{"writes":["digests"]}`), []byte(`{"v":1,"writes":["references"]}`)}, true},
+	}
+	for _, c := range cases {
+		if got := saysReferences(c.values); got != c.want {
+			t.Errorf("%s: says references %v, want %v", c.name, got, c.want)
+		}
+	}
+}