Merge pull request 'No merge is dropped, and the merge window lives on the bus (issues 973 step 2, 921)' (#367) from feat/merge-window-on-the-bus into main

Main refuses such a jail (fail2ban would refuse its whole configuration and
keep no bans), but no test held it: the only one sat untracked in an old
worktree. Issue 959.

cmd/mesh-controller/batches.go

@@ -27,9 +27,10 @@ import (
 // commit in it, with at most one commit per repository, the latest of its branch, which contains the earlier
 // ones. The walk names every merge it answers.
 //
-//   - One walk at a time. Merges that come while a walk runs join the next batch. A walk that started is never
+//   - One walk at a time beside a related one (ADR 0407 decision 4, issue 973). Merges that come while a walk
-//     superseded; one still waiting for its delivery's word is folded into the next batch's walk, which
+//     runs join the next batch; a closed batch that shares nothing with an open walk is cut beside it. A walk
-//     answers every merge it answered — the only supersession left.
+//     that started is never superseded; one still waiting for its delivery's word is folded into the walk of a
+//     batch related to it, which answers every merge it answered — the only supersession left.
 //   - A merge heard after a later merge of its branch was put in a walk is answered by that walk: a batch
 //     still assembling, an open walk that builds everything it moves, or a walk done that did, which then
 //     names it at once. Otherwise it joins the next batch.
@@ -39,8 +40,9 @@ import (
 //
 // **A merge waits only for a batch it shares something with** (novox/hq ADR 0304 decision 1, ADR 0306 decisions
 // 2 and 3, issue 525): it joins a batch it is related to, and one that is related to none and whose delivery
-// group has no member still to merge is a batch whose window is closed at once (batchToJoin). Walks stay one at
+// group has no member still to merge is a batch whose window is closed at once (batchToJoin). Such a batch is cut
-// a time: such a batch is cut when no walk is open, and queued behind the one that is otherwise.
+// at once, beside any open walk it shares nothing with, and queued behind the first one it is related to
+// (cutUnrelated).
 //
 // A merge on the controller's own path (a module whose walk waits for nobody's word) never shares a batch with
 // one that waits for mesh-delivery's: each kind has a batch of its own, an own-path batch is cut first and folds
@@ -61,13 +63,14 @@ var (
 // mergeWindowFileVar names the file the controller's window settings are written to.
 const mergeWindowFileVar = "MESH_MERGE_WINDOW_FILE"
 
-// batchCutEvery is how often the controller looks whether a batch's window closed.
-const batchCutEvery = 5 * time.Second
-
 // batchLateAfter is how long past its maximum a batch may still be assembling, with no walk open, before
 // the controller is said to have failed to cut it (ADR 0276 decision 9).
 const batchLateAfter = time.Minute
 
+// mergeGrace is how long before a batch's first merge a merge may have been made and still be told apart from one
+// made before the controller began keeping merges (owedLate).
+const mergeGrace = 10 * time.Minute
+
 // edgeGroupOrder is a delivery group's order made an edge of a walk's tiers (ADR 0276 decision 5): the
 // modules of the member after depend on the modules of the member before, and are gated on them.
 const edgeGroupOrder = "group-order"
@@ -125,8 +128,7 @@ func durationSetting(v any) (time.Duration, bool) {
 // hearMerge puts a merge the forge announced into the batch that answers it, at the moment it was heard,
 // and cuts what is due.
 func (f following) hearMerge(ctx context.Context, m link.SourceMoved, now time.Time) error {
-	// One merge acted on at a time, whoever hands it over: the bus, or the catch-up that reads back
+	// One merge acted on at a time, whoever hands it over (novox/hq issue 266).
-	// what the bus did not hand over (novox/hq issue 266).
 	actingOnMerges.Lock()
 	defer actingOnMerges.Unlock()
 
@@ -245,8 +247,8 @@ func (f following) hearMerge(ctx context.Context, m link.SourceMoved, now time.T
 }
 
 // owedLate says a merge that reads as history is still the record's to answer (review of this change): a cut
-// marks every module it moved seen at that moment, so a merge of that branch the bus lost and the catch-up
+// marks every module it moved seen at that moment, so a merge of that branch
-// hands over later — made before the cut, heard after it — reads as history and was dropped. One that touches
+// handed over late — made before the cut, heard after it — reads as history and was dropped. One that touches
 // a module the mesh holds, of a branch whose merges the controller keeps, made since it began keeping them,
 // is kept; one made before it began is old news, as before.
 func owedLate(ctx context.Context, inv *inventory.Inventory, m link.SourceMoved, merged time.Time,
@@ -340,8 +342,8 @@ func ownPath(touches []inventory.Entry) bool {
 //     names why): a new batch, which waits for the window, beside the ones it shares nothing with; its words say
 //     why.
 //   - none, and no member of its group is still to merge: a new batch whose window is closed at once. Cut at
-//     once means only that: walks stay one at a time (ADR 0276 decision 4), so with a walk open the batch is
+//     once means only that: with a walk open that it is related to the batch is queued behind it as any closed
-//     queued behind it as any closed batch is.
+//     batch is (ADR 0276 decision 4, ADR 0407 decision 4); beside one it shares nothing with it is cut.
 func batchToJoin(ctx context.Context, inv *inventory.Inventory, heard inventory.BatchedMerge, entries []inventory.Entry,
 	read map[string][]inventory.ReadRepository, now time.Time, own bool, waits string) (inventory.Plan, string, error) {
 	edges, err := inv.Dependencies(ctx)
@@ -406,7 +408,9 @@ func relatedBatches(ctx context.Context, inv *inventory.Inventory, merges []inve
 	}
 	carry := newFootprint()
 	if !own {
-		// An own-path batch folds no waiting walk (cutBatchesHeld); a catalogue batch may fold every one.
+		// An own-path batch folds no waiting walk (cutBatchesHeld); a catalogue batch folds those it is related to
+		// (cutUnrelated). Every one is counted here, which only makes a merge join a batch more readily: it waits
+		// for the window, never walks beside what it shares something with (issue 973).
 		plans, err := inv.OpenPlans(ctx)
 		if err != nil {
 			return nil, nil, nil, err
@@ -544,6 +548,11 @@ func newBatch(ctx context.Context, inv *inventory.Inventory, first inventory.Bat
 // keepBatch writes a batch as its merges and the clock say it now: the commits it holds, the merges they
 // answer, its window, and assembling or queued behind a walk. Saved only when that changed.
 func keepBatch(ctx context.Context, inv *inventory.Inventory, b *inventory.Plan, now time.Time, behind string) error {
+	return keepBatchBecause(ctx, inv, b, now, behind, "")
+}
+
+// keepBatchBecause is keepBatch for a batch queued behind a walk it is related to, which its words say (issue 973).
+func keepBatchBecause(ctx context.Context, inv *inventory.Inventory, b *inventory.Plan, now time.Time, behind, why string) error {
 	merges, err := inv.MergesOf(ctx, b.ID)
 	if err != nil {
 		return err
@@ -574,19 +583,25 @@ func keepBatch(ctx context.Context, inv *inventory.Inventory, b *inventory.Plan,
 		cutAtOnce = was.CutAtOnce                   // the merge's, until another joins it (ADR 0304 decision 1)
 		waits = was.Waits                           // its group's, as said when it was made (ADR 0306 decision 2)
 	}
-	b.Delivery.Batch = &inventory.PlanBatch{ClosesAt: closes, AtMost: latest, Behind: behind, Own: b.OwnPath(),
+	b.Delivery.Batch = &inventory.PlanBatch{ClosesAt: closes, AtMost: latest, Behind: behind, BehindWhy: why, Own: b.OwnPath(),
 		Closed: closed, Withheld: withheld, CutAtOnce: cutAtOnce, Waits: waits}
 	b.State = inventory.PlanAssembling
 	if behind != "" && windowClosed(*b, now) {
 		b.State = inventory.PlanQueued
 	} else {
-		b.Delivery.Batch.Behind = ""
+		b.Delivery.Batch.Behind, b.Delivery.Batch.BehindWhy = "", ""
 	}
 	b.Note = batchWords(*b, now)
 	if unNoted(*b) == before {
 		return nil
 	}
-	return inv.SavePlan(ctx, b)
+	if err := inv.SavePlan(ctx, b); err != nil {
+		return err
+	}
+	// The window's clock follows the record, only when the record moved: a merge made the quiet key again, a
+	// batch that opened made its most (ADR 0407 decision 1).
+	armWindow(ctx, *b, now)
+	return nil
 }
 
 // windowOf is when a batch's window closes, unless another merge comes, and when at the latest: the last
@@ -739,31 +754,10 @@ func nameMerges(ctx context.Context, inv *inventory.Inventory, p *inventory.Plan
 	return inv.SavePlan(ctx, p)
 }
 
-// batchCutter cuts what is due on a timer, until the context ends: a window closes with no merge to say so.
-func batchCutter(ctx context.Context, open *stores) {
-	tick := time.NewTicker(batchCutEvery)
-	defer tick.Stop()
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case <-tick.C:
-		}
-		release, err := open.inventory.HoldPlans(ctx, false)
-		if err != nil {
-			continue // whoever holds the plans moves them, and cuts at the end
-		}
-		if err := cutBatchesHeld(ctx, open, time.Now().UTC()); err != nil {
-			fmt.Printf("batches: %v\n", err)
-		}
-		release()
-	}
-}
-
 // cutBatchesHeld takes the step that is due, for a caller holding the plans: after a failed walk, the search
-// for the merge that brought the failure; with no walk open, the next merge of that search walked alone, or a
+// for the merge that brought the failure; with no walk open, the next merge of that search walked alone; else
-// closed batch cut into one walk with the waiting walks folded in; with a walk open, a closed batch said
+// each closed batch cut into a walk with the waiting walks it is related to folded in, unless it is related to a
-// queued behind it.
+// walk open, behind which it is said queued (cutUnrelated). While the search runs everything waits.
 func cutBatchesHeld(ctx context.Context, open *stores, now time.Time) error {
 	inv := open.inventory
 	if err := searchAfterFailures(ctx, inv); err != nil {
@@ -780,8 +774,7 @@ func cutBatchesHeld(ctx context.Context, open *stores, now time.Time) error {
 	if err != nil {
 		return err
 	}
-	var started *inventory.Plan
+	var started, waiting []inventory.Plan
-	var waiting []inventory.Plan
 	for i := range plans {
 		p := plans[i]
 		if p.Release != nil {
@@ -789,8 +782,8 @@ func cutBatchesHeld(ctx context.Context, open *stores, now time.Time) error {
 		}
 		if p.Waiting() {
 			waiting = append(waiting, p)
-		} else if started == nil {
+		} else {
-			started = &plans[i]
+			started = append(started, p)
 		}
 	}
 	batches, err := inv.Batches(ctx)
@@ -805,46 +798,116 @@ func cutBatchesHeld(ctx context.Context, open *stores, now time.Time) error {
 		}
 		return nil
 	}
-	if started != nil {
-		// **One walk at a time** (ADR 0276 decision 4): every batch waits behind it, whatever its window says.
-		return keepAll(started.ID)
-	}
 	if len(alone) > 0 {
+		// A failed walk's search walks its merges alone, one at a time: everything waits for it (kept as it was,
+		// issue 973).
+		if len(started) > 0 {
+			return keepAll(started[0].ID)
+		}
 		if len(waiting) > 0 {
 			// The waiting walk is the one open; the search walks after it, and the batches wait behind both.
 			return keepAll(waiting[0].ID)
 		}
 		return walkAlone(ctx, open, alone[0], now)
 	}
-	if err := keepAll(""); err != nil {
+	return cutUnrelated(ctx, open, batches, started, waiting, now)
+}
+
+// cutUnrelated cuts the closed batches that share nothing with an open walk, each into a walk beside the open
+// ones, and says the others queued behind the first walk they are related to (novox/hq ADR 0407 decision 4,
+// issue 973). A batch cut in this look counts as open for the batches after it.
+//
+//   - An own-path batch goes first and folds no waiting walk; a catalogue batch folds the waiting walks it is
+//     related to, and those it shares nothing with keep waiting.
+//   - **The controller's own path is related to everything** (decided for issue 973): a walk or batch that
+//     replaces the controller, the node-engine, the tool runner, the bus or mesh-delivery never runs beside
+//     another, and nothing runs beside it.
+//   - A batch whose window is still open is assembling; one related to an open walk is queued behind it, and
+//     ADR 0276 decision 4 stands for it: nothing supersedes a walk that started.
+func cutUnrelated(ctx context.Context, open *stores, batches, started, waiting []inventory.Plan, now time.Time) error {
+	inv := open.inventory
+	sort.SliceStable(batches, func(i, j int) bool { return batches[i].OwnPath() && !batches[j].OwnPath() })
+	var entries []inventory.Entry
+	var read map[string][]inventory.ReadRepository
+	var edges []inventory.Edge
+	judged := false
+	openWalks, err := openWalksOf(ctx, inv, started)
+	if err != nil {
 		return err
 	}
-	// The oldest batch whose window closed is cut, an own-path one before a catalogue one (it goes first, and
-	// the catalogue batch queues behind it rather than being cut and overtaken). Several of a kind may be open
-	// since a merge that shares nothing is cut at once (ADR 0304 decision 1): walks stay one at a time, so one
-	// is cut per look, and once its walk started the others are queued behind it.
-	sort.SliceStable(batches, func(i, j int) bool { return batches[i].OwnPath() && !batches[j].OwnPath() })
 	for i := range batches {
 		b := &batches[i]
 		if !windowClosed(*b, now) {
-			continue
+			if err := keepBatch(ctx, inv, b, now, ""); err != nil {
-		}
-		if b.OwnPath() {
-			// A walk on the controller's own path folds no walk waiting for its word (decided during the build,
-			// 2026-10-10): that walk keeps waiting beside it, and starts once this one ended (advanceOnce).
-			waiting = nil
-		}
-		if err := cutBatch(ctx, open, b, waiting, now); err != nil {
-			return err
-		}
-		// The batches left wait behind the walk just cut when it started; the next look does the same.
-		if b.Open() && !b.Waiting() {
-			if batches, err = inv.Batches(ctx); err != nil {
 				return err
 			}
-			return keepAll(b.ID)
+			continue
+		}
+		if !judged {
+			var err error
+			if entries, err = inv.Catalogued(ctx); err != nil {
+				return err
+			}
+			if read, err = readForPlanning(ctx, inv); err != nil {
+				return err
+			}
+			if edges, err = inv.Dependencies(ctx); err != nil {
+				return err
+			}
+			judged = true
+		}
+		held, err := inv.MergesOf(ctx, b.ID)
+		if err != nil {
+			return err
+		}
+		fp := footprintAsCut(held, entries, read, edges)
+		var fold, keep []inventory.Plan
+		if !b.OwnPath() {
+			for _, w := range waiting {
+				ww, err := openWalkOf(ctx, inv, w)
+				if err != nil {
+					return err
+				}
+				if ok, _ := related(fp, ww.fp, edges); ok {
+					fold = append(fold, w)
+					fp = unionFootprint(fp, ww.fp)
+				} else {
+					keep = append(keep, w)
+				}
+			}
+		}
+		if behind, why := relatedOpen(fp, b.OwnPath(), openWalks, edges); behind != "" {
+			if err := keepBatchBecause(ctx, inv, b, now, behind, why); err != nil {
+				return err
+			}
+			continue
+		}
+		if err := keepBatch(ctx, inv, b, now, ""); err != nil {
+			return err
+		}
+		if err := cutBatch(ctx, open, b, fold, now); err != nil {
+			return err
+		}
+		waiting = keep
+		if c := windowClockOf(); c != nil {
+			c.Forget(ctx, b.ID) // a batch cut has no window left to close
+		}
+		cut, err := inv.PlanByID(ctx, b.ID)
+		if err != nil {
+			return err
+		}
+		if !cut.Open() {
+			continue
+		}
+		w, err := openWalkOf(ctx, inv, cut)
+		if err != nil {
+			return err
+		}
+		if cut.Waiting() {
+			waiting = append(waiting, cut)
+		} else {
+			openWalks = append(openWalks, w)
 		}
-		return nil
 	}
 	return nil
 }
@@ -1396,7 +1459,11 @@ func batchWords(b inventory.Plan, now time.Time) string {
 		grouped += " (on the controller's own path)"
 	}
 	if b.State == inventory.PlanQueued {
-		return fmt.Sprintf("queued behind %s; grouped: %s; plan not yet calculated", w.Behind, grouped)
+		because := ""
+		if w.BehindWhy != "" {
+			because = " (" + w.BehindWhy + ")"
+		}
+		return fmt.Sprintf("queued behind %s%s; grouped: %s; plan not yet calculated", w.Behind, because, grouped)
 	}
 	if w.CutAtOnce {
 		return fmt.Sprintf("cut at once, sharing nothing with a batch assembling; grouped: %s; plan not yet calculated",

cmd/mesh-controller/batches_test.go

@@ -334,8 +334,8 @@ func captured(t *testing.T, run func() error) string {
 	return b.String()
 }
 
-// One walk at a time: a merge heard while a walk runs joins the next batch, which is cut when the walk ends;
+// A merge heard during a walk it is related to (the same repository, issue 973) joins the next batch, which is
-// the walk that started is never superseded.
+// cut when the walk ends; the walk that started is never superseded (ADR 0276 decision 4, ADR 0407 decision 4).
 func TestAMergeDuringAWalkJoinsTheNextBatch(t *testing.T) {
 	open := windowed(t)
 	asksWithPaths(t)
@@ -346,7 +346,7 @@ func TestAMergeDuringAWalkJoinsTheNextBatch(t *testing.T) {
 		t.Fatalf("the first batch was not cut: %+v", ws)
 	}
 	first := ws[0]
-	hear(t, open, repoMerge("two", "c2", t0.Add(2*time.Minute)), t0.Add(2*time.Minute))
+	hear(t, open, repoMerge("one", "c2", t0.Add(2*time.Minute)), t0.Add(2*time.Minute))
 	cutAt(t, open, t0.Add(5*time.Minute))
 	ws, bs := walks(t, open)
 	if len(ws) != 1 || len(bs) != 1 || bs[0].State != inventory.PlanQueued || bs[0].Delivery.Batch.Behind != first.ID {
@@ -391,7 +391,7 @@ func TestAWaitingWalkIsFoldedIntoTheNextBatch(t *testing.T) {
 		t.Fatalf("the walk does not wait for its word: %+v", ws)
 	}
 	waiting := ws[0]
-	hear(t, open, repoMerge("two", "c2", t0.Add(2*time.Minute)), t0.Add(2*time.Minute))
+	hear(t, open, repoMerge("one", "c2", t0.Add(2*time.Minute)), t0.Add(2*time.Minute))
 	cutAt(t, open, t0.Add(2*time.Minute+90*time.Second))
 	ws, bs := walks(t, open)
 	if len(bs) != 0 || len(ws) != 2 {
@@ -405,7 +405,7 @@ func TestAWaitingWalkIsFoldedIntoTheNextBatch(t *testing.T) {
 	if folded.State != inventory.PlanSuperseded || folded.Delivery.TakenOverBy != newer.ID {
 		t.Fatalf("the waiting walk is %s, taken over by %q", folded.State, folded.Delivery.TakenOverBy)
 	}
-	if len(newer.Delivery.Merges) != 2 || newer.CommitOf("novox/one") != "c1" || newer.CommitOf("novox/two") != "c2" {
+	if len(newer.Delivery.Merges) != 2 || newer.CommitOf("novox/one") != "c2" {
 		t.Fatalf("the newer walk answers %+v", newer.Delivery.Merges)
 	}
 	if _, in := newer.Modules["one"]; !in {
@@ -662,8 +662,8 @@ func TestALateMergeOfAOneModuleRepositoryIsNamed(t *testing.T) {
 	}
 }
 
-// One walk at a time holds against the delivery's word too: a waiting walk let go beside a started one takes the
+// A related walk waits for the delivery's word too: a waiting walk let go beside a started one it is related to
-// word and starts once that one ended, asking nothing before.
+// (the same repository, issue 973) takes the word and starts once that one ended, asking nothing before.
 func TestAWalkLetGoBesideAStartedOneStartsOnceItEnded(t *testing.T) {
 	open := windowed(t)
 	asked := asksWithPaths(t)
@@ -671,8 +671,8 @@ func TestAWalkLetGoBesideAStartedOneStartsOnceItEnded(t *testing.T) {
 	hear(t, open, repoMerge("one", "c1", t0), t0)
 	cutAt(t, open, t0.Add(2*time.Minute))
 	started, _ := walks(t, open)
-	waiting := inventory.Plan{ID: "plan-waiting", Repository: "novox/two", Branch: "main", Commit: "d1", Created: t0,
+	waiting := inventory.Plan{ID: "plan-waiting", Repository: "novox/one", Branch: "main", Commit: "d1", Created: t0,
-		State: inventory.PlanBuilding, Tiers: [][]string{{"two"}}, Modules: map[string]*inventory.PlanModule{"two": {}},
+		State: inventory.PlanBuilding, Tiers: [][]string{{"one"}}, Modules: map[string]*inventory.PlanModule{"one": {}},
 		Delivery: &inventory.PlanDelivery{Awaits: catalogue.DeliverySeat}}
 	if err := open.inventory.SavePlan(ctx, &waiting); err != nil {
 		t.Fatal(err)
@@ -729,49 +729,6 @@ func TestARemovedFileIsTheLastMergesWord(t *testing.T) {
 	}
 }
 
-// The catch-up hands over what the bus lost after its branch's later merges were cut, and the record keeps it:
-// an earlier merge is named by the walk that carried it; a merge made before a cut and heard after it, which
-// no later merge carries, joins the next batch — neither reads as history and is dropped (review of this change).
-func TestTheCatchUpKeepsWhatACutMadeHistory(t *testing.T) {
-	open := windowed(t)
-	asksWithPaths(t)
-	ctx := t.Context()
-	base := time.Now().UTC().Add(-time.Hour).Truncate(time.Second)
-	hear(t, open, repoMerge("one", "c2", base.Add(10*time.Second)), base.Add(11*time.Second))
-	cutAt(t, open, base.Add(2*time.Minute))
-	ws, _ := walks(t, open)
-	w := ws[0]
-	w.State = inventory.PlanDone
-	if err := open.inventory.SavePlan(ctx, &w); err != nil {
-		t.Fatal(err)
-	}
-	lost := []link.AnnouncedMerge{
-		{SourceMoved: repoMerge("one", "c1", base), At: base.Add(time.Second)},
-		{SourceMoved: repoMerge("one", "c3lost", base.Add(20*time.Second)), At: base.Add(21 * time.Second)},
-	}
-	catalogued := func(ctx context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error) {
-		entries, err := open.inventory.Catalogued(ctx)
-		if err != nil {
-			return nil, nil, err
-		}
-		read, err := readForPlanning(ctx, open.inventory)
-		return entries, read, err
-	}
-	if err := catchUpOnMerges(ctx, time.Now(), unheardMerges{announcedList(lost), open.inventory}, catalogued,
-		(following{open: open}).SourceMoved, t.Logf); err != nil {
-		t.Fatal(err)
-	}
-	got, _ := open.inventory.PlanByID(ctx, w.ID)
-	if len(got.Delivery.Merges) != 2 || !sameMerge(got.Delivery.Merges[0], inventory.PlanMerge{Repository: "novox/one",
-		Commit: "c1", Carried: "c2"}) {
-		t.Fatalf("the earlier merge the catch-up handed over is not named by the walk that carried it: %+v",
-			got.Delivery.Merges)
-	}
-	if _, kept, err := open.inventory.MergeOf(ctx, "novox/one", "c3lost"); err != nil || !kept {
-		t.Fatalf("the merge made before the cut and heard after it was dropped: %v %v", kept, err)
-	}
-}
-
 // A merge on the controller's own path never shares a batch with one that waits for mesh-delivery's word
 // (decided during the build, 2026-10-10): the two are batches of their own, cut one after the other; a walk on
 // the own path folds no waiting catalogue walk but batches its merges again behind it; the catalogue's walk

cmd/mesh-controller/beside_test.go

@@ -0,0 +1,106 @@
+package main
+
+import (
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// An unrelated closed batch is cut into a walk beside an open one, and that walk asks its first tier at once
+// (issue 973, novox/hq ADR 0407 decision 4): a hotfix is never held by an unrelated delivery.
+func TestAnUnrelatedBatchWalksBesideAnOpenWalk(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	ctx := t.Context()
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	cutAt(t, open, t0.Add(90*time.Second))
+	if ws, _ := walks(t, open); len(ws) != 1 || !ws[0].Open() {
+		t.Fatalf("the first walk is not open: %+v", ws)
+	}
+	heard := t0.Add(100 * time.Second)
+	hear(t, open, lone(repoMerge("two", "c2", heard)), heard)
+	advanceHeld(ctx, open)
+	ws, bs := walks(t, open)
+	if len(ws) != 2 || len(bs) != 0 {
+		t.Fatalf("%d walk(s) and %d batch(es); want the unrelated batch cut beside the open walk", len(ws), len(bs))
+	}
+	for _, w := range ws {
+		if !w.Open() || w.Waiting() || deferred(w) || !askedAny(w) {
+			t.Errorf("%s (%s) is not walking: open %v, deferred %v, asked %v, note %q", w.ID, w.Named(), w.Open(),
+				deferred(w), askedAny(w), w.Note)
+		}
+	}
+}
+
+// A closed batch related to an open walk is queued behind it with the reason, while one that shares nothing is
+// cut beside it; the walk that started is never superseded (ADR 0276 decision 4, issue 973).
+func TestARelatedBatchWaitsBehindTheOpenWalk(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	ctx := t.Context()
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	cutAt(t, open, t0.Add(90*time.Second))
+	first := func() inventory.Plan { ws, _ := walks(t, open); return ws[len(ws)-1] }()
+	heard := t0.Add(100 * time.Second)
+	hear(t, open, lone(repoMerge("one", "c2", heard)), heard)
+	hear(t, open, lone(repoMerge("three", "c3", heard.Add(time.Second))), heard.Add(time.Second))
+	ws, bs := walks(t, open)
+	if len(ws) != 2 || len(bs) != 1 {
+		t.Fatalf("%d walk(s) and %d batch(es); want the unrelated merge walking and the related one queued", len(ws), len(bs))
+	}
+	b := bs[0]
+	words := batchWords(b, heard)
+	if b.State != inventory.PlanQueued || b.Delivery.Batch.Behind != first.ID || b.CommitOf("novox/one") != "c2" ||
+		!strings.Contains(words, "queued behind "+first.ID) || !strings.Contains(words, "both change novox/one") {
+		t.Fatalf("the related batch is %s, behind %q, reading %q", b.State, b.Delivery.Batch.Behind, words)
+	}
+	if got, _ := open.inventory.PlanByID(ctx, first.ID); !got.Open() {
+		t.Fatalf("the started walk was %s by the related merge", got.State)
+	}
+	first.State = inventory.PlanDone
+	if err := open.inventory.SavePlan(ctx, &first); err != nil {
+		t.Fatal(err)
+	}
+	cutAt(t, open, heard.Add(2*time.Second))
+	if _, bs := walks(t, open); len(bs) != 0 {
+		t.Fatalf("the queued batch was not cut once the walk it waited for ended: %+v", bs)
+	}
+}
+
+// A batch on the controller's own path never walks beside another walk, and nothing walks beside it, whatever
+// the two share (issue 973): it replaces the controller, so it goes alone.
+func TestAnOwnPathBatchNeverWalksBeside(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	ctx := t.Context()
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	cutAt(t, open, t0.Add(90*time.Second))
+	ws, _ := walks(t, open)
+	first := ws[0]
+	heard := t0.Add(100 * time.Second)
+	hear(t, open, lone(repoMerge("mesh-controller", "k1", heard)), heard)
+	ws, bs := walks(t, open)
+	if len(ws) != 1 || len(bs) != 1 || !bs[0].OwnPath() || bs[0].State != inventory.PlanQueued ||
+		bs[0].Delivery.Batch.Behind != first.ID || !strings.Contains(batchWords(bs[0], heard), "own path") {
+		t.Fatalf("an own-path batch beside an open walk: %d walk(s), batches %+v", len(ws), bs)
+	}
+	first.State = inventory.PlanDone
+	if err := open.inventory.SavePlan(ctx, &first); err != nil {
+		t.Fatal(err)
+	}
+	cutAt(t, open, heard.Add(time.Second))
+	ws, _ = walks(t, open)
+	if len(ws) != 2 || !ws[0].Open() || ws[0].CommitOf("novox/mesh-controller") != "k1" {
+		t.Fatalf("the own-path batch was not cut once the walk ended: %+v", ws)
+	}
+	own := ws[0]
+	// Unrelated to the own-path walk by every footprint, a merge still waits for it.
+	later := heard.Add(time.Minute)
+	hear(t, open, lone(repoMerge("two", "c2", later)), later)
+	ws, bs = walks(t, open)
+	if len(ws) != 2 || len(bs) != 1 || bs[0].State != inventory.PlanQueued || bs[0].Delivery.Batch.Behind != own.ID {
+		t.Fatalf("a merge beside the own-path walk: %d walk(s), batches %+v", len(ws), bs)
+	}
+}

cmd/mesh-controller/delivery.go

@@ -136,8 +136,9 @@ func letGo(ctx context.Context, inv *inventory.Inventory, id, by, why string) (i
 		return p, fmt.Errorf("%s was let go by %s at %s already", p.ID, p.Delivery.By,
 			p.Delivery.Go.Local().Format("15:04:05"))
 	}
-	// **One walk at a time** (novox/hq ADR 0276): the word is taken, and the walk starts once no other walk is
+	// **One walk at a time beside a related one** (novox/hq ADR 0276, ADR 0407 decision 4, issue 973): the word is
-	// started (advanceOnce), so the delivery's owner says it once and is not refused.
+	// taken, and the walk starts once no started walk it is related to is open (advanceOnce), so the delivery's
+	// owner says it once and is not refused.
 	now := time.Now().UTC()
 	p.Delivery.Go, p.Delivery.By, p.Delivery.Why = &now, by, why
 	p.Note = "let go by " + by + "; its first tier is asked next"
@@ -268,7 +269,7 @@ func stopWalkWithout(ctx context.Context, inv *inventory.Inventory, id, by, why
 		append(moves, foldMoves...)); err != nil {
 		return p, said, fmt.Errorf("%s is not stopped, and none of its merges moved: %w", p.ID, err)
 	}
-	// What the batch holds and names now: written again by the cutter's next look should this fail.
+	// What the batch holds and names now: written again by the next look should this fail.
 	if err := keepBatch(ctx, inv, &target, now, ""); err != nil {
 		fmt.Printf("  %s holds its merges; its record is written again at the next look: %v\n", target.ID, err)
 	}

cmd/mesh-controller/facts.go

@@ -287,6 +287,12 @@ func gatherFacts(ctx context.Context, open *stores, busVersion string) (snapshot
 			return snapshot.Facts{}, err
 		}
 		m.Architecture, m.Kernel = reported.Architecture, reported.Kernel
+		// Whether its node-engine takes handed files is its contract (issue 969).
+		if stated, had, err := inv.HealthOf(ctx, n.Name); err != nil {
+			return snapshot.Facts{}, err
+		} else if had {
+			m.Contract = stated.Contract
+		}
 		outward, err := inv.OutwardLinksOf(ctx, n.Name)
 		if err != nil {
 			return snapshot.Facts{}, err

cmd/mesh-controller/lone_merge_test.go

@@ -227,8 +227,9 @@ func TestAMergeRelatedToTwoBatchesFoldsThem(t *testing.T) {
 	}
 }
 
-// With a walk open, a batch cut at once still waits for it, said queued behind it (walks stay one at a time, ADR
+// With a walk open that it is related to (the same repository, issue 973), a batch cut at once still waits for
-// 0276 decision 4); it is cut once that walk ended, with no window to wait for.
+// it, said queued behind it (ADR 0276 decision 4, ADR 0407 decision 4); it is cut once that walk ended, with no
+// window to wait for.
 func TestACutAtOnceBatchQueuesBehindAnOpenWalk(t *testing.T) {
 	open := windowed(t)
 	groupsWaiting(t, "pr/c1")
@@ -241,7 +242,7 @@ func TestACutAtOnceBatchQueuesBehindAnOpenWalk(t *testing.T) {
 	}
 	first := ws[0]
 	heard := t0.Add(100 * time.Second)
-	hear(t, open, lone(repoMerge("two", "c2", heard)), heard)
+	hear(t, open, lone(repoMerge("one", "c2", heard)), heard)
 	ws, bs := walks(t, open)
 	if len(ws) != 1 || len(bs) != 1 || bs[0].State != inventory.PlanQueued || bs[0].Delivery.Batch.Behind != first.ID ||
 		!bs[0].Delivery.Batch.CutAtOnce {
@@ -280,7 +281,8 @@ func TestAStopWalksTheOthersAgainInABatchOfTheirOwnWhenTheyShareNothing(t *testi
 	cutAt(t, open, t0.Add(2*time.Minute))
 	ws, _ := walks(t, open)
 	walk := ws[0]
-	hear(t, open, repoMerge("two", "two1cccccccc", t0.Add(3*time.Minute)), t0.Add(3*time.Minute))
+	// Related to the walk by its repository (queued behind it), and to nothing the stop walks again.
+	hear(t, open, catalogueMerge("cat2cccccccc", "notes", t0.Add(3*time.Minute)), t0.Add(3*time.Minute))
 	_, bs := walks(t, open)
 	if len(bs) != 1 {
 		t.Fatalf("the merge heard while the walk ran is in %d batches", len(bs))

cmd/mesh-controller/merge_gate.go

@@ -1017,6 +1017,13 @@ func raiseFromFacts(ctx context.Context, open *stores, f snapshot.Facts, shelf m
 			"kernel": m.Kernel, "capabilities": caps}); err != nil {
 			return notes, err
 		}
+		// What its node-engine takes, as the mesh composes it today (issue 969): without the statement the gate
+		// composes a moved module the old way on a machine the mesh hands it to.
+		if stated, ok := standInHealth(m, f.Taken); ok {
+			if _, err := inv.RecordHealth(ctx, stated); err != nil {
+				return notes, err
+			}
+		}
 		for _, record := range []func(context.Context, string, string) error{inv.RecordSealingKey, inv.RecordOverlayKey} {
 			key, err := standInKey()
 			if err != nil {
@@ -1355,3 +1362,12 @@ func firstOr(items, otherwise []string) string {
 	}
 	return "it says nothing more"
 }
+
+// standInHealth is the statement the gate gives a machine of the snapshot so that it composes as the mesh does:
+// only its contract, which says whether it takes handed files (issue 969). False for a snapshot that carries none.
+func standInHealth(m snapshot.Machine, taken time.Time) (inventory.NodeHealth, bool) {
+	if m.Contract == 0 {
+		return inventory.NodeHealth{}, false
+	}
+	return inventory.NodeHealth{Node: m.Name, Contract: m.Contract, SaidAt: taken, HeardAt: taken}, true
+}

cmd/mesh-controller/merge_gate_handing_test.go

@@ -0,0 +1,85 @@
+package main
+
+import (
+	"os"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	snapshot "github.com/novox/mesh-controller/internal/facts"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// Issue 969: the gate composes a machine as the mesh does, so a moved module is handed where the machine takes
+// handed files and composed the old way where it does not.
+func TestIssue969TheGateGivesAMachineItsContractSoItTakesHandedFiles(t *testing.T) {
+	taken := time.Date(2026, 10, 16, 12, 0, 0, 0, time.UTC)
+
+	stated, ok := standInHealth(snapshot.Machine{Name: "f35", Contract: link.HandedContract}, taken)
+	if !ok || !takesHanded(stated, ok) {
+		t.Fatalf("a machine whose engine takes handed files must be given that in the gate: %+v, %v", stated, ok)
+	}
+	old, ok := standInHealth(snapshot.Machine{Name: "e12", Contract: link.HandedContract - 1}, taken)
+	if !ok || takesHanded(old, ok) {
+		t.Fatalf("a machine whose engine is older must not take handed files: %+v, %v", old, ok)
+	}
+	none, ok := standInHealth(snapshot.Machine{Name: "d01"}, taken)
+	if ok || takesHanded(none, ok) {
+		t.Fatalf("a snapshot that carries no contract states nothing: %+v, %v", none, ok)
+	}
+}
+
+// Issue 969, through the store: a machine whose engine stated the handed contract is in the snapshot with it, the
+// gate raises it as taking handed files, and one whose statement is below it is raised as not taking them. Run in
+// the merge check with the other tests that need a store; the composition itself, handed here and refused there, is
+// the catalogue's TestAHandedModuleNamingARunTimeBoundKeyComposesOnlyWhereItIsHanded.
+func TestIssue969TheGateRaisesEachMachineAsItsEngineStatedItself(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	said := time.Date(2026, 10, 16, 12, 0, 0, 0, time.UTC)
+	for node, contract := range map[string]int{"anchor": link.HandedContract, "laptop": link.HandedContract - 1} {
+		if _, err := open.inventory.RecordHealth(ctx, inventory.NodeHealth{Node: node, Contract: contract,
+			SaidAt: said, HeardAt: said}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	f, err := gatherFacts(ctx, open, "2.11.17")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(f.Machines) != 2 {
+		t.Fatalf("the snapshot holds %d machines", len(f.Machines))
+	}
+	want := map[string]bool{}
+	for _, m := range f.Machines {
+		if m.Contract == 0 {
+			t.Fatalf("%s: the snapshot carries no contract", m.Name)
+		}
+		want[m.Name] = m.Contract >= link.HandedContract
+	}
+	handing := 0
+	for _, hands := range want {
+		if hands {
+			handing++
+		}
+	}
+	if handing != 1 {
+		t.Fatalf("one machine takes handed files in the snapshot, got %v", want)
+	}
+
+	raised, drop, err := throwawayStores(ctx, os.Getenv("MESH_TEST_POSTGRES"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer drop()
+	if _, err := raiseFromFacts(ctx, raised, f, map[string]catalogue.Manifest{}, map[string]inventory.Source{}); err != nil {
+		t.Fatal(err)
+	}
+	for name, hands := range want {
+		got, err := engineTakesHanded(ctx, raised.inventory, name)
+		if err != nil || got != hands {
+			t.Errorf("%s: the gate raises it as taking handed files = %v (%v), the mesh says %v", name, got, err, hands)
+		}
+	}
+}

cmd/mesh-controller/merge_on_the_bus_test.go

@@ -0,0 +1,171 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+	"github.com/novox/mesh-controller/internal/testbus"
+)
+
+// The merge path whole, against a real bus and the store (issue 973 and issue 921, novox/hq ADR 0407 decisions 1
+// and 3; ADR 0391's proof): the forge's merged event is published on the subject the controller consumes, the
+// controller's server hands it over and the window keys are on the bus. Store-backed: skips without one.
+
+// noListener hears no node: these tests send merges only.
+type noListener struct{ link.Listener }
+
+// theMergePathOn serves the controller's consume loop and its merge window on a test bus, and returns a way to
+// publish the forge's merged event.
+func theMergePathOn(t *testing.T, open *stores) func(link.SourceMoved) {
+	t.Helper()
+	js, err := broker.Dial(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+	if err := broker.AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	if err := js.EnsureControllerBuckets(); err != nil {
+		t.Fatal(err)
+	}
+	server := link.ConnectNats(js, nil, noListener{})
+	if err := server.Follows(following{open: open}); err != nil {
+		t.Fatal(err)
+	}
+	ctx, stop := context.WithCancel(context.Background())
+	done := make(chan struct{})
+	go func() { defer close(done); _ = server.Serve(ctx) }()
+	if err := startMergeWindow(ctx, open, js); err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() {
+		stop()
+		<-done
+		setWindowClock(nil)
+		windowLook.Store(nil)
+	})
+	return func(m link.SourceMoved) {
+		t.Helper()
+		body, err := json.Marshal(m)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if _, err := js.Context().Publish(broker.ControllerFollows[3], body); err != nil {
+			t.Fatal(err)
+		}
+	}
+}
+
+func becomes(t *testing.T, within time.Duration, what string, is func() bool) {
+	t.Helper()
+	deadline := time.Now().Add(within)
+	for time.Now().Before(deadline) {
+		if is() {
+			return
+		}
+		time.Sleep(50 * time.Millisecond)
+	}
+	t.Fatalf("%s did not happen within %s", what, within)
+}
+
+func mergeKept(open *stores, repo, commit string) bool {
+	_, known, err := open.inventory.MergeOf(context.Background(), repo, commit)
+	return err == nil && known
+}
+
+// Two merges of one repository heard while another controller holds the plans, and a rollout of that repository
+// open, are both recorded once the plans are free (handed over again by the bus, never dropped), join one batch,
+// and the next rollout carries both: no person released anything (issue 921).
+func TestMergesHeardWhileAnotherHolderHasThePlansAreRecordedAndCarried(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	was, wasPoll := inventory.HoldWaitFor, inventory.HoldPoll
+	inventory.HoldWaitFor, inventory.HoldPoll = time.Second, 20*time.Millisecond
+	t.Cleanup(func() { inventory.HoldWaitFor, inventory.HoldPoll = was, wasPoll })
+	// A window of a second: a merge joining a batch opens its window again (ADR 0407 decision 1), so the batch the
+	// two held merges make is cut when that window has closed and the rollout before it has ended, not before.
+	settings := t.TempDir() + "/merge-window.json"
+	if err := os.WriteFile(settings, []byte(`{"merge-window": "1s", "merge-window-at-most": "10m"}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv(mergeWindowFileVar, settings)
+	publish := theMergePathOn(t, open)
+	ctx := t.Context()
+
+	publish(repoMerge("one", "c1", time.Now()))
+	becomes(t, 15*time.Second, "the first rollout opening", func() bool {
+		ws, _ := walks(t, open)
+		return len(ws) == 1 && ws[0].Open()
+	})
+	ws, _ := walks(t, open)
+	first := ws[0]
+
+	release, err := open.inventory.HoldPlans(ctx, true) // the other controller works the plans
+	if err != nil {
+		t.Fatal(err)
+	}
+	publish(repoMerge("one", "c2", time.Now()))
+	publish(repoMerge("one", "c3", time.Now()))
+	time.Sleep(2500 * time.Millisecond) // both met the hold and were given back to the bus
+	if mergeKept(open, "novox/one", "c2") || mergeKept(open, "novox/one", "c3") {
+		t.Fatal("a merge was recorded while another holder had the plans")
+	}
+	release()
+	becomes(t, 30*time.Second, "both merges recorded after the hold", func() bool {
+		return mergeKept(open, "novox/one", "c2") && mergeKept(open, "novox/one", "c3")
+	})
+
+	first.State = inventory.PlanDone
+	if err := open.inventory.SavePlan(ctx, &first); err != nil {
+		t.Fatal(err)
+	}
+	// The rollout ended: the look that cuts what waited behind it, once the batch's window is closed too.
+	var bs []inventory.Plan
+	becomes(t, 15*time.Second, "the next rollout being cut with both merges", func() bool {
+		advanceHeld(ctx, open)
+		ws, bs = walks(t, open)
+		return len(bs) == 0 && len(ws) == 2
+	})
+	if ws[0].CommitOf("novox/one") != "c3" || len(ws[0].Delivery.Merges) != 2 {
+		t.Fatalf("the next rollout does not carry both merges: commit %q, merges %+v", ws[0].CommitOf("novox/one"),
+			ws[0].Delivery.Merges)
+	}
+}
+
+// A batch's window closes by the expiry of its key on the bus, heard by the watcher, and the batch is cut with
+// no cutter and no call to cut (ADR 0407 decision 1): a window of a second, a merge that waits for it.
+func TestABatchWindowClosesByTheExpiryOfItsKey(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	everyHeadWaits(t)
+	settings := t.TempDir() + "/merge-window.json"
+	if err := os.WriteFile(settings, []byte(`{"merge-window": "1s", "merge-window-at-most": "10m"}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv(mergeWindowFileVar, settings)
+	publish := theMergePathOn(t, open)
+
+	start := time.Now()
+	publish(repoMerge("one", "c1", start))
+	becomes(t, 10*time.Second, "the merge waiting in a batch", func() bool {
+		_, bs := walks(t, open)
+		return len(bs) == 1 && bs[0].State == inventory.PlanAssembling
+	})
+	if ws, _ := walks(t, open); len(ws) != 0 {
+		t.Fatalf("the batch was cut before its window closed: %+v", ws)
+	}
+	becomes(t, 15*time.Second, "the batch being cut by its key's expiry", func() bool {
+		ws, bs := walks(t, open)
+		return len(ws) == 1 && len(bs) == 0
+	})
+	if took := time.Since(start); took < time.Second {
+		t.Fatalf("the batch was cut after %s, before its window of a second was up", took)
+	}
+}

cmd/mesh-controller/merge_window_bus.go

@@ -0,0 +1,198 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"github.com/nats-io/nats.go/jetstream"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+)
+
+// The merge window's clock is on the bus (novox/hq ADR 0407 decision 1, issue 973): an open batch has a key
+// `<batch>.quiet` that lives the window and is made again by each merge into the batch, and a key `<batch>.most`
+// made once when the batch opens that lives the window's most. The first of the two to expire closes the batch,
+// and the watcher hears it: no timer polls the batches. The store still keeps the times (ClosesAt, AtMost) as the
+// record of when a window closed, which `plans` reads; the keys are what wakes the controller.
+
+// windowClock is the bus's side of the merge window.
+type windowClock interface {
+	// Arm makes a batch's keys: its quiet key again, to live quiet; its most key if it has none, to live most.
+	Arm(ctx context.Context, batch string, quiet, most time.Duration) error
+	// Forget takes a batch's keys away: it was cut, or closed, and nothing is to wake the controller for it.
+	Forget(ctx context.Context, batch string)
+}
+
+// windowLook is the look a window's expiry takes: the watcher's, and the fallback timer's.
+var windowLook atomic.Pointer[func(batch string)]
+
+var (
+	windowMu    sync.RWMutex
+	windowKeyed windowClock // nil in a process that is not the serving controller
+)
+
+// setWindowClock makes clock the merge window's; nil takes it away.
+func setWindowClock(c windowClock) {
+	windowMu.Lock()
+	defer windowMu.Unlock()
+	windowKeyed = c
+}
+
+func windowClockOf() windowClock {
+	windowMu.RLock()
+	defer windowMu.RUnlock()
+	return windowKeyed
+}
+
+// windowDeadlines are how long a batch's two keys are to live from now: what is left of its window and of its
+// most, each at least a second (the least a key can live). Not armed for a batch whose window is closed already
+// (by a stop, by its time, or because it is cut at once) or that is no batch.
+func windowDeadlines(b inventory.Plan, now time.Time) (quiet, most time.Duration, armed bool) {
+	if b.Delivery == nil || b.Delivery.Batch == nil || windowClosed(b, now) {
+		return 0, 0, false
+	}
+	// Rounded up to a whole second, at least one: the server rounds a key's time to live down to whole seconds, so
+	// a fraction would close the key before the record's time and the look would cut nothing.
+	least := func(d time.Duration) time.Duration {
+		if d < time.Second {
+			return time.Second
+		}
+		return (d + time.Second - 1) / time.Second * time.Second
+	}
+	w := b.Delivery.Batch
+	return least(w.ClosesAt.Sub(now)), least(w.AtMost.Sub(now)), true
+}
+
+// armWindow puts a batch's clock on the bus after its record was kept. Where the bus will not take it the window
+// closes by a timer of this batch alone, once, as long as is left: never silently never.
+func armWindow(ctx context.Context, b inventory.Plan, now time.Time) {
+	c := windowClockOf()
+	quiet, most, armed := windowDeadlines(b, now)
+	if c == nil || !armed {
+		return
+	}
+	if err := c.Arm(ctx, b.ID, quiet, most); err != nil {
+		fmt.Printf("batches: the window of %s could not be put on the bus (%v); it closes by a timer of its own\n", b.ID, err)
+		if look := windowLook.Load(); look != nil {
+			time.AfterFunc(quiet, func() { (*look)(b.ID) })
+		}
+	}
+}
+
+// busWindow is the clock on the controller's bucket.
+type busWindow struct {
+	kv jetstream.KeyValue
+}
+
+// newBusWindow opens the bucket and starts the watcher, which names each batch whose key expired to look until
+// the context ends. Only expiry wakes it: a key taken away by Forget or made again by Arm leaves a delete.
+func newBusWindow(ctx context.Context, api jetstream.JetStream, look func(batch string)) (*busWindow, error) {
+	kv, err := api.KeyValue(ctx, broker.MergeWindowBucket)
+	if err != nil {
+		return nil, fmt.Errorf("the merge window's bucket: %w", err)
+	}
+	w, err := kv.WatchAll(ctx, jetstream.UpdatesOnly())
+	if err != nil {
+		return nil, fmt.Errorf("watching the merge window: %w", err)
+	}
+	go func() {
+		defer w.Stop()
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case e, ok := <-w.Updates():
+				if !ok {
+					return
+				}
+				if e == nil || e.Operation() != jetstream.KeyValuePurge {
+					continue // a put, or a delete this controller made
+				}
+				if batch, _, found := strings.Cut(e.Key(), "."); found && look != nil {
+					look(batch)
+				}
+			}
+		}
+	}()
+	return &busWindow{kv: kv}, nil
+}
+
+func (w *busWindow) Arm(ctx context.Context, batch string, quiet, most time.Duration) error {
+	// The most is made once, when the batch opens: it is there already for every merge after the first.
+	if _, err := w.kv.Create(ctx, batch+".most", []byte("most"), jetstream.KeyTTL(most)); err != nil &&
+		!errors.Is(err, jetstream.ErrKeyExists) {
+		return err
+	}
+	// The quiet key is made again, not changed: a key's time to live is set when it is made.
+	if err := w.kv.Delete(ctx, batch+".quiet"); err != nil && !errors.Is(err, jetstream.ErrKeyNotFound) {
+		return err
+	}
+	if _, err := w.kv.Create(ctx, batch+".quiet", []byte("quiet"), jetstream.KeyTTL(quiet)); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (w *busWindow) Forget(ctx context.Context, batch string) {
+	for _, key := range []string{batch + ".quiet", batch + ".most"} {
+		_ = w.kv.Delete(ctx, key)
+	}
+}
+
+// lookAtBatches is the look an expired window asks for: the plans held, as a merge heard holds them, and the batches
+// whose window closed cut (cutBatchesHeld). Waiting for the plans rather than giving up: an expiry is heard once.
+func lookAtBatches(ctx context.Context, open *stores) {
+	release, err := open.inventory.HoldPlans(ctx, true)
+	if err != nil {
+		fmt.Printf("batches: a window closed and the plans could not be held for the look: %v\n", err)
+		return
+	}
+	defer release()
+	if err := cutBatchesHeld(ctx, open, time.Now().UTC()); err != nil {
+		fmt.Printf("batches: %v\n", err)
+	}
+}
+
+// startMergeWindow puts the merge window's clock on the bus for as long as the context lives: the controller is
+// the lease holder here, the only one that writes it. The watcher looks at the batches when a key expires; the
+// batches open now get their keys, and one look is taken, so a window that closed while no controller served is
+// cut (ADR 0407 decision 1). No timer polls them.
+func startMergeWindow(ctx context.Context, open *stores, js *broker.JetStream) error {
+	api, err := jetstream.New(js.Conn())
+	if err != nil {
+		return err
+	}
+	look := func(string) { lookAtBatches(ctx, open) }
+	windowLook.Store(&look)
+	clock, err := newBusWindow(ctx, api, look)
+	if err != nil {
+		windowLook.Store(nil)
+		return err
+	}
+	setWindowClock(clock)
+	go func() {
+		<-ctx.Done()
+		setWindowClock(nil)
+		windowLook.Store(nil)
+	}()
+	release, err := open.inventory.HoldPlans(ctx, true)
+	if err != nil {
+		return err
+	}
+	defer release()
+	batches, err := open.inventory.Batches(ctx)
+	if err != nil {
+		return err
+	}
+	now := time.Now().UTC()
+	for _, b := range batches {
+		armWindow(ctx, b, now)
+	}
+	return cutBatchesHeld(ctx, open, now)
+}

cmd/mesh-controller/merge_window_bus_test.go

@@ -0,0 +1,151 @@
+package main
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/nats-io/nats.go"
+	"github.com/nats-io/nats.go/jetstream"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/testbus"
+)
+
+// windowOnABus is the merge window's clock on a real bus, and the batches its watcher heard expire.
+func windowOnABus(t *testing.T) (*busWindow, <-chan string) {
+	t.Helper()
+	js, err := broker.Dial(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+	if err := js.EnsureControllerBuckets(); err != nil {
+		t.Fatal(err)
+	}
+	conn, err := nats.Connect(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(conn.Close)
+	api, err := jetstream.New(conn)
+	if err != nil {
+		t.Fatal(err)
+	}
+	heard := make(chan string, 16)
+	ctx, stop := context.WithCancel(context.Background())
+	t.Cleanup(stop)
+	w, err := newBusWindow(ctx, api, func(batch string) { heard <- batch })
+	if err != nil {
+		t.Fatal(err)
+	}
+	return w, heard
+}
+
+func hears(t *testing.T, heard <-chan string, within time.Duration) (string, bool) {
+	t.Helper()
+	select {
+	case b := <-heard:
+		return b, true
+	case <-time.After(within):
+		return "", false
+	}
+}
+
+// A batch's quiet key expires when no merge came for the window, and the watcher names the batch (issue 973,
+// novox/hq ADR 0407 decision 1).
+func TestTheQuietKeyExpiresAndTheWatcherNamesTheBatch(t *testing.T) {
+	w, heard := windowOnABus(t)
+	if err := w.Arm(t.Context(), "plan-1", time.Second, time.Minute); err != nil {
+		t.Fatal(err)
+	}
+	if b, ok := hears(t, heard, 10*time.Second); !ok || b != "plan-1" {
+		t.Fatalf("heard %q (%v); want plan-1", b, ok)
+	}
+}
+
+// A merge into the batch makes the quiet key again, so the window restarts: it does not expire at the first
+// deadline, and the batch's most closes it all the same.
+func TestAMergeRestartsTheQuietKeyAndTheMostKeyClosesAllTheSame(t *testing.T) {
+	w, heard := windowOnABus(t)
+	ctx := t.Context()
+	if err := w.Arm(ctx, "plan-2", 2*time.Second, 5*time.Second); err != nil {
+		t.Fatal(err)
+	}
+	time.Sleep(1200 * time.Millisecond)
+	if err := w.Arm(ctx, "plan-2", 2*time.Second, 5*time.Second); err != nil { // a merge: quiet again, most as it was
+		t.Fatal(err)
+	}
+	if b, ok := hears(t, heard, 1500*time.Millisecond); ok {
+		t.Fatalf("heard %q before the restarted window was up", b)
+	}
+	// Another merge keeps the window open; the most (5 s from the first) closes it at about 5 s.
+	start := time.Now()
+	for time.Since(start) < 2500*time.Millisecond {
+		if err := w.Arm(ctx, "plan-2", 2*time.Second, 5*time.Second); err != nil {
+			t.Fatal(err)
+		}
+		time.Sleep(700 * time.Millisecond)
+	}
+	if b, ok := hears(t, heard, 8*time.Second); !ok || b != "plan-2" {
+		t.Fatalf("the most did not close the batch: heard %q (%v)", b, ok)
+	}
+}
+
+// A batch cut is forgotten: its keys are taken away by a delete, which is no expiry and wakes nobody.
+func TestForgettingABatchWakesNobody(t *testing.T) {
+	w, heard := windowOnABus(t)
+	if err := w.Arm(t.Context(), "plan-3", time.Second, time.Minute); err != nil {
+		t.Fatal(err)
+	}
+	w.Forget(t.Context(), "plan-3")
+	if b, ok := hears(t, heard, 3*time.Second); ok {
+		t.Fatalf("a forgotten batch was heard expire: %q", b)
+	}
+}
+
+// The deadlines a batch's keys are made with: what is left of its window and of its most, at least a second;
+// none for a batch whose window is closed or that is no batch.
+func TestTheDeadlinesOfABatchsKeys(t *testing.T) {
+	now := time.Date(2026, 10, 17, 12, 0, 0, 0, time.UTC)
+	batch := func(closes, most time.Duration, closed bool) inventory.Plan {
+		return inventory.Plan{State: inventory.PlanAssembling, Delivery: &inventory.PlanDelivery{Batch: &inventory.PlanBatch{
+			ClosesAt: now.Add(closes), AtMost: now.Add(most), Closed: closed}}}
+	}
+	for _, c := range []struct {
+		what        string
+		b           inventory.Plan
+		quiet, most time.Duration
+		armed       bool
+	}{
+		{"open", batch(90*time.Second, 10*time.Minute, false), 90 * time.Second, 10 * time.Minute, true},
+		{"almost up", batch(100*time.Millisecond, 5*time.Second, false), time.Second, 5 * time.Second, true},
+		{"closed by a stop", batch(90*time.Second, 10*time.Minute, true), 0, 0, false},
+		{"up", batch(-time.Second, 10*time.Minute, false), 0, 0, false},
+		{"cut at once", func() inventory.Plan {
+			b := batch(90*time.Second, 10*time.Minute, false)
+			b.Delivery.Batch.CutAtOnce = true
+			return b
+		}(), 0, 0, false},
+		{"no batch", inventory.Plan{}, 0, 0, false},
+	} {
+		quiet, most, armed := windowDeadlines(c.b, now)
+		if armed != c.armed || quiet != c.quiet || most != c.most {
+			t.Errorf("%s: %v, %v, %v; want %v, %v, %v", c.what, quiet, most, armed, c.quiet, c.most, c.armed)
+		}
+	}
+}
+
+// A deadline that is not a whole number of seconds is rounded up, never down: the server rounds a key's time to live
+// down to whole seconds, so 89.999999 s would expire at 89 s, before the window is closed by the record's times, and
+// the look would cut nothing (review of issue 973).
+func TestAFractionalDeadlineIsRoundedUpToTheNextWholeSecond(t *testing.T) {
+	now := time.Date(2026, 10, 17, 12, 0, 0, 0, time.UTC)
+	b := inventory.Plan{State: inventory.PlanAssembling, Delivery: &inventory.PlanDelivery{Batch: &inventory.PlanBatch{
+		ClosesAt: now.Add(89*time.Second + 999999*time.Microsecond), AtMost: now.Add(599*time.Second + time.Nanosecond)}}}
+	quiet, most, armed := windowDeadlines(b, now)
+	if !armed || quiet != 90*time.Second || most != 600*time.Second {
+		t.Errorf("deadlines %v, %v (armed %v); want 1m30s, 10m0s", quiet, most, armed)
+	}
+}

cmd/mesh-controller/missed_merges.go

@@ -1,243 +0,0 @@
-package main
-
-import (
-	"context"
-	"fmt"
-	"sync"
-	"time"
-
-	"github.com/novox/mesh-controller/internal/inventory"
-	"github.com/novox/mesh-controller/internal/link"
-)
-
-// **A merge the bus announced and the controller never acted on is caught up** (novox/hq issue 266).
-//
-// The forge's poll announces every merge on the events stream, and the controller acts on what its
-// consumer there hands it. On 2026-10-06 one merge was on the stream and never handed over: the bus
-// server moved the consumer past it — a fault of consumers with several filters in the server the
-// mesh ran — and the controller, which only acts on what it is handed, said nothing. The modules
-// built from that repository stayed behind and were built by hand.
-//
-// So the stream is read back on a timer, on a consumer of its own filtered on merges alone, and every
-// announcement older than mergeGrace is judged as SourceMoved would judge it. **No record of what
-// was handled is kept, because none is needed**: acting on a merge marks every module it moved as
-// looked at since, so an announcement already acted on reads as history and moves nothing. One that
-// would still move something was never acted on — it is said, and acted on now.
-const (
-	// mergeGrace is how long an announcement is left to the controller's own consumer before it is
-	// judged missed. That consumer hands over one event at a time, and a merge waits behind a build
-	// outcome that is being acted on; acting on a merge itself asks builds and does not wait for them.
-	mergeGrace = 10 * time.Minute
-	// mergeLookBack is how far back a pass reads. A merge missed longer ago than this was missed by a
-	// controller that was not running this, and is the operator's to look at, not a surprise rebuild.
-	mergeLookBack = 24 * time.Hour
-	// mergeCatchUpEvery is how often the stream is read back.
-	mergeCatchUpEvery = 5 * time.Minute
-)
-
-// merges is what reads back the forge's announcements; the link server, or a test's list.
-type merges interface {
-	AnnouncedMerges(ctx context.Context, since time.Time) ([]link.AnnouncedMerge, error)
-}
-
-// unheardMerges is the announcements of merges the controller has not heard (novox/hq ADR 0276): a merge kept
-// in a batch is not acted on until its batch is cut, which can be past mergeGrace behind a long walk, and is
-// not missed for that.
-type unheardMerges struct {
-	merges
-	inv *inventory.Inventory
-}
-
-func (u unheardMerges) AnnouncedMerges(ctx context.Context, since time.Time) ([]link.AnnouncedMerge, error) {
-	all, err := u.merges.AnnouncedMerges(ctx, since)
-	if err != nil {
-		return nil, err
-	}
-	var out []link.AnnouncedMerge
-	for _, a := range all {
-		_, kept, err := u.inv.MergeOf(ctx, a.Owner+"/"+a.Repo, a.Commit)
-		if err != nil {
-			return nil, err
-		}
-		if !kept {
-			out = append(out, a)
-		}
-	}
-	return out, nil
-}
-
-// owedToTheRecord says a merge read as history is still owed to the record (novox/hq ADR 0276): the merges
-// reader knows, when it keeps them (unheardMerges).
-func owedToTheRecord(ctx context.Context, announced merges, m link.SourceMoved, entries []inventory.Entry,
-	read map[string][]inventory.ReadRepository) bool {
-	u, ok := announced.(unheardMerges)
-	if !ok {
-		return false
-	}
-	merged, err := time.Parse(time.RFC3339Nano, m.MergedAt)
-	if err != nil {
-		return false
-	}
-	raw := m
-	raw.MergedAt = ""
-	owed, err := owedLate(ctx, u.inv, m, merged, wouldMove(raw, entries, read))
-	return err == nil && owed
-}
-
-// catchingUpOnMerges reads back the forge's announcements on a timer, until the context ends.
-func catchingUpOnMerges(ctx context.Context, open *stores, announced merges) {
-	f := following{open: open}
-	catalogued := func(ctx context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error) {
-		entries, err := open.inventory.Catalogued(ctx)
-		if err != nil {
-			return nil, nil, err
-		}
-		read, err := readForPlanning(ctx, open.inventory)
-		return entries, read, err
-	}
-	failing := ""
-	tick := time.NewTicker(mergeCatchUpEvery)
-	defer tick.Stop()
-	for {
-		select {
-		case <-ctx.Done():
-			return
-		case <-tick.C:
-		}
-		watchedMerges.begin()
-		err := catchUpOnMerges(ctx, time.Now(), unheardMerges{announced, open.inventory}, catalogued, f.SourceMoved, func(format string, args ...any) {
-			fmt.Printf(format+"\n", args...)
-		})
-		// What the pass found is what S5 says (novox/hq to-be 45 §3); a pass that could not read
-		// says that instead, and leaves what the last one found standing.
-		watchedMerges.end(time.Now(), err)
-		// A pass that cannot read says so once, not every five minutes, and says when it reads again.
-		why := ""
-		if err != nil {
-			why = err.Error()
-		}
-		if why != failing {
-			if why != "" {
-				fmt.Printf("merges the bus may not have handed over cannot be looked for: %s\n", why)
-			} else {
-				fmt.Println("merges the bus may not have handed over are looked for again")
-			}
-			failing = why
-		}
-	}
-}
-
-// catchUpOnMerges is one pass: every announcement older than mergeGrace that acting on would still
-// move something is said and acted on, oldest first.
-//
-// Judged twice: once against the catalogue as the pass found it, and again just before acting,
-// because acting on an earlier missed merge of the same repository may have moved what a later one
-// would have.
-func catchUpOnMerges(ctx context.Context, now time.Time, announced merges,
-	catalogued func(context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error),
-	act func(context.Context, link.SourceMoved) error, say func(string, ...any)) error {
-
-	all, err := announced.AnnouncedMerges(ctx, now.Add(-mergeLookBack))
-	if err != nil {
-		return err
-	}
-	entries, read, err := catalogued(ctx)
-	if err != nil {
-		return err
-	}
-	for _, a := range all {
-		// A merge the forge said no time of is dated by its announcement, so a packaging module's look can
-		// make it history once acted on, and the catch-up does not act on it again every pass (ADR 0267).
-		if a.SourceMoved.MergedAt == "" && !a.At.IsZero() {
-			a.SourceMoved.MergedAt = a.At.UTC().Format(time.RFC3339)
-		}
-		if now.Sub(a.At) < mergeGrace {
-			continue
-		}
-		if len(wouldMove(a.SourceMoved, entries, read)) == 0 && !owedToTheRecord(ctx, announced, a.SourceMoved, entries, read) {
-			continue
-		}
-		if entries, read, err = catalogued(ctx); err != nil {
-			return err
-		}
-		moves := wouldMove(a.SourceMoved, entries, read)
-		if len(moves) == 0 {
-			if !owedToTheRecord(ctx, announced, a.SourceMoved, entries, read) {
-				continue
-			}
-			raw := a.SourceMoved
-			raw.MergedAt = ""
-			moves = wouldMove(raw, entries, read)
-		}
-		var names []string
-		for _, e := range moves {
-			names = append(names, e.Manifest.Module)
-		}
-		watchedMerges.found(missedMerge{Owner: a.Owner, Repo: a.Repo, Base: a.Base, Commit: a.Commit,
-			At: a.At, Modules: names})
-		say("%s/%s merged into %s (%.8s), announced %s ago, and the controller never acted on it: the bus "+
-			"did not hand the announcement over (novox/hq issue 266). %s %s behind it; acting on it now",
-			a.Owner, a.Repo, a.Base, a.Commit, now.Sub(a.At).Round(time.Minute), readableList(names),
-			isAre(len(names)))
-		if err := act(ctx, a.SourceMoved); err != nil {
-			say("%s/%s moved to %.8s and the mesh could not act on it: %v; the next pass tries again",
-				a.Owner, a.Repo, a.Commit, err)
-			continue
-		}
-		if entries, read, err = catalogued(ctx); err != nil {
-			return err
-		}
-	}
-	return nil
-}
-
-// missedMerge is one merge the bus announced and never handed over, as a pass found it.
-type missedMerge struct {
-	Owner, Repo, Base, Commit string
-	// At is when the bus took the announcement.
-	At      time.Time
-	Modules []string
-}
-
-// mergeWatch is what the passes found, for S5 (novox/hq to-be 45 §3): **a merge nothing read is
-// said**, urgent, even though the pass acts on it at once — the bus skipping a message is a fault of
-// the transport the mesh's every change rides on, and acting late is the repair, not the absence of
-// the fault. The next pass, finding it acted on, clears it.
-type mergeWatch struct {
-	mu      sync.Mutex
-	passed  time.Time
-	err     error
-	finding []missedMerge
-	missed  []missedMerge
-}
-
-var watchedMerges = &mergeWatch{}
-
-func (w *mergeWatch) begin() {
-	w.mu.Lock()
-	defer w.mu.Unlock()
-	w.finding = nil
-}
-
-func (w *mergeWatch) found(m missedMerge) {
-	w.mu.Lock()
-	defer w.mu.Unlock()
-	w.finding = append(w.finding, m)
-}
-
-func (w *mergeWatch) end(at time.Time, err error) {
-	w.mu.Lock()
-	defer w.mu.Unlock()
-	w.passed, w.err = at, err
-	if err == nil {
-		w.missed = w.finding
-	}
-}
-
-// last is when the last pass ended, what the last pass that read found, and what the last pass
-// could not read.
-func (w *mergeWatch) last() (time.Time, []missedMerge, error) {
-	w.mu.Lock()
-	defer w.mu.Unlock()
-	return w.passed, append([]missedMerge(nil), w.missed...), w.err
-}

cmd/mesh-controller/missed_merges_test.go

@@ -1,180 +0,0 @@
-package main
-
-import (
-	"context"
-	"errors"
-	"fmt"
-	"strings"
-	"testing"
-	"time"
-
-	"github.com/novox/mesh-controller/internal/inventory"
-	"github.com/novox/mesh-controller/internal/link"
-)
-
-// announcedList is the events stream's merges as a test gives them.
-type announcedList []link.AnnouncedMerge
-
-func (a announcedList) AnnouncedMerges(_ context.Context, since time.Time) ([]link.AnnouncedMerge, error) {
-	var out []link.AnnouncedMerge
-	for _, m := range a {
-		if !m.At.Before(since) {
-			out = append(out, m)
-		}
-	}
-	return out, nil
-}
-
-// aCatalogue is what the inventory holds, changed the way acting on a merge changes it: every module
-// the merge moved is marked as looked at (inventory.SourceMoved writes source_seen = now()).
-type aCatalogue struct {
-	entries []inventory.Entry
-	acted   []string
-	fail    error
-}
-
-func (c *aCatalogue) read(context.Context) ([]inventory.Entry, map[string][]inventory.ReadRepository, error) {
-	return append([]inventory.Entry(nil), c.entries...), nil, nil
-}
-
-func (c *aCatalogue) act(now func() time.Time) func(context.Context, link.SourceMoved) error {
-	return func(_ context.Context, m link.SourceMoved) error {
-		if c.fail != nil {
-			return c.fail
-		}
-		c.acted = append(c.acted, m.Repo+"@"+m.Commit[:8])
-		for _, moved := range wouldMove(m, c.entries, nil) {
-			for i := range c.entries {
-				if c.entries[i].Manifest.Module == moved.Manifest.Module {
-					c.entries[i].Source.Head = m.Commit
-					c.entries[i].Source.Seen = now()
-				}
-			}
-		}
-		return nil
-	}
-}
-
-func at(s string) time.Time {
-	t, err := time.Parse(time.RFC3339, s)
-	if err != nil {
-		panic(err)
-	}
-	return t
-}
-
-func announced(repo, commit, mergedAt, onTheBus string, paths ...string) link.AnnouncedMerge {
-	return link.AnnouncedMerge{
-		SourceMoved: link.SourceMoved{Owner: "novox", Repo: repo, Base: "main", Commit: commit,
-			MergedAt: mergedAt, Paths: paths,
-			CloneURL: "http://forge.internal:20000/novox/" + repo + ".git"},
-		At: at(onTheBus),
-	}
-}
-
-func built(module, repo, path, commit, seen string) inventory.Entry {
-	e := fromRepo(module, "http://forge.internal:20000/novox/"+repo+".git", path)
-	e.Source.BuiltFrom, e.Source.Head, e.Source.Seen = commit, commit, at(seen)
-	return e
-}
-
-// **novox/hq issue 266, as it happened.** The forge announced a merge of the tools repository on the
-// events stream; the bus never handed it to the controller, which acted on the merges around it and
-// not on this one, and said nothing. Read back from the stream, it is the one merge that would still
-// move something — so it is said and acted on, once, and only after the controller's own consumer
-// has had its time with it.
-func TestAMergeTheBusNeverHandedOverIsActedOnLate(t *testing.T) {
-	cat := &aCatalogue{entries: []inventory.Entry{
-		built("mesh-tools", "mesh-tools", "", "8b789578aaaaaaaa", "2026-10-04T15:24:32Z"),
-		built("node-tools", "mesh-tools", "node-tools", "8b789578aaaaaaaa", "2026-10-04T15:24:32Z"),
-		// Acted on when it was announced: looked at after it was merged.
-		built("gitea", "mesh-catalog", "modules/gitea", "5c2157b8bbbbbbbb", "2026-10-05T22:39:21Z"),
-	}}
-	stream := announcedList{
-		// Nothing the mesh holds is built from the records repository.
-		announced("hq", "88f7f79fcccccccc", "2026-10-05T22:43:00Z", "2026-10-05T22:43:04Z", "04-ISSUES/x.md"),
-		// Acted on: its module was looked at since.
-		announced("mesh-catalog", "78328d4adddddddd", "2026-10-05T22:39:00Z", "2026-10-05T22:39:21Z", "modules/gitea/x.ts"),
-		// Never handed over.
-		announced("mesh-tools", "9730bd89c3e48d0e", "2026-10-05T22:46:47Z", "2026-10-05T22:47:06Z",
-			"node-tools/internal/console/console.go"),
-	}
-	var said []string
-	say := func(format string, args ...any) { said = append(said, fmt.Sprintf(format, args...)) }
-	clock := at("2026-10-05T22:50:00Z")
-	now := func() time.Time { return clock }
-	pass := func() {
-		t.Helper()
-		if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(now), say); err != nil {
-			t.Fatal(err)
-		}
-	}
-
-	pass()
-	if len(cat.acted) != 0 {
-		t.Fatalf("a merge three minutes old was taken from the controller's own consumer: %v", cat.acted)
-	}
-
-	clock = at("2026-10-05T22:58:00Z")
-	pass()
-	if strings.Join(cat.acted, ",") != "mesh-tools@9730bd89" {
-		t.Fatalf("acted on %v, wanted the one merge never handed over", cat.acted)
-	}
-	if len(said) != 1 || !strings.Contains(said[0], "novox/mesh-tools merged into main (9730bd89)") ||
-		!strings.Contains(said[0], "mesh-tools and node-tools are behind it") {
-		t.Fatalf("the missed merge was not said as one: %q", said)
-	}
-
-	clock = at("2026-10-05T23:03:00Z")
-	pass()
-	if len(cat.acted) != 1 || len(said) != 1 {
-		t.Fatalf("a merge acted on was acted on again: %v %q", cat.acted, said)
-	}
-}
-
-// A merge that changed none of the held modules' files moves nothing, so it is never "missed"; one
-// that could not be acted on is said and tried again on the next pass.
-func TestAMissedMergeThatCouldNotBeActedOnIsTriedAgain(t *testing.T) {
-	cat := &aCatalogue{
-		entries: []inventory.Entry{built("gitea", "mesh-catalog", "modules/gitea", "5c2157b8bbbbbbbb", "2026-10-05T20:00:00Z")},
-		fail:    errors.New("the store is restarting"),
-	}
-	stream := announcedList{
-		announced("mesh-catalog", "aaaaaaaa11111111", "2026-10-05T21:00:00Z", "2026-10-05T21:00:10Z",
-			"modules/plex/module.json", "modules/plex/x.ts"),
-		announced("mesh-catalog", "bbbbbbbb22222222", "2026-10-05T21:10:00Z", "2026-10-05T21:10:10Z", "modules/gitea/x.ts"),
-	}
-	var said []string
-	say := func(format string, args ...any) { said = append(said, fmt.Sprintf(format, args...)) }
-	clock := at("2026-10-05T22:00:00Z")
-	now := func() time.Time { return clock }
-	if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(now), say); err != nil {
-		t.Fatal(err)
-	}
-	if len(said) != 2 || !strings.Contains(said[1], "could not act on it") {
-		t.Fatalf("a failed catch-up was not said: %q", said)
-	}
-	cat.fail = nil
-	clock = at("2026-10-05T22:05:00Z")
-	if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(now), say); err != nil {
-		t.Fatal(err)
-	}
-	if strings.Join(cat.acted, ",") != "mesh-catalog@bbbbbbbb" {
-		t.Fatalf("acted on %v, wanted only the merge that changed a held module", cat.acted)
-	}
-}
-
-// A merge older than the look-back is left to the operator: a controller that did not run this
-// missed it, and acting on it days later would be a surprise rebuild.
-func TestAMergeOlderThanTheLookBackIsLeftAlone(t *testing.T) {
-	cat := &aCatalogue{entries: []inventory.Entry{built("gitea", "mesh-catalog", "modules/gitea", "5c2157b8bbbbbbbb", "2026-10-01T00:00:00Z")}}
-	stream := announcedList{announced("mesh-catalog", "cccccccc33333333", "2026-10-03T00:00:00Z", "2026-10-03T00:00:05Z", "modules/gitea/x.ts")}
-	clock := at("2026-10-05T22:00:00Z")
-	if err := catchUpOnMerges(context.Background(), clock, stream, cat.read, cat.act(func() time.Time { return clock }),
-		func(string, ...any) {}); err != nil {
-		t.Fatal(err)
-	}
-	if len(cat.acted) != 0 {
-		t.Fatalf("a merge of three days ago was acted on: %v", cat.acted)
-	}
-}

cmd/mesh-controller/plain_words.go

@@ -429,12 +429,6 @@ var plainWordings = map[string]func(conditions.Observation) words{
 				"out) has not let it start yet.",
 			Resolved: "Resolved: the delivery is no longer waiting"}
 	}),
-	"merge-not-acted": worded(func(o conditions.Observation) words {
-		return words{Headline: "A merge was never picked up",
-			Explanation: "A pull request was merged, and the controller never heard of it, so nothing was built or " +
-				"sent. The mesh catches up on missed merges by itself.",
-			Resolved: "Resolved: the merge was picked up"}
-	}),
 	"ask-lost": worded(func(o conditions.Observation) words {
 		return words{Headline: "A build has no result",
 			Explanation: "A build was asked for and never answered. It is asked again with the next delivery.",

cmd/mesh-controller/plan.go

@@ -1126,7 +1126,13 @@ func engineTakesHanded(ctx context.Context, inv *inventory.Inventory, node strin
 	if err != nil {
 		return false, err
 	}
-	return had && stated.Contract >= link.HandedContract, nil
+	return takesHanded(stated, had), nil
+}
+
+// takesHanded is whether a machine's newest statement (had: it ever stated one) is of an engine that writes a
+// filled file on the provisioner client's word alone.
+func takesHanded(stated inventory.NodeHealth, had bool) bool {
+	return had && stated.Contract >= link.HandedContract
 }
 
 // provisionerAccount is the account the provisioner's client runs as on this machine (novox/hq ADR 0349 §4): the user

cmd/mesh-controller/plan_retry.go

@@ -197,12 +197,23 @@ func retryRefusal(p inventory.Plan, plans []inventory.Plan) error {
 	return oneWalkAtATime(p, plans)
 }
 
-// oneWalkAtATime refuses a retry while another walk is open, started or waiting for its word (novox/hq ADR
+// oneWalkAtATime refuses a retry while an open walk, started or waiting for its word, is related to it (novox/hq
-// 0276): a walk retried beside it would be two walks at once.
+// ADR 0276, ADR 0407 decision 4, issue 973): a walk retried beside a related one would be two walks on the same
+// things at once. A walk related to none may be retried beside them. The retried walk is judged by the modules it
+// plans, which already hold their dependents, and the repositories it carries.
 func oneWalkAtATime(p inventory.Plan, plans []inventory.Plan) error {
+	mine := footprintOfWalk(p, nil)
 	for _, q := range plans {
-		if q.ID != p.ID && q.Open() && q.Release == nil {
+		if q.ID == p.ID || !q.Open() || q.Release != nil {
-			return fmt.Errorf("%s is open (%s): one walk at a time — retry %s once it ended", q.ID, q.Named(), p.ID)
+			continue
+		}
+		if ownWalk(p) || ownWalk(q) {
+			return fmt.Errorf("%s is open (%s) and one of them is on the controller's own path: one walk at a time — "+
+				"retry %s once it ended", q.ID, q.Named(), p.ID)
+		}
+		if ok, why := related(mine, footprintOfWalk(q, nil), nil); ok {
+			return fmt.Errorf("%s is open (%s) and %s: one walk at a time beside a related one — retry %s once it ended",
+				q.ID, q.Named(), why, p.ID)
 		}
 	}
 	return nil

cmd/mesh-controller/probe_agent_root.go

@@ -320,7 +320,11 @@ func probeAgentRoot(ctx context.Context, d *doctor) ([]conditions.Observation, e
 		quiet, err := searchStillRunning(ctx, inv, node, n.AgentAccount, h, had, now)
 		return err == nil && quiet
 	}
-	return rootObservations(ctx, r, trustedMachines(r.entries), time.Now()), nil
+	obs := rootObservations(ctx, r, trustedMachines(r.entries), time.Now())
+	// A membership that lost its verified sender to a judgement that failed at a push gets it back here, once
+	// the machines pass (issue 930); the bound is one self-check.
+	giveVerifiedSendersBack(ctx, d)
+	return obs, nil
 }
 
 // rootObservations judges the machines and says each that fails.
@@ -365,14 +369,3 @@ func rootFreeAnswer(ctx context.Context, machines string, now time.Time) (any, e
 	}
 	return map[string]any{"machines": judgeRootFree(ctx, newRootReader(ctx, d.open.inventory, conn), names, now)}, nil
 }
-
-// rootFreeNow is the machines judged root-free, for composing a push's memberships: only those that pass.
-func rootFreeNow(ctx context.Context, r *rootReader, machines []string, now time.Time) map[string]bool {
-	free := map[string]bool{}
-	for _, v := range judgeRootFree(ctx, r, machines, now) {
-		if v.Free {
-			free[v.Machine] = true
-		}
-	}
-	return free
-}

cmd/mesh-controller/push.go

@@ -183,9 +183,12 @@ func serve(ctx context.Context) (err error) {
 	// Open plans move on a timer as well as on outcomes (novox/hq ADR 0162): a tier waiting for
 	// machines to report moves when they have, and a plan left by a replaced controller resumes.
 	go planTicker(ctx, open)
-	// And the merge window (novox/hq ADR 0276): a batch is cut into its walk when the window closes, which no
+	// And the merge window (novox/hq ADR 0276, ADR 0407): a batch is cut into its walk when the window closes,
-	// merge or outcome says.
+	// which no merge or outcome says. The window is two keys on the bus that expire and a watcher that hears it;
-	go batchCutter(ctx, open)
+	// nothing polls (issue 973).
+	if err := startMergeWindow(ctx, open, server.JetStream()); err != nil {
+		return fmt.Errorf("the merge window could not be put on the bus: %w", err)
+	}
 	// And the pending assignments settled on a tick of their own (novox/hq ADR 0261): made once their module
 	// is registered, ended with why when its build will not register it, raised and cleared as conditions.
 	// Never by a read.
@@ -198,9 +201,8 @@ func serve(ctx context.Context) (err error) {
 	if err := server.Follows(following{open: open}); err != nil {
 		return err
 	}
-	// And the merges the bus announced and never handed over, read back on a timer and acted on late
+	// No timer reads the merges back (novox/hq ADR 0407 decision 3, issue 973): the forge's event and the durable
-	// rather than never (novox/hq issue 266).
+	// consumer on the bus keep every merge, and one the controller cannot record yet is held and handed over again.
-	go catchingUpOnMerges(ctx, open, server)
 	// And a catalogue that has just started, asking for what it missed. The same type answers
 	// both: what a build meant and what the builds were are two questions about one record.
 	if err := server.Answers(following{open: open}); err != nil {
@@ -1333,8 +1335,14 @@ func issueMemberships(ctx context.Context, open *stores, server *link.Server, se
 	}
 	// Which machines are root-free now (novox/hq ADR 0259 §8): a channel's verified sender is composed for the
 	// router only from one, beside a router on one. Judged once per push, by the root-free verb's judgement.
-	records.RootFree = rootFreeNow(ctx, newRootReader(ctx, open.inventory, bus.Conn), records.Nodes, time.Now())
+	var whys map[string]string
+	records.RootFree, whys = rootVerdictsOf(judgeRootFree(ctx, newRootReader(ctx, open.inventory, bus.Conn),
+		records.Nodes, time.Now()))
 	where := broker.PlacementsOf(records, records.Interchangeable)
+	// What the judgement kept off is said, with its reason (novox/hq issue 930); the self-check gives it back.
+	for _, line := range droppedVerifiedSenders(records, where, whys) {
+		fmt.Println(line)
+	}
 	// **Every declared state's bucket, before the memberships that name it** (novox/hq ADR 0201). The
 	// raise at start asserts them too, but a module registered and assigned since would otherwise have
 	// its bucket only after the control plane next restarts — found the first time a module declared

cmd/mesh-controller/reissue_verified_sender.go

@@ -0,0 +1,224 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"slices"
+	"sort"
+	"strings"
+	"time"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A channel's verified sender is composed into a membership only while the machines are judged root-free at
+// that moment (novox/hq ADR 0259 §8), and the judgement fails closed when a read fails (novox/hq issue 930):
+// a push composed during a failed read issued the router a membership without it, and nothing composed it
+// again, so every approval was refused until somebody pushed. Two things follow.
+//
+// What a composition drops it says (droppedVerifiedSenders), with the reason the judgement gave.
+//
+// And the self-check issues it again (reissueVerifiedSenders): each time it judges the machines, a membership
+// that lacks a verified sender the composition would now give it is issued again with only its kinds changed.
+// **Bound: one self-check, doctorEvery (five minutes), after the judgement passes.** It only ever adds: a
+// machine that stops passing keeps what it was issued, as at a push, because the router asks the root-free
+// verb again before it honours an approval, and that is the check that holds.
+
+// droppedVerifiedSenders says, one line each, every kind that declares `verified-sender` and was composed
+// without it, naming the kind, its machine and why. whys is each machine's verdict reason (rootVerdict.Why).
+func droppedVerifiedSenders(r broker.Records, where broker.Placements, whys map[string]string) []string {
+	var routerNodes []string
+	for node, declared := range r.Assigned {
+		for _, d := range declared {
+			for _, s := range d.Holds {
+				if s.Name == routerSeat && !slices.Contains(routerNodes, node) {
+					routerNodes = append(routerNodes, node)
+				}
+			}
+		}
+	}
+	sort.Strings(routerNodes)
+	var out []string
+	for _, k := range where.Kinds {
+		declared := false
+		for _, d := range r.Assigned[k.Node] {
+			for _, s := range d.Holds {
+				if d.Module == k.Module && s.Name == k.Seat && s.Kind == k.Kind && slices.Contains(s.Capabilities, "verified-sender") {
+					declared = true
+				}
+			}
+		}
+		if !declared || slices.Contains(k.Capabilities, "verified-sender") {
+			continue
+		}
+		var why string
+		switch {
+		case !r.RootFree[k.Node]:
+			why = k.Node + " is not judged root-free: " + whys[k.Node]
+		case len(routerNodes) != 1:
+			why = fmt.Sprintf("the router is placed on %d machines, not one", len(routerNodes))
+		case !r.RootFree[routerNodes[0]]:
+			why = "the router's machine " + routerNodes[0] + " is not judged root-free: " + whys[routerNodes[0]]
+		default:
+			why = k.Module + " does not run as an account of its own"
+		}
+		out = append(out, fmt.Sprintf("  %s kind %s of %s on %s is composed without its verified sender: %s",
+			k.Seat, k.Kind, k.Module, k.Node, why))
+	}
+	return out
+}
+
+// withVerifiedSenders is a membership as issued with its seat traffic's kinds replaced by the composed ones,
+// and whether that changed it: only when a composed kind carries a verified sender the issued one lacks. Pure.
+func withVerifiedSenders(issued []byte, composed *broker.SeatTraffic) ([]byte, bool, error) {
+	if composed == nil {
+		return issued, false, nil
+	}
+	var fields map[string]json.RawMessage
+	if err := json.Unmarshal(issued, &fields); err != nil {
+		return nil, false, err
+	}
+	var traffic map[string]json.RawMessage
+	var had []broker.KindHeld
+	if raw, ok := fields["seat-traffic"]; ok {
+		if err := json.Unmarshal(raw, &traffic); err != nil {
+			return nil, false, err
+		}
+		if raw, ok := traffic["kinds"]; ok {
+			if err := json.Unmarshal(raw, &had); err != nil {
+				return nil, false, err
+			}
+		}
+	}
+	gains := false
+	for _, k := range composed.Kinds {
+		if !slices.Contains(k.Capabilities, "verified-sender") {
+			continue
+		}
+		held := false
+		for _, h := range had {
+			if h.Seat == k.Seat && h.Kind == k.Kind && h.Module == k.Module && h.Node == k.Node &&
+				slices.Contains(h.Capabilities, "verified-sender") {
+				held = true
+			}
+		}
+		gains = gains || !held
+	}
+	if !gains {
+		return issued, false, nil
+	}
+	if traffic == nil {
+		traffic = map[string]json.RawMessage{}
+	}
+	raw, err := json.Marshal(composed.Kinds)
+	if err != nil {
+		return nil, false, err
+	}
+	traffic["kinds"] = raw
+	if fields["seat-traffic"], err = json.Marshal(traffic); err != nil {
+		return nil, false, err
+	}
+	body, err := json.Marshal(fields)
+	return body, true, err
+}
+
+// reissueVerifiedSenders issues again each membership that lacks a verified sender the records now give it:
+// read as last issued, its kinds replaced, published only if no newer one was issued since (a push runs beside
+// the self-check, and its newer membership must not be overwritten by the older fields read here). One never issued is left to its own send. Failures are
+// said and passed over; the next self-check tries again.
+func reissueVerifiedSenders(r broker.Records, where broker.Placements,
+	last func(node, module string) ([]byte, uint64, error), publish func(node, module string, body []byte, seq uint64) error,
+	say func(format string, args ...any)) int {
+	issued := 0
+	for node, declared := range r.Assigned {
+		for _, d := range declared {
+			composed := broker.MembershipFor(node, d, where).SeatTraffic
+			if composed == nil || !slices.ContainsFunc(composed.Kinds, func(k broker.KindHeld) bool {
+				return slices.Contains(k.Capabilities, "verified-sender")
+			}) {
+				continue
+			}
+			was, seq, err := last(node, d.Module)
+			if err != nil {
+				if !errors.Is(err, nats.ErrMsgNotFound) {
+					say("  %s on %s: its membership could not be read to give its verified sender back: %v", d.Module, node, err)
+				}
+				continue
+			}
+			body, changed, err := withVerifiedSenders(was, composed)
+			if err != nil {
+				say("  %s on %s: its membership could not be read to give its verified sender back: %v", d.Module, node, err)
+				continue
+			}
+			if !changed {
+				continue
+			}
+			if err := publish(node, d.Module, body, seq); err != nil {
+				// A push issued another since it was read: that one is newer, and the next self-check looks again.
+				if strings.Contains(err.Error(), "wrong last sequence") {
+					say("  %s on %s: a newer membership was issued meanwhile; its verified sender is looked at again next time",
+						d.Module, node)
+					continue
+				}
+				say("  %s on %s: its verified sender could not be issued again: %v", d.Module, node, err)
+				continue
+			}
+			say("  %s on %s: issued its membership again, with the verified sender its machines now pass for", d.Module, node)
+			issued++
+		}
+	}
+	return issued
+}
+
+// rootVerdictsOf is each machine's verdict reason, for saying why one was not free.
+func rootVerdictsOf(verdicts []rootVerdict) (free map[string]bool, whys map[string]string) {
+	free, whys = map[string]bool{}, map[string]string{}
+	for _, v := range verdicts {
+		if v.Free {
+			free[v.Machine] = true
+		}
+		whys[v.Machine] = strings.TrimSpace(v.Why)
+	}
+	return free, whys
+}
+
+// giveVerifiedSendersBack is the self-check's step: every machine judged, the memberships that lost a verified
+// sender to a failed judgement issued again. Never fails the probe: what could not be done is said.
+func giveVerifiedSendersBack(ctx context.Context, d *doctor) {
+	if d.js == nil || d.open == nil || d.open.inventory == nil {
+		return
+	}
+	records, err := d.open.inventory.BusRecords(ctx)
+	if err != nil {
+		fmt.Printf("  the memberships could not be read to give a verified sender back: %v\n", err)
+		return
+	}
+	// Judged for every machine the records name, as a push does, not only those the self-check judged.
+	r := newRootReader(ctx, d.open.inventory, d.js.Conn())
+	all := judgeRootFree(ctx, r, records.Nodes, time.Now())
+	records.RootFree, _ = rootVerdictsOf(all)
+	where := broker.PlacementsOf(records, records.Interchangeable)
+	reissueVerifiedSendersOn(ctx, link.OverNATS{Conn: d.js.Conn(), JS: d.js.Context()}, records, where)
+}
+
+// reissueVerifiedSendersOn is reissueVerifiedSenders on the bus: memberships read from the AssignmentsStream and
+// published to it.
+func reissueVerifiedSendersOn(ctx context.Context, bus link.OverNATS, records broker.Records, where broker.Placements) int {
+	return reissueVerifiedSenders(records, where,
+		func(node, module string) ([]byte, uint64, error) {
+			m, err := bus.JS.GetLastMsg(broker.AssignmentsStream, broker.MembershipSubject(node, module))
+			if err != nil {
+				return nil, 0, err
+			}
+			return m.Data, m.Sequence, nil
+		},
+		func(node, module string, body []byte, seq uint64) error {
+			return bus.PublishMembershipAfter(ctx, node, module, body, seq)
+		},
+		func(format string, args ...any) { fmt.Printf(format+"\n", args...) })
+}

cmd/mesh-controller/reissue_verified_sender_test.go

@@ -0,0 +1,187 @@
+package main
+
+import (
+	"encoding/json"
+	"errors"
+	"strings"
+	"testing"
+
+	"github.com/nats-io/nats.go"
+
+	"github.com/novox/mesh-controller/internal/broker"
+	"github.com/novox/mesh-controller/internal/link"
+	"github.com/novox/mesh-controller/internal/testbus"
+)
+
+func verifiedSenderRecords(free map[string]bool) (broker.Records, broker.Declared) {
+	tg := broker.Seat{Name: "channel", Scope: "mesh", Accepts: []string{"show"}, Kinded: true, Kind: "telegram",
+		DeclaredBy: "messenger", Capabilities: []string{"choice", "verified-sender"}}
+	router := broker.Declared{Module: "messenger", RunsAs: "messenger", Holds: []broker.Seat{{Name: "operator-channel",
+		Scope: "mesh", Accepts: []string{"ask"}}}, Uses: []broker.Seat{{Name: "channel", Scope: "mesh", Kinded: true,
+		DeclaredBy: "messenger"}}}
+	return broker.Records{Nodes: []string{"novox"}, RootFree: free, Assigned: map[string][]broker.Declared{
+		"novox": {router, {Module: "telegram", RunsAs: "telegram", Holds: []broker.Seat{tg}}}}}, router
+}
+
+type fakeBus struct {
+	last      map[string][]byte
+	published map[string][]byte
+	seq       map[string]uint64
+	// between runs once, after a read and before the write: a push issuing meanwhile.
+	between func()
+}
+
+func (f *fakeBus) get(node, module string) ([]byte, uint64, error) {
+	if b, ok := f.last[node+"/"+module]; ok {
+		return b, f.seq[node+"/"+module], nil
+	}
+	return nil, 0, nats.ErrMsgNotFound
+}
+func (f *fakeBus) put(node, module string, body []byte, seq uint64) error {
+	if f.between != nil {
+		f.between()
+		f.between = nil
+	}
+	if seq != f.seq[node+"/"+module] {
+		return errors.New("nats: wrong last sequence")
+	}
+	f.published[node+"/"+module] = body
+	f.last[node+"/"+module] = body
+	f.seq[node+"/"+module]++
+	return nil
+}
+
+func hasVerified(t *testing.T, membership []byte) bool {
+	t.Helper()
+	var m broker.Membership
+	if err := json.Unmarshal(membership, &m); err != nil {
+		t.Fatal(err)
+	}
+	for _, k := range m.SeatTraffic.Kinds {
+		if k.Kind == "telegram" {
+			return strings.Contains(strings.Join(k.Capabilities, ","), "verified-sender")
+		}
+	}
+	t.Fatal("telegram not in the router's membership")
+	return false
+}
+
+// A membership composed while the judgement failed is issued again once it passes, with the verified sender
+// (novox/hq issue 930); while it fails, nothing is given back (ADR 0259 §8).
+func TestAMembershipThatLostItsVerifiedSenderGetsItBackWhenTheJudgementPasses(t *testing.T) {
+	failed, router := verifiedSenderRecords(nil)
+	whereFailed := broker.PlacementsOf(failed, nil)
+	bus := &fakeBus{last: map[string][]byte{}, published: map[string][]byte{}, seq: map[string]uint64{}}
+	body, _ := json.Marshal(broker.MembershipFor("novox", router, whereFailed))
+	bus.last["novox/messenger"] = body
+	if hasVerified(t, body) {
+		t.Fatal("composed with a failed judgement, and the verified sender is kept")
+	}
+	say := func(string, ...any) {}
+
+	if n := reissueVerifiedSenders(failed, whereFailed, bus.get, bus.put, say); n != 0 || len(bus.published) != 0 {
+		t.Fatalf("issued %d while the judgement still fails", n)
+	}
+	passing, _ := verifiedSenderRecords(map[string]bool{"novox": true})
+	wherePassing := broker.PlacementsOf(passing, nil)
+	if n := reissueVerifiedSenders(passing, wherePassing, bus.get, bus.put, say); n != 1 {
+		t.Fatalf("issued %d memberships after the judgement passed, want 1", n)
+	}
+	if !hasVerified(t, bus.published["novox/messenger"]) {
+		t.Fatal("the membership issued again lacks the verified sender")
+	}
+	if n := reissueVerifiedSenders(passing, wherePassing, bus.get, bus.put, say); n != 0 {
+		t.Fatalf("issued %d again with nothing changed", n)
+	}
+}
+
+// What a composition drops it says, naming the kind, the machine and the judgement's reason.
+func TestADroppedVerifiedSenderIsSaidWithItsReason(t *testing.T) {
+	records, _ := verifiedSenderRecords(nil)
+	where := broker.PlacementsOf(records, nil)
+	lines := droppedVerifiedSenders(records, where, map[string]string{"novox": "not measured: no verb heard"})
+	if len(lines) != 1 || !strings.Contains(lines[0], "telegram") || !strings.Contains(lines[0], "novox") ||
+		!strings.Contains(lines[0], "not measured: no verb heard") {
+		t.Fatalf("said %q", lines)
+	}
+	records, _ = verifiedSenderRecords(map[string]bool{"novox": true})
+	if lines := droppedVerifiedSenders(records, broker.PlacementsOf(records, nil), nil); len(lines) != 0 {
+		t.Fatalf("said %q of a kind that kept its verified sender", lines)
+	}
+}
+
+func TestAMembershipNeverIssuedIsLeftToItsOwnSend(t *testing.T) {
+	records, _ := verifiedSenderRecords(map[string]bool{"novox": true})
+	bus := &fakeBus{last: map[string][]byte{}, published: map[string][]byte{}, seq: map[string]uint64{}}
+	var said []string
+	n := reissueVerifiedSenders(records, broker.PlacementsOf(records, nil), bus.get, bus.put,
+		func(f string, a ...any) { said = append(said, f) })
+	if n != 0 || len(said) != 0 {
+		t.Fatalf("issued %d, said %v", n, said)
+	}
+}
+
+// **On a real bus** (novox/hq issue 930): a router's membership issued without the verified sender because the
+// judgement failed is, once the judgement passes, published again to the AssignmentsStream with it in its seat
+// traffic's kinds, and a second pass publishes nothing.
+func TestNatsAMembershipIssuedWithoutItsVerifiedSenderIsIssuedAgainWhenTheJudgementPasses(t *testing.T) {
+	js, err := broker.Dial(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(js.Close)
+	if err := broker.AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	bus := link.OverNATS{Conn: js.Conn(), JS: js.Context()}
+
+	failed, router := verifiedSenderRecords(nil)
+	body, err := json.Marshal(broker.MembershipFor("novox", router, broker.PlacementsOf(failed, nil)))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := bus.PublishMembership(t.Context(), "novox", "messenger", body); err != nil {
+		t.Fatal(err)
+	}
+	issued := func() []byte {
+		m, err := js.Context().GetLastMsg(broker.AssignmentsStream, broker.MembershipSubject("novox", "messenger"))
+		if err != nil {
+			t.Fatal(err)
+		}
+		return m.Data
+	}
+	if hasVerified(t, issued()) {
+		t.Fatal("issued with a failed judgement, and the verified sender is kept: the test proves nothing")
+	}
+	if n := reissueVerifiedSendersOn(t.Context(), bus, failed, broker.PlacementsOf(failed, nil)); n != 0 {
+		t.Fatalf("issued %d memberships while the judgement still fails", n)
+	}
+	passing, _ := verifiedSenderRecords(map[string]bool{"novox": true})
+	where := broker.PlacementsOf(passing, nil)
+	if n := reissueVerifiedSendersOn(t.Context(), bus, passing, where); n != 1 {
+		t.Fatalf("issued %d memberships after the judgement passed, want 1", n)
+	}
+	if !hasVerified(t, issued()) {
+		t.Fatal("the membership on the bus lacks the verified sender after the judgement passed")
+	}
+	if n := reissueVerifiedSendersOn(t.Context(), bus, passing, where); n != 0 {
+		t.Fatalf("a second pass issued %d memberships", n)
+	}
+}
+
+// A push issuing a newer membership between the self-check's read and its write is not overwritten (issue 930).
+func TestANewerMembershipIssuedMeanwhileIsNotOverwritten(t *testing.T) {
+	failed, router := verifiedSenderRecords(nil)
+	bus := &fakeBus{last: map[string][]byte{}, published: map[string][]byte{}, seq: map[string]uint64{}}
+	old, _ := json.Marshal(broker.MembershipFor("novox", router, broker.PlacementsOf(failed, nil)))
+	bus.last["novox/messenger"], bus.seq["novox/messenger"] = old, 1
+	newer := []byte(`{"node":"novox","module":"messenger","newer":true}`)
+	bus.between = func() { bus.last["novox/messenger"], bus.seq["novox/messenger"] = newer, 2 }
+	passing, _ := verifiedSenderRecords(map[string]bool{"novox": true})
+	if n := reissueVerifiedSenders(passing, broker.PlacementsOf(passing, nil), bus.get, bus.put, func(string, ...any) {}); n != 0 {
+		t.Fatalf("issued %d over a newer membership", n)
+	}
+	if string(bus.last["novox/messenger"]) != string(newer) || len(bus.published) != 0 {
+		t.Fatalf("the newer membership was overwritten: %s", bus.last["novox/messenger"])
+	}
+}

cmd/mesh-controller/related.go

@@ -1,8 +1,10 @@
 package main
 
 import (
+	"context"
 	"sort"
 	"strings"
+	"time"
 
 	"github.com/novox/mesh-controller/internal/inventory"
 )
@@ -11,7 +13,9 @@ import (
 // related when they share a repository, a module, a delivery group's head, or an edge of the module graph
 // between their modules. A module is every one the walk builds or sends: the modules its merges moved and the
 // modules that depend on them, not only the ones the merges touched. Unrelated changes may walk beside each
-// other (ADR 0304 decision 2, ADR 0282 decision 3); nothing walks so yet, and nothing but tests reads this.
+// other (ADR 0304 decision 2, ADR 0282 decision 3); a closed batch is cut beside an open walk it is
+// not related to, and a walk about to ask its first tier waits only for a started walk it is related to (issue 973,
+// ADR 0407 decision 4).
 
 // footprint is what one change reaches: its repositories (lower case), its modules, and its delivery group
 // heads (the head branch names of its merges). unknown says a merge of it was kept without the modules it
@@ -109,3 +113,83 @@ func firstShared(a, b map[string]bool) string {
 	sort.Strings(both)
 	return both[0]
 }
+
+// openWalk is a walk open beside a change: its id, what it reaches, and whether it is on the controller's own path.
+type openWalk struct {
+	id  string
+	fp  footprint
+	own bool
+}
+
+// relatedOpen is the first open walk a change is related to, with the reason; empty when it is related to none
+// and may walk beside them all (issue 973, ADR 0407 decision 4, ADR 0306 decision 3).
+//
+// **The controller's own path is related to everything** (decided for issue 973): a walk or batch that replaces
+// the controller, the node-engine, the tool runner, the bus or mesh-delivery never runs beside another, and
+// nothing runs beside it, whatever the footprints say.
+func relatedOpen(cand footprint, candOwn bool, open []openWalk, edges []inventory.Edge) (behind, why string) {
+	for _, w := range open {
+		if candOwn {
+			return w.id, "it is on the controller's own path"
+		}
+		if w.own {
+			return w.id, w.id + " is on the controller's own path"
+		}
+		if ok, why := related(cand, w.fp, edges); ok {
+			return w.id, why
+		}
+	}
+	return "", ""
+}
+
+// ownWalk says a walk is on the controller's own path: its batch says so, or it plans a module of that path
+// (issue 973).
+func ownWalk(p inventory.Plan) bool {
+	if p.OwnPath() {
+		return true
+	}
+	for name := range p.Modules {
+		if _, own := onTheControllersPath[name]; own {
+			return true
+		}
+	}
+	return false
+}
+
+// startedWalks are the open walks started beside the walk id, made at created: a merge's walk past its wait, not
+// the backlog's, with a tier asked, or on its way (let go, or waiting for nobody) before this one.
+func startedWalks(plans []inventory.Plan, id string, created time.Time) []inventory.Plan {
+	var out []inventory.Plan
+	for _, q := range plans {
+		if q.ID == id || q.Release != nil || q.Waiting() {
+			continue
+		}
+		if q.Tier > 0 || askedAny(q) || q.Created.Before(created) {
+			out = append(out, q)
+		}
+	}
+	return out
+}
+
+// openWalkOf is an open walk as a change is judged against it: what it reaches, from its planned modules and the
+// merges the record holds for it.
+func openWalkOf(ctx context.Context, inv *inventory.Inventory, p inventory.Plan) (openWalk, error) {
+	kept, err := inv.MergesOf(ctx, p.ID)
+	if err != nil {
+		return openWalk{}, err
+	}
+	return openWalk{id: p.ID, fp: footprintOfWalk(p, kept), own: ownWalk(p)}, nil
+}
+
+// openWalksOf are openWalkOf for each of plans.
+func openWalksOf(ctx context.Context, inv *inventory.Inventory, plans []inventory.Plan) ([]openWalk, error) {
+	var out []openWalk
+	for _, p := range plans {
+		w, err := openWalkOf(ctx, inv, p)
+		if err != nil {
+			return nil, err
+		}
+		out = append(out, w)
+	}
+	return out, nil
+}

cmd/mesh-controller/related_test.go

@@ -2,8 +2,11 @@ package main
 
 import (
 	"encoding/json"
+	"errors"
+	"fmt"
 	"strings"
 	"testing"
+	"time"
 
 	"github.com/novox/mesh-controller/internal/inventory"
 	"github.com/novox/mesh-controller/internal/link"
@@ -108,3 +111,125 @@ func TestABatchWhoseMovesAreNotKnownIsRelatedToEverything(t *testing.T) {
 		t.Fatalf("a known, unrelated batch read as related: %q", why)
 	}
 }
+
+// TestRelatedOpenNamesTheFirstOpenWalkAChangeIsRelatedTo: a change that shares nothing with any open walk is
+// held behind none (issue 973, ADR 0407 decision 4); one that shares a repository, a module, a graph edge or a
+// group head is held behind the first walk it shares it with, with the reason; one whose moves are not known
+// is held behind the first; a walk or change on the controller's own path is related to every other.
+func TestRelatedOpenNamesTheFirstOpenWalkAChangeIsRelatedTo(t *testing.T) {
+	edges := []inventory.Edge{{From: "nzbget", To: "base-image", Kind: inventory.EdgeStandsOn}}
+	walk := func(id, repo string, own bool, modules ...string) openWalk {
+		return openWalk{id: id, fp: footprintOfWalk(walkOf(repo, modules...), nil), own: own}
+	}
+	one, two := walk("plan-1", "novox/one", false, "notes"), walk("plan-2", "novox/two", false, "nzbget")
+	cand := func(repo, head string, moves ...string) footprint {
+		return footprintOfBatch([]inventory.BatchedMerge{heardMerge(repo, head, moves...)}, edges)
+	}
+	cases := []struct {
+		name   string
+		fp     footprint
+		own    bool
+		open   []openWalk
+		behind string
+		why    string
+	}{
+		{"nothing open", cand("novox/x", "fix/a", "app"), false, nil, "", ""},
+		{"unrelated to every open walk", cand("novox/x", "fix/a", "app"), false, []openWalk{one, two}, "", ""},
+		{"a repository", cand("novox/ONE", "fix/a", "app"), false, []openWalk{two, one}, "plan-1", "both change novox/one"},
+		{"a module through its dependents", cand("novox/x", "fix/a", "base-image"), false, []openWalk{one, two},
+			"plan-2", "both move nzbget"},
+		{"a graph edge", cand("novox/x", "fix/a", "nzbget"),
+			false, []openWalk{one, walk("plan-3", "novox/y", false, "base-image")}, "plan-3",
+			"nzbget depends on base-image (stands-on)"},
+		{"moves not known", cand("novox/x", "fix/a"), false, []openWalk{one, two}, "plan-1",
+			"the modules a merge of novox/x moves are not known until its batch is cut"},
+		{"the candidate on the own path", cand("novox/x", "fix/a", "app"), true, []openWalk{one}, "plan-1",
+			"it is on the controller's own path"},
+		{"an open walk on the own path", cand("novox/x", "fix/a", "app"), false,
+			[]openWalk{one, walk("plan-4", "novox/mesh-controller", true, "mesh-controller")}, "plan-4",
+			"plan-4 is on the controller's own path"},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			behind, why := relatedOpen(c.fp, c.own, c.open, edges)
+			if behind != c.behind || why != c.why {
+				t.Errorf("relatedOpen = %q, %q; want %q, %q", behind, why, c.behind, c.why)
+			}
+		})
+	}
+}
+
+// A walk is on the controller's own path when its batch says so or it plans a module of that path (issue 973).
+func TestAWalkOnTheControllersPathIsKnownByItsFlagOrItsModules(t *testing.T) {
+	if ownWalk(walkOf("novox/one", "notes")) {
+		t.Error("an ordinary walk read as on the controller's own path")
+	}
+	flagged := walkOf("novox/one", "notes")
+	flagged.Delivery = &inventory.PlanDelivery{Batch: &inventory.PlanBatch{Own: true}}
+	if !ownWalk(flagged) {
+		t.Error("a walk whose batch is own-path read as ordinary")
+	}
+	var module string
+	for m := range onTheControllersPath {
+		module = m
+	}
+	if !ownWalk(walkOf("novox/one", "notes", module)) {
+		t.Errorf("a walk planning %s read as ordinary", module)
+	}
+}
+
+// A walk about to ask its first tier waits only for a started walk it is related to (issue 973): started is a
+// tier asked, or a walk made before it; a waiting walk, the backlog's and itself are not.
+func TestWalksStartedBesideAreTheOpenOnesBeforeIt(t *testing.T) {
+	at := time.Date(2026, 10, 17, 12, 0, 0, 0, time.UTC)
+	mk := func(id string, created time.Time, tier int) inventory.Plan {
+		p := walkOf("novox/"+id, id)
+		p.ID, p.Created, p.Tier, p.State = id, created, tier, inventory.PlanBuilding
+		return p
+	}
+	older, newer, ticking := mk("older", at, 0), mk("newer", at.Add(time.Minute), 0), mk("ticking", at.Add(2*time.Minute), 1)
+	backlog := mk("backlog", at, 1)
+	backlog.Release = &inventory.PlanRelease{}
+	got := startedWalks([]inventory.Plan{older, newer, ticking, backlog, mk("me", at.Add(time.Minute), 0)}, "me", at.Add(time.Minute))
+	var ids []string
+	for _, p := range got {
+		ids = append(ids, p.ID)
+	}
+	if strings.Join(ids, ",") != "older,ticking" {
+		t.Errorf("started beside me = %v; want older, ticking", ids)
+	}
+}
+
+// A retry is refused beside an open walk it is related to and allowed beside one it shares nothing with
+// (issue 973); a retry on the controller's own path is refused beside any open walk.
+func TestARetryWaitsOnlyForAnOpenWalkItIsRelatedTo(t *testing.T) {
+	failed := walkOf("novox/one", "notes")
+	failed.ID, failed.State = "plan-1", inventory.PlanFailed
+	open := func(id, repo string, modules ...string) inventory.Plan {
+		p := walkOf(repo, modules...)
+		p.ID, p.State = id, inventory.PlanBuilding
+		return p
+	}
+	if err := oneWalkAtATime(failed, []inventory.Plan{open("plan-2", "novox/two", "app")}); err != nil {
+		t.Errorf("refused beside an unrelated walk: %v", err)
+	}
+	err := oneWalkAtATime(failed, []inventory.Plan{open("plan-3", "novox/three", "notes")})
+	if err == nil || !strings.Contains(err.Error(), "plan-3") || !strings.Contains(err.Error(), "both move notes") {
+		t.Errorf("a retry beside a walk moving the same module: %v", err)
+	}
+	failed.Delivery = &inventory.PlanDelivery{Batch: &inventory.PlanBatch{Own: true}}
+	if err := oneWalkAtATime(failed, []inventory.Plan{open("plan-2", "novox/two", "app")}); err == nil {
+		t.Error("an own-path retry went beside an open walk")
+	}
+}
+
+// Plans another controller is working are "not now", not a failure: the merge is held and handed over again
+// (issue 921).
+func TestPlansHeldByAnotherControllerAreNotNow(t *testing.T) {
+	if err := notNow(fmt.Errorf("x: %w", inventory.ErrPlansBusy)); !errors.Is(err, link.ErrTryAgain) {
+		t.Errorf("busy plans read as %v; want ErrTryAgain", err)
+	}
+	if err := notNow(errors.New("no such module")); errors.Is(err, link.ErrTryAgain) {
+		t.Errorf("another failure read as ErrTryAgain: %v", err)
+	}
+}

cmd/mesh-controller/release_plan.go

@@ -618,23 +618,32 @@ func planBuiltSaying(ctx context.Context, open *stores, module, commit, failed s
 	advanceHeld(ctx, open)
 }
 
-// startedBeside is the id of a started walk open beside this one — a merge's walk past its wait, not the
+// startedBeside is the id of a started walk open beside this one that it is related to — a merge's walk past
-// backlog's — or empty: one walk at a time (novox/hq ADR 0276).
+// its wait, not the backlog's — with why, or empty: a walk waits only for a started walk it shares something with
-func startedBeside(ctx context.Context, inv *inventory.Inventory, id string, created time.Time) (string, error) {
+// (novox/hq ADR 0276 decision 4, ADR 0407 decision 4, issue 973).
+func startedBeside(ctx context.Context, inv *inventory.Inventory, p inventory.Plan) (behind, why string, err error) {
 	plans, err := inv.OpenPlans(ctx)
 	if err != nil {
-		return "", err
+		return "", "", err
 	}
-	for _, q := range plans {
+	started := startedWalks(plans, p.ID, p.Created)
-		if q.ID == id || q.Release != nil || q.Waiting() {
+	if len(started) == 0 {
-			continue
+		return "", "", nil
-		}
-		// Started: a tier asked, or on its way (let go, or waiting for nobody) before this one.
-		if q.Tier > 0 || askedAny(q) || q.Created.Before(created) {
-			return q.ID, nil
-		}
 	}
-	return "", nil
+	edges, err := inv.Dependencies(ctx)
+	if err != nil {
+		return "", "", err
+	}
+	open, err := openWalksOf(ctx, inv, started)
+	if err != nil {
+		return "", "", err
+	}
+	mine, err := openWalkOf(ctx, inv, p)
+	if err != nil {
+		return "", "", err
+	}
+	behind, why = relatedOpen(mine.fp, mine.own, open, edges)
+	return behind, why, nil
 }
 
 // deferredNote begins the note of a walk deferred behind another.
@@ -693,7 +702,7 @@ func keepWalkPhases(ctx context.Context, open *stores) {
 // advanceHeld is advancePlans for a caller already holding the plans.
 func advanceHeld(ctx context.Context, open *stores) {
 	inv := open.inventory
-	// A walk that ended lets the next batch be cut at once, not at the cutter's next look (novox/hq ADR 0276).
+	// A walk that ended lets the next batch be cut at once, not at a window's next expiry (novox/hq ADR 0276, ADR 0407).
 	defer func() {
 		if err := cutBatchesHeld(ctx, open, time.Now().UTC()); err != nil {
 			fmt.Printf("batches: %v\n", err)
@@ -793,13 +802,14 @@ func advanceOnce(ctx context.Context, open *stores, p *inventory.Plan,
 	tier := p.Tiers[p.Tier]
 	// Not yet asked: ask — the whole tier, or what of it is still unasked (novox/hq issue 754).
 	if unasked := unaskedOf(*p); len(unasked) > 0 {
-		// **One walk at a time** (novox/hq ADR 0276): a walk about to ask its first tier while another started
+		// **One walk at a time beside a related one** (novox/hq ADR 0276, ADR 0407 decision 4, issue 973): a walk
-		// walk is open waits for that one to end, let go or not, and says so.
+		// about to ask its first tier while a started walk it is related to is open waits for that one to end,
+		// let go or not, and says so and why. A started walk it shares nothing with does not hold it.
 		if p.Tier == 0 && !askedInTier(*p) {
-			if behind, err := startedBeside(ctx, inv, p.ID, p.Created); err != nil {
+			if behind, why, err := startedBeside(ctx, inv, *p); err != nil {
 				return false, err
 			} else if behind != "" {
-				note := fmt.Sprintf("%s%s ended — one walk at a time", deferredNote, behind)
+				note := fmt.Sprintf("%s%s ended — %s", deferredNote, behind, why)
 				changed := p.Note != note
 				p.Note = note
 				return changed, nil
@@ -1722,7 +1732,7 @@ func planLineSaid(p inventory.Plan, now time.Time, pause pauseView, bound time.D
 		return fmt.Sprintf("%s · %s, %s, for %s", name, where, waitingSaid(p, waits[p.ID]), since)
 	}
 	if deferred(p) {
-		// Nor is waiting for the walk before it to end (one walk at a time, novox/hq ADR 0276).
+		// Nor is waiting for the walk before it to end (a related walk, novox/hq ADR 0276, ADR 0407).
 		return fmt.Sprintf("%s · %s, %s, for %s", name, where, p.Note, since)
 	}
 	if waiting, paused := pausedWaiting(p, pause, now); paused {

cmd/mesh-controller/signals.go

@@ -126,11 +126,6 @@ var signalsTable = []signalRow{
 		Kind: "controller-deaf", Severity: conditions.Urgent, Phase: 1,
 		needs: func(f *signalFacts) error { return f.loopErr }, watch: watchLoop,
 		newest: func(f *signalFacts) time.Time { return f.loop.took }},
-	{Row: "S5", Signal: "a merge announced becomes a plan, or nothing reads it", Emitter: "announcer → controller",
-		Trigger: "each merge", Bound: "10 min (the catch-up pass of issue 266)",
-		Kind: "merge-not-acted", Severity: conditions.Urgent, Phase: 1,
-		needs: func(f *signalFacts) error { return f.mergesErr }, watch: watchMerges,
-		newest: func(f *signalFacts) time.Time { return f.mergesPassed }},
 	{Row: "S6", Signal: "a build asked → its outcome", Emitter: "build seat", Trigger: "each ask",
 		Bound: "max(20 min, 3 × the p90 of measured builds), 1 h while nothing is measured, provisional — the " +
 			"build seat declares no timeout; and an ask the queue gave up on (dead) at once",
@@ -459,18 +454,6 @@ func watchLoop(f *signalFacts) []conditions.Observation {
 		Said: fmt.Sprintf("%d pending (%s), last taken %s", f.loop.pending, f.loop.where, since.UTC().Format(time.RFC3339))}}
 }
 
-func watchMerges(f *signalFacts) []conditions.Observation {
-	var out []conditions.Observation
-	for _, m := range f.merges {
-		out = append(out, conditions.Observation{Scope: conditions.ScopeMerge, ID: m.Repo + "." + short(m.Commit),
-			Kind: "merge-not-acted", Token: "not-acted", Severity: conditions.Urgent,
-			Summary: fmt.Sprintf("%s/%s merged into %s (%s) was never handed to the controller by the bus; "+
-				"acted on late by the catch-up", m.Owner, m.Repo, m.Base, short(m.Commit)),
-			Said: fmt.Sprintf("announced %s, %s behind it", m.At.UTC().Format(time.RFC3339), readableList(m.Modules))})
-	}
-	return out
-}
-
 func watchAsks(f *signalFacts) []conditions.Observation {
 	var out []conditions.Observation
 	for _, a := range f.asks {
@@ -686,7 +669,7 @@ func watchDeadLetters(f *signalFacts) []conditions.Observation {
 // controller, or a seat's holders.
 func consumerWho(stream, consumer string) string {
 	switch {
-	case consumer == broker.ControllerName:
+	case consumer == broker.ControllerName || consumer == broker.MergesConsumerName:
 		return "the controller"
 	case strings.HasPrefix(stream, "SEAT_") && strings.HasSuffix(consumer, "_worker"):
 		seat := strings.ToLower(strings.ReplaceAll(strings.TrimSuffix(strings.TrimPrefix(consumer, "SEAT_"), "_worker"), "_", "-"))

cmd/mesh-controller/signals_test.go

@@ -33,7 +33,6 @@ func calm(now time.Time) *signalFacts {
 		plans: []planFacts{{id: "plan-1", repository: "novox/app", commit: "c0ffee00", tier: 0, tiers: 2,
 			entered: now.Add(-time.Minute), bound: 30 * time.Minute, waiting: "building"}},
 		loop:          loopFacts{took: now.Add(-time.Second), pending: 1},
-		mergesPassed:  now.Add(-time.Minute),
 		selfCheck:     selfCheckFacts{last: now.Add(-time.Minute), every: 5 * time.Minute},
 		lostConsumers: map[string]bool{}, epochs: map[int64]inventory.Epoch{},
 		lease: leaseFacts{held: true, epoch: 57, renewed: now.Add(-2 * time.Second)},
@@ -72,12 +71,6 @@ var suppressions = map[string]suppression{
 		inside: func(f *signalFacts) { f.loop.took = f.now.Add(-119 * time.Second) },
 		past:   func(f *signalFacts) { f.loop.took = f.now.Add(-121 * time.Second) },
 	},
-	"S5": {
-		inside: func(f *signalFacts) {},
-		past: func(f *signalFacts) {
-			f.merges = []missedMerge{{Owner: "novox", Repo: "app", Base: "main", Commit: "c0ffee0011", At: f.now.Add(-11 * time.Minute)}}
-		},
-	},
 	"S6": {
 		inside: func(f *signalFacts) {
 			f.asks = []askFacts{{id: "build-1", seat: "node-build-agent", what: "novox/app", state: link.AskInFlight,

cmd/mesh-controller/upgrades.go

@@ -294,10 +294,10 @@ func (f following) Announceable(ctx context.Context) ([]link.Announcement, error
 	return out, nil
 }
 
-// notNow marks a store that could not be read right now, so the announcement is held rather than
+// notNow marks a store that could not be read right now, or plans another controller is working (issue 921), so
-// lost; anything else is returned as it was.
+// the announcement is held rather than lost; anything else is returned as it was.
 func notNow(err error) error {
-	if inventory.Unreachable(err) {
+	if inventory.Unreachable(err) || errors.Is(err, inventory.ErrPlansBusy) {
 		return fmt.Errorf("%w: %w", link.ErrTryAgain, err)
 	}
 	return err

cmd/mesh-controller/watchdogs.go

@@ -70,10 +70,6 @@ type signalFacts struct {
 	loop    loopFacts
 	loopErr error
 
-	merges       []missedMerge
-	mergesPassed time.Time
-	mergesErr    error
-
 	asks    []askFacts
 	asksErr error
 
@@ -358,10 +354,6 @@ func (w *watchdogs) gather(ctx context.Context) *signalFacts {
 		f.batches, f.plansErr = gatherBatches(ctx, inv, now)
 	}
 	f.loop, f.loopErr = w.gatherLoop()
-	f.mergesPassed, f.merges, f.mergesErr = watchedMerges.last()
-	if f.mergesErr == nil && !f.mergesPassed.IsZero() && now.Sub(f.mergesPassed) > 3*mergeCatchUpEvery {
-		f.mergesErr = fmt.Errorf("the catch-up of merges has not passed since %s", f.mergesPassed.UTC().Format(time.RFC3339))
-	}
 	f.asks, f.asksErr = w.gatherAsks(ctx, inv)
 	if open, err := w.keeper.Open(ctx); err != nil {
 		f.standingsErr = err

internal/broker/config_owner_live_test.go

@@ -160,3 +160,87 @@ func TestNatsOnlyMeshConfigsAccountPutsToItsBuckets(t *testing.T) {
 		}
 	}
 }
+
+// **The controller asks mesh-config's `config` in its own name** (issue 925): on a real server whose users are
+// composed from records that hold mesh-config on novox serving `config`, the controller's request on
+// `mesh.seat.mesh-config.call.config.controller` reaches mesh-config's account and its reply reaches the
+// controller's inbox; the same request under another caller's name is refused by the server.
+func TestNatsTheControllerAsksMeshConfigInItsOwnName(t *testing.T) {
+	hash, err := bcrypt.GenerateFromPassword([]byte("pw"), bcrypt.MinCost)
+	if err != nil {
+		t.Fatal(err)
+	}
+	seat := Seat{Name: "mesh-config", Scope: "mesh", Serves: []string{"config"}}
+	users, err := Users(Records{Nodes: []string{"novox"}, Assigned: map[string][]Declared{
+		"novox": {{Module: RuntimeModule}, {Module: "mesh-config", RunsAs: "mesh-config", Holds: []Seat{seat}}},
+	}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for i := range users {
+		users[i].PasswordHash = string(hash)
+	}
+	accounts, err := ComposeAccounts(users)
+	if err != nil {
+		t.Fatal(err)
+	}
+	conf := filepath.Join(t.TempDir(), "bus.conf")
+	if err := os.WriteFile(conf, []byte("listen: 127.0.0.1:-1\n"+accounts), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	opts, err := server.ProcessConfigFile(conf)
+	if err != nil {
+		t.Fatalf("the composed accounts do not parse: %v", err)
+	}
+	opts.NoLog, opts.NoSigs = true, true
+	s, err := server.NewServer(opts)
+	if err != nil {
+		t.Fatal(err)
+	}
+	go s.Start()
+	if !s.ReadyForConnections(10 * time.Second) {
+		t.Fatal("the bus did not come up")
+	}
+	t.Cleanup(s.Shutdown)
+	connect := func(user string) (*nats.Conn, chan error) {
+		t.Helper()
+		refused := make(chan error, 8)
+		nc, err := nats.Connect(s.ClientURL(), nats.UserInfo(user, "pw"), nats.CustomInboxPrefix("_INBOX."+user),
+			nats.ErrorHandler(func(_ *nats.Conn, _ *nats.Subscription, err error) { refused <- err }))
+		if err != nil {
+			t.Fatalf("%s cannot connect: %v", user, err)
+		}
+		t.Cleanup(nc.Close)
+		return nc, refused
+	}
+	owner, _ := connect("novox.mesh-config")
+	subject := CalledSubject("mesh.seat.mesh-config.tool.config", ControllerName)
+	if subject != "mesh.seat.mesh-config.call.config.controller" {
+		t.Fatalf("the subject is %s", subject)
+	}
+	if _, err := owner.Subscribe("mesh.seat.mesh-config.call.config.*", func(m *nats.Msg) {
+		_ = m.Respond([]byte(`{"ok":true}`))
+	}); err != nil {
+		t.Fatal(err)
+	}
+	if err := owner.Flush(); err != nil {
+		t.Fatal(err)
+	}
+	controller, refused := connect(ControllerName)
+	reply, err := controller.Request(subject, []byte(`{"action":"set"}`), 5*time.Second)
+	if err != nil || string(reply.Data) != `{"ok":true}` {
+		t.Fatalf("the controller's own-name request got %v (%v), want mesh-config's reply", reply, err)
+	}
+	if err := controller.Publish(CalledSubject("mesh.seat.mesh-config.tool.config", "anchor.node-tools"), nil); err != nil {
+		t.Fatal(err)
+	}
+	_ = controller.Flush()
+	select {
+	case err := <-refused:
+		if !errors.Is(err, nats.ErrPermissionViolation) {
+			t.Errorf("the server said %v, want a permissions violation", err)
+		}
+	case <-time.After(3 * time.Second):
+		t.Error("the server did not refuse the controller a request in another's name")
+	}
+}

internal/broker/controller_buckets.go

@@ -38,6 +38,11 @@ var (
 	// each ask by its id, its options and the actions they stand for, how it ended and whether the
 	// controller acted on its warrant — so a restart neither asks twice nor acts twice.
 	AskedBucket = BucketName(ControllerSeat, "asked")
+	// MergeWindowBucket holds the clock of every open batch's merge window (novox/hq ADR 0407 decision 1): per
+	// batch a key `<batch>.quiet`, made again by each merge into it and living the window, and a key
+	// `<batch>.most`, made once when it opens and living the window's most. The first to expire closes the
+	// batch, and the controller's watcher hears the delete marker. Written by the lease holder alone.
+	MergeWindowBucket = BucketName(ControllerSeat, "merge-window")
 )
 
 // The controller's record mirrors (novox/hq ADR 0289, design 54, design 57 package F3): a bucket per kind of
@@ -79,6 +84,10 @@ const (
 	MirrorMarkerTTL = 24 * time.Hour
 )
 
+// MergeWindowMarkerTTL is how long the marker an expired window key leaves is kept: long enough for a watcher
+// that is up to hear it, short enough to leave nothing behind.
+const MergeWindowMarkerTTL = time.Minute
+
 // AskedKeptFor is how long an ask is kept after it was made: a month, as the router keeps its own.
 const AskedKeptFor = 30 * 24 * time.Hour
 
@@ -106,13 +115,14 @@ const (
 func IsControllerBucket(bucket string) bool {
 	return bucket == CallsBucket || bucket == HandActsBucket || bucket == ConditionsBucket ||
 		bucket == ConditionHistoryBucket || bucket == LeaseBucket || bucket == AskedBucket ||
-		bucket == AssignmentsBucket || bucket == DeclaredConfigBucket || bucket == PredefinedBucket
+		bucket == AssignmentsBucket || bucket == DeclaredConfigBucket || bucket == PredefinedBucket ||
+		bucket == MergeWindowBucket
 }
 
 // ControllerBuckets are the controller's own buckets, in the order they are asserted.
 func ControllerBuckets() []string {
 	return []string{LeaseBucket, CallsBucket, HandActsBucket, ConditionsBucket, ConditionHistoryBucket, AskedBucket,
-		AssignmentsBucket, DeclaredConfigBucket, PredefinedBucket}
+		AssignmentsBucket, DeclaredConfigBucket, PredefinedBucket, MergeWindowBucket}
 }
 
 // ControllerBucketsAsserter is what raising the controller's buckets needs of a connection.
@@ -209,6 +219,21 @@ func (j *JetStream) EnsureControllerBuckets() error {
 	}); err != nil {
 		return fmt.Errorf("asserting bucket %s: %w", AskedBucket, err)
 	}
+	// **The merge window: keys that expire.** A key's own time to live is the window; the marker an expired key
+	// leaves is what a watcher hears, kept a minute (ADR 0407 decision 1).
+	if _, err := js.CreateOrUpdateKeyValue(ctx, jetstream.KeyValueConfig{
+		Bucket: MergeWindowBucket,
+		Description: "the clock of every open batch's merge window (novox/hq ADR 0407): `<batch>.quiet` lives the " +
+			"window and is made again by each merge, `<batch>.most` lives the most; the first to expire closes the " +
+			"batch; written by the controller alone",
+		History:        1,
+		MaxValueSize:   4 << 10,
+		MaxBytes:       8 << 20,
+		Storage:        jetstream.FileStorage,
+		LimitMarkerTTL: MergeWindowMarkerTTL,
+	}); err != nil {
+		return fmt.Errorf("asserting bucket %s: %w", MergeWindowBucket, err)
+	}
 	// **The record mirrors: no age** (design 54 §5). A key is removed when its record is, by a purge whose
 	// marker lives a day; a key that expired would be a record the bus forgot while the store still holds it.
 	for _, m := range []struct{ bucket, why string }{

internal/broker/controller_buckets_test.go

@@ -2,6 +2,7 @@ package broker
 
 import (
 	"context"
+	"github.com/nats-io/nats.go"
 	"slices"
 	"strings"
 	"testing"
@@ -226,3 +227,175 @@ func TestTheControllerMayAskTheVaultInItsOwnName(t *testing.T) {
 		}
 	}
 }
+
+// **The controller may ask mesh-config's `config` in its own name** (issue 925): move-config and the settings verb
+// set a trusted value through mesh-config, which takes it as the controller carries it only from the subject that
+// names the controller as caller. Granted when the records hold a holder serving the verb, and not otherwise;
+// the installer's first user list carries none.
+func TestTheControllerMayAskMeshConfigInItsOwnName(t *testing.T) {
+	config := Declared{Module: "mesh-config", Holds: []Seat{{Name: "mesh-config", Serves: []string{"config"}}}}
+	users, err := Users(Records{Nodes: []string{"novox"}, Assigned: map[string][]Declared{"novox": {config}}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if strings.Join(users[0].AsksTheConfig, ",") != "config" {
+		t.Fatalf("composed %v", users[0].AsksTheConfig)
+	}
+	p, err := PermissionsFor(users[0])
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !slices.Contains(p.Publish, "mesh.seat.mesh-config.call.config.controller") {
+		t.Errorf("the controller may not publish mesh.seat.mesh-config.call.config.controller: %v", p.Publish)
+	}
+	if !MayPublish(p, CalledSubject("mesh.seat.mesh-config.tool.config", ControllerName)) {
+		t.Error("the controller may not ask mesh-config's config in its own name")
+	}
+	if MayPublish(p, CalledSubject("mesh.seat.mesh-config.tool.config", "anchor.node-tools")) {
+		t.Error("the controller may ask mesh-config's config in another's name")
+	}
+	// No holder serving it composes no grant, nor does genesis.
+	other := Declared{Module: "mesh-config", Holds: []Seat{{Name: "mesh-config", Serves: []string{"list"}}}}
+	for _, r := range []Records{{Nodes: []string{"novox"}}, {Nodes: []string{"novox"}, Assigned: map[string][]Declared{"novox": {other}}}} {
+		users, _ = Users(r)
+		if len(users[0].AsksTheConfig) != 0 {
+			t.Errorf("composed %v without a holder serving config", users[0].AsksTheConfig)
+		}
+	}
+	genesis, _ := PermissionsFor(Principal{Kind: KindController, PasswordHash: "x"})
+	for _, s := range genesis.Publish {
+		if strings.Contains(s, "mesh-config") {
+			t.Errorf("the installer's controller may publish %s", s)
+		}
+	}
+}
+
+// **The merge window is a bucket whose keys expire** (novox/hq ADR 0407 decision 1, issue 973): a key made with a
+// time to live goes when it is up, and a watcher on the bucket hears it go. The controller may write it, and it is
+// no undeclared state.
+func TestTheMergeWindowsKeysExpireAndAWatcherHearsIt(t *testing.T) {
+	controller, err := PermissionsFor(Principal{Kind: KindController, PasswordHash: "x"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !slices.Contains(controller.Publish, "$KV."+MergeWindowBucket+".>") {
+		t.Error("the controller may not write the merge window's bucket")
+	}
+	if !IsControllerBucket(MergeWindowBucket) || !slices.Contains(ControllerBuckets(), MergeWindowBucket) {
+		t.Error("the merge window's bucket is not asserted as the controller's own")
+	}
+	js, err := Dial(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer js.Close()
+	if err := js.EnsureControllerBuckets(); err != nil {
+		t.Fatal(err)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Second)
+	defer cancel()
+	api, err := jetstream.New(js.Conn())
+	if err != nil {
+		t.Fatal(err)
+	}
+	kv, err := api.KeyValue(ctx, MergeWindowBucket)
+	if err != nil {
+		t.Fatal(err)
+	}
+	w, err := kv.WatchAll(ctx)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer w.Stop()
+	if _, err := kv.Create(ctx, "plan-1.quiet", []byte("x"), jetstream.KeyTTL(time.Second)); err != nil {
+		t.Fatal(err)
+	}
+	for {
+		select {
+		case e := <-w.Updates():
+			if e != nil && e.Key() == "plan-1.quiet" && e.Operation() != jetstream.KeyValuePut {
+				return
+			}
+		case <-ctx.Done():
+			t.Fatal("no watcher heard the key expire")
+		}
+	}
+}
+
+// **The merges have a consumer of their own, which a reset of the events consumer leaves alone** (issue 973, novox/hq
+// ADR 0407 decision 3): healer H4 re-makes the events consumer from now, dropping what it holds pending; a merge
+// waiting in the merges consumer is not that consumer's, and nothing would catch it up.
+func TestAResetOfTheEventsConsumerLeavesTheMergesAndTheirConsumerAlone(t *testing.T) {
+	js, err := Dial(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer js.Close()
+	if err := AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	if err := AssertMeshConsumers(js); err != nil {
+		t.Fatal(err)
+	}
+	for _, c := range MeshConsumers() {
+		if c.Stream == EventsStream && c.Name == MergesConsumerName {
+			if c.Resettable != "" || c.MaxDeliver != 0 || len(c.Filters) != 1 || c.Filters[0] != ControllerFollows[3] {
+				t.Errorf("the merges consumer is %+v; want one filter, no max-deliver, not resettable", c)
+			}
+		}
+		if c.Stream == EventsStream && c.Name == ControllerName {
+			for _, f := range c.Filters {
+				if f == ControllerFollows[3] {
+					t.Errorf("the events consumer still follows the merges, which the merges consumer takes")
+				}
+			}
+		}
+	}
+	if _, err := js.Context().Publish(ControllerFollows[3], []byte(`{"commit":"c1"}`)); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := js.Context().Publish(ControllerFollows[0], []byte(`{}`)); err != nil {
+		t.Fatal(err)
+	}
+	merges, err := js.Context().ConsumerInfo(EventsStream, MergesConsumerName)
+	if err != nil || merges.NumPending != 1 {
+		t.Fatalf("the merges consumer holds %v (%v); want the merge", merges, err)
+	}
+	if _, _, err := js.ResetConsumer(EventsStream, ControllerName); err != nil {
+		t.Fatal(err)
+	}
+	merges, err = js.Context().ConsumerInfo(EventsStream, MergesConsumerName)
+	if err != nil || merges.NumPending != 1 || merges.Delivered.Stream != 0 {
+		t.Fatalf("after the reset the merges consumer holds %+v (%v); want its merge left alone", merges, err)
+	}
+}
+
+// **Delivering this change loses no merge** (review of issue 973): a merge published under the old filter of the
+// `controller` consumer and never acknowledged by it is delivered by the merges consumer once the assertion has
+// made it, because it starts where `controller` has acknowledged up to, not from now.
+func TestAMergeTheOldConsumerHadNotAcknowledgedIsTheMergesConsumers(t *testing.T) {
+	js, err := Dial(testbus.URL(t))
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer js.Close()
+	if err := AssertMeshStreams(js); err != nil {
+		t.Fatal(err)
+	}
+	old := &nats.ConsumerConfig{Durable: ControllerName, AckPolicy: nats.AckExplicitPolicy, AckWait: 30 * time.Second,
+		MaxDeliver: 5, MaxAckPending: 1, DeliverSubject: "_DELIVER.controller.EVENTS", DeliverPolicy: nats.DeliverNewPolicy,
+		FilterSubjects: append(append([]string(nil), ControllerFollows...), AgainFilter(ControllerName))}
+	if _, err := js.Context().AddConsumer(EventsStream, old); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := js.Context().Publish(ControllerFollows[3], []byte(`{"commit":"c1"}`)); err != nil {
+		t.Fatal(err)
+	}
+	if err := AssertMeshConsumers(js); err != nil {
+		t.Fatal(err)
+	}
+	merges, err := js.Context().ConsumerInfo(EventsStream, MergesConsumerName)
+	if err != nil || merges.NumPending != 1 {
+		t.Fatalf("the merges consumer holds %v (%v); want the merge the old consumer never acknowledged", merges, err)
+	}
+}

internal/broker/derived.go

@@ -36,6 +36,12 @@ type Consumer struct {
 	// them over, and a loop that acts on one message at a time is the property the store window
 	// depends on. A queue group implies this, because a group has nothing to pull from.
 	Push bool
+	// TakesOverFrom names a consumer of the same stream whose filter this one takes over (issue 973): made while
+	// that consumer still filters this one's subject, it starts after what that consumer has acknowledged, so a
+	// message the old one held unacknowledged is not lost to the change. Otherwise it starts as FromNow says.
+	TakesOverFrom string
+	// DeliverSubject, when set, is where a push consumer delivers instead of DeliverSubjectFor's.
+	DeliverSubject string
 	// AckWaitSeconds before an unacknowledged delivery is redelivered.
 	AckWaitSeconds int
 	// MaxDeliver is how often a message is handed over before the consumer gives it up; zero for no

internal/broker/jetstream.go

@@ -343,6 +343,9 @@ func (j *JetStream) EnsureConsumer(c Consumer) error {
 		if c.FromNow {
 			want.DeliverPolicy = nats.DeliverNewPolicy
 		}
+		if from := j.ackFloorOfTheConsumerItTakesOver(c); from > 0 {
+			want.DeliverPolicy, want.OptStartSeq = nats.DeliverByStartSequencePolicy, from
+		}
 		if _, err := j.js.AddConsumer(c.Stream, want); err != nil {
 			return fmt.Errorf("creating consumer %s on %s: %w", c.Name, c.Stream, err)
 		}
@@ -352,6 +355,26 @@ func (j *JetStream) EnsureConsumer(c Consumer) error {
 	}
 }
 
+// ackFloorOfTheConsumerItTakesOver is the stream sequence a consumer made now starts at so as to lose nothing the
+// consumer it takes over from (Consumer.TakesOverFrom) has not acknowledged: that consumer's ack floor plus one, when
+// it exists and still filters this one's subject; zero otherwise.
+func (j *JetStream) ackFloorOfTheConsumerItTakesOver(c Consumer) uint64 {
+	if c.TakesOverFrom == "" || len(c.Filters) != 1 {
+		return 0
+	}
+	have, err := j.js.ConsumerInfo(c.Stream, c.TakesOverFrom)
+	if err != nil {
+		return 0
+	}
+	filters := append([]string{have.Config.FilterSubject}, have.Config.FilterSubjects...)
+	for _, f := range filters {
+		if f == c.Filters[0] {
+			return have.AckFloor.Stream + 1
+		}
+	}
+	return 0
+}
+
 // ConsumerState is what a reset says about a consumer, before and after.
 type ConsumerState struct {
 	DeliverPolicy string

internal/broker/nats.go

@@ -141,6 +141,12 @@ type Principal struct {
 	// controller asks. Composed from the records like AsksIssued, so the installer's first user list carries none.
 	AsksTheVault []string
 
+	// AsksTheConfig are the mesh-config verbs the controller asks in its own name (VerbsTheControllerAsksTheConfig,
+	// issue 925), those mesh-config's holder serves: on the subject that names its caller, so mesh-config takes a
+	// value as the controller carries it (ADR 0348 §8). Composed from the records like AsksTheVault, so the
+	// installer's first user list carries none.
+	AsksTheConfig []string
+
 	// AsksTheDesk says this principal is the vault's holder, which asks the desks' secret prompt in its own name
 	// (novox/hq ADR 0349 clause 9, issue 723): `node-launcher.secret` on any machine, on the subject that names
 	// its caller, and nothing else of that seat. With AsksIssued on the same principal, the vault's whole reach
@@ -316,6 +322,12 @@ var VerbsTheControllerAsksTheVault = []SeatVerb{{Seat: "mesh-vault", Verb: "acce
 	{Seat: "mesh-vault", Verb: "secret-ask"}, {Seat: "mesh-vault", Verb: "issued"}, {Seat: "mesh-vault", Verb: "rotate"},
 	{Seat: "mesh-vault", Verb: "move-asks"}, {Seat: "mesh-vault", Verb: "get"}}
 
+// VerbsTheControllerAsksTheConfig are the mesh-config seat's verbs the controller calls in its own name (issue 925):
+// `config`, the one verb mesh-config serves, which move-config uses to set each value and the settings verb to set,
+// clear and propose through mesh-config. mesh-config takes a value as the controller carries it only from the
+// subject that names the controller as caller (ADR 0348 §8); granted as Principal.AsksTheConfig says.
+var VerbsTheControllerAsksTheConfig = []SeatVerb{{Seat: "mesh-config", Verb: "config"}}
+
 // VaultMoveIn is the vault's verb a machine's node-engine asks in its own name to hand in a value sealed to its
 // machine (design 57's V3, issue 676); granted as Principal.MovesIntoTheVault says.
 var VaultMoveIn = SeatVerb{Seat: "mesh-vault", Verb: "move-in"}
@@ -356,6 +368,16 @@ var DegradableGrants = []DegradableGrant{
 		Subject: "mesh.seat." + VerbsTheControllerAsksTheOperatorChannel[0].Seat + ".tool." + VerbsTheControllerAsksTheOperatorChannel[0].Verb,
 		Reason:  "refused: the controller raises the secret-given condition as before (ADR 0380 §9)",
 	},
+	{
+		Subject: "$JS.ACK.EVENTS." + MergesConsumerName + ".>",
+		Reason: "carried only at genesis: the controller composes its own permissions and pushes them to the bus at " +
+			"its first push, which is when it acknowledges a merge (ADR 0407 decision 3, issue 973)",
+	},
+	{
+		Subject: "$KV." + MergeWindowBucket + ".>",
+		Reason: "refused: the controller closes a batch's window by a timer of its own for that batch, as the " +
+			"window setting says (ADR 0407 decision 1, issue 973)",
+	},
 }
 
 // perMachineEvents are a node-scoped seat's events about the holder itself, whose last token is the
@@ -623,7 +645,7 @@ func PermissionsFor(p Principal) (Permissions, error) {
 		// delivery from EVENTS — a consumer that cannot ack has every message redelivered for
 		// ever, refused by the list it already has.
 		sub = append(sub, ControllerFollows...)
-		pub = append(pub, "$JS.ACK.EVENTS."+ControllerName+".>")
+		pub = append(pub, "$JS.ACK.EVENTS."+ControllerName+".>", "$JS.ACK.EVENTS."+MergesConsumerName+".>")
 
 		// **Where an enrolment's answer goes**, and `allow_responses` does not cover it. That
 		// permits one reply to the reply subject of a message the user received — and a message a
@@ -1089,6 +1111,13 @@ func PermissionsFor(p Principal) (Permissions, error) {
 				pub = append(pub, "mesh.seat."+v.Seat+"."+CallKind+"."+v.Verb+"."+CallerToken(p.Username()))
 			}
 		}
+		// **And mesh-config's verb** (issue 925): it takes a trusted value from the controller alone, known from the
+		// subject's last token.
+		for _, v := range VerbsTheControllerAsksTheConfig {
+			if slices.Contains(p.AsksTheConfig, v.Verb) {
+				pub = append(pub, "mesh.seat."+v.Seat+"."+CallKind+"."+v.Verb+"."+CallerToken(p.Username()))
+			}
+		}
 	}
 
 	sort.Strings(pub)

internal/broker/streams.go

@@ -233,6 +233,18 @@ const ProvAppliedStream = "PROV_APPLIED"
 // all (ProvisionerAppliedSubject).
 const ProvAppliedSubjects = "prov.applied.>"
 
+// eventsTheControllerFollows are the events the controller's events consumer takes: all it follows but the forge's
+// merges, which the merges consumer takes (MergesConsumerName).
+func eventsTheControllerFollows() []string {
+	var out []string
+	for _, subject := range ControllerFollows {
+		if subject != ControllerFollows[3] {
+			out = append(out, subject)
+		}
+	}
+	return out
+}
+
 // DeliverSubjectFor is where a push consumer's messages land.
 //
 // **Per consumer, which means per stream as well as per name** (novox/hq 04-ISSUES/146). A push
@@ -248,6 +260,9 @@ const ProvAppliedSubjects = "prov.applied.>"
 // durable's name to its stream, and this subject was the one place that scoping was dropped. It
 // stays inside what a controller may already subscribe (`_DELIVER.controller.>`).
 func DeliverSubjectFor(c Consumer) string {
+	if c.DeliverSubject != "" {
+		return c.DeliverSubject
+	}
 	return "_DELIVER." + c.Name + "." + c.Stream
 }
 
@@ -317,6 +332,13 @@ func Overlaps() []string {
 // ack subject is derived from (nats.go: `$JS.ACK.<stream>.controller.>`).
 const ControllerName = "controller"
 
+// MergesConsumerName is the controller's durable consumer of the forge's merged event on EVENTS, apart from the
+// events consumer (novox/hq ADR 0407 decision 3, issue 973). A merge is never dropped, so this consumer has no
+// max-deliver, which would send a merge held five times to DEAD_LETTERS to wait for a person, and no healer
+// resets it, which would drop the merges it holds. A consumer of its own because both of those are properties of
+// a consumer, not of a subject.
+const MergesConsumerName = "controller-merges"
+
 // ControllerSeat is the role the control plane holds, and ControllerStates are the facts it states
 // under it (novox/hq ADR 0134).
 //
@@ -450,10 +472,27 @@ func MeshConsumers() []Consumer {
 			Why: "the controller is the single consumer of what nodes say; explicit ack and no " +
 				"max-deliver, because the store window's bound is the controller's own",
 		},
+		// The forge's merges (novox/hq ADR 0407 decision 3, issue 973): never dropped, so no max-deliver and no
+		// reset; a merge the controller cannot record yet is held with a growing delay until it can. **Asserted before
+		// the events consumer**, which stops filtering merges: the order of this list is the order of assertion.
+		{
+			Name:   MergesConsumerName,
+			Stream: "EVENTS",
+			// Within what the controller may subscribe already (`_DELIVER.controller.>`).
+			DeliverSubject: "_DELIVER." + ControllerName + ".merges",
+			Filters:        []string{ControllerFollows[3]},
+			Push:           true,
+			AckWaitSeconds: 30,
+			MaxAckPending:  1,
+			FromNow:        true,
+			TakesOverFrom:  ControllerName,
+			Why: "the forge's merges, one at a time; no max-deliver and never reset, because a merge is never " +
+				"dropped and no person releases one",
+		},
 		{
 			Name:           ControllerName,
 			Stream:         "EVENTS",
-			Filters:        append(append([]string(nil), ControllerFollows...), AgainFilter(ControllerName)),
+			Filters:        append(eventsTheControllerFollows(), AgainFilter(ControllerName)),
 			Push:           true,
 			AckWaitSeconds: 30,
 			MaxDeliver:     5,
@@ -463,12 +502,13 @@ func MeshConsumers() []Consumer {
 			MaxAckPending: 1,
 			FromNow:       true,
 			// **The one consumer the mesh resets by itself** (healer H4, issue 248): it fell a week
-			// behind once and held every merge after it. What a reset drops is caught up elsewhere —
+			// behind once and held every merge after it. What a reset drops is caught up elsewhere — a
-			// a merge by the catch-up pass that reads the forge (issue 266), a build's outcome from the
+			// build's outcome from the build records a plan settles from (issue 214), a provider's failing
-			// build records a plan settles from (issue 214), a provider's failing word by the provider
+			// word by the provider saying it again every quarter of an hour (ADR 0224). **Not the merges**:
-			// saying it again every quarter of an hour (ADR 0224).
+			// they are the merges consumer's below, which a reset leaves alone (ADR 0407 decision 3).
-			Resettable: "what it drops is caught up: merges by the catch-up pass (issue 266), build outcomes " +
+			Resettable: "what it drops is caught up: build outcomes from the build records (issue 214), a " +
-				"from the build records (issue 214), a provider's failing word said again (ADR 0224)",
+				"provider's failing word said again (ADR 0224); the forge's merges are not its, they have a " +
+				"consumer of their own that no reset touches",
 			Why: "the events the mesh's own controller reacts to, one at a time; after " +
 				"max-deliver it gives the event up, and the controller keeps it in DEAD_LETTERS until " +
 				"a person delivers it again or drops it",

internal/broker/testdata/composed.conf

@@ -24,7 +24,7 @@ accounts {
     jetstream: enabled
     users = [
       { user: "controller", password: "$2a$11$cccccccccccccccccccccc", permissions: {
-        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.DEAD_LETTER_NOTICES.controller.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "$KV.SEAT_MESH_BUILD_MACHINE_cancelled.>", "$KV.SEAT_NODE_BUILD_AGENT_cancelled.>", "$KV.mesh-controller_asked.>", "$KV.mesh-controller_assignments.>", "$KV.mesh-controller_calls.>", "$KV.mesh-controller_condition-history.>", "$KV.mesh-controller_conditions.>", "$KV.mesh-controller_declared-config.>", "$KV.mesh-controller_hand-acts.>", "$KV.mesh-controller_lease.>", "$KV.mesh-controller_predefined.>", "$SRV.INFO", "_INBOX.enrol.>", "mesh.again.>", "mesh.assignment.>", "mesh.events.dead.>", "mesh.mod.*.tool.>", "mesh.node.>", "mesh.seat.mesh-build-machine.accept.>", "mesh.seat.mesh-build-machine.tool.>", "mesh.seat.mesh-controller.event.applied", "mesh.seat.mesh-controller.event.assignment-changed", "mesh.seat.mesh-controller.event.built-before", "mesh.seat.mesh-controller.event.checked", "mesh.seat.mesh-controller.event.condition-changed", "mesh.seat.mesh-controller.event.condition-cleared", "mesh.seat.mesh-controller.event.condition-raised", "mesh.seat.mesh-controller.event.declared-config-changed", "mesh.seat.mesh-controller.event.doctor-heartbeat", "mesh.seat.mesh-controller.event.healer-acted", "mesh.seat.mesh-controller.event.plan-moved", "mesh.seat.mesh-controller.event.predefined-changed", "mesh.seat.mesh-controller.event.refused", "mesh.seat.mesh-controller.event.rolled-back", "mesh.seat.mesh-controller.event.secret-replaced", "mesh.seat.mesh-controller.tool.plans", "mesh.seat.mesh-delivery.tool.close", "mesh.seat.mesh-delivery.tool.groups", "mesh.seat.mesh-delivery.tool.release", "mesh.seat.mesh-delivery.tool.stalled", "mesh.seat.mesh-delivery.tool.stop", "mesh.seat.mesh-delivery.tool.times", "mesh.seat.node-backup.tool.backed-up.*", "mesh.seat.node-backup.tool.now.*", "mesh.seat.node-build-agent.accept.>", "mesh.seat.node-build-agent.tool.>", "mesh.seat.node-intrusion-prevention.tool.banned.*", "mesh.seat.node-launcher.tool.secret.*", "mesh.seat.node-service-manager.tool.restart.*", "mesh.seat.operator-channel.tool.tell"] }
+        publish: { allow: ["$JS.ACK.CONTROL.controller.>", "$JS.ACK.DEAD_LETTER_NOTICES.controller.>", "$JS.ACK.EVENTS.controller-merges.>", "$JS.ACK.EVENTS.controller.>", "$JS.API.>", "$KV.SEAT_MESH_BUILD_MACHINE_cancelled.>", "$KV.SEAT_NODE_BUILD_AGENT_cancelled.>", "$KV.mesh-controller_asked.>", "$KV.mesh-controller_assignments.>", "$KV.mesh-controller_calls.>", "$KV.mesh-controller_condition-history.>", "$KV.mesh-controller_conditions.>", "$KV.mesh-controller_declared-config.>", "$KV.mesh-controller_hand-acts.>", "$KV.mesh-controller_lease.>", "$KV.mesh-controller_merge-window.>", "$KV.mesh-controller_predefined.>", "$SRV.INFO", "_INBOX.enrol.>", "mesh.again.>", "mesh.assignment.>", "mesh.events.dead.>", "mesh.mod.*.tool.>", "mesh.node.>", "mesh.seat.mesh-build-machine.accept.>", "mesh.seat.mesh-build-machine.tool.>", "mesh.seat.mesh-controller.event.applied", "mesh.seat.mesh-controller.event.assignment-changed", "mesh.seat.mesh-controller.event.built-before", "mesh.seat.mesh-controller.event.checked", "mesh.seat.mesh-controller.event.condition-changed", "mesh.seat.mesh-controller.event.condition-cleared", "mesh.seat.mesh-controller.event.condition-raised", "mesh.seat.mesh-controller.event.declared-config-changed", "mesh.seat.mesh-controller.event.doctor-heartbeat", "mesh.seat.mesh-controller.event.healer-acted", "mesh.seat.mesh-controller.event.plan-moved", "mesh.seat.mesh-controller.event.predefined-changed", "mesh.seat.mesh-controller.event.refused", "mesh.seat.mesh-controller.event.rolled-back", "mesh.seat.mesh-controller.event.secret-replaced", "mesh.seat.mesh-controller.tool.plans", "mesh.seat.mesh-delivery.tool.close", "mesh.seat.mesh-delivery.tool.groups", "mesh.seat.mesh-delivery.tool.release", "mesh.seat.mesh-delivery.tool.stalled", "mesh.seat.mesh-delivery.tool.stop", "mesh.seat.mesh-delivery.tool.times", "mesh.seat.node-backup.tool.backed-up.*", "mesh.seat.node-backup.tool.now.*", "mesh.seat.node-build-agent.accept.>", "mesh.seat.node-build-agent.tool.>", "mesh.seat.node-intrusion-prevention.tool.banned.*", "mesh.seat.node-launcher.tool.secret.*", "mesh.seat.node-service-manager.tool.restart.*", "mesh.seat.operator-channel.tool.tell"] }
         subscribe: { allow: ["$JS.API.>", "$JS.EVENT.ADVISORY.CONSUMER.DELETED.>", "$JS.EVENT.ADVISORY.CONSUMER.MAX_DELIVERIES.>", "$SRV.INFO", "$SRV.INFO.mesh-controller", "$SRV.INFO.mesh-controller.>", "$SRV.PING", "$SRV.PING.mesh-controller", "$SRV.PING.mesh-controller.>", "$SRV.STATS", "$SRV.STATS.mesh-controller", "$SRV.STATS.mesh-controller.>", "_DELIVER.controller", "_DELIVER.controller.>", "_INBOX.controller.>", "mesh.control.>", "mesh.mod.*.event.provisioner.failing", "mesh.mod.*.event.provisioner.recovered", "mesh.mod.*.event.provisioner.retirement", "mesh.mod.gitea.event.pull.merged", "mesh.mod.gitea.event.pull.updated", "mesh.mod.mesh-catalog.event.catching-up", "mesh.mod.mesh-catalog.event.upgraded", "mesh.seat.mesh-build-machine.event.built", "mesh.seat.mesh-controller.tool.>", "mesh.seat.node-build-agent.event.built", "mesh.seat.operator-channel.event.decided.mesh-controller"] }
         allow_responses: { max: 1, ttl: "1m" }
       } }

internal/broker/users.go

@@ -102,7 +102,8 @@ type Records struct {
 // is a mesh that cannot be told anything, and there is no state of the records in which that is
 // correct.
 func Users(r Records) ([]Principal, error) {
-	out := []Principal{{Kind: KindController, Uses: asksSeatOf(r), AsksIssued: issuedOf(r), AsksTheVault: vaultVerbsOf(r)}}
+	out := []Principal{{Kind: KindController, Uses: asksSeatOf(r), AsksIssued: issuedOf(r), AsksTheVault: vaultVerbsOf(r),
+		AsksTheConfig: servedOf(r, VerbsTheControllerAsksTheConfig)}}
 
 	for _, node := range sortedCopy(r.Nodes) {
 		witness := false
@@ -290,13 +291,16 @@ func holdsSeat(d Declared, seat string) bool {
 
 // vaultVerbsOf are the vault's intake verbs its holder serves, as the records say (issue 675): what the controller
 // is granted to ask it in its own name. None while no holder serves them.
-func vaultVerbsOf(r Records) []string {
+func vaultVerbsOf(r Records) []string { return servedOf(r, VerbsTheControllerAsksTheVault) }
+
+// servedOf are the verbs of asked that a seat's holder serves, as the records say, by name.
+func servedOf(r Records, asked []SeatVerb) []string {
 	seen := map[string]bool{}
 	var out []string
 	for _, node := range sortedCopy(r.Nodes) {
 		for _, d := range r.Assigned[node] {
 			for _, s := range d.Holds {
-				for _, v := range VerbsTheControllerAsksTheVault {
+				for _, v := range asked {
 					if s.Name == v.Seat && slices.Contains(s.Serves, v.Verb) && !seen[v.Verb] {
 						seen[v.Verb] = true
 						out = append(out, v.Verb)

internal/broker/writers.go

@@ -117,8 +117,11 @@ var WritersTable = []WriterRow{
 		Subjects: []string{"mesh.control.*.report", "mesh.control.*.health", "mesh.control.*.cli"}, Writes: ownMachine},
 	{State: "the controller lease", Writer: "the controller instance holding it", KeptIn: "key-value " + LeaseBucket,
 		Others: "a candidate waits", Subjects: kvOf(LeaseBucket), Writes: isController},
+	// And the clock of an open batch's merge window, a plan before it is one (novox/hq ADR 0407 decision 1): keys that
+	// expire in a bucket of their own, written by the lease holder.
 	{State: "plans and their tiers", Writer: "controller (lease holder), compare-and-set on the plan's revision",
-		KeptIn: "the controller's store", Others: "read through plans"},
+		KeptIn: "the controller's store; a batch's merge window in key-value " + MergeWindowBucket, Others: "read through plans",
+		Subjects: kvOf(MergeWindowBucket), Writes: isController},
 	{State: "conditions", Writer: "controller", KeptIn: "key-value " + ConditionsBucket + " (and its history, " +
 		ConditionHistoryBucket + "); how often each was observed, when last and its newest evidence in the " +
 		"controller's store, so a look that changes nothing writes no bucket", Others: "raise or clear only through observations the controller reads",

internal/catalogue/bound.go

@@ -111,7 +111,7 @@ func boundNeed(n Needed, bound Chosen, catalogue map[string]Manifest, node Node,
 		}
 		held.From, held.At, held.Module = node.Name, at, m.Module
 		held.Serves, held.SharedOwn, held.Identity = servedByOne(m, n.Name), sharedByOne(m, n.Name), m.IdentityBoundOf(n.Name)
-		held.CredentialForm = m.CredentialFormOf(n.Name)
+		held.CredentialForm, held.CredentialTool = m.CredentialFormOf(n.Name), m.CredentialToolOf(n.Name)
 		return held, true, ""
 	}
 	matching := bound.among(world.Offered[n.Name])
@@ -125,7 +125,7 @@ func boundNeed(n Needed, bound Chosen, catalogue map[string]Manifest, node Node,
 	identity := DefaultIdentityBound
 	if pm, known := catalogue[p.Module]; known {
 		held.SharedOwn, _ = pm.SharedCredentialOf(n.Name)
-		held.CredentialForm = pm.CredentialFormOf(n.Name)
+		held.CredentialForm, held.CredentialTool = pm.CredentialFormOf(n.Name), pm.CredentialToolOf(n.Name)
 		identity = pm.IdentityBoundOf(n.Name)
 	}
 	held.From, held.At, held.Module, held.Serves, held.Identity = p.Node, p.At, p.Module, p.Serves, identity

internal/catalogue/handed.go

@@ -104,6 +104,9 @@ type PlanProvision struct {
 	Provider       string `json:"provider"`
 	ProviderModule string `json:"provider-module"`
 	CredentialForm string `json:"credential-form,omitempty"`
+	// CredentialTool is, for an offer whose credential the provider's software issues (issue 928), the provider's tool
+	// the vault asks for each holder's value; empty for any other form.
+	CredentialTool string `json:"credential-tool,omitempty"`
 	Path           string `json:"path,omitempty"`
 	// Contributes is what the consumer's manifest tells its provider for this provision (Manifest.Contributes: a
 	// database's name, its extensions), so the provider's helper learns it from the bus (issue 776). Nil where the
@@ -134,7 +137,7 @@ func provisionsOf(needs []Needed, module, as string, contributes map[string]map[
 			continue
 		}
 		p := PlanProvision{Provision: n.Name, Local: n.Local, Provider: n.From, ProviderModule: n.Module,
-			CredentialForm: n.CredentialForm, Contributes: maps.Clone(contributes[n.Name]),
+			CredentialForm: n.CredentialForm, CredentialTool: n.CredentialTool, Contributes: maps.Clone(contributes[n.Name]),
 			At: n.At, From: n.From, As: as}
 		// **A credential only a template names** (issue 788): named like a file's, so the client fills it from the
 		// vault's row for this holder.

internal/catalogue/handed_bound_test.go

@@ -123,3 +123,37 @@ func TestAHandedModulesLocalCredentialsOfTheVaultAreNamedByLocalInItsPlan(t *tes
 		t.Fatalf("the plan names %v, want admin and internal-token", names)
 	}
 }
+
+// **Issue 969: a key its provider serves only at run time.** A moved module's file may name `${bound:…}` of a
+// provision whose provider tells it at run time (ssh-client's known_host of the forge). Where the machine takes
+// handed files the client fills it and the module composes; where it does not the controller must fill it from what
+// the provider serves statically, cannot, and refuses it with the very words an unmarked module is refused with
+// (ADR 0349 §17: the old way, byte for byte).
+func TestAHandedModuleNamingARunTimeBoundKeyComposesOnlyWhereItIsHanded(t *testing.T) {
+	text := "Host ${bound:postgres-database:at}\n  KnownHost ${bound:postgres-database:known_host}\n"
+	moved := strings.Replace(botMoved(), `"own-secrets":{"token":"${dir:state}/token"}`,
+		`"requires":["postgres-database"],"own-secrets":{"token":"${dir:state}/token"}`, 1)
+	unmoved := strings.Replace(moved, `"moved":true,`, ``, 1)
+	if unmoved == moved {
+		t.Fatal("the twin is still marked moved")
+	}
+
+	c, got := composeWith(t, withDatabase(t, moved, text, true), handingTo("store:5000"))
+	if strings.Contains(got, "known_host") || len(c.Filled) == 0 {
+		t.Fatalf("handed, the module is filled by the client and the controller names nothing: %v %s", c.Filled, got)
+	}
+
+	for name, with := range map[string]Rendering{
+		"an engine that does not take a filled file alone": {Provisioner: "mesh-provisioner-client", ArtifactStore: "store:5000"},
+		"no provisioner client on the machine":             {Hands: true, ArtifactStore: "store:5000"},
+	} {
+		_, err := withDatabase(t, moved, text, true).Compose(sealedForBot(with))
+		if err == nil || !strings.Contains(err.Error(), `asks its postgres-database binding for "known_host"`) {
+			t.Fatalf("%s: the key is not refused: %v", name, err)
+		}
+		_, before := withDatabase(t, unmoved, text, false).Compose(sealedForBot(with))
+		if before == nil || before.Error() != err.Error() {
+			t.Fatalf("%s: refused otherwise than the unmarked module: %v / %v", name, err, before)
+		}
+	}
+}

internal/catalogue/handed_providers_test.go

@@ -242,3 +242,33 @@ func TestAnOfferAnsweredByTheHelperIsReadAndHandsNoFile(t *testing.T) {
 		t.Fatalf("the flag changed what is composed or handed:\nbefore: %s\nnow:    %s\n%v %v", before, got, c2.Filled, c2.Plans)
 	}
 }
+
+// A plan entry names the tool of an offer whose credential the provider's software issues (issue 928), so the vault
+// asks that tool for each holder; an offer of any other form names none.
+func TestAHandedModulesPlanNamesTheCredentialToolOfASoftwareIssuedOffer(t *testing.T) {
+	for credential, want := range map[string]string{
+		``:                                      "",
+		`,"credential":{"own":"admin"}`:         "",
+		`,"credential":{"issued-by":"outside"}`: "",
+		`,"credential":{"issued-by":"software","tool":"gitea_agent_token"}`: "gitea_agent_token",
+	} {
+		raw := `{"module":"store","version":"1","provides":[{"name":"postgres-database","scope":"mesh"` +
+			credential + `}],"serves":{"postgres-database":{"port":5432}}`
+		if strings.Contains(credential, `"own"`) {
+			raw += `,"own-secrets":{"admin":{"path":"/var/lib/store/admin","taken":"at-start"}}`
+		}
+		node := workstation()
+		node.At = "workstation.internal"
+		consumer := notifierConsumer(t, true, false)
+		r, err := Resolve(shelf(consumer, mustParse(t, raw+`}`)), []string{consumer.Module}, node, twoStores("anchor"))
+		if err != nil {
+			t.Fatal(err)
+		}
+		r.Account = "op"
+		c, _ := composeWith(t, r, handingTo("store:5000"))
+		got := c.Plans["notify"].Provisions
+		if len(got) != 1 || got[0].CredentialTool != want {
+			t.Errorf("an offer with %q: the plan names %+v, want the tool %q", credential, got, want)
+		}
+	}
+}

internal/catalogue/jail_problems_test.go

@@ -0,0 +1,49 @@
+package catalogue
+
+import (
+	"strings"
+	"testing"
+)
+
+// A jail whose pattern names <HOST> twice is refused by name: fail2ban would read it as a duplicate
+// capture group and refuse its whole configuration, leaving the machine with no bans at all — which
+// is what happened on the control node the day this was built (novox/hq ADR 0179).
+func TestAJailsPatternNamesTheHostOncePerShape(t *testing.T) {
+	twice := Manifest{Module: "route-proxy", Jails: []Jail{{
+		Name:      "route-proxy",
+		Failregex: `^.*(?:refused for <HOST>:\d+|asked from <HOST>:\d+)$`,
+		Jail:      "backend = systemd",
+	}}}
+	problems := twice.jailProblems()
+	if len(problems) != 1 || !strings.Contains(problems[0], "names <HOST> 2 times") {
+		t.Fatalf("a pattern naming <HOST> twice was not refused by name: %v", problems)
+	}
+	if !strings.Contains(problems[0], "no bans at all") {
+		t.Errorf("the refusal does not say what it costs: %q", problems[0])
+	}
+
+	// The same two shapes, one pattern each, is how it is written — and is accepted.
+	split := Manifest{Module: "route-proxy", Jails: []Jail{{
+		Name:      "route-proxy",
+		Failregex: "^.*refused for <HOST>:\\d+$\n            ^.*asked from <HOST>:\\d+$",
+		Jail:      "backend = systemd",
+	}}}
+	if problems := split.jailProblems(); len(problems) != 0 {
+		t.Fatalf("one <HOST> per pattern was refused: %v", problems)
+	}
+}
+
+// A jail with no name, no pattern, or a name another jail of the same module already took is refused.
+func TestAJailIsNamedOnceAndSaysWhatAFailureLooksLike(t *testing.T) {
+	m := Manifest{Module: "mailu", Jails: []Jail{
+		{Name: "", Failregex: "^.*rip=<HOST>$"},
+		{Name: "mailu-front", Failregex: "  "},
+		{Name: "mailu-front", Failregex: "^.*rip=<HOST>$"},
+	}}
+	problems := strings.Join(m.jailProblems(), "\n")
+	for _, want := range []string{"a jail with no name", "says nothing a failed attempt looks like", "two jails called"} {
+		if !strings.Contains(problems, want) {
+			t.Errorf("missing %q in:\n%s", want, problems)
+		}
+	}
+}

internal/catalogue/manifest.go

@@ -303,6 +303,13 @@ func (m Manifest) IssuedCredentialOf(provision string) (string, bool) {
 	return "", false
 }
 
+// CredentialToolOf is the tool a plan names for a provision, set only where the offer says the provider's software
+// issues its credential (issue 928); empty otherwise.
+func (m Manifest) CredentialToolOf(provision string) string {
+	tool, _ := m.IssuedCredentialOf(provision)
+	return tool
+}
+
 // IssuedOutsideOf is whether this module's offer of a provision says only a person outside the mesh can give its
 // credential (novox/hq ADR 0337): the mesh never makes a value for it, and a consumer holds one only once a person
 // gave it with `secret accept`.

internal/catalogue/resolve.go

@@ -230,6 +230,8 @@ type Needed struct {
 	// CredentialForm is the credential form the provider's offer declares (Manifest.CredentialFormOf, issue 768),
 	// named in a handed consumer's plan for the vault; empty where the provider's manifest is not in hand.
 	CredentialForm string
+	// CredentialTool is the tool of the offer that the provider's software issues the credential through (issue 928).
+	CredentialTool string
 	// Manager is set when this holder is a refreshable-grant licence's MANAGER, delivered the refresh
 	// token rather than an access token (novox/hq ADR 0050). It changes one thing downstream: an empty
 	// Sealed is tolerated — the manager has not adopted a refresh token yet, which is a real waiting
@@ -543,7 +545,7 @@ func Resolve(catalogue map[string]Manifest, assigned []string, node Node, world
 				needs = append(needs, Needed{
 					Name: want, From: node.Name, At: at, Module: by.Module,
 					Serves: servedByOne(by, want), For: because[want],
-					SharedOwn: sharedByOne(by, want), CredentialForm: by.CredentialFormOf(want),
+					SharedOwn: sharedByOne(by, want), CredentialForm: by.CredentialFormOf(want), CredentialTool: by.CredentialToolOf(want),
 					Identity: by.IdentityBoundOf(want)})
 			} else if served := servedByOne(by, want); len(served) > 0 {
 				// Answered here with no credential to mint, but the provider serves facts the
@@ -565,7 +567,7 @@ func Resolve(catalogue map[string]Manifest, assigned []string, node Node, world
 				}
 				needs = append(needs, Needed{
 					Name: want, From: node.Name, At: at, Serves: served, For: because[want], Module: by.Module,
-					CredentialForm: by.CredentialFormOf(want)})
+					CredentialForm: by.CredentialFormOf(want), CredentialTool: by.CredentialToolOf(want)})
 			}
 			continue
 		}
@@ -594,17 +596,17 @@ func Resolve(catalogue map[string]Manifest, assigned []string, node Node, world
 						node.Name, want, p.Node, meshNetwork))
 					return
 				}
-				shared, form := "", ""
+				shared, form, tool := "", "", ""
 				// A provider whose definition is not in hand is held to the tightest bound the
 				// mesh knows, not to none: what it keeps is not known here (ADR 0225).
 				bound := DefaultIdentityBound
 				if pm, known := catalogue[p.Module]; known {
 					shared, _ = pm.SharedCredentialOf(want)
-					form = pm.CredentialFormOf(want)
+					form, tool = pm.CredentialFormOf(want), pm.CredentialToolOf(want)
 					bound = pm.IdentityBoundOf(want)
 				}
 				needs = append(needs, Needed{Name: want, From: p.Node, At: p.At, Module: p.Module,
-					Serves: p.Serves, For: because[want], SharedOwn: shared, CredentialForm: form, Identity: bound,
+					Serves: p.Serves, For: because[want], SharedOwn: shared, CredentialForm: form, CredentialTool: tool, Identity: bound,
 					LeftOut: p.LeftOut})
 			}
 			switch {

internal/catalogue/ssh_client_test.go

@@ -1,19 +1,66 @@
 package catalogue
 
 import (
+	"io/fs"
+	"os"
+	"path/filepath"
+	"regexp"
 	"strings"
 	"testing"
+
+	"github.com/novox/mesh-controller/internal/beside"
 )
 
+// boundKeys are every key the files of a module ask of ${bound:forge-account:<key>}, read from the module's own
+// tree on the shelf, so the test offers what ssh-client asks however often it asks one more (issue 985).
+func boundKeys(t *testing.T, module, provision string) map[string]bool {
+	t.Helper()
+	ask := regexp.MustCompile(`\$\{bound:` + regexp.QuoteMeta(provision) + `:([A-Za-z0-9_.-]+)\}`)
+	keys := map[string]bool{}
+	root := filepath.Join(beside.Catalogue(t), module)
+	err := filepath.WalkDir(root, func(path string, d fs.DirEntry, err error) error {
+		if err != nil || d.IsDir() || strings.HasSuffix(path, "_test.go") {
+			return err
+		}
+		body, err := os.ReadFile(path)
+		if err != nil {
+			return err
+		}
+		for _, m := range ask.FindAllStringSubmatch(string(body), -1) {
+			keys[m[1]] = true
+		}
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("reading what %s asks of %s: %v", module, provision, err)
+	}
+	return keys
+}
+
 // The ssh-client module, composed as a machine receives it (novox/hq to-be 29, research 027/03): a
 // region at the START of the operator's ~/.ssh/config that includes ~/.ssh/config.d/* — first,
 // because ssh takes the first value it finds for each option — and the mesh's Host blocks as the
 // whole of ~/.ssh/config.d/00-mesh, every other node with its account. ~/.ssh and ~/.ssh/config.d
 // are created 0700 and owned by the account; the operator's own config below the region is kept.
 func TestSSHClientOwnsTheOperatorsSSHConfig(t *testing.T) {
-	shelf := shelf(catalogueManifest(t, "ssh-client"))
+	shelf := shelf(catalogueManifest(t, "ssh-client"), catalogueManifest(t, "gitea"))
+	// ssh-client requires forge-account since mesh-catalog #340 (novox/hq issue 955), answered by the forge's
+	// holder on another node; this test is about the operator's config, so the forge is offered from there
+	// (issue 985).
+	// Every key ssh-client's files ask is offered, read from them (boundKeys); a few have a realistic value.
+	serves := map[string]any{}
+	for key := range boundKeys(t, "ssh-client", "forge-account") {
+		serves[key] = "sample-" + key
+	}
+	for key, value := range map[string]any{"ssh_host": "marge.internal", "ssh_port": 222,
+		"known_host": "[marge.internal]:222 ssh-ed25519 AAAA", "ssh_base": "ssh://git@marge.internal:222/",
+		"https_base": "https://git.example.org/"} {
+		serves[key] = value
+	}
+	forge := World{Offered: map[string][]Provider{
+		"forge-account": {{Node: "marge", At: "marge.internal", Module: "gitea", Serves: serves}}}}
 	got, err := Resolve(shelf, []string{"ssh-client"},
-		Node{Name: "homer", At: "homer.internal", Account: "jo", AccountHome: "/home/jo"}, World{})
+		Node{Name: "homer", At: "homer.internal", Account: "jo", AccountHome: "/home/jo"}, forge)
 	if err != nil {
 		t.Fatal(err)
 	}

internal/facts/facts.go

@@ -112,6 +112,10 @@ type Machine struct {
 	// NodeEngine and NodeTools are the builds it runs.
 	NodeEngine string `json:"node-engine,omitempty"`
 	NodeTools  string `json:"node-tools,omitempty"`
+	// Contract is the contract of the node-engine's newest health statement (0 when it never stated one): what
+	// says whether the machine takes handed files (novox/hq issue 969). A check that composed it without would
+	// judge a moved module the old way where the mesh hands it.
+	Contract int `json:"contract,omitempty"`
 	// Capabilities are what it reported it can do; the detail kept only where it is a version.
 	Capabilities []Capability `json:"capabilities,omitempty"`
 	// Site, Hub, Public: where it is on the private network — its site (a pseudonym), whether it is the

internal/inventory/plans.go

@@ -119,6 +119,8 @@ type PlanBatch struct {
 	ClosesAt time.Time `json:"closes_at"`
 	AtMost   time.Time `json:"at_most"`
 	Behind   string    `json:"behind,omitempty"`
+	// BehindWhy says why the batch waits behind that walk: what the two share (issue 973).
+	BehindWhy string `json:"behind_why,omitempty"`
 	// Own says the batch holds merges on the controller's own path, whose walk waits for nobody's word: such
 	// a merge never shares a batch with one that waits for mesh-delivery's (ADR 0276, decided during the
 	// build, 2026-10-10).

internal/inventory/provisioning.go

@@ -81,18 +81,10 @@ type ProvisionedCollection struct {
 }
 
 // ProvisionedCredential is one provision the module requires, under one local name, and its provider (issue 760);
-// with a name, a credential issued for this holder alone (issue 757).
+// with a name, a credential issued for this holder alone (issue 757). It is the composition's own type, not a copy:
-type ProvisionedCredential struct {
+// the plan is kept and read back as the one it was sent as, so no field of it (contributes, at, from, as, names) can
-	Name           string `json:"name,omitempty"`
+// be lost between the two (issue 923).
-	Provision      string `json:"provision"`
+type ProvisionedCredential = catalogue.PlanProvision
-	Local          string `json:"local,omitempty"`
-	Provider       string `json:"provider"`
-	ProviderModule string `json:"provider-module"`
-	// CredentialForm is the form the provider's offer declares for its credential (issue 768): shared, pair,
-	// issued or given; empty when the provider's manifest was not in hand.
-	CredentialForm string `json:"credential-form,omitempty"`
-	Path           string `json:"path,omitempty"`
-}
 
 // ProvisionedSecret is one of the module's own secrets, by its name in the manifest. Path, when set, is the file
 // holding the secret alone (the manifest's `own-secrets`), which the client fills with its value and nothing else.

internal/inventory/provisioning_test.go

@@ -352,3 +352,42 @@ func TestAPlanMayHoldSeveralNamedItemsOfOneCollectionAndRefusesABadOne(t *testin
 		}
 	}
 }
+
+// **A kept plan's provisions keep every field the composition sets** (issue 923): the names, the place, the machine
+// and the identity the provider's helper writes into the consumer's entry, and what the manifest contributes, are
+// read back as they were sent, so the assignments mirror carries them to the provider.
+func TestAKeptPlansProvisionsKeepContributesAtFromAsAndNames(t *testing.T) {
+	sent := catalogue.PlanProvision{Provision: "route", Provider: "novox", ProviderModule: "route-proxy",
+		Contributes: map[string]any{"database": "gitea"}, At: "10.0.0.1", From: "novox", As: "gitea@novox",
+		Names: map[string]string{"name-web": "git.example.org", "internal-name": "gitea.mesh"}}
+	raw, err := json.Marshal(map[string]any{"gitea": map[string]any{
+		"provisions": []catalogue.PlanProvision{sent}, "required-by": "forge"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var kept map[string]keptPlan
+	if err := json.Unmarshal(raw, &kept); err != nil {
+		t.Fatal(err)
+	}
+	got := kept["gitea"].Provisions
+	if len(got) != 1 || !reflect.DeepEqual(got[0], sent) {
+		t.Fatalf("the kept plan's provisions are %+v, want %+v", got, sent)
+	}
+}
+
+// A kept plan's provision keeps the tool the vault asks for a software-issued credential (issue 928).
+func TestAKeptPlansProvisionKeepsItsCredentialTool(t *testing.T) {
+	sent := catalogue.PlanProvision{Provision: "gitea-read", Provider: "novox", ProviderModule: "gitea",
+		CredentialForm: catalogue.CredentialFormIssued, CredentialTool: "gitea_agent_token"}
+	raw, err := json.Marshal(map[string]any{"claude-code": map[string]any{"provisions": []catalogue.PlanProvision{sent}}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var kept map[string]keptPlan
+	if err := json.Unmarshal(raw, &kept); err != nil {
+		t.Fatal(err)
+	}
+	if got := kept["claude-code"].Provisions; len(got) != 1 || got[0].CredentialTool != "gitea_agent_token" {
+		t.Fatalf("the kept plan's provisions are %+v, want the tool kept", got)
+	}
+}

internal/link/advisories.go

@@ -145,7 +145,7 @@ func MeshNamed(stream, name string) bool {
 		return true
 	case strings.HasPrefix(stream, "SEAT_"):
 		return strings.HasSuffix(name, "_worker")
-	case name == broker.ControllerName:
+	case name == broker.ControllerName || name == broker.MergesConsumerName:
 		return true
 	case stream == broker.EventsStream:
 		return strings.Contains(name, "_")
@@ -156,7 +156,7 @@ func MeshNamed(stream, name string) bool {
 // ConsumerInWords is a durable consumer as the mesh says it: whose, and for what.
 func ConsumerInWords(stream, name string) string {
 	switch {
-	case name == broker.ControllerName:
+	case name == broker.ControllerName || name == broker.MergesConsumerName:
 		return "the controller's consumer on " + stream
 	case stream == "NODES":
 		return "how " + name + " hears what it should be (its declaration consumer)"

internal/link/bus.go

@@ -179,6 +179,19 @@ func (b OverNATS) PublishMembership(ctx context.Context, node, module string, bo
 	return nil
 }
 
+// PublishMembershipAfter issues a membership only if the one read at sequence seq is still the newest of its
+// subject: a push that issued another since is not overwritten, and the refusal comes back (novox/hq issue 930).
+func (b OverNATS) PublishMembershipAfter(ctx context.Context, node, module string, body []byte, seq uint64) error {
+	ctx, cancel := context.WithTimeout(ctx, MembershipWait)
+	defer cancel()
+	_, err := b.JS.Publish(broker.MembershipSubject(node, module), body, nats.Context(ctx),
+		nats.ExpectLastSequencePerSubject(seq))
+	if err != nil {
+		return fmt.Errorf("issuing %s on %s its membership after sequence %d: %w", module, node, seq, err)
+	}
+	return nil
+}
+
 func (b OverNATS) PublishDeclaration(ctx context.Context, node string, body []byte) error {
 	_, err := b.JS.Publish(DeclareSubject(node), body, nats.Context(ctx))
 	if err != nil {

internal/link/receive_fake_test.go

@@ -68,6 +68,7 @@ type fakeControl struct {
 	on          *fakeInbound
 	redelivered bool
 	since       time.Time
+	holdFails   error
 }
 
 func (m *fakeControl) Kind() string                         { return m.kind }
@@ -88,6 +89,9 @@ func (m *fakeControl) HeldFor() time.Duration {
 }
 
 func (m *fakeControl) Hold(time.Duration) error {
+	if m.holdFails != nil {
+		return m.holdFails
+	}
 	if m.since.IsZero() {
 		m.since = time.Now()
 	}

internal/link/receive_nats.go

@@ -148,6 +148,24 @@ func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Con
 		defer func() { _ = followed.Unsubscribe() }()
 	}
 
+	// And the forge's merges, off a consumer of their own: no max-deliver and no reset, so a merge held for the
+	// plans waits for them however long (novox/hq ADR 0407 decision 3). Bound to the same channel: one loop.
+	if n.follows[KindSourceMoved] {
+		if events == nil {
+			events = make(chan *nats.Msg, Prefetch)
+		}
+		merges, err := standingBy(ctx, log.Default(), "EVENTS", func() (*nats.Subscription, error) {
+			return js.ChanSubscribe(MergedSubject, events, nats.Bind("EVENTS", broker.MergesConsumerName))
+		})
+		if err != nil {
+			return fmt.Errorf("subscribing to the forge's merges: %w", err)
+		}
+		if merges == nil {
+			return nil
+		}
+		defer func() { _ = merges.Unsubscribe() }()
+	}
+
 	// A connection that dropped is said, not discovered. A controller whose bus connection is gone
 	// is a mesh where nothing can be told anything.
 	gone := make(chan error, 1)

internal/link/receive_nats_test.go

@@ -590,3 +590,77 @@ func TestNatsAReportThatIsNotNewsIsNotStated(t *testing.T) {
 type sameAgain struct{}
 
 func (sameAgain) Heard(context.Context, Report) (bool, error) { return false, nil }
+
+// manyTimesBusy is an upgrader whose plans are busy for its first merges, as another controller's hold is.
+type manyTimesBusy struct {
+	toldAbout
+	busy  int
+	calls int
+}
+
+func (u *manyTimesBusy) SourceMoved(context.Context, SourceMoved) error {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	u.calls++
+	if u.calls <= u.busy {
+		return errors.Join(ErrTryAgain, errors.New("another controller is working the plans"))
+	}
+	return nil
+}
+
+func (u *manyTimesBusy) called() int {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	return u.calls
+}
+
+// **A merge held more than five times is still handed over and recorded** (issue 973, novox/hq ADR 0407 decision
+// 3): the merges have a durable consumer of their own with no max-deliver, so they never go to DEAD_LETTERS to wait
+// for a person; the delay grows, and is capped.
+func TestNatsAMergeHeldMoreThanFiveTimesIsStillHandedOverAndRecorded(t *testing.T) {
+	first, most := MergeHeldFirst, MergeHeldMost
+	MergeHeldFirst, MergeHeldMost = 5*time.Millisecond, 40*time.Millisecond
+	t.Cleanup(func() { MergeHeldFirst, MergeHeldMost = first, most })
+	js := aBus(t)
+	up := &manyTimesBusy{busy: 8}
+	s := &Server{inbound: Nats(js), bus: OverNATS{Conn: js.Conn(), JS: js.Context()}, log: quiet()}
+	if err := s.Follows(up); err != nil {
+		t.Fatal(err)
+	}
+	ctx, stop := context.WithCancel(context.Background())
+	defer stop()
+	go func() { _ = s.Serve(ctx) }()
+	eventually(t, "the merges consumer being made", func() bool {
+		_, err := js.Context().ConsumerInfo("EVENTS", broker.MergesConsumerName)
+		return err == nil
+	})
+
+	body, _ := json.Marshal(SourceMoved{Owner: "novox", Repo: "one", Base: "main", Commit: "abcdef0123"})
+	if _, err := js.Context().Publish(MergedSubject, body); err != nil {
+		t.Fatal(err)
+	}
+	eventually(t, "the ninth hand-over recording the merge", func() bool { return up.called() == 9 })
+	eventually(t, "the merge being acknowledged", func() bool {
+		info, err := js.Context().ConsumerInfo("EVENTS", broker.MergesConsumerName)
+		return err == nil && info.NumAckPending == 0 && info.NumPending == 0
+	})
+	if info, err := js.Context().StreamInfo(broker.DeadLettersStream); err == nil && info.State.Msgs != 0 {
+		t.Fatalf("%d message(s) in dead letters: a merge was given up on", info.State.Msgs)
+	}
+}
+
+// The delay a held merge waits grows from the first to the most and stays there.
+func TestAHeldMergesDelayGrowsAndIsCapped(t *testing.T) {
+	var got []time.Duration
+	for _, held := range []time.Duration{0, 4 * time.Second, 10 * time.Second, 25 * time.Second, 70 * time.Second,
+		200 * time.Second, 20 * time.Minute} {
+		got = append(got, MergeRedeliverAfter(held))
+	}
+	want := []time.Duration{5 * time.Second, 5 * time.Second, 10 * time.Second, 25 * time.Second, 70 * time.Second,
+		200 * time.Second, 5 * time.Minute}
+	for i := range want {
+		if got[i] != want[i] {
+			t.Errorf("held %d: waits %v; want %v", i, got[i], want[i])
+		}
+	}
+}

internal/link/serve.go

@@ -636,8 +636,25 @@ func short(commit string) string {
 	return commit
 }
 
-// sourceMoved acts on the forge's announcement of a merge. Taken whatever happens: a build that
+// MergeHeldFirst is how long a merge the controller could not record yet waits before the bus hands it over
-// fails is reported by the build itself, and re-delivering the merge would only re-fail it.
+// again, and MergeHeldMost the longest it ever waits: the delay grows with how long the merge has been held, from
+// the first to the most (novox/hq ADR 0407 decision 3). Variables so a test need not wait.
+var (
+	MergeHeldFirst = 5 * time.Second
+	MergeHeldMost  = 5 * time.Minute
+)
+
+// MergeRedeliverAfter is the delay for a merge held for the time given: that time, so the delay about doubles
+// with each hold, between MergeHeldFirst and MergeHeldMost.
+func MergeRedeliverAfter(heldFor time.Duration) time.Duration {
+	return min(max(heldFor, MergeHeldFirst), MergeHeldMost)
+}
+
+// sourceMoved acts on the forge's announcement of a merge. **A merge is never dropped** (issue 921, novox/hq ADR
+// 0407 decision 3): one the controller could not record for now (the plans held by another controller, the store
+// away: ErrTryAgain) is held and handed over again until it is recorded; the controller giving up on it would
+// lose a change for good. Anything else is an answer and is taken: a build that fails is reported by the build
+// itself, and re-delivering the merge would only re-fail it.
 func (s *Server) sourceMoved(ctx context.Context, m Control) {
 	var moved SourceMoved
 	if err := json.Unmarshal(m.Body(), &moved); err != nil {
@@ -650,7 +667,25 @@ func (s *Server) sourceMoved(ctx context.Context, m Control) {
 		_ = m.Took()
 		return
 	}
-	if err := s.upgrader.SourceMoved(ctx, moved); err != nil {
+	err := s.upgrader.SourceMoved(ctx, moved)
+	// Left unsettled when the controller is ending, whatever error that gave (the plans' hold returns a bare
+	// context.Canceled): the bus hands it to whatever consumes next, and taking it would lose the merge.
+	if err != nil && ctx.Err() != nil {
+		return
+	}
+	if err != nil && errors.Is(err, ErrTryAgain) {
+		// Also left unsettled when the bus will not hold it.
+		delay := MergeRedeliverAfter(m.HeldFor())
+		if held := m.Hold(delay); held != nil {
+			s.log.Printf("%s/%s moved to %.8s and could not be held (%v): %v; left for the bus to hand over again",
+				moved.Owner, moved.Repo, moved.Commit, held, err)
+			return
+		}
+		s.log.Printf("%s/%s moved to %.8s and is held: %v; handed over again in %s", moved.Owner, moved.Repo,
+			moved.Commit, err, delay)
+		return
+	}
+	if err != nil {
 		s.log.Printf("%s/%s moved to %.8s and the mesh could not act on it: %v", moved.Owner, moved.Repo, moved.Commit, err)
 	}
 	_ = m.Took()

internal/link/store_window_test.go

@@ -106,3 +106,61 @@ func TestAMessageHandledDuringShutdownIsLeftForTheBus(t *testing.T) {
 		t.Fatalf("a build result handled during shutdown was settled, and so lost: %+v", *to)
 	}
 }
+
+// A merge the controller could not record for now (the plans held by another controller, the store away) is held
+// and handed over again, never taken: the bus keeps it until it is recorded (issue 921). Any other failure is an
+// answer and is taken, as before.
+func TestAMergeThatCouldNotBeRecordedNowIsHeldNotDropped(t *testing.T) {
+	merged := SourceMoved{Owner: "novox", Repo: "one", Base: "main", Commit: "abcdef0123"}
+	for _, c := range []struct {
+		what string
+		err  error
+		want func(*settled) bool
+	}{
+		{"the plans held by another controller", errors.Join(ErrTryAgain, errors.New("another controller is working the plans")),
+			func(s *settled) bool { return s.unsettled() }},
+		{"the store away", errors.Join(ErrTryAgain, restarting), func(s *settled) bool { return s.unsettled() }},
+		{"cannot act", errors.New("no such module"), func(s *settled) bool { return s.acked && !s.nacked }},
+		{"kept", nil, func(s *settled) bool { return s.acked && !s.nacked }},
+	} {
+		s, in := serving()
+		s.upgrader = upgradesWith{err: c.err}
+		to := &settled{}
+		held := in.sends(t, to, KindSourceMoved, merged).(*fakeControl)
+		s.act(context.Background(), held)
+		if !c.want(to) {
+			t.Errorf("%s: a merge was settled as %+v", c.what, *to)
+		}
+		if c.err != nil && errors.Is(c.err, ErrTryAgain) && len(in.held) != 1 {
+			t.Errorf("%s: the merge was not held to be handed over again", c.what)
+		}
+	}
+}
+
+// A merge being held when the controller is replaced (its context ended) or when the bus will not hold it is left
+// unsettled, never taken: the bus hands it to whatever consumes next (review of issue 921).
+func TestAMergeIsLeftUnsettledWhenTheControllerEndsOrTheHoldFails(t *testing.T) {
+	merged := SourceMoved{Owner: "novox", Repo: "one", Base: "main", Commit: "abcdef0123"}
+	try := errors.Join(ErrTryAgain, errors.New("another controller is working the plans"))
+
+	// The plans' hold returns a bare context.Canceled when the controller is replaced, not ErrTryAgain.
+	s, in := serving()
+	s.upgrader = upgradesWith{err: context.Canceled}
+	to := &settled{}
+	ended, end := context.WithCancel(context.Background())
+	end()
+	s.act(ended, in.sends(t, to, KindSourceMoved, merged))
+	if !to.unsettled() {
+		t.Errorf("a merge was settled as %+v when the controller's context ended", *to)
+	}
+
+	s, in = serving()
+	s.upgrader = upgradesWith{err: try}
+	to = &settled{}
+	m := in.sends(t, to, KindSourceMoved, merged).(*fakeControl)
+	m.holdFails = errors.New("the bus will not hold it")
+	s.act(context.Background(), m)
+	if !to.unsettled() {
+		t.Errorf("a merge was settled as %+v when it could not be held", *to)
+	}
+}

testdata/store-tests

@@ -8,6 +8,7 @@
 ./cmd/mesh-controller TestAStopWithoutAMergeLeavesItOutOfABatch
 ./cmd/mesh-controller TestSavingPlansAndMovingMergesIsOneAct
 ./cmd/mesh-controller TestAPlanFollowedFromTheStoreAnswersWhenItMoves
+./cmd/mesh-controller TestIssue969TheGateRaisesEachMachineAsItsEngineStatedItself
 ./internal/inventory TestTheLastSendSaysWhatTheMachineAnswered
 ./internal/inventory TestAReportIsTheLastSendsAnswerOnlyWhenItIsAboutThatSend
 ./cmd/mesh-controller TestNodeShowAndStatusReadTheLastSendFromTheStore
@@ -127,3 +128,8 @@
 ./cmd/mesh-controller TestAGroupsCompositionIntoAReleaseBranchIsJudgedByItsOwnController
 ./cmd/mesh-controller TestAGroupsCompositionJudgedWithTheValidatorIntoAReleaseBranchIsJudgedByItsOwnController
 ./cmd/mesh-controller TestAssigningTheClientIssuesItsAccount
+./cmd/mesh-controller TestAnUnrelatedBatchWalksBesideAnOpenWalk
+./cmd/mesh-controller TestARelatedBatchWaitsBehindTheOpenWalk
+./cmd/mesh-controller TestAnOwnPathBatchNeverWalksBeside
+./cmd/mesh-controller TestMergesHeardWhileAnotherHolderHasThePlansAreRecordedAndCarried
+./cmd/mesh-controller TestABatchWindowClosesByTheExpiryOfItsKey