Assemble merges in a rolling window and walk each batch once (hq ADR 0276, issue 362)

Every merge opened a walk and the next merge of the branch superseded it: two catalogue merges 18 s apart left a walk no delivery held, and the operator started it by hand 58 minutes later. A merge now joins the open batch, kept in the store (migration 0089), which is cut into one walk when no merge came for merge-window (90 s) or at merge-window-at-most (10 min): one commit per repository, the latest of its branch, with every file the batch's merges changed. One walk at a time; a started walk is never superseded, a waiting one is folded into the next. The walk names every merge it answers on the wire (delivery.merges, taken_over_by, batch). A failed walk walks its earlier merges alone, newest first, until one is delivered. A delivery group's order becomes tier edges inside the walk. plans shows the batch assembling; S18 and S19 bound its waits; S16 names the merges a waiting walk answers.
2026-10-10 13:20:04 +02:00
parent 8a53532d89
commit 96fc4209d3
24 changed files with 2110 additions and 563 deletions
@@ -0,0 +1,537 @@
+package main
+
+import (
+	"bytes"
+	"io"
+	"os"
+	"slices"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// The suite's tests of a merge, but for the merge window's own (this file's), plan the merge as it is heard: a
+// window of nothing closes at once, so a merge with no walk open is cut into its walk at once, as every merge
+// was planned before novox/hq ADR 0276. The window's tests set it through the controller's settings.
+func init() { mergeWindowDefault = 0 }
+
+// windowed is a mesh whose controller's settings give a merge window of 90 seconds and at most 10 minutes, as its
+// settings file says them, and
+// modules built from the catalogue (app, notes) and from three repositories of their own.
+func windowed(t *testing.T) *stores {
+	t.Helper()
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	settings := t.TempDir() + "/merge-window.json"
+	if err := os.WriteFile(settings, []byte(`{"merge-window": "90s", "merge-window-at-most": "10m"}`), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv(mergeWindowFileVar, settings)
+	if err := inv.RegisterModule(ctx, catalogue.Manifest{Module: "mesh-controller", Version: "1"},
+		inventory.Source{Repository: "novox/mesh-controller", Seat: "git", Ref: "main", BuiltFrom: "c0",
+			Head: "c0"}); err != nil {
+		t.Fatal(err)
+	}
+	for _, m := range []struct{ module, repository, path string }{
+		{"app", "novox/mesh-catalog", "modules/app"},
+		{"notes", "novox/mesh-catalog", "modules/notes"},
+		{"one", "novox/one", ""},
+		{"two", "novox/two", ""},
+		{"three", "novox/three", ""},
+	} {
+		if err := inv.RegisterModule(ctx, catalogue.Manifest{Module: m.module, Version: "1"},
+			inventory.Source{Repository: m.repository, Seat: "git", Path: m.path, Ref: "main", BuiltFrom: "c0",
+				Head: "c0"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	return open
+}
+
+// t0 is the moment a test's clock starts: merges are dated by it, and heard after it.
+var t0 = time.Now().UTC().Truncate(time.Second)
+
+// catalogueMerge is a merge of the catalogue changing one module's directory, made at a moment.
+func catalogueMerge(commit, module string, made time.Time) link.SourceMoved {
+	return link.SourceMoved{Owner: "novox", Repo: "mesh-catalog", Base: "main", Commit: commit,
+		MergedAt: made.Format(time.RFC3339Nano), Paths: []string{"modules/" + module + "/module.json"},
+		ModuleDirs: []string{"modules/" + module}, ModuleDirsSaid: true}
+}
+
+// repoMerge is a merge of a repository of one module's own.
+func repoMerge(repo, commit string, made time.Time) link.SourceMoved {
+	return link.SourceMoved{Owner: "novox", Repo: repo, Base: "main", Commit: commit,
+		MergedAt: made.Format(time.RFC3339Nano), Paths: []string{"main.go"}}
+}
+
+// hear is a merge heard at a moment.
+func hear(t *testing.T, open *stores, m link.SourceMoved, now time.Time) {
+	t.Helper()
+	if err := (following{open: open}).hearMerge(t.Context(), m, now); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// cutAt is the cutter looking at a moment.
+func cutAt(t *testing.T, open *stores, now time.Time) {
+	t.Helper()
+	if err := cutBatchesHeld(t.Context(), open, now); err != nil {
+		t.Fatal(err)
+	}
+}
+
+// walks is every plan record that is a walk, newest first, and the batches not yet cut.
+func walks(t *testing.T, open *stores) (walks, batches []inventory.Plan) {
+	t.Helper()
+	recent, err := open.inventory.RecentPlans(t.Context(), 50)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, p := range recent {
+		if p.Batch() {
+			batches = append(batches, p)
+		} else {
+			walks = append(walks, p)
+		}
+	}
+	return walks, batches
+}
+
+// The replay of issue 362: two catalogue merges 18 seconds apart are one batch, cut once into one walk at the
+// later commit, which names both merges — the earlier carried by the later.
+func TestTwoMergesSecondsApartAreOneWalkAtTheLaterCommit(t *testing.T) {
+	open := windowed(t)
+	asked := asksWithPaths(t)
+	claude := catalogueMerge("553b7191claude", "app", t0)
+	dunst := catalogueMerge("48bda475dunst", "notes", t0.Add(17*time.Second))
+	hear(t, open, claude, t0.Add(time.Second))
+	hear(t, open, dunst, t0.Add(18*time.Second))
+	if ws, bs := walks(t, open); len(ws) != 0 || len(bs) != 1 {
+		t.Fatalf("within the window: %d walk(s), %d batch(es)", len(ws), len(bs))
+	}
+	cutAt(t, open, t0.Add(18*time.Second+89*time.Second))
+	if ws, _ := walks(t, open); len(ws) != 0 {
+		t.Fatalf("cut before the window closed: %+v", ws)
+	}
+	cutAt(t, open, t0.Add(18*time.Second+90*time.Second))
+	ws, bs := walks(t, open)
+	if len(ws) != 1 || len(bs) != 0 {
+		t.Fatalf("after the window: %d walk(s), %d batch(es)", len(ws), len(bs))
+	}
+	w := ws[0]
+	if w.Commit != dunst.Commit || len(w.Commits) != 1 || w.Commits[0].Commit != dunst.Commit {
+		t.Fatalf("the walk is at %s %+v, not the later commit", w.Commit, w.Commits)
+	}
+	for _, m := range []string{"app", "notes"} {
+		if _, in := w.Modules[m]; !in {
+			t.Fatalf("the walk does not build %s, which one of its merges moved: %v", m, w.Modules)
+		}
+	}
+	want := []inventory.PlanMerge{{Repository: "novox/mesh-catalog", Commit: claude.Commit, Carried: dunst.Commit},
+		{Repository: "novox/mesh-catalog", Commit: dunst.Commit}}
+	if w.Delivery == nil || !slices.Equal(w.Delivery.Merges, want) {
+		t.Fatalf("the walk answers %+v, want %+v", w.Delivery, want)
+	}
+	if len(*asked) != 2 {
+		t.Fatalf("the walk asked %v", *asked)
+	}
+	// Heard again, as the bus may hand it over twice: never a second merge, nor a second walk.
+	hear(t, open, claude, t0.Add(5*time.Minute))
+	if ws, bs := walks(t, open); len(ws) != 1 || len(bs) != 0 {
+		t.Fatalf("a merge heard twice made %d walk(s) and %d batch(es)", len(ws), len(bs))
+	}
+}
+
+// Merges of three repositories in one window are one walk, one commit for each.
+func TestThreeRepositoriesInOneWindowAreOneWalk(t *testing.T) {
+	open := windowed(t)
+	asksRecorded(t)
+	for i, r := range []string{"one", "two", "three"} {
+		hear(t, open, repoMerge(r, "c-"+r, t0.Add(time.Duration(i)*20*time.Second)),
+			t0.Add(time.Duration(i)*20*time.Second+time.Second))
+	}
+	cutAt(t, open, t0.Add(41*time.Second+90*time.Second))
+	ws, bs := walks(t, open)
+	if len(ws) != 1 || len(bs) != 0 {
+		t.Fatalf("%d walk(s), %d batch(es)", len(ws), len(bs))
+	}
+	w := ws[0]
+	if len(w.Commits) != 3 || len(w.Delivery.Merges) != 3 {
+		t.Fatalf("the walk carries %+v and answers %+v", w.Commits, w.Delivery.Merges)
+	}
+	for _, r := range []string{"one", "two", "three"} {
+		if w.CommitOf("novox/"+r) != "c-"+r {
+			t.Fatalf("the walk carries %s at %q", r, w.CommitOf("novox/"+r))
+		}
+		if _, in := w.Modules[r]; !in {
+			t.Fatalf("the walk does not build %s: %v", r, w.Modules)
+		}
+	}
+	// Each module's ask is recorded at its own repository's commit.
+	for _, r := range []string{"one", "two", "three"} {
+		q, found, err := open.inventory.BuildRequestByID(t.Context(), w.Modules[r].Build)
+		if err != nil || !found || q.Commit != "c-"+r {
+			t.Fatalf("%s's build was not recorded at its own commit: %+v %v %v", r, q, found, err)
+		}
+	}
+}
+
+// Each merge restarts the window: a merge at second 80 keeps the batch open until second 170; the window
+// closes at most ten minutes after its first merge however busy.
+func TestAMergeRestartsTheWindowAndTheMaximumClosesIt(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	hear(t, open, repoMerge("two", "c2", t0.Add(80*time.Second)), t0.Add(80*time.Second))
+	cutAt(t, open, t0.Add(100*time.Second))
+	if ws, _ := walks(t, open); len(ws) != 0 {
+		t.Fatal("the window closed 90 seconds after its first merge, not after its last")
+	}
+	cutAt(t, open, t0.Add(170*time.Second))
+	if ws, _ := walks(t, open); len(ws) != 1 || len(ws[0].Commits) != 2 {
+		t.Fatalf("the window did not close 90 seconds after its last merge: %+v", ws)
+	}
+
+	busy := windowed(t)
+	asksWithPaths(t)
+	var last time.Time
+	for i := 0; i < 12; i++ {
+		last = t0.Add(time.Duration(i) * time.Minute)
+		repo := []string{"one", "two", "three"}[i%3]
+		hear(t, busy, repoMerge(repo, "c"+string(rune('a'+i)), last), last)
+		if ws, _ := walks(t, busy); len(ws) > 0 {
+			if i != 10 {
+				t.Fatalf("a busy window closed at merge %d (minute %d), not at its maximum", i, i)
+			}
+			break
+		}
+	}
+	ws, _ := walks(t, busy)
+	if len(ws) != 1 {
+		t.Fatalf("ten minutes of merges a minute apart were never cut: %d walk(s)", len(ws))
+	}
+}
+
+// `plans` lists the batch being assembled first, in the operator's words: how long is left, when at the latest,
+// what is grouped, and that the plan is not yet calculated. It says so on the bus as plan-moved, too.
+func TestTheAssemblingBatchIsShown(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	var said []inventory.Plan
+	was := inventory.PlanSaved
+	inventory.PlanSaved = func(p inventory.Plan) { said = append(said, p) }
+	t.Cleanup(func() { inventory.PlanSaved = was })
+	now := time.Now().UTC()
+	hear(t, open, catalogueMerge("553b7191claude", "app", now.Add(-20*time.Second)), now.Add(-19*time.Second))
+	hear(t, open, catalogueMerge("48bda475dunst", "notes", now.Add(-2*time.Second)), now.Add(-time.Second))
+	hear(t, open, repoMerge("one", "a6bc0931one", now), now)
+	out := captured(t, func() error { return plansCommand(t.Context(), nil) })
+	first := strings.SplitN(out, "\n", 2)[0]
+	for _, want := range []string{"assembling: ", " s left (at the latest ", "grouped: novox/mesh-catalog@48bda475 " +
+		"(answers 553b7191), novox/one@a6bc0931; plan not yet calculated"} {
+		if !strings.Contains(first, want) {
+			t.Fatalf("plans' first line %q does not say %q", first, want)
+		}
+	}
+	if len(said) == 0 || said[len(said)-1].State != inventory.PlanAssembling || said[len(said)-1].Delivery.Batch == nil ||
+		len(said[len(said)-1].Delivery.Merges) != 3 {
+		t.Fatalf("the batch was not said as assembling with its merges: %+v", said)
+	}
+}
+
+// captured is what a command printed.
+func captured(t *testing.T, run func() error) string {
+	t.Helper()
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatal(err)
+	}
+	was := os.Stdout
+	os.Stdout = w
+	runErr := run()
+	os.Stdout = was
+	_ = w.Close()
+	var b bytes.Buffer
+	_, _ = io.Copy(&b, r)
+	if runErr != nil {
+		t.Fatal(runErr)
+	}
+	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;
+// the walk that started is never superseded.
+func TestAMergeDuringAWalkJoinsTheNextBatch(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	cutAt(t, open, t0.Add(90*time.Second))
+	ws, _ := walks(t, open)
+	if len(ws) != 1 || !ws[0].Open() {
+		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))
+	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 {
+		t.Fatalf("the merge during the walk: %d walk(s), batches %+v", len(ws), bs)
+	}
+	if got, _ := open.inventory.PlanByID(t.Context(), first.ID); !got.Open() {
+		t.Fatalf("the started walk was %s by the next merge", got.State)
+	}
+	if !strings.HasPrefix(batchWords(bs[0], t0.Add(5*time.Minute)), "queued behind "+first.ID) {
+		t.Fatalf("a queued batch reads %q", batchWords(bs[0], t0.Add(5*time.Minute)))
+	}
+	first.State = inventory.PlanDone
+	if err := open.inventory.SavePlan(t.Context(), &first); err != nil {
+		t.Fatal(err)
+	}
+	cutAt(t, open, t0.Add(6*time.Minute))
+	ws, bs = walks(t, open)
+	if len(ws) != 2 || len(bs) != 0 || ws[0].Commit != "c2" {
+		t.Fatalf("the queued batch was not cut when the walk ended: %+v %+v", ws, bs)
+	}
+}
+
+// A walk waiting for its delivery's word is folded into the next batch's walk, which answers its merges; it
+// names the walk that took it over.
+func TestAWaitingWalkIsFoldedIntoTheNextBatch(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	ctx := t.Context()
+	if err := open.inventory.RegisterModule(ctx, catalogue.Manifest{Module: "mesh-delivery", Version: "1",
+		Claims: []catalogue.Claim{{Name: catalogue.DeliverySeat, Scope: catalogue.ScopeMesh}}},
+		inventory.Source{Repository: "novox/mesh-catalog", Seat: "git", Path: "modules/mesh-delivery", Ref: "main",
+			BuiltFrom: "c0", Head: "c0"}); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := open.inventory.Assign(ctx, "anchor", "mesh-delivery"); err != nil {
+		t.Fatal(err)
+	}
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	cutAt(t, open, t0.Add(90*time.Second))
+	ws, _ := walks(t, open)
+	if len(ws) != 1 || !ws[0].Waiting() {
+		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))
+	cutAt(t, open, t0.Add(2*time.Minute+90*time.Second))
+	ws, bs := walks(t, open)
+	if len(bs) != 0 || len(ws) != 2 {
+		t.Fatalf("%d walk(s), %d batch(es)", len(ws), len(bs))
+	}
+	folded, err := open.inventory.PlanByID(ctx, waiting.ID)
+	if err != nil {
+		t.Fatal(err)
+	}
+	newer := ws[0]
+	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" {
+		t.Fatalf("the newer walk answers %+v", newer.Delivery.Merges)
+	}
+	if _, in := newer.Modules["one"]; !in {
+		t.Fatalf("what the folded walk was to build is not built: %v", newer.Modules)
+	}
+}
+
+// A merge heard after a later merge of its branch was walked is answered by that walk when it builds all it
+// moves — named at once on a walk done — and joins the next batch when it does not.
+func TestALateMergeIsAnsweredByTheWalkOfTheLaterOne(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	ctx := t.Context()
+	later := catalogueMerge("48bda475dunst", "notes", t0.Add(17*time.Second))
+	hear(t, open, later, t0.Add(18*time.Second))
+	cutAt(t, open, t0.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)
+	}
+	// Moves only notes, which the done walk built from the branch after it: answered there, at once.
+	hear(t, open, catalogueMerge("553b7191early", "notes", t0), t0.Add(15*time.Minute))
+	got, _ := open.inventory.PlanByID(ctx, w.ID)
+	if got.State != inventory.PlanDone || len(got.Delivery.Merges) != 2 ||
+		got.Delivery.Merges[0] != (inventory.PlanMerge{Repository: "novox/mesh-catalog", Commit: "553b7191early",
+			Carried: later.Commit}) {
+		t.Fatalf("the late merge is not named by the walk that carried it: %+v", got.Delivery.Merges)
+	}
+	// Moves app, which that walk never built: the next batch walks it.
+	hear(t, open, catalogueMerge("1111aaaaapp", "app", t0.Add(time.Second)), t0.Add(16*time.Minute))
+	_, bs := walks(t, open)
+	if len(bs) != 1 || bs[0].Delivery.Merges[0].Commit != "1111aaaaapp" {
+		t.Fatalf("a late merge moving what the walk never built did not join the next batch: %+v", bs)
+	}
+}
+
+// A failed walk marks nothing delivered: its earlier merges are walked alone, on their own commit, the newest
+// first; the first delivered answers the older ones, and the search stops.
+func TestAFailedWalkWalksItsEarlierMergesAlone(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	ctx := t.Context()
+	oldest := catalogueMerge("aaaa0001", "app", t0)
+	middle := catalogueMerge("bbbb0002", "app", t0.Add(10*time.Second))
+	newest := catalogueMerge("cccc0003", "notes", t0.Add(20*time.Second))
+	for i, m := range []link.SourceMoved{oldest, middle, newest} {
+		hear(t, open, m, t0.Add(time.Duration(i)*10*time.Second+time.Second))
+	}
+	cutAt(t, open, t0.Add(2*time.Minute))
+	ws, _ := walks(t, open)
+	failed := ws[0]
+	failed.State, failed.Note = inventory.PlanFailed, "notes failed to build in tier 0"
+	if err := open.inventory.SavePlan(ctx, &failed); err != nil {
+		t.Fatal(err)
+	}
+	cutAt(t, open, t0.Add(3*time.Minute))
+	got, _ := open.inventory.PlanByID(ctx, failed.ID)
+	if got.State != inventory.PlanFailed || len(got.Delivery.Merges) != 3 {
+		t.Fatalf("the failed walk's record changed: %s %+v", got.State, got.Delivery.Merges)
+	}
+	ws, _ = walks(t, open)
+	alone := ws[0]
+	if alone.ID == failed.ID || alone.Commit != middle.Commit || len(alone.Delivery.Merges) != 1 ||
+		alone.Delivery.Merges[0] != (inventory.PlanMerge{Repository: "novox/mesh-catalog", Commit: middle.Commit}) {
+		t.Fatalf("the newest earlier merge was not walked alone on its own commit: %+v", alone)
+	}
+	alone.State = inventory.PlanDone
+	if err := open.inventory.SavePlan(ctx, &alone); err != nil {
+		t.Fatal(err)
+	}
+	cutAt(t, open, t0.Add(4*time.Minute))
+	ws, bs := walks(t, open)
+	if ws[0].ID != alone.ID || len(bs) != 0 {
+		t.Fatalf("the search went on after a merge was delivered: %+v", ws[0])
+	}
+	done, _ := open.inventory.PlanByID(ctx, alone.ID)
+	if len(done.Delivery.Merges) != 2 || done.Delivery.Merges[0] != (inventory.PlanMerge{Repository: "novox/mesh-catalog",
+		Commit: oldest.Commit, Carried: middle.Commit}) {
+		t.Fatalf("the oldest merge is not answered by the walk that delivered the one after it: %+v", done.Delivery.Merges)
+	}
+	// A stopped walk starts no search: a person ended it.
+	hear(t, open, catalogueMerge("dddd0004", "app", t0.Add(5*time.Minute)), t0.Add(5*time.Minute))
+	hear(t, open, catalogueMerge("eeee0005", "notes", t0.Add(5*time.Minute+time.Second)), t0.Add(5*time.Minute+time.Second))
+	cutAt(t, open, t0.Add(8*time.Minute))
+	ws, _ = walks(t, open)
+	if _, err := stopWalk(ctx, open.inventory, ws[0].ID, "mesh-delivery for jochen", "not now"); err != nil {
+		t.Fatal(err)
+	}
+	cutAt(t, open, t0.Add(9*time.Minute))
+	if again, _ := walks(t, open); again[0].ID != ws[0].ID {
+		t.Fatalf("a stopped walk's earlier merge was walked alone: %+v", again[0])
+	}
+}
+
+// A delivery group's order holds inside the walk (ADR 0249 unchanged): merges of two repositories from one head
+// branch name, the node-engine's before the controller's, are tiered so; without the group, one tier.
+func TestAGroupsOrderBecomesTiersInsideTheWalk(t *testing.T) {
+	for _, grouped := range []bool{true, false} {
+		open := windowed(t)
+		asksWithPaths(t)
+		ctx := t.Context()
+		for _, m := range []string{"mesh-host"} {
+			if err := open.inventory.RegisterModule(ctx, catalogue.Manifest{Module: m, Version: "1"},
+				inventory.Source{Repository: "novox/" + m, Seat: "git", Ref: "main", BuiltFrom: "c0", Head: "c0"}); err != nil {
+				t.Fatal(err)
+			}
+		}
+		host := repoMerge("mesh-host", "h1", t0)
+		ctl := repoMerge("mesh-controller", "k1", t0.Add(time.Second))
+		host.Head, ctl.Head = "feat/together", "feat/together"
+		if !grouped {
+			ctl.Head = "feat/alone"
+		}
+		hear(t, open, ctl, t0.Add(2*time.Second))
+		hear(t, open, host, t0.Add(3*time.Second))
+		cutAt(t, open, t0.Add(2*time.Minute))
+		ws, _ := walks(t, open)
+		if len(ws) != 1 {
+			t.Fatalf("%d walks", len(ws))
+		}
+		tiers := ws[0].Tiers
+		if grouped {
+			if len(tiers) != 2 || !slices.Equal(tiers[0], []string{"mesh-host"}) ||
+				!slices.Equal(tiers[1], []string{"mesh-controller"}) {
+				t.Fatalf("the group's order is not the walk's tiers: %v", tiers)
+			}
+		} else if len(tiers) != 1 {
+			t.Fatalf("two merges of no group were ordered: %v", tiers)
+		}
+	}
+}
+
+// A restarted controller resumes the window where it stood: the batch, its merges and their times are in the
+// store, a merge handed over again is not doubled, and the batch is cut when its window closes.
+func TestABatchSurvivesARestart(t *testing.T) {
+	open := windowed(t)
+	asksWithPaths(t)
+	hear(t, open, repoMerge("one", "c1", t0), t0)
+	hear(t, open, repoMerge("two", "c2", t0.Add(30*time.Second)), t0.Add(30*time.Second))
+
+	// A new controller: nothing of the first one's but the store.
+	again, err := openStores(t.Context())
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(again.Close)
+	hear(t, again, repoMerge("one", "c1", t0), t0.Add(60*time.Second)) // the bus hands it over again
+	cutAt(t, again, t0.Add(119*time.Second))
+	if ws, bs := walks(t, again); len(ws) != 0 || len(bs) != 1 || len(bs[0].Delivery.Merges) != 2 {
+		t.Fatalf("the restarted controller's batch: %d walk(s), %+v", len(ws), bs)
+	}
+	cutAt(t, again, t0.Add(120*time.Second))
+	ws, bs := walks(t, again)
+	if len(ws) != 1 || len(bs) != 0 || len(ws[0].Delivery.Merges) != 2 {
+		t.Fatalf("the restarted controller did not cut the batch at its window: %+v %+v", ws, bs)
+	}
+}
+
+// The window's settings reach the controller in its settings file, read at every look; one that says no
+// duration, or no file, is the default.
+func TestTheWindowIsTheControllersSetting(t *testing.T) {
+	file := t.TempDir() + "/merge-window.json"
+	t.Setenv(mergeWindowFileVar, file)
+	for _, c := range []struct {
+		content        string
+		window, atMost time.Duration
+	}{
+		{`{"merge-window": "60s", "merge-window-at-most": "5m"}`, time.Minute, 5 * time.Minute},
+		{`{"merge-window": 30, "merge-window-at-most": ""}`, 30 * time.Second, mergeWindowAtMostDefault},
+		{`{"merge-window": "soon"}`, mergeWindowDefault, mergeWindowAtMostDefault},
+	} {
+		if err := os.WriteFile(file, []byte(c.content), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		if w, m := mergeWindowOf(); w != c.window || m != c.atMost {
+			t.Errorf("%s reads as %s and %s, want %s and %s", c.content, w, m, c.window, c.atMost)
+		}
+	}
+	t.Setenv(mergeWindowFileVar, file+".gone")
+	if w, m := mergeWindowOf(); w != mergeWindowDefault || m != mergeWindowAtMostDefault {
+		t.Errorf("no file reads as %s and %s", w, m)
+	}
+}
+
+// S16 names every merge a waiting walk answers, never one commit a supersession may have ended (issue 362).
+func TestAWaitingWalkNamesTheMergesItAnswers(t *testing.T) {
+	f := calm(t0)
+	f.waits = []waitFacts{{id: "plan-2", repository: "novox/mesh-catalog", commit: "48bda475dunst", awaits: "mesh-delivery",
+		since: t0.Add(-31 * time.Minute), merges: []inventory.PlanMerge{
+			{Repository: "novox/mesh-catalog", Commit: "553b7191claude", Carried: "48bda475dunst"},
+			{Repository: "novox/mesh-catalog", Commit: "48bda475dunst"}}}}
+	got := watchWaits(f)
+	if len(got) != 1 || !strings.Contains(got[0].Summary, "novox/mesh-catalog@553b7191") ||
+		!strings.Contains(got[0].Summary, "novox/mesh-catalog@48bda475") {
+		t.Fatalf("the waiting walk does not name both merges: %+v", got)
+	}
+}