Cut a closed batch beside the open walks it shares nothing with, queue a related one with the reason, fold only related waiting walks (issue 973)

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
-//     superseded; one still waiting for its delivery's word is folded into the next batch's walk, which
-//     answers every merge it answered — the only supersession left.
+//   - One walk at a time beside a related one (ADR 0407 decision 4, issue 973). Merges that come while a walk
+//     runs join the next batch; a closed batch that shares nothing with an open walk is cut beside it. A walk
+//     that started is never superseded; one still waiting for its delivery's word is folded into the walk of a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
-// a time: such a batch is cut when no walk is open, and queued behind the one that is otherwise.
+// group has no member still to merge is a batch whose window is closed at once (batchToJoin). Such a batch is cut
+// 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
@@ -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
-//     queued behind it as any closed batch is.
+//     once means only that: with a walk open that it is related to the batch is queued behind it as any closed
+//     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
@@ -574,13 +583,13 @@ 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 {
@@ -761,9 +770,9 @@ func batchCutter(ctx context.Context, open *stores) {
 }
 
 // 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
-// closed batch cut into one walk with the waiting walks folded in; with a walk open, a closed batch said
-// queued behind it.
+// for the merge that brought the failure; with no walk open, the next merge of that search walked alone; else
+// each closed batch cut into a walk with the waiting walks it is related to folded in, unless it is related to a
+// 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 +789,7 @@ func cutBatchesHeld(ctx context.Context, open *stores, now time.Time) error {
 	if err != nil {
 		return err
 	}
-	var started *inventory.Plan
-	var waiting []inventory.Plan
+	var started, waiting []inventory.Plan
 	for i := range plans {
 		p := plans[i]
 		if p.Release != nil {
@@ -789,8 +797,8 @@ func cutBatchesHeld(ctx context.Context, open *stores, now time.Time) error {
 		}
 		if p.Waiting() {
 			waiting = append(waiting, p)
-		} else if started == nil {
-			started = &plans[i]
+		} else {
+			started = append(started, p)
 		}
 	}
 	batches, err := inv.Batches(ctx)
@@ -805,46 +813,113 @@ 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 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 {
+			if err := keepBatch(ctx, inv, b, now, ""); 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
+		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 +1471,11 @@ 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",