A gate judges its own send and the build it sent, and never puts the controller back behind its store (hq issue 352)

On 2026-10-09 a release's gate on the control node read the machine's
report against a newer send another plan had just made there, failed
three builds the machine had reported healthy, and put them back on
every machine to a controller older than the store's schema; that
controller then passed the newer plan's gate from its own health.

- A gate keeps what its send carried (digest, sequence) and reads the
  report against it; a report on the last send is on it too.
- A gate judges only the build the machine was last sent: another build
  there supersedes the judging — no verdict, nothing put back.
- A controller is told its build (MESH_CONTROLLER_VERSION, ${version}
  in a process's env) and records how far it reads the store's schema;
  a put-back to a build that reaches less, or never said, is refused
  and the current build kept, said as urgent.
- A release's open gate holds other sends of its modules there, and a
  plan's own first send waits on it.

cmd/mesh-controller/gate.go

@@ -87,6 +87,9 @@ const (
 	healthWaiting
 	healthNotYet
 	healthBroken
+	// healthSuperseded is a judging that cannot go on: the machine was sent another build of the module
+	// after the gate's send (novox/hq issue 352). No verdict on the build judged, and nothing put back.
+	healthSuperseded
 )
 
 // served is what one machine's node tools answered the bus's discovery with.
@@ -122,6 +125,39 @@ type gateFacts struct {
 	// groupsAdded is, per module, whether the move judged puts an account in a group its previous build did
 	// not (issue 318 review): the only move whose wait for a new login is excused.
 	groupsAdded map[string]bool
+	// sent is, per machine, the declaration the gate's own send carried there (novox/hq issue 352): a
+	// report is held against it, never against the send made last. sentBuilds is what each machine was
+	// last sent of every module, and judged the commit of each module this gate judges: a machine last
+	// sent another build of the module is not running the build judged.
+	sent       map[string]inventory.SentDeclaration
+	sentBuilds map[string]map[string]string
+	commits    map[string]string
+}
+
+// reportedOn says a machine's last report is on what the gate sent it (novox/hq issue 352): on that
+// declaration, or one it was sent after it — or, for a gate kept before sends were kept on it, on the
+// declaration last sent. On 2026-10-09 a release's gate read the control node's report against a newer
+// send another plan had just made there, and failed three builds the machine had reported healthy as
+// "has not reported on what it was sent".
+func (f gateFacts) reportedOn(machine string, r inventory.Reported) bool {
+	if sent, kept := f.sent[machine]; kept {
+		return sent.ReportsOn(r)
+	}
+	return r.Current
+}
+
+// supersededOn says the machine was last sent another build of the module than the one this gate judges
+// (novox/hq issue 352): the judging cannot go on, whatever the machine reports. On 2026-10-09 a controller
+// put back by one gate judged another gate's newer controller build passed on the same machine, reading
+// the put-back build's health as the newer one's.
+func (f gateFacts) supersededOn(module, machine string) (string, bool) {
+	judged, known := f.commits[module]
+	sent, has := f.sentBuilds[machine][module]
+	if !known || !has || judged == "" || sent == "" || sameCommit(sent, judged) {
+		return "", false
+	}
+	return fmt.Sprintf("%s was sent %s %s after this gate's %s: the build judged no longer runs there, and "+
+		"this judging is superseded by that send's", machine, module, short(sent), short(judged)), true
 }
 
 // gatherGateFacts reads what a judging needs, from the store, the bus and this controller's memory. A
@@ -199,9 +235,12 @@ func judgeHealth(module, component string, m catalogue.Manifest, machine string,
 		return healthBroken, fmt.Sprintf("the witness on %s judged the %s %s and %s: %s", machine, r.Component,
 			short(r.From), r.Outcome, r.Why)
 	}
+	if why, superseded := f.supersededOn(module, machine); superseded {
+		return healthSuperseded, why
+	}
 	r, said := f.reports[machine]
 	switch {
-	case !said || r.At == nil || !r.Current:
+	case !said || r.At == nil || !f.reportedOn(machine, r):
 		return healthNotYet, fmt.Sprintf("%s has not reported on what it was sent", machine)
 	case r.Outcome == inventory.OutcomeFailed || r.Outcome == inventory.OutcomeRefused:
 		return healthBroken, fmt.Sprintf("%s %s what it was sent", machine, r.Outcome)
@@ -438,6 +477,21 @@ func judgeMoves(ctx context.Context, open *stores, g *inventory.PlanGate, pairs
 		return "", err
 	}
 	facts.groupsAdded = movesAddingGroups(ctx, open.inventory, g, pairs, shelf)
+	facts.sent = g.Sent
+	facts.commits, facts.sentBuilds = judgedCommits(g, pairs), map[string]map[string]string{}
+	// A module this gate put back at once (putBackBroken) was sent its earlier build by the gate itself:
+	// not another send, and not a judging superseded.
+	for _, m := range g.Returned {
+		delete(facts.commits, m)
+	}
+	for _, j := range pairs {
+		if _, read := facts.sentBuilds[j.node]; read {
+			continue
+		}
+		if builds, known, err := open.inventory.SentBuilds(ctx, j.node); err == nil && known {
+			facts.sentBuilds[j.node] = builds
+		}
+	}
 	// **What is wrong with a machine itself is the machine's** (novox/hq issue 281): read once for each
 	// machine judged, apart from what is wrong with a module there, and never pinned on the module the
 	// gate happens to be kept on.
@@ -474,6 +528,13 @@ func judgeMoves(ctx context.Context, open *stores, g *inventory.PlanGate, pairs
 		if _, seen := reading[j.module]; !seen {
 			modules = append(modules, j.module)
 		}
+		if h == healthSuperseded {
+			// Decided at once (novox/hq issue 352): nothing of this gate's can be judged on a machine that
+			// was sent another build of it, and nothing is put back — the later send is what runs there.
+			g.Failing, g.Last = nil, ""
+			decide(g, inventory.GateSuperseded, said, now)
+			return g.Verdict, nil
+		}
 		if h == healthBroken && !slices.Contains(g.Broken, j.module) {
 			g.Broken = append(g.Broken, j.module)
 			if g.BrokenWhy == "" {
@@ -577,6 +638,40 @@ func judgeMoves(ctx context.Context, open *stores, g *inventory.PlanGate, pairs
 	return g.Verdict, nil
 }
 
+// judgedCommits is the commit of each module a gate judges: the gate's own To for its module, and each
+// carried move's. Pure.
+func judgedCommits(g *inventory.PlanGate, pairs []judged) map[string]string {
+	out := map[string]string{}
+	for _, c := range g.Carried {
+		if c.To != "" {
+			out[c.Module] = c.To
+		}
+	}
+	if g.To != "" {
+		for _, j := range pairs {
+			if _, has := out[j.module]; !has && !slices.ContainsFunc(g.Carried, func(c inventory.CarriedMove) bool { return c.Module == j.module }) {
+				out[j.module] = g.To
+			}
+		}
+	}
+	return out
+}
+
+// sentNow is what each machine was just sent, read after a send for the gate to keep (novox/hq issue
+// 352): a machine whose send is not on record is left out, and its report is read as before.
+func sentNow(ctx context.Context, inv *inventory.Inventory, machines []string) map[string]inventory.SentDeclaration {
+	out := map[string]inventory.SentDeclaration{}
+	for _, n := range machines {
+		if s, found, err := inv.SentTo(ctx, n); err == nil && found {
+			out[n] = s
+		}
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
 // whyFor is a passing gate's why as one module's verdict says it: the send's, and that module's own wait
 // for a person, never another's (issue 318 review).
 func whyFor(g *inventory.PlanGate, module string) string {
@@ -802,6 +897,16 @@ func gateFailed(ctx context.Context, open *stores, p *inventory.Plan, module str
 			"back", module, short(state.Previous), inventory.KeptBuilds))
 		return
 	}
+	if g.Component == lease.ComponentController {
+		// **The controller is never put back to a build older than the store's schema** (novox/hq issue
+		// 352): the build before it carries fewer migrations than the failed one applied, starts behind
+		// its own records, and judges the next gate with what it can read. The current build is kept and
+		// the condition says so; a person decides.
+		if why, ok := controllerSchemaAllows(ctx, inv, previous); !ok {
+			notBack(why)
+			return
+		}
+	}
 	if err := inv.RestoreModule(ctx, previous); err != nil {
 		notBack(err.Error())
 		return
@@ -835,6 +940,53 @@ func gateFailed(ctx context.Context, open *stores, p *inventory.Plan, module str
 	sayRollback(ctx, open, module, g, "")
 }
 
+// controllerSchemaAllows says the store's schema lets this build of the controller be put back: the
+// build recorded, when it served, a reach at or past the highest migration the store has applied. One
+// that never recorded a reach is not proved safe, and is refused as such (novox/hq issue 352). Why
+// says what is kept and why when it is not.
+func controllerSchemaAllows(ctx context.Context, inv *inventory.Inventory, previous inventory.Build) (string, bool) {
+	applied, err := inv.SchemaApplied(ctx)
+	if err != nil {
+		return "what the store's schema reaches cannot be read, so whether the build before it can read it is not " +
+			"known; the current build is kept: " + err.Error(), false
+	}
+	build := versionOfBuild(previous)
+	if build == "" {
+		return fmt.Sprintf("the build before it (%s) names no bundle to know it by, so whether it can read the store's "+
+			"schema (migration %04d) is not known; the current build is kept, and a person decides", short(previous.Commit), applied), false
+	}
+	reach, known, err := inv.SchemaReachOf(ctx, build)
+	if err != nil {
+		return "what the build before it knows of the store's schema cannot be read; the current build is kept: " + err.Error(), false
+	}
+	if !known {
+		return fmt.Sprintf("the build before it (%s, %s) never recorded how far it reads the store's schema — a "+
+			"controller records that when it serves — so it is not proved to read migration %04d, which the store "+
+			"has applied; a controller older than its store starts behind its own records and judges with what it "+
+			"can read, so the current build is kept, and a person decides", short(previous.Commit), build, applied), false
+	}
+	if reach < applied {
+		return fmt.Sprintf("the build before it (%s, %s) reads the store's schema up to migration %04d, and the store "+
+			"is at %04d: a controller older than its store starts behind its own records and judges with what it "+
+			"can read, so the current build is kept, and a person decides", short(previous.Commit), build, reach, applied), false
+	}
+	return "", true
+}
+
+// versionOfBuild is the version a build's bundle is delivered as — its archive's digest, short, as the
+// catalogue names it (`${version}`) — read from the build's artifacts; empty when none is a bundle.
+func versionOfBuild(b inventory.Build) string {
+	for _, a := range b.Made {
+		if a.Kind != catalogue.ArtifactBundle && a.Kind != catalogue.ArtifactArchive {
+			continue
+		}
+		if _, hex, found := strings.Cut(a.Reference, "sha256:"); found && len(hex) >= 12 {
+			return hex[:12]
+		}
+	}
+	return ""
+}
+
 // rollbacks is what a failed send puts back, sent together (novox/hq issue 281): a gate that judged one
 // send judges what it moved as one, and what it found wanting goes back in one send per machine — not
 // in a send for each module, which is the churn that failed the gate in the first place.