Hold a consumer's findings under its unhealthy provider, and say them once there (hq ADR 0240, to-be 48 Phase C)

With twelve consumers of the database provision, one provider down would be
twelve conditions for one fault and twelve gates failed for something none of
them did. A consumer's check names the provision it exercises; while the
provider composed for it — its recorded binding, or the machine its credential
comes from — is unhealthy on the record, what that check finds raises nothing
of its own: the provider's condition lists it as waiting and is urgent, and
the consumer's gate waits, past its bound too, rather than putting a build
back. Liveness findings and checks naming no provision stay the consumer's own,
and once the provider is healthy a consumer still failing is raised at once.

cmd/mesh-controller/check.go

@@ -8,6 +8,7 @@ import (
 	"path/filepath"
 	"sort"
 	"strings"
+	"time"
 
 	"github.com/novox/mesh-controller/internal/catalogue"
 )
@@ -107,6 +108,28 @@ func moduleCheckFor(paths []string, longestMachine int, out io.Writer) error {
 		names = append(names, name)
 	}
 	sort.Strings(names)
+
+	// **Every long-running resource says how it is ready** (novox/hq ADR 0240 rule 8): warned until the
+	// date, refused from it. The count is the catalogue's: its merge check keeps the number and lets a
+	// change lower it, never raise it.
+	undeclared := 0
+	required := !checkNow().Before(catalogue.HealthRequiredFrom)
+	for _, name := range names {
+		missing := catalogue.Undeclared(shelf[name])
+		undeclared += len(missing)
+		if len(missing) == 0 {
+			continue
+		}
+		if required {
+			for _, id := range missing {
+				fmt.Fprintf(out, "%s: %s stays up and does not say how it is ready: a long-running resource declares "+
+					"health since %s (novox/hq ADR 0240 rule 8)\n", name, id, catalogue.HealthRequiredFrom.Format("2006-01-02"))
+			}
+			failed += len(missing)
+			faulted[name] = true
+		}
+	}
+
 	for _, name := range names {
 		m := shelf[name]
 		if faulted[name] {
@@ -138,8 +161,24 @@ func moduleCheckFor(paths []string, longestMachine int, out io.Writer) error {
 			}
 			fmt.Fprintf(out, ", keeps %s", strings.Join(kept, ", "))
 		}
+		// How what it runs is ready (ADR 0240): each declared check, and what is judged by liveness alone.
+		var checks []string
+		for _, r := range m.Resources {
+			if h, has, _ := catalogue.ReadHealth(r); has {
+				checks = append(checks, fmt.Sprintf("%v by %s", r["id"], catalogue.HealthWords(h)))
+			}
+		}
+		if len(checks) > 0 {
+			fmt.Fprintf(out, ", ready: %s", strings.Join(checks, "; "))
+		}
+		if missing := catalogue.Undeclared(m); len(missing) > 0 {
+			fmt.Fprintf(out, "; WARNING: %s stay(s) up and say(s) not how it is ready — judged by liveness alone, "+
+				"refused from %s (ADR 0240 rule 8)", strings.Join(missing, ", "), catalogue.HealthRequiredFrom.Format("2006-01-02"))
+		}
 		fmt.Fprintln(out)
 	}
+	// The count the catalogue keeps (ADR 0240 rule 8), in a line its merge check reads.
+	fmt.Fprintf(out, "%s %d\n", UndeclaredHealthLine, undeclared)
 	if failed > 0 {
 		return fmt.Errorf("%d problem(s) in %d manifest(s)", failed, len(paths))
 	}
@@ -150,6 +189,13 @@ func moduleCheckFor(paths []string, longestMachine int, out io.Writer) error {
 	return nil
 }
 
+// UndeclaredHealthLine starts the line `module check` says the count of long-running resources without
+// `health` in, over the manifests given: the catalogue's merge check compares it with the number it keeps.
+const UndeclaredHealthLine = "long-running resources without health:"
+
+// checkNow is the clock `module check` judges the date by; a test sets it.
+var checkNow = time.Now
+
 func joinInvokes(invokes []string) string {
 	if len(invokes) == 1 && invokes[0] == "*" {
 		return "every tool"

cmd/mesh-controller/check_here.go

@@ -17,7 +17,7 @@ import (
 )
 
 // `check-here` is a pull request's merge check run on the machine at hand **exactly as the build seat
-// runs it** (novox/hq issue 286): the same code (builder.Check), the same ask the controller would make of
+// runs it** (novox/hq issue 283): the same code (builder.Check), the same ask the controller would make of
 // the seat — the gate's modules and judge by the planner's own answer over the facts snapshot, the
 // repositories beside it at the refs the snapshot says the seat clones them at — in the toolchain image
 // the mesh holds, as the user the build seat runs as, against a throwaway store and bus of the versions
@@ -126,7 +126,7 @@ func checkHereCommand(ctx context.Context, args []string) error {
 	}
 	if len(toolchains) == 0 {
 		return errors.New("the facts snapshot names no toolchain: it was taken by a controller from before " +
-			"issue 286, and the seat's toolchain cannot be known here")
+			"issue 283, and the seat's toolchain cannot be known here")
 	}
 	beside := map[string]builder.Beside{}
 	for d, ref := range f.Beside {

cmd/mesh-controller/facts.go

@@ -406,11 +406,11 @@ func gatherFacts(ctx context.Context, open *stores, busVersion string) (snapshot
 		if err != nil {
 			return snapshot.Facts{}, err
 		}
-		mod := snapshot.Module{Name: e.Manifest.Module, Repository: e.Source.Repository, Path: e.Source.Path,
+		mod := snapshot.Module{Name: e.Manifest.Module, Repository: snapshot.RepositoryName(e.Source.Repository), Path: e.Source.Path,
 			Commit: e.Source.BuiltFrom, Provided: e.Provided, RollOut: current[e.Manifest.Module].RollOut,
 			Manifest: raw}
 		for _, r := range read[e.Manifest.Module] {
-			mod.Reads = append(mod.Reads, r.Repository)
+			mod.Reads = append(mod.Reads, snapshot.RepositoryName(r.Repository))
 		}
 		f.Modules = append(f.Modules, mod)
 		if e.Provided || e.Source.Repository == "" {
@@ -426,7 +426,8 @@ func gatherFacts(ctx context.Context, open *stores, busVersion string) (snapshot
 		if commit == "" {
 			commit = s.BuiltFrom
 		}
-		f.Sources = append(f.Sources, snapshot.Source{Repository: repository, Commit: commit, Modules: count[repository]})
+		f.Sources = append(f.Sources, snapshot.Source{Repository: snapshot.RepositoryName(repository), Commit: commit,
+			Modules: count[repository]})
 	}
 	for _, e := range edges {
 		f.Edges = append(f.Edges, snapshot.Edge{From: e.From, To: e.To, Kind: e.Kind})

cmd/mesh-controller/gate.go

@@ -77,6 +77,10 @@ type health int
 
 const (
 	healthGood health = iota
+	// healthWaiting is not a pass and not a fault: what the module's checks find waits on an unhealthy
+	// provider (ADR 0240 rule 5), so the judging waits — past the bound too — rather than putting back a
+	// build for something it did not do.
+	healthWaiting
 	healthNotYet
 	healthBroken
 )
@@ -105,6 +109,12 @@ type gateFacts struct {
 	// holder is who holds the controller lease, for judging the controller.
 	holder    *lease.Holder
 	holderErr error
+	// health is each machine's newest health statement (ADR 0240); a machine absent never stated one.
+	// healthErr is why they could not be read.
+	health    map[string]inventory.NodeHealth
+	healthErr error
+	// heldOn is, per "<module>@<machine>", the provider its findings are held under (ADR 0240 rule 5).
+	heldOn map[string]string
 }
 
 // gatherGateFacts reads what a judging needs, from the store, the bus and this controller's memory. A
@@ -127,6 +137,9 @@ var gatherGateFacts = func(ctx context.Context, open *stores, component string)
 	for _, n := range nodes {
 		f.engines[n.Name] = n.HostVersion
 	}
+	// What each machine says of its long-running resources (ADR 0240): unreadable is said, never read as
+	// healthy.
+	f.health, f.healthErr = inv.Healths(ctx)
 	if d := doctorFrom; d != nil {
 		if d.keeper != nil {
 			f.judged = true
@@ -140,6 +153,17 @@ var gatherGateFacts = func(ctx context.Context, open *stores, component string)
 	} else {
 		f.servedErr = errors.New("this process does not serve the mesh, so it cannot ask the bus who serves what")
 	}
+	// Whose findings wait on an unhealthy provider (ADR 0240 rule 5): their gates wait, not fail.
+	if f.healthErr == nil && f.openErr == nil {
+		if hold, err := readHolding(ctx, inv, f.open); err == nil {
+			f.heldOn = map[string]string{}
+			for machine := range f.health {
+				for module, p := range hold.heldModules(machine) {
+					f.heldOn[module+"@"+machine] = p.Module + " on " + p.Node
+				}
+			}
+		}
+	}
 	if theLease != nil {
 		h, found, err := theLease.holder(ctx)
 		switch {
@@ -193,7 +217,9 @@ func judgeHealth(module, component string, m catalogue.Manifest, machine string,
 			if !onIt {
 				continue
 			}
-			if c.Subject.Scope == conditions.ScopeMachine || slices.Contains(strings.Split(c.Subject.ID, "."), module) {
+			// A module's own health condition names it whole, its name's dots and all (ADR 0240).
+			ownHealth := c.Subject.Scope == conditions.ScopeModule && c.Subject.ID == module+"."+machine
+			if c.Subject.Scope == conditions.ScopeMachine || ownHealth || slices.Contains(strings.Split(c.Subject.ID, "."), module) {
 				return healthNotYet, fmt.Sprintf("raised since it was sent: %s — %s", c.Key, c.Summary)
 			}
 		}
@@ -241,6 +267,11 @@ func judgeHealth(module, component string, m catalogue.Manifest, machine string,
 				return healthNotYet, fmt.Sprintf("the node tools on %s do not serve %s's tools", machine, module)
 			}
 		}
+		// **And what it runs is stated healthy** (ADR 0240 §4): every long-running resource of it on that
+		// machine, in a statement heard since the send. A resource still starting makes the judging wait.
+		if h, why := moduleHealthWord(module, machine, since, f); h != healthGood {
+			return h, why
+		}
 	}
 	return healthGood, ""
 }
@@ -421,6 +452,10 @@ func judgeMoves(ctx context.Context, open *stores, g *inventory.PlanGate, pairs
 		// What broke is put back; what was only not yet healthy beside it is too — they moved together.
 		g.Failing = failing
 		decide(g, inventory.GateFailed, why, now)
+	case worst == healthWaiting:
+		// Waiting on a provider that is unhealthy: not a pass, and not a failure at the bound either —
+		// the provider's own condition says what is wrong (ADR 0240 rule 5).
+		g.Passes, g.LastPass, g.Last, g.Failing = 0, nil, why, failing
 	case worst == healthNotYet:
 		g.Passes, g.LastPass, g.Last, g.Failing = 0, nil, why, failing
 		if now.Sub(*g.Since) > gateBound {

cmd/mesh-controller/gate_test.go

@@ -346,6 +346,48 @@ func TestTheHealthDefinitions(t *testing.T) {
 		!strings.Contains(why, "first run") {
 		t.Errorf("a controller not ready is %v (%s)", h, why)
 	}
+	// What the module runs (novox/hq ADR 0240 §4): a judging passes only when every long-running
+	// resource of it on that machine is stated healthy since the send; starting and unhealthy wait.
+	f = applied("anchor")
+	stated := func(state, reason string, heard time.Time) {
+		f.health = map[string]inventory.NodeHealth{"anchor": {Node: "anchor", Contract: 1, SaidAt: heard, HeardAt: heard,
+			Resources: []inventory.ResourceHealth{
+				{Module: "app", Resource: "app.server", Kind: "container", Target: "app-server", State: state, Reason: reason},
+				{Module: "other", Resource: "other.server", Kind: "container", State: link.StateUnhealthy, Reason: "down"}}}}
+	}
+	plain := catalogue.Manifest{Module: "app"}
+	if h, why := judgeHealth("app", "", plain, "anchor", since, f); h != healthGood {
+		t.Errorf("a machine whose engine states no health is judged as before, not %v (%s)", h, why)
+	}
+	stated(link.StateStarting, "", now)
+	if h, why := judgeHealth("app", "", plain, "anchor", since, f); h != healthNotYet || !strings.Contains(why, "starting") {
+		t.Errorf("a resource still starting is not yet a pass: %v (%s)", h, why)
+	}
+	stated(link.StateUnhealthy, "restarting", now)
+	if h, why := judgeHealth("app", "", plain, "anchor", since, f); h != healthNotYet || !strings.Contains(why, "restarting") {
+		t.Errorf("a resource unhealthy fails the judging: %v (%s)", h, why)
+	}
+	stated(link.StateHealthy, "", since.Add(-time.Minute))
+	if h, why := judgeHealth("app", "", plain, "anchor", since, f); h != healthNotYet {
+		t.Errorf("a statement from before the send says nothing of the new build: %v (%s)", h, why)
+	}
+	stated(link.StateHealthy, "", now)
+	if h, why := judgeHealth("app", "", plain, "anchor", since, f); h != healthGood {
+		t.Errorf("every resource of it healthy since the send — another module's state is not its — is %v (%s)", h, why)
+	}
+	f.healthErr = errors.New("the store is away")
+	if h, _ := judgeHealth("app", "", plain, "anchor", since, f); h != healthNotYet {
+		t.Errorf("health that cannot be read is never read as healthy: %v", h)
+	}
+	// And the condition it raises after the send holds it, as every condition about it does.
+	f.healthErr = nil
+	f.judged = true
+	f.open = []conditions.Condition{{Key: "module.app.anchor.unhealthy", Kind: kindModuleUnhealthy, Subject: conditions.Subject{
+		Scope: conditions.ScopeModule, ID: "app.anchor", Machine: "anchor"}, Raised: now, Summary: "app on anchor is not healthy"}}
+	if h, why := judgeHealth("app", "", plain, "anchor", since, f); h != healthNotYet || !strings.Contains(why, "module.app.anchor.unhealthy") {
+		t.Errorf("a module held on its own unhealthy condition is %v (%s)", h, why)
+	}
+
 	// A machine that refused what it was sent: broken.
 	f = applied("anchor")
 	r := f.reports["anchor"]

cmd/mesh-controller/handacts.go

@@ -41,6 +41,11 @@ type handActVerb struct {
 // causeLeakedInLogs is the cause a rotation after a value was printed into a log gives.
 const causeLeakedInLogs = "leaked-in-logs"
 
+// causeDrill is the cause of every act `hand-act drill` records, and the one cause `hand-act record`
+// refuses: a drill has its own verb, so whether an act was a drill is said by the verb a person chose,
+// never by a word typed into a repair's cause (novox/hq issue 292).
+const causeDrill = "drill"
+
 // handActVerbs is every verb that writes the hand-act log (novox/hq to-be 45 §7). **S15 reads it**:
 // an act recorded by a verb whose entry names a decision is the mesh working as decided, never a
 // repair, and does not count toward `healer-wanted` — whatever cause it gives. A verb not listed, or
@@ -57,8 +62,14 @@ var handActVerbs = []handActVerb{
 	{Verb: "broker consumer-reset"},
 	// Silencing the same condition twice says the condition, or what it watches, wants mending.
 	{Verb: "conditions silence"},
-	// An act done outside the mesh: the mesh cannot tell a repair from a decision there, so it counts.
-	{Verb: "hand-act record"},
+	// An act done outside the mesh: the mesh cannot tell a repair from a decision there, so it counts —
+	// except a drill recorded through it before `hand-act drill` existed (2026-10-07). It refuses the
+	// cause since, so no act recorded through it now carries it.
+	{Verb: "hand-act record", Decision: "a drill recorded before `hand-act drill` existed: a person's " +
+		"deliberate test, never a repair", DecidedFor: []string{causeDrill}},
+	// A drill: something broken on purpose to see the mesh raise and clear it. A person's test, never a
+	// repair, however often it is run.
+	{Verb: "hand-act drill", Decision: "a drill is a person's deliberate test of the mesh, never a repair"},
 	// A person's decisions by design.
 	{Verb: "retire approve", Decision: "nothing is retired past its bound without a person (ADR 0230)"},
 	{Verb: "retire reject", Decision: "keeping a consumer active is a person's word (ADR 0230)"},
@@ -174,6 +185,10 @@ func handActCommand(ctx context.Context, args []string) error {
 			return errors.New("hand-act record says the cause too: --cause <word>, the word a second " +
 				"act for the same reason will use — it is how a repair done twice is found")
 		}
+		if strings.EqualFold(strings.TrimSpace(*f.cause), causeDrill) {
+			return errors.New("a drill is not recorded as a repair: hand-act drill <what was done> --why <text> " +
+				"records it as the person's deliberate test it is, which no healer is wanted for. Nothing was recorded")
+		}
 		act := link.HandAct{Verb: "hand-act record", Args: []string{what}, Why: strings.TrimSpace(*f.why),
 			Cause: strings.TrimSpace(*f.cause), Condition: strings.TrimSpace(*f.condition)}
 		return onTheBus(func(conn *nats.Conn) error {
@@ -186,8 +201,12 @@ func handActCommand(ctx context.Context, args []string) error {
 			return nil
 		})
 	}
+	if len(args) > 0 && args[0] == "drill" {
+		return handActDrill(ctx, args[1:])
+	}
 	if len(args) > 0 && args[0] != "list" && !strings.HasPrefix(args[0], "-") {
-		return errors.New("hand-act record <what> --why <text> --cause <word> | hand-acts [--days N] [--json]")
+		return errors.New("hand-act record <what> --why <text> --cause <word> | hand-act drill <what> --why <text> " +
+			"| hand-acts [--days N] [--json]")
 	}
 	if len(args) > 0 && args[0] == "list" {
 		args = args[1:]
@@ -239,6 +258,38 @@ func handActCommand(ctx context.Context, args []string) error {
 	})
 }
 
+// handActDrill is `hand-act drill`: an act done on purpose to test the mesh — a module stopped, a
+// process killed — recorded so the conditions it raises are read as the drill they are. Its cause is
+// always causeDrill and S15 never counts it (handActVerbs).
+func handActDrill(ctx context.Context, args []string) error {
+	set := flag.NewFlagSet("hand-act drill", flag.ContinueOnError)
+	why := set.String("why", "", "what the drill tests — recorded in the hand-act log (novox/hq to-be 45 §7)")
+	condition := set.String("condition", "", "the key of the condition the drill is meant to raise, if any")
+	positionals, err := parseAround(set, args)
+	if err != nil {
+		return err
+	}
+	what := strings.TrimSpace(strings.Join(positionals, " "))
+	if what == "" {
+		return errors.New("hand-act drill <what was done on purpose> --why <what it tests> [--condition <key>]")
+	}
+	if strings.TrimSpace(*why) == "" {
+		return errors.New("a drill says what it tests: --why <text> (recorded in the hand-act log, novox/hq " +
+			"to-be 45 §7). Nothing was recorded")
+	}
+	act := link.HandAct{Verb: "hand-act drill", Args: []string{what}, Why: strings.TrimSpace(*why),
+		Cause: causeDrill, Condition: strings.TrimSpace(*condition)}
+	return onTheBus(func(conn *nats.Conn) error {
+		written, err := link.RecordHandAct(ctx, conn, act)
+		if err != nil {
+			return fmt.Errorf("the drill could not be recorded: %w", err)
+		}
+		fmt.Printf("recorded as %s: %s drilled %q, because %q — a drill, which no healer is wanted for\n",
+			written.ID, written.By, what, written.Why)
+		return nil
+	})
+}
+
 // repairs are the acts that are not a person's decision by design: what S15 counts.
 func repairs(acts []link.HandAct) []link.HandAct {
 	out := make([]link.HandAct, 0, len(acts))

cmd/mesh-controller/handacts_test.go

@@ -92,3 +92,32 @@ func TestDurationsAreSummarisedPerSubject(t *testing.T) {
 		t.Fatalf("a minute between words suggests %q", got[1].Suggests)
 	}
 }
+
+// **A drill has its own verb** — `hand-act drill`, the seat's `drill` — and `hand-act record` refuses
+// the cause, so a repair cannot pass for a drill by the word it gives.
+func TestADrillIsRecordedThroughItsOwnVerb(t *testing.T) {
+	err := handActCommand(context.Background(), []string{"record", "stopped searxng", "--why", "a test", "--cause", "drill"})
+	if err == nil || !strings.Contains(err.Error(), "hand-act drill") {
+		t.Errorf("hand-act record --cause drill was not sent to the drill verb: %v", err)
+	}
+	if err := handActCommand(context.Background(), []string{"drill", "stopped searxng"}); err == nil ||
+		!strings.Contains(err.Error(), "--why") {
+		t.Errorf("a drill without what it tests: %v", err)
+	}
+	if err := handActCommand(context.Background(), []string{"drill", "--why", "a test"}); err == nil ||
+		!strings.Contains(err.Error(), "hand-act drill <what") {
+		t.Errorf("a drill without what was done: %v", err)
+	}
+	argv, err := argvFor("drill", map[string]any{"what": "stopped searxng", "why": "ADR 0240 phase A",
+		"condition": "machine.ace.module.searxng.unhealthy"})
+	if want := "hand-act drill stopped searxng --why ADR 0240 phase A --condition machine.ace.module.searxng.unhealthy"; err != nil ||
+		strings.Join(argv, " ") != want {
+		t.Errorf("the seat's drill: %v %v, want %q", argv, err, want)
+	}
+	if _, err := argvFor("drill", map[string]any{"what": "stopped searxng"}); err == nil {
+		t.Error("the seat's drill without why was not refused")
+	}
+	if repairingCommand([]string{"hand-act", "drill", "x"}) != "hand-act drill" {
+		t.Error("a drill through `command` is not held to why")
+	}
+}

cmd/mesh-controller/health_check_test.go

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"bytes"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+)
+
+// Every catalogue module that runs something long-lived declares how it is ready (novox/hq ADR 0240 rule
+// 8): `module check` warns and counts the undeclared before the date, and refuses them from it.
+func TestModuleCheckCountsTheUndeclaredAndRefusesThemFromTheDate(t *testing.T) {
+	dir := t.TempDir()
+	digest := "@sha256:" + strings.Repeat("a", 64)
+	path := filepath.Join(dir, "module.json")
+	os.WriteFile(path, []byte(`{"module":"web","listens":[{"name":"web","port":80,"from":"mesh"}],"resources":[
+	  {"id":"server","type":"container","name":"web","image":"registry.example/web`+digest+`","ports":["80"],
+	   "health":{"kind":"http","endpoint":"web"}},
+	  {"id":"worker","type":"container","name":"web-worker","image":"registry.example/web`+digest+`"},
+	  {"id":"seed","type":"container","name":"web-seed","image":"registry.example/web`+digest+`","run-once":true}]}`), 0o600)
+	defer func() { checkNow = time.Now }()
+
+	checkNow = func() time.Time { return catalogue.HealthRequiredFrom.Add(-time.Hour) }
+	var out bytes.Buffer
+	if err := moduleCheck([]string{path}, &out); err != nil {
+		t.Fatalf("refused before the date: %v\n%s", err, out.String())
+	}
+	for _, want := range []string{"ready: server by http / on web every 30s", "WARNING: worker stay(s) up",
+		catalogue.HealthRequiredFrom.Format("2006-01-02"), UndeclaredHealthLine + " 1"} {
+		if !strings.Contains(out.String(), want) {
+			t.Errorf("the check does not say %q:\n%s", want, out.String())
+		}
+	}
+
+	checkNow = func() time.Time { return catalogue.HealthRequiredFrom }
+	out.Reset()
+	if err := moduleCheck([]string{path}, &out); err == nil {
+		t.Fatalf("a long-running resource without health passed after the date:\n%s", out.String())
+	}
+	if !strings.Contains(out.String(), "web: worker stays up and does not say how it is ready") {
+		t.Errorf("the refusal does not name the resource:\n%s", out.String())
+	}
+}

cmd/mesh-controller/main.go

@@ -123,7 +123,7 @@ func run() error {
 	// The merge gate: every machine of the snapshot composed with a change (novox/hq to-be 45 §9).
 	case "merge-gate":
 		return mergeGateCommand(ctx, args[1:])
-	// A pull request's merge check run here exactly as the build seat runs it (novox/hq issue 286).
+	// A pull request's merge check run here exactly as the build seat runs it (novox/hq issue 283).
 	case "check-here":
 		return checkHereCommand(ctx, args[1:])
 	case "upgrade":

cmd/mesh-controller/merge_gate.go

@@ -108,7 +108,7 @@ type mergeVerdict struct {
 	Facts    time.Time `json:"facts"`
 	Failures []string  `json:"failures,omitempty"`
 	// Errors are what kept the gate from judging: a machine the mesh composes that the gate could not
-	// raise as it is. Any one makes the verdict an error — never a pass (novox/hq issue 285).
+	// raise as it is. Any one makes the verdict an error — never a pass (novox/hq issue 282).
 	Errors   []string          `json:"errors,omitempty"`
 	Warnings []string          `json:"warnings,omitempty"`
 	Notes    []string          `json:"notes,omitempty"`
@@ -222,7 +222,7 @@ func mergeGateCommand(ctx context.Context, args []string) error {
 }
 
 // errMergeGateCouldNotJudge is the gate's own error: the mesh as it is could not be raised, so nothing
-// the change does to it can be seen. Never a pass (novox/hq issue 285).
+// the change does to it can be seen. Never a pass (novox/hq issue 282).
 var errMergeGateCouldNotJudge = errors.New("the merge gate could not judge the change")
 
 // errMergeGateFailed is the gate's own failure: the verdict says why, so the error says nothing more.
@@ -351,7 +351,7 @@ func judgeChange(ctx context.Context, in mergeCheckInput) (mergeVerdict, error)
 			v.Failures = append(v.Failures, fmt.Sprintf("%s: nothing could be sent to it with this change — %s",
 				gm.Described, firstOr(newOnly(gm.Change.Problems, gm.Base.Problems), gm.Change.Problems)))
 		case !gm.Base.Composes && m.Declaration.Composes:
-			// **The mesh composes it and the gate could not raise it as it is** (novox/hq issue 285): a fact
+			// **The mesh composes it and the gate could not raise it as it is** (novox/hq issue 282): a fact
 			// the snapshot does not carry, or one the gate does not raise. Then the change is judged against a
 			// machine that is not the mesh's — broken against broken, which passes whatever the change does —
 			// so the gate cannot judge, and says so: an error, never a pass.

cmd/mesh-controller/merge_gate_test.go

@@ -323,11 +323,11 @@ func busMesh(t *testing.T) snapshot.Facts {
 	return f
 }
 
-// **Issue 285**: every machine running a module on the bus failed to compose in the gate's store, with
+// **Issue 282**: every machine running a module on the bus failed to compose in the gate's store, with
 // the change and without — the store held the module's credential and no account for it, which
 // composition refuses (issue 203) — and the gate passed every change, "0 of 4 compose". The account the
 // mesh issued is raised with its credential, so the machine composes in the gate as on the mesh.
-func TestIssue285AModuleOnTheBusComposesInTheGate(t *testing.T) {
+func TestIssue282AModuleOnTheBusComposesInTheGate(t *testing.T) {
 	f := busMesh(t)
 	v := gateJudged(t, f, "")
 	if v.Verdict != "pass" {
@@ -343,10 +343,10 @@ func TestIssue285AModuleOnTheBusComposesInTheGate(t *testing.T) {
 	}
 }
 
-// **Issue 285**: a machine the mesh composes that the gate cannot raise as it is leaves the change judged
+// **Issue 282**: a machine the mesh composes that the gate cannot raise as it is leaves the change judged
 // against a machine that is not the mesh's — broken against broken, which passes whatever the change does.
 // That is an error, never a pass.
-func TestIssue285AMachineTheGateCannotRaiseIsAnErrorNeverAPass(t *testing.T) {
+func TestIssue282AMachineTheGateCannotRaiseIsAnErrorNeverAPass(t *testing.T) {
 	f := busMesh(t)
 	for i := range f.Machines {
 		// A fact the snapshot does not carry: the module's credential, gone from the laptop's.
@@ -386,7 +386,7 @@ func TestAWithheldPathIsStoodInForByAPath(t *testing.T) {
 	}
 }
 
-// **Issue 286**: a check run by hand clones beside a change what the seat clones — one rule, read by the
+// **Issue 283**: a check run by hand clones beside a change what the seat clones — one rule, read by the
 // controller's ask and by the facts — and finds the repository it checks from its origin.
 func TestACheckByHandClonesWhatTheSeatClones(t *testing.T) {
 	for dir, refs := range map[string]map[string]string{

cmd/mesh-controller/module_health.go

@@ -0,0 +1,325 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"strings"
+	"sync/atomic"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/conditions"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A module says how it is healthy, and the node-engine judges it (novox/hq ADR 0240, to-be 48 §4 and §5,
+// Phase A).
+//
+// **The node-engine owns every verdict; the controller keeps the last word and raises the condition.** A
+// machine states, in every report and as an event between reports, the state of every long-running
+// resource it runs for a module. The controller keeps the newest statement per machine (node_health), and
+// raises `module.<module>.<machine>.unhealthy` when two statements in a row say a resource of the module
+// is unhealthy — one statement is listed as unconfirmed, as the self-check does a finding one look can be
+// wrong about (to-be 45 §4, issue 277) — and clears it on the first that does not. The release gate reads
+// the stated health: a judging passes a module only when every long-running resource of it on that
+// machine is stated healthy, so a resource still starting is not yet a pass.
+//
+// **An engine older than the judging states nothing**, and its machine's health is not known: never
+// healthy, never a reason to raise anything, and the gate judges it as it did before.
+
+// The condition a module's health raises.
+const (
+	kindModuleUnhealthy = "module-unhealthy"
+	// sourceHealth is what raised it: the machine's own statement.
+	sourceHealth = "health"
+	// moduleUnhealthyUrgentAfter is how long it stands before it is urgent (to-be 48 §4).
+	moduleUnhealthyUrgentAfter = 4 * time.Hour
+	// moduleUnhealthyAfter is how many statements in a row raise it.
+	moduleUnhealthyAfter = 2
+)
+
+// healthRefused counts the statements refused as older than the one kept, for the log and a test.
+var healthRefused atomic.Int64
+
+// moduleHealth keeps what the machines state, for the link (link.Healths).
+type moduleHealth struct {
+	inv    *inventory.Inventory
+	keeper func() *conditions.Keeper
+}
+
+func (m moduleHealth) Stated(ctx context.Context, node string, h link.Health) error {
+	return stateHealth(ctx, m.inv, m.keeper(), node, h, time.Now())
+}
+
+// stateHealth keeps one machine's statement and raises or clears its modules' conditions from it. An
+// older statement than the one kept is refused, by when the engine looked.
+func stateHealth(ctx context.Context, inv *inventory.Inventory, k *conditions.Keeper, node string, h link.Health,
+	now time.Time) error {
+	if h.Contract == 0 {
+		return nil
+	}
+	prev, had, err := inv.HealthOf(ctx, node)
+	if err != nil {
+		return err
+	}
+	if had && h.At.Before(prev.SaidAt) {
+		healthRefused.Add(1)
+		return nil
+	}
+	unhealthy := map[string][]inventory.ResourceHealth{}
+	resources := make([]inventory.ResourceHealth, 0, len(h.Resources))
+	for _, r := range h.Resources {
+		kept := inventory.ResourceHealth{Module: r.Module, Resource: r.Resource, Kind: r.Kind, Target: r.Target,
+			State: r.State, Reason: r.Reason, Since: r.Since, Streak: r.Streak, Restarts: r.Restarts,
+			Check: r.Check, Needs: r.Needs}
+		resources = append(resources, kept)
+		if r.State == link.StateUnhealthy && r.Module != "" {
+			unhealthy[r.Module] = append(unhealthy[r.Module], kept)
+		}
+	}
+	streaks := map[string]int{}
+	for module := range unhealthy {
+		streaks[module] = prev.Streaks[module] + 1
+	}
+	stored, err := inv.RecordHealth(ctx, inventory.NodeHealth{Node: node, Contract: h.Contract, SaidAt: h.At,
+		HeardAt: now, Resources: resources, Streaks: streaks})
+	if err != nil || !stored {
+		if err == nil {
+			healthRefused.Add(1)
+		}
+		return err
+	}
+	if k == nil {
+		return nil
+	}
+	return judgeModuleHealth(ctx, inv, k, node, unhealthy, streaks, now)
+}
+
+// judgeModuleHealth raises a module's condition on a machine on the second statement in a row that says a
+// resource of it is unhealthy — or on the first while it is already open — and clears every one this
+// statement no longer says. **A consumer whose findings wait on an unhealthy provider is held** (to-be 48
+// §6): raised as nothing of its own, listed at the provider's condition, which is urgent while anyone
+// waits on it.
+func judgeModuleHealth(ctx context.Context, inv *inventory.Inventory, k *conditions.Keeper, node string,
+	unhealthy map[string][]inventory.ResourceHealth, streaks map[string]int, now time.Time) error {
+	open, err := k.Open(ctx)
+	if err != nil {
+		return err
+	}
+	standing := map[string]conditions.Condition{}
+	for _, c := range open {
+		if c.Kind == kindModuleUnhealthy && c.Subject.Machine == node {
+			standing[c.Key] = c
+		}
+	}
+	var hold *holding
+	if inv != nil {
+		if hold, err = readHolding(ctx, inv, open); err != nil {
+			return err
+		}
+	}
+	var problems []string
+	modules := make([]string, 0, len(unhealthy))
+	for m := range unhealthy {
+		modules = append(modules, m)
+	}
+	sort.Strings(modules)
+	seen := map[string]bool{}
+	heldOn := map[string]string{}
+	providers := map[catalogue.Chosen]bool{}
+	for _, m := range modules {
+		o := moduleUnhealthyObservation(m, node, unhealthy[m])
+		if hold != nil {
+			if p, held := hold.heldUnder(node, m, unhealthy[m]); held {
+				// Held under the provider's condition: nothing of its own, and the provider's says it waits.
+				heldOn[o.Key()] = p.Module + " on " + p.Node
+				providers[p] = true
+				continue
+			}
+			if waiters := hold.waitersOn(catalogue.Chosen{Node: node, Module: m}); len(waiters) > 0 {
+				o.Severity = conditions.Urgent
+				o.Said += "; " + waitingWords(waiters)
+				o.Summary += fmt.Sprintf("; %d consumer(s) wait on it", len(waiters))
+			}
+		}
+		seen[o.Key()] = true
+		c, isOpen := standing[o.Key()]
+		if streaks[m] < moduleUnhealthyAfter && !isOpen {
+			continue // unconfirmed: one statement can be wrong; `node show` lists it
+		}
+		if isOpen && now.Sub(c.Raised) >= moduleUnhealthyUrgentAfter {
+			o.Severity = conditions.Urgent
+		}
+		if _, err := k.Observe(ctx, o); err != nil {
+			problems = append(problems, err.Error())
+		}
+	}
+	for key, c := range standing {
+		if seen[key] {
+			continue
+		}
+		module := strings.TrimSuffix(c.Subject.ID, "."+node)
+		why := fmt.Sprintf("%s says no resource of %s is unhealthy", node, module)
+		if on, held := heldOn[key]; held {
+			why = fmt.Sprintf("what %s finds on %s waits on %s, which is unhealthy: held under its condition", module, node, on)
+		}
+		if _, err := k.Clear(ctx, key, why); err != nil {
+			problems = append(problems, err.Error())
+		}
+	}
+	// And each provider a consumer here now waits on, when its own condition is open: said again with who
+	// waits on it, so the wait is listed at the provider whichever machine's statement arrived first.
+	for p := range providers {
+		if err := sayWaiters(ctx, k, hold, p, now); err != nil {
+			problems = append(problems, err.Error())
+		}
+	}
+	if len(problems) > 0 {
+		return fmt.Errorf("%s", strings.Join(problems, "; "))
+	}
+	return nil
+}
+
+// sayWaiters observes a provider's open condition again, with who waits on it, from its machine's newest
+// statement. Nothing when its condition is not open: it is raised by its own statements, on its own looks.
+func sayWaiters(ctx context.Context, k *conditions.Keeper, hold *holding, p catalogue.Chosen, now time.Time) error {
+	var raisedAt *conditions.Condition
+	for i, c := range hold.open {
+		if c.Key == moduleUnhealthyKey(p.Module, p.Node) {
+			raisedAt = &hold.open[i]
+		}
+	}
+	if raisedAt == nil {
+		return nil
+	}
+	var rs []inventory.ResourceHealth
+	for _, r := range hold.healths[p.Node].Resources {
+		if r.Module == p.Module && r.State == link.StateUnhealthy {
+			rs = append(rs, r)
+		}
+	}
+	if len(rs) == 0 {
+		return nil
+	}
+	o := moduleUnhealthyObservation(p.Module, p.Node, rs)
+	if waiters := hold.waitersOn(p); len(waiters) > 0 {
+		o.Severity = conditions.Urgent
+		o.Said += "; " + waitingWords(waiters)
+		o.Summary += fmt.Sprintf("; %d consumer(s) wait on it", len(waiters))
+	}
+	_, err := k.Observe(ctx, o)
+	return err
+}
+
+// moduleUnhealthyObservation is a module unhealthy on a machine, in words: the summary names the module,
+// the machine and what is wrong with each resource; the detail — targets, streaks, since — is evidence.
+func moduleUnhealthyObservation(module, node string, rs []inventory.ResourceHealth) conditions.Observation {
+	var words, said []string
+	for _, r := range rs {
+		words = append(words, fmt.Sprintf("its %s %s %s", r.Kind, r.Resource, reasonWords(r)))
+		said = append(said, fmt.Sprintf("%s (%s %s): %s, %d look(s) in a row, %d restart(s) counted, since %s",
+			r.Resource, r.Kind, r.Target, orNotSaid(r.Reason), r.Streak, r.Restarts,
+			r.Since.UTC().Format("2006-01-02 15:04:05 MST")))
+	}
+	return conditions.Observation{Scope: conditions.ScopeModule, ID: module + "." + node, Token: "unhealthy",
+		Kind: kindModuleUnhealthy, Machine: node, Severity: conditions.Warning, Source: sourceHealth,
+		Summary: fmt.Sprintf("%s on %s is not healthy: %s", module, node, strings.Join(words, "; ")),
+		Said:    strings.Join(said, "; ")}
+}
+
+// reasonWords is why a resource is unhealthy, as a person reads it.
+func reasonWords(r inventory.ResourceHealth) string {
+	switch r.Reason {
+	case "restarting":
+		return fmt.Sprintf("keeps restarting (%d restart(s) counted)", r.Restarts)
+	case "down":
+		return "is not running"
+	case "":
+		return "is unhealthy"
+	}
+	// What a declared check found says an endpoint, a path or an address: evidence, never the summary the
+	// operator's channel carries (ADR 0234 §6). The summary names the check.
+	if r.Check != "" {
+		return "fails its " + r.Check + " check"
+	}
+	return "is unhealthy: " + r.Reason
+}
+
+func orNotSaid(s string) string {
+	if s == "" {
+		return "no reason said"
+	}
+	return s
+}
+
+// moduleHealthWord is the gate's reading of a module's stated health on a machine (ADR 0240 §4, ADR 0236
+// §2 as amended): good when every long-running resource of it is stated healthy in a statement heard since
+// the send; not yet otherwise, saying which. A machine that never stated health is judged as before.
+func moduleHealthWord(module, machine string, since time.Time, f gateFacts) (health, string) {
+	if f.healthErr != nil {
+		return healthNotYet, "what " + machine + " says of its resources' health cannot be read: " + firstLine(f.healthErr.Error())
+	}
+	h, states := f.health[machine]
+	if !states {
+		return healthGood, ""
+	}
+	if h.HeardAt.Before(since) {
+		return healthNotYet, fmt.Sprintf("%s has not said how what %s runs is since it was sent", machine, module)
+	}
+	for _, r := range h.Resources {
+		if r.Module != module {
+			continue
+		}
+		switch r.State {
+		case link.StateHealthy:
+		case link.StateStarting:
+			return healthNotYet, fmt.Sprintf("its %s %s on %s is still starting", r.Kind, r.Resource, machine)
+		case link.StateUnhealthy:
+			if on, held := f.heldOn[module+"@"+machine]; held {
+				return healthWaiting, fmt.Sprintf("its %s %s on %s waits on %s, which is unhealthy", r.Kind,
+					r.Resource, machine, on)
+			}
+			return healthNotYet, fmt.Sprintf("its %s %s on %s %s", r.Kind, r.Resource, machine, reasonWords(r))
+		default:
+			return healthNotYet, fmt.Sprintf("its %s %s on %s is %s%s", r.Kind, r.Resource, machine, r.State,
+				reasonAfter(r.Reason))
+		}
+	}
+	return healthGood, ""
+}
+
+func reasonAfter(s string) string {
+	if s == "" {
+		return ""
+	}
+	return ": " + s
+}
+
+// healthLines is what `node show` says of a machine's long-running resources: each with its state and
+// since when, an unhealthy one said once marked unconfirmed.
+func healthLines(h inventory.NodeHealth, had bool, now time.Time) []string {
+	if !had {
+		return []string{"  its node-engine does not say how what it runs is — it is older than the judging (ADR 0240)"}
+	}
+	if len(h.Resources) == 0 {
+		return []string{fmt.Sprintf("  it runs nothing long-lived for a module (said %s ago)", roughly(now.Sub(h.HeardAt)))}
+	}
+	out := []string{fmt.Sprintf("  what it runs, as it said %s ago:", roughly(now.Sub(h.HeardAt)))}
+	for _, r := range h.Resources {
+		line := fmt.Sprintf("    %-10s %-34s %s %s, since %s", r.State, r.Resource, r.Kind, r.Target,
+			r.Since.Local().Format("2006-01-02 15:04"))
+		if r.Reason != "" {
+			line += " — " + r.Reason
+		}
+		if r.Restarts > 0 {
+			line += fmt.Sprintf(", %d restart(s) counted", r.Restarts)
+		}
+		if r.State == link.StateUnhealthy && h.Streaks[r.Module] < moduleUnhealthyAfter {
+			line += " (unconfirmed: said once)"
+		}
+		out = append(out, line)
+	}
+	return out
+}

cmd/mesh-controller/module_health_test.go

@@ -0,0 +1,196 @@
+package main
+
+import (
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/conditions"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A module's stated health, as the controller keeps it and raises from it (novox/hq ADR 0240, "how it is
+// checked", rule 4): one unhealthy statement raises nothing and is listed unconfirmed; two raise; a
+// healthy one clears; a condition from before the send clears at the new build's start; an older
+// statement is refused; and an engine that states nothing raises nothing.
+
+var h0 = time.Date(2026, 10, 7, 12, 0, 0, 0, time.UTC)
+
+func aStatement(at time.Time, states ...string) link.Health {
+	h := link.Health{Contract: link.LivenessContract, At: at}
+	for i, s := range states {
+		r := link.ResourceHealth{Module: "letta", Resource: "letta.server", Kind: "container", Target: "letta-server",
+			State: s, Since: at}
+		if i > 0 {
+			r.Module, r.Resource, r.Target = "mqtt", "mqtt.broker", "mosquitto.service"
+		}
+		if s == link.StateUnhealthy {
+			r.Reason, r.Restarts, r.Streak = "restarting", 4, 2
+		}
+		h.Resources = append(h.Resources, r)
+	}
+	return h
+}
+
+func TestTwoUnhealthyStatementsRaiseTheModulesConditionAndAHealthyOneClearsIt(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	k := conditionsFrom
+	const key = "module.letta.anchor.unhealthy"
+	openKeys := func() []string {
+		t.Helper()
+		list, err := k.Open(ctx)
+		if err != nil {
+			t.Fatal(err)
+		}
+		var keys []string
+		for _, c := range list {
+			keys = append(keys, c.Key)
+		}
+		return keys
+	}
+
+	if err := stateHealth(ctx, inv, k, "anchor", aStatement(h0, link.StateHealthy, link.StateHealthy), h0); err != nil {
+		t.Fatal(err)
+	}
+	// One statement: nothing raised, and `node show` lists it as unconfirmed.
+	if err := stateHealth(ctx, inv, k, "anchor", aStatement(h0.Add(time.Minute), link.StateUnhealthy, link.StateHealthy),
+		h0.Add(time.Minute)); err != nil {
+		t.Fatal(err)
+	}
+	if keys := openKeys(); len(keys) != 0 {
+		t.Fatalf("one statement raised %v", keys)
+	}
+	kept, had, err := inv.HealthOf(ctx, "anchor")
+	if err != nil || !had {
+		t.Fatalf("the statement was not kept: %v %v", had, err)
+	}
+	if lines := strings.Join(healthLines(kept, had, h0.Add(time.Minute)), "\n"); !strings.Contains(lines, "unconfirmed") ||
+		!strings.Contains(lines, "letta.server") {
+		t.Fatalf("node show does not list the first statement as unconfirmed:\n%s", lines)
+	}
+
+	// The second in a row raises it — the module's own, never the other module's on the machine.
+	if err := stateHealth(ctx, inv, k, "anchor", aStatement(h0.Add(2*time.Minute), link.StateUnhealthy, link.StateHealthy),
+		h0.Add(2*time.Minute)); err != nil {
+		t.Fatal(err)
+	}
+	if keys := openKeys(); len(keys) != 1 || keys[0] != key {
+		t.Fatalf("two statements raised %v, not %s", keys, key)
+	}
+	c, _, _ := k.Get(ctx, key)
+	if c.Severity != conditions.Warning || c.Resolver != conditions.ResolverSelf || c.Subject.Machine != "anchor" ||
+		!strings.Contains(c.Summary, "letta on anchor") || !strings.Contains(c.Summary, "keeps restarting") ||
+		!strings.Contains(c.Evidence[0].Said, "letta-server") {
+		t.Fatalf("the condition does not say it in words with its evidence: %+v", c)
+	}
+
+	// Standing four hours, it is urgent.
+	if err := stateHealth(ctx, inv, k, "anchor", aStatement(h0.Add(5*time.Hour), link.StateUnhealthy, link.StateHealthy),
+		time.Now().Add(5*time.Hour)); err != nil {
+		t.Fatal(err)
+	}
+	if c, _, _ := k.Get(ctx, key); c.Severity != conditions.Urgent {
+		t.Fatalf("unhealthy for four hours is still %s", c.Severity)
+	}
+
+	// A new build's start — every start begins in `starting` — clears it: what follows is the new build's.
+	if err := stateHealth(ctx, inv, k, "anchor", aStatement(h0.Add(6*time.Hour), link.StateStarting, link.StateHealthy),
+		h0.Add(6*time.Hour)); err != nil {
+		t.Fatal(err)
+	}
+	if keys := openKeys(); len(keys) != 0 {
+		t.Fatalf("the first statement that says no resource is unhealthy did not clear it: %v", keys)
+	}
+	// And the streak starts again: one unhealthy statement after it raises nothing.
+	if err := stateHealth(ctx, inv, k, "anchor", aStatement(h0.Add(7*time.Hour), link.StateUnhealthy), h0.Add(7*time.Hour)); err != nil {
+		t.Fatal(err)
+	}
+	if keys := openKeys(); len(keys) != 0 {
+		t.Fatalf("one statement after a clearing raised %v", keys)
+	}
+}
+
+func TestAnOlderHealthStatementIsRefused(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	before := healthRefused.Load()
+	if err := stateHealth(ctx, inv, nil, "anchor", aStatement(h0.Add(time.Minute), link.StateHealthy), h0); err != nil {
+		t.Fatal(err)
+	}
+	// An event said before the report that overtook it arrives late.
+	if err := stateHealth(ctx, inv, nil, "anchor", aStatement(h0, link.StateUnhealthy), h0); err != nil {
+		t.Fatal(err)
+	}
+	kept, _, err := inv.HealthOf(ctx, "anchor")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !kept.SaidAt.Equal(h0.Add(time.Minute)) || kept.Resources[0].State != link.StateHealthy {
+		t.Fatalf("the older statement replaced the newer: %+v", kept)
+	}
+	if healthRefused.Load() != before+1 {
+		t.Fatalf("the refusal was not counted")
+	}
+}
+
+// **An engine older than the judging states nothing**: its reports raise nothing, keep nothing, and the
+// gate judges its machine as before — never healthy for having said nothing, never unhealthy.
+func TestAReportWithNoHealthRaisesAndKeepsNothing(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	l := nudgingListener{Enrolment: link.Enrolment{Inventory: open.inventory}}
+	if _, err := l.Heard(ctx, link.Report{Node: "anchor", Declared: "d1", Applied: []string{"letta.server"}}); err != nil {
+		t.Fatal(err)
+	}
+	if _, had, err := open.inventory.HealthOf(ctx, "anchor"); err != nil || had {
+		t.Fatalf("health kept for an engine that said none: %v %v", had, err)
+	}
+	if lines := healthLines(inventory.NodeHealth{}, false, time.Now()); !strings.Contains(lines[0], "older than the judging") {
+		t.Fatalf("node show: %v", lines)
+	}
+	// And one that does, through the report, is kept.
+	h := aStatement(time.Now().UTC(), link.StateHealthy)
+	if _, err := l.Heard(ctx, link.Report{Node: "anchor", Declared: "d1", Health: &h}); err != nil {
+		t.Fatal(err)
+	}
+	if kept, had, err := open.inventory.HealthOf(ctx, "anchor"); err != nil || !had || len(kept.Resources) != 1 {
+		t.Fatalf("the report's health was not kept: %+v %v %v", kept, had, err)
+	}
+}
+
+// A declared `health` is sent only to an engine whose own statement says it reads it (novox/hq ADR 0240
+// Phase B): an older engine is strict and would refuse the whole declaration for the field.
+func TestHealthIsSentOnlyToAnEngineThatSaysItReadsIt(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	reads := func() bool {
+		t.Helper()
+		got, err := engineReadsHealth(ctx, inv, "anchor")
+		if err != nil {
+			t.Fatal(err)
+		}
+		return got
+	}
+	if reads() {
+		t.Fatal("an engine that never stated anything is sent health")
+	}
+	if err := stateHealth(ctx, inv, nil, "anchor", aStatement(h0, link.StateHealthy), h0); err != nil {
+		t.Fatal(err)
+	}
+	if reads() {
+		t.Fatal("an engine judging liveness alone is sent health")
+	}
+	later := aStatement(h0.Add(time.Minute), link.StateHealthy)
+	later.Contract = link.ReadinessContract
+	if err := stateHealth(ctx, inv, nil, "anchor", later, h0.Add(time.Minute)); err != nil {
+		t.Fatal(err)
+	}
+	if !reads() {
+		t.Fatal("an engine that reads health is not sent it")
+	}
+}

cmd/mesh-controller/nodes.go

@@ -537,6 +537,17 @@ func showNode(ctx context.Context, inv *inventory.Inventory, name string) error
 		}
 	}
 
+	// What it runs for its modules, and how each is (novox/hq ADR 0240): "is it working" answered without
+	// a terminal on the machine.
+	if h, had, err := inv.HealthOf(ctx, name); err != nil {
+		fmt.Printf("\n  what it says of what it runs could NOT be read: %v\n", err)
+	} else {
+		fmt.Println()
+		for _, line := range healthLines(h, had, time.Now()) {
+			fmt.Println(line)
+		}
+	}
+
 	held, err := inv.Profile(ctx, name)
 	if err != nil {
 		return err

cmd/mesh-controller/plan.go

@@ -17,6 +17,7 @@ import (
 	"github.com/novox/mesh-controller/internal/catalogue"
 	"github.com/novox/mesh-controller/internal/inventory"
 	"github.com/novox/mesh-controller/internal/licences"
+	"github.com/novox/mesh-controller/internal/link"
 	"github.com/novox/mesh-controller/internal/overlay"
 )
 
@@ -860,7 +861,12 @@ func renderingFor(ctx context.Context, open *stores, node string,
 	if artifactStore != "" {
 		reach = map[string]string{"mesh-artifact-store": artifactStore}
 	}
+	readsHealth, err := engineReadsHealth(ctx, inv, node)
+	if err != nil {
+		return catalogue.Rendering{}, inventory.Node{}, err
+	}
 	return catalogue.Rendering{
+		ReadsHealth:   readsHealth,
 		BusMembership: memberships[node],
 		Settings:      settings, Generators: gens, Grants: grants, Needed: needed, Foreseen: foreseen, Ports: ports,
 		Certificate: certificate, Authority: authority, Mesh: private, Names: names,
@@ -872,6 +878,17 @@ func renderingFor(ctx context.Context, open *stores, node string,
 	}, record, nil
 }
 
+// engineReadsHealth says whether a machine's node-engine reads a declared `health` (novox/hq ADR 0240
+// Phase B), by its own newest statement: one older, or one that never stated anything, is not sent the
+// field, because it parses strictly and would refuse the whole declaration for it.
+func engineReadsHealth(ctx context.Context, inv *inventory.Inventory, node string) (bool, error) {
+	stated, had, err := inv.HealthOf(ctx, node)
+	if err != nil {
+		return false, err
+	}
+	return had && stated.Contract >= link.ReadinessContract, nil
+}
+
 // zonesInTheMesh is every zone a module in the mesh declares, where the mesh placed it (novox/hq ADR
 // 0199): the zone settled from that node's settings, the node's private address, the port the
 // answering listen is published on there.

cmd/mesh-controller/provider_hold.go

@@ -0,0 +1,192 @@
+package main
+
+import (
+	"context"
+	"fmt"
+	"sort"
+	"strings"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/conditions"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A provider down is said once, at the provider (novox/hq ADR 0240 rule 5, to-be 48 §6, Phase C).
+//
+// With twelve consumers of the database provision and thirty-six of a route, one provider down would be
+// twelve conditions for one fault and twelve gates failed for something none of them did. So a consumer's
+// check names, in `needs`, the provision it exercises; while **that provision's provider for this
+// consumer** — the one the controller composed the consumer against: its recorded binding (ADR 0232), or
+// the provider its credential for the provision is from — is unhealthy on the record, what the check finds
+// is held under the provider's condition: listed there as waiting on it, raised as nothing of its own, and
+// the consumer's gate waits rather than fails. The provider's condition is urgent while consumers wait.
+//
+// **Only what the check finds is held.** A consumer that is down or restarting is its own, whatever its
+// provider does; so is anything a check that names no provision finds, and anything found while the
+// provider is healthy. A machine-level fault is never pinned on a module (issue 281), and this does not
+// change that.
+
+// holding is what one reading of the record needs to say who waits on whom: every machine's newest
+// statement, the open conditions, and the catalogue — read once, asked many times.
+type holding struct {
+	ctx     context.Context
+	inv     *inventory.Inventory
+	healths map[string]inventory.NodeHealth
+	open    []conditions.Condition
+	shelf   map[string]catalogue.Manifest
+	// providers memoises providerFor by machine, consumer and provision.
+	providers map[string]providerLookup
+}
+
+type providerLookup struct {
+	chosen catalogue.Chosen
+	ok     bool
+}
+
+// readHolding reads what the hold is judged from.
+func readHolding(ctx context.Context, inv *inventory.Inventory, open []conditions.Condition) (*holding, error) {
+	healths, err := inv.Healths(ctx)
+	if err != nil {
+		return nil, err
+	}
+	return &holding{ctx: ctx, inv: inv, healths: healths, open: open, providers: map[string]providerLookup{}}, nil
+}
+
+// heldFinding says a resource's state is a finding of its declared check that names a provision: what
+// may be held. Down and restarting are liveness, the resource's own.
+func heldFinding(r inventory.ResourceHealth) bool {
+	return r.State == link.StateUnhealthy && r.Check != "" && r.Needs != "" &&
+		r.Reason != "down" && r.Reason != "restarting"
+}
+
+// providerFor is the provider composed for a consumer's provision: its recorded binding, else the
+// machine its credential for the provision comes from and the module there that provides it.
+func (h *holding) providerFor(machine, consumer, provision string) (catalogue.Chosen, bool) {
+	key := machine + "\x00" + consumer + "\x00" + provision
+	if p, known := h.providers[key]; known {
+		return p.chosen, p.ok
+	}
+	chosen, ok := h.lookUpProvider(machine, consumer, provision)
+	h.providers[key] = providerLookup{chosen, ok}
+	return chosen, ok
+}
+
+func (h *holding) lookUpProvider(machine, consumer, provision string) (catalogue.Chosen, bool) {
+	if bound, err := h.inv.BindingsFor(h.ctx, machine); err == nil {
+		if c, ok := bound[consumer][provision]; ok && c.Node != "" && c.Module != "" {
+			return c, true
+		}
+	}
+	secrets, err := h.inv.SecretsOf(h.ctx, machine, consumer)
+	if err != nil {
+		return catalogue.Chosen{}, false
+	}
+	for _, s := range secrets {
+		if s.Name != provision || s.Provider == "" {
+			continue
+		}
+		if h.shelf == nil {
+			if h.shelf, err = h.inv.Catalogue(h.ctx); err != nil {
+				return catalogue.Chosen{}, false
+			}
+		}
+		assigned, err := h.inv.Assigned(h.ctx, s.Provider)
+		if err != nil {
+			return catalogue.Chosen{}, false
+		}
+		for _, module := range assigned {
+			for _, offer := range h.shelf[module].Offers() {
+				if offer == provision {
+					return catalogue.Chosen{Node: s.Provider, Module: module}, true
+				}
+			}
+		}
+	}
+	return catalogue.Chosen{}, false
+}
+
+// unhealthy says a provider is unhealthy on the record: its condition is open, or its machine's newest
+// statement says a resource of it is unhealthy.
+func (h *holding) unhealthy(p catalogue.Chosen) bool {
+	key := moduleUnhealthyKey(p.Module, p.Node)
+	for _, c := range h.open {
+		if c.Key == key {
+			return true
+		}
+	}
+	for _, r := range h.healths[p.Node].Resources {
+		if r.Module == p.Module && r.State == link.StateUnhealthy {
+			return true
+		}
+	}
+	return false
+}
+
+// heldUnder is the provider a consumer's unhealthy resources wait on: when every one of them is a finding
+// of a check naming a provision whose provider for this consumer is unhealthy on the record. False when any
+// is the consumer's own.
+func (h *holding) heldUnder(machine, module string, rs []inventory.ResourceHealth) (catalogue.Chosen, bool) {
+	var on catalogue.Chosen
+	for _, r := range rs {
+		if r.State != link.StateUnhealthy {
+			continue
+		}
+		if !heldFinding(r) {
+			return catalogue.Chosen{}, false
+		}
+		p, ok := h.providerFor(machine, module, r.Needs)
+		if !ok || (p.Node == machine && p.Module == module) || !h.unhealthy(p) {
+			return catalogue.Chosen{}, false
+		}
+		on = p
+	}
+	return on, on.Module != ""
+}
+
+// waitersOn is every consumer held under a provider, as "<module> on <machine>", sorted.
+func (h *holding) waitersOn(p catalogue.Chosen) []string {
+	var out []string
+	for machine, nh := range h.healths {
+		byModule := map[string][]inventory.ResourceHealth{}
+		for _, r := range nh.Resources {
+			if r.Module != "" && r.State == link.StateUnhealthy {
+				byModule[r.Module] = append(byModule[r.Module], r)
+			}
+		}
+		for module, rs := range byModule {
+			if on, held := h.heldUnder(machine, module, rs); held && on == p {
+				out = append(out, module+" on "+machine)
+			}
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// heldModules is, for one machine's statement, each module whose finding is held, with the provider.
+func (h *holding) heldModules(machine string) map[string]catalogue.Chosen {
+	out := map[string]catalogue.Chosen{}
+	byModule := map[string][]inventory.ResourceHealth{}
+	for _, r := range h.healths[machine].Resources {
+		if r.Module != "" && r.State == link.StateUnhealthy {
+			byModule[r.Module] = append(byModule[r.Module], r)
+		}
+	}
+	for module, rs := range byModule {
+		if on, held := h.heldUnder(machine, module, rs); held {
+			out[module] = on
+		}
+	}
+	return out
+}
+
+// moduleUnhealthyKey is a module's health condition's key on a machine.
+func moduleUnhealthyKey(module, machine string) string {
+	return conditions.ScopeModule + "." + module + "." + machine + ".unhealthy"
+}
+
+// waitingWords is the provider's evidence that consumers wait on it.
+func waitingWords(waiters []string) string {
+	return fmt.Sprintf("waiting on it — %s", strings.Join(waiters, ", "))
+}

cmd/mesh-controller/provider_hold_test.go

@@ -0,0 +1,172 @@
+package main
+
+import (
+	"context"
+	"slices"
+	"strings"
+	"testing"
+	"time"
+
+	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/conditions"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// A provider down is said once, at the provider (novox/hq ADR 0240 rule 5, "how it is checked"): one
+// unhealthy database provider and three consumers failing their checks that need it — one condition, at the
+// provider, urgent, the consumers listed as waiting; their gates wait, not fail; once the provider is
+// healthy, a consumer still failing is its own. A consumer failing while its provider is healthy is raised
+// on its own from the start.
+func TestOneProviderDownAndThreeConsumersFailingAreOneCondition(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	k := conditionsFrom
+	register(t, open, catalogue.Manifest{Module: "db", Version: "1",
+		Provides: []catalogue.Offer{{Name: "postgres-database", Scope: catalogue.ScopeMesh}}})
+	consumers := []string{"shop", "wiki", "crm"}
+	for _, c := range consumers {
+		register(t, open, catalogue.Manifest{Module: c, Version: "1", Requires: []string{"postgres-database"}})
+	}
+	if _, err := inv.Assign(ctx, "anchor", "db"); err != nil {
+		t.Fatal(err)
+	}
+	machineOf := map[string]string{"shop": "laptop", "wiki": "laptop", "crm": "anchor"}
+	for _, c := range consumers {
+		if _, err := inv.Assign(ctx, machineOf[c], c); err != nil {
+			t.Fatal(err)
+		}
+		if err := inv.RecordBindings(ctx, machineOf[c], []inventory.Binding{{Machine: machineOf[c], Consumer: c,
+			Provision: "postgres-database", Provider: catalogue.Chosen{Node: "anchor", Module: "db"}}}); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	at := h0
+	say := func(machine string, rs ...link.ResourceHealth) {
+		t.Helper()
+		at = at.Add(time.Second)
+		for i := range rs {
+			rs[i].Since = at
+		}
+		if err := stateHealth(ctx, inv, k, machine, link.Health{Contract: link.ReadinessContract, At: at, Resources: rs}, at); err != nil {
+			t.Fatal(err)
+		}
+	}
+	dbDown := link.ResourceHealth{Module: "db", Resource: "db.server", Kind: "container", Target: "db", State: link.StateUnhealthy,
+		Reason: "its command: the database refuses connections", Check: "exec"}
+	dbUp := dbDown
+	dbUp.State, dbUp.Reason = link.StateHealthy, ""
+	failing := func(module string) link.ResourceHealth {
+		return link.ResourceHealth{Module: module, Resource: module + ".web", Kind: "container", Target: module,
+			State: link.StateUnhealthy, Reason: "http /health on web: answered 500", Check: "http", Needs: "postgres-database"}
+	}
+	fine := func(module string) link.ResourceHealth {
+		r := failing(module)
+		r.State, r.Reason = link.StateHealthy, ""
+		return r
+	}
+	openKeys := func() []conditions.Condition {
+		t.Helper()
+		list, err := k.Open(ctx)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return list
+	}
+
+	// Two looks of the provider down and its consumers failing, in either order on each machine.
+	for look := 0; look < 2; look++ {
+		say("laptop", failing("shop"), failing("wiki"))
+		say("anchor", dbDown, failing("crm"))
+	}
+	raised := openKeys()
+	if len(raised) != 1 || raised[0].Key != "module.db.anchor.unhealthy" {
+		var keys []string
+		for _, c := range raised {
+			keys = append(keys, c.Key)
+		}
+		t.Fatalf("one provider down and three consumers failing raised %v; want the provider's alone", keys)
+	}
+	c := raised[0]
+	if c.Severity != conditions.Urgent {
+		t.Errorf("consumers wait on the provider and its condition is %s", c.Severity)
+	}
+	said := c.Evidence[0].Said
+	for _, w := range []string{"shop on laptop", "wiki on laptop", "crm on anchor"} {
+		if !strings.Contains(said, w) {
+			t.Errorf("the provider's condition does not list %s as waiting on it: %s", w, said)
+		}
+	}
+
+	// Their gates wait, not fail: the judging is neither a pass nor a fault, past the bound too.
+	f, err := gatherGateFacts(ctx, open, "")
+	if err != nil {
+		t.Fatal(err)
+	}
+	f.open, f.openErr, f.judged = raised, nil, false
+	for _, m := range consumers {
+		h, why := moduleHealthWord(m, machineOf[m], h0, f)
+		if h != healthWaiting || !strings.Contains(why, "waits on db on anchor") {
+			t.Errorf("%s's gate: %v %q; want it waiting on its provider", m, h, why)
+		}
+	}
+	// And a whole judging past the bound puts nothing back: it waits.
+	long := h0.Add(-time.Hour)
+	for _, m := range []string{"anchor", "laptop"} {
+		f.reports[m] = inventory.Reported{Node: m, Outcome: inventory.OutcomeApplied, At: &f.now, Current: true}
+	}
+	gatherWas := gatherGateFacts
+	defer func() { gatherGateFacts = gatherWas }()
+	gatherGateFacts = func(context.Context, *stores, string) (gateFacts, error) { return f, nil }
+	g := &inventory.PlanGate{Machines: []string{"laptop"}, Since: &long}
+	verdict, err := judgeMoves(ctx, open, g, []judged{{module: "shop", node: "laptop"}}, time.Now())
+	if err != nil || verdict != "" || !strings.Contains(g.Last, "waits on db on anchor") {
+		t.Fatalf("a held consumer's gate past its bound: verdict %q (%v), last %q; want it waiting", verdict, err, g.Last)
+	}
+
+	// The provider healthy again: a consumer still failing is now its own, at once (its streak stood).
+	say("anchor", dbUp, fine("crm"))
+	say("laptop", failing("shop"), fine("wiki"))
+	var keys []string
+	for _, c := range openKeys() {
+		keys = append(keys, c.Key)
+	}
+	if !slices.Equal(keys, []string{"module.shop.laptop.unhealthy"}) {
+		t.Fatalf("after the provider recovered: %v; want shop's own and nothing else", keys)
+	}
+}
+
+// A consumer failing while its provider is healthy is raised on its own.
+func TestAConsumerFailingBesideAHealthyProviderIsItsOwn(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+	register(t, open, catalogue.Manifest{Module: "db", Version: "1",
+		Provides: []catalogue.Offer{{Name: "postgres-database", Scope: catalogue.ScopeMesh}}})
+	register(t, open, catalogue.Manifest{Module: "shop", Version: "1", Requires: []string{"postgres-database"}})
+	for _, a := range [][2]string{{"anchor", "db"}, {"laptop", "shop"}} {
+		if _, err := inv.Assign(ctx, a[0], a[1]); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := inv.RecordBindings(ctx, "laptop", []inventory.Binding{{Machine: "laptop", Consumer: "shop",
+		Provision: "postgres-database", Provider: catalogue.Chosen{Node: "anchor", Module: "db"}}}); err != nil {
+		t.Fatal(err)
+	}
+	healthy := link.ResourceHealth{Module: "db", Resource: "db.server", Kind: "container", Target: "db", State: link.StateHealthy, Check: "exec"}
+	shop := link.ResourceHealth{Module: "shop", Resource: "shop.web", Kind: "container", Target: "shop", State: link.StateUnhealthy,
+		Reason: "http /health on web: answered 500", Check: "http", Needs: "postgres-database"}
+	for i := 1; i <= 2; i++ {
+		at := h0.Add(time.Duration(i) * time.Minute)
+		_ = stateHealth(ctx, inv, conditionsFrom, "anchor", link.Health{Contract: 2, At: at, Resources: []link.ResourceHealth{healthy}}, at)
+		if err := stateHealth(ctx, inv, conditionsFrom, "laptop", link.Health{Contract: 2, At: at, Resources: []link.ResourceHealth{shop}}, at); err != nil {
+			t.Fatal(err)
+		}
+	}
+	list, _ := conditionsFrom.Open(ctx)
+	if len(list) != 1 || list[0].Key != "module.shop.laptop.unhealthy" {
+		t.Fatalf("a consumer failing beside a healthy provider raised %v", list)
+	}
+}

cmd/mesh-controller/push.go

@@ -128,6 +128,10 @@ func serve(ctx context.Context) (err error) {
 			stopActing()
 		case <-ctx.Done():
 		}
+		// Stopping, whichever way: a verb arriving from here is refused as a handover, so its caller
+		// asks the controller after this one rather than have a command started and killed with this
+		// process (novox/hq issue 289).
+		handingOver.Store(true)
 	}()
 	defer func() {
 		select {
@@ -182,6 +186,9 @@ func serve(ctx context.Context) (err error) {
 	if err := server.Watches(standings{keeper: func() *conditions.Keeper { return conditionsFrom }}); err != nil {
 		return err
 	}
+	// And what each machine says of what it runs between its reports, kept and raised from (novox/hq ADR
+	// 0240): the gate, `node show` and the conditions read it.
+	server.Hears(moduleHealth{inv: inv, keeper: func() *conditions.Keeper { return conditionsFrom }})
 	// And what they say about consumers the mesh stopped asking for: one waiting for a person is an
 	// urgent condition, and an act asked of a provider some other way is recorded by hand (ADR 0230).
 	if err := server.KeepsRetirements(retirements{keeper: func() *conditions.Keeper { return conditionsFrom },

cmd/mesh-controller/replays_test.go

@@ -1,9 +1,18 @@
 package main
 
 import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+	"slices"
 	"testing"
+	"time"
 
 	"github.com/novox/mesh-controller/internal/catalogue"
+	"github.com/novox/mesh-controller/internal/conditions"
+	"github.com/novox/mesh-controller/internal/inventory"
+	"github.com/novox/mesh-controller/internal/link"
 )
 
 // The replays of the controller's incidents (novox/hq to-be 45 §9, M9): each a scripted replay of what
@@ -111,3 +120,224 @@ func TestReplay273AConsumerBesideItsStoreStaysBoundToIt(t *testing.T) {
 		t.Fatalf("the resolver was bound to %q; its seat is held on anchor (issue 258)", network)
 	}
 }
+
+// **R-crashloop — a module that crash-loops after it applied fails its gate on its first machine** (novox/hq
+// ADR 0240; research 032 §6). On the home server the agent server restarted about a hundred times while the
+// mesh read it applied, its tools served and no condition raised — the gate judged a module by what the
+// mesh saw from outside, and nothing looked at what it ran. The outcome asserted: a build whose container
+// exits at start never passes its gate on the first machine, is put back there at the bound, and never
+// reaches the second.
+//
+// The machine is heard as a node-engine says it: its report of the apply, then the same account said again
+// later, each carrying what the engine states of what it runs — `starting` as the apply ends, then the
+// crash loop. What it states of the crash loop is MESH_REPLAY_STATEMENT when the lab's replay ran the
+// engine against a real container (mesh-lab replays, R-crashloop), and otherwise what the engine said of
+// one, kept below. Heard as bytes, so an older controller reads them as it reads any report — this file is
+// written only with what the controller had before the judging, for the prover to lay over that commit.
+func TestReplayCrashLoopFailsItsGateOnTheFirstMachine(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	inv := open.inventory
+
+	crashLoop := []byte(`{"contract":1,"at":"2026-10-07T00:00:00Z","resources":[{"module":"app","resource":"app.server",` +
+		`"kind":"container","target":"app-server","state":"unhealthy","reason":"restarting","since":"2026-10-07T00:00:00Z",` +
+		`"streak":5,"restarts":2}]}`)
+	if path := os.Getenv("MESH_REPLAY_STATEMENT"); path != "" {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			t.Fatalf("the engine's statement of the crash loop: %v", err)
+		}
+		crashLoop = raw
+	}
+	// `null` is an engine that states nothing — older than the judging — and its reports carry no health.
+	var stated map[string]any
+	if err := json.Unmarshal(crashLoop, &stated); err != nil {
+		t.Fatal(err)
+	}
+
+	for _, b := range []inventory.Build{
+		{ID: "build-1", Module: "app", Commit: "c1", Repository: "novox/mesh-catalog", Path: "modules/app",
+			Asked: time.Now().Add(-2 * time.Hour), At: time.Now().Add(-2 * time.Hour)},
+		{ID: "build-2", Module: "app", Commit: "c2", Repository: "novox/mesh-catalog", Path: "modules/app",
+			Asked: time.Now().Add(-time.Minute), At: time.Now().Add(-time.Minute)},
+	} {
+		manifest, _ := json.Marshal(catalogue.Manifest{Module: "app", Version: b.Commit})
+		b.Manifest = manifest
+		if err := inv.RecordBuild(ctx, b); err != nil {
+			t.Fatal(err)
+		}
+	}
+	registerAt := func(commit string, asked time.Time) {
+		if err := inv.RegisterModule(ctx, catalogue.Manifest{Module: "app", Version: commit},
+			inventory.Source{Repository: "novox/mesh-catalog", Seat: "git", Path: "modules/app", BuiltFrom: commit,
+				Head: commit, Asked: asked}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	registerAt("c1", time.Now().Add(-2*time.Hour))
+	for _, n := range []string{"anchor", "laptop"} {
+		if _, err := inv.Assign(ctx, n, "app"); err != nil {
+			t.Fatal(err)
+		}
+		if err := inv.RecordSent(ctx, nodeID(t, open, n), "d-"+n+"-c1", map[string]string{"app": "c1"}); err != nil {
+			t.Fatal(err)
+		}
+	}
+	registerAt("c2", time.Now().Add(-time.Minute))
+
+	// The machine: what it is sent is applied, and its report says what runs is starting.
+	listener := nudgingListener{Enrolment: link.Enrolment{Inventory: inv}}
+	sequence := int64(0)
+	say := func(node, digest, state string) {
+		t.Helper()
+		sequence++
+		health := map[string]any{}
+		for k, v := range stated {
+			health[k] = v
+		}
+		health["at"] = time.Now().UTC().Format(time.RFC3339Nano)
+		if state != "" && stated != nil {
+			resources := []any{}
+			for _, r := range stated["resources"].([]any) {
+				kept := map[string]any{}
+				for k, v := range r.(map[string]any) {
+					kept[k] = v
+				}
+				kept["state"], kept["reason"] = state, ""
+				resources = append(resources, kept)
+			}
+			health["resources"] = resources
+		}
+		said := map[string]any{"node": node, "applied": []string{"app.server"}, "declared": digest,
+			"report_sequence": sequence}
+		if stated != nil {
+			said["health"] = health
+		}
+		body, _ := json.Marshal(said)
+		var report link.Report
+		if err := json.Unmarshal(body, &report); err != nil {
+			t.Fatal(err)
+		}
+		if _, err := listener.Heard(ctx, report); err != nil {
+			t.Fatal(err)
+		}
+	}
+	var sent [][]string
+	last := map[string]string{}
+	n := 0
+	wasSend := sendRollout
+	sendRollout = func(ctx context.Context, open *stores, names []string) ([]string, error) {
+		sent = append(sent, append([]string(nil), names...))
+		current, err := open.inventory.CurrentBuilds(ctx)
+		if err != nil {
+			return nil, err
+		}
+		for _, node := range names {
+			n++
+			digest := fmt.Sprintf("d-%s-%d", node, n)
+			if err := open.inventory.RecordSent(ctx, nodeID(t, open, node), digest, map[string]string{"app": current["app"].Commit}); err != nil {
+				return nil, err
+			}
+			last[node] = digest
+			say(node, digest, "starting")
+		}
+		return names, nil
+	}
+	t.Cleanup(func() { sendRollout = wasSend })
+	wasSettle, wasEvery, wasBound := gateSettle, gateEvery, gateBound
+	gateSettle, gateEvery = 0, 0
+	t.Cleanup(func() { gateSettle, gateEvery, gateBound = wasSettle, wasEvery, wasBound })
+
+	built := time.Now().UTC()
+	plan := inventory.Plan{ID: "plan-crashloop", Repository: "novox/mesh-catalog", Branch: "main", Commit: "c2",
+		Created: built, State: inventory.PlanBuilding, Tiers: [][]string{{"app"}},
+		Modules: map[string]*inventory.PlanModule{"app": {State: "built", BuiltAt: &built, Commit: "c2", Build: "build-2"}}}
+	if err := inv.SavePlan(ctx, &plan); err != nil {
+		t.Fatal(err)
+	}
+
+	advancePlans(ctx, open) // the first machine is sent the new build, and starts it
+	if len(sent) == 0 || !slices.Equal(sent[0], []string{"anchor"}) {
+		t.Fatalf("sent %v, not the first machine first", sent)
+	}
+	advancePlans(ctx, open) // judged while it starts
+	// Its container exits at start, and the runtime restarts it: the machine says so, again and again.
+	for i := 0; i < 6 && len(sent) == 1; i++ {
+		say("anchor", last["anchor"], "")
+		advancePlans(ctx, open)
+	}
+	gateBound = -time.Second // and the bound passes
+	advancePlans(ctx, open)
+
+	p, err := inv.PlanByID(ctx, "plan-crashloop")
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate := p.Modules["app"].Gate
+	for _, names := range sent {
+		if slices.Contains(names, "laptop") {
+			t.Fatalf("the crash loop passed its gate on anchor and was sent to laptop: sent %v, the gate %+v", sent, gate)
+		}
+	}
+	if gate == nil || gate.Verdict != inventory.GateFailed || p.State != inventory.PlanFailed {
+		t.Fatalf("a crash-looping build did not fail its gate on the first machine: the plan is %s (%s), its gate %+v",
+			p.State, p.Note, gate)
+	}
+	if current, _ := inv.CurrentBuilds(ctx); current["app"].Commit != "c1" {
+		t.Fatalf("the module is registered at %s, not put back to c1", current["app"].Commit)
+	}
+}
+
+// **R145 — a web application that accepts TCP and answers nothing is raised within two looks** (novox/hq
+// ADR 0240 rule 4 and Phase B, issue 145). For eleven hours a web application's port was open and its
+// program ran while every request hung, and the mesh said its machine was healthy; a person found it.
+// Liveness cannot see it and a TCP check cannot either: the port is open. The module's declared HTTP check
+// can. The engine's half (mesh-host internal/liveness TestReplaySilentWebAppIsSaidUnhealthy) states what it
+// found looking at such a program; here the controller hears that statement on two looks in a row and
+// raises the module's condition — the second, never the first. `null` is an engine older than the
+// readiness check: it states the program alive, and nothing is raised.
+func TestReplaySilentWebAppIsRaisedWithinTwoLooks(t *testing.T) {
+	open := aMesh(t)
+	ctx := t.Context()
+	silent := []byte(`{"contract":2,"at":"2026-10-07T00:00:00Z","resources":[{"module":"app","resource":"app.server",` +
+		`"kind":"container","target":"app-server","state":"unhealthy","reason":"http / on web: no answer within 5s",` +
+		`"since":"2026-10-07T00:00:00Z","streak":3,"check":"http"}]}`)
+	if path := os.Getenv("MESH_REPLAY_STATEMENT"); path != "" {
+		raw, err := os.ReadFile(path)
+		if err != nil {
+			t.Fatalf("the engine's statement of the silent web application: %v", err)
+		}
+		silent = raw
+	}
+	var h link.Health
+	if err := json.Unmarshal(silent, &h); err != nil {
+		t.Fatal(err)
+	}
+	if h.Contract == 0 {
+		t.Fatal("the engine states nothing of the silent web application: it is older than the judging")
+	}
+	open1 := func() []conditions.Condition {
+		t.Helper()
+		list, err := conditionsFrom.Open(ctx)
+		if err != nil {
+			t.Fatal(err)
+		}
+		return list
+	}
+	for look := 1; look <= 2; look++ {
+		at := time.Now().Add(time.Duration(look) * time.Second)
+		h.At = at
+		if err := stateHealth(ctx, open.inventory, conditionsFrom, "anchor", h, at); err != nil {
+			t.Fatal(err)
+		}
+		raised := open1()
+		switch {
+		case look == 1 && len(raised) != 0:
+			t.Fatalf("one look raised %v", raised[0].Key)
+		case look == 2 && (len(raised) != 1 || raised[0].Key != "module.app.anchor.unhealthy"):
+			t.Fatalf("two looks in a row did not raise the module's condition: %v", raised)
+		case look == 2:
+			t.Logf("raised on the second look: %s", raised[0].Summary)
+		}
+	}
+}

cmd/mesh-controller/seatverbs.go

@@ -456,6 +456,15 @@ func (a *verbArguments) commandLine() ([]string, error) {
 			argv = append(argv, "--condition", c)
 		}
 		return argv, nil
+	case "drill":
+		if err := need("what", "why"); err != nil {
+			return nil, err
+		}
+		argv := []string{"hand-act", "drill", str("what"), "--why", str("why")}
+		if c := str("condition"); c != "" {
+			argv = append(argv, "--condition", c)
+		}
+		return argv, nil
 	case "hand-acts":
 		argv := []string{"hand-acts", "--json"}
 		if d := str("days"); d != "" {
@@ -723,6 +732,8 @@ func repairingCommand(argv []string) string {
 		return "plans " + argv[1]
 	case argv[0] == "broker" && len(argv) > 1 && argv[1] == "consumer-reset":
 		return "broker consumer-reset"
+	case argv[0] == "hand-act" && len(argv) > 1 && argv[1] == "drill":
+		return "hand-act drill"
 	case argv[0] == "hand-act":
 		return "hand-act record"
 	case argv[0] == "conditions" && len(argv) > 1 && argv[1] == "silence":
@@ -740,12 +751,18 @@ func isWhyFlag(word string) bool {
 }
 
 // runVerb runs this binary with the given command line and gathers what it said.
+//
+// **This binary is the image this process runs, never the file at the path it started from**
+// (novox/hq issue 289): the node-engine's witness moves a running build aside when it places the next
+// one, into a directory only it may enter, and deletes it once the next is proved — while this process
+// may still be serving. A verb run from the path then failed with "permission denied" for the seconds
+// the old controller still answered. selfCommand runs what this process is running, wherever its file
+// went; a verb it still cannot start is refused as a handover, which the caller asks again.
 func runVerb(ctx context.Context, argv []string) (verbAnswer, error) {
-	self, err := os.Executable()
-	if err != nil {
-		return verbAnswer{}, err
+	if handingOver.Load() {
+		return verbAnswer{}, fmt.Errorf("%w: this controller is stopping and runs no new command", link.ErrHandingOver)
 	}
-	cmd := exec.CommandContext(ctx, self, argv...)
+	cmd := selfCommand(ctx, argv)
 	// The same environment: the stores' credentials, the bus, the broker — everything a command run
 	// from a shell in this container would have, because it is that.
 	cmd.Env = os.Environ()
@@ -773,6 +790,12 @@ func runVerb(ctx context.Context, argv []string) (verbAnswer, error) {
 	var exit *exec.ExitError
 	if runErr != nil && !errors.As(runErr, &exit) {
 		// Not the command refusing — the command not running at all, which is this process's fault.
+		if errors.Is(runErr, os.ErrPermission) || errors.Is(runErr, os.ErrNotExist) {
+			// Its own image unreachable: a build replaced under a process that has not yet stopped.
+			// Refused as a handover, so the caller asks the controller that follows.
+			return answer, fmt.Errorf("%w: could not run %s from this controller's own build: %v",
+				link.ErrHandingOver, strings.Join(argv, " "), runErr)
+		}
 		return answer, fmt.Errorf("could not run %s: %w", strings.Join(argv, " "), runErr)
 	}
 	return answer, nil

cmd/mesh-controller/selfexec.go

@@ -0,0 +1,46 @@
+package main
+
+import (
+	"context"
+	"os"
+	"os/exec"
+	"runtime"
+	"sync/atomic"
+)
+
+// startedFrom is the path this process's executable had when it started: what a verb's command line
+// is named in a process listing, and what is run where the image cannot be named otherwise.
+var startedFrom, _ = os.Executable()
+
+// ownImage is how a process names the executable it is running, as long as it runs, wherever the file
+// has gone since. On Linux the kernel keeps it: /proc/self/exe is the running image itself, not a path,
+// so it is valid after the file is renamed into a directory this process may not enter, or deleted —
+// which is what the node-engine's witness does to a build it replaces (novox/hq issue 289). Read in the
+// child, it names the child's image, which until the exec is this process's.
+//
+// os.Executable reads the same link and returns the path it points at *now*: correct at start and
+// wrong the moment the file moves, which is the fault. A variable so a test can name another.
+var ownImage = func() string {
+	if runtime.GOOS == "linux" {
+		if _, err := os.Stat("/proc/self/exe"); err == nil {
+			return "/proc/self/exe"
+		}
+	}
+	return startedFrom
+}
+
+// selfCommand is this binary run with a command line: the image this process runs, named in a process
+// listing as the path it started from.
+func selfCommand(ctx context.Context, argv []string) *exec.Cmd {
+	cmd := exec.CommandContext(ctx, ownImage(), argv...)
+	if startedFrom != "" {
+		cmd.Args[0] = startedFrom
+	}
+	return cmd
+}
+
+// handingOver is set when the serving controller begins to stop — a signal from its supervisor, a lease
+// lost. From then a verb that would run a command is refused as a handover rather than started and
+// killed with this process: the caller asks again, and the controller after this one answers
+// (novox/hq issue 289).
+var handingOver atomic.Bool

cmd/mesh-controller/selfexec_test.go

@@ -0,0 +1,177 @@
+package main
+
+import (
+	"bufio"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-controller/internal/link"
+)
+
+// The roles a copy of this test binary plays in TestAVerbRunsWhileItsBuildIsMovedOrDeleted.
+const selfExecRole = "MESH_CONTROLLER_SELFEXEC_ROLE"
+
+// TestSelfExecServer is a controller standing in, when run as one: it waits until its build has been
+// moved, then runs a verb as the seat runs one, and prints what came of it as JSON.
+func TestSelfExecServer(t *testing.T) {
+	if os.Getenv(selfExecRole) != "server" {
+		t.Skip("run by TestAVerbRunsWhileItsBuildIsMovedOrDeleted")
+	}
+	line, _ := bufio.NewReader(os.Stdin).ReadString('\n')
+	if strings.TrimSpace(line) != "go" {
+		t.Fatalf("told %q", line)
+	}
+	if err := os.Setenv(selfExecRole, "verb"); err != nil {
+		t.Fatal(err)
+	}
+	answer, err := runVerb(t.Context(), []string{"-test.run", "^TestSelfExecVerb$", "-test.v"})
+	said := map[string]any{"ok": answer.OK, "output": answer.Output}
+	if err != nil {
+		said["error"] = err.Error()
+	}
+	body, _ := json.Marshal(said)
+	fmt.Println("ANSWER " + string(body))
+}
+
+// TestSelfExecVerb is the verb, when run as one.
+func TestSelfExecVerb(t *testing.T) {
+	if os.Getenv(selfExecRole) != "verb" {
+		t.Skip("run by TestSelfExecServer")
+	}
+	fmt.Println("the verb ran")
+}
+
+// **A verb runs while the build it was started from is moved where its user may not go, or deleted**
+// (novox/hq issue 289). The node-engine's witness moves a running controller's build into a directory
+// only it may enter as it places the next, and deletes it once the next is proved; the controller
+// still serving in between ran its verbs from the path and answered "permission denied".
+//
+// A copy of this test binary is the controller: started from one place, moved into a directory closed
+// to everyone (or deleted), and only then asked to run a verb.
+func TestAVerbRunsWhileItsBuildIsMovedOrDeleted(t *testing.T) {
+	if runtime.GOOS != "linux" {
+		t.Skip("the image is named through /proc on Linux only")
+	}
+	self, err := os.Executable()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, how := range []string{"moved into a closed directory", "deleted"} {
+		t.Run(how, func(t *testing.T) {
+			root := t.TempDir()
+			placed := filepath.Join(root, "mesh-controller", "mesh-controller")
+			if err := os.MkdirAll(filepath.Dir(placed), 0o755); err != nil {
+				t.Fatal(err)
+			}
+			copyFile(t, self, placed)
+
+			server := exec.Command(placed, "-test.run", "^TestSelfExecServer$", "-test.v")
+			server.Env = append(os.Environ(), selfExecRole+"=server")
+			stdin, err := server.StdinPipe()
+			if err != nil {
+				t.Fatal(err)
+			}
+			stdout, err := server.StdoutPipe()
+			if err != nil {
+				t.Fatal(err)
+			}
+			server.Stderr = os.Stderr
+			if err := server.Start(); err != nil {
+				t.Fatal(err)
+			}
+			t.Cleanup(func() { _ = server.Process.Kill(); _ = server.Wait() })
+
+			// What the witness does to a running build, once the process runs.
+			switch how {
+			case "deleted":
+				if err := os.RemoveAll(filepath.Dir(placed)); err != nil {
+					t.Fatal(err)
+				}
+			default:
+				kept := filepath.Join(root, ".witness", "mesh-controller", "previous")
+				if err := os.MkdirAll(filepath.Dir(kept), 0o700); err != nil {
+					t.Fatal(err)
+				}
+				if err := os.Rename(filepath.Dir(placed), kept); err != nil {
+					t.Fatal(err)
+				}
+				if err := os.Chmod(filepath.Join(root, ".witness"), 0); err != nil {
+					t.Fatal(err)
+				}
+				t.Cleanup(func() { _ = os.Chmod(filepath.Join(root, ".witness"), 0o700) })
+			}
+			if _, err := io.WriteString(stdin, "go\n"); err != nil {
+				t.Fatal(err)
+			}
+			out, _ := io.ReadAll(stdout)
+			_ = server.Wait()
+			var said struct {
+				OK     bool   `json:"ok"`
+				Output string `json:"output"`
+				Error  string `json:"error"`
+			}
+			found := false
+			for _, line := range strings.Split(string(out), "\n") {
+				if rest, ok := strings.CutPrefix(line, "ANSWER "); ok {
+					found = json.Unmarshal([]byte(rest), &said) == nil
+				}
+			}
+			if !found {
+				t.Fatalf("the controller standing in answered nothing:\n%s", out)
+			}
+			if said.Error != "" || !said.OK || !strings.Contains(said.Output, "the verb ran") {
+				t.Fatalf("the verb did not run from the controller's own image after its build was %s: %+v", how, said)
+			}
+		})
+	}
+}
+
+// A verb the controller cannot start from its own build is refused as a handover — marked so its
+// caller asks again — never a permission error; and so is one arriving once the controller is stopping.
+func TestAVerbThatCannotRunIsRefusedAsAHandover(t *testing.T) {
+	closed := filepath.Join(t.TempDir(), "closed")
+	if err := os.MkdirAll(closed, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	was := ownImage
+	ownImage = func() string { return filepath.Join(closed, "gone", "mesh-controller") }
+	t.Cleanup(func() { ownImage = was })
+	_, err := runVerb(t.Context(), []string{"status"})
+	if !errors.Is(err, link.ErrHandingOver) {
+		t.Fatalf("a build that is not there was answered %v, not a handover", err)
+	}
+
+	ownImage = was
+	handingOver.Store(true)
+	t.Cleanup(func() { handingOver.Store(false) })
+	if _, err := runVerb(t.Context(), []string{"-test.run", "^$"}); !errors.Is(err, link.ErrHandingOver) {
+		t.Fatalf("a controller stopping ran a verb: %v", err)
+	}
+}
+
+func copyFile(t *testing.T, from, to string) {
+	t.Helper()
+	in, err := os.Open(from)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer in.Close()
+	out, err := os.OpenFile(to, os.O_CREATE|os.O_WRONLY|os.O_TRUNC, 0o755)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := io.Copy(out, in); err != nil {
+		t.Fatal(err)
+	}
+	if err := out.Close(); err != nil {
+		t.Fatal(err)
+	}
+}

cmd/mesh-controller/signals_test.go

@@ -550,3 +550,47 @@ func TestAWalkWaitingFourHoursIsUrgentAndNamesTheWayOn(t *testing.T) {
 		t.Fatalf("it does not say how on: %q", got[0].Summary)
 	}
 }
+
+// **A drill is a person's deliberate test, never a repair** — the log of 2026-10-07: two drills of ADR
+// 0240 recorded through `hand-act record --cause drill` before `hand-act drill` existed raised
+// `mesh.hand-acts.drill.healer-wanted`. A drill through its own verb never counts, the two recorded
+// before it clear on the next tick, and the cause word alone on any other verb still counts — whether
+// an act is a drill is said by the verb chosen, not by a word typed into a repair's cause.
+func TestADrillIsNoRepairAndItsHealerWantedClears(t *testing.T) {
+	now := time.Date(2026, 10, 7, 12, 0, 0, 0, time.UTC)
+	f := calm(now)
+	f.handActs = []link.HandAct{actOf(now.Add(-26*time.Hour), "hand-act drill", causeDrill),
+		actOf(now.Add(-time.Hour), "hand-act drill", causeDrill), actOf(now.Add(-time.Minute), "hand-act drill", causeDrill)}
+	if got := watchHandActs(f); len(got) != 0 {
+		t.Errorf("drills repeated asked for a healer: %+v", got)
+	}
+	for _, verb := range []string{"push", "conditions silence", "plans close", "a verb nobody listed"} {
+		f := calm(now)
+		f.handActs = []link.HandAct{actOf(now.Add(-26*time.Hour), verb, causeDrill), actOf(now.Add(-time.Hour), verb, causeDrill)}
+		if got := watchHandActs(f); len(got) != 1 {
+			t.Errorf("%s with the cause %q passed for a drill: %+v", verb, causeDrill, got)
+		}
+	}
+
+	store := conditions.NewInMemory()
+	k := conditions.NewKeeper(t.Context(), conditions.Options{Store: store, History: store,
+		Teller: &conditions.Told{}, Now: func() time.Time { return now }})
+	defer k.Close(context.Background())
+	before := []conditions.Observation{{Scope: conditions.ScopeMesh, ID: "hand-acts." + causeDrill,
+		Token: "healer-wanted", Kind: "healer-wanted", Severity: conditions.Warning,
+		Summary: "\"drill\" was repaired by hand 2 times in 14 days"}}
+	if err := k.Reconcile(t.Context(), "S15", before); err != nil {
+		t.Fatal(err)
+	}
+	if open, _ := k.Open(t.Context()); len(open) != 1 {
+		t.Fatalf("the condition of the build before was not open: %+v", open)
+	}
+	f = calm(now)
+	f.handActs = []link.HandAct{actOf(now.Add(-11*time.Hour), "hand-act record", causeDrill),
+		actOf(now.Add(-30*time.Minute), "hand-act record", causeDrill)}
+	w := &watchdogs{keeper: k, started: now.Add(-time.Hour)}
+	w.see(t.Context(), f)
+	if open, _ := k.Open(t.Context()); len(open) != 0 {
+		t.Fatalf("a healer-wanted for two drills stayed open: %+v", open)
+	}
+}

cmd/mesh-controller/status_summary.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"fmt"
+	"os"
 	"sync"
 	"time"
 
@@ -204,6 +205,15 @@ func (l nudgingListener) Heard(ctx context.Context, report link.Report) (bool, e
 	if news {
 		l.summary.nudge()
 	}
+	// What it says of its long-running resources (novox/hq ADR 0240): kept, and its modules' conditions
+	// raised or cleared from it. Only from an account of the machine the store took.
+	if err == nil && report.Health != nil && report.Superseded == "" && report.Rekey == nil && report.Node != "" &&
+		l.Enrolment.Inventory != nil {
+		if herr := stateHealth(ctx, l.Enrolment.Inventory, conditionsFrom, report.Node, *report.Health, now); herr != nil {
+			fmt.Fprintf(os.Stderr, "mesh-controller: could not keep what %s says of its resources' health: %v\n",
+				report.Node, herr)
+		}
+	}
 	if err == nil && l.open != nil && startedWell(report) {
 		// Off the report's path: replacing a given value sends the machine, and a report waits for
 		// nothing it caused (novox/hq ADR 0228).

go.mod

@@ -35,4 +35,4 @@ require (
 // committed. Every build (the build agent's `go build`, the Dockerfile) compiles from vendor/ and
 // fetches nothing; go refuses to build when vendor/ and this file disagree, so a pin moved without
 // `go mod vendor` fails loudly, at once, everywhere.
-replace github.com/novox/mesh-host => git.novox.be/novox/mesh-host v0.0.0-20261006095519-3e80b7ae325e
+replace github.com/novox/mesh-host => git.novox.be/novox/mesh-host v0.0.0-20261007120832-bdd44154ccac

go.sum

@@ -1,5 +1,7 @@
 git.novox.be/novox/mesh-host v0.0.0-20261006095519-3e80b7ae325e h1:g9h4QRaAMg5yaJLwqtb0FoOs23DVGUYpW6qvnQ3oY5A=
 git.novox.be/novox/mesh-host v0.0.0-20261006095519-3e80b7ae325e/go.mod h1:VlilMCRZ5yyNXg7SNigNBLr0Gt32jrGw5KSNq5JAVYs=
+git.novox.be/novox/mesh-host v0.0.0-20261007120832-bdd44154ccac h1:KvnKtJ2rWeIE/t4GweK+JL0OjKSNxsrVP3/nMdpii8o=
+git.novox.be/novox/mesh-host v0.0.0-20261007120832-bdd44154ccac/go.mod h1:VlilMCRZ5yyNXg7SNigNBLr0Gt32jrGw5KSNq5JAVYs=
 github.com/antithesishq/antithesis-sdk-go v0.7.0-default-no-op h1:Z/MZK75wC/NSrkgqeNIa7jexam9uWzhLmFTSCPI/kn0=
 github.com/antithesishq/antithesis-sdk-go v0.7.0-default-no-op/go.mod h1:FQyySiasQQM8735Ddel3MRojmy4dA1IqCeyJ5jmPMbI=
 github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=

internal/broker/nats.go

@@ -122,6 +122,10 @@ type Principal struct {
 	// key, and nothing else of the bucket, so it can judge a new controller build and put the previous
 	// one back.
 	WitnessesController bool
+	// Checks are the tools a machine principal's node-engine asks as a declared health check, each
+	// `<module>.<tool>` (novox/hq ADR 0240, to-be 48 §3): asked of the instance on its own machine and
+	// nowhere else.
+	Checks []string
 
 	// PasswordHash is the bcrypt hash the mesh minted. The plaintext is sealed to the principal
 	// and never appears here: this file is written to a node's disk and read by a server, and a
@@ -1021,5 +1025,26 @@ func WitnessSubjects(p Principal) []string {
 	if p.WitnessesController {
 		out = append(out, lease.LeaseReadSubject(LeaseBucket))
 	}
+	return append(out, CheckSubjects(p)...)
+}
+
+// CheckSubjects are the tools a machine's node-engine asks as declared health checks (novox/hq ADR 0240,
+// to-be 48 §3): each `<module>.<tool>` on this machine's instance — `mesh.mod.<module>.tool.<tool>.<node>`
+// — and never the plain subject, which any machine's instance may answer. Sorted and once each.
+func CheckSubjects(p Principal) []string {
+	seen := map[string]bool{}
+	var out []string
+	for _, c := range p.Checks {
+		module, tool, ok := strings.Cut(c, ".")
+		if !ok || !safeSubject.MatchString(module) || !safeSubject.MatchString(tool) {
+			continue
+		}
+		subject := "mesh.mod." + module + ".tool." + tool + "." + p.Node
+		if !seen[subject] {
+			seen[subject] = true
+			out = append(out, subject)
+		}
+	}
+	sort.Strings(out)
 	return out
 }

internal/broker/users.go

@@ -44,6 +44,9 @@ type Declared struct {
 	// SnapshotsTheBus says the module holds mesh-broker — it is the bus — and so is the one module
 	// granted the snapshot API, to copy the bus's streams for the night's backup (novox/hq ADR 0235).
 	SnapshotsTheBus bool
+	// Checks are the module's own tools its health asks, each `<module>.<tool>` (novox/hq ADR 0240, to-be
+	// 48 §3): the machine's node-engine asks them of its own node tools, and is granted that and no more.
+	Checks []string
 }
 
 // Records is what composing a user list needs to know about the mesh, and nothing more.
@@ -73,12 +76,14 @@ func Users(r Records) ([]Principal, error) {
 
 	for _, node := range sortedCopy(r.Nodes) {
 		witness := false
+		var checks []string
 		for _, d := range r.Assigned[node] {
 			if d.Module == controllerModule {
 				witness = true
 			}
+			checks = append(checks, d.Checks...)
 		}
-		out = append(out, Principal{Kind: KindNode, Node: node, WitnessesController: witness})
+		out = append(out, Principal{Kind: KindNode, Node: node, WitnessesController: witness, Checks: checks})
 		// **Where the runtime is assigned, the machine gets one runtime principal in place of the
 		// runtime module's own** (novox/hq ADR 0175, to-be 38). It carries every module on the
 		// node: its serving grants are the union of theirs. Every other module keeps its own

internal/broker/witness_grants_test.go

@@ -35,3 +35,31 @@ func TestTheWitnessIsGrantedWhatItReadsAndNoMore(t *testing.T) {
 		}
 	}
 }
+
+// A module's health that asks one of its own tools is asked by the machine's node-engine, of the instance
+// on its own machine and nowhere else (novox/hq ADR 0240, to-be 48 §3).
+func TestTheEngineIsGrantedTheToolsItsModulesHealthAsks(t *testing.T) {
+	users, err := Users(Records{Nodes: []string{"control", "edge"},
+		Assigned: map[string][]Declared{"edge": {{Module: "keycloak", Checks: []string{"keycloak.keycloak_admin_health"}}}}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, u := range users {
+		if u.Kind != KindNode {
+			continue
+		}
+		perms, err := PermissionsFor(u)
+		if err != nil {
+			t.Fatal(err)
+		}
+		mine := "mesh.mod.keycloak.tool.keycloak_admin_health.edge"
+		if got := slices.Contains(perms.Publish, mine); got != (u.Node == "edge") {
+			t.Errorf("%s may ask keycloak's health tool on edge: %v", u.Node, got)
+		}
+		for _, p := range perms.Publish {
+			if p == "mesh.mod.keycloak.tool.keycloak_admin_health" || p == "mesh.mod.keycloak.tool.>" {
+				t.Errorf("%s may ask the tool of any machine: %s", u.Node, p)
+			}
+		}
+	}
+}

internal/broker/writers.go

@@ -86,7 +86,9 @@ var WritersTable = []WriterRow{
 		Others: "read", Subjects: []string{"mesh.node.*.declare"}, Writes: isController},
 	{State: "a machine's applied state and its report", Writer: "the node-engine's apply queue",
 		KeptIn: "the machine; the report on the bus", Others: "the reconcile and a delivery enqueue, never apply",
-		Subjects: []string{"mesh.control.*.report"}, Writes: ownMachine},
+		// And its health statement between reports (novox/hq ADR 0240): the same writer stating the same
+		// machine, inside the grant it already had (`mesh.control.<its own>.>`).
+		Subjects: []string{"mesh.control.*.report", "mesh.control.*.health"}, 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},
 	{State: "plans and their tiers", Writer: "controller (lease holder), compare-and-set on the plan's revision",

internal/builder/check.go

@@ -298,6 +298,13 @@ func Check(ctx context.Context, run Runner, spec CheckSpec, workspace, registry
 	say("check", "the facts of %s (%s): %d machine(s), the bus at %s, the store at %s", f.Taken.Format(time.RFC3339),
 		short(strings.TrimPrefix(digest, "sha256:")), len(f.Machines), f.Versions.Bus, f.Versions.Store)
 
+	// **What an earlier delivery of this same ask left is removed first** (novox/hq issue 285): an ask is
+	// redelivered when the holder that took it stopped mid-check — the build agent itself updated by the
+	// rollout it is checking — and its containers, named by the ask's id, are still there. Raising the
+	// store again under that name was refused, and the check said it could not run.
+	if n, err := RemoveContainersOf(ctx, run, spec.ID); err == nil && n > 0 {
+		say("check", "removed %d throwaway container(s) an earlier delivery of this check left", n)
+	}
 	// The throwaway store and bus, of the versions the mesh runs, removed whatever happens.
 	defer func() {
 		removing, done := context.WithTimeout(context.Background(), time.Minute)
@@ -519,21 +526,24 @@ func gateLayer(ctx context.Context, spec CheckSpec, tree, root, gate, verdictFil
 			case "fail":
 				return "fail", said.Summary
 			case "error":
-				// The gate could not raise the mesh as it is (novox/hq issue 285): said in its own words.
+				// The gate could not raise the mesh as it is (novox/hq issue 282): said in its own words.
 				return "error", said.Summary
 			}
 		}
 		// The gate could not judge: not the change's fault, and never a pass.
 		return "error", "the merge gate could not judge the change: " + lastLine(out.String())
 	}
-	var said struct {
-		Verdict string `json:"verdict"`
-		Summary string `json:"summary"`
-	}
 	raw, err := os.ReadFile(verdictFile)
-	if err != nil || json.Unmarshal(raw, &said) != nil || said.Verdict == "" {
+	if err != nil {
 		return "error", "the merge gate said no verdict"
 	}
+	said, ok := readGateVerdict(raw)
+	if !ok {
+		return "error", "the merge gate said no verdict"
+	}
+	if verdict, summary, raised := gateRaisedTheMesh(said); !raised {
+		return verdict, summary
+	}
 
 	// 3. **The replays of what the mesh runs** (to-be 45 §9, M9): mesh-lab's, from its main — reviewed code,
 	// so given the container runtime the resolver replay raises containers with — against the bus of the
@@ -683,6 +693,58 @@ func newProblems(change, base string) []string {
 	return out
 }
 
+// gateVerdict is what of the gate's verdict the seat reads: the verdict, and every machine — whether it
+// composes on the mesh and whether it composed in the gate's store without the change.
+type gateVerdict struct {
+	Verdict  string `json:"verdict"`
+	Summary  string `json:"summary"`
+	Machines []struct {
+		Described string `json:"described"`
+		Live      bool   `json:"live-composes"`
+		Base      struct {
+			Composes bool `json:"composes"`
+		} `json:"base"`
+	} `json:"machines"`
+}
+
+// readGateVerdict reads the verdict the gate wrote, from its first line that opens a JSON document: a
+// judge from before the verdict was alone on its output printed what composition said ahead of it.
+func readGateVerdict(raw []byte) (gateVerdict, bool) {
+	var v gateVerdict
+	text := string(raw)
+	if i := strings.Index(text, "\n{"); i >= 0 && !strings.HasPrefix(strings.TrimSpace(text), "{") {
+		text = text[i+1:]
+	}
+	if json.Unmarshal([]byte(text), &v) != nil || v.Verdict == "" {
+		return gateVerdict{}, false
+	}
+	return v, true
+}
+
+// gateRaisedTheMesh is the seat's own reading of a verdict that passes (novox/hq issue 285): **a machine
+// the mesh composes that did not compose in the gate's store without the change makes the gate an error,
+// never a pass** — the change was judged against a machine that is not the mesh's, broken against broken.
+// The judge says so itself since issue 285, but the judge is the controller the mesh runs, and one from
+// before it passed such a verdict; read here too, the rule holds whatever judged. It answers false, with
+// the verdict to report, when the gate did not raise the mesh.
+func gateRaisedTheMesh(v gateVerdict) (string, string, bool) {
+	if v.Verdict != "pass" && v.Verdict != "warning" {
+		return "", "", true
+	}
+	var unraised []string
+	for _, m := range v.Machines {
+		if m.Live && !m.Base.Composes {
+			unraised = append(unraised, m.Described)
+		}
+	}
+	if len(unraised) == 0 {
+		return "", "", true
+	}
+	return "error", fmt.Sprintf("the mesh as it is could not be raised, so the change cannot be judged against it: "+
+		"%d of %d machines compose on the mesh and not in the gate (the first: %s) — novox/hq issue 285",
+		len(unraised), len(v.Machines), unraised[0]), false
+}
+
 // gitShow is a file as a ref has it.
 func gitShow(ctx context.Context, dir, ref, file string) ([]byte, error) {
 	cmd := exec.CommandContext(ctx, "git", "show", ref+":"+filepath.ToSlash(file))
@@ -835,7 +897,7 @@ func (t *tail) String() string {
 
 // whatFailed is the line of a failed script's output that says what failed, for the status a pull request
 // shows: the first failing test, the first failing package, the files not formatted — a bare "FAIL" or a
-// file's name said nothing a reader could act on (novox/hq issue 286) — and the last line otherwise.
+// file's name said nothing a reader could act on (novox/hq issue 283) — and the last line otherwise.
 func whatFailed(s string) string {
 	lines := strings.Split(strings.TrimSpace(s), "\n")
 	for i, line := range lines {

internal/builder/check_test.go

@@ -415,7 +415,7 @@ func TestABuildSaysWhetherItsCommitIsOnTheTrunk(t *testing.T) {
 	}
 }
 
-// **Issue 286**: a failed merge-check.sh was said by its last line — a bare "FAIL", or the name of a file
+// **Issue 283**: a failed merge-check.sh was said by its last line — a bare "FAIL", or the name of a file
 // gofmt listed — which named nothing a reader could act on. The status says what failed.
 func TestAFailedScriptIsSaidByWhatFailed(t *testing.T) {
 	for out, want := range map[string]string{
@@ -429,3 +429,54 @@ func TestAFailedScriptIsSaidByWhatFailed(t *testing.T) {
 		}
 	}
 }
+
+// **Issue 285**: a judge from before the rule passed "every machine composes with the change as it did
+// without (0 of 4 compose)". The seat reads the machines itself: a machine the mesh composes that did not
+// compose in the gate's store makes the gate an error, whatever judged it.
+func TestTheSeatCallsAGateThatRaisedNoMachineAnError(t *testing.T) {
+	old := "bus users without a credential: controller\n" + `{"verdict":"pass","summary":"every machine composes with the change as it did without (0 of 2 compose)",
+		"machines":[{"described":"the hub","live-composes":true,"base":{"composes":false}},
+		{"described":"a machine","live-composes":true,"base":{"composes":false}}]}`
+	v, ok := readGateVerdict([]byte(old))
+	if !ok {
+		t.Fatal("a verdict after a line of composition's was not read")
+	}
+	verdict, summary, raised := gateRaisedTheMesh(v)
+	if raised || verdict != "error" || !strings.Contains(summary, "2 of 2 machines") || !strings.Contains(summary, "the hub") {
+		t.Errorf("0 of 2 composing is %q %q", verdict, summary)
+	}
+	good := `{"verdict":"pass","summary":"(2 of 2 compose)","machines":[{"described":"the hub","live-composes":true,"base":{"composes":true}},
+		{"described":"a laptop","live-composes":false,"base":{"composes":false}}]}`
+	v, _ = readGateVerdict([]byte(good))
+	if _, _, raised := gateRaisedTheMesh(v); !raised {
+		t.Error("a machine that does not compose on the mesh either was read as the gate's failure")
+	}
+	failed := `{"verdict":"fail","summary":"x","machines":[{"described":"the hub","live-composes":true,"base":{"composes":false}}]}`
+	v, _ = readGateVerdict([]byte(failed))
+	if _, _, raised := gateRaisedTheMesh(v); !raised {
+		t.Error("a failing verdict is the change's, and stands")
+	}
+}
+
+// **Issue 285**: an ask redelivered after its holder stopped mid-check found its own throwaway store still
+// there under its name, and the check said it could not run. What an earlier delivery left goes first.
+func TestARedeliveredCheckRemovesWhatItsEarlierDeliveryLeft(t *testing.T) {
+	registry := checkEnvironment(t)
+	repo, head := aCheckedRepository(t, map[string]string{CheckScript: "echo checked\n"})
+	id := fmt.Sprintf("check-again-%d", time.Now().UnixNano())
+	if out, err := exec.Command("docker", "run", "-d", "--rm", "--label", BuildLabel+"="+id, "--name", id+"-store",
+		"postgres:17-alpine", "sleep", "300").CombinedOutput(); err != nil {
+		t.Skipf("no container for the earlier delivery: %v %s", err, out)
+	}
+	v, err := Check(t.Context(), Command, CheckSpec{ID: id, Repository: repo, Ref: head, Owner: "novox",
+		Repo: "hq", Number: 7, Toolchain: goToolchain}, t.TempDir(), registry, GitCredential{}, nil)
+	if err != nil {
+		t.Fatalf("a redelivered check could not run: %v", err)
+	}
+	if v.Repo == nil || v.Repo.Verdict != "pass" {
+		t.Errorf("the repository's check answered %+v", v.Repo)
+	}
+	if left := labelled(id); len(left) > 0 {
+		t.Errorf("the check left %d container(s) behind", len(left))
+	}
+}

internal/catalogue/declaration.go

@@ -235,6 +235,12 @@ type Rendering struct {
 	// mounts and environment. A node setting fixed at installation; empty means the default,
 	// /var/lib — see dir_into.go.
 	DataRoot string
+
+	// ReadsHealth says this machine's node-engine reads a resource's `health` (novox/hq ADR 0240 Phase B:
+	// its statement's contract is link.ReadinessContract or later). An older engine parses strictly and
+	// refuses a field it does not know, whole — so to it the field is not sent, and what it runs is
+	// judged by liveness alone.
+	ReadsHealth bool
 }
 
 // machinePort is where a module's port lives on this machine, or the port itself when the mesh has
@@ -957,6 +963,9 @@ func (r Resolution) compose(with Rendering, owner map[string]string,
 				return nil, err
 			}
 			publishedOn(copied, m.Module, with)
+			// How it is ready, in the node-engine's words: its endpoint as the port this machine
+			// published it on — or not sent at all to an engine older than the field (ADR 0240).
+			healthInto(copied, m, with)
 			// The account's environment and every module's shell code, where this module holds the
 			// seat that places them (novox/hq ADR 0203, ADR 0204). Gathered from every module on
 			// the node, as the jails are, and **last of every placeholder pass**: shell code is a

internal/catalogue/health.go

@@ -0,0 +1,460 @@
+package catalogue
+
+import (
+	"fmt"
+	"sort"
+	"strconv"
+	"strings"
+	"time"
+)
+
+// A module says how each long-running resource is ready (novox/hq ADR 0240 rule 2, to-be 48 §2 and §8,
+// Phase B).
+//
+// **On the resource, beside its other fields**: one kind and its timing. The node-engine judges every
+// long-running resource alive with no declaration (Phase A); this is how a module says what *ready* means
+// for one — the image's own check adopted by name, an HTTP request to a declared endpoint, a TCP connect,
+// a command in the container, the unit's own readiness, or one of the module's own tools.
+//
+// **An endpoint is named, never a port or an address**: the check follows the machine's port for that
+// endpoint as the endpoint does, so a port this machine gave elsewhere moves the check with it. The
+// controller composes the name into the port the machine published it on, and sends the field only to a
+// node-engine that reads it (link.ReadinessContract): an older, strict engine refuses a field it does not
+// know, whole.
+//
+// **The bounds are the record's**: an interval not under ten seconds, a timeout under the interval, at
+// least two failing looks in a row, and a grace plus the failing looks within five minutes — so a
+// resource broken from its start is said within the gate's ten. Refused here, near the author, and again
+// by the node-engine, far away, in the same words.
+
+// HealthField is the resource field.
+const HealthField = "health"
+
+// The kinds of check (to-be 48 §2).
+const (
+	HealthRuntime = "runtime"
+	HealthHTTP    = "http"
+	HealthTCP     = "tcp"
+	HealthExec    = "exec"
+	HealthUnit    = "unit"
+	HealthTool    = "tool"
+)
+
+// The bounds and the defaults (ADR 0240 rule 2).
+const (
+	HealthIntervalDefault = 30 * time.Second
+	HealthIntervalFloor   = 10 * time.Second
+	HealthTimeoutDefault  = 5 * time.Second
+	HealthLooksDefault    = 3
+	HealthLooksFloor      = 2
+	HealthGraceDefault    = 60 * time.Second
+	// HealthWithin is the most a grace and the failing looks may take together: a resource broken
+	// from its start is said within the gate's ten minutes with room for its judgings.
+	HealthWithin = 5 * time.Minute
+)
+
+// HealthRequiredFrom is when `module check` refuses a long-running resource without `health` (ADR 0240
+// rule 8): six weeks after liveness was first judged live (2026-10-07), unless the catalogue's count of
+// undeclared resources reached zero first — which its own counter enforces by never letting it rise.
+var HealthRequiredFrom = time.Date(2026, 11, 18, 0, 0, 0, 0, time.UTC)
+
+// Health is one resource's declaration, read.
+type Health struct {
+	Kind string
+	// Endpoint is the `listens` name an http or tcp check looks at.
+	Endpoint string
+	// Path, Status, Body and Scheme are an http check's: the path asked, the status expected (zero: any
+	// status under 400), a text the answer must hold, and http or https.
+	Path   string
+	Status int
+	Body   string
+	Scheme string
+	// Command is an exec check's command, run by the container's shell.
+	Command string
+	// Tool is a tool check's tool, one of the module's own.
+	Tool string
+	// The timing, with the defaults applied.
+	Interval, Timeout, Grace time.Duration
+	Looks                    int
+	// Needs is the provision the check exercises (to-be 48 §6): while its provider for this consumer is
+	// unhealthy, what this check finds is held under the provider's condition.
+	Needs string
+}
+
+// healthKeys are the keys a `health` field may carry; anything else is refused by name.
+var healthKeys = map[string]bool{"kind": true, "endpoint": true, "path": true, "status": true, "body": true,
+	"scheme": true, "command": true, "tool": true, "interval": true, "timeout": true, "looks": true,
+	"grace": true, "needs": true}
+
+// healthAddressKeys are what a check may not be aimed by: a port or an address does not follow the
+// machine's port for the endpoint, and a manifest is the same on every machine.
+var healthAddressKeys = map[string]bool{"port": true, "address": true, "host": true, "url": true, "ip": true}
+
+// LongRunning says whether a manifest resource stays up: a container that is no step and on no schedule,
+// a service stated running, a process that is no step and on no schedule (ADR 0240 rule 1).
+func LongRunning(r map[string]any) bool {
+	switch fmt.Sprint(r["type"]) {
+	case "container", "process":
+		if once, _ := r["run-once"].(bool); once {
+			return false
+		}
+		return r["schedule"] == nil
+	case "service":
+		return fmt.Sprint(r["state"]) == "running"
+	}
+	return false
+}
+
+// ReadHealth reads a resource's `health` field, defaults applied. False when it carries none.
+func ReadHealth(r map[string]any) (Health, bool, []string) {
+	raw, present := r[HealthField]
+	if !present {
+		return Health{}, false, nil
+	}
+	id := fmt.Sprint(r["id"])
+	fields, ok := raw.(map[string]any)
+	if !ok {
+		return Health{}, true, []string{fmt.Sprintf("%s: health is %T; it is an object with a kind and its timing", id, raw)}
+	}
+	var problems []string
+	keys := make([]string, 0, len(fields))
+	for k := range fields {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+	for _, k := range keys {
+		switch {
+		case healthAddressKeys[k]:
+			problems = append(problems, fmt.Sprintf("%s: its health names a %s; a check names an endpoint the module "+
+				"declares under listens, by its name, so it follows the port this machine gives it (ADR 0240 rule 2)", id, k))
+		case !healthKeys[k]:
+			problems = append(problems, fmt.Sprintf("%s: its health says %q, which a health check does not have", id, k))
+		}
+	}
+	text := func(key string) string {
+		v, present := fields[key]
+		if !present {
+			return ""
+		}
+		s, ok := v.(string)
+		if !ok {
+			problems = append(problems, fmt.Sprintf("%s: its health's %s is %T; it is text", id, key, v))
+		}
+		return strings.TrimSpace(s)
+	}
+	duration := func(key string, fallback time.Duration) time.Duration {
+		s := text(key)
+		if s == "" {
+			return fallback
+		}
+		d, err := time.ParseDuration(s)
+		if err != nil || d < 0 {
+			problems = append(problems, fmt.Sprintf("%s: its health's %s is %q; it is a duration such as \"30s\"", id, key, s))
+			return fallback
+		}
+		return d
+	}
+	number := func(key string, fallback int) int {
+		v, present := fields[key]
+		if !present {
+			return fallback
+		}
+		f, ok := v.(float64)
+		if !ok || f != float64(int(f)) {
+			problems = append(problems, fmt.Sprintf("%s: its health's %s is %v; it is a whole number", id, key, v))
+			return fallback
+		}
+		return int(f)
+	}
+	h := Health{Kind: text("kind"), Endpoint: text("endpoint"), Path: text("path"), Body: text("body"),
+		Scheme: text("scheme"), Command: text("command"), Tool: text("tool"), Needs: text("needs"),
+		Interval: duration("interval", HealthIntervalDefault), Timeout: duration("timeout", HealthTimeoutDefault),
+		Grace: duration("grace", HealthGraceDefault), Looks: number("looks", HealthLooksDefault),
+		Status: number("status", 0)}
+	return h, true, problems
+}
+
+// healthProblems is everything wrong with a manifest's `health` fields, in the manifest's words (to-be
+// 48 §8): a field on something that does not stay up, a kind its resource cannot have, an endpoint the
+// module does not declare, a port or an address, a timing outside its bounds, a tool the module does not
+// serve, and a tool check with no check of another kind beside it on the module.
+func healthProblems(m Manifest) []string {
+	var problems []string
+	endpoints := map[string]Listening{}
+	for _, l := range m.Listens {
+		if n := strings.TrimSpace(l.Name); n != "" {
+			endpoints[n] = l
+		}
+	}
+	tools := map[string]bool{}
+	for _, t := range m.Tools {
+		tools[t] = true
+	}
+	wants := map[string]bool{}
+	for _, w := range m.Wants() {
+		wants[w] = true
+	}
+	var toolChecks []string
+	otherKinds := 0
+	for _, r := range m.Resources {
+		h, has, read := ReadHealth(r)
+		if !has {
+			continue
+		}
+		id, typ := fmt.Sprint(r["id"]), fmt.Sprint(r["type"])
+		where := m.Module + ": " + id
+		for _, p := range read {
+			problems = append(problems, m.Module+": "+p)
+		}
+		if !LongRunning(r) {
+			problems = append(problems, fmt.Sprintf("%s declares health and does not stay up: a step, anything on a "+
+				"schedule and a service not stated running are judged by their step and their schedule (ADR 0240 rule 1)", where))
+			continue
+		}
+		switch h.Kind {
+		case HealthRuntime, HealthExec:
+			if typ != "container" {
+				problems = append(problems, fmt.Sprintf("%s is a %s and its health is %q, which only a container has: "+
+					"the runtime runs it inside the container", where, typ, h.Kind))
+			}
+		case HealthUnit:
+			if typ == "container" {
+				problems = append(problems, fmt.Sprintf("%s is a container and its health is %q, which is a service's "+
+					"or a process's own readiness", where, h.Kind))
+			}
+		case HealthHTTP, HealthTCP, HealthTool:
+		case "":
+			problems = append(problems, fmt.Sprintf("%s declares health with no kind: %s", where, healthKindsWords()))
+		default:
+			problems = append(problems, fmt.Sprintf("%s declares health of kind %q: %s", where, h.Kind, healthKindsWords()))
+		}
+		switch h.Kind {
+		case HealthHTTP, HealthTCP:
+			l, declared := endpoints[h.Endpoint]
+			switch {
+			case h.Endpoint == "":
+				problems = append(problems, fmt.Sprintf("%s's %s check names no endpoint: it names one the module "+
+					"declares under listens, by its name", where, h.Kind))
+			case !declared:
+				problems = append(problems, fmt.Sprintf("%s's %s check names the endpoint %q, which %s does not declare "+
+					"under listens (%s)", where, h.Kind, h.Endpoint, m.Module, namedEndpointsWords(endpoints)))
+			case l.At() != "tcp":
+				problems = append(problems, fmt.Sprintf("%s's %s check names %q, which is %s: a check connects over tcp",
+					where, h.Kind, h.Endpoint, l.At()))
+			}
+		default:
+			if h.Endpoint != "" {
+				problems = append(problems, fmt.Sprintf("%s's %s check names an endpoint, which only an http or tcp check "+
+					"looks at", where, h.Kind))
+			}
+		}
+		if h.Kind != HealthHTTP && (h.Path != "" || h.Status != 0 || h.Body != "" || h.Scheme != "") {
+			problems = append(problems, fmt.Sprintf("%s's %s check says a path, a status, a body or a scheme, which "+
+				"only an http check has", where, h.Kind))
+		}
+		if h.Kind == HealthHTTP {
+			if h.Path != "" && !strings.HasPrefix(h.Path, "/") {
+				problems = append(problems, fmt.Sprintf("%s's http check asks %q; a path starts with /", where, h.Path))
+			}
+			if h.Status != 0 && (h.Status < 100 || h.Status > 599) {
+				problems = append(problems, fmt.Sprintf("%s's http check expects status %d, which is not one", where, h.Status))
+			}
+			if h.Scheme != "" && h.Scheme != "http" && h.Scheme != "https" {
+				problems = append(problems, fmt.Sprintf("%s's http check is over %q; it is http or https", where, h.Scheme))
+			}
+		}
+		if (h.Command != "") != (h.Kind == HealthExec) {
+			if h.Kind == HealthExec {
+				problems = append(problems, fmt.Sprintf("%s's exec check says no command", where))
+			} else {
+				problems = append(problems, fmt.Sprintf("%s's %s check says a command, which only an exec check runs",
+					where, h.Kind))
+			}
+		}
+		if h.Kind == HealthTool {
+			switch {
+			case h.Tool == "":
+				problems = append(problems, fmt.Sprintf("%s's tool check names no tool", where))
+			case !tools[h.Tool]:
+				problems = append(problems, fmt.Sprintf("%s's tool check asks %q, which %s does not serve (its tools: %s)",
+					where, h.Tool, m.Module, orNoneWords(m.Tools)))
+			}
+			toolChecks = append(toolChecks, id)
+		} else {
+			if h.Tool != "" {
+				problems = append(problems, fmt.Sprintf("%s's %s check names a tool, which only a tool check asks", where, h.Kind))
+			}
+			if h.Kind != "" {
+				otherKinds++
+			}
+		}
+		if h.Needs != "" && !wants[h.Needs] {
+			problems = append(problems, fmt.Sprintf("%s's check needs %q, which %s does not require: a check names "+
+				"the provision it exercises, among those the module requires", where, h.Needs, m.Module))
+		}
+		problems = append(problems, healthTimingProblems(where, h)...)
+	}
+	if len(toolChecks) > 0 && otherKinds == 0 {
+		problems = append(problems, fmt.Sprintf("%s judges itself only by its own tool (%s): a tool check is for "+
+			"function no endpoint shows, and only beside a check of another kind the module does not run itself "+
+			"(ADR 0227 rule 8, ADR 0240 rule 2)", m.Module, strings.Join(toolChecks, ", ")))
+	}
+	return problems
+}
+
+// healthTimingProblems holds a check's timing to its bounds.
+func healthTimingProblems(where string, h Health) []string {
+	var problems []string
+	if h.Interval < HealthIntervalFloor {
+		problems = append(problems, fmt.Sprintf("%s looks every %s; a check looks no more often than every %s — "+
+			"the mesh is a guest on the machine (ADR 0240 rule 2)", where, h.Interval, HealthIntervalFloor))
+	}
+	if h.Timeout <= 0 || h.Timeout >= h.Interval {
+		problems = append(problems, fmt.Sprintf("%s gives a look %s, which must be more than nothing and under its "+
+			"interval of %s", where, h.Timeout, h.Interval))
+	}
+	if h.Looks < HealthLooksFloor {
+		problems = append(problems, fmt.Sprintf("%s is unhealthy after %d failing look(s); it is at least %d — one "+
+			"look can be wrong (issue 277)", where, h.Looks, HealthLooksFloor))
+	}
+	if h.Grace < 0 {
+		problems = append(problems, fmt.Sprintf("%s has a grace of %s", where, h.Grace))
+	}
+	if h.Looks >= HealthLooksFloor && h.Interval >= HealthIntervalFloor {
+		if took := h.Grace + time.Duration(h.Looks)*h.Interval; took > HealthWithin {
+			problems = append(problems, fmt.Sprintf("%s is said unhealthy at the earliest %s after it starts (a grace "+
+				"of %s and %d looks every %s); it is at most %s, so a resource broken from its start is said within "+
+				"the gate's bound", where, took, h.Grace, h.Looks, h.Interval, HealthWithin))
+		}
+	}
+	return problems
+}
+
+func healthKindsWords() string {
+	return "a check is runtime (the image's own, adopted by name), http, tcp, exec, unit or tool"
+}
+
+func namedEndpointsWords(endpoints map[string]Listening) string {
+	if len(endpoints) == 0 {
+		return "it names none"
+	}
+	names := make([]string, 0, len(endpoints))
+	for n := range endpoints {
+		names = append(names, n)
+	}
+	sort.Strings(names)
+	return "it names " + strings.Join(names, ", ")
+}
+
+func orNoneWords(names []string) string {
+	if len(names) == 0 {
+		return "none"
+	}
+	return strings.Join(names, ", ")
+}
+
+// Undeclared is every long-running resource of the manifest without `health`, by id (ADR 0240 rule 8):
+// what the catalogue's count counts.
+func Undeclared(m Manifest) []string {
+	var out []string
+	for _, r := range m.Resources {
+		if !LongRunning(r) {
+			continue
+		}
+		if _, has := r[HealthField]; !has {
+			out = append(out, fmt.Sprint(r["id"]))
+		}
+	}
+	return out
+}
+
+// HealthChecks is every tool a module's health asks, as `<module>.<tool>`: what a machine's node-engine
+// is granted to ask of its own node tools (to-be 48 §3).
+func HealthChecks(m Manifest) []string {
+	var out []string
+	for _, r := range m.Resources {
+		h, has, _ := ReadHealth(r)
+		if has && h.Kind == HealthTool && h.Tool != "" && LongRunning(r) {
+			out = append(out, m.Module+"."+h.Tool)
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// healthInto composes a resource's `health` into the node-engine's words, or takes it away (to-be 48 §2,
+// §3): for an engine that reads it, the endpoint becomes the port this machine published it on and the
+// defaults are written out; for one that does not — older and strict — the field is not sent, and the
+// resource is judged by liveness alone, as before.
+func healthInto(resource map[string]any, m Manifest, with Rendering) {
+	if _, has := resource[HealthField]; !has {
+		return
+	}
+	if !with.ReadsHealth {
+		delete(resource, HealthField)
+		return
+	}
+	h, _, _ := ReadHealth(resource)
+	out := map[string]any{"kind": h.Kind, "interval": h.Interval.String(), "timeout": h.Timeout.String(),
+		"looks": h.Looks, "grace": h.Grace.String()}
+	if h.Endpoint != "" {
+		out["endpoint"] = h.Endpoint
+		if port, ok := EndpointPort(m, h.Endpoint); ok {
+			out["port"] = with.machinePort(m.Module, port)
+		}
+	}
+	if h.Kind == HealthHTTP {
+		path := h.Path
+		if path == "" {
+			path = "/"
+		}
+		out["path"] = path
+		if h.Status != 0 {
+			out["status"] = h.Status
+		}
+		if h.Body != "" {
+			out["body"] = h.Body
+		}
+		if h.Scheme != "" {
+			out["scheme"] = h.Scheme
+		}
+	}
+	if h.Command != "" {
+		out["command"] = h.Command
+	}
+	if h.Tool != "" {
+		out["tool"] = h.Tool
+	}
+	if h.Needs != "" {
+		out["needs"] = h.Needs
+	}
+	resource[HealthField] = out
+}
+
+// HealthWords is a declared check in a line, for `module check` and `node show`.
+func HealthWords(h Health) string {
+	var what string
+	switch h.Kind {
+	case HealthHTTP:
+		what = "http " + orSlash(h.Path) + " on " + h.Endpoint
+		if h.Status != 0 {
+			what += " expecting " + strconv.Itoa(h.Status)
+		}
+	case HealthTCP:
+		what = "tcp on " + h.Endpoint
+	case HealthTool:
+		what = "its tool " + h.Tool
+	case HealthRuntime:
+		what = "its image's own check"
+	default:
+		what = h.Kind
+	}
+	return fmt.Sprintf("%s every %s", what, h.Interval)
+}
+
+func orSlash(p string) string {
+	if p == "" {
+		return "/"
+	}
+	return p
+}

internal/catalogue/health_test.go

@@ -0,0 +1,217 @@
+package catalogue
+
+import (
+	"encoding/json"
+	"strings"
+	"testing"
+
+	"github.com/novox/mesh-host/validate"
+)
+
+// A module says how each long-running resource is ready (novox/hq ADR 0240 rule 2, to-be 48 §8): `module
+// check` refuses each part out of its bounds, each endpoint named by a port or an address, a tool the
+// module does not serve, and a tool check alone — a test per refusal.
+
+const healthDigest = "@sha256:aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa"
+
+// healthManifest is a module of one web container, one running service and one step, with the health
+// given on the container (or on the resource named by on).
+func healthManifest(t *testing.T, on string, health map[string]any, more ...map[string]any) []byte {
+	t.Helper()
+	resources := []map[string]any{
+		{"id": "server", "type": "container", "name": "app-server", "image": "registry.example/app" + healthDigest,
+			"ports": []any{"8080"}},
+		{"id": "daemon", "type": "service", "unit": "app.service", "state": "running"},
+		{"id": "seed", "type": "container", "name": "app-seed", "image": "registry.example/app" + healthDigest,
+			"run-once": true},
+		{"id": "sweep", "type": "container", "name": "app-sweep", "image": "registry.example/app" + healthDigest,
+			"schedule": "0 3 * * *"},
+	}
+	resources = append(resources, more...)
+	for _, r := range resources {
+		if r["id"] == on && health != nil {
+			r["health"] = health
+		}
+	}
+	m := map[string]any{"module": "app", "requires": []any{"postgres-database"}, "tools": []any{"app_status"},
+		"listens": []any{
+			map[string]any{"name": "web", "port": 8080, "from": "mesh"},
+			map[string]any{"name": "beacon", "port": 9999, "protocol": "udp", "from": "mesh"},
+		},
+		"resources": resources}
+	raw, err := json.Marshal(m)
+	if err != nil {
+		t.Fatal(err)
+	}
+	return raw
+}
+
+func TestAWellFormedHealthIsAccepted(t *testing.T) {
+	for _, h := range []map[string]any{
+		{"kind": "http", "endpoint": "web", "path": "/healthz", "status": 200, "body": "ok", "needs": "postgres-database"},
+		{"kind": "tcp", "endpoint": "web", "interval": "10s", "timeout": "2s", "looks": 2, "grace": "0s"},
+		{"kind": "runtime"},
+		{"kind": "exec", "command": "pg_isready -q", "grace": "4m", "looks": 2, "interval": "30s"},
+	} {
+		if _, err := ParseManifest(healthManifest(t, "server", h)); err != nil {
+			t.Errorf("%v was refused: %v", h, err)
+		}
+	}
+	if _, err := ParseManifest(healthManifest(t, "daemon", map[string]any{"kind": "unit"})); err != nil {
+		t.Errorf("a service's own readiness was refused: %v", err)
+	}
+	// A tool beside a check of another kind on the module.
+	raw := healthManifest(t, "server", map[string]any{"kind": "http", "endpoint": "web"},
+		map[string]any{"id": "admin", "type": "service", "unit": "app-admin.service", "state": "running",
+			"health": map[string]any{"kind": "tool", "tool": "app_status"}})
+	if _, err := ParseManifest(raw); err != nil {
+		t.Errorf("a tool check beside an http check was refused: %v", err)
+	}
+}
+
+func TestEveryOutOfBoundsHealthIsRefusedByName(t *testing.T) {
+	cases := []struct {
+		name   string
+		on     string
+		health map[string]any
+		says   string
+	}{
+		{"no kind", "server", map[string]any{"endpoint": "web"}, "with no kind"},
+		{"an unknown kind", "server", map[string]any{"kind": "ping"}, `of kind "ping"`},
+		{"a port", "server", map[string]any{"kind": "tcp", "port": 8080}, "names a port"},
+		{"an address", "server", map[string]any{"kind": "http", "endpoint": "web", "address": "127.0.0.1"}, "names a address"},
+		{"a url", "server", map[string]any{"kind": "http", "url": "http://localhost:8080/"}, "names a url"},
+		{"an unknown key", "server", map[string]any{"kind": "tcp", "endpoint": "web", "retries": 3}, `"retries"`},
+		{"no endpoint", "server", map[string]any{"kind": "http"}, "names no endpoint"},
+		{"an undeclared endpoint", "server", map[string]any{"kind": "tcp", "endpoint": "admin"}, "does not declare"},
+		{"a udp endpoint", "server", map[string]any{"kind": "tcp", "endpoint": "beacon"}, "a check connects over tcp"},
+		{"an interval under the floor", "server", map[string]any{"kind": "tcp", "endpoint": "web", "interval": "5s", "timeout": "1s"}, "no more often than every 10s"},
+		{"a timeout of the interval", "server", map[string]any{"kind": "tcp", "endpoint": "web", "interval": "10s", "timeout": "10s"}, "under its interval"},
+		{"one failing look", "server", map[string]any{"kind": "tcp", "endpoint": "web", "looks": 1}, "at least 2"},
+		{"a grace and looks past five minutes", "server", map[string]any{"kind": "tcp", "endpoint": "web", "grace": "4m", "looks": 3, "interval": "30s"}, "at most 5m0s"},
+		{"a duration that is not one", "server", map[string]any{"kind": "tcp", "endpoint": "web", "interval": "often"}, "is a duration"},
+		{"looks that are not a number", "server", map[string]any{"kind": "tcp", "endpoint": "web", "looks": "three"}, "whole number"},
+		{"a path without a slash", "server", map[string]any{"kind": "http", "endpoint": "web", "path": "health"}, "starts with /"},
+		{"a status that is not one", "server", map[string]any{"kind": "http", "endpoint": "web", "status": 700}, "is not one"},
+		{"a scheme that is not one", "server", map[string]any{"kind": "http", "endpoint": "web", "scheme": "ftp"}, "http or https"},
+		{"a status on a tcp check", "server", map[string]any{"kind": "tcp", "endpoint": "web", "status": 200}, "only an http check has"},
+		{"an exec with no command", "server", map[string]any{"kind": "exec"}, "says no command"},
+		{"a command on an http check", "server", map[string]any{"kind": "http", "endpoint": "web", "command": "true"}, "only an exec check runs"},
+		{"a runtime check on a service", "daemon", map[string]any{"kind": "runtime"}, "only a container has"},
+		{"a unit check on a container", "server", map[string]any{"kind": "unit"}, "a service's or a process's"},
+		{"a tool the module does not serve", "server", map[string]any{"kind": "tool", "tool": "app_admin"}, "does not serve"},
+		{"a tool check alone", "server", map[string]any{"kind": "tool", "tool": "app_status"}, "judges itself only by its own tool"},
+		{"needs not required", "server", map[string]any{"kind": "tcp", "endpoint": "web", "needs": "redis"}, "does not require"},
+		{"health on a step", "seed", map[string]any{"kind": "runtime"}, "does not stay up"},
+		{"health on a schedule", "sweep", map[string]any{"kind": "runtime"}, "does not stay up"},
+		{"health that is not an object", "server", nil, "is an object"},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			raw := healthManifest(t, c.on, c.health)
+			if c.health == nil {
+				raw = []byte(strings.Replace(string(raw), `"name":"app-server"`, `"name":"app-server","health":"tcp"`, 1))
+			}
+			_, err := ParseManifest(raw)
+			if err == nil {
+				t.Fatalf("%s was accepted", c.name)
+			}
+			if !strings.Contains(err.Error(), c.says) {
+				t.Errorf("refused, but not for saying %q: %v", c.says, err)
+			}
+		})
+	}
+}
+
+func TestHealthIsComposedAsThePortThisMachineGaveTheEndpoint(t *testing.T) {
+	m, err := ParseManifest(healthManifest(t, "server", map[string]any{"kind": "http", "endpoint": "web", "path": "/healthz",
+		"needs": "postgres-database"}))
+	if err != nil {
+		t.Fatal(err)
+	}
+	compose := func(with Rendering) map[string]any {
+		out, err := Resolution{Node: "laptop", Modules: []Manifest{m}}.Declaration(with)
+		if err != nil {
+			t.Fatal(err)
+		}
+		at := indexOfID(out, "app.server")
+		if at < 0 {
+			t.Fatal("the container was lost")
+		}
+		return out[at]
+	}
+	// An engine older than the field is not sent it: it would refuse the whole declaration.
+	if h, sent := compose(Rendering{Ports: map[string]map[int]int{"app": {8080: 31001}}})["health"]; sent {
+		t.Fatalf("health was sent to an engine that does not read it: %v", h)
+	}
+	got := compose(Rendering{ReadsHealth: true, Ports: map[string]map[int]int{"app": {8080: 31001}}})["health"].(map[string]any)
+	if got["port"] != 31001 || got["endpoint"] != "web" || got["path"] != "/healthz" || got["needs"] != "postgres-database" {
+		t.Errorf("composed as %v", got)
+	}
+	if got["interval"] != "30s" || got["timeout"] != "5s" || got["looks"] != 3 || got["grace"] != "1m0s" {
+		t.Errorf("the defaults were not written out: %v", got)
+	}
+	// The port this machine gives the endpoint moves, and the check moves with it.
+	moved := compose(Rendering{ReadsHealth: true, Ports: map[string]map[int]int{"app": {8080: 31002}}})["health"].(map[string]any)
+	if moved["port"] != 31002 {
+		t.Errorf("after the port moved the check still dials %v", moved["port"])
+	}
+	// And the catalogue's manifest is untouched by composing it.
+	if _, ok := m.Resources[0]["health"].(map[string]any)["port"]; ok {
+		t.Error("composing wrote the port into the catalogue's own manifest")
+	}
+}
+
+func TestTheUndeclaredAreTheLongRunningWithoutHealth(t *testing.T) {
+	m, err := ParseManifest(healthManifest(t, "server", map[string]any{"kind": "runtime"}))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := strings.Join(Undeclared(m), ","); got != "daemon" {
+		t.Errorf("undeclared: %q; the step and the schedule are not long-running, the server declares", got)
+	}
+}
+
+func TestATooledHealthIsGrantedToTheEngine(t *testing.T) {
+	raw := healthManifest(t, "server", map[string]any{"kind": "http", "endpoint": "web"},
+		map[string]any{"id": "admin", "type": "service", "unit": "app-admin.service", "state": "running",
+			"health": map[string]any{"kind": "tool", "tool": "app_status"}})
+	m, err := ParseManifest(raw)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if got := HealthChecks(m); len(got) != 1 || got[0] != "app.app_status" {
+		t.Errorf("checks %v", got)
+	}
+}
+
+// What the controller composes the node-engine takes: every kind, composed for an engine that reads it,
+// passes the engine's own validator (mesh-host/validate) — one set of words on both sides.
+func TestEveryComposedHealthIsOneTheNodeEngineTakes(t *testing.T) {
+	for _, h := range []map[string]any{
+		{"kind": "http", "endpoint": "web", "path": "/healthz", "status": 200, "body": "ok", "needs": "postgres-database"},
+		{"kind": "tcp", "endpoint": "web", "interval": "10s", "timeout": "2s", "looks": 2, "grace": "0s"},
+		{"kind": "runtime"},
+		{"kind": "exec", "command": "pg_isready -q"},
+	} {
+		m, err := ParseManifest(healthManifest(t, "server", h,
+			map[string]any{"id": "admin", "type": "service", "unit": "app-admin.service", "state": "running",
+				"health": map[string]any{"kind": "tool", "tool": "app_status"}}))
+		if err != nil {
+			t.Fatal(err)
+		}
+		m.Resources[1]["health"] = map[string]any{"kind": "unit"}
+		out, err := Resolution{Node: "laptop", Modules: []Manifest{m}}.Declaration(Rendering{ReadsHealth: true,
+			Ports: map[string]map[int]int{"app": {8080: 31001}}})
+		if err != nil {
+			t.Fatal(err)
+		}
+		body, err := json.Marshal(map[string]any{"declaration": validate.Version, "resources": out})
+		if err != nil {
+			t.Fatal(err)
+		}
+		if problems := validate.Declaration(body); len(problems) > 0 {
+			t.Errorf("%v composed into something the node-engine refuses: %v", h, problems)
+		}
+	}
+}

internal/catalogue/manifest.go

@@ -2036,6 +2036,9 @@ func ParseManifest(raw []byte) (Manifest, error) {
 	}
 
 	problems = append(problems, zoneProblems(m)...)
+	// How each long-running resource is ready (novox/hq ADR 0240 rule 2): said near its author, in the
+	// words the node-engine would refuse it in far away.
+	problems = append(problems, healthProblems(m)...)
 	if len(problems) > 0 {
 		sort.Strings(problems)
 		return Manifest{}, fmt.Errorf("this manifest cannot be used:\n  - %s",

internal/catalogue/verbs.go

@@ -262,6 +262,14 @@ var ControllerVerbs = []Verb{
 			"cause":     "the cause in a word, or a condition's kind — the word a second act for the same reason uses",
 			"condition": "the key of the condition it addressed, if any (optional)",
 		}, []string{"what", "why", "cause"})},
+	{Name: "drill", Description: "Record a drill — something broken on purpose to see the mesh raise and clear it: " +
+		"a module stopped, a process killed — with what it tests, in the hand-act log. A drill is a person's " +
+		"deliberate test, never a repair: no healer is wanted for it however often it is run (S15).",
+		Input: schema(map[string]string{
+			"what":      "what was done on purpose, in a line",
+			"why":       "what the drill tests",
+			"condition": "the key of the condition the drill is meant to raise, if any (optional)",
+		}, []string{"what", "why"})},
 	{Name: "hand-acts", Description: "What was done by hand lately — pushes, plans ended, consumers re-made, acts " +
 		"recorded — who, why and the cause of each, and which causes repeat: each repeat is a healer the mesh lacks.",
 		Input: schema(map[string]string{"days": "how many days back (default 14)"}, nil)},

internal/conditions/condition.go

@@ -46,11 +46,14 @@ const (
 	ScopeMesh     = "mesh"
 	// ScopeDelivery is a delivery mesh-delivery owns, by its id (novox/hq ADR 0239).
 	ScopeDelivery = "delivery"
+	// ScopeModule is a module on a machine, by `<module>.<machine>`: what it runs is not healthy there
+	// (novox/hq ADR 0240).
+	ScopeModule = "module"
 )
 
 // Scopes is every scope, in the order a person reads them.
 var Scopes = []string{ScopeMachine, ScopePlan, ScopeCall, ScopeBuild, ScopeMerge, ScopeProvider,
-	ScopeSeat, ScopeBus, ScopeCore, ScopeProbe, ScopeMesh, ScopeDelivery}
+	ScopeSeat, ScopeBus, ScopeCore, ScopeProbe, ScopeMesh, ScopeDelivery, ScopeModule}
 
 // Who resolves a condition.
 const (

internal/facts/facts.go

@@ -64,7 +64,7 @@ type Facts struct {
 	Edges []Edge `json:"edges,omitempty"`
 	// Beside is the ref each repository is cloned at beside a merge check, by the directory it is found
 	// under — the commits the mesh runs, the catalogue's main — so a check run by hand reads the siblings
-	// the build seat reads (novox/hq issue 286).
+	// the build seat reads (novox/hq issue 283).
 	Beside map[string]string `json:"beside,omitempty"`
 }
 
@@ -85,7 +85,7 @@ type Versions struct {
 	// Toolchains are the toolchain images a merge check runs in, by language, as the mesh holds them —
 	// with no address: the artifact store the snapshot is read from is where they are pulled from. What
 	// a repository's merge-check.sh runs in on the build seat, and so what it must run in anywhere else
-	// (novox/hq issue 286): two releases of one compiler disagree, down to how gofmt lays out a file.
+	// (novox/hq issue 283): two releases of one compiler disagree, down to how gofmt lays out a file.
 	Toolchains map[string]string `json:"toolchains,omitempty"`
 }
 

internal/facts/facts_test.go

@@ -172,3 +172,18 @@ func TestAWithheldPathStaysAPath(t *testing.T) {
 		t.Errorf("a key became %q", got)
 	}
 }
+
+// **Issue 288**: a repository was kept as the URL it was cloned from, the forge's address with it.
+func TestARepositoryIsNamedWithoutTheForge(t *testing.T) {
+	for in, want := range map[string]string{
+		"http://forge.internal:3000/owner/repo.git": "owner/repo",
+		"ssh://git@forge.internal:222/owner/repo":   "owner/repo",
+		"git@forge.internal:owner/repo.git":         "owner/repo",
+		"owner/repo":                                "owner/repo",
+		"":                                          "",
+	} {
+		if got := RepositoryName(in); got != want {
+			t.Errorf("%q is named %q, not %q", in, got, want)
+		}
+	}
+}

internal/facts/scrub.go

@@ -321,3 +321,31 @@ func standIn(run string) string {
 	}
 	return run
 }
+
+// RepositoryName is a repository as `owner/repository`, without the forge's address it was cloned from:
+// http://forge.internal:3000/owner/repo.git → owner/repo (novox/hq issue 288). What a check matches a
+// pull request's repository by is its owner and name, never the forge's address, so nothing is lost.
+func RepositoryName(repository string) string {
+	r := strings.TrimSuffix(strings.TrimSuffix(strings.TrimSpace(repository), "/"), ".git")
+	if r == "" {
+		return ""
+	}
+	if _, rest, found := strings.Cut(r, "://"); found {
+		r = rest
+		if _, path, found := strings.Cut(r, "/"); found {
+			r = path
+		} else {
+			return ""
+		}
+	} else if at := strings.Index(r, "@"); at >= 0 {
+		// scp-like: git@forge:owner/repo
+		if _, path, found := strings.Cut(r[at+1:], ":"); found {
+			r = path
+		}
+	}
+	parts := strings.Split(strings.Trim(r, "/"), "/")
+	if len(parts) >= 2 {
+		return parts[len(parts)-2] + "/" + parts[len(parts)-1]
+	}
+	return parts[len(parts)-1]
+}

internal/inventory/busrecords.go

@@ -137,6 +137,8 @@ func declaredFor(m catalogue.Manifest, seats map[string]catalogue.SeatDeclaratio
 		// And the state it keeps and reads (novox/hq ADR 0201).
 		State: bucketsOf(m),
 		Reads: m.Reads,
+		// And the tools its health asks (novox/hq ADR 0240): the machine's node-engine is granted them.
+		Checks: catalogue.HealthChecks(m),
 	}
 	// Whether it can be given an account at all: delivered as its own secret named broker, so one
 	// that declares none has nowhere to read it (novox/hq issue 195).

internal/inventory/health.go

@@ -0,0 +1,124 @@
+package inventory
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"time"
+
+	"github.com/jackc/pgx/v5"
+)
+
+// What each machine says of its long-running resources (novox/hq ADR 0240, to-be 48 §4): the newest
+// statement per machine, and per module how many statements in a row said a resource of it was unhealthy.
+
+// ResourceHealth is one long-running resource's state as the node-engine said it.
+type ResourceHealth struct {
+	Module   string    `json:"module"`
+	Resource string    `json:"resource"`
+	Kind     string    `json:"kind"`
+	Target   string    `json:"target"`
+	State    string    `json:"state"`
+	Reason   string    `json:"reason,omitempty"`
+	Since    time.Time `json:"since"`
+	Streak   int       `json:"streak,omitempty"`
+	Restarts int       `json:"restarts,omitempty"`
+	// Check is the declared check's kind, empty for liveness alone; Needs the provision it exercises
+	// (ADR 0240 Phase B, to-be 48 §6).
+	Check string `json:"check,omitempty"`
+	Needs string `json:"needs,omitempty"`
+}
+
+// NodeHealth is a machine's newest statement, as kept.
+type NodeHealth struct {
+	Node     string
+	Contract int
+	// SaidAt is when the engine looked (the machine's clock); HeardAt when this controller heard it.
+	SaidAt    time.Time
+	HeardAt   time.Time
+	Resources []ResourceHealth
+	// Streaks is, per module, how many statements in a row said a resource of it was unhealthy.
+	Streaks map[string]int
+}
+
+// HealthOf is a machine's newest statement; false when its node-engine has never stated one.
+func (i *Inventory) HealthOf(ctx context.Context, nodeName string) (NodeHealth, bool, error) {
+	all, err := i.healths(ctx, nodeName)
+	if err != nil {
+		return NodeHealth{}, false, err
+	}
+	h, ok := all[nodeName]
+	return h, ok, nil
+}
+
+// Healths is every machine's newest statement, by machine. A machine absent never stated one: its
+// node-engine is older than the judging, and its health is not known.
+func (i *Inventory) Healths(ctx context.Context) (map[string]NodeHealth, error) {
+	return i.healths(ctx, "")
+}
+
+func (i *Inventory) healths(ctx context.Context, only string) (map[string]NodeHealth, error) {
+	rows, err := i.store.Pool().Query(ctx,
+		`select n.name, h.contract, h.said_at, h.heard_at, h.resources, h.streaks
+		 from node_health h join node n on n.id = h.node
+		 where $1 = '' or n.name = $1`, only)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	out := map[string]NodeHealth{}
+	for rows.Next() {
+		var h NodeHealth
+		var resources, streaks []byte
+		if err := rows.Scan(&h.Node, &h.Contract, &h.SaidAt, &h.HeardAt, &resources, &streaks); err != nil {
+			return nil, err
+		}
+		if err := json.Unmarshal(resources, &h.Resources); err != nil {
+			return nil, fmt.Errorf("%s's health cannot be read: %w", h.Node, err)
+		}
+		if err := json.Unmarshal(streaks, &h.Streaks); err != nil {
+			return nil, fmt.Errorf("%s's health cannot be read: %w", h.Node, err)
+		}
+		out[h.Node] = h
+	}
+	return out, rows.Err()
+}
+
+// RecordHealth keeps a machine's statement in place of the one kept — unless the one kept is newer, by
+// when the engine looked: then nothing is written, and false says it was refused as older.
+func (i *Inventory) RecordHealth(ctx context.Context, h NodeHealth) (bool, error) {
+	if h.Resources == nil {
+		h.Resources = []ResourceHealth{}
+	}
+	if h.Streaks == nil {
+		h.Streaks = map[string]int{}
+	}
+	resources, err := json.Marshal(h.Resources)
+	if err != nil {
+		return false, err
+	}
+	streaks, err := json.Marshal(h.Streaks)
+	if err != nil {
+		return false, err
+	}
+	heard := h.HeardAt
+	if heard.IsZero() {
+		heard = time.Now()
+	}
+	var node string
+	err = i.store.Pool().QueryRow(ctx,
+		`insert into node_health (node, contract, said_at, heard_at, resources, streaks)
+		 select id, $2, $3, $4, $5, $6 from node where name = $1
+		 on conflict (node) do update set contract = excluded.contract, said_at = excluded.said_at,
+		     heard_at = excluded.heard_at, resources = excluded.resources, streaks = excluded.streaks
+		   where node_health.said_at <= excluded.said_at
+		 returning node`, h.Node, h.Contract, h.SaidAt, heard, resources, streaks).Scan(&node)
+	if errors.Is(err, pgx.ErrNoRows) {
+		if _, nerr := i.NodeByName(ctx, h.Node); nerr != nil {
+			return false, nerr
+		}
+		return false, nil
+	}
+	return err == nil, err
+}

internal/inventory/migrations/0076-a-module-says-how-it-is-healthy.sql

@@ -0,0 +1,26 @@
+-- A module says how it is healthy, and the node-engine judges it (novox/hq ADR 0240, to-be 48 Phase A).
+--
+-- Every machine's node-engine states the health of every long-running resource it runs for a module — a
+-- container that stays up, a process that stays up, a service stated running — in every report and as an
+-- event between reports. The controller keeps the newest statement per machine, here, so the release
+-- gate, `node show` and a controller started again all read the same word; and with it, per module, how
+-- many statements in a row said a resource of it was unhealthy — `module.<module>.<machine>.unhealthy` is
+-- raised on the second (ADR 0240 §4).
+--
+-- One row per machine, replaced, as node_report is: the question is the machine's state now. A machine
+-- whose node-engine is older than the judging has no row, and its health is not known — never healthy,
+-- never unhealthy.
+create table node_health (
+    node      uuid        primary key references node(id) on delete cascade,
+    -- The statement's version (the engine's liveness contract).
+    contract  int         not null,
+    -- When the node-engine looked, on the machine's clock: the order of its statements. An older one
+    -- than this is refused.
+    said_at   timestamptz not null,
+    -- When this controller heard it, on its own: what "since the send" is judged by.
+    heard_at  timestamptz not null default now(),
+    -- [{module, resource, kind, target, state, reason, since, streak, restarts}], as the engine said them.
+    resources jsonb       not null default '[]',
+    -- {module: statements in a row saying a resource of it is unhealthy}.
+    streaks   jsonb       not null default '{}'
+);

internal/link/bus.go

@@ -75,8 +75,16 @@ const (
 	// ToolsAliveSubjects is every machine's node tools saying they are there (novox/hq to-be 45 §3,
 	// S11): core NATS like the host's, for the same reason.
 	ToolsAliveSubjects = "mesh.control.*.tools-alive"
+
+	// HealthSubjects is every machine's health statement between its reports (novox/hq ADR 0240): core
+	// NATS like the heartbeat, because a statement lost is said again within a minute while anything is
+	// not healthy, and the next report carries it whatever happens.
+	HealthSubjects = "mesh.control.*.health"
 )
 
+// HealthSubject is one machine's health statement.
+func HealthSubject(node string) string { return "mesh.control." + node + ".health" }
+
 // ReportSubject is where one node says what it did. On the CONTROL stream, because it is the
 // message the store-window guarantee is about (ADR 0083).
 func ReportSubject(node string) string { return "mesh.control." + node + ".report" }

internal/link/calls.go

@@ -29,6 +29,15 @@ import (
 // then and with "still running as call <id>" when it does not; and every call is kept here, with
 // what came of it, including an answer the bus refused, so `calls` can say it.
 
+// ErrHandingOver is a call refused because the controller answering it is being replaced: stopping, or
+// unable to run its own build because the node-engine has moved it aside (novox/hq issue 289). Nothing
+// was done, and the controller after it answers the same call — so the answer carries
+// `"retry": RetryHandingOver`, and a caller asks once more.
+var ErrHandingOver = errors.New("the controller is handing over to the next one; nothing was done, ask again")
+
+// RetryHandingOver is the `retry` mark of an answer refused by ErrHandingOver.
+const RetryHandingOver = "handing-over"
+
 // AnswerWithin is how long a call runs before its caller is answered that it is still running. Well
 // inside the shortest wait of a caller the mesh ships (the console's thirty seconds) and the bus's
 // own window for an answer (broker.ResponseTTL), so the one answer a call has is never late for
@@ -524,7 +533,10 @@ func (l *CallLog) serveCall(seat, verb string, args json.RawMessage, reply strin
 		var body []byte
 		result, err := handle(ctx, args)
 		failed := err != nil
-		if err != nil {
+		if errors.Is(err, ErrHandingOver) {
+			// Marked as well as said, so a caller asks again without reading the words (issue 289).
+			body, _ = json.Marshal(map[string]any{"error": err.Error(), "retry": RetryHandingOver})
+		} else if err != nil {
 			body, _ = json.Marshal(map[string]any{"error": err.Error()})
 		} else if body, err = json.Marshal(map[string]any{"result": result}); err != nil {
 			failed = true

internal/link/calls_test.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"log"
 	"strings"
 	"sync"
@@ -177,3 +178,24 @@ func recentOf(l *CallLog) []Call {
 	out, _ := l.Recent()
 	return out
 }
+
+// A call refused because its controller is handing over says so in a mark as well as in words, so the
+// caller asks once more without parsing a sentence (novox/hq issue 289).
+func TestAHandoverRefusalIsMarkedRetryable(t *testing.T) {
+	l, a := NewCallLog(), newAnswers(t)
+	l.serveCall("mesh-controller", "rotate", nil, "_INBOX.x.9", func(context.Context, json.RawMessage) (any, error) {
+		return nil, fmt.Errorf("%w: stopping", ErrHandingOver)
+	}, a.respond, nil)
+	got := a.only()
+	if got["retry"] != RetryHandingOver || !strings.Contains(fmt.Sprint(got["error"]), "handing over") {
+		t.Fatalf("answered %v", got)
+	}
+
+	l, a = NewCallLog(), newAnswers(t)
+	l.serveCall("mesh-controller", "rotate", nil, "_INBOX.x.10", func(context.Context, json.RawMessage) (any, error) {
+		return nil, errors.New("refused for its own reason")
+	}, a.respond, nil)
+	if _, marked := a.only()["retry"]; marked {
+		t.Fatal("an ordinary refusal was marked to be asked again")
+	}
+}

internal/link/contracts.go

@@ -35,6 +35,10 @@ var Contracts = map[string]Contract{
 	KindHeartbeat: {Unordered: "a word that the machine is there: the newest heard is the newest said, and one " +
 		"lost is the next one"},
 	KindToolsHeartbeat: {Unordered: "a word that the node tools are there, as a machine's heartbeat"},
+	KindHealth: {Ordered: "by the time the node-engine looked (`at`, on the machine's clock): a statement older " +
+		"than the one kept for that machine is refused, so an event arriving after a newer report does not undo it " +
+		"(novox/hq ADR 0240)",
+		Tests: []string{"TestAnOlderHealthStatementIsRefused"}},
 	KindEnrolment: {Unordered: "a request answered once, under a token spent once: a second presentation is " +
 		"refused by the token, not by an order (issue 083)",
 		Tests: []string{"TestAnEnrolmentMetByAHeldTokenIsAskedToTryAgain"}},

internal/link/contracts_test.go

@@ -16,7 +16,7 @@ import (
 func TestEveryConsumedKindHasAContract(t *testing.T) {
 	// What the controller can be handed, from the subjects it is granted and the ones it derives.
 	subjects := []string{EnrolSubject, BuiltSubject, ReportSubject("anchor"), AliveSubject("anchor"),
-		ToolsAliveSubject("anchor"), "mesh.mod.postgres.event.provisioner.failing"}
+		ToolsAliveSubject("anchor"), HealthSubject("anchor"), "mesh.mod.postgres.event.provisioner.failing"}
 	subjects = append(subjects, broker.ControllerFollows...)
 	kinds := map[string]bool{}
 	for _, s := range subjects {

internal/link/health.go

@@ -0,0 +1,56 @@
+package link
+
+import (
+	"context"
+	"encoding/json"
+	"strings"
+)
+
+// A machine's health statement (novox/hq ADR 0240, to-be 48 §4).
+//
+// **The node-engine owns every verdict; the controller keeps the last word per machine.** A statement
+// arrives in every report and as an event between reports — on each change, and again every minute while
+// a resource is not healthy. What the controller does with it — keep it, refuse an older one, raise
+// `module.<module>.<machine>.unhealthy` on the second statement that says so — is the Healths it is given.
+
+// Healths keeps what machines state of their long-running resources.
+type Healths interface {
+	// Stated keeps one machine's statement, from a report or the event. An older statement than the one
+	// kept is refused there, by its time on the machine.
+	Stated(ctx context.Context, node string, h Health) error
+}
+
+// Hears says where the machines' health statements are kept. Their subscription is the heartbeats':
+// core, and always made, so nothing is asked of the bus here.
+func (s *Server) Hears(h Healths) { s.healths = h }
+
+// healthSaid acts on one health event. Core NATS, so there is nothing to hold: a statement that could not
+// be kept is said in the log, and the next one — a minute away while anything is not healthy — is kept.
+func (s *Server) healthSaid(ctx context.Context, m Control) {
+	defer func() { _ = m.Took() }()
+	var said HealthSaid
+	if err := json.Unmarshal(m.Body(), &said); err != nil || said.Health.Contract == 0 {
+		return
+	}
+	// The machine is the one in the subject the bus let it publish on, never the body's.
+	node, ok := nodeOfHealth(m.Subject())
+	if !ok || s.healths == nil {
+		return
+	}
+	if err := s.healths.Stated(ctx, node, said.Health); err != nil {
+		s.log.Printf("could not keep what %s says of its resources' health: %v", node, err)
+	}
+}
+
+// nodeOfHealth is the machine a health statement names in its subject.
+func nodeOfHealth(subject string) (string, bool) {
+	rest, ok := strings.CutPrefix(subject, "mesh.control.")
+	if !ok {
+		return "", false
+	}
+	node, kind, ok := strings.Cut(rest, ".")
+	if !ok || kind != "health" || node == "" || strings.Contains(node, ".") {
+		return "", false
+	}
+	return node, true
+}

internal/link/health_test.go

@@ -0,0 +1,37 @@
+package link
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+// A health event is kept as the machine its subject names says it, whatever its body claims, and taken
+// (novox/hq ADR 0240): core NATS, nothing to hold.
+
+type keptHealths struct{ by map[string]Health }
+
+func (k *keptHealths) Stated(_ context.Context, node string, h Health) error {
+	k.by[node] = h
+	return nil
+}
+
+func TestAHealthEventIsKeptAsTheMachineInItsSubject(t *testing.T) {
+	s, in := serving()
+	kept := &keptHealths{by: map[string]Health{}}
+	s.Hears(kept)
+	to := &settled{}
+	m := in.sends(t, to, KindHealth, HealthSaid{Node: "laptop", Health: Health{Contract: LivenessContract, At: time.Now(),
+		Resources: []ResourceHealth{{Module: "letta", Resource: "letta.server", State: StateUnhealthy}}}}).(*fakeControl)
+	m.subject = HealthSubject("anchor")
+	s.act(t.Context(), m)
+	if !to.acked {
+		t.Fatal("a health event was not taken")
+	}
+	if _, lied := kept.by["laptop"]; lied || len(kept.by["anchor"].Resources) != 1 {
+		t.Fatalf("kept %+v: the machine is the subject's, never the body's", kept.by)
+	}
+	if kind, ok := kindOfSubject(HealthSubject("anchor")); !ok || kind != KindHealth {
+		t.Fatalf("%s is not read as a health statement", HealthSubject("anchor"))
+	}
+}

internal/link/protocol.go

@@ -262,6 +262,13 @@ type Report struct {
 	// `witness` and `not-reversible` are sent only to a machine whose report carries it.
 	Witness int `json:"witness,omitempty"`
 
+	// Health is the machine's word on every long-running resource it runs for a module (novox/hq ADR
+	// 0240, to-be 48 §4; mesh-host's internal/liveness): its state, since when, its failing streak and
+	// the restarts its node-engine counted. **Absent from an engine older than the judging**, which is
+	// read as "not known" — never as healthy, never as a reason to raise anything; present with no
+	// resources from a machine that runs nothing long-lived.
+	Health *Health `json:"health,omitempty"`
+
 	// Rekey is a node taking a found tunnel's key as its overlay key after enrolment (novox/hq
 	// ADR 0105). A report carrying one is not an account of the machine: it moves the node's
 	// overlay key and tunnel and nothing else.
@@ -404,3 +411,54 @@ func EnrolProof(secret string, public []byte, overlay, sealing, serving string)
 	return []byte("novox-mesh-enrol\x00" + secret + "\x00" + base64.StdEncoding.EncodeToString(public) +
 		"\x00" + overlay + "\x00" + sealing + "\x00" + serving)
 }
+
+// LivenessContract is the version of the health statement this controller reads (ADR 0240 Phase A).
+const LivenessContract = 1
+
+// ReadinessContract is the statement of an engine that also reads a resource's declared `health` and
+// judges it (ADR 0240 Phase B): only to an engine whose statement says this or later is the field sent,
+// because an older one parses strictly and would refuse the whole declaration for it.
+const ReadinessContract = 2
+
+// Health is one statement of a machine's long-running resources (to-be 48 §4): in every report, as the
+// event HealthSubject between reports on each change, and again every minute while one is not healthy.
+// The node-engine's own (mesh-host internal/link Health); a test on each side holds the field names.
+type Health struct {
+	Contract int `json:"contract"`
+	// At is when the engine looked, on the machine's clock: the order of its statements.
+	At        time.Time        `json:"at"`
+	Resources []ResourceHealth `json:"resources"`
+}
+
+// The states a resource is said in (ADR 0240 §4).
+const (
+	StateHealthy   = "healthy"
+	StateUnhealthy = "unhealthy"
+	StateStarting  = "starting"
+	StateHeld      = "held"
+	StateUnknown   = "unknown"
+)
+
+// ResourceHealth is one long-running resource's state.
+type ResourceHealth struct {
+	Module   string    `json:"module"`
+	Resource string    `json:"resource"`
+	Kind     string    `json:"kind"`
+	Target   string    `json:"target"`
+	State    string    `json:"state"`
+	Reason   string    `json:"reason,omitempty"`
+	Since    time.Time `json:"since"`
+	Streak   int       `json:"streak,omitempty"`
+	Restarts int       `json:"restarts,omitempty"`
+	// Check is the declared check's kind (ADR 0240 Phase B), empty for a resource judged by liveness
+	// alone; Needs is the provision the check exercises (to-be 48 §6).
+	Check string `json:"check,omitempty"`
+	Needs string `json:"needs,omitempty"`
+}
+
+// HealthSaid is the health event's body: the machine and its statement. The machine is read from the
+// subject the bus let it publish on, never from here.
+type HealthSaid struct {
+	Node   string `json:"node"`
+	Health Health `json:"health"`
+}

internal/link/protocol_test.go

@@ -25,6 +25,14 @@ func TestTheWireFormatIsExactlyTheseFieldNames(t *testing.T) {
 			[]string{"protocol", "address", "port", "by", "published", "container-port"}},
 		{EnrolRequest{Node: "n", Secret: "s", PublicKey: []byte("k")},
 			[]string{"node", "secret", "public_key"}},
+		// novox/hq ADR 0240: every long-running resource's health, in every report and in its own event.
+		{Report{Node: "n", Health: &Health{Contract: LivenessContract}}, []string{"node", "health"}},
+		{Health{Contract: LivenessContract, Resources: []ResourceHealth{}}, []string{"contract", "at", "resources"}},
+		{ResourceHealth{Module: "m", Resource: "m.r", Kind: "container", Target: "t", State: StateUnhealthy,
+			Reason: "restarting", Streak: 2, Restarts: 3, Check: "http", Needs: "postgres-database"},
+			[]string{"module", "resource", "kind", "target", "state", "reason", "since", "streak", "restarts",
+				"check", "needs"}},
+		{HealthSaid{Node: "n"}, []string{"node", "health"}},
 	} {
 		raw, err := json.Marshal(c.value)
 		if err != nil {

internal/link/receive.go

@@ -30,8 +30,10 @@ const (
 	KindHeartbeat = "heartbeat"
 	// KindToolsHeartbeat is a machine's node tools saying they are there (novox/hq to-be 45 S11).
 	KindToolsHeartbeat = "tools-heartbeat"
-	KindBuilt          = "built"
-	KindModuleMoved    = "module-moved"
+	// KindHealth is a machine's statement of its long-running resources' health (novox/hq ADR 0240).
+	KindHealth      = "health"
+	KindBuilt       = "built"
+	KindModuleMoved = "module-moved"
 	// KindSourceMoved is the forge announcing a merge: a source moved, and what it produces is
 	// built without anybody telling the mesh (novox/hq 04-ISSUES/131).
 	KindSourceMoved = "source-moved"

internal/link/receive_nats.go

@@ -108,6 +108,13 @@ func (n *natsInbound) Receive(ctx context.Context, act func(context.Context, Con
 		return fmt.Errorf("subscribing to the node tools' heartbeats: %w", err)
 	}
 	defer func() { _ = toolsAlive.Unsubscribe() }()
+	// And what each machine says of its long-running resources between reports (novox/hq ADR 0240): core
+	// too, and said again while it matters.
+	health, err := conn.ChanSubscribe(HealthSubjects, beats)
+	if err != nil {
+		return fmt.Errorf("subscribing to the machines' health: %w", err)
+	}
+	defer func() { _ = health.Unsubscribe() }()
 
 	// The events the controller follows, when something is listening for them.
 	var events chan *nats.Msg
@@ -239,6 +246,8 @@ func kindOfSubject(subject string) (string, bool) {
 			return KindHeartbeat, true
 		case "tools-alive":
 			return KindToolsHeartbeat, true
+		case "health":
+			return KindHealth, true
 		}
 	}
 	switch subject {

internal/link/serve.go

@@ -94,6 +94,8 @@ type Server struct {
 	retirements Retirements
 	// checker asks for a pull request's merge check (novox/hq to-be 45 §9).
 	checker Checker
+	// healths keeps what machines say of their long-running resources (novox/hq ADR 0240).
+	healths Healths
 
 	log *log.Logger
 	// giveUp is how long one message is held for the store; zero means GiveUpAfter.
@@ -194,6 +196,8 @@ func (s *Server) act(ctx context.Context, m Control) {
 		s.heartbeat(m)
 	case KindToolsHeartbeat:
 		s.toolsHeartbeat(m)
+	case KindHealth:
+		s.healthSaid(ctx, m)
 	case KindBuilt:
 		s.wasBuilt(ctx, m)
 	case KindModuleMoved:

module.json

@@ -60,6 +60,7 @@
     "pause",
     "resume",
     "hand-act",
+    "drill",
     "hand-acts",
     "durations",
     "conditions",

vendor/github.com/novox/mesh-host/internal/declaration/declaration.go

@@ -590,6 +590,23 @@ type Process struct {
 	// For a process that stays up; a step or a scheduled run is not running a moment later by
 	// design, so there is nothing to hand over to.
 	Replaces []string `json:"replaces,omitempty"`
+
+	// Witness is how the node-engine judges a new build of this process, and restores the build
+	// before it when the new one is not healthy in bound (novox/hq to-be 45 §8): "lease" — the
+	// controller this machine started holds the controller's lease; "ping" — this machine's runtime
+	// answers the services protocol's PING; "none". Absent is the default for the process's name:
+	// the mesh's two core processes are judged, nothing else is. For a process that stays up.
+	Witness string `json:"witness,omitempty"`
+
+	// NotReversible says why this build may not be rolled back, when it may not: the build before it
+	// would run against what this one changes — a migration it runs that the older build cannot read
+	// (to-be 45 §8, rule 8). A build so declared that is not healthy in bound is left running and said
+	// as urgent; the build before it is never started against the newer data.
+	NotReversible string `json:"not-reversible,omitempty"`
+
+	// Health is how this resource is ready (novox/hq ADR 0240 rule 2, Phase B): one kind and its
+	// timing, judged by the node-engine beside liveness. Absent: judged alive or not, and nothing more.
+	Health *Health `json:"health,omitempty"`
 }
 
 func (d *Process) Identity() string { return d.ID }
@@ -621,7 +638,7 @@ func ProcessNameProblem(name string) string {
 }
 
 func (d *Process) validate(where string, _ bool) []string {
-	var problems []string
+	problems := d.Health.problems(where, false, !d.RunOnce && d.Schedule == "")
 	if problem := ProcessNameProblem(d.Name); problem != "" {
 		problems = append(problems, where+": "+problem)
 	}
@@ -637,6 +654,19 @@ func (d *Process) validate(where string, _ bool) []string {
 	if len(d.Run) == 0 {
 		problems = append(problems, where+": a process needs to say what to run")
 	}
+	switch d.Witness {
+	case "", "lease", "ping", "none":
+	default:
+		problems = append(problems, fmt.Sprintf("%s: witness %q is not one this host keeps: lease, ping or none",
+			where, d.Witness))
+	}
+	if d.Witness != "" && d.Witness != "none" && (d.RunOnce || d.Schedule != "") {
+		problems = append(problems, where+": a witness judges a process that stays up; a step or a "+
+			"scheduled run is not running between its runs")
+	}
+	if strings.ContainsAny(d.NotReversible, "\n\r") {
+		problems = append(problems, where+": not-reversible is one line")
+	}
 	for _, part := range d.Run {
 		if part == "" {
 			problems = append(problems, where+": a process command has an empty element")
@@ -756,6 +786,10 @@ type Service struct {
 	// said by the controller, which knows the found tunnel's key is this node's own: without that,
 	// starting this unit on the found one's port would drop every peer's packets.
 	TakesOver *TakeOver `json:"takes-over,omitempty"`
+
+	// Health is how this resource is ready (novox/hq ADR 0240 rule 2, Phase B): one kind and its
+	// timing, judged by the node-engine beside liveness. Absent: judged alive or not, and nothing more.
+	Health *Health `json:"health,omitempty"`
 }
 
 // TakeOver is a found tunnel a service replaces: its interface, the unit that raised it, and its
@@ -784,7 +818,7 @@ const (
 func (s *Service) UserScoped() bool { return s.Scope == ScopeUser }
 
 func (s *Service) validate(where string, _ bool) []string {
-	var problems []string
+	problems := s.Health.problems(where, false, s.State == "running")
 	if s.Unit == "" {
 		problems = append(problems, where+": a service needs a unit")
 	}
@@ -1096,6 +1130,10 @@ type Container struct {
 	// offline job says *before*, not *instead of*; a recurring window is the case order cannot
 	// express, and the only one this serves.
 	WhileStopped []string `json:"while-stopped,omitempty"`
+
+	// Health is how this resource is ready (novox/hq ADR 0240 rule 2, Phase B): one kind and its
+	// timing, judged by the node-engine beside liveness. Absent: judged alive or not, and nothing more.
+	Health *Health `json:"health,omitempty"`
 }
 
 func (c *Container) Identity() string { return c.ID }
@@ -1103,7 +1141,7 @@ func (c *Container) Kind() Type       { return TypeContainer }
 func (c *Container) Target() string   { return c.Name }
 
 func (c *Container) validate(where string, _ bool) []string {
-	var problems []string
+	problems := c.Health.problems(where, true, !c.RunOnce && c.Schedule == "")
 	if c.Name == "" {
 		problems = append(problems, where+": a container needs a name")
 	}

vendor/github.com/novox/mesh-host/internal/declaration/health.go

@@ -0,0 +1,188 @@
+package declaration
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"strings"
+	"time"
+)
+
+// Health is how a long-running resource is ready, as the controller composed it from the module's
+// `health` (novox/hq ADR 0240 rule 2, to-be 48 §2–§3, Phase B): one kind and its timing, the endpoint
+// already the port this machine published it on.
+//
+// **The node-engine runs every kind and owns every verdict.** http and tcp it makes itself, from the
+// machine to the port; unit it reads from the service manager it already reads; exec and runtime it hands
+// to the container runtime as the container's own check, with this timing, and reads the state; tool it
+// asks of its own node tools. Nothing else on the machine sets a container's check.
+//
+// Refused here as the controller refuses it near the author, in the same bounds: an engine that took a
+// check it could not judge would say a module ready that nothing looked at.
+type Health struct {
+	Kind string `json:"kind"`
+	// Endpoint is the module's name for what Port is: for the words a verdict is said in.
+	Endpoint string `json:"endpoint,omitempty"`
+	Port     int    `json:"port,omitempty"`
+	Path     string `json:"path,omitempty"`
+	Status   int    `json:"status,omitempty"`
+	Body     string `json:"body,omitempty"`
+	Scheme   string `json:"scheme,omitempty"`
+	Command  string `json:"command,omitempty"`
+	Tool     string `json:"tool,omitempty"`
+	Interval string `json:"interval"`
+	Timeout  string `json:"timeout"`
+	Looks    int    `json:"looks"`
+	Grace    string `json:"grace"`
+	// Needs is the provision the check exercises (to-be 48 §6): said with every verdict, so the
+	// controller can hold what it finds under the provider's own condition.
+	Needs string `json:"needs,omitempty"`
+}
+
+// UnmarshalJSON reads a health strictly, as everything a declaration carries is read: a field this host
+// does not know is a part of the check the controller believes it asked for, and nothing would look at it.
+func (h *Health) UnmarshalJSON(raw []byte) error {
+	type plain Health
+	var p plain
+	dec := json.NewDecoder(bytes.NewReader(raw))
+	dec.DisallowUnknownFields()
+	if err := dec.Decode(&p); err != nil {
+		return fmt.Errorf("health: %w", err)
+	}
+	*h = Health(p)
+	return nil
+}
+
+// The kinds.
+const (
+	HealthRuntime = "runtime"
+	HealthHTTP    = "http"
+	HealthTCP     = "tcp"
+	HealthExec    = "exec"
+	HealthUnit    = "unit"
+	HealthTool    = "tool"
+)
+
+// The bounds (ADR 0240 rule 2) — the controller's, held again here.
+const (
+	HealthIntervalFloor = 10 * time.Second
+	HealthLooksFloor    = 2
+	HealthWithin        = 5 * time.Minute
+)
+
+// Every, Within and GraceOf are the timing, read. Validated on arrival, so a parse error here is
+// impossible on a declaration that was accepted; it reads as zero.
+func (h *Health) Every() time.Duration   { d, _ := time.ParseDuration(h.Interval); return d }
+func (h *Health) Within() time.Duration  { d, _ := time.ParseDuration(h.Timeout); return d }
+func (h *Health) GraceOf() time.Duration { d, _ := time.ParseDuration(h.Grace); return d }
+
+// RunByRuntime says the container runtime runs this check as the container's own: exec and runtime.
+func (h *Health) RunByRuntime() bool {
+	return h != nil && (h.Kind == HealthExec || h.Kind == HealthRuntime)
+}
+
+// Words is the check in a few words, as a verdict is said: "http /healthz on web".
+func (h *Health) Words() string {
+	switch h.Kind {
+	case HealthHTTP:
+		return "http " + h.Path + " on " + orPort(h.Endpoint, h.Port)
+	case HealthTCP:
+		return "tcp on " + orPort(h.Endpoint, h.Port)
+	case HealthTool:
+		return "its tool " + h.Tool
+	case HealthRuntime:
+		return "its image's own check"
+	case HealthExec:
+		return "its command"
+	case HealthUnit:
+		return "its unit"
+	}
+	return h.Kind
+}
+
+func orPort(endpoint string, port int) string {
+	if endpoint != "" {
+		return endpoint
+	}
+	return fmt.Sprint(port)
+}
+
+// problems holds a resource's health to its kind and bounds. container says whether the resource is a
+// container; longRunning whether it stays up.
+func (h *Health) problems(where string, container, longRunning bool) []string {
+	if h == nil {
+		return nil
+	}
+	var problems []string
+	say := func(format string, args ...any) {
+		problems = append(problems, where+": "+fmt.Sprintf(format, args...))
+	}
+	if !longRunning {
+		say("health is judged on what stays up; a step or a scheduled run is judged by its own outcome")
+	}
+	switch h.Kind {
+	case HealthHTTP, HealthTCP:
+		if h.Port < 1 || h.Port > 65535 {
+			say("a %s check needs the port it looks at", h.Kind)
+		}
+	case HealthExec:
+		if !container {
+			say("an exec check runs inside a container")
+		}
+		if strings.TrimSpace(h.Command) == "" {
+			say("an exec check needs a command")
+		}
+	case HealthRuntime:
+		if !container {
+			say("a runtime check is a container image's own")
+		}
+	case HealthUnit:
+		if container {
+			say("a unit check is a service's or a process's own")
+		}
+	case HealthTool:
+		if strings.TrimSpace(h.Tool) == "" {
+			say("a tool check names the tool")
+		}
+	default:
+		say("health of kind %q; it is runtime, http, tcp, exec, unit or tool", h.Kind)
+	}
+	if h.Kind == HealthHTTP {
+		if !strings.HasPrefix(h.Path, "/") {
+			say("an http check asks a path starting with /")
+		}
+		if h.Status != 0 && (h.Status < 100 || h.Status > 599) {
+			say("an http check expects status %d, which is not one", h.Status)
+		}
+		if h.Scheme != "" && h.Scheme != "http" && h.Scheme != "https" {
+			say("an http check is over http or https, not %q", h.Scheme)
+		}
+	}
+	if strings.ContainsAny(h.Command, "\n\r") {
+		say("an exec check's command is one line")
+	}
+	every, everyErr := time.ParseDuration(h.Interval)
+	within, withinErr := time.ParseDuration(h.Timeout)
+	grace, graceErr := time.ParseDuration(h.Grace)
+	switch {
+	case everyErr != nil || withinErr != nil || graceErr != nil:
+		say("health's interval, timeout and grace are durations")
+	default:
+		if every < HealthIntervalFloor {
+			say("a check looks no more often than every %s, not every %s", HealthIntervalFloor, every)
+		}
+		if within <= 0 || within >= every {
+			say("a look takes more than nothing and less than its interval")
+		}
+		if grace < 0 {
+			say("a grace is not negative")
+		}
+		if h.Looks >= HealthLooksFloor && grace+time.Duration(h.Looks)*every > HealthWithin {
+			say("a grace and the failing looks take at most %s", HealthWithin)
+		}
+	}
+	if h.Looks < HealthLooksFloor {
+		say("a check is unhealthy after at least %d failing looks, not %d", HealthLooksFloor, h.Looks)
+	}
+	return problems
+}

vendor/modules.txt

@@ -75,7 +75,7 @@ github.com/nats-io/nkeys
 # github.com/nats-io/nuid v1.0.1
 ## explicit
 github.com/nats-io/nuid
-# github.com/novox/mesh-host v0.0.0 => git.novox.be/novox/mesh-host v0.0.0-20261006095519-3e80b7ae325e
+# github.com/novox/mesh-host v0.0.0 => git.novox.be/novox/mesh-host v0.0.0-20261007120832-bdd44154ccac
 ## explicit; go 1.26.0
 github.com/novox/mesh-host/internal/declaration
 github.com/novox/mesh-host/validate
@@ -133,4 +133,4 @@ golang.org/x/text/width
 # golang.org/x/time v0.15.0
 ## explicit; go 1.25.0
 golang.org/x/time/rate
-# github.com/novox/mesh-host => git.novox.be/novox/mesh-host v0.0.0-20261006095519-3e80b7ae325e
+# github.com/novox/mesh-host => git.novox.be/novox/mesh-host v0.0.0-20261007120832-bdd44154ccac