hold the plugin ceiling at the start, not on a ticker

Warm was an argument to Evict, so the count was TRIMMED once a minute rather than
held. That is a bound only while starts arrive slower than the sweep, and the
thing that starts fastest is exactly the thing that needs bounding: a host whose
fan-out door asks every subsystem at once started ~100 children inside 90
seconds, held all of them until the next sweep, stopped answering its own
liveness probe, and was killed. The ceiling appeared only in the logs of a
container that was already gone.

So Warm is Config now and makeRoom applies it in startOnDemand, before another
process exists: if the host is at its ceiling, the least recently used evictable
plugin is stopped to make room. Age keeps its own sweep — steady state and burst
are different bounds and both belong to the same policy.

Adoption is what makes it possible. A plugin's `app` is the definition that
Load'ed it, which for a composed service is a sub-app holding one plugin and no
ceiling, so asking it "how many are running" answers 1 and "what is the budget"
answers 0 — measured, and the reason a first attempt at this was inert. The root
of a composition is known in exactly one place, where the generation computes its
hosts, so that is where every plugin learns its owner.

Both halves are proven necessary by mutation: dropping the start-path call, or
dropping the adoption, each turns the burst test from peak 4 into peak 12 against
a ceiling of 4. Race-clean over two runs; full suite green.

Co-authored-by: Hanzo Dev <[email protected]>

generation.go

@@ -95,6 +95,17 @@ func (a *App) build() (*generation, error) {
 	g.serve = g.router.Handler()
 	g.ops = composeOps(occ)
 	g.hosts = computeHosts(a, occ)
+	// Adopt: every plugin in this composition learns which app is the ROOT of it.
+	// This is the only place that knows, and a plugin needs it to ask the two
+	// questions its own definition cannot answer — how many processes are running
+	// across the whole host, and what the host's ceiling is.
+	for _, h := range g.hosts {
+		h.plugMu.Lock()
+		for _, pl := range h.plugins {
+			pl.owner.Store(a)
+		}
+		h.plugMu.Unlock()
+	}
 	g.routes = len(g.router.GetRoutes(false))
 	return g, nil
 }

idle.go

@@ -45,13 +45,21 @@ func (p Plugin) idleAfter() time.Duration { return p.IdleAfter }
 //
 // One pass, no goroutine — the caller decides the cadence, so a host under test
 // can drive it directly and a host in production can put it on the ticker it
-// already has. warm is an ARGUMENT rather than a field because the number of
+// already has.
-// processes a host may hold is a fact about the host's memory budget, which the
+//
-// host knows and a library cannot.
+// THE COUNT IS NOT THIS FUNCTION'S TO ENFORCE ALONE, and an earlier version of
+// this comment claimed otherwise: warm was an ARGUMENT here, on the reasoning
+// that the number is a fact about the host and not the library. The reasoning
+// held; the placement did not. A bound consulted once a minute is outrun by
+// anything that starts faster, and a fan-out door that asks every subsystem at
+// once does — a hundred children inside ninety seconds, so the process held them
+// all until the next sweep, stopped answering its own liveness probe, and was
+// killed with the ceiling never applied. Warm is state now and makeRoom applies
+// it at the start path; this still runs it, because age and count are one sweep.
 //
 // Draining uses the plugin's own Drain, so a request in flight when the sweep
 // lands finishes on the old process exactly as it would across a Reload.
-func (a *App) Evict(warm int) int {
+func (a *App) Evict() int {
 	now := time.Now()
 	stopped := 0
 	for _, p := range a.pluginSet() {
@@ -59,7 +67,61 @@ func (a *App) Evict(warm int) int {
 			stopped++
 		}
 	}
-	return stopped + a.evictOver(warm, now)
+	return stopped + a.evictOver(a.warm, now)
+}
+
+// host is the ROOT app of the composition this plugin belongs to, which is the
+// only app that can answer how many processes are running and what the budget is.
+// Falls back to the defining app for a plugin whose composition was never built —
+// a Load'ed service used directly in a test, where the two are the same thing.
+func (p *plugin) host() *App {
+	if o := p.owner.Load(); o != nil {
+		return o
+	}
+	return p.app
+}
+
+// makeRoom is the ceiling, and it runs where a process is about to exist rather
+// than on a ticker. Called from the start path with the starter's own lock held,
+// so it must never consider the starter itself: evicting p here would deadlock on
+// p.mu, and p is the one plugin about to be needed anyway.
+//
+// This is the half a sweep cannot do. Evict trims to warm once a minute, which is
+// a bound only while starts arrive slower than that. Measured on a host whose MCP
+// door asks every subsystem at once: ~100 children started inside 90 seconds, the
+// pod stopped answering its own liveness probe, and the kubelet killed it — the
+// "ceiling" observed only in the logs of a container that was already gone.
+func (a *App) makeRoom(starter *plugin) {
+	if a.warm <= 0 {
+		return
+	}
+	now := time.Now()
+	for {
+		var (
+			live      int
+			evictable []*plugin
+		)
+		for _, p := range a.pluginSet() {
+			if p.cur.Load() == nil {
+				continue
+			}
+			live++
+			if p != starter && p.spec.Lazy && p.spec.idleAfter() > 0 {
+				evictable = append(evictable, p)
+			}
+		}
+		// The starter is not running yet, so room for it means strictly under.
+		if live < a.warm || len(evictable) == 0 {
+			return
+		}
+		slices.SortFunc(evictable, func(x, y *plugin) int {
+			return cmp.Compare(x.lastUse.Load(), y.lastUse.Load())
+		})
+		cold := evictable[0]
+		if !cold.evict("room", now.Sub(time.Unix(0, cold.lastUse.Load()))) {
+			return // already down; re-reading would spin
+		}
+	}
 }
 
 // pluginSet is every plugin this host and its composed hosts hold, read once
@@ -187,7 +249,7 @@ func (p *plugin) evict(reason string, idle time.Duration) bool {
 // Reap runs Evict on a ticker until the returned function is called. A host
 // that wants the behaviour writes one line; a host that does not is unaffected,
 // because nothing here runs unless it is asked for.
-func (a *App) Reap(every time.Duration, warm int) (stop func()) {
+func (a *App) Reap(every time.Duration) (stop func()) {
 	if every <= 0 {
 		every = time.Minute
 	}
@@ -202,7 +264,7 @@ func (a *App) Reap(every time.Duration, warm int) (stop func()) {
 			case <-done:
 				return
 			case <-t.C:
-				a.Evict(warm)
+				a.Evict()
 			}
 		}
 	}()

idle_bench_test.go

@@ -51,7 +51,7 @@ func TestIdle_ReclaimsRealMemory(t *testing.T) {
 	warmRSS := totalRSS(t, warmPIDs)
 
 	time.Sleep(80 * time.Millisecond)
-	if n := app.Evict(0); n != plugins {
+	if n := app.Evict(); n != plugins {
 		t.Fatalf("evicted %d of %d", n, plugins)
 	}
 	if !settles(3*time.Second, func() bool { return len(runningPIDs(t, app)) == 0 }) {

idle_test.go

@@ -35,7 +35,7 @@ func TestIdle_EvictsAndComesBack(t *testing.T) {
 
 	// Fresh from a request, so it is NOT idle. An evictor that stops a plugin
 	// that just served would be a request-latency bug that only shows in prod.
-	if n := app.Evict(0); n != 0 {
+	if n := app.Evict(); n != 0 {
 		t.Fatalf("evicted %d plugins that had just served", n)
 	}
 	if !app.Plugins()[0].Running {
@@ -43,7 +43,7 @@ func TestIdle_EvictsAndComesBack(t *testing.T) {
 	}
 
 	time.Sleep(80 * time.Millisecond)
-	if n := app.Evict(0); n != 1 {
+	if n := app.Evict(); n != 1 {
 		t.Fatalf("Evict stopped %d, want 1 after idling past IdleAfter", n)
 	}
 
@@ -86,7 +86,7 @@ func TestIdle_ZeroNeverEvicts(t *testing.T) {
 	pid := app.Plugins()[0].PID
 
 	time.Sleep(60 * time.Millisecond)
-	if n := app.Evict(0); n != 0 {
+	if n := app.Evict(); n != 0 {
 		t.Fatalf("evicted %d with IdleAfter unset — zero must mean never", n)
 	}
 	if got := app.Plugins()[0]; !got.Running || got.PID != pid {
@@ -120,7 +120,7 @@ func TestIdle_EagerIsNeverEvicted(t *testing.T) {
 		t.Fatal("eager plugin did not serve")
 	}
 	time.Sleep(40 * time.Millisecond)
-	if n := app.Evict(0); n != 0 {
+	if n := app.Evict(); n != 0 {
 		t.Fatalf("evicted %d eager plugins — IdleAfter applies to lazy only", n)
 	}
 	if got := app.Plugins()[0]; !got.Running || got.PID != pid {
@@ -146,7 +146,7 @@ func TestIdle_ReapRunsAndStops(t *testing.T) {
 	if status, _ := call(t, app, "GET", "/v1/demo/version", ""); status != 200 {
 		t.Fatal("first request failed")
 	}
-	stop := app.Reap(10*time.Millisecond, 0)
+	stop := app.Reap(10 * time.Millisecond)
 	if !settles(3*time.Second, func() bool { return !app.Plugins()[0].Running }) {
 		t.Fatal("Reap never evicted an idle plugin")
 	}

load.go

@@ -208,7 +208,12 @@ type plugin struct {
 	prefixes []string // every subtree this plugin answers
 	spec     Plugin
 
-	app      *App // for logging and supervision when started on demand
+	app *App // for logging and supervision when started on demand
+	// owner is the ROOT app of the composition this plugin ended up in, stamped
+	// when a generation is built. app is the definition that Load'ed it, which for
+	// a composed service is a sub-app holding one plugin and no ceiling — so it is
+	// the wrong thing to ask "how many are running" or "what is the budget".
+	owner    atomic.Pointer[App]
 	cur      atomic.Pointer[instance]
 	mu       sync.Mutex
 	reloads  atomic.Int64
@@ -328,6 +333,10 @@ func (p *plugin) startOnDemand() (Client, string) {
 	if p.closed || p.disabled.Load() {
 		return nil, ""
 	}
+	// The ceiling is applied HERE, before another process exists, so warm bounds
+	// what this host holds rather than what a sweep restores a minute later.
+	p.host().makeRoom(p)
+
 	in, err := start(p.spec)
 	if err != nil {
 		p.app.logger.Error("zip lazy plugin failed to start", "name", p.name, "err", err)

warm_test.go

@@ -8,20 +8,20 @@ import (
 	"github.com/zap-proto/zip"
 )
 
-// warmHost loads n lazy plugins that never age out, so anything these tests see
+// The ceiling is enforced where a plugin STARTS, so most of these tests never
-// evicted was evicted by the COUNT bound and not by IdleAfter. Every one is
+// call Evict: they ask for plugins and read how many processes exist. That is the
-// warmed by a request, because a plugin nothing asked for is not running and
+// design — a bound restored by a sweep is not a bound, because the burst that
-// would not count toward the bound in the first place.
+// matters is over before the sweep runs.
 //
 // Dir is sockDir(t) and not t.TempDir(): a unix socket path is bounded at ~108
-// bytes, and t.TempDir() spells the TEST'S NAME into it, so a descriptive name
+// bytes and t.TempDir() spells the TEST'S NAME into it, so a descriptive name
 // pushes the child's listen past the limit and it dies with `bind: invalid
 // argument` — which surfaces as a 503 that reads exactly like a lazy plugin
 // declining to start.
-func warmHost(t *testing.T, n int, idle time.Duration) *zip.App {
+func warmHost(t *testing.T, n, warm int, idle time.Duration) *zip.App {
 	t.Helper()
 	bin := buildPlugin(t, "v1")
-	app := zip.New(zip.Config{AppName: "host", DisableStartupMessage: true})
+	app := zip.New(zip.Config{AppName: "host", DisableStartupMessage: true, Warm: warm})
 	for i := range n {
 		name := fmt.Sprintf("demo%d", i)
 		app.Use(must(zip.Load(
@@ -32,91 +32,115 @@ func warmHost(t *testing.T, n int, idle time.Duration) *zip.App {
 		)))
 	}
 	t.Cleanup(func() { _ = app.Shutdown() })
-
-	// Warmed oldest-first with a gap, so last-use order is demo0 < demo1 < ...
-	// and the coldest is deterministic rather than whatever the clock did.
-	for i := range n {
-		if status, _ := call(t, app, "GET", fmt.Sprintf("/v1/demo%d/version", i), ""); status != 200 {
-			t.Fatalf("demo%d did not serve", i)
-		}
-		time.Sleep(2 * time.Millisecond)
-	}
-	if got := len(runningPIDs(t, app)); got != n {
-		t.Fatalf("warmed %d plugins, want %d", got, n)
-	}
 	return app
 }
 
-// TestWarm_StopsTheColdestUntilUnderTheBound is the bound doing its job, and it
+// ask serves one request per plugin, oldest first with a gap, so last-use order is
-// checks WHICH were stopped rather than only how many: a policy that evicted the
+// demo0 < demo1 < ... and "the coldest" is deterministic rather than whatever the
-// hottest would satisfy a count assertion and be exactly backwards.
+// clock did. It returns the PEAK number of processes seen along the way, which is
-func TestWarm_StopsTheColdestUntilUnderTheBound(t *testing.T) {
+// the figure a trim-afterwards implementation gets wrong.
+func ask(t *testing.T, app *zip.App, n int) (peak int) {
+	t.Helper()
+	for i := range n {
+		if status, body := call(t, app, "GET", fmt.Sprintf("/v1/demo%d/version", i), ""); status != 200 {
+			t.Fatalf("demo%d did not serve: status=%d body=%q", i, status, body)
+		}
+		if live := len(runningPIDs(t, app)); live > peak {
+			peak = live
+		}
+		time.Sleep(2 * time.Millisecond)
+	}
+	return peak
+}
+
+// TestWarm_CeilingHoldsDuringABurst is the property a sweep cannot provide, and
+// the one whose absence took a production pod down. NO sweep runs here — no Reap,
+// no Evict — so the only thing that can hold the line is the start path. The count
+// is read after EVERY request, because the end is exactly where a
+// trim-afterwards implementation also looks correct.
+func TestWarm_CeilingHoldsDuringABurst(t *testing.T) {
+	const plugins, warm = 12, 4
+	app := warmHost(t, plugins, warm, time.Hour) // an hour: age can reclaim nothing
+
+	peak := ask(t, app, plugins)
+	if peak > warm {
+		t.Fatalf("peak %d processes against a ceiling of %d — the bound is being "+
+			"restored by a sweep instead of enforced at the start", peak, warm)
+	}
+	if peak != warm { // or this proves only that twelve fit
+		t.Errorf("peak was %d, want exactly the ceiling %d", peak, warm)
+	}
+	t.Logf("%d sequential cold starts, peak live %d, ceiling %d", plugins, peak, warm)
+
+	// A ceiling is not an Unload: the route evicted first still answers.
+	if status, _ := call(t, app, "GET", "/v1/demo0/version", ""); status != 200 {
+		t.Fatal("the coldest plugin stopped answering after being evicted for room")
+	}
+}
+
+// TestWarm_KeepsTheMostRecentlyUsed checks WHICH survive, not just how many: a
+// policy that shed the hottest would satisfy a count assertion and be exactly
+// backwards.
+func TestWarm_KeepsTheMostRecentlyUsed(t *testing.T) {
 	const plugins, warm = 6, 2
-	app := warmHost(t, plugins, time.Hour) // an hour: nothing can age out here
+	app := warmHost(t, plugins, warm, time.Hour)
+	ask(t, app, plugins)
 
-	if n := app.Evict(warm); n != plugins-warm {
-		t.Fatalf("Evict(%d) stopped %d of %d, want %d", warm, n, plugins, plugins-warm)
-	}
-	if !settles(3*time.Second, func() bool { return len(runningPIDs(t, app)) == warm }) {
-		t.Fatalf("%d still running, want %d", len(runningPIDs(t, app)), warm)
-	}
-
-	// The survivors must be the two most recently used, demo4 and demo5.
 	live := map[string]bool{}
 	for _, p := range app.Plugins() {
 		if p.Running {
 			live[p.Name] = true
 		}
 	}
-	for _, name := range []string{"demo4", "demo5"} {
+	for _, n := range []string{"demo4", "demo5"} {
-		if !live[name] {
+		if !live[n] {
-			t.Errorf("%s was evicted, but it was among the %d most recently used", name, warm)
+			t.Errorf("%s is not running, but it is among the %d most recently used", n, warm)
 		}
 	}
-	for _, name := range []string{"demo0", "demo1", "demo2", "demo3"} {
+	for _, n := range []string{"demo0", "demo1", "demo2", "demo3"} {
-		if live[name] {
+		if live[n] {
-			t.Errorf("%s survived, but it was colder than the survivors", name)
+			t.Errorf("%s survived, but it is colder than the survivors", n)
 		}
 	}
 }
 
-// TestWarm_ZeroIsUnbounded pins the default. A host that never passes a bound
+// TestWarm_ZeroIsUnbounded pins the default: a host that states no ceiling keeps
-// must keep the behaviour it had before the bound existed.
+// the behaviour it had before one existed.
 func TestWarm_ZeroIsUnbounded(t *testing.T) {
 	const plugins = 4
-	app := warmHost(t, plugins, time.Hour)
+	app := warmHost(t, plugins, 0, time.Hour)
 
-	if n := app.Evict(0); n != 0 {
+	if peak := ask(t, app, plugins); peak != plugins {
-		t.Fatalf("Evict(0) stopped %d — zero must mean unbounded", n)
+		t.Fatalf("peak %d of %d with no ceiling — zero must mean unbounded", peak, plugins)
 	}
-	if got := len(runningPIDs(t, app)); got != plugins {
+	if n := app.Evict(); n != 0 {
-		t.Fatalf("%d running after Evict(0), want %d", got, plugins)
+		t.Fatalf("Evict stopped %d with no ceiling and nothing idle", n)
 	}
 }
 
-// TestWarm_UnderTheBoundStopsNothing: the bound is a ceiling, not a target. A
+// TestWarm_UnderTheCeilingStopsNothing: it is a ceiling, not a target. A host
-// host holding fewer processes than it may must not pay a cold start to reach it.
+// holding fewer processes than it may must not pay a cold start to reach it.
-func TestWarm_UnderTheBoundStopsNothing(t *testing.T) {
+func TestWarm_UnderTheCeilingStopsNothing(t *testing.T) {
-	app := warmHost(t, 2, time.Hour)
+	app := warmHost(t, 2, 5, time.Hour)
-
+	if peak := ask(t, app, 2); peak != 2 {
-	if n := app.Evict(5); n != 0 {
+		t.Fatalf("peak %d, want 2", peak)
-		t.Fatalf("Evict(5) stopped %d with only 2 running", n)
 	}
-	if got := len(runningPIDs(t, app)); got != 2 {
+	if n := app.Evict(); n != 0 {
-		t.Fatalf("%d running, want 2", got)
+		t.Fatalf("Evict stopped %d with 2 running under a ceiling of 5", n)
 	}
 }
 
-// TestWarm_ExemptPluginsCountButAreNotStopped is the honest half of the bound.
+// TestWarm_ExemptPluginsCountButAreNotStopped is the honest half. IdleAfter zero
-// IdleAfter zero means never-evict, and such a plugin holds memory like any
+// means never-evict, and such a plugin holds memory like any other — so it must
-// other — so it must count against the budget (or the budget is authorised
+// COUNT against the ceiling (or the budget is authorised twice) and must still not
-// twice) and must still not be stopped (or "never" meant nothing).
+// be stopped (or "never" meant nothing). When the exempt alone fill the ceiling the
+// host goes OVER it rather than shedding the identity service every other call
+// goes through.
 func TestWarm_ExemptPluginsCountButAreNotStopped(t *testing.T) {
 	bin := buildPlugin(t, "v1")
-	app := zip.New(zip.Config{AppName: "host", DisableStartupMessage: true})
+	app := zip.New(zip.Config{AppName: "host", DisableStartupMessage: true, Warm: 2})
-	// Two exempt (IdleAfter unset) and two evictable.
 	for _, p := range []zip.Plugin{
-		{Name: "keep0", Bin: bin, Dir: sockDir(t), Lazy: true},
+		{Name: "keep0", Bin: bin, Dir: sockDir(t), Lazy: true}, // IdleAfter unset: exempt
 		{Name: "keep1", Bin: bin, Dir: sockDir(t), Lazy: true},
 		{Name: "drop0", Bin: bin, Dir: sockDir(t), Lazy: true, IdleAfter: time.Hour},
 		{Name: "drop1", Bin: bin, Dir: sockDir(t), Lazy: true, IdleAfter: time.Hour},
@@ -124,6 +148,7 @@ func TestWarm_ExemptPluginsCountButAreNotStopped(t *testing.T) {
 		app.Use(must(zip.Load(p, "/v1/"+p.Name)))
 	}
 	defer func() { _ = app.Shutdown() }()
+
 	for _, n := range []string{"keep0", "keep1", "drop0", "drop1"} {
 		if status, body := call(t, app, "GET", "/v1/"+n+"/version", ""); status != 200 {
 			t.Fatalf("%s did not serve: status=%d body=%q", n, status, body)
@@ -131,79 +156,45 @@ func TestWarm_ExemptPluginsCountButAreNotStopped(t *testing.T) {
 		time.Sleep(2 * time.Millisecond)
 	}
 
-	// Four running, bound of two: the only two it MAY stop are the evictable
-	// ones, so it stops both and stays over the bound rather than touching the
-	// exempt pair.
-	if n := app.Evict(2); n != 2 {
-		t.Fatalf("Evict(2) stopped %d, want 2 (both evictable)", n)
-	}
-	if !settles(3*time.Second, func() bool { return len(runningPIDs(t, app)) == 2 }) {
-		t.Fatalf("running=%d, want the 2 exempt", len(runningPIDs(t, app)))
-	}
 	for _, p := range app.Plugins() {
 		switch p.Name {
 		case "keep0", "keep1":
 			if !p.Running {
 				t.Errorf("%s was evicted, but IdleAfter unset means never", p.Name)
 			}
-		case "drop0", "drop1":
+		case "drop0":
 			if p.Running {
-				t.Errorf("%s survived a bound of 2 with 4 running", p.Name)
+				t.Error("drop0 survived: drop1's start should have taken the coldest evictable")
 			}
 		}
 	}
-
+	// Two exempt plus the one evictable that just served: over the ceiling of 2,
-	// And a bound BELOW the exempt count reclaims nothing further, rather than
+	// and correctly so.
-	// looping or stopping something it may not.
+	if live := len(runningPIDs(t, app)); live != 3 {
-	if n := app.Evict(1); n != 0 {
+		t.Fatalf("%d running, want 3 — two exempt over the ceiling plus the newest", live)
-		t.Fatalf("Evict(1) stopped %d when only exempt plugins remain", n)
 	}
-}
+	// A sweep does not shed the exempt pair either.
-
+	app.Evict()
-// TestWarm_EvictedRoutesStillAnswer: the bound must not take a route down. This
+	if !settles(2*time.Second, func() bool { return len(runningPIDs(t, app)) == 2 }) {
-// is what separates it from Unload, and it is the property a caller notices.
+		t.Fatalf("after a sweep: %d running, want the 2 exempt", len(runningPIDs(t, app)))
-func TestWarm_EvictedRoutesStillAnswer(t *testing.T) {
-	app := warmHost(t, 4, time.Hour)
-	before := map[string]int{}
-	for _, p := range app.Plugins() {
-		before[p.Name] = p.PID
-	}
-	if n := app.Evict(1); n != 3 {
-		t.Fatalf("Evict(1) stopped %d of 4, want 3", n)
-	}
-	if !settles(3*time.Second, func() bool { return len(runningPIDs(t, app)) == 1 }) {
-		t.Fatal("did not settle at the bound")
-	}
-
-	// demo0 was the coldest, so it was certainly evicted. Asking again must
-	// serve, from a DIFFERENT process.
-	status, body := call(t, app, "GET", "/v1/demo0/version", "")
-	if status != 200 {
-		t.Fatalf("demo0 after eviction: status=%d body=%q", status, body)
 	}
 	for _, p := range app.Plugins() {
-		if p.Name == "demo0" {
+		if (p.Name == "keep0" || p.Name == "keep1") && !p.Running {
-			if !p.Running || p.PID == 0 {
+			t.Errorf("%s was evicted by the sweep, but it is exempt", p.Name)
-				t.Fatalf("demo0 did not restart on demand: %+v", p)
-			}
-			if p.PID == before["demo0"] {
-				t.Fatalf("same pid %d — nothing was actually stopped", p.PID)
-			}
 		}
 	}
 }
 
-// TestWarm_BothBoundsInOnePass: age and count are one sweep, and the count is
+// TestWarm_AgeStillReclaimsUnderTheCeiling: the two bounds are one sweep, and age
-// applied to what age LEFT. Without that ordering the returned figure and the
+// reclaims plugins the ceiling had no reason to touch.
-// live set disagree.
+func TestWarm_AgeStillReclaimsUnderTheCeiling(t *testing.T) {
-func TestWarm_BothBoundsInOnePass(t *testing.T) {
+	const plugins, warm = 3, 5 // ceiling never reached
-	const plugins, warm = 5, 3
+	app := warmHost(t, plugins, warm, 40*time.Millisecond)
-	app := warmHost(t, plugins, 40*time.Millisecond)
+	ask(t, app, plugins)
 
 	time.Sleep(80 * time.Millisecond) // every one is now past IdleAfter
-	if n := app.Evict(warm); n != plugins {
+	if n := app.Evict(); n != plugins {
-		t.Fatalf("Evict(%d) stopped %d, want all %d — age runs first and takes them all",
+		t.Fatalf("Evict stopped %d, want all %d on age alone", n, plugins)
-			warm, n, plugins)
 	}
 	if !settles(3*time.Second, func() bool { return len(runningPIDs(t, app)) == 0 }) {
 		t.Fatalf("%d still running", len(runningPIDs(t, app)))

zip.go

@@ -103,6 +103,16 @@ type Config struct {
 	// DisableStartupMessage suppresses Fiber's startup banner.
 	DisableStartupMessage bool
 
+	// Warm is how many plugin processes this app may hold at once. Zero means
+	// unbounded, the historical behaviour. It is enforced where a plugin STARTS,
+	// so it is a ceiling rather than something a sweep restores a minute later,
+	// and again by [App.Evict] alongside each plugin's IdleAfter.
+	//
+	// The host states it because the number is a fact about the host's memory: the
+	// processes are its own children in its own cgroup, and a library cannot know
+	// how much of it they may have.
+	Warm int
+
 	// ErrorHandler is the catch-all error handler. Defaults to zip.errorHandler
 	// which renders {error, code, status} JSON.
 	ErrorHandler fiber.ErrorHandler
@@ -258,6 +268,12 @@ type App struct {
 	open   *plugin
 	plugMu sync.Mutex
 
+	// warm is Config.Warm: how many plugin processes may run at once. It is state
+	// rather than an argument to Evict because the ceiling must be known where a
+	// process is STARTED — a bound only a periodic sweep consults is a bound a
+	// burst outruns. See makeRoom.
+	warm int
+
 	// mcpList is the whole tools/list array, rendered by installMCP and re-rendered
 	// by any later load: own ops ++ every plugin catalogue, sorted by name. Serving
 	// bytes is what makes the most-called MCP method free. Atomic because a Load
@@ -343,6 +359,7 @@ func newApp(cfg Config) *App {
 		logger:    cfg.Logger,
 		loader:    cfg.Loader,
 		born:      time.Now(),
+		warm:      cfg.Warm,
 		telemetry: newTelemetry(cfg),
 	}
 }