fix(ws-server): close self-fire restart feedback loop (internal#544)

Three-layer cohesive fix for the 2026-05-19 ~00:05-00:09Z 4x reprov thrash
class observed on prod-Reviewer + prod-Researcher: a single secrets PUT
fanned out into 4x stop+provision cycles per workspace within 4 min,
each stopping the just-launched (still-pending) EC2 of the previous
cycle. Root-caused via Loki (provision.ec2_started / ec2_stopped pairs).

Empirical chain (all in workspace-server/internal/handlers/):
1. secrets.go SetSecret → go h.restartFunc → coalesceRestart cycle.
2. runRestartCycle sets url='' synchronously, then async provisions EC2.
3. During 20-30s pending window: url='' AND cpProv.IsRunning()==false
   — indistinguishable from a dead container.
4. Canvas /delegations poll OR the trailing restart-context probe fires
   ProxyA2A → maybeMarkContainerDead OR preflightContainerHealth →
   RestartByID → loop.
5. coalesceRestart's pending flag drains by running ANOTHER full cycle
   → ec2_stopped of the just-booted instance → re-provision.

Fix (single PR, three interdependent layers):

L1) Restart-aware health probes — workspace_restart.go exposes
    isRestarting(workspaceID) bool. Both maybeMarkContainerDead and
    preflightContainerHealth early-return false/nil while a restart
    cycle is in flight. Breaks the self-fire at the probe layer.

L2) Restart-context probe gate — sendRestartContext now requires
    url != '' AND last_heartbeat_at > restart_start_ts before firing
    the trailing ProxyA2A probe. Adds waitForFreshHeartbeat() next to
    waitForWorkspaceOnline. Belt-and-suspenders so the probe never
    tries until the new container is actually addressable.

L3) RestartByID debounce — silent-drop successive RestartByID calls
    within restartDebounceWindow=60s of restartStartedAt. Not coalesce
    (which would still drain to another full cycle). Drop is observable
    via restartByIDDropCounter (atomic.Uint64) + the dropped log line.
    Only programmatic path; HTTP Restart handler is unaffected.

Tests:
- TestIsRestarting_{FalseWhenNoStateEntry,TrueWhileCycleRunning}
- TestMaybeMarkContainerDead_SkippedWhileRestarting (L1)
- TestPreflightContainerHealth_SkippedWhileRestarting (L1)
- TestRestartByID_DebounceSilentDrop (L3, counter assertion)
- TestRestartByID_DebounceExpiresAfterWindow (L3, window release)
- TestRestartByID_SingleProvisionPerRestart (regression — asserts
  exactly 1 cycle per trigger, with 4 dropped self-fire probes)

Existing coalesce/restart/preflight/maybeMarkContainerDead tests
remain green. Full handlers suite: ok in 15.8s.

Closes internal#544.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

workspace-server/internal/handlers/a2a_proxy_helpers.go

@@ -168,6 +168,21 @@ func (h *WorkspaceHandler) maybeMarkContainerDead(ctx context.Context, workspace
 	if !h.HasProvisioner() {
 		return false
 	}
+	// Restart-aware short-circuit: during the 20-30s EC2-pending window of
+	// an in-flight restart, the workspace's url='' and IsRunning() returns
+	// false → looks indistinguishable from a dead container. Pre-fix this
+	// fired a fresh RestartByID for the just-launched instance, which
+	// coalesceRestart's pending-flag drained by running ANOTHER full
+	// stop+provision cycle (= ec2_stopped of the still-pending instance
+	// → re-provision). That's the 4x reprov thrash class. Skip the
+	// container-dead path while a restart is in flight; the in-flight
+	// restart's own provisionWorkspaceAutoSync will surface a real failure
+	// (markProvisionFailed) if the new container never comes up. Issue
+	// internal#544.
+	if isRestarting(workspaceID) {
+		log.Printf("ProxyA2A: maybeMarkContainerDead skipped for %s — restart already in flight (self-fire guard)", workspaceID)
+		return false
+	}
 
 	var running bool
 	var inspectErr error
@@ -223,6 +238,18 @@ func (h *WorkspaceHandler) maybeMarkContainerDead(ctx context.Context, workspace
 // shape post-EC2-replace (see molecule-controlplane#20 incident
 // 2026-05-07) where the reconciler hasn't respawned the agent yet.
 func (h *WorkspaceHandler) preflightContainerHealth(ctx context.Context, workspaceID string) *proxyA2AError {
+	// Restart-aware short-circuit (mirror of maybeMarkContainerDead): if a
+	// restart cycle is in flight for this workspace, do not run the
+	// IsRunning probe — it would observe the EC2-pending state as "not
+	// running" and trigger RestartByID for an already-restarting workspace,
+	// closing the self-fire loop. Returning nil lets the optimistic
+	// forward proceed; the upstream Do() call will fail with a connection
+	// error or 502, and the *post-restart* reactive path can decide what
+	// to do once the cycle has actually completed. Issue internal#544.
+	if isRestarting(workspaceID) {
+		log.Printf("ProxyA2A preflight: %s — skipped, restart already in flight (self-fire guard)", workspaceID)
+		return nil
+	}
 	running, err := h.provisioner.IsRunning(ctx, workspaceID)
 	if err != nil {
 		// Transient daemon error. Provisioner.IsRunning returns (true, err)

workspace-server/internal/handlers/restart_context.go

@@ -180,6 +180,42 @@ func waitForWorkspaceOnline(ctx context.Context, workspaceID string, timeout tim
 	return false
 }
 
+// waitForFreshHeartbeat polls until the workspace has BOTH a non-empty
+// url AND a last_heartbeat_at strictly after restartStartTs (i.e. the
+// heartbeat we observe is NEW, not the stale pre-restart one carried
+// across through the row update). Returns false on timeout or DB error.
+//
+// This is the Layer 2 gate for the 2026-05-19 ws-server self-fire restart
+// loop fix. status='online' can flip while url='' is still in place (the
+// status update happens in /registry/register; url is set at the same
+// time but the read here may see a transient interleaving) and pre-fix
+// the trailing restart-context probe could fire against a half-registered
+// row, triggering the upstream-502 → maybeMarkContainerDead → self-fire
+// chain we're closing. The url + heartbeat-freshness check is the
+// strict, correlated end-state assertion that says "the new container is
+// actually addressable" — not just "some heartbeat happened".
+func waitForFreshHeartbeat(ctx context.Context, workspaceID string, restartStartTs time.Time, timeout time.Duration) bool {
+	deadline := time.Now().Add(timeout)
+	for time.Now().Before(deadline) {
+		var url sql.NullString
+		var lastHB sql.NullTime
+		err := db.DB.QueryRowContext(ctx,
+			`SELECT url, last_heartbeat_at FROM workspaces WHERE id = $1`, workspaceID,
+		).Scan(&url, &lastHB)
+		if err == nil &&
+			url.Valid && url.String != "" &&
+			lastHB.Valid && lastHB.Time.After(restartStartTs) {
+			return true
+		}
+		select {
+		case <-ctx.Done():
+			return false
+		case <-time.After(restartContextOnlinePollInterval):
+		}
+	}
+	return false
+}
+
 // buildRestartA2APayload wraps the rendered context string in the
 // JSON-RPC 2.0 / A2A message/send shape that the proxy already knows
 // how to normalize. Returns the marshalled body ready for ProxyA2ARequest.
@@ -220,6 +256,22 @@ func (h *WorkspaceHandler) sendRestartContext(workspaceID string, data restartCo
 		log.Printf("restart-context: workspace %s did not come online within %s — dropping context message", workspaceID, restartContextOnlineTimeout)
 		return
 	}
+	// Self-fire guard (Layer 2 of the 2026-05-19 ws-server self-fire fix):
+	// status='online' alone is not enough to safely fire the trailing
+	// ProxyA2ARequest. The workspace must also have:
+	//   - url != ''                            (the new container's URL has been registered)
+	//   - last_heartbeat_at > data.RestartAt   (the heartbeat we're seeing is NEW, not stale)
+	// Without those, ProxyA2ARequest can fail with a connect error or
+	// upstream 502, hit handleA2ADispatchError → maybeMarkContainerDead →
+	// RestartByID → self-fire. The Layer 1 isRestarting gate already
+	// covers that, but this is a belt-and-suspenders so the probe never
+	// even tries until the new container is actually addressable. Best-
+	// effort: if the DB read errors out we proceed (preserves the legacy
+	// behaviour of "online means online").
+	if !waitForFreshHeartbeat(ctx, workspaceID, data.RestartAt, restartContextOnlineTimeout) {
+		log.Printf("restart-context: workspace %s online but no fresh heartbeat or empty url — dropping context message (self-fire guard)", workspaceID)
+		return
+	}
 
 	text := buildRestartContextMessage(data)
 	body, err := buildRestartA2APayload(text)

workspace-server/internal/handlers/workspace_restart.go

@@ -8,6 +8,7 @@ import (
 	"runtime/debug"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"github.com/Molecule-AI/molecule-monorepo/platform/internal/db"
@@ -39,12 +40,57 @@ type restartState struct {
 	mu      sync.Mutex
 	running bool // true while a restart cycle is in flight
 	pending bool // set by any caller that arrived during the in-flight cycle
+	// restartStartedAt records the wall-clock when the most recent cycle
+	// flipped running=true. Used by the self-fire debounce (internal#544,
+	// the ws-server self-fire restart feedback loop seen in prod-Reviewer/
+	// Researcher 2026-05-19 ~00:05Z 4x reprov thrash): any RestartByID
+	// arriving within restartDebounceWindow of this timestamp is silently
+	// dropped so a probe firing during the EC2-pending window can't
+	// re-trigger a fresh full cycle on the just-launched instance.
+	restartStartedAt time.Time
 }
 
 // restartStates is a per-workspace map of *restartState. Each workspace gets
 // its own entry so unrelated workspaces don't serialize on each other.
 var restartStates sync.Map // map[workspaceID]*restartState
 
+// restartDebounceWindow is the silent-drop window for successive RestartByID
+// calls. Sized to cover the typical EC2 pending → online interval (20-30s)
+// with a margin so a probe firing during the just-after-online but still-
+// flaky heartbeat window also gets dropped. Bigger than that would block
+// legitimate "Restart failed, retry" recoveries; smaller would let the
+// 4x thrash class through. Package-level so tests can shrink it.
+var restartDebounceWindow = 60 * time.Second
+
+// restartByIDDropCounter is incremented every time RestartByID drops a call
+// inside the debounce window. Exposed as a package-level atomic counter so
+// (a) tests can assert the drop fired, (b) ops can grep logs for the drop
+// log line + the counter snapshot in a future /admin/metrics endpoint.
+// Not a Prometheus metric because the platform doesn't pull metrics from
+// workspace-server yet — that's a separate RFC.
+var restartByIDDropCounter atomic.Uint64
+
+// isRestarting reports whether a restart cycle is currently in flight for
+// the workspace. Callers that have their own "container looks dead" probe
+// MUST consult this before triggering a restart, because during the
+// 20-30s EC2-pending window the workspace's url='' and IsRunning()=false
+// looks identical to a dead container — and any restart-triggering probe
+// (maybeMarkContainerDead from canvas /delegations poll, or the trailing
+// restart-context probe at the end of runRestartCycle) will set
+// pending=true and the outer coalesceRestart loop will drain by running
+// ANOTHER full cycle, ec2_stopped of the just-booted instance →
+// re-provision. That's the self-fire loop closed by this gate.
+func isRestarting(workspaceID string) bool {
+	sv, ok := restartStates.Load(workspaceID)
+	if !ok {
+		return false
+	}
+	state := sv.(*restartState)
+	state.mu.Lock()
+	defer state.mu.Unlock()
+	return state.running
+}
+
 // isParentPaused checks if any ancestor of the workspace is paused.
 func isParentPaused(ctx context.Context, workspaceID string) (bool, string) {
 	var parentID *string
@@ -376,9 +422,45 @@ func (h *WorkspaceHandler) RestartByID(workspaceID string) {
 	if !h.HasProvisioner() {
 		return
 	}
+	// Self-fire debounce: drop (not coalesce) successive RestartByID calls
+	// within restartDebounceWindow of the most recent cycle's start. This
+	// is the load-bearing protection against the 4x reprov thrash class —
+	// coalesceRestart's pending-flag would otherwise drain by running
+	// ANOTHER full cycle of stop+provision on the just-launched EC2 (still
+	// in the pending state), which is the self-fire we're closing.
+	//
+	// Only applies to RestartByID (programmatic — secrets handler,
+	// maybeMarkContainerDead, preflightContainerHealth). The HTTP Restart
+	// handler in workspace_restart.go's Restart() bypasses this path and
+	// calls RestartWorkspaceAutoOpts directly, so user-initiated restart
+	// clicks are unaffected.
+	if shouldDebounceRestart(workspaceID) {
+		restartByIDDropCounter.Add(1)
+		log.Printf("RestartByID: %s — dropped (within %s self-fire debounce window; total dropped=%d)",
+			workspaceID, restartDebounceWindow, restartByIDDropCounter.Load())
+		return
+	}
 	coalesceRestart(workspaceID, func() { h.runRestartCycle(workspaceID) })
 }
 
+// shouldDebounceRestart reports whether the most recent cycle for this
+// workspace started within restartDebounceWindow. Read-only on
+// restartState; the actual restartStartedAt stamp is written in
+// coalesceRestart when running flips false→true.
+func shouldDebounceRestart(workspaceID string) bool {
+	sv, ok := restartStates.Load(workspaceID)
+	if !ok {
+		return false
+	}
+	state := sv.(*restartState)
+	state.mu.Lock()
+	defer state.mu.Unlock()
+	if state.restartStartedAt.IsZero() {
+		return false
+	}
+	return time.Since(state.restartStartedAt) < restartDebounceWindow
+}
+
 // coalesceRestart implements the pending-flag gate around an arbitrary cycle
 // function. Extracted from RestartByID for direct unit testing — the cycle
 // function in production is `runRestartCycle`, but tests pass a counter to
@@ -398,6 +480,12 @@ func coalesceRestart(workspaceID string, cycle func()) {
 		return
 	}
 	state.running = true
+	// Stamp the start time so the RestartByID debounce can drop any
+	// self-fire probe that hits within restartDebounceWindow. Only the
+	// false→true edge stamps; the drain-loop's inner cycles re-use the
+	// same start (they're effectively one "restart event" from the
+	// debounce's POV).
+	state.restartStartedAt = time.Now()
 	state.mu.Unlock()
 
 	// Always clear running on exit — including panic — so a panicking

workspace-server/internal/handlers/workspace_restart_self_fire_test.go

@@ -0,0 +1,297 @@
+package handlers
+
+// Tests for the 2026-05-19 ws-server self-fire restart feedback loop fix.
+//
+// Empirical chain reproduced (prod-Reviewer/Researcher 4x reprov thrash
+// 2026-05-19 ~00:05-00:09Z, root-caused via Loki):
+//
+//  1. POST /secrets → go h.restartFunc(workspaceID) (secrets.go:264).
+//  2. runRestartCycle sets url='' synchronously, then async provisions EC2
+//     (workspace_restart.go).
+//  3. During 20-30s window while EC2 is `pending` (codex first heartbeat
+//     not yet landed): workspaces.url='' AND IsRunning=false.
+//  4. Any ProxyA2A (canvas /delegations poll OR the restart-context probe
+//     at the end of runRestartCycle) → maybeMarkContainerDead sees the
+//     container-dead state → calls RestartByID → loop.
+//  5. coalesceRestart sets pending=true, drains by running ANOTHER full
+//     cycle → provision.ec2_stopped of the just-booted instance →
+//     re-provision.
+//
+// Fix: three interdependent layers.
+//
+//  L1) isRestarting() gate in maybeMarkContainerDead +
+//      preflightContainerHealth — early-return false/nil so the probe
+//      can't trigger a fresh RestartByID while a restart is in flight.
+//  L2) sendRestartContext requires url != '' AND last_heartbeat_at >
+//      restart_start_ts before firing the trailing ProxyA2A probe.
+//  L3) RestartByID silently drops successive calls within
+//      restartDebounceWindow of restartStartedAt, with a counter for
+//      observability.
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/DATA-DOG/go-sqlmock"
+)
+
+// resetSelfFireState wipes all the per-workspace mutation state these
+// tests touch, plus the package-level drop counter, so the test is
+// hermetic regardless of ordering.
+func resetSelfFireState(workspaceID string) {
+	restartStates.Delete(workspaceID)
+	restartByIDDropCounter.Store(0)
+}
+
+// markRestarting forces restartStates into "cycle in flight" without
+// running an actual cycle, so the tests can isolate the gate behaviour
+// without the full provision pipeline. Returns a finish() that flips
+// running=false (mimicking coalesceRestart's deferred state-clear).
+func markRestarting(workspaceID string) (finish func()) {
+	sv, _ := restartStates.LoadOrStore(workspaceID, &restartState{})
+	state := sv.(*restartState)
+	state.mu.Lock()
+	state.running = true
+	state.restartStartedAt = time.Now()
+	state.mu.Unlock()
+	return func() {
+		state.mu.Lock()
+		state.running = false
+		state.mu.Unlock()
+	}
+}
+
+// TestIsRestarting_FalseWhenNoStateEntry — baseline: a workspace that
+// has never been restarted reports !isRestarting. Pinning this so a
+// future LoadOrStore refactor can't silently start returning true for
+// unknown workspaces.
+func TestIsRestarting_FalseWhenNoStateEntry(t *testing.T) {
+	const wsID = "self-fire-ws-never"
+	resetSelfFireState(wsID)
+	if isRestarting(wsID) {
+		t.Fatal("isRestarting must return false for a workspace with no state entry")
+	}
+}
+
+// TestIsRestarting_TrueWhileCycleRunning — the load-bearing invariant
+// that Layer 1 depends on. While running=true, isRestarting must report
+// true; the moment it flips to false, isRestarting must report false.
+func TestIsRestarting_TrueWhileCycleRunning(t *testing.T) {
+	const wsID = "self-fire-ws-in-flight"
+	resetSelfFireState(wsID)
+
+	finish := markRestarting(wsID)
+	if !isRestarting(wsID) {
+		t.Fatal("isRestarting must return true while running=true")
+	}
+	finish()
+	if isRestarting(wsID) {
+		t.Fatal("isRestarting must return false after running flips back to false")
+	}
+}
+
+// TestMaybeMarkContainerDead_SkippedWhileRestarting — Layer 1 for the
+// reactive path. With isRestarting=true the function must early-return
+// false WITHOUT invoking IsRunning, hitting the DB UPDATE, or kicking
+// a RestartByID goroutine. If any of those side-effects fire we'd
+// re-arm the self-fire loop the gate exists to close.
+func TestMaybeMarkContainerDead_SkippedWhileRestarting(t *testing.T) {
+	const wsID = "self-fire-ws-mmcd"
+	resetSelfFireState(wsID)
+	mock := setupTestDB(t) // sqlmock with strict expectation matching
+
+	// Workspace row read inside maybeMarkContainerDead — this happens
+	// BEFORE the isRestarting gate in the current implementation, so
+	// allow exactly one SELECT runtime row.
+	mock.ExpectQuery(`SELECT COALESCE\(runtime, 'langgraph'\) FROM workspaces WHERE id =`).
+		WithArgs(wsID).
+		WillReturnRows(sqlmock.NewRows([]string{"runtime"}).AddRow("claude-code"))
+
+	// Gate flipped: must early-return without doing anything else.
+	finish := markRestarting(wsID)
+	defer finish()
+
+	stub := &preflightLocalProv{running: false, err: nil}
+	h := newSelfFireHandler(t)
+	h.provisioner = stub
+
+	if got := h.maybeMarkContainerDead(context.Background(), wsID); got != false {
+		t.Errorf("maybeMarkContainerDead must return false while restarting, got %v", got)
+	}
+	if stub.calls != 0 {
+		t.Errorf("IsRunning must not be called while restarting (Layer 1 gate broken); got %d calls", stub.calls)
+	}
+}
+
+// TestPreflightContainerHealth_SkippedWhileRestarting — Layer 1 for the
+// proactive path. Same shape as above: with restart in flight, return
+// nil (let the optimistic forward proceed) and DO NOT call IsRunning.
+// The forward will fail with a connect error; the post-restart reactive
+// path can decide what to do then, by which point the EC2 has either
+// come up (no more failures) or markProvisionFailed has fired.
+func TestPreflightContainerHealth_SkippedWhileRestarting(t *testing.T) {
+	const wsID = "self-fire-ws-preflight"
+	resetSelfFireState(wsID)
+	_ = setupTestDB(t)
+
+	finish := markRestarting(wsID)
+	defer finish()
+
+	stub := &preflightLocalProv{running: false, err: nil}
+	h := newSelfFireHandler(t)
+	h.provisioner = stub
+
+	if err := h.preflightContainerHealth(context.Background(), wsID); err != nil {
+		t.Errorf("preflightContainerHealth must return nil while restarting, got %+v", err)
+	}
+	if stub.calls != 0 {
+		t.Errorf("IsRunning must not be called while restarting (Layer 1 gate broken); got %d calls", stub.calls)
+	}
+}
+
+// TestRestartByID_DebounceSilentDrop — Layer 3. After a cycle starts,
+// any RestartByID arriving within restartDebounceWindow MUST be dropped
+// silently — not coalesced (which would still drain to another cycle).
+// The drop counter must increment by exactly one per dropped call so
+// ops can see how often the self-fire would have fired pre-fix.
+func TestRestartByID_DebounceSilentDrop(t *testing.T) {
+	const wsID = "self-fire-ws-debounce"
+	resetSelfFireState(wsID)
+
+	// Stamp restartStartedAt = now, running=false (simulates the "just
+	// finished" window where the loop would re-fire pre-fix).
+	sv, _ := restartStates.LoadOrStore(wsID, &restartState{})
+	state := sv.(*restartState)
+	state.mu.Lock()
+	state.restartStartedAt = time.Now()
+	state.running = false
+	state.mu.Unlock()
+
+	// Counter baseline.
+	if got := restartByIDDropCounter.Load(); got != 0 {
+		t.Fatalf("expected drop counter 0 at start, got %d", got)
+	}
+
+	// Five rapid-fire RestartByID calls should all drop (the maximum
+	// observed pre-fix was 4x — pinning >=4 here keeps the regression
+	// shape true to the prod incident).
+	h := newSelfFireHandler(t)
+	stub := &preflightLocalProv{running: true, err: nil}
+	h.provisioner = stub
+	for i := 0; i < 5; i++ {
+		h.RestartByID(wsID)
+	}
+
+	if got := restartByIDDropCounter.Load(); got != 5 {
+		t.Errorf("expected 5 drops within debounce window, got %d", got)
+	}
+
+	// shouldDebounceRestart itself must report true for the same window.
+	if !shouldDebounceRestart(wsID) {
+		t.Error("shouldDebounceRestart must return true within window")
+	}
+}
+
+// TestRestartByID_DebounceExpiresAfterWindow — outside the window, the
+// debounce must release: a legitimate later restart (e.g. user clicked
+// Restart again after waiting) must proceed to coalesceRestart. We
+// shrink restartDebounceWindow to 1ms for the duration of this test so
+// we don't sleep a full 60s in CI.
+func TestRestartByID_DebounceExpiresAfterWindow(t *testing.T) {
+	const wsID = "self-fire-ws-debounce-release"
+	resetSelfFireState(wsID)
+
+	orig := restartDebounceWindow
+	restartDebounceWindow = 5 * time.Millisecond
+	defer func() { restartDebounceWindow = orig }()
+
+	// Stamp inside the window.
+	sv, _ := restartStates.LoadOrStore(wsID, &restartState{})
+	state := sv.(*restartState)
+	state.mu.Lock()
+	state.restartStartedAt = time.Now()
+	state.running = false
+	state.mu.Unlock()
+
+	if !shouldDebounceRestart(wsID) {
+		t.Fatal("within 5ms window must debounce")
+	}
+
+	// Sleep past the window. Use a small margin to avoid clock-skew
+	// flakes on slow CI hosts.
+	time.Sleep(20 * time.Millisecond)
+
+	if shouldDebounceRestart(wsID) {
+		t.Fatal("after 20ms (4x window) must no longer debounce")
+	}
+}
+
+// TestRestartByID_SingleProvisionPerRestart — the regression test for
+// the prod incident: a SINGLE secrets PUT (which is the trigger shape)
+// must produce exactly ONE coalesceRestart cycle, not four. Models the
+// full chain: secrets handler → RestartByID → coalesceRestart → cycle
+// runs → during the cycle window, simulated probes call RestartByID
+// again. With all three layers in place, the probes are dropped and the
+// total cycle count stays at 1.
+func TestRestartByID_SingleProvisionPerRestart(t *testing.T) {
+	const wsID = "self-fire-ws-single-provision"
+	resetSelfFireState(wsID)
+
+	// In-flight gate that mimics the EC2-pending window. The cycle
+	// blocks on cycleProceed so we can fire the simulated probes while
+	// running=true.
+	var cycleCount atomic.Int32
+	cycleStarted := make(chan struct{}, 1)
+	cycleProceed := make(chan struct{})
+
+	cycle := func() {
+		n := cycleCount.Add(1)
+		if n == 1 {
+			cycleStarted <- struct{}{}
+			<-cycleProceed
+		}
+	}
+
+	// Kick the first cycle via coalesceRestart (this is what RestartByID
+	// would do post-debounce-check).
+	done := make(chan struct{})
+	go func() {
+		coalesceRestart(wsID, cycle)
+		close(done)
+	}()
+	<-cycleStarted
+
+	// Simulate the 4 probe-driven RestartByID calls observed in prod.
+	// Each must drop because we're within the debounce window AND a
+	// cycle is in flight.
+	h := newSelfFireHandler(t)
+	stub := &preflightLocalProv{running: true, err: nil}
+	h.provisioner = stub
+	for i := 0; i < 4; i++ {
+		h.RestartByID(wsID)
+	}
+
+	// Release the cycle.
+	close(cycleProceed)
+	<-done
+
+	if got := cycleCount.Load(); got != 1 {
+		t.Errorf("expected exactly 1 provision cycle for a single trigger "+
+			"(self-fire fix), got %d — regression of the prod 4x reprov thrash class",
+			got)
+	}
+	if got := restartByIDDropCounter.Load(); got != 4 {
+		t.Errorf("expected 4 self-fire probes dropped, got %d "+
+			"(observability counter must record the saved cycles)", got)
+	}
+}
+
+// newSelfFireHandler constructs a minimal *WorkspaceHandler suitable for
+// the Layer-1 gate tests. Wraps the boilerplate so the per-test setup
+// stays focused on the assertion.
+func newSelfFireHandler(t *testing.T) *WorkspaceHandler {
+	t.Helper()
+	return NewWorkspaceHandler(newTestBroadcaster(), nil, "http://localhost:8080", t.TempDir())
+}