fix: terminate telemetry work on disposal

Give the best-effort telemetry adapter a terminal ACTIVE/DISPOSED lifecycle.
dispose() now removes the pagehide listener, clears the queue, invalidates
scheduled callback generations and aborts the in-flight sink; emit after
dispose is a no-op and a sink that ignores the abort cannot reschedule or
update post-dispose state. flush() joins the active delivery instead of
resolving early, and runtime infrastructure teardown disposes telemetry first.

Telemetry and diagnostics capacities are validated at construction against a
documented ceiling, so NaN or Infinity can no longer disable eviction.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs/architecture/decisions/VD-07-diagnostics-and-telemetry-exporter.md

@@ -58,6 +58,20 @@ vendor 결정 전에는 안전한 기본값이 아니다.
 7. queue full, invalid event/context, serialization과 sink failure는 제한된
    reason bucket으로 집계한다. drop observer의 failure는 다시 telemetry를
    발행하지 않는 nonrecursive 경계다.
+7-1. telemetry adapter lifecycle은 `ACTIVE | DISPOSED` 둘뿐이다. `dispose()`는
+   한 번만 전이하고 `pagehide` listener 제거, queue 비우기, scheduled callback
+   generation 무효화, in-flight sink `AbortController` abort를 모두 수행한다.
+   dispose 뒤 `emit()`은 no-op이고 새 flush는 스케줄되지 않으며, abort를 무시한
+   sink가 늦게 settle해도 post-dispose delivery state를 갱신하거나 재스케줄하지
+   못한다. 종료 중 drop telemetry를 재귀적으로 발행하지 않는다.
+7-2. `flush()`는 active delivery promise를 join한다. 이미 진행 중인 flush가
+   있으면 같은 promise를 반환하므로 `await flush()`는 실제 settle을 뜻한다.
+7-3. runtime `infrastructure.dispose()`는 diagnostics/state dependency를 파괴하기
+   전에 `telemetry.dispose()`를 먼저 호출한다.
+7-4. queue/entry capacity는 construction-time 계약이다. `Number.isSafeInteger`가
+   아니거나 1 미만이거나 문서화된 ceiling(각각 `MAX_TELEMETRY_QUEUE`,
+   `MAX_DIAGNOSTIC_ENTRIES` = 10,000)을 넘으면 `TypeError`로 거절한다. NaN/Infinity가
+   조용히 eviction을 비활성화하는 경로를 남기지 않는다.
 8. diagnostics와 telemetry failure는 제품 흐름, HTTP 결과, route transition,
    storage/cache fallback과 React error surface를 바꾸지 않는다.
 9. mount 전 bootstrap failure는 안전한 build/config/error kind만 별도 evidence로

docs/operations/adapter-remediation-ledger.md

@@ -75,14 +75,14 @@ Rollout state starts at `NOT_STARTED`; documented-unimplemented items start at
 | N-01 | Live V3 path | `corepack pnpm exec vitest run tests/integration/http-execution-v3-observability.test.ts` | `fix: restore V3 HTTP observability` | `FIXED_NOT_RELEASED` | diagnostics/telemetry producer gate regression | Red 5/5 failed → green 5/5; `check:diagnostics` PASS (8 diagnostics, 5 telemetry producers); `check:types` PASS; `check:architecture` PASS |
 | N-02 | Live V3 path | `corepack pnpm exec vitest run tests/integration/http-execution-v3-auth-profile.test.ts` | `fix: enforce installed HTTP auth profiles` | `FIXED_NOT_RELEASED` | authenticated request 4xx spike after profile enforcement | Red suite failed to load (`installRestAuthProfileRegistry` absent) → green 7/7; `check:types` PASS; `check:architecture` PASS; `lint` PASS; unit+integration+features 1560 passed with only the pre-existing environmental `ci-artifact-contract` failures |
 | N-03 | Live V3 path | `corepack pnpm exec vitest run tests/unit/http-execution-v3.test.ts -t "retry-time fence"` | `fix: preserve command effect certainty across retries` | `FIXED_NOT_RELEASED` | command effect verdict regression | Red reproduced `SCOPE_FENCED` with `NOT_STARTED` after one dispatched attempt → green `MAYBE_APPLIED`; lattice table 9/9; `check:types` PASS; `lint` PASS |
-| N-04 | Live composition teardown | `corepack pnpm exec vitest run tests/unit/telemetry.test.ts` | — | `NOT_STARTED` | telemetry delivery loss after teardown change | — |
+| N-04 | Live composition teardown | `corepack pnpm exec vitest run tests/unit/telemetry.test.ts tests/unit/runtime-adapters.test.ts` | `fix: terminate telemetry work on disposal` | `FIXED_NOT_RELEASED` | telemetry delivery loss after teardown change | Red 10 failed (5 lifecycle + 5 capacity) → green 34/34; `check:diagnostics` PASS; `check:types` PASS; `check:architecture` PASS; `lint` PASS |
 | N-05 | Rollout blocker (sidecar not composed) | `corepack pnpm exec vitest run tests/unit/conditional-validator-store.test.ts` | — | `NOT_STARTED` | persisted validator key incompatibility | — |
 | N-06 | Legacy V2 rollback seam | `corepack pnpm exec vitest run tests/integration/http-client.test.ts` | — | `NOT_STARTED` | legacy keyed command rejection spike | — |
 | N-07 | Legacy V2 rollback seam | `corepack pnpm exec vitest run tests/integration/auth-recovery.test.ts` | — | `NOT_STARTED` | legacy credential timeout regression | — |
 | N-08 | Legacy V2 rollback seam | `corepack pnpm exec vitest run tests/unit/bounded-json-compatibility.test.ts` | — | `NOT_STARTED` | legacy JSON failure-code drift | — |
 | N-09 | Live cross-context host | `corepack pnpm exec vitest run tests/unit/cross-tab-invalidation.test.ts` | — | `NOT_STARTED` | cross-tab invalidation drop | — |
 | N-10 | Cursor runtime `AVAILABLE_NOT_COMPOSED` | `corepack pnpm exec vitest run tests/unit/cursor-pagination-runtime.test.ts` | — | `NOT_STARTED` | pagination abort semantics change | — |
-| N-11 | Live composition | `corepack pnpm exec vitest run tests/unit/diagnostics.test.ts` | — | `NOT_STARTED` | capacity rejection on valid composition | — |
+| N-11 | Live composition | `corepack pnpm exec vitest run tests/unit/telemetry.test.ts -t capacity` | `fix: terminate telemetry work on disposal` | `FIXED_NOT_RELEASED` | capacity rejection on valid composition | Red 5/5 capacity cases → green; ceilings documented in VD-07 §7-4 |
 
 ### Storage and browser files (`docs/reviews/adapters/03-storage-and-browser-files.md`)
 

src/adapters/diagnostics/bounded-diagnostics.ts

@@ -6,11 +6,15 @@ import {
   type DiagnosticRecordInput,
 } from "../../contracts/diagnostics.ts";
 import { projectTelemetryEvent } from "../../contracts/telemetry.ts";
+import { assertBoundedCapacity } from "../telemetry/best-effort-telemetry.ts";
 
 export const noOpDiagnostics: DiagnosticsPort = Object.freeze({
   record() {},
 });
 
+/** N-11. Documented absolute ceiling for bounded in-memory evidence. */
+export const MAX_DIAGNOSTIC_ENTRIES = 10_000;
+
 export function createDiagnosticsAdapter(
   options: Readonly<{
     maxEntries?: number;
@@ -18,7 +22,11 @@ export function createDiagnosticsAdapter(
     sink?: (record: DiagnosticRecord) => void;
   }> = {},
 ) {
-  const maxEntries = Math.max(1, options.maxEntries ?? 100);
+  const maxEntries = assertBoundedCapacity(
+    options.maxEntries ?? 100,
+    MAX_DIAGNOSTIC_ENTRIES,
+    "diagnostics maxEntries",
+  );
   const entries: DiagnosticRecord[] = [];
   const droppedReasons = new Map<string, number>();
 

src/adapters/telemetry/best-effort-telemetry.ts

@@ -37,6 +37,31 @@ export const noOpTelemetry: TelemetryAdapter = Object.freeze({
   dispose: () => {},
 });
 
+/**
+ * N-04. Disposal is terminal: there is no durable queue and no resurrection.
+ */
+type TelemetryLifecycle = "ACTIVE" | "DISPOSED";
+
+/** N-11. Documented absolute ceiling for the in-memory best-effort queue. */
+export const MAX_TELEMETRY_QUEUE = 10_000;
+
+/**
+ * N-11. A non-finite or fractional capacity silently disables eviction, so it is
+ * a construction-time configuration error rather than a runtime drop.
+ */
+export function assertBoundedCapacity(
+  value: number,
+  ceiling: number,
+  label: string,
+): number {
+  if (!Number.isSafeInteger(value) || value < 1 || value > ceiling) {
+    throw new TypeError(
+      `${label} must be a safe integer between 1 and ${ceiling}`,
+    );
+  }
+  return value;
+}
+
 export function createTelemetryAdapter(
   options: TelemetryAdapterOptions,
 ): TelemetryAdapter {
@@ -46,11 +71,19 @@ export function createTelemetryAdapter(
 
   const endpoint = options.endpoint;
   const fetcher = options.fetcher ?? fetch;
-  const maxQueue = Math.max(1, options.maxQueue ?? 100);
+  const maxQueue = assertBoundedCapacity(
+    options.maxQueue ?? 100,
+    MAX_TELEMETRY_QUEUE,
+    "telemetry maxQueue",
+  );
   const schedule = options.schedule ?? queueMicrotask;
   const queue: TelemetryEvent[] = [];
+  let lifecycleState: TelemetryLifecycle = "ACTIVE";
+  /** Scheduled callbacks captured before disposal must not run afterwards. */
+  let scheduleGeneration = 0;
   let scheduled = false;
-  let flushing = false;
+  let activeFlush: Promise<void> | null = null;
+  let activeSink: AbortController | null = null;
   let dropped = 0;
   const dropReasons = new Map<string, number>();
   let lastDeliveryEvidence: TelemetryEvent | null = null;
@@ -93,9 +126,12 @@ export function createTelemetryAdapter(
   }
 
   function scheduleFlush(): void {
-    if (scheduled) return;
+    if (scheduled || lifecycleState === "DISPOSED") return;
     scheduled = true;
+    const generation = scheduleGeneration;
     schedule(() => {
+      // A callback captured before disposal belongs to a dead generation.
+      if (generation !== scheduleGeneration) return;
       scheduled = false;
       void flush();
     });
@@ -105,6 +141,7 @@ export function createTelemetryAdapter(
     eventName: TelemetryEventName,
     attributes: Record<string, unknown>,
   ): void {
+    if (lifecycleState === "DISPOSED") return;
     const projected = projectTelemetryEvent(
       eventName,
       attributes,
@@ -124,26 +161,44 @@ export function createTelemetryAdapter(
     scheduleFlush();
   }
 
-  async function flush(): Promise<void> {
+  /**
-    if (flushing || queue.length === 0) return;
+   * `flush()` joins the active delivery instead of resolving immediately, so an
-    flushing = true;
+   * awaited flush really means "the in-flight batch has settled".
-    const batch = queue.splice(0, queue.length);
+   */
-    try {
+  function flush(): Promise<void> {
-      const response = await fetcher(endpoint, {
+    if (activeFlush) return activeFlush;
-        method: "POST",
+    if (lifecycleState === "DISPOSED" || queue.length === 0) {
-        headers: { "Content-Type": "application/json" },
+      return Promise.resolve();
-        body: JSON.stringify({ events: batch }),
-        keepalive: true,
-      });
-      if (!response.ok) recordDrop("sink-failure", batch.length);
-    } catch {
-      recordDrop("sink-failure", batch.length);
-    } finally {
-      flushing = false;
-      if (queue.length > 0) {
-        scheduleFlush();
-      }
     }
+    const generation = scheduleGeneration;
+    const run = async () => {
+      const batch = queue.splice(0, queue.length);
+      const controller = new AbortController();
+      activeSink = controller;
+      try {
+        const response = await fetcher(endpoint, {
+          method: "POST",
+          headers: { "Content-Type": "application/json" },
+          body: JSON.stringify({ events: batch }),
+          keepalive: true,
+          signal: controller.signal,
+        });
+        // A sink that ignored the abort must not update post-dispose state.
+        if (generation !== scheduleGeneration) return;
+        if (!response.ok) recordDrop("sink-failure", batch.length);
+      } catch {
+        if (generation !== scheduleGeneration) return;
+        recordDrop("sink-failure", batch.length);
+      } finally {
+        if (activeSink === controller) activeSink = null;
+        activeFlush = null;
+        if (generation === scheduleGeneration && queue.length > 0) {
+          scheduleFlush();
+        }
+      }
+    };
+    activeFlush = run();
+    return activeFlush;
   }
 
   const flushBeforePageExit = () => {
@@ -151,8 +206,20 @@ export function createTelemetryAdapter(
   };
   lifecycle?.addEventListener("pagehide", flushBeforePageExit);
 
+  /**
+   * N-04. Terminal disposal: one state transition, no further admission, no
+   * further scheduling, and no recursive drop telemetry while shutting down.
+   */
   function dispose(): void {
+    if (lifecycleState === "DISPOSED") return;
+    lifecycleState = "DISPOSED";
+    scheduleGeneration += 1;
+    scheduled = false;
     lifecycle?.removeEventListener("pagehide", flushBeforePageExit);
+    queue.length = 0;
+    activeSink?.abort();
+    activeSink = null;
+    activeFlush = null;
   }
 
   return Object.freeze({

src/bootstrap/runtime-adapters.ts

@@ -480,6 +480,10 @@ export async function createRuntimeAdapters(
       crossContextInvalidationStatus: () =>
         serverStateGeneration.getSnapshot().crossContextStatus(),
       dispose() {
+        // N-04. Telemetry is torn down first: it must stop scheduling and
+        // delivering before the diagnostics and state dependencies it observes
+        // are destroyed.
+        telemetry.dispose();
         conditionalValidators.clear();
         serverStateScope.dispose();
         serverStateGeneration.dispose();

tests/unit/runtime-adapters.test.ts

@@ -138,6 +138,38 @@ describe("runtime adapter composition", () => {
     adapters.infrastructure.dispose();
   });
 
+  it("runtime infrastructure disposal disposes telemetry first", async () => {
+    const fetcher = vi.fn(async () => new Response(null, { status: 204 }));
+    const adapters = await createRuntimeAdapters({
+      runtime: {
+        ...runtime,
+        config: {
+          ...runtime.config,
+          TELEMETRY_ENABLED: true,
+          TELEMETRY_ENDPOINT: "https://telemetry.test/events",
+        },
+      } as Runtime,
+      release,
+      host: {},
+      fetcher: fetcher as unknown as typeof fetch,
+    });
+
+    adapters.outputPorts.telemetry.emit("api.request.failed", {
+      error_kind: "SERVER_FAILURE",
+      http_status_group: "5xx",
+      attempt_count_bucket: "1",
+      route_id: "TEST_ROUTE",
+    });
+    expect(adapters.outputPorts.telemetry.pendingCount()).toBe(1);
+
+    adapters.infrastructure.dispose();
+
+    expect(adapters.outputPorts.telemetry.pendingCount()).toBe(0);
+    await Promise.resolve();
+    await Promise.resolve();
+    expect(fetcher).not.toHaveBeenCalled();
+  });
+
   it("replaces the QueryClient and coordinator for each session generation", async () => {
     const adapters = await createRuntimeAdapters({ runtime, release, host: {} });
     const previousClient = adapters.infrastructure.queryClient;

tests/unit/telemetry.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it, vi } from "vitest";
 
+import { createDiagnosticsAdapter } from "../../src/adapters/diagnostics/bounded-diagnostics.ts";
 import {
   createTelemetryAdapter,
   safeTraceparent,
@@ -236,4 +237,122 @@ describe("best-effort telemetry adapter", () => {
     expect(adapter.pendingCount()).toBe(0);
     expect(remove).toHaveBeenCalledWith("pagehide", expect.any(Function));
   });
+
+  it("drops queued events and invalidates scheduled callbacks on dispose", async () => {
+    const callbacks: Array<() => void> = [];
+    const fetcher = vi.fn(async () => new Response(null, { status: 204 }));
+    const adapter = createTelemetryAdapter({
+      enabled: true,
+      endpoint: "https://telemetry.test/events",
+      schedule: (callback) => callbacks.push(callback),
+      fetcher,
+    });
+    adapter.emit("api.request.failed", validAttributes);
+    expect(adapter.pendingCount()).toBe(1);
+
+    adapter.dispose();
+    expect(adapter.pendingCount()).toBe(0);
+
+    for (const callback of callbacks) callback();
+    await Promise.resolve();
+    expect(fetcher).not.toHaveBeenCalled();
+    // Disposal is a silent shutdown, not a recursive drop event.
+    expect(adapter.dropReasons()).toEqual({});
+  });
+
+  it("ignores emit after dispose", async () => {
+    const fetcher = vi.fn(async () => new Response(null, { status: 204 }));
+    const adapter = createTelemetryAdapter({
+      enabled: true,
+      endpoint: "https://telemetry.test/events",
+      schedule: () => {},
+      fetcher,
+    });
+    adapter.dispose();
+    adapter.emit("api.request.failed", validAttributes);
+    await adapter.flush();
+
+    expect(adapter.pendingCount()).toBe(0);
+    expect(fetcher).not.toHaveBeenCalled();
+  });
+
+  it("aborts an in-flight sink and prevents post-dispose rescheduling", async () => {
+    const callbacks: Array<() => void> = [];
+    let observedSignal: AbortSignal | undefined;
+    let releaseSink: (() => void) | undefined;
+    const fetcher = vi.fn(async (_input: unknown, init?: RequestInit) => {
+      observedSignal = init?.signal ?? undefined;
+      await new Promise<void>((resolve) => {
+        releaseSink = resolve;
+      });
+      return new Response(null, { status: 204 });
+    });
+    const adapter = createTelemetryAdapter({
+      enabled: true,
+      endpoint: "https://telemetry.test/events",
+      schedule: (callback) => callbacks.push(callback),
+      fetcher: fetcher as unknown as typeof fetch,
+    });
+    adapter.emit("api.request.failed", validAttributes);
+    callbacks.splice(0).forEach((callback) => callback());
+    await vi.waitFor(() => expect(fetcher).toHaveBeenCalledOnce());
+
+    adapter.dispose();
+    expect(observedSignal?.aborted).toBe(true);
+
+    // The sink ignored the abort and settles late.
+    releaseSink?.();
+    await Promise.resolve();
+    await Promise.resolve();
+    expect(callbacks).toHaveLength(0);
+    expect(fetcher).toHaveBeenCalledOnce();
+  });
+
+  it("joins an already active flush", async () => {
+    let releaseSink: (() => void) | undefined;
+    const fetcher = vi.fn(async () => {
+      await new Promise<void>((resolve) => {
+        releaseSink = resolve;
+      });
+      return new Response(null, { status: 204 });
+    });
+    const adapter = createTelemetryAdapter({
+      enabled: true,
+      endpoint: "https://telemetry.test/events",
+      schedule: () => {},
+      fetcher: fetcher as unknown as typeof fetch,
+    });
+    adapter.emit("api.request.failed", validAttributes);
+
+    const first = adapter.flush();
+    await vi.waitFor(() => expect(fetcher).toHaveBeenCalledOnce());
+    let secondSettled = false;
+    const second = adapter.flush().then(() => {
+      secondSettled = true;
+    });
+    await Promise.resolve();
+    expect(secondSettled).toBe(false);
+
+    releaseSink?.();
+    await first;
+    await second;
+    expect(secondSettled).toBe(true);
+    adapter.dispose();
+  });
+
+  it.each([Number.NaN, Number.POSITIVE_INFINITY, 0, -1, 1.5])(
+    "rejects invalid telemetry and diagnostics capacity %s",
+    (value) => {
+      expect(() =>
+        createTelemetryAdapter({
+          enabled: true,
+          endpoint: "https://telemetry.test/events",
+          maxQueue: value,
+        }),
+      ).toThrow(TypeError);
+      expect(() => createDiagnosticsAdapter({ maxEntries: value })).toThrow(
+        TypeError,
+      );
+    },
+  );
 });