apps/api/src/server.ts

@@ -130,6 +130,13 @@ export interface StartApiOptions {
    * test can prove real timeout behavior deterministically without waiting out the production
    * values. Never set by the executable entry point. */
   readonly nodeServerTimeoutOverride?: Partial<NodeServerTimeouts>;
+  /** Test-only escape hatch: narrows the graceful HTTP drain deadline
+   * (`HTTP_DRAIN_TIMEOUT_MS`) so a forced close can be proven in milliseconds. Never set by the
+   * executable entry point. */
+  readonly drainTimeoutMsOverride?: number;
+  /** Test-only escape hatch: narrows the per-module stop/rollback hook deadline
+   * (`MODULE_STOP_HOOK_TIMEOUT_MS`). Never set by the executable entry point. */
+  readonly moduleHookTimeoutMsOverride?: number;
 }
 
 async function safeCloseDatabase(handle: ApiDatabaseHandle): Promise<void> {
@@ -152,6 +159,94 @@ async function safeStopModuleRuntime(moduleRuntime: StartedModuleRuntime): Promi
   }
 }
 
+/**
+ * The platform's fixed graceful-drain deadline. In-flight HTTP work has this long to finish after
+ * the listener stops accepting new connections; after it, connections are closed so process
+ * shutdown stays bounded. Not configurable in this slice.
+ */
+export const HTTP_DRAIN_TIMEOUT_MS = 30_000;
+
+/**
+ * A shutdown that did not complete cleanly. Carries only fixed booleans and the already-sanitized
+ * module-runtime error — never a request, socket, URL, credential, SQL, or stack detail — so an
+ * operator can see *which* stages failed without any of it becoming a disclosure channel.
+ */
+export class ApiShutdownError extends Error {
+  readonly drainTimedOut: boolean;
+  readonly moduleStopFailed: boolean;
+  readonly moduleStopError: unknown;
+
+  constructor(drainTimedOut: boolean, moduleStopError: unknown) {
+    super("The API did not shut down cleanly.");
+    this.name = "ApiShutdownError";
+    this.drainTimedOut = drainTimedOut;
+    this.moduleStopFailed = moduleStopError !== undefined;
+    this.moduleStopError = moduleStopError;
+    this.stack = `${this.name}: ${this.message}`;
+    Object.freeze(this);
+  }
+}
+
+/**
+ * Stops accepting new connections, then waits up to `timeoutMs` for existing HTTP work to finish.
+ *
+ * At the deadline it closes idle and then still-active connections using Node's own supported
+ * public `Server` APIs (`closeIdleConnections`/`closeAllConnections`, both called only after
+ * `server.close()`), and stops waiting for the close callback. No private socket or server field is
+ * touched.
+ *
+ * Force-closing a connection ends the transport. It does **not** prove that arbitrary in-process
+ * handler JavaScript was canceled — a handler promise may still be running and may still settle.
+ * What is guaranteed is that shutdown no longer waits for it, and that the pool closes afterwards,
+ * so its late database work fails.
+ */
+/** `ServerType` is a union that also covers HTTP/2 servers, which do not expose these two methods.
+ * `serve()` here always creates a plain `http.Server`, which does; narrowing with a real capability
+ * check keeps that honest instead of asserting a cast. */
+interface ConnectionClosingServer {
+  closeAllConnections(): void;
+  closeIdleConnections(): void;
+}
+
+function supportsConnectionClosing(
+  server: ServerType,
+): server is ServerType & ConnectionClosingServer {
+  const candidate = server as Partial<ConnectionClosingServer>;
+  return (
+    typeof candidate.closeIdleConnections === "function" &&
+    typeof candidate.closeAllConnections === "function"
+  );
+}
+
+async function drainHttpServer(
+  server: ServerType,
+  timeoutMs: number,
+): Promise<"drained" | "forced"> {
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  const closed = new Promise<void>((resolve) => {
+    server.close(() => resolve());
+  });
+  try {
+    const outcome = await Promise.race([
+      closed.then(() => "drained" as const),
+      new Promise<"forced">((resolve) => {
+        timer = setTimeout(() => resolve("forced"), timeoutMs);
+      }),
+    ]);
+    if (outcome === "drained") {
+      return "drained";
+    }
+    if (supportsConnectionClosing(server)) {
+      server.closeIdleConnections();
+      server.closeAllConnections();
+    }
+    return "forced";
+  } finally {
+    clearTimeout(timer);
+    // The close callback may still fire later; nothing awaits it, and it carries no rejection.
+  }
+}
+
 /**
  * Binds the Node HTTP server. Two platform-owned transport concerns are applied at this exact
  * boundary, below Hono entirely: `wrapFetchWithClientAddress` resolves and sanitizes the caller's
@@ -292,7 +387,11 @@ export async function startApi(
   // handle to stop here, only the pool this failure leaves open.
   let moduleRuntime: StartedModuleRuntime;
   try {
-    moduleRuntime = await startDiscoveredModuleRuntime(inventory.modules, services);
+    moduleRuntime = await startDiscoveredModuleRuntime(inventory.modules, services, {
+      ...(options.moduleHookTimeoutMsOverride === undefined
+        ? {}
+        : { hookTimeoutMs: options.moduleHookTimeoutMsOverride }),
+    });
   } catch (error) {
     await safeCloseDatabase(databaseHandle);
     throw error;
@@ -345,13 +444,12 @@ export async function startApi(
   // code can run against it afterward. Unlike the startup-cleanup paths above, a module-stop
   // failure here is not swallowed — it is a genuine shutdown result a caller should observe — but
   // database cleanup still always runs first, so a stop failure never leaves the pool open.
+  const drainTimeoutMs = options.drainTimeoutMsOverride ?? HTTP_DRAIN_TIMEOUT_MS;
   let shutdownPromise: Promise<void> | undefined;
   const shutdown = (): Promise<void> => {
     if (!shutdownPromise) {
       shutdownPromise = (async () => {
-        await new Promise<void>((resolve) => {
-          server.close(() => resolve());
-        });
+        const drainOutcome = await drainHttpServer(server, drainTimeoutMs);
         let moduleStopError: unknown;
         try {
           await moduleRuntime.stop();
@@ -359,8 +457,14 @@ export async function startApi(
           moduleStopError = error;
         }
         await safeCloseDatabase(databaseHandle);
-        if (moduleStopError) {
-          throw moduleStopError;
+        // Every stage above ran regardless of what the earlier ones did, so a drain timeout or a
+        // module-stop failure can never leave the pool open. Both failures are reported together,
+        // as fixed metadata: a forced drain is a genuine shutdown failure, not a success.
+        if (drainOutcome === "forced" || moduleStopError) {
+          throw new ApiShutdownError(
+            drainOutcome === "forced",
+            moduleStopError === undefined ? undefined : moduleStopError,
+          );
         }
       })();
     }

apps/api/tests/api-http.integration.ts

@@ -336,6 +336,8 @@ function apiOrigin(running: RunningApi): string {
 
 interface StartConfiguredApiOptions {
   readonly clientAddress?: ClientAddressConfigInput;
+  /** Test-only narrowing of the platform's 30-second graceful HTTP drain deadline. */
+  readonly drainTimeoutMsOverride?: number;
   readonly modulesDirectoryOverride?: string;
   readonly nodeServerTimeoutOverride?: Partial<NodeServerTimeouts>;
   /** Defaults to `true`: almost every existing scenario in this suite signs accounts up over HTTP
@@ -391,6 +393,9 @@ async function startConfiguredApi(
     ...(options.nodeServerTimeoutOverride === undefined
       ? {}
       : { nodeServerTimeoutOverride: options.nodeServerTimeoutOverride }),
+    ...(options.drainTimeoutMsOverride === undefined
+      ? {}
+      : { drainTimeoutMsOverride: options.drainTimeoutMsOverride }),
   });
 }
 
@@ -2725,3 +2730,114 @@ describe.sequential("live API — Node header/request timeouts and client errors
     expect(logs.join("\n")).not.toContain("zzz");
   });
 });
+
+/**
+ * Pass 4: shutdown is finite even when HTTP work does not settle.
+ *
+ * These drive a real bound listener and a real socket rather than a mock, because the behavior
+ * under test is what Node's own `server.close()` does and does not do. Every socket, server, and
+ * disposable database is closed in `finally`, and each scenario carries its own outer deadline so
+ * a regression fails rather than hangs.
+ */
+describe.sequential("live API — bounded HTTP drain and shutdown", () => {
+  async function withDeadline<T>(operation: Promise<T>, timeoutMs: number): Promise<T> {
+    let timer: ReturnType<typeof setTimeout> | undefined;
+    try {
+      return await Promise.race([
+        operation,
+        new Promise<never>((_resolve, reject) => {
+          timer = setTimeout(
+            () => reject(harnessError("shutdown did not settle in time")),
+            timeoutMs,
+          );
+        }),
+      ]);
+    } finally {
+      clearTimeout(timer);
+    }
+  }
+
+  it("serves a real request and then drains cleanly, well inside the deadline", async () => {
+    const database = await createDisposableDatabase("drain_ok");
+    await applyCoreMigrations(database.url);
+    const running = await startConfiguredApi(database.url, { drainTimeoutMsOverride: 5_000 });
+    try {
+      const origin = apiOrigin(running);
+      const response = await fetch(`${origin}/api/v1/principal`);
+      expect(response.status).toBe(200);
+      expect(await response.json()).toEqual({ kind: "anonymous" });
+
+      // Nothing is stalled, so the graceful path completes and shutdown reports success — the
+      // drain deadline is never reached and no connection is force-closed.
+      await expect(withDeadline(running.shutdown(), 20_000)).resolves.toBeUndefined();
+
+      // The listener really is closed afterwards.
+      await expect(fetch(`${origin}/api/v1/principal`)).rejects.toBeTruthy();
+    } finally {
+      await running.shutdown().catch(() => undefined);
+      await dropDatabase(database.name);
+    }
+  });
+
+  it("force-closes a stalled connection at the drain deadline and still reaches pool closure", async () => {
+    const database = await createDisposableDatabase("drain_forced");
+    await applyCoreMigrations(database.url);
+    // A short, test-only drain deadline; production keeps HTTP_DRAIN_TIMEOUT_MS unchanged.
+    const running = await startConfiguredApi(database.url, { drainTimeoutMsOverride: 250 });
+    const address = running.server.address();
+    if (!address || typeof address !== "object") {
+      throw harnessError("The API integration server has no bound address.");
+    }
+    const socket = net.connect({ host: "127.0.0.1", port: address.port });
+    try {
+      // A real, deliberately incomplete request: headers are never finished, so Node keeps the
+      // connection open and `server.close()` alone would wait indefinitely.
+      await new Promise<void>((resolve, reject) => {
+        socket.once("error", reject);
+        socket.on("data", () => undefined);
+        socket.write("GET /api/v1/principal HTTP/1.1\r\nHost: 127.0.0.1\r\n", () => resolve());
+      });
+
+      const closed = new Promise<void>((resolve) => {
+        socket.once("close", () => resolve());
+      });
+
+      let captured: unknown;
+      try {
+        await withDeadline(running.shutdown(), 20_000);
+      } catch (error) {
+        captured = error;
+      }
+
+      // A forced drain is a genuine shutdown failure, never reported as success.
+      expect(captured).toMatchObject({ drainTimedOut: true, name: "ApiShutdownError" });
+      // ...and it still carries nothing about the request, socket, or database.
+      const publicText = `${String(captured)}\n${(captured as Error).stack ?? ""}`;
+      expect(publicText).not.toContain("127.0.0.1");
+      expect(publicText).not.toContain(database.name);
+      // The connection was actually closed by the platform, not left dangling.
+      await withDeadline(closed, 10_000);
+    } finally {
+      socket.destroy();
+      await running.shutdown().catch(() => undefined);
+      await dropDatabase(database.name);
+    }
+  });
+
+  it("shares one memoized shutdown across concurrent callers", async () => {
+    const database = await createDisposableDatabase("drain_once");
+    await applyCoreMigrations(database.url);
+    const running = await startConfiguredApi(database.url, { drainTimeoutMsOverride: 5_000 });
+    try {
+      const [first, second, third] = await withDeadline(
+        Promise.all([running.shutdown(), running.shutdown(), running.shutdown()]),
+        20_000,
+      );
+      expect([first, second, third]).toEqual([undefined, undefined, undefined]);
+      // A later caller observes the same completed result rather than restarting cleanup.
+      await expect(running.shutdown()).resolves.toBeUndefined();
+    } finally {
+      await dropDatabase(database.name);
+    }
+  });
+});

apps/api/tests/lifecycle.test.ts

@@ -1067,11 +1067,14 @@ describe("Ekklesia API module runtime startup and shutdown", () => {
       issues: [expect.objectContaining({ code: "MODULE_START_FAILED" })],
     });
 
+    // The module whose start hook failed is rolled back first through its own stop hook, then the
+    // module that had already started, in reverse order.
     expect(recorder.events).toEqual([
       "one:register",
       "two:register",
       "one:start",
       "two:start",
+      "two:stop",
       "one:stop",
     ]);
     expect(poolConstructor).toHaveBeenCalledTimes(1);
@@ -1177,8 +1180,14 @@ describe("Ekklesia API module runtime startup and shutdown", () => {
 
     const running = await startApi(validConfig(), { modulesDirectoryOverride: modulesDirectory });
 
+    // Shutdown now aggregates which stages failed; the module-runtime error is carried intact.
     await expect(running.shutdown()).rejects.toMatchObject({
-      issues: [expect.objectContaining({ code: "MODULE_STOP_FAILED", moduleId: "stop-fails" })],
+      drainTimedOut: false,
+      moduleStopError: {
+        issues: [expect.objectContaining({ code: "MODULE_STOP_FAILED", moduleId: "stop-fails" })],
+      },
+      moduleStopFailed: true,
+      name: "ApiShutdownError",
     });
     expect(poolEnd).toHaveBeenCalledTimes(1);
   });

apps/api/tests/module-database.integration.ts

@@ -450,8 +450,15 @@ describe.sequential("live API — module database explicit migration ownership",
       await applyCoreMigrations(database.url);
       resetRecorder();
 
+      // The module whose `start` failed is now rolled back through its own `stop` hook first. That
+      // hook also writes to the table its missing migration would have created, so it fails too —
+      // which is exactly the documented reason a `stop` hook must tolerate partial initialization.
+      // Both outcomes are reported, in order, and neither is swallowed.
       await expect(startMigratedApi(database.url, fixtureModulesDirectory)).rejects.toMatchObject({
-        issues: [expect.objectContaining({ code: "MODULE_START_FAILED", moduleId: "db-facade" })],
+        issues: [
+          expect.objectContaining({ code: "MODULE_START_FAILED", moduleId: "db-facade" }),
+          expect.objectContaining({ code: "MODULE_ROLLBACK_FAILED", moduleId: "db-facade" }),
+        ],
       });
 
       // No schema inference, no automatic table creation as a side effect of the failed attempt.

core/src/modules/runtime.ts

@@ -34,6 +34,7 @@ export type ModuleRuntimeIssueCode =
   | "MODULE_CAPABILITY_UNREGISTERED"
   | "MODULE_REGISTER_FAILED"
   | "MODULE_ROLLBACK_FAILED"
+  | "MODULE_ROLLBACK_TIMEOUT"
   | "MODULE_ROUTE_DEFINITION_INVALID"
   | "MODULE_ROUTE_DUPLICATE"
   | "MODULE_ROUTE_MAX_BODY_BYTES_INVALID"
@@ -47,7 +48,8 @@ export type ModuleRuntimeIssueCode =
   | "MODULE_RUNTIME_IMPORT_FAILED"
   | "MODULE_RUNTIME_VERSION_MISMATCH"
   | "MODULE_START_FAILED"
-  | "MODULE_STOP_FAILED";
+  | "MODULE_STOP_FAILED"
+  | "MODULE_STOP_TIMEOUT";
 
 export interface ModuleRuntimeIssue {
   readonly code: ModuleRuntimeIssueCode;
@@ -74,7 +76,16 @@ export interface StartedModuleRuntime {
   stop(): Promise<void>;
 }
 
-export interface StartModuleRuntimeOptions extends ModuleDiscoveryOptions {
+export interface ModuleRuntimeTimeoutOptions {
+  /** Test-only narrowing of `MODULE_STOP_HOOK_TIMEOUT_MS`, applied to both normal stop and startup
+   * rollback, so a deadline can be proven in milliseconds instead of ten seconds. Production
+   * composition never sets this. */
+  readonly hookTimeoutMs?: number;
+}
+
+export interface StartModuleRuntimeOptions
+  extends ModuleDiscoveryOptions,
+    ModuleRuntimeTimeoutOptions {
   /** The same frozen services object every started module's `start`/`stop` hooks and (once
    * `apps/api` mounts routes) HTTP handlers receive for this runtime — a required production
    * dependency, never an optional test-only global. See `ModuleRuntimeServices`. */
@@ -489,18 +500,93 @@ async function registerModules(
   }
 }
 
+/**
+ * The platform's fixed wait for one module's `stop` hook, during either a normal stop or a startup
+ * rollback. Not configurable in this slice — no environment variable, manifest field, or runtime
+ * service — so every deployment bounds cleanup identically.
+ *
+ * This bounds *waiting*, not execution. An in-process promise cannot be canceled: a hook that
+ * exceeds this deadline keeps running and may settle later, but it no longer delays the rest of
+ * shutdown, and the pool closes after the bounded sequence, so its late database work fails. The
+ * platform does not sandbox, terminate, or undo a trusted module's JavaScript.
+ */
+export const MODULE_STOP_HOOK_TIMEOUT_MS = 10_000;
+
+interface StopModulesOptions {
+  /** Test-only narrowing of `MODULE_STOP_HOOK_TIMEOUT_MS` so a deadline can be proven without
+   * waiting out the production value. Production composition never sets this. */
+  readonly hookTimeoutMs?: number;
+}
+
+/**
+ * Awaits one hook under a deadline. The losing branch is always handled: the timer is cleared when
+ * the hook wins, and a hook that loses gets a permanent no-op rejection handler attached so a late
+ * throw can never surface as an unhandled rejection. A timed-out hook is never awaited again.
+ */
+async function awaitHookWithinDeadline(
+  hook: Promise<void>,
+  timeoutMs: number,
+): Promise<"completed" | "timed-out"> {
+  let timer: ReturnType<typeof setTimeout> | undefined;
+  // Attached before the race so the rejection is considered handled even if the timeout wins.
+  const observed = hook.then(
+    () => "completed" as const,
+    (error: unknown) => {
+      throw error;
+    },
+  );
+  observed.catch(() => undefined);
+  try {
+    return await Promise.race([
+      observed,
+      new Promise<"timed-out">((resolve) => {
+        timer = setTimeout(() => resolve("timed-out"), timeoutMs);
+      }),
+    ]);
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+/**
+ * Stops modules in reverse order, giving each hook a bounded slice of time and continuing to the
+ * next module whatever the previous one did. Every applicable hook is invoked at most once.
+ *
+ * The whole sequence is finite because the module count is finite and fixed at discovery: its
+ * worst case is `startedModuleCount * MODULE_STOP_HOOK_TIMEOUT_MS`, not one global ten-second
+ * bound.
+ */
 async function stopModules(
   modules: readonly LoadedModule[],
   registry: StagedCapabilityRegistry,
   phase: "rollback" | "stop",
   services: ModuleRuntimeServices,
+  options: StopModulesOptions = {},
 ): Promise<readonly ModuleRuntimeIssue[]> {
+  const timeoutMs = options.hookTimeoutMs ?? MODULE_STOP_HOOK_TIMEOUT_MS;
   const issues: ModuleRuntimeIssue[] = [];
   for (const module of [...modules].reverse()) {
+    const moduleId = module.discovered.manifest.id;
+    const stop = module.definition.stop;
+    if (!stop) {
+      // No synthetic callback and no timer for a module that declares no stop hook.
+      continue;
+    }
     try {
-      await module.definition.stop?.(
-        registry.createRuntimeContext(module.discovered, phase, services),
+      const outcome = await awaitHookWithinDeadline(
+        Promise.resolve(stop(registry.createRuntimeContext(module.discovered, phase, services))),
+        timeoutMs,
       );
+      if (outcome === "timed-out") {
+        issues.push(
+          runtimeIssue(
+            phase === "rollback" ? "MODULE_ROLLBACK_TIMEOUT" : "MODULE_STOP_TIMEOUT",
+            phase,
+            moduleId,
+            `Module ${moduleId} ${phase} did not complete within the platform deadline.`,
+          ),
+        );
+      }
     } catch (error) {
       if (error instanceof ModuleRuntimeSignal) {
         issues.push(error.issue);
@@ -509,8 +595,8 @@ async function stopModules(
           runtimeIssue(
             phase === "rollback" ? "MODULE_ROLLBACK_FAILED" : "MODULE_STOP_FAILED",
             phase,
-            module.discovered.manifest.id,
-            `Module ${module.discovered.manifest.id} ${phase} failed.`,
+            moduleId,
+            `Module ${moduleId} ${phase} failed.`,
           ),
         );
       }
@@ -523,14 +609,21 @@ async function startModules(
   modules: readonly LoadedModule[],
   registry: StagedCapabilityRegistry,
   services: ModuleRuntimeServices,
+  options: StopModulesOptions = {},
 ): Promise<void> {
-  const started: LoadedModule[] = [];
+  // A module joins the rollback set the moment its `start` hook is *invoked*, not when it
+  // completes. A hook that acquired a connection, registered a listener, or recorded partial
+  // initialization before throwing still has cleanup to do, and previously only the modules that
+  // had already finished were rolled back — leaving the failing module's own resources behind. Its
+  // `stop` hook therefore runs first, ahead of the earlier modules in reverse order, which is why
+  // module authors must make `stop` safe after a partially completed `start`.
+  const rollbackCandidates: LoadedModule[] = [];
   for (const module of modules) {
+    rollbackCandidates.push(module);
     try {
       await module.definition.start?.(
         registry.createRuntimeContext(module.discovered, "start", services),
       );
-      started.push(module);
     } catch (error) {
       const startIssue =
         error instanceof ModuleRuntimeSignal
@@ -541,7 +634,13 @@ async function startModules(
               module.discovered.manifest.id,
               `Module ${module.discovered.manifest.id} startup failed.`,
             );
-      const rollbackIssues = await stopModules(started, registry, "rollback", services);
+      const rollbackIssues = await stopModules(
+        rollbackCandidates,
+        registry,
+        "rollback",
+        services,
+        options,
+      );
       throw new ModuleRuntimeError([startIssue, ...rollbackIssues]);
     }
   }
@@ -558,12 +657,13 @@ async function startModules(
 export async function startDiscoveredModuleRuntime(
   discoveredModules: readonly DiscoveredModule[],
   services: ModuleRuntimeServices,
+  timeouts: ModuleRuntimeTimeoutOptions = {},
 ): Promise<StartedModuleRuntime> {
   const modules = await loadRuntimeDefinitions(discoveredModules);
   const capabilityRegistry = new StagedCapabilityRegistry();
   const httpRoutes = new StagedModuleHttpRoutes();
   await registerModules(modules, capabilityRegistry, httpRoutes);
-  await startModules(modules, capabilityRegistry, services);
+  await startModules(modules, capabilityRegistry, services, timeouts);
 
   const capabilities = capabilityRegistry.publish();
   const routes = httpRoutes.publish();
@@ -583,7 +683,7 @@ export async function startDiscoveredModuleRuntime(
     routes,
     stop: (): Promise<void> => {
       stopPromise ??= (async () => {
-        const issues = await stopModules(modules, capabilityRegistry, "stop", services);
+        const issues = await stopModules(modules, capabilityRegistry, "stop", services, timeouts);
         if (issues.length > 0) {
           throw new ModuleRuntimeError(issues);
         }
@@ -597,5 +697,7 @@ export async function startModuleRuntime(
   options: StartModuleRuntimeOptions,
 ): Promise<StartedModuleRuntime> {
   const discovered = await discoverModules(options);
-  return startDiscoveredModuleRuntime(discovered, options.services);
+  return startDiscoveredModuleRuntime(discovered, options.services, {
+    ...(options.hookTimeoutMs === undefined ? {} : { hookTimeoutMs: options.hookTimeoutMs }),
+  });
 }

core/tests/module-runtime.test.ts

@@ -369,10 +369,13 @@ describe("module runtime", () => {
     const error = await captureRuntimeError(modulesDirectory);
 
     expect(runtimeIssueCodes(error)).toEqual(["MODULE_START_FAILED", "MODULE_ROLLBACK_FAILED"]);
+    // The module whose `start` threw is rolled back *first*, then the previously started modules in
+    // reverse order — its own hook may have partial initialization to release before it threw.
     expect(recorder.events).toEqual([
       "one:start",
       "two:start",
       "three:start",
+      "three:stop",
       "two:stop",
       "one:stop",
     ]);
@@ -1151,3 +1154,200 @@ describe("module runtime services", () => {
     expect(vi.mocked(services.database.transaction)).not.toHaveBeenCalled();
   });
 });
+
+/**
+ * Pass 4: a module whose `start` hook fails owns cleanup of whatever it managed to acquire, and
+ * every stop/rollback hook runs under a finite platform deadline.
+ */
+describe("module startup rollback and bounded stop hooks", () => {
+  it("releases a partially initialized resource through the failing module's own stop hook", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    const recorder = createEventRecorder();
+    await createRuntimeModule(modulesDirectory, "partial", {
+      serverSource: `let acquired = false;
+      export default {
+        id: "partial", version: "0.1.0",
+        start() {
+          acquired = true;
+          ${recorder.expression}.push("partial:acquired");
+          throw new Error("secret-token at /private/path");
+        },
+        stop() {
+          if (acquired) { acquired = false; ${recorder.expression}.push("partial:released"); }
+        }
+      };`,
+    });
+
+    const error = await captureRuntimeError(modulesDirectory);
+
+    expect(runtimeIssueCodes(error)).toEqual(["MODULE_START_FAILED"]);
+    expect(recorder.events).toEqual(["partial:acquired", "partial:released"]);
+    expect(JSON.stringify(error.issues)).not.toContain("secret-token");
+    expect(JSON.stringify(error.issues)).not.toContain("/private/path");
+  });
+
+  it("reports the failing module's own rollback failure and still rolls back earlier modules", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    const recorder = createEventRecorder();
+    await createRuntimeModule(modulesDirectory, "earlier", {
+      serverSource: `export default {
+        id: "earlier", version: "0.1.0",
+        stop() { ${recorder.expression}.push("earlier:stop"); }
+      };`,
+    });
+    await createRuntimeModule(modulesDirectory, "failing", {
+      manifest: manifestFor("failing", {
+        dependencies: { modules: { required: { earlier: "^0.1.0" } } },
+      }),
+      serverSource: `export default {
+        id: "failing", version: "0.1.0",
+        start() { throw new Error("start blew up"); },
+        stop() { ${recorder.expression}.push("failing:stop"); throw new Error("rollback blew up"); }
+      };`,
+    });
+
+    const error = await captureRuntimeError(modulesDirectory);
+
+    expect(runtimeIssueCodes(error)).toEqual(["MODULE_START_FAILED", "MODULE_ROLLBACK_FAILED"]);
+    expect(recorder.events).toEqual(["failing:stop", "earlier:stop"]);
+  });
+
+  it("gives a ModuleRuntimeSignal thrown during start the same rollback treatment", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    const recorder = createEventRecorder();
+    // `capabilities.require` for a capability this module never declared throws a
+    // `ModuleRuntimeSignal` from inside `start`, rather than an ordinary Error.
+    await createRuntimeModule(modulesDirectory, "signal", {
+      serverSource: `export default {
+        id: "signal", version: "0.1.0",
+        start({ capabilities }) { capabilities.require("undeclared.thing"); },
+        stop() { ${recorder.expression}.push("signal:stop"); }
+      };`,
+    });
+
+    const error = await captureRuntimeError(modulesDirectory);
+
+    expect(error.issues[0]?.phase).toBe("start");
+    expect(recorder.events).toEqual(["signal:stop"]);
+  });
+
+  it("invokes each stop hook at most once across rollback", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    const recorder = createEventRecorder();
+    await createRuntimeModule(modulesDirectory, "counted", {
+      serverSource: `export default {
+        id: "counted", version: "0.1.0",
+        start() { throw new Error("nope"); },
+        stop() { ${recorder.expression}.push("counted:stop"); }
+      };`,
+    });
+
+    await captureRuntimeError(modulesDirectory);
+
+    expect(recorder.events.filter((event) => event === "counted:stop")).toHaveLength(1);
+  });
+
+  it("bounds a never-settling rollback hook and continues with the earlier module", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    const recorder = createEventRecorder();
+    await createRuntimeModule(modulesDirectory, "earlier", {
+      serverSource: `export default {
+        id: "earlier", version: "0.1.0",
+        stop() { ${recorder.expression}.push("earlier:stop"); }
+      };`,
+    });
+    await createRuntimeModule(modulesDirectory, "hanging", {
+      manifest: manifestFor("hanging", {
+        dependencies: { modules: { required: { earlier: "^0.1.0" } } },
+      }),
+      serverSource: `export default {
+        id: "hanging", version: "0.1.0",
+        start() { throw new Error("nope"); },
+        stop() { ${recorder.expression}.push("hanging:stop"); return new Promise(() => {}); }
+      };`,
+    });
+
+    // A short test-only deadline; production uses MODULE_STOP_HOOK_TIMEOUT_MS unchanged.
+    let captured: ModuleRuntimeError | undefined;
+    try {
+      await startModuleRuntime({ ...runtimeOptions(modulesDirectory), hookTimeoutMs: 25 });
+    } catch (error) {
+      captured = error as ModuleRuntimeError;
+    }
+
+    expect(captured).toBeInstanceOf(ModuleRuntimeError);
+    expect(runtimeIssueCodes(captured as ModuleRuntimeError)).toEqual([
+      "MODULE_START_FAILED",
+      "MODULE_ROLLBACK_TIMEOUT",
+    ]);
+    // The timed-out hook did not block the earlier module's rollback.
+    expect(recorder.events).toEqual(["hanging:stop", "earlier:stop"]);
+  });
+
+  it("bounds a never-settling normal stop hook and continues with the earlier module", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    const recorder = createEventRecorder();
+    await createRuntimeModule(modulesDirectory, "earlier", {
+      serverSource: `export default {
+        id: "earlier", version: "0.1.0",
+        stop() { ${recorder.expression}.push("earlier:stop"); }
+      };`,
+    });
+    await createRuntimeModule(modulesDirectory, "hanging", {
+      manifest: manifestFor("hanging", {
+        dependencies: { modules: { required: { earlier: "^0.1.0" } } },
+      }),
+      serverSource: `export default {
+        id: "hanging", version: "0.1.0",
+        stop() { ${recorder.expression}.push("hanging:stop"); return new Promise(() => {}); }
+      };`,
+    });
+
+    const runtime = await startModuleRuntime({
+      ...runtimeOptions(modulesDirectory),
+      hookTimeoutMs: 25,
+    });
+    let captured: ModuleRuntimeError | undefined;
+    try {
+      await runtime.stop();
+    } catch (error) {
+      captured = error as ModuleRuntimeError;
+    }
+
+    expect(runtimeIssueCodes(captured as ModuleRuntimeError)).toEqual(["MODULE_STOP_TIMEOUT"]);
+    expect(recorder.events).toEqual(["hanging:stop", "earlier:stop"]);
+    // Repeated and concurrent callers share the memoized promise and never restart cleanup.
+    const [first, second] = await Promise.allSettled([runtime.stop(), runtime.stop()]);
+    expect(first.status).toBe("rejected");
+    expect(second.status).toBe("rejected");
+    expect(recorder.events.filter((event) => event === "hanging:stop")).toHaveLength(1);
+  });
+
+  it("never lets a hook that rejects after its deadline become an unhandled rejection", async () => {
+    const modulesDirectory = await createModulesDirectory();
+    await createRuntimeModule(modulesDirectory, "late", {
+      serverSource: `export default {
+        id: "late", version: "0.1.0",
+        stop() { return new Promise((_resolve, reject) => { setTimeout(() => reject(new Error("late")), 40); }); }
+      };`,
+    });
+
+    const unhandled: unknown[] = [];
+    const onUnhandled = (reason: unknown): void => {
+      unhandled.push(reason);
+    };
+    process.on("unhandledRejection", onUnhandled);
+    try {
+      const runtime = await startModuleRuntime({
+        ...runtimeOptions(modulesDirectory),
+        hookTimeoutMs: 10,
+      });
+      await expect(runtime.stop()).rejects.toBeInstanceOf(ModuleRuntimeError);
+      // Give the late rejection time to fire, then confirm nothing escaped.
+      await new Promise((resolve) => setTimeout(resolve, 80));
+      expect(unhandled).toEqual([]);
+    } finally {
+      process.off("unhandledRejection", onUnhandled);
+    }
+  });
+});

docs/development/api.md

@@ -585,8 +585,9 @@ failure propagates: module-services/identity/authorization construction failing
 was started yet — the frozen `ModuleRuntimeServices` object is built in the same step, from the
 same pool); the server failing to bind its listen address; or an app-construction failure, both of
 which also stop the module runtime first, since it may have already started successfully. A module
-load/register/start failure uses Core's own existing rollback (already-started modules stop, in
-reverse order, before the failure propagates) and then closes the pool — there is no separate
+load/register/start failure uses Core's own rollback — the module whose `start` hook failed is
+stopped first, through its own `stop` hook, then already-started modules in reverse order, each
+under the same bounded per-hook deadline — before the failure propagates, and then closes the pool — there is no separate
 runtime handle to stop in that case, since `startDiscoveredModuleRuntime` never returned one.
 
 ## A real signup → sign-in → principal → sign-out transcript

docs/development/testing.md

@@ -152,6 +152,20 @@ key material is ever committed to this repository** — and the disposable conta
 `afterAll` even when an assertion fails. See
 [ADR 0018](../architecture/decisions/0018-identity-session-resolution-and-database-transport.md).
 
+### Which kind of timing proof lives where
+
+Pass 4 introduced several deadlines, and their tests deliberately differ in kind:
+
+- **Unit deadline tests** (`core/tests/module-runtime.test.ts`) drive module stop and rollback
+  hooks through a short, test-only `hookTimeoutMs` override with manually controlled
+  never-settling promises — no real waiting, no sleeps, and an explicit `unhandledRejection` check
+  proving a hook that rejects after its deadline stays handled.
+- **Real raw-HTTP force-close** (`apps/api/tests/api-http.integration.ts`) binds the real listener,
+  opens a real socket, writes a deliberately incomplete request, and proves shutdown force-closes
+  it at a short `drainTimeoutMsOverride` and still reaches pool closure.
+Both use a short **test-only override**; the production defaults (10 s per module hook, 30 s HTTP
+drain) are asserted separately as constants and are never waited out.
+
 CI runs the same command in the dedicated `migration-smoke` job. After this change is merged, add
 the following pull-request status context to `main` branch protection:
 

docs/module-development/README.md

@@ -180,6 +180,21 @@ Entrypoint imports must have no top-level side effects. Modules are trusted back
 evaluation cannot be rolled back. Keep registration declarative; open resources in `start` and
 close them in `stop`.
 
+### `stop()` may run after a partially completed `start()`
+
+If your `start` hook throws, **your own `stop` hook runs first**, ahead of the previously started
+modules being rolled back in reverse order. That is deliberate: a hook that opened a connection,
+registered a listener, or recorded partial initialization before failing is the only code that knows
+how to release it. Write `stop` so it is safe and idempotent about resources it may or may not have
+acquired — guard on what you actually took, and never assume `start` completed.
+
+Stop and rollback hooks have finite time: the platform waits at most **10 seconds** per module hook,
+then records a timeout and continues with the next module in reverse order. Settle promptly. A hook
+that exceeds its deadline is not killed — the platform simply stops waiting for it — but the shared
+PostgreSQL pool closes after the bounded stop sequence, so **a timed-out hook cannot rely on
+database availability afterwards**, and any late query will fail. The platform cannot undo side
+effects your module performed outside its stop hook.
+
 ## HTTP routes
 
 A module registers plain Fetch (`Request` in, `Response` out) route handlers during `register`,