Fix restart sentinel internal continuations (#88161)

* fix restart sentinel internal continuations

* update gateway prompt snapshots

* stabilize sandbox browser audit timer tests

* drive sandbox audit timeouts deterministically

* drive gh-read timeout tests deterministically

* drive label-open-issues timeout tests deterministically

* document deterministic timeout test timers

* test: preserve deterministic timer setup after rebase

test/scripts/gh-read.test.ts

@@ -1,4 +1,4 @@
-import { beforeEach, describe, expect, it, vi } from "vitest";
+import { afterEach, describe, expect, it, vi } from "vitest";
 import {
   buildReadPermissions,
   githubJson,
@@ -11,7 +11,7 @@ import {
 } from "../../scripts/gh-read.js";
 
 describe("gh-read helpers", () => {
-  beforeEach(() => {
+  afterEach(() => {
     vi.useRealTimers();
   });
 
@@ -59,19 +59,27 @@ describe("gh-read helpers", () => {
   });
 
   it("aborts stalled GitHub API fetches at the request timeout", async () => {
-    vi.useFakeTimers();
     let signal: AbortSignal | undefined;
+    let markFetchStarted!: () => void;
+    const fetchStarted = new Promise<void>((resolve) => {
+      markFetchStarted = resolve;
+    });
+
+    vi.useFakeTimers();
     const request = githubJson("/app", "token", undefined, {
       timeoutMs: 5,
       fetchImpl: ((_url, init) => {
         signal = init?.signal ?? undefined;
+        markFetchStarted();
         return new Promise(() => {});
       }) as typeof fetch,
     });
+    const rejection = expect(request).rejects.toThrow(/GitHub API GET \/app exceeded timeout/u);
 
-    const rejected = expect(request).rejects.toThrow(/GitHub API GET \/app exceeded timeout/u);
+    await fetchStarted;
     await vi.advanceTimersByTimeAsync(5);
-    await rejected;
+
+    await rejection;
     expect(signal?.aborted).toBe(true);
   });
 
@@ -82,12 +90,13 @@ describe("gh-read helpers", () => {
       timeoutMs: 5,
       fetchImpl: (() => Promise.resolve(response)) as typeof fetch,
     });
-
-    const rejected = expect(request).rejects.toThrow(
+    const rejection = expect(request).rejects.toThrow(
       /GitHub API GET \/app\/installations exceeded timeout/u,
     );
+
     await vi.advanceTimersByTimeAsync(5);
-    await rejected;
+
+    await rejection;
   });
 
   it("bounds GitHub API error response bodies", async () => {

test/scripts/label-open-issues.test.ts

@@ -1,4 +1,4 @@
-import { describe, expect, it } from "vitest";
+import { afterEach, describe, expect, it, vi } from "vitest";
 import { testing } from "../../scripts/label-open-issues.ts";
 
 const labelItem = {
@@ -9,6 +9,12 @@ const labelItem = {
 };
 
 describe("label-open-issues helpers", () => {
+  // Timeout tests below advance fake timers explicitly so CI shard load cannot
+  // turn a bounded request-timeout assertion into a wall-clock wait.
+  afterEach(() => {
+    vi.useRealTimers();
+  });
+
   it("classifies items from OpenAI structured response text", async () => {
     const response = new Response(
       JSON.stringify({
@@ -37,34 +43,49 @@ describe("label-open-issues helpers", () => {
 
   it("aborts stalled OpenAI classification fetches at the request timeout", async () => {
     let signal: AbortSignal | undefined;
+    let markFetchStarted!: () => void;
+    const fetchStarted = new Promise<void>((resolve) => {
+      markFetchStarted = resolve;
+    });
+
+    vi.useFakeTimers();
     const request = testing.classifyItem(labelItem, "issue", {
       apiKey: "test-key",
       model: "test-model",
       timeoutMs: 5,
       fetchImpl: ((_url, init) => {
         signal = init?.signal ?? undefined;
+        markFetchStarted();
         return new Promise(() => {});
       }) as typeof fetch,
     });
-
-    await expect(request).rejects.toThrow(
+    const rejection = expect(request).rejects.toThrow(
       /OpenAI issue label classification request exceeded timeout/u,
     );
+
+    await fetchStarted;
+    await vi.advanceTimersByTimeAsync(5);
+
+    await rejection;
     expect(signal?.aborted).toBe(true);
   });
 
   it("times out stalled OpenAI classification body reads", async () => {
     const response = new Response(new ReadableStream({}), { status: 200 });
+    vi.useFakeTimers();
     const request = testing.classifyItem(labelItem, "issue", {
       apiKey: "test-key",
       model: "test-model",
       timeoutMs: 5,
       fetchImpl: (() => Promise.resolve(response)) as typeof fetch,
     });
-
-    await expect(request).rejects.toThrow(
+    const rejection = expect(request).rejects.toThrow(
       /OpenAI issue label classification request exceeded timeout/u,
     );
+
+    await vi.advanceTimersByTimeAsync(5);
+
+    await rejection;
   });
 
   it("bounds OpenAI error response bodies", async () => {