diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index fe5b8f2..bd28061 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -108,6 +108,21 @@ jobs:
 
       - run: pnpm install --frozen-lockfile
 
+      # The other two jobs set ELECTRON_SKIP_BINARY_DOWNLOAD, and all three
+      # share one pnpm store cache keyed on the lockfile, so whichever of them
+      # saves that cache first can hand this job an electron package whose
+      # postinstall was skipped and whose binary is therefore missing. The
+      # symptom is "Electron failed to install correctly" out of
+      # electron.launch, and because it depends on which parallel job wins the
+      # cache write, it appears as flake rather than as a broken job.
+      #
+      # install.js is the package's own postinstall. It checks the installed
+      # version first and downloads only what is missing, reusing the binary
+      # cache restored above, so running it here is idempotent and costs
+      # nothing when the store cache was clean.
+      - name: Repair the Electron binary if the shared store cache skipped it
+        run: node node_modules/electron/install.js
+
       - run: pnpm e2e --project=electron
 
       - uses: actions/upload-artifact@v4
diff --git a/e2e/journeys/create-book.spec.ts b/e2e/journeys/create-book.spec.ts
index 608959c..51734e3 100644
--- a/e2e/journeys/create-book.spec.ts
+++ b/e2e/journeys/create-book.spec.ts
@@ -97,16 +97,28 @@ test('revises the table of contents, persists it, and generates chapter 1 on app
 
   // Persistence, checked on disk rather than on screen: the revision wrote
   // through to toc.yml, not just into the streaming buffer the page renders.
-  const toc = await bookRepository(app.dataDir).getToc(id)
-  expect(toc.chapters.map(chapter => chapter.title)).toEqual(TOC_REVISED_CHAPTERS.map(chapter => chapter.title))
+  //
+  // Polled rather than read once, because the two are not simultaneous.
+  // Titles appear on screen while `revise-toc.ts` is still streaming, and the
+  // parse and the save both happen after the stream ends, so a bare read here
+  // races a write that is strictly later. Polling waits for that write on a
+  // deadline, the same way a web-first assertion waits for the screen, and it
+  // weakens nothing because the expected value is unchanged.
+  await expect.poll(async () => (await bookRepository(app.dataDir).getToc(id)).chapters.map(chapter => chapter.title))
+    .toEqual(TOC_REVISED_CHAPTERS.map(chapter => chapter.title))
 
   await wizard(page).approveToc()
 
   await expect(page.getByText(chapterMarker(1)).first()).toBeVisible()
 
-  const book = await readBook(app.dataDir, id)
-  expect(book.generatedUpTo).toBe(1)
-  expect(book.status).toBe('reading')
+  // Same reasoning, and this is the one that actually caught it. `start-book.ts`
+  // streams the chapter, saves it, generates its quiz, and only then writes
+  // generatedUpTo and status in its finalize step. Chapter 1's prose is on
+  // screen well before that, so reading the book once at this point asks a
+  // question the server has not answered yet. It won on every local run and on
+  // the pull request, and lost on the busier post-merge runner.
+  await expect.poll(async () => (await readBook(app.dataDir, id)).generatedUpTo).toBe(1)
+  expect((await readBook(app.dataDir, id)).status).toBe('reading')
 
   // And the model path /start actually drives: a skill classification call
   // and a chapter-1 stream, not just the revise-toc stream from earlier.