test(desktop): finish moving non-Electron coverage out of the E2E tier

.github/workflows/ci.yml

@@ -91,6 +91,12 @@ jobs:
       - name: Check Windows test inventory
         run: npm run windows:inventory
 
+      # Deliberately here and not in the e2e job: that job is conditional, and a
+      # tier this guard only checks when the tier already ran is a guard that
+      # grows back (#4761). Pure Node, no build -- it reads the spec sources.
+      - name: Check Electron e2e budget
+        run: npm run check:e2e-budget
+
       # Runs on the PR merge result: after a sibling protocol change lands on
       # main with the same epoch text, the silently merged tree still carries
       # the current base parent's epoch and this fails instead of shipping two
@@ -357,37 +363,7 @@ jobs:
 
       - name: Desktop e2e
         if: steps.plan.outputs.e2e == 'true'
-        run: |
-          set -euo pipefail
-          display_base=90
-          worker_count=4
-          xvfb_pids=()
-          cleanup() {
-            kill "${xvfb_pids[@]}" 2>/dev/null || true
-            wait "${xvfb_pids[@]}" 2>/dev/null || true
-          }
-          trap cleanup EXIT
-          for ((index = 0; index < worker_count; index += 1)); do
-            display=$((display_base + index))
-            Xvfb ":$display" -screen 0 1280x1024x24 -nolisten tcp \
-              >"$RUNNER_TEMP/xvfb-$display.log" 2>&1 &
-            xvfb_pids+=("$!")
-          done
-          for ((index = 0; index < worker_count; index += 1)); do
-            display=$((display_base + index))
-            for _ in {1..50}; do
-              [[ -S "/tmp/.X11-unix/X$display" ]] && break
-              kill -0 "${xvfb_pids[$index]}" 2>/dev/null || break
-              sleep 0.1
-            done
-            if [[ ! -S "/tmp/.X11-unix/X$display" ]]; then
-              cat "$RUNNER_TEMP/xvfb-$display.log"
-              exit 1
-            fi
-          done
-          MAKA_E2E_X_DISPLAY_BASE="$display_base" \
-            npm exec -w @maka/desktop -- playwright test \
-              --config e2e/playwright.config.ts --workers="$worker_count"
+        run: xvfb-run -a npm exec -w @maka/desktop -- playwright test --config e2e/playwright.config.ts
 
       # Playwright keeps a trace, a video and a screenshot for every failed
       # test. Without this they die with the runner, and an e2e flake can only

apps/desktop/e2e-budget.json

@@ -0,0 +1,73 @@
+{
+  "policy": [
+    "A test belongs in this tier only when it needs a real Electron window: an OS dialog, a main-process round trip, input released outside the window, persistence across a renderer reload, or a cross-process lifecycle.",
+    "Layout geometry, renderer state machines and pure CSS go to a Storybook play function; main-process logic goes to node --test. See #4761.",
+    "Every spec below records the Electron-owned mechanism it needs. If you cannot name one, the test does not belong here."
+  ],
+  "specs": {
+    "composer-directory-reference.spec.ts": {
+      "tests": 1,
+      "electron": "the folder reference has to survive a renderer reload and still agree with the Host's session record"
+    },
+    "context-window-save.spec.ts": {
+      "tests": 1,
+      "electron": "the saved window is read back from the Host's connection snapshot, not from renderer state"
+    },
+    "new-task-reload.spec.ts": {
+      "tests": 1,
+      "electron": "renderer reload is the whole contract: an explicit new task must not reopen history"
+    },
+    "proxy-password-editing.spec.ts": {
+      "tests": 1,
+      "electron": "the password never reaches the renderer; only the Host can report passwordConfigured and authenticate offline"
+    },
+    "quote-window-boundary.spec.ts": {
+      "tests": 1,
+      "electron": "a pointer capture released outside the window -- there is no window edge to leave in a browser tab"
+    },
+    "session-draft-focus.spec.ts": {
+      "tests": 1,
+      "electron": "needs a second real Session (Host round trip) to switch to; the focus and draft-restore halves alone would not earn a window"
+    },
+    "session-workbar.spec.ts": {
+      "tests": 5,
+      "electron": "Git changes re-read on window focus, terminal PTY ownership across Sessions, Side Chat's fork lifecycle, and a first send that has to reach the Host; the composer-usage test is renderer-only and rides along on those windows until app-shell.tsx's composer-to-workbar wiring has a story host"
+    },
+    "settings.spec.ts": {
+      "tests": 4,
+      "electron": "the preload makaE2eLatch holds the settings chunk mid-load, and the rename it commits is a Host write; the workbar-chrome test rides along on that window and would move to a story the day app-shell.tsx's settings wiring has one"
+    },
+    "sidebar-project-reload.spec.ts": {
+      "tests": 1,
+      "electron": "rail grouping has to be rebuilt from persisted project state after a renderer reload"
+    },
+    "skill-draft-lifecycle.spec.ts": {
+      "tests": 2,
+      "electron": "revision retry and cancel are Host-owned draft transitions across a parent and a child Session"
+    },
+    "slash-command-compact.spec.ts": {
+      "tests": 1,
+      "electron": "/compact is a Host round trip that ends in a real terminal Turn state"
+    },
+    "streaming-remount.spec.ts": {
+      "tests": 4,
+      "electron": "observation seeding, reconnect and settle are Host subscriptions surviving a renderer remount"
+    },
+    "transcript-scroll-cost.spec.ts": {
+      "tests": 3,
+      "electron": "the perf budget is measured from CDP wheel input and the browser's own render skipping"
+    },
+    "workhub-layout.spec.ts": {
+      "tests": 1,
+      "electron": "Coordination startup failure comes from the Host, and recovery needs a default model written back to it"
+    },
+    "workhub-reconstruction.spec.ts": {
+      "tests": 2,
+      "electron": "delegation linkage is rebuilt by the Host across navigation and across Sessions"
+    },
+    "zh-tw-locale.spec.ts": {
+      "tests": 1,
+      "electron": "the locale is a persisted Host setting and only takes effect through a renderer reload"
+    }
+  }
+}

apps/desktop/e2e/fixtures.ts

@@ -523,34 +523,13 @@ type E2eTestFixtures = {
   parentRemovalWindow: Page;
   railRenderWindow: Page;
   promptRailWindow: Page;
-  threadSearchWindow: Page;
   requestHeaderRowWindow: Page;
   newTaskTargetWindow: Page;
   directoryReferenceWindow: { page: Page; folder: string };
   accessibilityNarrativeWindow: Page;
 };
 
-type E2eWorkerFixtures = {
-  isolatedDisplay: void;
-};
-
-export const test = base.extend<E2eTestFixtures, E2eWorkerFixtures>({
-  isolatedDisplay: [async ({}, use, workerInfo) => {
-    const base = process.env.MAKA_E2E_X_DISPLAY_BASE;
-    if (base === undefined) {
-      await use();
-      return;
-    }
-    if (!/^\d+$/.test(base)) throw new Error(`Invalid E2E X display base: ${base}`);
-    const previous = process.env.DISPLAY;
-    process.env.DISPLAY = `:${Number(base) + workerInfo.parallelIndex}`;
-    try {
-      await use();
-    } finally {
-      if (previous === undefined) delete process.env.DISPLAY;
-      else process.env.DISPLAY = previous;
-    }
-  }, { scope: 'worker', auto: true }],
+export const test = base.extend<E2eTestFixtures>({
   directoryReferenceWindow: async ({}, use) => {
     await withE2eWindow(
       { seed: true, readinessSelector: COMPOSER_INPUT, locale: 'zh-CN', showWindow: true },
@@ -661,18 +640,6 @@ export const test = base.extend<E2eTestFixtures, E2eWorkerFixtures>({
       showWindow: true,
     }, use);
   },
-  // The same seeded transcript, on a window of its own. Search reads the Host
-  // through the bridge and renders nothing, so it needs neither the warm
-  // window's compositor nor its between-test reset — and taking it off the
-  // reused window is what retires the readiness gate's cross-test bleed (#4707).
-  threadSearchWindow: async ({}, use) => {
-    await withE2eWindow({
-      seed: false,
-      readinessSelector: '[data-turn-id]',
-      e2eFixtureScenario: 'chat-prompt-rail',
-      locale: 'zh-CN',
-    }, use);
-  },
   // Settings → 模型, where `no-models` is the seeded openai-compatible relay —
   // the connection type whose detail page owns the custom request headers
   // editor. Shown, because what this window is for is a rendered box

apps/desktop/e2e/playwright.config.ts

@@ -22,26 +22,19 @@ import { defineConfig } from '@playwright/test';
 /**
  * Playwright config for the desktop Electron E2E suite.
  *
- * Most tests launch a real Electron window backed by the deterministic fake
- * backend (MAKA_E2E=1) against their OWN throwaway userData dir. The read-only
- * prompt-rail scenario instead keeps its Electron + Host composition warm for
- * a worker and resets its Host range and renderer state per test. Keep the
- * local default at one worker: concurrent windows share OS focus, invalidating
- * geometry and focus contracts. Developers can still pass `--workers`
- * explicitly for a subset that has neither concern.
+ * Every test launches a real Electron window backed by the deterministic fake
+ * backend (MAKA_E2E=1) against its OWN throwaway userData dir. One worker,
+ * everywhere: what is left in this tier is here because it needs a native
+ * window, and concurrent windows share OS focus, which invalidates exactly the
+ * focus, pointer and geometry contracts that kept these tests here.
  * Deliberately no test count here — the previous note carried a stale one that
  * outlived two rounds of pruning. `playwright test --list` is the only figure
  * that cannot rot.
  *
- * CI gives every Playwright worker an isolated X display, so they overlap
- * without sharing focus or a compositor. Local parallelism is opt-in for the
- * same reason.
- *
  * Run from apps/desktop via `npm run e2e`, which builds the app first.
  */
 export default defineConfig({
   testDir: '.',
-  fullyParallel: true,
   workers: 1,
   // CI publishes no Playwright report that consumes Git metadata. Disable its
   // best-effort shallow-history fetch, which otherwise waits on a fixed timeout.