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

[Astro-Han]

Summary

The end of #4761: nothing is left in the Electron tier that does not need a native window, and the parallel harness that tier was carrying is gone. The tier goes from 35 tests / 17 files to 30 tests / 16 files, ~2.5 minutes on one worker.

The last geometry cases become stories. Five contracts were still costing an Electron window each, and none read anything Electron owns — they are layout and renderer state, measured in Chromium either way:

contract	now pinned by
the browser address field tracks the workbar column between 480px and the 320px floor	Product/Session Workbar/Browser Address Field Tracks Column Width
the tool picker's shortcut keys stay inside the panel at the floor	Product/Session Workbar/Tool Picker At Column Floor
Side Chat's composer resolves the conversation's --radius-chat and keeps its send button inside the card when the model chip runs long	Product/Session Workbar/Side Chat At Column Floor
the workbar's collapse toggle is one control across two bands: centre-aligned with the tab strip, shifted by the platform's caption width, parked at the same x/y in the titlebar, restoring the open face rather than the picker	Product/Shell Official AppShell/Workbar Collapse Keeps One Toggle In Place
WorkHub's submitted-work button keeps the target's project name inside its own box	Product/WorkHub/Submitted Work Keeps Target Metadata Inside

The titlebar case is the one that looked Electron-shaped, because the toggle moves between the workbar's own bar and the titlebar's right cluster and workbar/shell.css pads the bar with the titlebar strip's gutter so it lands on the same x in both. app-shell.stories.tsx already mounts the real titlebar over the real shell grid, so it hosts both bands; the story's collapse runs through reduceWorkbarLayout, the reducer the app dispatches into, so "restore brings back the face" is the production rule firing rather than a story-local useState. ComposedShell gains two optional props to let a story drive the titlebar's restore affordance it was previously rendering as a static collapsed={false}.

WorkHub gets its first story group. WorkHubSurface already takes its controller as a prop — the seam production fills with createWorkHubController — so a story serves one pinned projection instead of driving a live Coordination Session.

One E2E spec was pure duplication. fixture-thread-search.spec.ts launched a window to call window.maka.search.thread and assert a content hit carries its turn id. runtime-host-search-ipc-main.test.ts already asserts exactly that — same query, same 用户消息 summary, same four-field target — plus the title-hit and transcript-failure cases the E2E never covered. Deleted with no replacement, along with the window fixture it was the sole user of.

The parallel display harness retires. #4523 bought E2E wall-clock with four hand-built Xvfb displays and --workers=4. That parallelism is what invalidated the premises these tests were written against — a fresh renderer per test, exclusive OS focus, unthrottled frames — which is the flake class #4761 exists to remove. The config has answered workers: 1 since; CI was building four displays for one worker to use one of. The step is now the same one-line xvfb-run every other Electron job uses, and MAKA_E2E_X_DISPLAY_BASE plus its worker fixture go with it.

What stays is now written down, not asserted. apps/desktop/e2e-budget.json records each of the 16 remaining specs, its test count, and the Electron-owned mechanism it needs — the OS folder chooser, /compact's Host round trip, a pointer capture released outside the window, renderer-reload persistence, WorkHub's delegation lifecycle, revision-draft branching, the CDP wheel input the transcript perf budget measures. npm run check:e2e-budget fails on a spec that is not in the budget, a count that drifted, or an empty reason; the admission rule sits at the top of the same file and is printed back on failure. There is no regenerate mode — the failure names the number to type, and the reason has to be written by hand, so a new spec is justified in the diff or the guard stays red.

It runs in the planning lane rather than the e2e job, because that job is conditional and a guard that only checks the tier when the tier already ran is a guard that grows back. Pure Node, no build. Its discovery mirrors Playwright's own default testMatch — recursive, .spec.ts and .test.ts — because playwright.config.ts sets testDir: '.' and overrides neither; a test( it cannot attribute to a line of its own (a loop, a helper) is refused rather than counted as zero.

Not folded into a shared ratchet with check-renderer-architecture: the 3.5k lines there are Babel analysis and --base baseline rematerialization, and what the two share — read a ledger, diff, exit 1 — is under twenty lines each, over different subjects (monotonic debt vs an exact inventory). Worth revisiting at a third real instance.

Refs #4761

Verification

• Every new story assertion was mutation-checked: deleting the production rule it pins turns that story red, and the failure names that rule's own measurement.
  • .maka-browser-address-field's flex/min-width → expected 154 to be greater than 250
  • the launcher frame and list min-width: 0 → expected 1304 to be <= 1280
  • a forked --_chat-composer-radius → '12px' to be '28px'
  • the left-controls min-width: 0 → the send button 40px past the card's right edge
  • .workhub-submitted > button's height: auto → the project name 17px below the button
  • the caption term dropped from the workbar bar's padding-inline → expected 1224 to be close to 1144, received difference is 80
  • align-items: flex-start on the tab strip → centre line off by 2px
  • reduceWorkbarLayout's collapse clearing panels → the picker comes back instead of the face
  • [+] dispatching open-launcher instead of opening the face menu → Unable to find role="menu"
• npm run smoke:storybook -w @maka/desktop — 310 stories, 337 theme renders, all pass.
• npx playwright test --config e2e/playwright.config.ts --list — 30 tests in 16 files; e2e/session-workbar.spec.ts re-run on macOS, 5 passed.
• An adversarial review of the branch found three real defects, all fixed here: the budget guard's discovery surface was narrower than Playwright's in three silent ways (subdirectory, .test.ts, generated tests) — root cause was that the guard and the ledger were written from one reading of today's tier and validated against that same reading; deleting the thread-search E2E cost two assertions the main-process test did not carry (a composed Desktop session key where the fake returned a bare Host id, and a sequence that any projection satisfied at 0) — thread-search.ts forced to sequence: 0 now turns the unit test red; and the migrated titlebar story checked the workbar frame rather than the face inside it, which carries its own hidden.
• An ablation pass over the whole change removed two story assertions that could never fail (toBeVisible on the workbar card while the face menu is open; comparing the restored tab's label in a story with one face) and four unused pieces of the guard (--write, the test.only/test.skip/test.fixme and hook-form tables, the unread version field). Two more mutations were run to confirm what stayed does bite: WorkbarTitlebarActions rendering while the column is open, and a 4px shift of the titlebar's right cluster.
• npm run check:e2e-budget — 30 tests in 16 files; mutation-checked by adding an unjustified spec (not in the budget) and an extra test in an existing one (records 4 test(s), the file has 5).
• node scripts/audit-alignment.mjs — all fixtures clean.
• npm run format / npm run lint — clean.
• npm run typecheck -w @maka/desktop — the three pre-existing Astryx surface errors (ChatLayoutProps.autoScroll, MarkdownProps.settledText, SideNavItemProps.trailingAction) reproduce on main and are untouched; nothing else reports.

AI use

Select exactly one:

• No generative tool made a substantive contribution
• Generative tooling made a substantive contribution

Tool(s) and scope: Claude Code moved the tests, removed the harness and wrote this PR.

Checklist

• Tests cover the change and fail without it
• Lint, format, typecheck and the affected suites pass locally

Does this PR entail a change in behavior?

• Yes
• No

[Astro-Han]

Reviewed current head 57b1dae273438220df279fcf6e2075a2130435b4 (OPEN). NO-GO — one P2 test-coverage gap below; the integration gate is also still open (test/package pending). No new P0/P1 production correctness, permission, or concurrency issues found.

P2 — the deleted Electron search test's Host-session key projection is uncovered

The deleted fixture-thread-search.spec.ts:21-55 verified through the real Electron/preload/Host fixture chain that window.maka.search.thread returns desktopSessionKey, turnId=turn-prompt-rail-3, and sequence=4. The replacements do not cover preload.ts:3003-3018's Host-session key projection: runtime-host-search-ipc-main.test.ts:28-90 calls the main handler directly with a raw session id and single-message sequence=0; runtime-host-guest-ipc-preload.test.ts:65-89's search stub returns an empty array. If that mapping regresses, search-result navigation fails before parsing at preload.ts:419-438. Fix: add a non-empty search-hit preload bridge test asserting JSON.stringify([hostId, sessionId]), turnId, and nonzero sequence — no need to restore the E2E as-is.

Scope and limits

check:e2e-budget (30 tests/16 files), node --check, git diff --check pass; merge-tree against latest origin/main=d2d7efe6 clean. Exact-head hosted audit, package-linux, and both Runtime Host owner checks green; test/package still pending. Local Storybook/typecheck/E2E could not run independently (missing tsc and Playwright deps).

---

Automated review notice: This comment was posted by an automated review agent operated by Astro-Han. It is not an independent human review and does not replace one.

简体中文

评审结论来自自动化审查流程；发布者没有读这份 diff，核的是当前 head 有没有漂移、以及 exact-head 的门禁状态。当前 head 是 57b1dae，未关闭。一条 P2：删掉的 Electron 搜索用例覆盖的 key 投影丢了，替代测试没覆盖到。修好再合。

[jackwener]

Approving at exact head e8fa69a1. Checked the counts, the five destinations and the one outright deletion; all hold. No P0–P2.

• The tier numbers are exact. 16 e2e/*.spec.ts, 30 test( calls, and 16 entries in e2e-budget.json — the budget describes the tier that is actually there, not the one being described.
• The five geometry contracts land where the table says: BrowserAddressFieldTracksColumnWidth, ToolPickerAtColumnFloor and SideChatAtColumnFloor in session-workbar.stories.tsx, WorkbarCollapseKeepsOneToggleInPlace in app-shell.stories.tsx, SubmittedWorkKeepsTargetMetadataInside in workhub.stories.tsx.
• The titlebar case is the one worth spot-checking, and it survives the check. Its play distinguishes the workbar frame from the face's sibling overlay panel — a visible frame is not a visible face — and states "not the picker" as a reachability claim via queryByRole rather than a presence check, since the launcher stays mounted behind the face. That is a story reasoning about the real composition, not a re-skin of the assertion.
• The deletion is justified, which is the claim that usually is not. fixture-thread-search.spec.ts is removed with no replacement because runtime-host-search-ipc-main.test.ts already asserts the same thing: a content hit whose summary is 用户消息 and whose target passes a four-field deepEqual, plus the title-hit case the E2E never had. The destination is stronger than what left.
• The new guard is actually wired: check:e2e-budget is a root script, check-e2e-budget.mjs ships with its own check-e2e-budget.test.mjs, and it runs from ci.yml. Putting it in the planning lane rather than the conditional e2e job is the right call for the reason given — a guard that only checks the tier when the tier already ran is not a guard.

The reasoning for retiring the four-display harness is sound on its own terms: the parallelism invalidated the premises these tests were written against, and the config has answered workers: 1 since anyway.

简体中文

在 exact head e8fa69a1 上批准。数目、五处去处、以及那一处彻底删除都核过,全部成立。 无 P0–P2。

• 层级数字是精确的:16 个 e2e/*.spec.ts、30 个 test(、e2e-budget.json 16 条 —— 预算描述的是真实存在的那个层,不是被描述的那个。
• 五个几何契约都落在表格所说的位置(workbar 三个、app-shell 一个、workhub 一个)。
• titlebar 那条最值得抽查,而它经得起抽查:它的 play 区分了 workbar frame 与 face 的兄弟 overlay 面板(frame 可见不等于 face 可见),并把「不是 picker」表述为可达性主张、用 queryByRole 判断,因为 launcher 仍挂在 face 背后。这是一条在对真实组合做推理的 story,不是把断言换层皮。
• 那处删除是有据的,而这恰恰是通常不成立的那类主张。 fixture-thread-search.spec.ts 被无替代删除,因为 runtime-host-search-ipc-main.test.ts 已经断言同样的事:摘要为 用户消息 的内容命中,且 target 通过四字段 deepEqual,外加 E2E 从未有过的标题命中。去处比离开的更强。
• 新守卫是真的接上了:check:e2e-budget 是根脚本,check-e2e-budget.mjs 自带 check-e2e-budget.test.mjs,并由 ci.yml 触发。放在 planning lane 而不是条件式 e2e job 里是对的,理由如正文所说 —— 只有在该层已经跑过时才检查该层的守卫,不是守卫。

退役四显示器并行装置的理由本身也成立:那种并行恰恰使这些测试所依据的前提失效,而且配置此后一直是 workers: 1。

---

Automated review notice: This comment was posted by an automated review agent operated by jackwener. It is not an independent human review and does not replace one.

[Astro-Han]

Reviewed current head e8fa69a1b2fdd10bf958a7d34e3f9b8f103845be (OPEN, MERGEABLE/BLOCKED awaiting human review). NO-GO — one P2 test-coverage gap below, still open. Prior conclusions were re-verified against this head, not carried over: the new commit strengthens main-process search tests (composed Desktop session key, nonzero sequence) and Workbar hide/restore assertions, but the gap below survives. Checks: audit, package, package-linux, both owner checks green; test still pending.

P2 — the empty search stub still leaves the key projection uncovered

runtime-host-guest-ipc-preload.test.ts:58-89's search:thread stub still returns an empty array — no coverage of preload.ts:3003-3025's Host-session key projection for a non-empty search hit. A regression there still breaks search-result navigation before parsing at preload.ts:419-438. Fix: one non-empty preload bridge test asserting JSON.stringify([hostId, sessionId]), turnId, and nonzero sequence.

Scope and limits

Merge-tree against latest origin/main=d2d7efe6 clean. Local full verification tooling gaps as before; no GitHub review published from the lane for this head yet (this synthesis is the published record).

---

Automated review notice: This comment was posted by an automated review agent operated by Astro-Han. It is not an independent human review and does not replace one.

简体中文

评审结论来自自动化审查流程；发布者没有读这份 diff，核的是当前 head 有没有漂移、以及 exact-head 的门禁状态。当前 head 是 e8fa69a，未关闭。新提交补强了主进程测试，但空 stub 那条覆盖缺口还在。修好再合。