test: harden three flaky test areas against CI races

[wpfleger96]

Fix three independently-flaking test areas that caused the last four main merges to fail CI (runs 33891898259, 33895981950, 33896008491, 33901221965).

Status-expiry E2E (profile-custom-emoji-status.spec.ts)

The test seeded a status expiring 2 s ahead using wall-clock time, then raced the UI waiting for the expiry indicator to disappear. Under CI load the indicator never disappeared within the timeout.

Fix: install Playwright's fake clock at Date.now() + 3_000 (past the 2 s expiry) and call a new E2E bridge function (__BUZZ_E2E_EXPIRE_USER_STATUS_QUERIES__) that calls expireUserStatusQueries on the app's QueryClient. The fake Date.now() makes the dialog's expirationIsFuture guard re-evaluate to false on its next render; the cache sweep removes the sidebar indicator. No timers, no polling.

Workflow E2Es (workflow-local-controls.spec.ts, workflows.spec.ts)

One-shot boundingBox() reads during inspector-open transitions produced null or reversed coordinates, causing alternating assertion failures.

Fix: await all three operator buttons visible + waitForAnimations before reading geometry; wrap layout assertions in expect.poll to re-measure until stable. In workflows.spec.ts, await the "Trigger event" button visible before clicking it (the inspector may still be animating in after clicking "Trigger: Message Posted").

Rust cancellation test (fake_llm.rs)

h.send("session/cancel", ...) is request-like; releasing the round-2 gate before the cancel ACK arrived let round 2 win the race, producing stopReason: null.

Fix: drain stdout into a buffer until the cancel ACK is received, then release the gate. The cancel token is guaranteed set before round 2 can respond. Buffered messages are replayed through the same classification logic.

Pre-existing biome lint (utilities.css)

Add biome-ignore for dynamic-range-limit (emerging CSS property, Chromium 133+ / WebKit 26+) that was blocking the pre-commit hook for all committers. Unrelated to the flake fixes.

---

Local verification

Rust cancellation test — 25/25 iterations:

test cancelled_turn_with_usage_emits_notification_before_response ... ok  (×25)

Full hook suite passed clean on push (desktop-check, desktop-typecheck, desktop-test, desktop-tauri-checks, rust-tests).

[github-actions]

🔐 Codex Security Review

Note: This is an automated, security-focused review generated by Codex.
Use it as a supplement to human review; false positives are possible.

Scope

• Exact PR diff: 4beffef6979347f7ebdc760705865bdd04d23508...a14d1e3c2afb6725ae4d4d47f162ff94de98fcf4
• Model: gpt-5.6-sol

💡 Click "edited" above to see earlier reviews for this PR.

---

Review Summary

Overall Risk: NONE

No concrete security, correctness, or reliability issues found in the authorized PR range. Changes are limited to test synchronization, a test-only status-expiry hook, and a CSS lint suppression.

Findings

No concrete security, correctness, or reliability findings were identified.

Notes

• Static inspection only; repository scripts, builds, and tests were not executed as required.

---

_{Generated by Codex Security Review |
Requested by: @wpfleger96 |
Workflow run}

[TheSentinel454]

Review feedback at exact head 5a3a393c57e52a621732382332eca9240c213313:

P1 — status-expiry E2E bypasses the production scheduler

desktop/tests/e2e/profile-custom-emoji-status.spec.ts installs a fake clock after app startup and directly invokes __BUZZ_E2E_EXPIRE_USER_STATUS_QUERIES__ through desktop/src/testing/e2eBridge.ts. That bridge calls the cache sweep with an artificial future timestamp, bypassing the production scheduleExpiration → setTimeout lifecycle in desktop/src/features/user-status/hooks.ts.

As a result, the test would still pass if the scheduler stopped arming or firing. The sweep already has unit coverage; this E2E should bind the production lifecycle seam, consistent with TESTING.md:25-31.

Suggested fix: install Playwright's clock before page.goto, seed a status expiring at fixed time +2s, fast-forward the clock, and observe automatic sidebar removal while preserving the unsaved draft. The E2E should fail if scheduleExpiration no longer invokes its timer.

P2 — cancel-ACK assertion is tautological

In crates/buzz-agent/tests/fake_llm.rs, the drain loop can exit only after observing the cancel response, after which saw_cancel_ok is assigned true. The later assert!(saw_cancel_ok, ...) therefore cannot fail and suggests independent coverage that the control flow already provides.

Please remove saw_cancel_ok and its assertion while retaining the ACK-gated loop. The cancellation ordering mechanism itself looks sound.

The workflow polling/readiness changes also look sound; they replace one-shot visibility reads with condition-based synchronization.

[Chessing234]

the expire-via-e2e-bridge + fake clock approach is solid. one hang risk: the cancel-ack drain is now an unbounded loop { recv } — if the agent never acks cancel the test wedges instead of failing the old 40-iter budget. worth a timeout or iter cap around that drain.

[Chessing234]

poll helpers that click on every failed check can toggle the combobox/inspector if isVisible flickers. also the cancel-ack drain has no timeout now — if the ack never shows up that test hangs instead of failing.