test(desktop): await Plan turn settlement

[liugddx]

Summary

Stabilize the Plan/catalog-refresh Desktop e2e scenarios by waiting for Runtime Host's authoritative terminal turn ownership before latching the next session-catalog read.

• adds one shared barrier that requires runningTurnIds to be known-empty;
• applies it to the four Plan scenarios that first create a session through the fake backend;
• preserves the existing pending refresh, rejected refresh, queued latest-intent, and deletion-race assertions.

Why

The fake backend response can become visible before Runtime Host publishes terminal turn ownership. The tests previously treated that text as a completion barrier, then observed the Plan row enabled briefly before a late running projection disabled it between the assertion and click.

This reproduced on two unrelated #3789 heads:

• run 32867280955 timed out on the second Plan click in the latest-intent scenario;
• run 32870919058 timed out on the first Plan click in the committed-state scenario.

The new barrier uses the same known-empty runningTurnIds authority already used by the Desktop streaming e2e suite. It runs before sessions.list is latched, so the latch still owns only the Plan commit's catalog refresh.

Verification

• npm run lint
• npm run format:check
• npm run build
• npm run typecheck
• npx knip --workspace apps/desktop
• git diff --check
• Playwright discovery compiles and selects the focused scenario correctly.
• Hosted test passed in 9m11s, including the complete Desktop e2e suite.

Local Windows Electron did not create its first test window, so the focused runtime assertion requires the hosted Linux/xvfb CI lane. The failure occurred before the test body and is not counted as passing evidence.

AI use

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

Tool and scope: Codex analyzed the two hosted traces, added the test-only settlement barrier, and ran the listed local gates.

Fixes #3814.

Checklist

• Tests cover the timing boundary that failed in hosted CI
• Lint, format, build, typecheck, knip, and diff checks pass locally
• Hosted test passes on exact head e62c5756e

Does this PR entail a change in behavior?

• Yes
• No - this changes only Desktop e2e synchronization.

[liugddx]

@Astro-Han Hosted exact-head test is green, including the complete Desktop e2e suite. This PR keeps the existing catalog-failure assertions and adds only the authoritative runningTurnIds settlement barrier before the latch. Could you please review when available?

[Astro-Han]

Update on cf56ebb: GO — test-only stability fix waits for authoritative terminal state, hosted test green.

简体中文

仅测试，无阻断。

[liugddx]

Fresh exact-head est is green on �62c5756e after rebasing onto main=8ff9528eb, including the full Desktop e2e suite. #3814 is ready for maintainer review and merge.

[liugddx]

@Astro-Han Current exact head �62c5756e is rebased onto main=8ff9528eb. Hosted est run 32918399089 is green, including all Desktop e2e tests. The barrier is test-only and #3814 remains open until this PR merges.

[Astro-Han]

I reviewed this head and found no blocking issues.

Rebase-only change from cf56ebb; test-only Plan turn settlement still waits for authoritative empty runningTurnIds before latching catalog. Hosted test: SUCCESS (32918399089).

No P0-P3.

简体中文

该头无阻断。

---

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.