CI flake: AgentRunWorktreeStartTests/testProductionCapturePreservesBoundStoredSelectionInsteadOfCanonicalUI fails intermittently on shard 4

[moonray]

Symptom

RepoPromptTests.AgentRunWorktreeStartTests/testProductionCapturePreservesBoundStoredSelectionInsteadOfCanonicalUI (Tests/RepoPromptTests/MCP/AgentRunWorktreeStartTests.swift:844, assertion at :918) fails intermittently on the Build and Test (app shard 4) CI job, then passes on a re-run / on other PRs.

Failure

AgentRunWorktreeStartTests.swift:918: error: -[RepoPromptTests.AgentRunWorktreeStartTests testProductionCapturePreservesBoundStoredSelectionInsteadOfCanonicalUI] :
  failed - Expected production source capture: The launching tab's Oracle review context could not be captured: The launching selection changed while its frozen review capability was being created.
Test Case '...testProductionCapturePreservesBoundStoredSelectionInsteadOfCanonicalUI' failed (2.433 seconds).

Evidence this is a flake (not a deterministic regression)

• Fails on PR Title fresh Codex threads on the direct streaming path #370, shard 4 — run 28918071953 (job 85789092344).
• Passes on PR fix(settings): persist global defaults across restarts and propagate across windows #418, shard 4 — run 28920242736 (job 85795722053), same test, shard green in 10m52s.
• Passes locally on a warm run.

Same test, opposite outcomes on two PRs built against the same main → non-deterministic.

Likely cause

The failure message points to a race: "the launching selection changed while its frozen review capability was being created." That's timing-sensitive concurrency between the launching tab's selection state and the creation of its frozen review capability — it tips over under hosted-runner scheduling/load but holds on a warm local machine. The test is tagged deterministic_lifecycle_integration, lifecycle owner WindowStatesManager.

Impact

• Blocks otherwise-green PRs: Title fresh Codex threads on the direct streaming path #370 is currently red solely because of this single shard-4 failure (shards 1–3, Style, Provider Tests, Secret Scan all pass), which also delays its re-approval/merge.
• Erodes the CI signal and forces manual re-runs.

Suggested investigation

• Stabilize the production source-capture path around AgentRunWorktreeStartTests.swift:918 so frozen-review-capability creation observes a stable selection snapshot, or quiesce/await before asserting.
• If the race is inherent, make the test deterministically resilient (deterministic setup + explicit await) rather than leaving it timing-dependent.

(Filing without labels — filer lacks label permission on this repo. A maintainer please add e.g. bug / area:agent-mode.)

[moonray]

Aggregating failing tests across a sample of 18 recent failed CI runs (9 on main, 9 on PRs), this test is the top CI noise source:

• 11/18 failed runs — including 4 on main (runs 28793247539, 28777769030, 28750986647, 28746237308).
• Identical failure each time (the frozen-review-capability / selection race).
• Next-most-frequent flake in the sample is 2×; everything else 1×.

So this single test accounts for the majority of recent red CI, and is the main thing keeping PR #370 red. Stabilizing it (or quarantining/auto-retrying it until fixed) would clear the bulk of the noise.

Tracked under the umbrella issue in #425.