🤖 tests: flaky TaskService waitForWorkspaceTurn completion-race test times out under full-suite load

[ThomasK33]

Summary

TaskService > waitForWorkspaceTurn handles completion racing with waiter registration (src/node/services/taskService.test.ts) fails intermittently in the full make test-unit run. It fails with:

error: Timed out waiting for workspace turn
      at <anonymous> (src/node/services/workspaceTurnManager.ts:2683:38)
(fail) TaskService > waitForWorkspaceTurn handles completion racing with waiter registration [108.00ms]

Evidence

1. Full-suite run on 2026-09-23: Bun 1.3.5, Linux x64, BUN_JSC_useWasmIPInt=0, stock GIT_TEMPLATE_DIR. Result: 20,108 pass, 9 skip, 1 fail. This test was the only failure.
2. The same test passed 3/3 when run in isolation (bun test src/node/services/taskService.test.ts -t "waitForWorkspaceTurn handles completion racing").
3. The branch where it failed (🤖 feat: let agents discover selectable models with models_list #4346) does not modify taskService.ts or workspaceTurnManager.ts.

Suspected cause (unverified)

The failure took about 100 ms, which suggests the test passes a very short timeoutMs. waitForWorkspaceTurn arms that timer before its asynchronous getWorkspaceTurn read settles. Under full-suite load (event loop and I/O contention), the timer can fire before the completion or record read is observed, so the test measures scheduling latency instead of the race it intends to cover. A deterministic fix would synchronize on the completion/registration signal rather than a wall-clock timeout.

Acceptance

• The test deterministically exercises the completion-vs-registration race without depending on a short wall-clock timeout.
• The test is stable under load, for example bun test --rerun-each=50 on the file and a full make test-unit.
• Product behavior changes only if a real race in waitForWorkspaceTurn is found.

---

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high • Cost: $255.89

[linear-code]

MUX-170