flake: track remaining dev-watcher and MCP App gate failures

[ScriptedAlchemy]

Summary

Track the two remaining known pnpm check:local-ci flakes after the #110 temp-root isolation fix and the #111 / PR #120 overview reload-frame fix:

1. dev-watcher.test.ts / Chokidar create-event coalescing behavior. Under a loaded full integration pool, a delayed filesystem event can overlap an explicit rebuild and advance lifecycle state sooner than the assertion expects.
2. mcp-app-real.e2e.test.ts polling. The real RSC runtime timeline App test can miss its operation response within the default 1-second expect.poll budget under gate load.

Gate evidence from PR #113

Three serialized local-CI attempts were run after targeted RSC runtime and example verification passed:

• Attempt 1 at f784024a: Node 22 hit the dev-workbench watcher/lifecycle variant; Node 24 hit a runtime-playground provider-identity wait; Node 26 passed. Both failures passed immediately in isolated reruns.
• Attempt 2 at f784024a: Node 22 and Node 24 hit the same watcher coalescing lifecycle assertion; Node 26 passed. The complete 30-test dev-workbench file passed in isolation.
• Attempt 3 at 8463d022: Node 22 hit mcp-app-real.e2e.test.ts :: opens the real RSC runtime timeline App from provider-owned run evidence at expect.poll(...).toHaveLength(1) with a 1-second budget; Node 24 and Node 26 passed. The exact test passed immediately in isolation (14.6s total).

In every attempt, unit tests, package lint/typecheck, the RSC example pool, release checks, and eval:spot passed. These failures are load-sensitive and do not reproduce in isolation.

Acceptance

• Add deterministic synchronization or event de-duplication for the Chokidar create/rebuild path so the full pool and isolated execution observe the same lifecycle boundary.
• Replace the MCP App test's fixed short polling assumption with a bounded readiness signal or an appropriately scaled wait tied to the existing test time-scale mechanism.
• Demonstrate repeated full-pool passes across Node 22.19, 24, and 26 under representative local-CI load.

[ScriptedAlchemy]

Mitigation landed in #138: both tracked flakes now carry per-test rstest retries ({ retry: 2 }) —

• dev-watcher.test.ts :: waits for the real watcher root before reporting create, change, and delete source inputs
• mcp-app-real.e2e.test.ts :: opens the real RSC runtime timeline App from provider-owned run evidence

This only absorbs the load-sensitive failures in gate runs; the root causes (chokidar create/rebuild synchronization and the MCP App polling budget) remain open and this issue stays open to track them per the acceptance criteria.