Runtime App teardown acknowledgement lost under slow Node 26 CI runners

[ScriptedAlchemy]

Recurring CI failure (Node 26 Verify on #18, #20, #22; also one Node 22 run): packages/workbench/tests/mcp-app-real.e2e.test.ts > opens the real RSC runtime timeline App from provider-owned run evidence fails at line ~1362 with Matcher did not succeed in 60000ms on expect.poll(thirdAcknowledgement, { timeout: 15_000 * timeScale }).

Sequence: the third Runtime App binding is created, window.location.hash = '#mcp' navigates away, the ui/resource-teardown request is observed reaching the third frame, but the frame's JSON-RPC acknowledgement never arrives. Passes locally with CI=1 time scaling, and the same flow passes on Node 22/24 most of the time.

Hypothesis: on slow runners the sandboxed App frame is unmounted by the navigation before it posts its teardown response, so the lifecycle owner waits on an acknowledgement that can no longer be delivered (the design's teardown contract awaits the exact ack before DELETE). If so this is a race in the close path, not a budget shortfall - raising the wait will not fix it. Worth a deterministic repro that delays the frame's ack past unmount.

Evidence: run 33228321926 job 99036290040 (attempt 1), annotations and log excerpt in the job.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EMWczsXAkj7fC5ssSxGK43

[ScriptedAlchemy]

Closed-issue re-audit (lane G4) — VERIFIED on main c94df11c, with one test gap filled.

• Relay fix from fix: stop watched-file write races in e2e suites and host message drops in the Runtime App relay #35 (cf02915d) is intact: packages/agent-bundle/src/dev/runtime-client-surface-proxy.ts queues host→App messages during the handshake (maxPendingHostMessages = 32, L204; queue at L334–343; flush on ui/notifications/initialized at L355–359). Teardown-ack budget is gracefulTeardownTimeoutMs = 10_000 with the teardownAckTimeoutMs test seam (packages/workbench/src/mcp/runtime-app-bridge.ts L60, L133, L564–569); no commits touched the bridge since fix: stop watched-file write races in e2e suites and host message drops in the Runtime App relay #35.
• Flake burn-in: packages/workbench/tests/mcp-app-real.e2e.test.ts :: opens the real RSC runtime timeline App from provider-owned run evidence (the failing test from this issue) ran 5/5 green on af1c1855 (prebuilt integration pool).
• Gap: fix: stop watched-file write races in e2e suites and host message drops in the Runtime App relay #35 covered the relay queue only through that e2e — no unit proof existed. PR fix(docs,test): closed-issue audit G4 — stale Claude cwd/preview-peer docs, restore rendered-skill docs, pin the #23 handshake queue #368 (merge c94df11c) adds runtime-client-surface-proxy.test.ts :: queues host requests relayed during the App handshake and flushes them once the App reports initialized, driving the served shell deterministically: early/teardown requests queue, foreign-origin and non-RPC traffic is still rejected, the queue is bounded at 32, everything flushes in order on initialized, post-init traffic relays immediately. 27/27 in the file; CI Verify/Release/Examples green.