flake: RSC runtime generation activation can wedge or go permanently stale after a repair rebuild (Node 24 CI, 2x on 2026-08-29)

[ScriptedAlchemy]

Two Verify (Node 24) failures on 2026-08-29 share one signature: after a source edit, the fixture compiler finishes a successful rebuild, but the Workbench page never observes a new data-runtime-generation — no runtime.generation.activated (and no .failed) ever reaches the client.

Occurrences

1. Main, merge of fix: stop watched-file write races in e2e suites and host message drops in the Runtime App relay #35 (cf02915d) — run 33266629560: runtime-playground-capture.test.ts timed out at 600s with zero output (the capture child had no overall deadline; opaque). Node 22.19/26 passed.
2. PR test: fail the runtime capture fast with phase attribution instead of an opaque timeout #37 branch (same tree + capture watchdog) — run 33271302280: runtime-playground-hmr.e2e.test.ts > activates an edited RSC generation and replays the selected hook without replacing the document failed at expect.poll(generation).not.toBe(failedGeneration) after 120s. Log timeline:
   • 19:43:51 broken write compiles, fails (expected; page shows failure, last-good retained)
   • 19:43:53 repair write lands (staged rename verified on disk by the test)
   • 19:43:54 ready built in 0.44s (widget) / ready built in 0.95s (rsc) — rebuild succeeded
   • next 119s: no second compile cycle, no activated/failed event, page generation never changes

Analysis

The wedge is inside the activation pipeline in examples/rsc-agent-runtime/src/dev/rsbuild-runtime-session.ts (#activate), between the compiler's after-hook and the runtime.generation.activated emit. Candidate mechanisms, in suspicion order:

1. Hang in materializeRuntimeGeneration or #mcpRegistry.prepareActivationReconcile — neither is bounded; a wedged MCP server restart on a contended 2-core runner emits nothing, and the page has no recovery trigger. This also explains occurrence 1: fixture.close() awaits #providerTail, so a wedged activation can hang cleanup, which the capture script (pre-test: fail the runtime capture fast with phase attribution instead of an opaque timeout #37) surfaced as a silent 600s timeout.
2. Spurious permanent supersession — #activationGuard.check() fails if any live attempt with a higher sequence exists in the window between guard.wait() returning and check() running. A no-op attempt (duplicate watch event compiling identical content) settles without raising the cohort revision and without producing its own candidate; if it also classifies cohortChanged=false against the plugin's capturedCohort before the superseded activation's failure callback resets it (rsbuild.config.ts line ~150), nothing ever activates and there is no retry. (Less likely for occurrence 2: the log shows no second compile cycle.)

Suggested next steps

• Bound materializeRuntimeGeneration + prepareActivationReconcile (or make them #closed-aware) and emit runtime.generation.failed on expiry so the page recovers via its existing bootstrap path.
• Make the plugin's same-revision dedup (capturedCohort) confirmed-only (set on activated outcome) so a no-op rebuild racing a failed activation re-captures instead of dropping the cohort permanently.
• test: fail the runtime capture fast with phase attribution instead of an opaque timeout #37's capture watchdog will attribute the wedge phase on the next capture-test occurrence; consider a burn-in of runtime-playground-hmr.e2e.test.ts on Node 24 to reproduce.

Both occurrences were Node 24 only; the same commits passed Node 22.19 and Node 26, and the full parallel integration pool passes locally on a many-core machine — consistent with a contention-sensitive race, not a Node 24 incompatibility.

[ScriptedAlchemy]

Closed-issue re-audit (lane G4) — VERIFIED on main af1c1855/c94df11c, no fix needed.

• Both fix(rsc-runtime): root-cause the generation-activation wedge after a repair rebuild #39 mechanisms survive fix(dev): #75 — completed MultiStats cohorts are the authoritative compile identity, no FIFO hook pairing #83's cohort-identity rewrite: bounded activation phases (examples/rsc-agent-runtime/src/dev/rsbuild-runtime-session.ts L138 defaultActivationPhaseBudgetMs = 30_000 * runtimeTimeScale, L2460–2462 withinDeadline, beforeActivationCommit seam L592/L2498) and supersession only on real-generation ordinals (rsbuild.config.ts L184–197: no before/after hook pairing).
• Flake burn-in, 5 consecutive runs on af1c1855, all green (5 targeted tests each): dev-provider.integration.test.ts :: commits an activation while a later compile is still in flight at the commit check (L1545), :: commits an activation after a later broken attempt fails inside the commit window (L1603), :: collapses an in-flight activation into a compile observed during its guard wait (L1711), :: fails a wedged activation reconcile within the budget and releases its late reservation (L1829), :: emits one owned App reload for changed output and none for an unchanged-output recompile (L280). Companion checkpoint/materializer unit files: 30/30.