Why the merge train was dequeued

PR #24053 (merge-train/spartan-v5 → v5-next) was dequeued from the merge queue because its merge-queue-heavy CI run failed. Tracing the failure:

• CI3 merge_group run 27568575541 → shard x4-full (ci-full-no-test-cache) failed.
• The failing test was yarn-project/world-state/src/native/native_world_state.test.ts (code: 1), specifically:

  NativeWorldState › Pending and Proven chain › does not fail when a delayed-close fork is destroyed by a reorg before its close fires

• Failing assertion (native_world_state.test.ts:966):

  expect((ws as any).instance.queues.has(forkId)).toBe(false);  // Expected: false, Received: true
  

  The warnSpy assertion on the preceding line passed; only the queue-cleanup assertion failed.

This test and the world-state instance/facade code it exercises are not modified by this train — the test already exists unchanged in v5-next. So this is a pre-existing timing flake that happened to fail in this grind and blocked the train.

Root cause

The delayed-close path is correct and always eventually cleans up:

• asyncDispose() schedules sleep(closeDelayMs).then(() => close()) (fire-and-forget).
• close() → doClose() → instance.call(DELETE_FORK), and the per-fork queue entry is deleted in call()'s finally regardless of whether the native side rejects with Fork not found (which it does here, because the reorg/unwind already destroyed the native fork).

The test then waited a fixed sleep(closeDelayMs * 3) (3s) before asserting the queue was gone. Under a loaded merge-queue grind, the scheduled close (fired at +closeDelayMs) plus its async DELETE_FORK round-trip and queue.stop() can take longer than that fixed margin, so the queue entry was still present when the assertion ran — Received true.

Fix

Replace the racy fixed sleep with a deterministic wait on the exact asserted condition:

await retryUntil(() => !(ws as any).instance.queues.has(forkId), 'delayed-close fork queue cleanup', 30, 0.1);

This polls until the queue entry is removed (the cleanup runs inside the same close path), with a 30s timeout. It preserves the regression intent: if cleanup genuinely stopped happening, retryUntil throws a clear Timeout awaiting delayed-close fork queue cleanup instead of a bare toBe(false). The production code is unchanged; no test is skipped or weakened. The now-unused sleep import is swapped for retryUntil.

Verification

This is a test-only, single-statement timing change. I was unable to run ./bootstrap.sh ci in this environment — the workspace has no node_modules and no barretenberg/native world_state_napi build artifacts, so exercising this native test would require a multi-hour C++ + TS build that isn't proportionate to a wait-deterministically change. The change is type-checked by inspection: retryUntil(fn, name, timeoutSec, intervalSec) matches the foundation signature, and the import path (@aztec/foundation/retry) is the one used across yarn-project. CI on this PR will validate it.

Created by claudebox · group: slackbot