[steipete]

Diagnostic only; do not merge.

The unchanged parked-worker immediate-GC case added by #129 fails on the signed Windows ARM64 executable from 6aba42d75d, while the prior qualified 807211511f executable passes. Previous native traces show a complete wake/timer/GC-end sequence without a finalizer enqueue.

This draft runs two crossed builds in one native ARM64 diagnostic iteration:

Bun source	Published engine
6aba42d75d	a2976f02 (prior qualified engine)
807211511f	419b5a04 (current engine)

Each build uses the complete checksum-pinned engine manifest from the corresponding source revision. The native job checks the actual executable revision, engine revision, PE architecture, and unchanged regression-case hash. Ten interleaved trials of each crossed build run alongside the original signed executables as controls. No timeout or assertion is changed.

A separate dispatch tests an environment-gated call to the existing clearStackBelowCaller immediately before the asynchronous collection request, comparing the same executable with that call disabled and enabled. The helper and its clearing size are unchanged. A passing diagnostic would support the hypothesis, not prove a stale root or a production fix.

The existing release workflow is temporarily reduced to dispatch-only diagnostic jobs, with no publication or signing jobs. Other build/lint jobs are suppressed on this diagnostic branch. The investigation has a maximum of four ARM64 iterations; this draft will close unmerged afterward. No WebKit source or release is changed.

The first crossed-build attempt stopped at compilation: the coverage/profiler function-range tuple and block-query APIs differ across engine versions. The next crossed attempt uses small, explicit adapters only in CodeCoverage.cpp and JSInspectorProfiler.cpp; the unchanged test runs with coverage disabled. Adapter hashes and the exact modified-path list are retained in build metadata. These adapters do not alter GC, allocation, event-loop, or worker code.

[steipete]

Closing this diagnostic draft unmerged. The native comparison places the regression on the engine side:

Native Windows ARM64 executable	Pass	Original 10 s timeout
Original signed S7	10	0
Original signed S8	0	10
S8 Bun + S7 engine	10	0
S7 Bun + new engine	0	10

Crossed comparison, run 40166390479. Both crossed builds use the complete checksum-pinned engine manifest. Coverage-only adapters reconcile profiler APIs; the test runs without coverage. No GC, worker, or test code changes are used in this comparison.

The final async-clearing comparison, run 40169137156, reuses the same diagnostic executable: clearing OFF timed out 10/10; clearing ON passed 10/10. Original signed controls again produced S8 10/10 timeouts and S7 10/10 passes. All 80 completed trial logs, executable identities/hashes, and engine/adapter manifest hashes were checked. The original test and deadline are unchanged.

The helper's effect supports stack- or timing-sensitive retention but does not identify a collection-time root or an exact causal engine commit. The earlier failing traces complete wake delivery, timer firing, and GC end without finalizer enqueue. A retained dump identifies the marked fixture victim, but contains no pointer into it in the captured stacks or general-purpose/SIMD registers; that later snapshot cannot establish why it was marked. The requested Windows register/stack-capture files and Heap::gatherStackRoots are unchanged across the old/new engine replays and published pins.

The first crossed attempt failed compilation on the profiler API differences. The first async native attempt failed before any trial because of an artifact-directory assumption. Those setup failures are recorded, and their repairs received Codex review. No runtime fix is proposed here. An engine repair needs its own investigation and WebKit release; this draft and its diagnostic branch must not be merged.