Severity: low-medium — undocumented divergence, corpus-sound today.
The async form of subtask.cancel parks on a determinacy wait
(intrinsics/async_builtins.ts:493-530) so the BLOCKED/resolved answer
matches the reference's synchronous delivery under the engine's microtask
hop. But while parked, Store.tick may resume any ready thread of the
store — sibling-task effects become observable across a single built-in
call, which no reference schedule produces (definitions.py:2489-2493: the
async form returns [BLOCKED] with no suspension).
The inline rationale is sound for the current corpus (interleaved siblings
were already at block points), but "atomic built-in" is a spec property,
not a schedule choice, and this weakening is absent from
docs/architecture.md §6's divergence list (unlike FIFO).
Work: (1) record the divergence in §6 with the benignity argument
(orchestrator authors), or narrow candidate admission during this park to
the callee's threads if the argument doesn't survive scrutiny; (2) add the
seeded-shuffle regression the review asked for on the adjacent suspicion:
cancel-bracket timing — Task.requestCancellation (task/mod.ts:421-428)
closes the enter/leave bracket around SuspensionPoint.resume, but the
resumed wasm runs microtasks later, and a concurrent host export call
enters synchronously (exec/boundary.ts:1002-1011, 1185) without consulting
resumingThread. No divergence was constructed; pin it with a targeted
DELTIC_SCHED_SEED test across seeds.
Found in the 2026-08-10 adversarial conformance review (runtime + translator vs component-model @ 73b7ad5 definitions.py; wasmtime-environ 47.0.3 treated as trusted).
Severity: low-medium — undocumented divergence, corpus-sound today.
The async form of
subtask.cancelparks on a determinacy wait(intrinsics/async_builtins.ts:493-530) so the BLOCKED/resolved answer
matches the reference's synchronous delivery under the engine's microtask
hop. But while parked,
Store.tickmay resume any ready thread of thestore — sibling-task effects become observable across a single built-in
call, which no reference schedule produces (definitions.py:2489-2493: the
async form returns [BLOCKED] with no suspension).
The inline rationale is sound for the current corpus (interleaved siblings
were already at block points), but "atomic built-in" is a spec property,
not a schedule choice, and this weakening is absent from
docs/architecture.md §6's divergence list (unlike FIFO).
Work: (1) record the divergence in §6 with the benignity argument
(orchestrator authors), or narrow candidate admission during this park to
the callee's threads if the argument doesn't survive scrutiny; (2) add the
seeded-shuffle regression the review asked for on the adjacent suspicion:
cancel-bracket timing —
Task.requestCancellation(task/mod.ts:421-428)closes the enter/leave bracket around
SuspensionPoint.resume, but theresumed wasm runs microtasks later, and a concurrent host export call
enters synchronously (exec/boundary.ts:1002-1011, 1185) without consulting
resumingThread. No divergence was constructed; pin it with a targetedDELTIC_SCHED_SEEDtest across seeds.Found in the 2026-08-10 adversarial conformance review (runtime + translator vs component-model @ 73b7ad5 definitions.py; wasmtime-environ 47.0.3 treated as trusted).