async: FACT task.return skips the LiftOptions memory equality check (self-compares)

[lannbot]

From the 2026-08-20 conformance review. The code comment at the site is honest and detailed but is the only record; this issue is the tracking home it lacks.

The divergence

canon_task_return must check the options at the task.return site against the ones the task was lifted with — definitions.py:2389 via LiftOptions.equal (:643-646), which compares string encoding AND memory. For FACT passthrough tasks (task.factPassthrough), runtime/src/intrinsics/async_builtins.ts ~163-171 substitutes the memory half with a self-comparison (opts.memory vs itself, always true): the plan only carries the adapter's view of the lift memory, so a real comparison would false-reject legitimate shapes (e.g. cross-abi-calls.wast).

Net effect: a FACT-path task.return with a genuinely mismatched memory is accepted where the reference traps. wasmtime's own check is one-sided in the same way, but wasmtime corroborates rather than overrides.

Ask

Either:

1. Thread the true lift-site memory through the plan so the check can be real — a descriptor-IR/plan-format contract change (versioned event, orchestrator-owned per the repo protocol); or
2. Adjudicate the acceptance as a named divergence in docs/architecture.md (with the wasmtime-parity argument), and cross-link this issue from the code comment.

Classification from the review: C, high confidence the check is skipped, medium confidence on how much it matters in practice (requires a malformed-but-validating plan or an adapter/lift memory mismatch to observe).