task core: CM#705 reentrance residue — subtask.cancel claim window, mid-hop reads, wrapped-implies-marked (#279)

[lannbot]

Three reentrance.wast rows the retired FACT stub used to mask:

• :837 — canon_subtask_cancel sets has_sync_waiter BEFORE on_cancel()
  for both forms, so a reentrant waitable.join on the subtask traps;
  ours set it only inside the sync blocking branch. Also aligned
  trap_if(subtask.in_waitable_set()) to the reference (unconditional)
  and the two trap texts to the suite's wording.
• :517 — under jspi, subtask.cancel read the callee's state while its
  cancelled thread was still mid-microtask-hop (resolved, but still
  holding the exclusive slot), so the next cancel saw the instance
  excluded and reported BLOCKED. Both forms now park until the callee is
  determinate (threads done or parked) — resolved() alone is not
  sufficient — through one shared park replacing the duplicated async
  branch.
• :429 — $MC imported only b's async-start-call (Suspending-wrapped
  but deliberately not marking its importer), so it had no promising
  entry and the wrap tripped jspi pin (c) with SuspendError before any
  deadlock verdict. Wrapped now implies marked, as a second evidence
  tier (wrapFuncs) that gives guest modules a promising entry without
  promoting FACT adapter pass-throughs (which regressed
  drop-subtask.wast:140 under a seed). intrinsics.md and jspi/bridge.ts
  updated to the new rule.

Conformance 1284/0 (+3); the cm705-reentrance xfail class is retired.

Closes #279.