runtime: fix canonical call context and drop delivery semantics

[lannbot]

Summary

Fix spec-confirmed bugs found through the supplementary Wasmtime async corpus, and adopt the operator-authorized drop-delivery correction from component-model PR719.

• Give synchronous FACT calls and host-boundary realloc their required fresh canonical task/thread context. Preserve logical identity separately from physical JSPI continuation ownership, including deferred admission, nested yield, cancellation, and post-hop trap cleanup.
• Resolve async-start post-return through the post-return table, not the callback table. Apply post-return restrictions only at verified FACT post_return import edges, preserving normal references to the same function.
• Report a genuinely admission-blocked async callee as STARTING without draining its unrelated lock holder. Preserve eager FACT behavior and allow legal core-start synchronous calls without inventing an outer task.
• Accept declaration-only components containing unused blocking imports; validate wrapping against actual constructed entries.
• Deliver stream DROPPED at event consumption after peer drop while preserving progress; retain completed future values and upgrade cancelled future events only when the peer is gone. Assumption pinned to PR719 head35e9769957627c2bee5cd445b998b08b3c652c86; Component Model submodule unchanged. Its larger paired-end refactor is not adopted.
• Avoid O(count) allocation for guest-to-guest payloadless streams. Trap invalid error-context result areas before realloc instead of leaking an internal assertion.

Real translated regressions cover nested/deferred admission and suspension, after-hop trap cleanup, cancellation while waiting, post-return restrictions and import-edge locality. Drop-delivery tests exercise detached original buffers, exact/partial/zero progress, real wait/poll/cancel/drop paths, and max-count packing. Mutation checks proved the delivery and nested-suspension regressions fail without their fixes.

Coverage delta

Supplementary Wasmtime: 29 formerly failed commands now pass:

• 9 declaration-only instantiations
• 17 context/task assertions and cascades
• 3 bounds/count assertions

New census: 88 files / 850 commands; 544 passed (343 setup,201 assertions),128 classified failures,4 deferred-thread skips,174 excluded commands, zero unexpected outcomes or infrastructure failures. Diagnostic-only mismatches remain classified; no broad message equivalences were added. Related gaps remain tracked by #372.

Official corpus: sync-streams:208 now passes. Removed that exact xfail and adjusted aggregate counts; retained WebKit's independent multi-memory cascade overlay.

Verification

• Independent scoped reviews and final integration review: LGTM.
• All recipes of just gates < /dev/null completed across the full run and resumed remaining recipes. Initial lint/type and old test-identity failures were corrected; a later shell run stopped only at the proven stale aggregate count, then just shells browsers smoke-tls smoke-c0 < /dev/null passed after updating it.
• Runtime:1037 passed,0 failed,3 existing ignored. Seeded runtime:1034 passed,0 failed,6 existing ignored.
• Official corpus:1286 passed commands,125 xfail,0 unexpected failures; seeded1283 passed. Existing deferred/profile exclusions remain.
• Supplementary Wasmtime FIFO/seed and public guest gates passed.
• Pinned SpiderMonkey/Node and findings-only Bun match expectations; JSC pinned skipped on arm64. Chromium/Firefox page,worker,shared-worker and OPFS lanes passed.
• TLS/C0 smokes passed with existing not-applicable/translation-only cases. No consumer edits were made; full before/after snapshots were recorded for the main gate run, but the resumed smoke run did not obtain a complete fresh pre-run snapshot.
• Logs: /tmp/opencode/wasmtime-async-fixes-gates.log, /tmp/opencode/wasmtime-async-fixes-remaining-gates.log.

No manifests, protocol marks, or convention goldens changed. Automerge will be armed and watched through delivery.