File the upstream findings (IROH-1; component-model tracker)

[lann]

Findings against foreign repos are tracked in-repo and await filing (operator's task — they target repositories this project doesn't own):

• upstream-consumer-findings.md: IROH-1 — polymorph-iroh's endpoint holds a RefCell borrow across a wit_bindgen::block_on yield in the TLS CertificateVerify signing path; any concurrently-parked task then panics (already borrowed → unreachable trap). Guest bug on any host; near-deterministic repro under deltic's scheduler (see also Re-examine the JSPI bridge exclusivity divergence against definitions.py #5). File against polymorph-iroh with the fix sketch (take the borrow inside drain's inner steps, or move signing out of the borrowed region).
• upstream-component-model-repo-findings.md: the standing spec/reference findings (CM-3 cancel-write supersession, CM-4, NOTE-1, vestigial $async? on resource.drop, dead canon_backpressure_set in definitions.py, and the rest of the tracker) → WebAssembly/component-model issues/PRs per the tracker's own filing notes.

The tracker files remain the source of truth; this issue just makes the filing task visible.

[lann]

Both drafts are now filing-ready (commit 953cd5b): CM-4 gained the spec-history review, the no-CI-cross-check structural finding, and the wasmtime determinism evidence (exams/wasmtime-exclusivity/); IROH-1 was rewritten with the resolution-boundary semantics — pre-resolution borrows safe, post-resolution blocks are interleaving points per the suite-pinned semantics, wasmtime green by deterministic schedule rather than guarantee. Recommend filing CM-4 first and referencing it from the IROH-1 issue so the iroh fix discussion doesn't relitigate the semantics. (#5 closed into this.)

[lann]

Refinement to the IROH-1 draft (2a236f0), from the operator's question "does wasmtime deterministically avoid this for all programs or only the suite?": only the suite (closed systems — no clocks/IO, interleavings forced by rendezvous). For real programs the block-window's runnable set is fed by wall-clock inputs, so wasmtime's green endpoint row is timing stability (native sign ≈ zero-width window vs deltic's crypto.subtle Promise hop), not determinism and not avoidance. The draft now also notes the shelter erodes under iroh's own roadmap (non-extractable keys ⇒ slower async signers on every host) and proposes the falsifying repro: ~1 ms of host-sign latency under wasmtime should reproduce the collision there.

[lann]

CM-4 is now filing-complete with executable artifacts (see the commit just pushed): exams/wasmtime-exclusivity/cm4-run-tests.patch (the asked-for obviously-reasonable diff to the reference's own test file — fails at the suite-pinned expectation under pristine definitions.py), cm4-reference-fix.patch (the resolution-scoped gate in the reference's model — and the stronger finding it exposed: test_callback_interleaving's second progress-free poll window encodes the hold-semantics, so the spec repo's unit tests and wast suite pin OPPOSITE semantics), root-cause.md (full trace), verify-cm4.sh (all four legs from pristine copies). Suggested filing shape: lead with the two-corpora contradiction + the no-CI-cross-check structural gap; attach both patches; note the harness's hang-after-failure property so reviewers run under timeout.

[lann]

CM-4 is adjudicated (2026-08-10): upstream test defect — sync-streams.wast:145 overfits wasmtime's deferred-entry scheduler policy; no reference/spec change proposed. The filing kit (distilled finding, proposed schedule-agnostic test fix, CI suggestion, repro pointers) lives on #43; the in-tree exam kit was retired by #46 (archived at 4f3351f:exams/wasmtime-exclusivity/). Tracker entry: upstream-component-model-repo-findings.md CM-4.