docs(review): v0.66.0 clean-room cold review of 2348afe0

[avrabe]

Clean-room cold review of 2348afe0 (release(v0.66.0)), performed with no stake in the release shipping and no PR body/CHANGELOG prose trusted as evidence on its own — every number re-derived directly from the tree.

Result up front

Zero shipped-byte defects. Every finding is in the release's account of its own work, not in a compiler output, encoder byte, or frozen golden.

Lead finding — FALSE

"84 of 89 pinned parity entries (585 of 595 assertions)" contradicts the release's own scorecard fifteen lines above it ("known_open_pins went 104 -> 27 -> 127. It CLOSED 77 wrong-answer pins..."). Independently extracted the KNOWN dict from scripts/repro/selector_parity_197_differential.py at both the v0.65.0 tag and HEAD via ast.literal_eval: the true closure is 77 of 89 entries (567 of 595 assertions), and the true residual is 12 entries / 28 assertions across three issues (#1204 6/15, #1206 1/2, #1215 5/11) — not "5 entries / 11 assertions" as claimed in CHANGELOG.md, RQ-66-BOTHWRONG.yaml, and this PR's own body. #1204 and #1206 are untouched by this release and still compute wrong answers live (confirmed by running the differential against a freshly built binary) — neither is disclosed in the residuals list. Root cause: the artifact conflates "the five issues RQ-66-BOTHWRONG addresses" (82 entries/578 assertions) with "the whole tracked population" (89/595, which also includes #1204/#1206), plus a bare arithmetic slip (595-11=584, not 585).

Also found

• RQ-66-UNWATCHED's artifact status was never updated after its work shipped — still status: proposed, no landed:, verified-by: still pre-work "PENDING" — even though claims.yaml and CHANGELOG.md both describe its four new oracles as landed. This is a second, live instance of the exact gate hole this release itself filed as #1250 (which fixed three other artifacts but missed this fourth). Consequence: only 6 of 8 artifacts actually carry status: implemented — the release commit subject and this PR's own title ("8 artifacts, 7 implemented") aren't supported by the tree today. The fix is the artifact field, not the title — the work genuinely shipped.
• A leftover CHANGELOG.md [Unreleased] section (above [0.66.0], never folded in — confirmed via git show 2348afe0 -- CHANGELOG.md, a pure insertion below it) states a stale emulation floor, 367751, vs. the live 367759 (oracle_wiring_check.py, matching this PR's own body). Exactly explained: ra003_join_1230_differential.py (one of RQ-66-UNWATCHED's new oracles) adds 8, and 367751 + 8 = 367759.
• Two stale comments in claims.yaml (no gate impact — verified against claim_check.py's own logic): the known_open_pins ratchet's prose says "104 -> 27 -> 133" / "+29" in two places, which matches neither the pinned value: 127 nor CHANGELOG.md's own correct "+23"; and "THE POPULATION TRIPWIRE" comment says "six tables (6+2)" where the tree has ten (the machine-checked expect: 12 is itself correct).
• "Five frozen goldens re-frozen" matches no reading — three files, seven distinct pinned-value changes, five in one file alone. The executing-differential claim (flight_seam_differential.py, unicorn vs. wasmtime) was built and run for real and is confirmed genuine.
• RQ-66-PINDEBT's verified-by: is also still pre-work "PENDING" text despite its landed: field reporting the work as done.
• Claim 13's "six merges so far" does not appear anywhere in CHANGELOG.md, this PR's body, or any artifact — it's a stale number in the review's own input briefing, not a release claim, so it is reported as coverage rather than a defect. The "0-line diff" half of that claim is confirmed universally across all 41 recorded merges.

Confirmed clean

The other nine claims (bug-attribution for #1210/#1211/#1215, #1209 memory64 refusal and unmeasured reach cost, the ARM census's load64.wast-alone departure, the memory64-family filename-regex false positives, the pin-count rise disclosed plainly as a headline, the waiver mechanism, mutants_untested, the mutation survey's 27% figure, the VARVE hazard's RECORDED-not-WATCHED disposition, and ARCHMODEL's cut-time re-verification) were independently re-derived and hold. Full detail, evidence, and exact file/line citations in docs/reviews/v0.66-cold-review.md.

Refs #1189 #1210 #1229 #1236

🤖 Generated with Claude Code

https://claude.ai/code/session_01YJK5LZZEkV5smCY1jKn18L

[avrabe]

Superseded by #1253, which carries this review document verbatim plus the corrections it demanded.

Closing rather than merging both, because #1253 must land as one commit: the review's findings and the fixes for them belong in the same change, so that a reader of the tag sees the false claim and its correction together rather than having to reconstruct the order.

This review's findings are not dismissed — every one is acted on in #1253:

• the false "84 of 89 / 585 of 595" headline, corrected to the derived 77 of 103 / 567 of 622 in all five places it had propagated, each now naming its base;
• the understated residual, expanded to 12 entries / 28 assertions and disclosing optimized path clobbers R11 (callee-saved; the direct selector's linear-memory base) without saving it — a direct-routed caller's next load reads through garbage (caller(0) = 1792, wasmtime 7) #1204 and optimized path: loop label placed after the decrement of a pre-loop-defined local — tests/wast countdown(5) never terminates, exit 0 #1206 as live undisclosed wrong answers;
• RQ-66-UNWATCHED still proposed after shipping, now implemented with landed: — so the release title's "7 implemented" is true rather than aspirational;
• the stale [Unreleased] emulation floor and the "five frozen goldens" count.

Thank you for the base-naming diagnosis in particular. "84 of 89" was a real measurement against a base nobody wrote down, and identifying that — rather than just flagging the number — is what made the fix correct everywhere instead of in one file.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!