Independent frozen select adapter review — ACCEPT (offline)

Reviewed candidate f977ded63e9bc13d12a3ea66966502a8abbd693e, based on main 606caff1de558f185487aed5e31ede159262ba7e. The candidate is integrated unchanged. This accepts its source/model/host adapter semantics, not an original game capture or a new native qualification. No compiler, native module, guest, test box, public deployment, worker or subagent was used.

The historical result, attempt and identity come from the unchanged instance/export-bound snapshot and oracle. Later live EAX/ESP/EIP have their own batch/time metadata and explicit later-state label. The immutable hook can be observed after subsequent guest instructions. No execution fence, short batches, completion-time fdsets or completion-time live EIP are required or claimed. The accepted native qualification on base main is not rerun.

Source and placement findings

The two production additions are the adapter and the exact runner transformer. Their candidate blobs remain 9ada19e7e0570ae841e4f82cfb4cf9cd7fc4d7b7 and d91571577ae28db2cd696ac068b65d958fe0b1e1. Hook SHA-256 remains b7c77c20b12cd582d6fcfbe246e1f0abdcb644266fc3f161bf5d683d93b13930. No WAT, hook, oracle, scheduler, pacing or input changes occur.

The transform requires runner SHA-256 4d7b9587676db38fcf3314807acb2b2a42b5f09d2c98fcf3dccb8a6cde9e1809 and four unique anchors. Full transform/restore returns exactly the pinned 685960-byte source; changed source and duplicate/missing anchors reject. Setup follows actual instantiation at original line 4850, before ctx.exports and PE/DLL startup. Observation follows normal return at original lines 10363, 10831 and 10876. Main suspension still skips execution and observation; the parallel swallow and serial break paths retain their original guest behavior. Diagnostic observation catches are internal to the observer. A throwing run does not observe. No new wrapper executes or alters guest calls.

The idle shortcut requires started=1 and every other diagnostic scalar zero; the unchanged-park shortcut compares every field. Neither creates a historical result. Changed snapshots go through the original fail-closed oracle. Missing attempts and counter gaps reject. Closure precedes live reads and persistence, so failure cannot replay a previously accepted completion. The diagnostic catch reports failure and disables observation; it does not claim success.

Evidence

Receipts, full bounded stdout/stderr, generated real JSONL, supervisor and resource measurements are retained beside this report in maelstrom-select-frozen-independent-integrate-20261024-evidence/.

Check Outcome
Supplied adapter suite 12 groups pass, including live-result and live-fence baselines rejecting
Unchanged hook/model suite 11 groups pass
Unchanged hook repair suite 16 groups pass
Independent idle/park sequence repeated idle/unchanged parks followed by genuine attempt 2 completion pass
Independent missed attempts unseen initial and intermediate parks reject and close
Independent malformed states fault, capped, wrong API, counter/token/post gap and pending identity change reject
Independent lifecycle invalid batches, failed start, duplicate start, closure and replay controls pass
Actual generated sink two separate instances append two LF-terminated records; both independently parse and preserve result/later state
Actual run fragments sink/readLive exceptions preserve guest continuation, guest exception identity and swallow/break distinction at all three sites
No option/suspension opaque revoked owner untouched; suspended main does not run or observe
Independent rejecting mutation suppress oracle rejection in process only: unseen-attempt regression fails with missing expected counter-gap exception

The independent suite has eight groups and uses the unchanged source evaluator. The rejecting mutation never edits candidate production files. Wrong instance, source, forged/copy/mutated/replayed capture, cap and closed-source cases also pass in the unchanged hook repair suite. No broad swallowed rejection passed these controls. Ordinary sink failures produce a diagnostic and no successful record; persisted evidence is only available after an actual successful write.

Reproduce in this checkout, supplying a new empty writable directory for the independent persisted sink:

node --max-old-space-size=256 test/test-maelstrom-select-frozen-record.cjs \
  /home/user/wine-assembly-shadow-man-span-static-canonical-oct23 \
  /tmp/maelstrom-head-yield-observer-oct24/ops/run.js
node --max-old-space-size=256 test/test-maelstrom-select-completion-hook.cjs \
  /home/user/wine-assembly-shadow-man-span-static-canonical-oct23
node --max-old-space-size=256 test/test-maelstrom-select-hook-repair.cjs \
  /home/user/wine-assembly-shadow-man-span-static-canonical-oct23
node --max-old-space-size=256 test/test-maelstrom-select-frozen-independent.cjs \
  /home/user/wine-assembly-shadow-man-span-static-canonical-oct23 \
  /tmp/maelstrom-head-yield-observer-oct24/ops/run.js NEW_EMPTY_DIRECTORY

Initial review bookkeeping errors are retained: an unnecessary source-root HEAD equality assertion failed (HEAD is d8cd8c54); all six consumed source files were then verified byte-equal to assigned b41a7c5 and their hook pins. The first hook test lacked its JSON pin fixture in the sparse set; adding that unchanged dependency passed. An initial suspected sink escaping problem was disproved by generated-code execution and real LF-framed persistence. None was a candidate defect or access denial. The original failed hook receipt is retained.

Integration and limits

Exclusive shared integration lock covers reconciliation, final checks and publication. Actual remote main was exactly expected 606caff1 and the candidate task ref exactly f977ded6 before integration. The owned sparse repository uses flattened dependency-first read-only alternates, without fetch, clone, repack, object copy or full-tree materialization. All 23 candidate changed blobs are verified unchanged in the integration commit. The added report, independent suite and evidence are reviewed and committed, then relevant suites are run on that exact committed tree before atomic fast-forward publication of task and main. Final SHA/tree/ref/blob receipts reside in the owned ops directory; publication is conditional on those checks. A changed main blocks publication.

The 64 MiB accounting includes checkout/Git/staging/temp/evidence/publication, full external session and launcher logs, positive SQLite growth, future growth and 8 MiB reserve. Installed tools and existing read-only inputs are separately accounted. Cheap execution uses 256 MiB Node heap, 60-second CPU/wall limits, bounded output/capture and process-group TERM/KILL cleanup. The existing exact filename-bearing tree-mount receipt is reused; no mount or guest is executed. Strict dedicated Git SSH trust configuration is preserved.

Runtime arrival, retries/completion in the original game, fdsets/timeval, thread/route state, gameplay, overhead and runtime equivalence remain unknown. Nested DLL-loader runs and worker instances are outside this adapter's three explicit run sites; missed attempts still reject. No new native qualification is inferred from source, model or host-fragment controls. No box is assigned: bx_vwsk6745 is coordinator-reported stopped and unqueried here; hostbx_69aem736 is untouched; lifecycle remains coordinator-owned.