Census-fixed callback native acceptance — 2026-10-20

Historical source last updated, 2026-10-20. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

Tested source: 1d58b6c5497d6315976ca9d506f66c84612681cc. Fixture: 2c902c79f362560112ca30d41e4332688046d55a, with only test/test-edit-winhelp-callback-native.js, test/fixtures/edit-winhelp-native/validate.js, and docs/re-notes/edit-winhelp-native-fixture-20261020.md overlaid. No JS COM candidate overlay was applied.

The reused /home/user/callback-native-fullvm tree initially matched all 6,430 prior source blobs plus three fixture blobs. Only the three-file reviewed census commit delta was applied. All 6,432 exact candidate Git blobs and three fixture blobs matched before the build, after the build, and after all runs. Complete staged manifests are retained. This candidate matched remote main at the initial verification; later main changes are not qualified by this run (see main-final.txt and the final receipt).

The canonical bash tools/build.sh passed all gates, compilation, and data overlap checks, producing a 1,874,960-byte module: a50b56a2dfc26174231cde9492b8ab2fb715eaecad6fd8e9e5b8445b3c88e361. The build directory was empty before this build. The older 8f6 module was not used. Environment: Linux x86-64, Node v24.18.1, V8 13.6.233.17-node.50; full versions and uname are retained. Existing pinned dependencies were reused; package-lock, installed lock hashes, and the preceding successful npm ci --ignore-scripts log are retained.

Actual VM scenario Result Observed acceptance
edit-ordinary PASS Real callback entries, text/result, registers/ESP, second call, 70 successful siblings
edit-nested PASS Outer and nested restoration, text/result, second call, 70 successful siblings
edit-error PASS Error 123, zero result/empty text, restoration, 70 successful siblings
edit-timeout PASS Callback entered and timed out, error 1, restoration, 70 successful siblings
help-ordinary FAIL Public API result 0 rather than 1; no DLL entry recorded
help-nested FAIL Public API result 0 rather than 1; no DLL entry recorded
help-timeout FAIL Public API result 0 rather than 1; no DLL entry or timeout marker recorded

All seven guest processes exited normally with code 0, no signal or parent process error. The three WinHelp harness processes exited 1 on the assertion. Their observations show registration accepted (1), caller completion, 70 sibling attempts but zero successful siblings, and zero DLL entry, argument, and computed-result fields. Therefore these runs do not qualify WinHelp callback restoration, nesting, or abandonment. The cause is unresolved; no failure diagnosis or repair loop was undertaken. Their raw per-case receipt leaves its case verdict PENDING because the assertion threw; its overall status is FAIL. replay.js independently reproduces those failures from saved observations without launching a VM.

The fixture's separate offline validator passed: seven generated scenarios with 21 disassemblies, seven argument negatives, 14 ABI negatives, 14 missing-evidence negatives, seven synthetic positives, and 84 synthetic oracle negatives. These checks are retained separately and are not native results.

Execution ran 12:45:18–12:45:48 UTC on sole runtime bx_6czhvg6m. Build wall bound: 600 seconds plus 10-second kill grace. Each native runner used 120 seconds / 4,000 batches / batch size 1,000, a 150-second fixture parent bound, and a 170-second outer bound plus 10-second kill grace. No wall bound fired. Each scenario ran once, including the remaining independent scenarios after a failure; none was retried.

Self-contained evidence is in scratch/runs/20261020-callback-native-census-fixed/: the actual module, all seven generated PEs/DLLs/overlays, raw logs and receipts, offline controls, canonical build log, exact manifests, source delta, oracle source snapshots, and reproduction commands. The archive and all 176 recovered file hashes were independently verified; 11 snapshot Git blobs were checked against the manifests. python3 verify-evidence.py checks integrity, and node replay.js rechecks all seven saved observation verdicts without runtime access. To rebuild, recover the listed source blobs from the exact Git revision on a build host, overlay the three fixture blobs, install the pinned lock, then use execute.sh with fresh evidence directories. Full source/assets remain staged remotely; the local bundle contains manifests and the relevant snapshots, not another full checkout.

All 203 preserved Army/ICM evidence/module/archive hashes matched after execution. Evidence was recovered and verified before runtime release at 12:47:26 UTC. The box was left running with TTL 15:53 UTC. No runtime work remains queued by this worker.

Final evidence and sparse checkout occupy under 8 MB; free disk remained above 2 GiB. A staging mistake briefly created an 81 MB local full-source archive, violating the 20 MB task cap; it was immediately removed and only the compact delta was transferred. The disk floor was not crossed. SSH transfer was denied with 403; permitted Boat execution API transfer was used instead, without altering permissions.

Existing fixture limits remain: no exhausted-capacity, wrong-token/epoch, spill/SEH, asynchronous-resumption, slice-64-success, mixed-nesting, or native Windows comparison. This does not qualify Circus gameplay, JS COM callbacks, or whole-jump readiness.