Exact select hook: canonical build and controlled native qualification

CONTROLLED_NATIVE_PASS; independent review pending. One compiler process compiled both diagnostic modules; one native process instantiated the tail module once with a fixed 512 MiB shared linear arena and Node heap limit 512 MiB. No original game, test box, download, deployment or main integration occurred.

Accepted hook base: 01c4a60d50afe394a0afc9fda1267e6a508e7821, tree 2a9a16c9dad6eba99c5c79b02c0796b92186a692. Standing canonical source/compiler revision: b41a7c5f4b752d1f0bdab1141a11d4fed4dc6f9a at /home/user/wine-assembly-shadow-man-span-static-canonical-oct23, reused in place. The accepted hook file SHA-256 remains b7c77c20b12cd582d6fcfbe246e1f0abdcb644266fc3f161bf5d683d93b13930.

Actual modules and build

tools/maelstrom-select-hook-canonical-build.cjs verifies the retained 140-file, 21,207,647-byte manifest, all six hook pins and accepted transformer hash. It uses watxSourceClosure / compileClosure from the standing compiler. Only the accepted transformed winsock and dispatcher text replaces closure VFS entries; all f, src/f, ./f aliases agree. The single src/main.watx entry reaches 128 files, including both transformed files. Compiler byte-to-text decoding is preserved, including ASCII comment normalization; raw transformed and compiler-text hashes are separately recorded. No source was materialized or modified.

Options remain production, standardWat, no runtimeBuiltins, strictDeclarations, tailCalls true/false, no names, shake or extra sources. Compatibility uses the unchanged accepted dispatch-trampoline overlay. Both modules receive the original wine-region-layout stamping recipe and pass synchronous native Module validation with the unchanged historical 8 MiB constructor guard preloaded. Guard SHA-256: 47ce2e2b835be33eab6b2b24f027ecaab8ceb61c559890e51548c78199beb197. Compiler loader SHA-256: d04ed1d5ca74fc1b6f10357576a38e726d58f0ad908ccf34a29de35052aa1a33; closure helper SHA-256: 60d2f039abeb4a940405c885cb507ee98649eb3659969580c336a7302e5bb86c. Full compiler stages, helper, closure and transformed input seals are in build-inputs.json; compiler-produced full inventories are in build-outputs.json.

Static acceptance and controls

The unchanged accepted static helper independently reconstructs all 269 storage regions and named spans from the source and region mirror, including DIRECT_WINDOW. Its serialization derives 9a3bc9149d58209c; the declared fingerprint is not used as independent proof. Full source imports and signatures (291, including memory), 2,245 function exports/signatures, memory import/export, noStart and exact custom section payloads match both binaries. Seventeen getters are () -> i32; msc_start is () -> (). Both modes add exactly eighteen private mutable i32 zero globals compared with their respective sealed canonical baseline. Memory and data sections are byte-identical to the original; no host import or guest storage was added. Compatibility's original two trampoline globals are preserved.

Six binary-record controls reject a new host import, changed getter signature and changed layout payload for each module. Four focused emitted-source controls reject completion before the actual EAX store, park-as-completion, missing EAX store and erased completion. Two native-evidence controls reject park completion using stale EAX and a completion retaining pre-store EAX. These are rejecting source/record mutations, not additional native mutant executions. Accepted implementation/source/binaries remain unchanged.

Actual native acceptance

Cheap preflight checks exact signatures, narrowly pins the two retained native initializer recipes, verifies disjoint code/return/stack/fdset/timeval fixtures, and rejects a deliberately contaminated counter fixture (972 remains a count, never an execution inference). Every actual fixture rechecks fresh counter values [0,0] before its run. The exact native mount argv and intended mount-file list are empty; the retained canonical-parser receipt for that unchanged configuration also covers ordinary and comma-bearing control filenames. Trusted SSH configuration is retained verbatim in configuration.json and checkout.json.

Fresh admission and exclusive /tmp/wine-ops-heavy-runtime.lock precede the sole native process. A synchronous guarded Module and real Instance run actual production dispatcher/handler code, with a native socket allocation and a documented EOF-ready socket-state fixture. JS does not implement select, its dispatcher, readiness filter, stack handling, result store or diagnostic globals. Only lifecycle notifications, controlled get_ticks, and the socket teardown notification are supported host imports; unexpected behavioral imports throw. WSA is inactive, so the real vsock_pump gate returns without network work.

Seven scenarios run OFF and then ON, using disjoint fixtures on the same instance:

Twenty real run(1) calls include three pending-yield control invocations per lane. Each active case increments actual API dispatch count once. Pending controls add no API call, host call, register change or diagnostic completion. One-block budgets prevent the subsequent return block from executing; its stale-result sentinel and fresh return counter are checked separately. Counters alone are not called execution proof. Normalized OFF/ON registers, ESP/EIP, fdset contents, yield/halt state, original host-call sequences and cache deltas match exactly. No diagnostic operation clears yields, advances clocks, pumps sockets or writes guest memory. Fixture setup and scheduler-resume setters are explicit test actions.

The accepted snapshot(e) and createOracle(instance,e) capture authentic native exports at safe synchronous boundaries. They accept park token 1 from thunk entry, park token 2 from inline CALL and positive completion token 3 from inline JMP. The final native snapshot has API 934, thunk 0x75040f8, thread 1, compiler-placed REGFILE 86228992, saved return 4784672, preESP 121709056, postESP 121709080, two parks, one completion, result 1. Pre-run EAX was 0x5a170123; the diagnostic result matches the actual production EAX store rather than that stale value. Forged copies, stale authentic captures, wrong instance, wrong export source and reused authentic completion reject. The first completion remains frozen through later actual calls. See native-result.json and native terminal/capture receipts.

Limits, resources and recovery

This qualifies controlled production pathways only. First-completion freezing means the later zero returns are behavioral checks, not zero-result diagnostic captures. Actual same-frame retry completion, exceptions within select, unsupported direct test_call_select after start, compatibility instantiation, unobserved intra-invocation properties and overhead remain unqualified. No original-game/API-result causality is established; the original-game runtime gate remains unresolved pending coordinator review and an authorized runtime task.

Initial precheckout admission passed with 62,750,720 projected bytes of 67,108,864, including current full WAL, session/launcher files, positive SQLite growth, publication, future growth and 8 MiB reserve; memory was 3,186,044,928 bytes against 2,684,354,560. Every execution/publication stage remeasures. Completed materialized bytes replace their remaining-growth projection rather than being double charged. Installed Node/compiler and standing readonly inputs execute in place. Both actual modules are explicitly retained in this evidence directory as well as owned ops.

Compiler limits: CPU/wall 900 s, oldspace 2048 MiB, aggregate output 524,288 bytes. Native: CPU 90 s/wall 60 s, heap/arena 512 MiB each, output 131,072 bytes. Cheap commands: CPU/wall 60 s, Node heap 256 MiB. All supervisor subprocesses have a 24 MiB regular-file limit, subreaper and process-group TERM/KILL/reap within 5 s. Compiler elapsed 5.205 s, peak child RSS 461,602,816 bytes; native elapsed 0.048 s, peak child RSS 143,114,240 bytes. Both exit 0 with no surviving group.

Retained bookkeeping failures: the first static checker compared compat globals against the tail baseline; it was corrected to the exact sealed compat baseline. The first focused ordering checker expected one extra closing parenthesis; only that checker was corrected. Original failing captures remain. The first timeout board watcher lacked an explicit regular-file guard and was terminated; its log and correction receipt remain charged. Guard identity was recorded after compiler dispatch, although the unchanged guard was preloaded before the driver. Read-only receipt searches included missing paths and truncated inspection output; those session/launcher bytes remain charged. No compiler/backend fix or native retry. The previous worker's failed memory admission and all prior receipts remain in their original paths; this task does not erase those charges or limitations.

The owned root is /tmp/maelstrom-select-hook-recovered-canonical-native-oct24. ops/supervise.py and account.py are retained here for recovery, with commands in terminal receipts. To reproduce in a separately authorized task, restore these ops scripts beside its isolated repo, use the standing inputs in place, take fresh admission, and run the focused build/static/preflight/native drivers. One-shot attempt receipts forbid unchanged replay in this task. Actual execution finished before 2026-10-25 00:15 UTC. No box was assigned or queried; bx_69aem736 is untouched, state/expiry unqueried, lifecycle and shutdown remain coordinator-owned.

The task branch is task/maelstrom-select-hook-recovered-canonical-native-oct24. External ops retain final publication/ref/blob seals, resource and cleanup receipts; the append-only repository messageboard records the actual commit/remote branch. Root tracked files/index/refs and all donors are readonly. Claude idle. Stop after publication; independent review is next.