GTA2 Worker COMPLETE confirmation repair
Status: implementation published on main as
6d8b95c2cb4403b71d8e6ed2a21222d97fd3d12c; complete published-source
main/Worker qualification PASS. The single subsequent original software attempt
failed native USER health at bridge installation before capture or input. Its
endpoint is permanently closed. No WebGL attempt or box was launched.
Evidence: scratch/runs/20261116-gta2-worker-complete-confirmation/.
The predecessor report is 5eeacc6c064cc207817d365b8ab30566e8c111b9.
Its main baseline completed at 335/422/379/383 ms. Worker baseline stalled on
module recovery at 500 ms; its candidate ARM at 466 ms did not COMPLETE.
Independent expiry/signaling/quiescence were 507/508/509 ms. Both old endpoints
remain permanently closed. Exact inherited source pins and relevant binary
patches were retained before this task's edits.
Heavy work began only after explicit GL and Sims QI final releases, personal hash/wait/PID/group checks and free locks. GL: 959 direct references, 166 process and 65 thread waits with both EOFs, 176 PID/group absences. Sims QI: 2,635 indexed artifacts, 487 actual waits with both EOFs, 684 PID/group absences. A temporary handoff allowed the coordinator's remote preservation merge; its 14 actual waits and both EOFs, artifact hashes and current PID/group absence were personally checked. Heavy work resumed on published main c1ccb2eb3b8fa241ff688a622873d194a6f85806. Foreign WIP, including guest-thread-host364bbcd1, is not adopted.
The controller now starts its independent deadline service before asynchronous configuration IO. Confirmation and the original parent final checks overlap; every check still joins before COMPLETE and physical DOWN. Event notifications replace repeated polling. EOF/cancellation is checked again after confirmation. A dedicated phase pipe retains actual observations if the watchdog kills the child. After an independently delivered UP, a read-only original-renderer witness records physical UP before the unchanged missing-schedule refusal. The fixed 500 ms COMPLETE, 1,750 ms independent UP and 3 s aggregate limits are unchanged, as are owner, nonce, process/thread generation, renderer, source/module and physical-input predicates.
The independent child's loopback transport uses builtin HTTP/socket IO and bounded buffered WebSocket text framing. It verifies the upgrade and refuses invalid masking, UTF-8, continuation and control frames. It recovers and hashes all current module wire bytes on every gesture; no native/module observation is cached. Ordinary parent transport behavior remains unchanged.
The new owned verified-source-view preload validates each immutable Git blob, keeps its bytes private, and records its SHA-256. Each final source-pin invocation compares every immutable digest in that same verified view after freshly reading and checking the preload's current and exact executed source provenance. Owned mutable helpers, the preload itself, modules and run inputs always use fresh IO and hashes; immutable-view identity changes and missing provenance refuse. This removes repeated immutable source copies/hashes without removing a final pin predicate. All original final-check invocation sites remain. The view is prepared in a sealed transient memfd; source recovery is Git SHA plus exact binary patches, never a source-tree or WASM snapshot.
Actual candidate progression:
- candidate-worker-1: first COMPLETE 453 ms. Next gesture's 3 MB module response arrived at 499 ms; no ARM/DOWN, independent expiry 505 ms. Permanently closed.
- transport-worker-1: COMPLETE 435/481 ms for the two modals. Third child confirmed at 473 ms but the parent's joined checks crossed 500 ms; no third DOWN. Permanently closed. This evidence motivated the immutable-view change.
- immutable-worker-1: four actual COMPLETE receipts at 232/366/235/204 ms; two personally reviewed visible native modals, public DOWN1/UP3, all physical keys UP and native ExitProcess(0).
- immutable-main-1: four COMPLETE receipts at 173/243/168/142 ms, same visible modals, public DOWN1/UP3 and native ExitProcess(0). An expired initial review was refused locally before a command; a fresh full capture was reviewed.
Every authored input above followed personal inspection of its current raw PNG, backing, logical, displayed and public GetPixel evidence. Genuine native DLL lifecycle evidence is retained. All completed fixture endpoints were closed and actual child waits/both EOFs retained. Fixture native ExitProcess is distinct from subsequent browser host teardown and does not qualify original GTA2 Quit.
Focused checks pass: 18 standalone authored controls, current source guards, bridge/page/refusal models, actual preinitialization refusal controls, seven real-socket transport cases and seven immutable-view provenance/fresh-input controls. The actual compiled deadline service killed a blocked victim with expiry/signaling/quiescence 501/502/504 ms and zero input requests. Initial and final shared and HEAD-plus-whole-owned full flock builds pass. The existing actual paired-controller protocol controls pass. In the actual Worker stalled-parent-after-DOWN control, independent physical UP was witnessed 1,329 ms after DOWN while the parent remained blocked. Missing-schedule refusal remained mandatory, followed by local replay/rebind/after-UNKNOWN refusal with zero later wire requests. The complete published main/Worker live fault checklist now passes, as detailed below. After failure the test wrapper attempts replay/rebind/later requests through the same terminal API and requires local refusal with unchanged actual wire logs.
Published-source qualification uses the same committed implementation and freshly
verified HEAD module SHA-256
89dcaf9ce39a950edb36ab12301d96733be5b17b269f0d1f5f7599953885f37c.
Four actual COMPLETE times are main 139/114/126/81 ms, Worker
267/490/232/206 ms, each including independent postarm confirmation and both
original final source-check invocations. Every immutable source ledger records
two final checks per release child. Both fixtures retain genuine COMCTL/COMMCTRL
lifecycle, two personally reviewed VISIBLE native modals, public DOWN1/UP3,
all physical keys UP and native ExitProcess(0). The slowest Worker result has
only 10 ms margin; these results do not promise universal runtime timing.
All 50 published native fault cases pass (24 main, 26 Worker): stale review, missing/changed source, wrong generation, source/module changes before and after ARM, owner/renderer/handler/queue/context/retired owner, focus/reload, duplicate ARM, bind after ARM, early EOF, late ACK, child start failure, parent cancellation, stalled parent before ARM and after real DOWN; Worker also tests retired Worker owner/endpoint. Actual terminal API replay/rebind/after-UNKNOWN calls all refuse locally with unchanged wire logs. Only the two intentional after-DOWN stalls send input; real independently witnessed physical UP is 1471/1428 ms for main/Worker, with the unchanged missing-schedule refusal. The matrix retains 396 actual wait/both-EOF receipts and 394 recorded current PID/group absences. Actual late-ACK expiry/signaling/quiescence is main 501/502/503 ms and Worker 501/501/503 ms. Worker reload quiescence is 579 ms after expiry at 501 ms; the COMPLETE deadline remains 500 ms and no input is sent. The Worker-owner child had already exited when the watchdog expired, so its receipt is not represented as a watchdog signal. All causal failures and raw phase logs remain.
The single ordinary original software attempt followed the completed published
qualification. Its 900 s launch limit, independent 920 s watchdog and 180 s Quit
reserve were prearmed; 71 existing media files were hash-verified without copies.
The actual native Worker bridge's initial snapshot() failed with
Error: unhealthy native USER context during binding installation. This occurs
before the startup capture or command/input loop. No authored input, nonce or
release configuration was issued. No startup image exists to review. Failure
caused permanent UNKNOWN, zero allowed later requests, browser closure and
actual controller/native/watchdog/child waits with both EOFs. The endpoint will
not be retried, rebound, replayed or reconnected. No diagnosis beyond that
observed native USER readiness contract is claimed. Walking, car entry, driving
and original native Quit were not reached; host teardown is not native Quit.
The paired WebGL run therefore remains prohibited.
Final evidence is result.json, runtime-release.json, qualification.json,
matrix-audit.json and original-result.json under the run directory. Source
recovery uses the exact implementation Git SHA and relevant pre-edit/candidate
binary patches, including originally untracked owned files. No source tree,
media or WASM snapshots are stored. Selected named screenshot references are in
result.json; no frame or capture bundle was created. Final shared restoration
uses the ordinary full flock build, preserving the qualified HEAD module's hash
and rebuild recipe rather than archiving its bytes. All owned runtime endpoints
are closed, actual waits joined and current PID/groups checked before the final
SOURCE_BUILD_RUNTIME release. No worker or box was created.