Recoil native HUD gate, 2026-11-08

Run: scratch/runs/20261108-recoil-native-hud-gate/.

This worker repairs the retained v2 compound driver in a new run directory. The prior pinned v1, retained v2 and their raw failures are unchanged. The power-control/forcefield/onward-access objective is not accepted by driver qualification. The single original ended at a new dependency-packaging boundary before Resume input. The substantive driver repair is locally qualified; the native power-control objective remains unproved.

Reproduced driver failures and repair

baseline-regressions.json reproduces two failures in the retained v2 bytes: a synchronous HUD probe can delay the overall timer callback and admit a gameplay DOWN after the absolute deadline; a delayed C DOWN acknowledgement can extend the planned 80-ms C pulse beyond 1.1 seconds. These are harness failures, not demonstrated emulator divergences.

The newly owned compound-v2.cjs checks the absolute deadline before each gameplay DOWN and releases C on its own 80-ms timer, independently of its acknowledgement. Arrow releases have their own timer and the overall release/Escape remains independently armed before Resume. A gameplay DOWN also requires the qualified image to remain at most 250 ms old. A delayed C acknowledgement therefore cannot admit an ordinary action using expired HUD evidence. Releases can follow a deadline; further gameplay DOWNs cannot.

hudProbe reads the full owner/module identity before and after the native canvas image, saves its PNG and hash, and retains the 250-ms freshness check. It first requires a nonliving pre-Resume image, rejecting an already visible HUD as a cached-backdrop admission. It then requires two living HUD candidates with distinct green timer-glyph masks. Hashing only the glyph mask prevents a changing sky or cursor from impersonating timer progress. Repeated captures of the same HUD cannot qualify. Missing health, unsupported dimensions, the retained death capture and identity changes reject gameplay input. This is a specific 640×480 Software HUD gate, not target recognition or a general game-state oracle.

Qualification and source identity

The selected published closure is 49874a81c5ae180c20a15c6bb169d9041f028cae, including protocol 663f1c139 and the browser journal decoder repair. WAT sources are read from Git and compiled in an in-memory VFS. Published bytes are served through the retained Git provider; peer WIP is excluded from the selected runtime. Git SHA, binary diff and new-input patches provide recovery; no source/media archive is made.

The real registration qualifies 106 manifest entries + EXE + manifest = 108 pins / 76,019,820 bytes. Empty, missing, duplicate and wrong staged EXE cases and the invalid zero-file report reject. All 113 HTTP dependency pins pass. The unchanged bound-input, ordinary-keys, native-pointer and transport control helpers match their retained SHA-256 values; their existing controls are reused rather than claimed as new original-game evidence.

repair-regression-final.log covers twelve timing/failure cases, including synchronous overdue probes, an overdue interval between two ordinary keys, independent per-key releases, slow sampling, changed identity, stale images, late/absent HUD and partial UNKNOWN. Recorded native image tests reject an unchanged cached HUD and death; two distinct timer images qualify.

The full browser driver is tested separately from these helper tests. Its generated fixture uses production renderer-input and an actual recovered Worker module, displaying explicitly labeled retained menu/HUD images. This does not launch Recoil or prove Recoil consumes an input. The first integrated attempt's Chrome-exit/driver-exit assertion mistake and its captures are retained, alongside the corrected final attempt. Browser normal exit and a driver's expected failure are different facts.

Original-game interpretation

Host dispatch times, actual native key consumption and first observed native menu are separate observations. A first sampled HUD after Resume is an upper bound on that visible transition, not an exact simulation-resumption time. Generated fixture events cannot establish native Recoil C or movement consumption. The prior loading display's 1.3 rocket-site text and historical 1.1 power-control brief still require an ordinary fresh briefing/map before any mission-objective claim. Full27, parity, audio and simulation FPS remain open.

Integrated results and dependency repair

final-integrated-qualification.json records eleven full-driver cases: delayed, absent, late, cached-after, cached-baseline, changed-owner, stale, slow-1100-ms, partial UNKNOWN, per-key C and per-key arrow/overall timers. Nine unchanged results were reused. A 50-ms timer-image fixture could change between identity reads and correctly admitted no gameplay; only that fixture interval was corrected to 400 ms for the two timer cases. Both raw failures are retained. The delayed C acknowledgement case released C after 81 ms and rejected the following arrow because its HUD evidence had expired. The slow-probe case also demonstrates that host UP dispatch is not a native release deadline when the browser event loop stalls.

native-final-qualification.json records the actual published Calculator Edit → Copy route through the complete v2 driver and recovered production module 6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0. Its unsupported HUD admitted zero gameplay keys; independent Escape and visible menu closure were checked. These local tests did not prove the remote dependency closure: the original pin scanner followed relative requires and missed the bare pngjs import. The local installation concealed that omission. This qualification failure belongs to this worker.

After the original stopped, browser-run-v3.cjs was created separately; the actual original v2 bytes and inventory remain unchanged. V3 eagerly loads the HUD decoder before creating a browser. prepare-dependency-repair.cjs pins 23 reachable pngjs 7.0.0 files (63,864 bytes), checks every installed byte against the lockfile's SHA-512-verified npm artifact in memory, and writes a 755-file repaired runtime manifest. No npm archive or package tree is copied to evidence. dependency-closure.json retains the package URL, integrity, archive hash and dependency edges for recovery. The repaired stage byte check passes all 755 pins against the same selected published closure.

hermetic-modules.cjs refuses resolution of any unstaged nonbuiltin module. The missing-dependency full-entry regression exits before creating browser output; the good full integrated delayed-HUD case passes with that loader, including trusted input consumption in the generated fixture, independent releases and clean Chrome exit. The unchanged native route and eleven timing controls above are reused; they are not relabeled as new original evidence. The separate provisioning reference supports npm pins, but deliberately refuses execution in this completed task. No second box was allocated.

Single original: observations and stopping boundary

The one UUID-idempotent original was bx_qfgwy3ra, created at 08:35:46.821Z. It used the selected 730-file closure, independently checked in the remote runtime, and the actual recovered module hash above. A detached observable 25-minute watchdog was armed before allocation, with deadline 09:00:47Z; initial box expiry was 09:01:46.835Z.

Fresh pixels established Software mode, default name user, Medium, ordinary native Start/New Game, the loading sequence and a living tank with health 100 and timer 00:00:04 (opening-camera.png). Ordinary loading/briefing frames showed “1.2 Missile Control” with SAM-site text/map and “1.4 Go to Landing Site.” This demonstrates cycling subobjective text; it does not resolve the historical 1.1 power-control brief versus 1.3 rocket-site display. No mission objective is accepted.

Command 25 dispatched and independently released Escape; native-pause.png first visibly establishes the native pause menu. Command 26 produced the fresh highlighted Resume image. Command 27 attempted the first HUD-gated compound action but failed while requiring pngjs in its pre-Resume baseline probe. Its error is Cannot find module 'pngjs'; no Resume, C or arrow was sent by that compound. Admission had already marked possible dispatch, so transport conservatively became sticky UNKNOWN with no exact command ACK. The browser closed normally, and there was no retry, death/restart sequence, fire, targeted defense, private guest call, state/clock edit or hotpatch.

Commands 0–26 have exact ACKs, including command 9's stale-review refusal before effects and a fresh review afterward. Command 27 is UNKNOWN; it must not be recast as an acknowledged zero-effect action. Source ordering supports the pre-input failure, while the transport record remains ambiguous by design. No native HUD-gate-before-C, C consumption, ordinary movement consumption, control identification, forcefield change or onward access was proved. The last sampled health is not a final-state claim. A read-only presentation check found the logical canvas and exclusive source identical; the earlier OffscreenCanvas inspection error is retained and does not establish an emulator divergence.

The terminal collector verified 15 process IDs absent, HTTP/CDP ports closed, profile/writer gone and last observed buttons/arrows released. All 323 unique remote artifacts / 25,728,292 bytes were retrieved and independently rehashed before requesting normal shutdown. terminal-retrieval-verification.json and final-stop-receipt.json retain the checks and lifecycle. The box became archived at 08:50:52.535Z, final expiry null; stop receipt stop_48ae8d97f58840d0 completed at 08:50:52.558Z, forced false, error null. Normal and watchdog supervisors exited cleanly. No old box was resumed or deleted.

The new concrete boundary is dependency qualification on an isolated runtime, not the former early-Resume delay. A future separately authorized original must use the repaired closure and hermetic loader, then establish the native HUD gate before C and ordinary action consumption plus independent UP and visible MENU closure. Fresh ordinary briefing/map evidence must resolve 1.1 before approaching a reviewed ramp/control. No production emulator edit was justified, and no production build or public deployment is claimed here.