Recoil menu budget objective, 2026-11-08
Run: scratch/runs/20261108-recoil-menu-budget-objective/.
The sole fresh original successfully used the inherited explicit 500 ms
Escape command and reached native MENU. Its one ordinary-M compound then
failed to leave MENU: all eleven HUD probes had the same MENU pixels, so the
unchanged 4,500 ms gate sent zero M/C/arrow DOWNs. The first new defect is
accepting positioned: true and host virtual cursor coordinates as sufficient
Resume placement despite a visibly off-target guest cursor. This task closes on
a tested generic admission repair for that defect. Native M consumption and
reviewed ramp/control movement remain open; no second original or replay ran.
Before allocation, 788 providers passed exact size/SHA-256 checks. The actual
explicit-manifest preflight loaded the real hermetic preloader, HUD decoder and
native-menu helper with an intentionally wrong ambient PIN_FILE overridden.
The six actual Chrome menu-budget cases, full-browser menu-entry case, prior
750 ms source-stall/seven snapshot-identity cases and inherited suites were
reused through eight hash-checked control records, without standalone reruns.
browser-menu.cjs, native-menu-entry.cjs, bound-input.cjs and compound-v4.cjs
were reused byte for byte in the new run. Both complete guest/render byte checks
per HUD probe, exact instances/epochs, uncached validation, final co-located
pixels, temporal pair, 250 ms freshness and immutable deadlines remain intact.
The selected published source is
49874a81c5ae180c20a15c6bb169d9041f028cae, with actual guest and renderer module
6cf120d0d3b3db1823a8105eb26ef52a9093a7f871c37a46117475e213fda1c0.
The checkout identity at preparation was
aee5fec9cb3138ab8e170d28b43e978b5eeaa6cf; peer WIP was excluded. The remote
launch checked all 788 pins again after 789 individual uploads including the
manifest. Only one original launcher was started, PID 6406 at 10:32:15.251Z.
Reviewed native Software/OK, Game > Start, NEW GAME, default user/Medium and START
led to Island Storm. The 10:34:25 loading capture shows 1.1 Operation Short
Circuit; the passive 10:35:01 capture shows 1.4 Go to Landing Site. No briefing
cycling input was sent. Opening camera, transport and disembark captures remain
separate. The misleading filename living-check still depicts transport,
not a living HUD. The first reviewed living HUD is review/hud-check.png at
10:37:58.897Z, health 100, timer 00:00:05, tank and ramp ahead.
The native Escape request explicitly specifies 500 ms. Its persisted success
receipt records start 10:38:13.648Z and independent release at 10:38:14.149Z.
The 10:38:16.090Z capture visibly shows native MENU. No 100 ms request was
extended or retried. A later ordinary relative-pointer movement requested
[80,-186]; passive renderer coordinates became [400,60]. I incorrectly
interpreted those coordinates as proof of Resume placement. In the actual
resume-ready.png screenshot, the visible guest cursor hotspot is approximately
[430,224], above the reviewed Resume interior [380,247,504,279] in screenshot
pixels. The native 640x480 HUD-probe image independently shows the arrow above
the button. Host virtual coordinates and the guest-drawn cursor therefore
cannot substitute for each other here; the exact cause of their difference is
not established.
The single ordinary-map command requests Resume, m for 500 ms and gateMs 4500.
Its compound starts 10:38:57.123Z, gateDue 10:39:01.623Z. The acknowledged Resume
mouse DOWN/UP does not change the MENU image. Eleven probes, including the
baseline, retain one image hash; none is stale or a HUD candidate. The gate
closes at its original deadline, releases input and sends its ordinary Escape.
A final in-flight probe finishes afterward, without extending admission. The
10:39:03.108Z capture remains MENU. There are 21 commands and 21 exact ACKs,
including quit, zero original UNKNOWNs, zero gameplay DOWNs and no attack,
forced scene, guest write, private handler, configuration, clock or seed change.
No additional Resume or M attempt follows this newly demonstrated defect.
The new small native-menu-position.cjs requires a concrete visual-review
witness: exact capture SHA-256, review reference, decoded PNG dimensions,
visible guest cursor hotspot and reviewed Resume interior rectangle. It checks
actual PNG bytes/CRC, bounds and hotspot containment before any Resume effect.
A bare positioned: true, mismatched image or off-target hotspot fails closed.
This is a validator for explicit visual review, not an automatic cursor
recognizer. The reviewer must identify the actual guest cursor and button;
it cannot establish guest input consumption or explain cursor sensitivity.
browser-position.cjs integrates this guard before admission/input dispatch.
position-runtime-pins.json and launch-position.py provide the corrected
explicit-manifest closure. All 790 providers, real HUD decoder and new helper
pass the no-browser preflight. The original browser/manifest are preserved.
No production source was changed or production build claimed.
test-position.cjs reproduces the false boolean acceptance using the retained
actual original capture, rejects its visible mismatch with zero effects, accepts
a valid reviewed witness and checks seven negative witness cases. Two new actual
full-browser generated regressions exercise the integration. The off-target case
rejects before Resume: no input admission, no gameplay keys, only the initial
MENU phase, terminal no-replay closure and complete local cleanup. The positive
case uses a personally reviewed identical generated MENU image with hotspot
[410,50] inside [360,45,471,75], passes the unchanged 4,500 ms temporal HUD
gate and sends one trusted M DOWN. Its independent UP arrives about 495 ms
later; a second safe UP occurs at compound closure, leaving physical state empty.
These are generated integration cases, not original Recoil consumption evidence.
The initial positive verifier incorrectly expected exactly one UP. Its failed
diagnostic is retained; verify-position-good.cjs validates the already completed
run with the correct two-UP behavior, without another runtime execution.
Loading, opening camera and MENU are distinct observed phases. The interval between the reviewed HUD and Escape is only an enclosing wall interval, not an exact active-simulation measurement. MENU pixels during the failed gate are observed; exact MENU entry/resume instants and any unobserved interval remain unknown. No freeze, audio, parity, full27, simulation-FPS or game-FPS claim follows.
All 283 remote references / 23,098,993 bytes / 228 unique hashes were matched or retrieved, then independently reread and hashed before delegated stop. All 41 remote PNG references pass CRC decoding. The first immediate cleanup assertion found the launcher still present; that failure is retained separately. The later read-only custody check confirms launcher, monitor and Chrome PIDs absent, HTTP/CDP ports closed, profile/writer absent and last observed keys and buttons released. The added remote custody process records lock release. Before-close browser descendants and that added custody PID were not individually rechecked before stop; this is a custody-evidence limitation, not a claim of per-PID verification. Final box archival is independently confirmed.
The sole UUID-idempotent box is bx_8g6a8tbf, created 10:31:55Z, initial expiry
10:57:55.317Z. Observable independent normal-stop and 25-minute watchdog processes
were armed before allocation; watchdog deadline was 10:56:55.130Z. After verified
retrieval, normal delegated shutdown archives the box at 10:39:47.823Z, expiry
null. Receipt stop_bb855221d9ce6dd5 completes 10:39:47.843Z, forced false,
error null. Both lifecycle children exit 0 and supervisor children are empty.
The remote-only custody lock was added during the original; startup was not
lock-wrapped, which is explicitly retained as a lifecycle limitation. No local
runtime lock was held during provisioning, remote supervision or retrieval.
Each new local browser fixture held /tmp/wine-runtime.lock only while running;
both browsers, known PIDs, ports, profiles and writers drained, then locks released.
No old box resume/delete, public deployment, archive, media copy or subworker was
used. Free disk stayed above 5 GB.
verification.json, closure-check.json, position-launch-preflight.json, the
original receipts and both new browser qualifications retain the checks and
limits above. Git SHA, relevant binary diff, small owned-input/correction patches
and existing media paths/hashes provide recovery. Owned evidence containment, missing-artifact and schema checks pass; the global
checker retains seven unrelated containment errors in older peer runs.
Only this note is published;
evidence and helper candidates remain in the owned ignored run. Next requires a
separately authorized original with a visually verified Resume hotspot before
ordinary M, followed by reviewed movement only after a visible consequence.