Battlezone (demo)

Activision, 1998, Direct3D (its own plugin\dx5draw.dll renderer) tank action/RTS hybrid. Candidate battlezone-demo, app battlezone_demo, local only.

Source and layout

Route (headless, --batch-size=100000)

Emulator bug it found: heap fragmentation under 1 MB

Every ad repaint pass allocates 4 KB (file buffer), 1072 bytes (BITMAPINFO, never freed) and 921,608 bytes (pixels), then frees the 4 KB and the pixels. Large blocks only got their own arena from 1 MB up, so the 900 KB block went back on the ordinary free list. The next pass carved the leaked BITMAPINFO from it, which left it too short for the next 900 KB request. Every pass took fresh space; about 335 passes in, the heap returned NULL and the main menu background drew grey. $HEAP_LARGE is now the Windows heap's VirtualMemoryThreshold, 0x7F000 (508 KB). test-heap-large-block-arena.js replays the pattern 1200 times.

Controls (data\Input.map)

Q forward, S back, W turbo, A/D strafe, keypad 4/6 steer (RightArrow in the map is the keypad key; GreyRightArrow etc. move the track camera), F/C pitch, E jump, left mouse or keypad 0 (Insert, distinct from GreyInsert) fires, right mouse or keypad Enter cycles weapons, Space is the reticle command, Alt the target command, 1-9 the command menus, Esc the main menu. Verified headless: keypad 4 turns, Q drives, the grey Left arrow does not steer. The touch layout maps the stick to Q/S/keypad 4/6.

Browser

The default lazy file policy streamed addon\misn*.trn, but those files are INI text read by GetPrivateProfileInt at mission launch, which can't park: VfsPendingError. trn is now in lib/app-files.js EAGER_EXTENSIONS.

Original combat and vehicle entry (2026-11-01)

Run scratch/runs/20261101-battlezone-original-combat-control/ reviews one unchanged registered original, native local CLI, cooperative backend, 640×480: --app=battlezone_demo --no-build --batch-size=100000 --control-stdin --frozen. The GL predecessor's fresh final/result and 08:47:07 runtime/source/build RELEASE were checked before the 08:48:54 gate; actual main was b05826e4a6d67eab69ddbacb00a685c019b474c5. The historical accepted Battlezone run directory was absent locally; the retained note/board established driving only, so it did not replace this combat check.

The accepted build was reused after all 4,392 accepted inputs and 141 prior read hashes matched. Actual runtime compilation observed module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. All 153 local fixture files, including the browser manifest and original companions, matched before/after; the 3,732,992-byte EXE matched 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. All 4,444 source pins and 153 actual runtime file-read hashes also matched. Read observations cover synchronous file reads and complete bounded descriptor reads, not every possible OS/asynchronous read. These are actual runtime reads, not compilation-time source reads relabeled at final flush. The observer's separate actual sync/async, whole shared 512 MB readonly, clock identity, and emitted late-10,001 retention checks passed.

Ordinary route: ad click at batch 280; initially black main-menu captures repainted on a held Single Player press at 670/release 695. Stars and Stripes at 920/945, Launch at 1020/1045, visible mission at 1745. Briefing instructions say to get in a vehicle at Luna Outpost 3 and point the reticle at vehicles and press I for additional information. I produced no reviewed information panel; Space on the distant light tank remained “No Action.” No entry key was invented.

Firing accepted: the early green arcs while aiming at the friendly tank were ambiguous and are not counted as projectile evidence. A later ground aim shows a blue shot streak and orange impact flare. The clean stationary series is stationary-fire-before (6730), held100 (6830), held250 (6980), released (7030), and settled (7330), all PNGs with matching input-state records. Left-button down was queued at 6730 and up at 6980, both at 320,240. The impact area differs by 1,313 pixels while held and returns exactly to its baseline after settling. Two terrain patches and the left vehicle patch are pixel-identical across all five captures: the shot effect is not camera motion or selection alone. The pilot ammo bar does not visibly decrease. Radar, small HUD changes, and upper-screen rendering artifacts are not counted as combat effects. The earlier ground-fire-held1000.png also shows the blue streak, but its aim changed during that press, so it is not the fixed-camera comparison.

Vehicle entry and vehicle fire accepted: Q held from 7930 to 8430 walked into the nearby vehicle. vehicle-info-held250.png still shows the pilot's Plasma Rifle/Sniper Rifle; approach500.png, vehicle-entry-released.png (8530), and vehicle-entry-settled.png (8730) show AT-Stabber, Thermal Hornet, and Solar Flare instead. This ordinary forward approach boarded the vehicle without a separate entry key. The HUD transition supports entry beyond camera movement. A subsequent centered left-button hold from 8730 to 8980 visibly fires the AT-Stabber: tank-fire-before/held100/held250/released/settled.png show ammo 83 → 77 → 67 → 66 → 66, with yellow shot streaks while held. The last decrement occurs before release processing finishes. Vehicle terrain continues to drift slightly after the approach; this series is not a claim of a fixed camera, and its ammo count is the independent firing outcome.

Actual input evidence is the bounded raw dequeue/route/DispatchMessage trace and frozen queue/counter records, not merely scheduled commands. The mouse DirectInput FIFO head remained zero and is explicitly not claimed as consumed; the window-message path consumed the button edges. No guest game structures were inspected. One original exited 0 after 551.64 seconds, with no crash/trap. Own runtime/driver/watcher PIDs, process group, and child temps were absent afterward; minimum free disk was 2,270,924,800 bytes. Box ID/state/ expiry are N/A. No source, test, helper, backend, registration, or default was changed. The note and ignored run evidence are the only deliverables.

Target-hit / unit-command attempt: prerequisite stop (2026-11-01)

Run scratch/runs/20261101-battlezone-target-hit-command/ stopped before launching an original. It adds no target-hit, visible unit damage, or unit command outcome to the accepted entry/firing evidence above. Battlezone remains a local candidate (LOCAL_CANDIDATE_APPS), absent public DESKTOP_APPS; no registration changed.

The valid gate opened at 14:32:52 UTC, after the exact quake2-eight-save-restoration DONE/PUSH RELEASE messageboard action naming runtime/source/build, fresh terminal launcher final.txt/result.json, and direct predecessor PID/process-group cleanup checks. WAIT entries and quoted conditions were not accepted as releases. Dispatch main was 8552125804a6994de385d7f00858e789fd59024f; actual main at the gate was fbb83926ba331962cc85e2c44f1b623e778e262f. No runtime, build, or test ran before this gate. The 25-minute gate deadline was 14:57:52 UTC.

All 4,392 accepted build-input pins and 141 prior read pins matched, including canonical module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. The actual before/after inventory contained 4,444 source pins and all 153 fixture files, including the manifest and companions; hashes and fixture tree membership stayed unchanged. The EXE remained 3,732,992 bytes, SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. These establish build eligibility and file identity, not an actual original module load or runtime-read observation in this attempt.

First blocker: the scratch launcher invoked offline.cjs without setting SWEEP_CASE_DIR. The observer takes its output directory from that variable; observe.cjs:44 called path.join with undefined in save(), producing TypeError [ERR_INVALID_ARG_TYPE] and Node exit 1 at 14:32:57 UTC. This was a worker harness prerequisite failure, not a Battlezone compatibility failure. The sync/async instance exercises preceded the throw, but the emitted late-10,001 check and final whole-512-MB readonly verification did not complete. The observer contract is not passed for this attempt.

Per the first-failure stop condition, the harness was not repaired or rerun. There were zero original launches, zero gameplay inputs, and no new game captures, so no consumed-control, released-control, matched-neutral, target-health, hit, kill, or command-response claim is made. The bundled briefing's Recycler/Scavenger objective and Input.map were read in scratch preparation, but the prepared command route was never executed. Existing vehicle entry and firing/ammo decrement remain the earlier accepted results. Local native box ID/state/expiry are N/A. No tracked source, helper, test, configuration, backend, or registration changed. Evidence retains the gate, pin comparisons, exact exception, and cleanup/publication records. A future separately authorized attempt needs a complete observer validation before launching the original; this assignment stops at the prerequisite failure.

Corrected observer and reticle unit outcome (2026-11-01)

Run scratch/runs/20261101-battlezone-target-hit-corrected/ adds an ordinary reticle command with an observable Scavenger approach. It does not add target damage, a kill, destination arrival, or mission completion. The earlier vehicle-entry and firing/ammo results are not counted again. Battlezone remains in LOCAL_CANDIDATE_APPS, absent public DESKTOP_APPS.

The exact urban-chaos-confrontation DONE/PUSH RELEASE action naming runtime/source/build, fresh terminal launcher final/result, and directly checked predecessor PID/group cleanup opened the gate at 14:49:06 UTC. Dispatch was ad94d102f5b4a239e86d7b38970455dada45dabe; actual main was c6fba571254a85245166115a08880cec00dc911d. The gate deadline was 15:14:06 UTC. Before that gate, preparation stayed in scratch without tests or runtime.

The authorized scratch correction sets SWEEP_CASE_DIR=__dirname before offline.cjs requires the observer. Its actual synchronous/asynchronous execution, whole shared 512-MB readonly hash, native/import clock identity, and emitted late-10,001 retention checks all passed before game launch. All 4,392 accepted build inputs and 141 prior read pins matched, permitting --no-build. Actual runtime compilation observed module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. Before/after inventories matched all 4,444 source pins and 153 original files, including manifest, companions, and tree membership. The EXE remained 3,732,992 bytes, SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. All 153 observed runtime-read hashes matched. Observation covers synchronous string/URL reads and complete bounded descriptor reads, not every possible asynchronous/direct OS read or compilation-time source read.

One original used the registered native cooperative CLI with --app=battlezone_demo --no-build --batch-size=100000 --control-stdin --frozen. The first 280-batch request returned after 147 seconds, within its 180-second deadline; this was slow startup, not a timeout or established freeze. Reviewed captures then followed ad 280/305, Single Player 326/351, Stars and Stripes 372/397, Launch 418/443, and visible mission at 1364. The actual briefing orders the player to have the Recycler build a Scavenger; Input.map supplies numbered menus and Space as the reticle command.

Build outcome: 5 opened the Recycler menu at 1410/1435. It offered 1 Scavenger for four scrap. Pressing 1 at 1502/1527 changed scrap 05/30 to 01/30 and enabled the Utility menu. Utility 3 then listed 1 Scavenger; selecting it opened its Follow Me / Go To Nav / Scavenge / Pick Me Up menu. This supports a built, selectable unit, beyond an unchanged menu selection.

Reticle outcome: the first Space command at 1846/1871 displayed Follow Light Tank and closed the menu, but the unit was already moving during the neutral interval; that interval alone is inconclusive. Subsequent ordinary relative mouse input established a terrain reticle. With the Scavenger selected, terrain-command-ready.png (4764) explicitly offers Space Go To. A matched 100/1/300-batch neutral bracket precedes Space down at 5165 and up at 5265. The menu closes, and a yellow vehicle crosses the reticle area in reticle-goto-settled.png (5566); that sprite alone is not the final identity or arrival evidence.

A second, nearer terrain point is visible as a green ground highlight in near-goto-ready.png (6408), again with Space Go To and the selected Scavenger. The matched neutral bracket is 6408/6508/6509/6809; the command bracket is reticle-near-goto-before/held/released/settled.png at 6809/6909/6910/7210. Space down/up were consumed, with no steering, translation, or mouse movement during either bracket or the later approach. Later fixed-view captures identify Scavenger with a world marker and a decreasing range: near-arrival1800.png (9010) shows 663 m and near-arrival2700.png (9910) shows 384 m. Its marker and radar position also change while the foreground/base view remains visually fixed. Together with the native Go To affordance and consumed command, this is an observable commanded-unit approach, not camera drift, projectiles, or menu closure alone. Arrival and the later escort/harvesting objective were not established. Early live range readings in tens of metres were incorrect; the retained captures support hundreds of metres and are authoritative.

All 32 key edges (16 balanced pairs) match the actual dequeue and DispatchMessage order, with each key dequeue routed. Mouse/menu edges also have raw consumption records; native dialog processing is not uniformly a main-window DispatchMessage path. Final scheduled/pending/cross-thread/ renderer queues are empty; all used sampled physical/async keys, including left-button VK1, are false. The numeric renderer mouse-button mask was sampled unknown, not zero; pressed latches remain true for some used keys. Guest post-queue and SDL consumption are unknown. No guest game structures were inspected. All 122 retained PNGs passed CRC/deflate checks, and the full runtime gzip validated.

An optional quantitative pixel comparison failed because Python Pillow was absent (ModuleNotFoundError: No module named PIL). It was not installed, repaired, or retried; gameplay stopped and no quantitative scene-difference claim is made. The retained matched neutral and command captures were visually reviewed. This analysis dependency is not a Battlezone crash or defect.

The original exited 0 after 719.25 seconds at batch 9910, without a crash/trap. Minimum controller-sampled free disk was 2,046,517,248 bytes; the auxiliary guard remained active through publication. Finite sampling cannot establish unsampled disk history. Own runtime/helpers/watchers/groups and new child temps were checked absent after shutdown. Local native box ID/state/expiry are N/A; own shutdown was explicitly delegated. Only this note and ignored scratch evidence changed; no tracked source/helper/test/configuration/backend or registration changed.

Scavenger arrival attempt: black native dialog (2026-11-01)

Run scratch/runs/20261101-battlezone-scavenger-arrival/ extends the identified commanded Scavenger approach to 180 m, then stops at a black native-dialog transition. It establishes neither destination arrival nor scavenging/escort progress. The already accepted construction and 663 m → 384 m approach are setup, not new success. Battlezone remains a LOCAL_CANDIDATE_APPS entry, absent public DESKTOP_APPS; no registration changed.

The gate opened at 15:32:43 UTC, after the exact author urban-chaos-small-step-confrontation posted DONE/PUSH RELEASE naming runtime/source/build, both fresh terminal launcher files existed with exit 0, and predecessor PID/group/child-temp cleanup was checked directly. No WAIT or generic release substring qualified. Dispatch main was 60500ae55fa28dc658402496afb6e36f88d4296f; actual main was 88cf06d749e4eff628e660b331b8030651edc063. The gate deadline was 15:57:43 UTC, within the overall 17:15 UTC limit. Preparation before the gate was scratch-only, without runtime or tests.

All 4,392 accepted build inputs and 141 prior read pins matched, allowing --no-build. The corrected scratch wiring sets SWEEP_CASE_DIR=__dirname before requiring the observer. Its actual sync/async execution, whole shared 512-MB readonly hash, native/import clock identity, and emitted late-10,001 retention checks passed before launching the original. The actual runtime compile observed module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. Before/after inventories matched 4,444 source pins and all 153 fixture files, including manifest, companions, and fixture tree membership. The original EXE remained 3,732,992 bytes, SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. All 153 observed runtime-read hashes matched. Read coverage is synchronous string/URL reads and complete bounded descriptor reads, not all asynchronous or direct OS reads; runtime identity is not compiler-source-read evidence.

One unchanged original used the registered native cooperative CLI: --app=battlezone_demo --no-build --batch-size=100000 --control-stdin --frozen. Actual captures were reviewed through the ad/menu route, Stars and Stripes briefing, and mission. The live briefing orders vehicle entry, building a Scavenger, and escorting it while it gathers bio-metal. The bundled Input.map was read; no bindings were invented. Reusing the accepted route reproduced scrap 05/30 → 01/30 and a selectable Scavenger whose menu includes 3 Scavenge. No Scavenge order was issued before the blocker.

The near terrain point visibly offers Space Go To at batch 6408. The retained matched neutral bracket is 6408/6508/6509/6809; reticle-near-goto-before/held/released/settled.png is 6809/6909/6910/7210, with Space down at 6809 and up at 6909. There is no translation, steering, or mouse movement during these brackets or the later wait. The reused route reaches the identified 384 m marker at batch 9910. New reviewed captures then show 265 m at 10210 and 180 m at 10510, with the unit marker/radar moving in the same scene. These remain approach evidence, not arrival.

First visible-feedback blocker: between batch 10510 and the next capture at 10810, the runtime logs CreateDialog for HWND 0x10018, a 427×274-dialog-unit template, with control 1237. The capture becomes entirely black. A readonly renderer window inventory later shows that dialog visible, owned by BattleZone HWND 0x10001, with an empty title and 640×480 surface. The retained bounded no-input capture sequence at 10810/11110/11410/11710 stays black; a standard-library PNG/zlib decoder confirms zero nonblack RGB pixels in each. This does not reveal the dialog's message or establish whether it represents defeat, an objective result, or another game state. No arrival, harvesting, scrap gain, target hit/kill, or mission completion is claimed. Gameplay inputs stopped; there was no dialog bypass, repair, retry, or second original. The runtime remained responsive to inspection and quit, so black output is not labeled a crash or freeze.

All 32 key edges / 16 balanced pairs match actual dequeue and DispatchMessage order, and every key dequeue has a route record. Mouse/menu edges retain raw consumption evidence; native menu/dialog handling does not uniformly use main-window DispatchMessage. Final batch 11710 has empty known scheduled/pending/cross-thread/renderer queues. All used sampled physical and async keys, including left-button VK1, are false; some pressed latches remain true. The numeric renderer mouse-button mask, guest post-queue, and SDL consumption remain unknown. No guest game structures were inspected. Retained PNG CRC/deflate and full runtime gzip checks passed. Image analysis used Python's standard library only, without Pillow or dependency installation.

The original exited 0 after 367.89 seconds, with no recorded crash/trap. The supervising tool session separately reported status 143 after the gate log recorded HELPERS STOPPED; its cause is not established and is not relabeled as the original game exit. Controller-sampled minimum free disk was 2,036,707,328 bytes, above the 2,000,000,000-byte floor; guard and publication samples are retained. Finite sampling does not establish unsampled history. Own native runtime shutdown was explicitly delegated; box ID/state/expiry are N/A. No tracked source, helper, test, configuration, backend, or registration changed. Only this note and ignored scratch evidence are deliverables. A future investigation would need to explain the black dialog before accepting arrival or scavenging; this assignment stops at that blocker.

Black dialog identified; ordinary onward Scavenge order (2026-11-01)

Run scratch/runs/20261101-battlezone-black-dialog/ identifies the formerly black dialog as a MISSION COMPLETED screen and recovers its actual native presentation. Ordinary Next Mission → Launch → Utility → Scavenger → Scavenge subsequently reaches the second mission and issues its displayed Scavenge order. A disk-floor breach stops observation before a harvesting, scrap-gain, arrival, or escort outcome is established. No second original ran.

Preparation read the original PE resources, imports, aligned disassembly, saved actual window inventory, and bundled Input.map. The changed hypothesis was H1: the CLI main canvas omits the native dialog backing, rather than the guest drawing a black dialog. Two templates initially matched the logged geometry: resource 130, static control 1237, END MISSION, and resource 131, the same control/geometry labeled BACK. Both are 427×274 dialog units; control 86,7 70×14 becomes logged pixel geometry 129,11 105×21. Control ID or geometry alone therefore could not establish meaning.

The exact urban-chaos-bounded-wait author release, fresh terminal launcher final.txt/result.json (exit 0), and directly verified predecessor cleanup opened the gate at 15:57:55 UTC. Dispatch was a1c43c412f2406bb58081ad4fa654b36b2973af3; actual main was repinned to 4d9446b48ccc83286c5ed82ee61e14f2b3f303e0, with deadline 16:22:55 UTC. All 4,392 accepted build inputs and 141 read pins matched. Corrected SWEEP_CASE_DIR observer validation exercised actual sync/async execution, whole shared 512-MB readonly hashing, native/import clock identity, and emitted late-10,001 retention before the original. Offline records are separately named. Actual original compilation observed module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. All 4,444 source pins, 153 fixture files/tree members, and 156 observed runtime-read hashes matched afterward. The EXE remains 3,732,992 bytes, SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. Read coverage remains synchronous string/URL and complete bounded descriptor reads, not all async/direct OS reads or compiler-source-read evidence.

The unchanged original used --app=battlezone_demo --no-build --batch-size=100000 --control-stdin --frozen, reusing the predecessor's actual controller route. Construction and the 663→384→265→180 m approach are setup. At 10810, readonly dialog_get_proc_export(0x10018) returns 0x5bd890. The PE call at 0x5bd72e invokes imported DialogBoxParamA with template 130 and that callback. Its initialization references custom label NEXT MISSION (original VA 0x626fc4) and MISSION COMPLETED (0x626fac), explaining why the template's static label alone is insufficient.

native-dialog-direct.png reads the existing dialog _backCanvas without painting or guest writes. Existing dlg-png at the next batch produces native-dialog-scheduled.png: both visibly show MISSION COMPLETED, the debrief, and NEXT MISSION. Their decoded RGB bytes match exactly, with 261,547 nonblack pixels; the corresponding main-canvas capture has zero. The matched neutral native captures also match exactly. This establishes the displayed mission result, without independently proving the first Scavenger's arrival, collection, or escort mechanics. The debrief describes Soviet forces, the loss of a Recycler at Eagle's Nest 1, and recall of combat units; it is not an inferred defeat message.

The narrow capture reproducer is the actual game's DialogBoxParamA(130, callback=0x5bd890) reached by accepted-route.json, then frozen {action:"png",path:...} versus the existing dlg-png:PATH plus one step, or a direct readonly backing encode as in inspect-dialog.py. No repaint, callback patch, memory write, or fabricated dialog is needed. The available repair target is test/run.js controlPng (around line 7388): it encodes renderer.canvas via canvasToPng without the renderer.repaint() used by the scheduled png handler around line 10141. canvasToPng only invokes the DirectDraw present hook. The capture discrepancy is proved; a generic repair and standalone API fixture were not implemented or tested here. No guest callback/JS/concurrency/SEH algorithm was altered.

The Next Mission neutral bracket is 10811/10836/10837/10857; ordinary mouse down/up at visible 90,15 produces 10857/10882/10883/10903 and the next briefing (native HWND 0x1001a, callback 0x5b4a10). After a matched 25/1/20 neutral bracket, clicking visible Launch at 550,15 produces 11249/11274/11275/11295. Actual onward gameplay is reviewed at 12195: scrap 08/30, pilot 10/13, Command Tower at 260 m. Utility 3 lists two Scavengers. Selecting 1 displays 3 Scavenge; pressing 3 has bracket 12425/12450/12451/12471, preceded by a matched 25/1/20 neutral bracket. The consumed order closes the command menu; scrap remains 08/30. This is an issued native Scavenge order, not evidence of completed collection.

First substantive blocker: during the next requested 300-batch no-input wait, controller-sampled free disk fell to 1,978,392,576 bytes, below the required 2,000,000,000-byte floor. The guard requested SIGTERM; the original handled it orderly and exited 0 after 594.73 seconds at batch 12744, before that step replied. No retry, second original, foreign cleanup, bypass, or further gameplay followed. Own redundant evidence was losslessly compacted or removed with recovery commands; sampled free disk recovered above the floor. The continuous-disk requirement failed; later recovery does not erase it. The supervising gate session separately returned 143 after HELPERS STOPPED; that is not the original's exit status. No game crash/freeze is established.

All 38 key edges / 19 pairs match actual dequeue, route, and main-window DispatchMessage order. Mouse actions retain raw native-dialog consumption records. The last complete state sample is 12471, after Scavenge release: known scheduled/pending/cross-thread/renderer queues empty, all used sampled physical/async keys and VK1 false. Pressed latches remain true for some keys; numeric mouse-button mask, guest post-queue and SDL consumption are unknown. This is not an at-shutdown state sample. 210 PNGs pass CRC/deflate checks; the runtime gzip validates. Analysis uses stdlib PNG/zlib, without Pillow. Own PIDs/groups/watchers/child temps were checked absent; native local box ID/state/expiry are N/A and own shutdown was delegated. Only this note and ignored evidence changed; no source/helper/test/configuration/backend or registration changed. No FPS, audio, browser/GPU, or full-suite claim is made.

2026-11-01: extra repaint disproved; onward Scavenge observation interrupted

Run: scratch/runs/20261101-battlezone-dialog-capture-harvest/. The exact urban-chaos-ready-wait-repair release, fresh launcher terminal files, and direct cleanup passed at 16:35:46 UTC. Input main 6f4fa1d5ecb75f03aea8e0b354dbe7c051ca3c5f was repinned to a17c5a7c85aae38a97653b801da68de6147048df; the runtime deadline was 17:00:46 UTC. Scratch preparation preceded the gate; build, tests and the single original followed it. The mandatory locked build, existing test/test-control-stdin-frozen-cli.js, and observer validation passed. No production capture correction or new regression test was committed: the proposed extra repaint does not repair the real capture path.

At frozen 10810, the actual native dialog again shows MISSION COMPLETED (resource 130, control 1237, callback 0x5bd890). Both control PNG and an explicit renderer.repaint() followed by control PNG remain completely black. Scheduled png at the following batch is also black. All three corresponding native backing captures have identical RGB hashes and 261,547 nonblack pixels, matching the predecessor. Whole 512-MB memory hash, EIP, API count, and batch are unchanged around the frozen control capture and explicit repaint. The scheduled pair separately requires one guest batch; it is not presented as a frozen comparison.

The earlier suggestion that controlPng merely lacks the scheduled handler's repaint is incomplete: canvasToPng → presentForRead → dxPresentHook already calls renderer.repaint(). The actual visible dialog's GDI write sequence is 18584, while the main DirectDraw layer's write sequence is 18585, with coverSeq 0. Publication is not held. The exclusive compositor in lib/renderer.js gates window backings against the newest present sequence; this rejects the dialog backing at this checkpoint. No bounded test/run.js-only correction was established; changing renderer freshness rules is outside the assigned file ownership. No renderer rule, callback, guest memory, concurrency or SEH algorithm was changed; no replacement frame or game-specific fallback was added. Existing native backing capture remains the readable evidence route.

Ordinary Next Mission at 90,15 reaches the reviewed briefing at 10903; ordinary Launch at 550,15 reaches onward gameplay at 11895, with scrap 08/30. The live Utility menu lists two Scavengers; selecting 1 shows 3 Scavenge. The Scavenge action bracket is 12125/12150/12151/12171, preceded by the matched 25/1/20 neutral interval. The actual key release is dequeued, routed and dispatched at 12153. Reviewed no-input checkpoints 12471 and 12771 (300 and 600 batches after settling) still show 08/30. This establishes the issued order and bounded observation, not collection, scrap gain, or escort mechanics.

First changed blocker: the driver and disk monitor terminated unexpectedly; their tool sessions returned 143, while the original's exit hook recorded code 0. At discovery around 16:46:22 UTC, the original and both helpers were absent. Shutdown cause is unestablished, and no game crash/freeze is inferred. The driver did not write its normal exit.json. No second original or further input followed. The surviving owned board watcher was explicitly stopped. Last complete input-state evidence is 12771, not a shutdown sample: known scheduled/pending/cross-thread/renderer queues are empty; used physical and async keys, including VK1, are false. Pressed latches remain true for some keys; numeric mouse mask, guest post queue and SDL consumption remain unknown.

The original launched with 2,237,816,832 bytes free. All recorded disk samples exceed 2 GB; the recorded minimum is 2,181,066,752 bytes. Monitoring ended unexpectedly too, so continuity across that final gap is not established. The truncated original gzip is retained unchanged; log-recovery.json gives a lossless recovery recipe and hash for the exact retained prefix. The separately named recovered-prefix gzip validates; missing shutdown output is not invented. Terminal journals were losslessly compressed with recovery recipes.

All 4,062 scoped source/module pins, 153 fixture files and manifest members, and 156 observed runtime-read hashes match afterward. This source scope is tracked/untracked JS/CJS/MJS/JSON/WAT/WATX/shell/Python/WASM files under src/lib/tools/test, plus host/package/module pins; it is not the predecessor's 4,444-entry scope. The actual compilation again observes module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. Observer validation executes sync/async modules against whole 512-MB readonly memory, preserves clock identity, and retains late event 10,001. Runtime read coverage is synchronous string/URL and complete bounded descriptor reads, not all asynchronous/direct OS or compiler-source reads. All 38 key edges / 19 pairs match dequeue/route/dispatch order; 81 PNGs pass CRC/deflate checks (four are explicitly historical references). Known owned PIDs/groups and matching helpers are absent, with no new child temps. Native local box ID/state/expiry are N/A; own shutdown was explicitly delegated. Only this note and ignored evidence changed. No FPS, audio, browser/GPU or full-suite claim.

2026-11-01: generic exclusive freshness repaired; native launch blocked

Run: scratch/runs/20261101-battlezone-renderer-dialog-freshness/. The exact urban-chaos-guard-gameplay author release, fresh launcher final/result (exit 0), and direct owned-process/group cleanup passed at 17:14:58 UTC. Input 6ca46f8ffced6b6e4331e8fa4efa33e4d91c6872 was repinned to main f20e19b24651f2e37fe53b422378205328338270. The gate deadline was 17:39:58 UTC, inside the overall 18:00 UTC limit. Only scratch preparation preceded this gate; source edits, build and tests followed it. No original was launched in this run.

The generic renderer now measures exclusive GDI backing freshness against the last covering presentation (coverSeq), consistently with shared native children. A newer DirectDraw write through a window-bound clipper does not overwrite other visible window pixels. Comparing their backings with that write's writeSeq incorrectly drops the active native dialog. The correction applies to top-level and own-child backings in both the screen compositor and its post-processing source. Existing visibility, ancestor clipping and stack order remain authoritative. The last real covering sequence survives later clipped writes, so previously covered backings remain rejected until an actual newer GDI paint. There is no game-specific branch, modal exemption, unconditional overlay or replacement frame. The complete production delta changes three freshness predicates and comments; callback/context, worker publication holds, concurrency and SEH algorithms are unchanged.

test/test-renderer-dialog-freshness.js exercises the actual renderer's lazy exclusive repaint, post-processing source and ordinary PNG encoding with already-painted fixture backings. The final test fails against the exact pre-change renderer: the newer clipped presentation replaces the expected dialog pixel with the underlying primary pixel. It passes against the correction, including real newer covering presentation, later clipped write without stale revival, newer native repaint, hidden dialog/child, modeless owned popup, and overlapping front-window order. An initial child fixture omitted the absolute screen-geometry exports; those were supplied before the final old-source negative and passing regression. These pixel checks establish compositor behavior, not fresh Battlezone gameplay.

The mandatory flock /tmp/wine-build.lock bash tools/build.sh passed, as did the new regression and focused existing native-child-overlay, owned-popup clipping, debug scaling, GPU client-origin, worker repaint-boundary, lossless-PNG, frozen-control-CLI and callback-census tests, plus census negative controls. Only the declared lib/renderer.js:host-file census seal was refreshed after full-delta review; ownership and nonlocal holds remain pending. The observed regression actually read renderer SHA-256 a7c304c678a6a91622b693c8fb5416ace4c002245c3653b65fb18fdaf59d0924; all 31 observed read pins match afterward. All 4,063 scoped source/module pins and 153 fixture paths are unchanged between the post-repair preflight snapshots. This includes original bl_rel.exe 3,732,992 bytes / SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a and original Input.map. The local built module remains dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4; there is no newly loaded original-runtime module to attest. The scratch observer validation executes sync/async modules against whole readonly 512-MB memory, retains event 10,001 and preserves clock identity; actual monotonic clock reads are bounded by the before/after readings.

Prior driver/monitor status 143 was investigated using the retained termination record and launcher code. Original exit hook 0 does not establish a game crash; the historical signal sender/reason remains unknown. Urban's published diagnosis likewise leaves that cause unknown and records its detached finite supervisor's normal exit after 6,196 watchdog samples. This run prepared the same detached-session/finite-deadline approach with signal logging, heartbeat checks and owned shutdown. Its detached gate/disk watcher survived its short launcher; disk sampling remained above 2 GB. Finite sampling does not prove disk state between samples.

First native prerequisite failure, 17:17:48 UTC: immediate launch headroom was 2,179,108,864 bytes, below the required 2,200,000,000. Original attempts stopped with zero originals, no retry and no foreign cleanup. The independently tested production repair is retained. Fresh standard capture matching the readable actual mission-complete dialog, ordinary Next Mission/Launch, onward gameplay and any harvesting remain unverified with this correction. Historical dialog/native and black standard PNGs were inspected as prior evidence only. There are no new game input, release, neutral-dispatch, FPS, audio or full-suite claims. Owned watchers/helpers were stopped and checked absent; native box ID/state/expiry are N/A, with delegated own shutdown and ownership handed back.

2026-11-01: independent heartbeat stops fixed-capture preparation

Run scratch/runs/20261101-battlezone-fixed-capture-harvest/ launched zero originals. Fresh standard mission-complete capture and mission 2 arrival, harvesting or another objective change remain unverified.

Only scratch controller/route preparation ran while waiting for the exact urban-chaos-native-encounter release. The accepted startup/mission route, original Input.map, prior black/native dialog evidence, renderer repair notes and Urban detached supervisor were read. A detached primary supervisor and a separate detached heartbeat/disk enforcer were started. At 17:28:28 UTC, independent sample 465 observed primary counter 454 at 1.119649 seconds old, exceeding its one-second freshness limit. The enforcer stopped the owned supervisor with SIGTERM. The supervisor subsequently wrote counter 455 and recorded that signal; the cause of the heartbeat delay is unknown. This is a preparation guard failure, with no game process or game crash. The first-failure rule stopped all attempts; there was no retry, original, build, regression execution or observer execution.

The supervisor signal handler overwrote the initiating guard-failure.json. The initiating traceback survives in independent-exit.json, and the exact sample/timestamps survive in independent-guard.json; the stop record explains this overwrite. Primary sampling recorded a minimum of 2,178,494,464 bytes free; independent sampling recorded 2,178,514,944 bytes. Both exceed the 2,000,000,000-byte floor. Finite sampling, including the observed stale interval, does not establish unsampled disk history. No launch-headroom success is claimed because no launch was attempted.

After the failure, the exact Urban author release, fresh launcher final/result (exit 0), and direct 21 PID / 19 group absence checks allowed documentation and publication only. Dispatch main ebd313fd0 was repinned to actual main 2bc06c8a987c8b924e7b9c34447435089bbda2c0; no runtime gate was opened. Static post-failure read pins retain the original EXE, original map, renderer, regression, CLI and built-module hashes. These are local file reads, not fresh runtime/module/read-observer evidence. Previously accepted renderer/build results were consulted without duplicating acceptance tests.

Both owned PIDs and groups are absent, matching helpers and new child temps are absent, and delegated shutdown is complete. Local native box ID/state/expiry are N/A. Recovery is by reading the retained scratch scripts and JSON records; the runtime maximum remains one, but this assignment permits no new attempt. Only this note and ignored evidence changed. No FPS, audio or full-suite claim.

2026-11-01: guarded capture stopped at observer launcher prerequisite

Run scratch/runs/20261101-battlezone-guarded-capture-harvest/ launched zero originals. The repaired standard mission-complete capture and mission 2 arrival, harvesting or another objective change remain unverified.

Scratch preparation recovered the accepted startup/mission route, original Input.map, native/standard comparison and ordinary onward Scavenge controls. The exact urban-chaos-guard-diagnostic-encounter author release, fresh launcher final/result (exit 0), and direct cleanup passed at 17:49:57 UTC. Dispatch 7d47a40e5f518743ea7db2a5d4b6fb596084f1b5 was repinned to actual main c32bdc566aca2652f1d54a85b7173818b1105777; the runtime deadline was 18:14:57 UTC, within the overall 19:45 UTC limit. All 4,063 accepted source/module pins matched again at the gate, allowing reuse of the renderer worker's locked build and focused tests. No fresh build or regression ran.

After the gate, the tested Urban scratch guard implementation was adopted. Its retained healthy, delayed, stale, dead-process, replaced-identity and explicit-failure-file controls passed. Three-second monotonic freshness and process start-tick identity checks retain direct independent disk/process/deadline enforcement. First guard failure uses exclusive file creation, separate from later rejection and shutdown records. The prior one-second guard was not reused. Detached primary and independent guards advanced without a guard failure.

First prerequisite failure, 17:50:18 UTC: this worker mistakenly passed Node's offline.cjs observer validation to the detached Python runpy helper. Python rejected its first JavaScript line with SyntaxError; the observer never executed. helper-observer-terminal.json retains the full error, and firstfailure.json preserves the initiating failure separately from shutdown. The first-failure rule stopped attempts: no retry, original, gameplay input, loaded-module observation, fresh observer success or final game-release sample. This is a worker launcher error, not a renderer regression, guard failure or game crash. The prepared capture comparison was never executed.

Both guards stopped normally on the explicit owned stop request, after 131 primary / 130 independent samples. Each recorded a minimum of 2,158,096,384 bytes free, above the 2,000,000,000-byte floor. Finite samples do not prove unsampled disk history. No original launch-headroom success is claimed. Static file hashes do not establish loaded runtime identity or observed runtime reads. Owned process/group absence and temporary-file checks are retained in cleanup.json; local native box ID/state/expiry are N/A and own shutdown was explicitly delegated.

Recovery instructions preserve the failed scripts and require using Node for the .cjs observer in a separately authorized attempt, with a cheap detached launcher check before any original. No corrected launcher was replayed here. Only this note and ignored evidence changed. The remaining task is the actual standard/native dialog comparison followed by visible mission 2 objective progress. No FPS, audio or full-suite claim.

2026-11-01: standard mission-complete capture verified; mission 2 turret built

Run scratch/runs/20261101-battlezone-node-controller-gameplay/ verifies the repaired ordinary standard PNG, then advances mission 2 through resource gain, Recycler arrival/deployment and a completed defensive Turret build. One unchanged original ran; no production source or test changed.

At frozen batch 10810, dialog-standard.png visibly reads MISSION COMPLETED, the debrief and NEXT MISSION. The contemporaneous real native dialog backing (dialog-standard-native.png, HWND 0x10018, callback 0x5bd890, resource 130) has exactly the same decoded RGB bytes: 261,547 nonblack pixels, SHA-256 ff54a979824a960bdbb1b52bdd83fec55cac7d79ee35e4f6a6501cdbbaa1587c. This is the standard capture succeeding, with the backing used only for comparison. Whole 512-MB memory hash, EIP, API count and batch are unchanged around the frozen comparison. No repaint message, memory patch or replacement frame was used. The first mission's individual collection/escort mechanism is still not independently established by its displayed result.

Ordinary Next Mission at 90,15, then Launch at 550,15, reaches mission 2 at 11894 with scrap 08/30, pilot 10/13, and the Command Tower at 260 m. The live briefing explicitly orders fighting attackers and using the Recycler to build defensive turrets. Utility 3 lists two Scavengers; selecting 1 exposes 3 Scavenge. The matched neutral bracket is 12078/12103/12104/12124; the Scavenge input bracket is 12124/12149/12150/12170. At 13970, after 1,800 no-input batches, scrap still reads 08/30, but pilot capacity has risen to 10/15 and 5 Recycler has become available. Its live menu confirms a Recycler with Go To Geyser.

The selected Recycler's reticle explicitly offers Space Go To Geyser. The Space input bracket 14108/14133/14134/14154 follows the matched neutral 14062/14087/14088/14108. During the ensuing no-input observation, the large Recycler moves into the foreground. Scrap is 08/30 at 15054, then 14/30 at 15354 and remains 14/30 at 15954. Selecting Recycler again at 16046 identifies it at 29 m with a construction menu and 0 Pack Up, establishing arrival/deployment rather than only an issued movement order. This menu offers 4 Turret, costing 05 scrap. Pressing 4 has matched neutral 16046/16071/16072/16092 and input 16092/16117/16118/16138; scrap changes 14/30 → 09/30. After another 1,200 no-input batches, Defense becomes available and scrap rises to 12/30. The reviewed Defense inventory and final standard PNG at 17430 list 1 Turret. These establish a concrete mission objective change and resource gains beyond the predecessor's unchanged 08/30. The exact Scavenger delivery that caused each gain was not visually identified; do not equate the HUD increments alone with a proved harvesting mechanism or mission 2 victory.

Scratch preparation fixed the predecessor's interpreter mistake with explicit Node/Python argv arrays. Cheap detached controls retained durable exit-0, exit-7 and timeout receipts before any original. The exact urban-chaos-detached-controller-encounter author release, fresh launcher final/result (exit 0), and direct cleanup opened the gate at 18:09:17 UTC. Dispatch 19f8b03404cbb6a79bdedf02c33f9e39255610dc was repinned to actual main 7699f3fb00184b5d46c2cd98ad03569c5017318d; deadline 18:34:17 UTC. All 4,063 accepted source/module pins matched, allowing reuse of the renderer repair's locked build and focused regression/renderer/capture/census tests. Identical observer/validation bytes also allow reuse of the accepted sync/async whole-512-MB readonly, actual monotonic clock-bounds and late-10,001 retention validation. No duplicate build, production test or offline observer execution ran. The fresh frozen comparison supplies separate whole-memory readonly evidence for the actual dialog capture.

Urban's newer retained runtime failure exposed a 3.012139-second stale published primary heartbeat with matching PID/start identity; its publication delay cause remains unknown. This concrete scratch dependency was resolved before the original: six-second bounded freshness and atomic heartbeat replacement without per-sample fsync passed healthy advancing I/O, measured 3.5-second healthy delay, 6.15-second stale, dead-process, replaced-identity and explicit-failure controls. Immutable first-failure and terminal receipts retain durable atomic-link/fsync behavior. The detached primary supervisor and independent direct disk/process/deadline enforcement remain separate sessions. No unchanged one-second or failing three-second policy was replayed, and no historical signal sender was inferred.

The coordinator's direct same-task amendments at 18:11:51 / 18:16:44 UTC replace the original 2,150,000,000-byte launch margin with 2,100,000,000; the worker explicitly acknowledged this before launch at 18:17:07. Short board-tail checks initially missed the earlier direct answer; the full coordination tail recovered it. The 2,020,000,000-byte early stop, 2,000,000,000-byte floor and deadline stayed unchanged. Actual launch headroom was 2,135,273,472 bytes. Controller-sampled minimum was 2,049,708,032 bytes; independent minimum 2,049,712,128, primary minimum 2,089,832,448. Primary/independent histories retain 4,575/4,573 samples with largest timestamp gaps 2.228/2.403 seconds. Both guards stopped normally, without a failure. Finite samples do not establish unsampled disk history.

The original and driver both have fresh OS wait receipts 0, after 726.545 seconds of original runtime, with normal quit before the gate deadline; the supervisor and independent guard also have normal terminal receipts. All 18 actual detached commands have durable successful terminal receipts. All 48 key edges / 24 balanced pairs match actual dequeue, route and main-window DispatchMessage order. Final 17430 has empty known scheduled/pending/cross-thread/renderer queues, used physical/async keys and VK1 released. Pressed latches may remain set; numeric mouse-button mask, guest post queue and SDL consumption are unknown.

Before/after 4,063 source/module pins, all 153 fixture paths/tree members and 156 observed runtime-read hashes match. Actual compilation observes module SHA-256 dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4; the renderer read remains a7c304c678a6a91622b693c8fb5416ace4c002245c3653b65fb18fdaf59d0924. Original bl_rel.exe remains 3,732,992 bytes / SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a, with original Input.map. Read coverage is sync string/URL and complete bounded descriptor reads, not every async/direct OS read or compiler-source read evidence. 128 PNGs pass CRC/complete-deflate checks and the complete runtime gzip validates. Selected game milestones were visually inspected. Closed own journals were losslessly compressed with byte hashes and recovery commands; RECOVERY.txt preserves the controller recipes. Owned process, group, watcher and new-temp absence is checked directly in cleanup.json. Native box ID/state/expiry are N/A; delegated own shutdown is complete. Only this note and ignored evidence changed. No FPS, audio or full-suite claim.

2026-11-01: mission 2 defense attempt stopped by heartbeat read-order race

Run scratch/runs/20261101-battlezone-mission2-defense/ adds no game progress beyond the predecessor's deployed Recycler and purchased Turret. One original launched; the first real guard failure stopped it during startup, before the accepted route sent any game input. No retry ran.

The exact urban-chaos-allocator-probe-nov01 worker 01a19858-8ae8-77b0-92ac-ca7ae8890f24 author release, fresh launcher final/result and directly absent owned PIDs/groups opened the gate at 18:40:09 UTC. Dispatch 39242f4b76421b468157ee51d7593037fcb527de was repinned to main f1f89cb096421bbcd8ed13d2fa19ae50346fd5e7; the fixed deadline was 19:05:09. Disk initially prevented launch. Coordinator-owned lossless cleanup subsequently met the unchanged 2,100,000,000-byte launch margin, with the 2,020,000,000-byte stop and 2,000,000,000-byte floor retained. Launch main was f37832b9091a0b4774ad9296deebc7c752573e56, with 2,101,334,016 bytes free.

At 18:49:54.288751 UTC, immutable guard-failure.json records future-monotonic-heartbeat, age −0.009168345 seconds, with matching primary PID/start identity. The unchanged guard sampled its comparison clock at 1606203.522861289, then read a concurrently published heartbeat stamped 1606203.532029634. Its source samples the clock before reading the file; the retained values expose that read-order race. Later explicit-failure rejections and shutdown records remain separate from this initiating failure. The six-second stale limit was not exceeded. No guard correction or relaxed threshold was tried in this assignment.

Original 4036505 has OS wait 0 after 87.513 seconds, with the log ending at 258 batches and an orderly SIGTERM shutdown. Driver OS wait is −15 from owned supervisor cleanup. There are zero game-input records; the initial 280-batch step never acknowledged completion. No fresh PNG, mission 2 combat/objective evidence, native save or final frozen key sample exists. Previously reviewed standard PNGs and the original Input.map informed preparation only. This is a harness blocker, not evidence of a game crash.

All 4,063 source/module pins, 153 fixture paths/tree members, and 128 observed runtime-read hashes match. The actual compiled module remains dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4. Exact pins allowed reuse of the accepted locked build, focused tests and readonly observer validation; detached Node/Python exit-0, exit-7 and timeout controls passed in scratch. No production build, test or source changed. Primary/independent histories contain 2,300/2,297 samples, minimum free disk 2,080,223,232 bytes; finite samples do not establish unsampled history. Owned process/group, watcher and new-temp absence is checked directly. Native box ID/state/expiry are N/A; delegated own shutdown is complete.

RECOVERY.txt preserves the exact route and blocker. Before a separately authorized original attempt, correct and test the heartbeat comparison's read ordering while retaining genuine future/stale, identity, failure-file and independent disk/deadline enforcement. Then advance the existing Turret through ordinary positioning/deployment to visible defense combat or additional orders. No FPS, audio or full-suite claim.

2026-11-01: scratch heartbeat ordering repaired; second Turret placed and native saves retained

Run scratch/runs/20261101-battlezone-heartbeat-order-mission2/ fixes the demonstrated scratch guard race and advances one original through a second Turret's perimeter placement. Identified defense combat, a kill, mission 2 completion and a changed objective remain unverified. No production source, test, fixture, TODOS or STATUS file changed in this assignment.

guard.py now reads and parses the heartbeat before sampling fresh comparison clocks. race-regression.py deterministically publishes a heartbeat between the old clock sample and read: old ordering rejects age −0.010 seconds, corrected ordering accepts age 0. A genuinely future timestamp still rejects. The unchanged six-second policy passes the accepted healthy 3.5-second control and rejects 6.15-second staleness, dead/replaced PID identity and an explicit failure file. Immutable first-failure and direct disk/deadline negatives pass. The independent guard and supervisor bytes are unchanged; they retain direct disk, deadline and process checks. Accepted detached Node/Python dispatch exit-0, exit-7 and timeout evidence is reused with exact unchanged helper inputs.

Exact urban-chaos-empty-arena-repair-nov01 author worker 01a1986d-057c-7040-9e19-4255fef2e7e3 released runtime/source/build; fresh launcher final/result, run final/result and directly absent recorded owned processes/groups opened the gate at 19:02:26 UTC, deadline 19:27:26. Input 02b44b8944e97cd518c7349119037eb43bedccb5 was repinned to published main 0f9dfbb45eff6ded35a15bcf3a316c700594d831. The consumed production WASM is 7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0. Its allocator-only source/module delta is covered by that author's final locked build, repaired production/source regressions and five focused allocator tests; other accepted renderer/controller/observer inputs match. This run does not require the obsolete dd43e4a module. Original bl_rel.exe remains SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a.

The first replay wrapper encountered a scratch command-lock collision: the long-lived detached supervisor held the same lock. It failed before spawning a replay child. Its initiating failure and terminal receipt remain intact. Finite route commands then used a separate receipt/lock directory with the unchanged detached helper. The same original continued; no retry or guard failure occurred. The accepted route reached 10810, where the ordinary standard mission-complete PNG again exactly matches native dialog RGB and the frozen whole-memory/EIP/API/batch comparison passes. Next Mission, Launch, Scavenge, Recycler deployment and the first Turret build follow the proven route.

At 17522, the selected first Turret is 30 m away with Pack Up available. Its reticle offers Defend Storage Crates; ordinary Space is bracketed at 17568/17593/17594/17614, with matched neutral and actual standard captures. Later views show hull loss, earlier red radar contacts and streak/impact-like effects. These do not identify an attacking unit or prove a Turret kill. Forward-only/released-key observations also shift the view substantially; physical/async release alone does not establish guest SDL control consumption. Target acquisition and a clear view of actual defense combat remain the next specific gameplay blocker, rather than heartbeat freshness.

The second ordinary Turret purchase at 26560/26585/26586/26606 consumes scrap 30/30 → 25/30. Defense inventory at 27898 lists two Turrets. The shipped O objective display at 27990 explicitly calls for a defensive perimeter around the base and protection of the Command Tower and Solar Arrays; it is the current objective, not a newly completed one. The ordinary Backspace overview and pointer expose Space: Go To on nearby terrain. The second Turret receives that command at 28312/28337/28338/28358. After no-input observation, its selected standard capture at 32784 shows it 68 m from the player at the commanded point, with full displayed health and Pack Up available. This establishes useful ordinary perimeter placement beyond the predecessor's first build, without injected state or progress.

Ordinary Escape → Save Game → native Save As produces C:\m2 (81,024 bytes) and, after the perimeter placement, C:\m3 (66,050 bytes). The latter's SHA-256 is e5506fc41132b395c3e7de455a070996d159890cf93b3e581f3343c93dd8d3da. native-save-latest.json maps copied game-written VFS bytes to both guest paths; RECOVERY.txt gives the exact-file mount and ordinary Load Game proposal. The dialog did not append .SAV. Reload is untested; these are native saves, not VM snapshots. Only bytes were copied out; no guest/VFS mutation or reload ran.

Original 4052619 and driver 4052618 both have OS wait 0, original runtime 1,383.154 seconds. Supervisors also terminate normally. Final 33222 has empty known scheduled/pending/cross-thread/renderer queues and all used sampled physical/async keys plus VK1 released. There are 100 balanced key edges/50 pairs, 96 dequeued and 90 main-window DispatchMessage edges; the exact native Save As text/pause-release omissions are retained, with remaining edges matched in order. Mouse-button mask, guest post queue and SDL consumption remain unknown. No blanket all-key-dispatch claim is made.

Before/after 4,064 source/module pins, 153 fixture paths/tree members and 269 observed runtime-read hashes match; actual compilation observes the released module above. All 500 PNGs pass CRC/complete-deflate validation. Primary/independent histories contain 5,526/5,524 samples, sampled minima 2,024,783,872/2,024,787,968 bytes, with no guard failure. Launch free space was 2,114,555,904 bytes; launch 2.100 GB, stop 2.020 GB and floor 2.000 GB are unchanged. Finite samples do not establish unsampled history. Closed own journals were losslessly compressed with SHA roundtrips; same-byte PNG hardlinks retain every pathname and byte. Direct cleanup finds all 110 recorded owned PIDs/groups absent, no own helpers or new child temp directories. Native box ID/state/expiry are N/A; delegated own shutdown is complete. No further original ran. No FPS, audio or full-suite claim.

2026-11-01: native mission 2 reload verified; Fighter identified before disk stop

Run scratch/runs/20261101-battlezone-native-save-defense/ verifies ordinary native reload of the predecessor's C:\m3 save. Defense combat, a kill and the next objective remain unverified: the first genuine disk guard failure ended this run before the proposed next Turret command. No second original ran. Only this note is a tracked change; no production source, tests, fixtures, TODOS or STATUS files changed in this assignment.

The copied game-written save is exactly 66,050 bytes, SHA-256 e5506fc41132b395c3e7de455a070996d159890cf93b3e581f3343c93dd8d3da. native-m3.bin, save-mount.json and mounted-save-readonly.json preserve the host copy, structured argv and the actual mounted VFS bytes/hash. The launch passes one argv element --vfs-mount=HOSTPATH=c:\m3, preserving the backslash. This is native saved-game data, not a VM snapshot; no guest edits, injected progress or memory-state restoration occurred.

Initial Single Player → Load Saved Game is inert at both the label and button center. The same original therefore performs only a minimal initial mission launch, then Escape → Load Game, which opens the actual native Open dialog at 2008. No earlier mission progress is replayed. Its *.SAV filter hides the extensionless file, so ordinary input enters the full c:\m3 path at 2146–2176 and clicks Open at 2222/2247/2248/2268. After loading, the neutral standard capture at 3168 shows two Turrets in Defense, scrap 30/30, pilot 09/14, ammo 83, and the perimeter/protection objective. Ordinary Backspace overview and selection of Turret 2 produce 3352: the selected Turret is deployed 68 m away, with full displayed health and Pack Up available. These actual standard captures match the predecessor's inventory/perimeter progression; no pixel-identical or complete session restoration claim is made. See reload-verification.json.

Ordinary centering, F/C pitch, keypad steering and Q approach eventually expose the base and an enemy. After the shipped T frontal-target command, standard 7580 explicitly names a red Fighter, with the world target at about 121 m. Defense → Turret 2 at 7764 retains the identified Fighter at 107 m; the selected Turret's HUD offers Alt: No Action Fighter, rather than an attack command. The final reviewed 8198 capture still identifies Fighter. This establishes enemy acquisition beyond the predecessor's anonymous radar/effects evidence. It does not establish Turret fire, damage, a kill or an objective change. The proposed Space Go To command was rejected before any input was sent because the original had already stopped.

Exact urban-chaos-repaired-encounter-nov01 author 01a1987f-b5d1-7301-8472-b2038489254e release, fresh launcher/run final/result and direct owned process/group absence opened the gate at 19:38:52 UTC, with a 20:03:51 runtime deadline. Published main was repinned to c157668dc924a80c52a1fc28440a702eb8680cf9. Actual compilation observes 7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0; the accepted allocator build/regression/five-focused-test evidence is reused with unchanged inputs. Original bl_rel.exe remains SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. All 4,064 source/module pins, 153 fixture paths/tree members, and 271 observed runtime-read hashes match before/after. The unchanged readonly observer's accepted sync/async whole-memory/clock checks are reused. All 312 PNGs pass CRC and complete-deflate/scanline-length checks.

The scratch guard retains read/parse-before-clock ordering and the six-second freshness policy. Deterministic old-race rejection/new-order acceptance, genuine future, healthy 3.5-second delay, stale 6.15-second delay, dead/replaced identity, explicit failure, immutable first failure and direct disk/deadline controls pass. Detached Node/Python exit-0, exit-7 and timeout controls pass. Before the original, scratch shutdown is corrected so primary signals only the runtime, once; independent falls back only if primary is dead; the driver survives to reap the original, and supervisor finally records its actual driver wait. A simultaneous-failure dummy test records original OS wait −15, driver OS wait 0 and supervisor OS wait 0.

Scratch command failures are retained, without replacing their initiating receipts: a pre-original root-lock probe spawned no child; an ad wrapper's 120-second limit overlapped slow startup and expired during neutral observation before its click; a read-only inspection collided with an active command and spawned no child; an omitted center-mouse helper exited before input. Separate finite receipt directories and serial command completion then continued the same original. None is relabeled as an original retry or genuine guard failure.

First runtime blocker: at 19:54:33.322536 UTC, the driver's direct disk check measures 2,004,004,864 bytes, below the unchanged 2,020,000,000-byte stop threshold and above the 2,000,000,000-byte floor at that observation. Launch free space was 2,147,147,776 bytes; launch threshold remains 2,100,000,000 bytes. The writer and duration of the transient drop are unknown. Primary/independent histories contain 3,580/3,576 finite samples, with minima 2,058,579,968/2,054,955,008 bytes; these histories do not capture the lower initiating direct sample or establish unsampled history. No attempt continued after this first genuine failure, and no new native save was written. The verified native-m3.bin remains the useful recovery save.

Original 4093317 has actual OS wait −15, elapsed 901.199 seconds; driver 4093316 has actual OS wait 0, and the detached supervisor OS wait is 0. The observer's separate Node exit-hook value is 0, not a substitute for the original's OS wait. The revised scratch shutdown retained both actual waits during the real failure. The last complete live sample is 8198, not a post-stop final release sample: known scheduled/pending/ cross-thread/renderer queues are empty and sampled used physical/async keys plus VK1 are released. Tab/VK9 was not sampled. The trace has 40 balanced key edges, 35 dequeued and 34 main-window DispatchMessage edges; exact native Open text/pause omissions are retained in verification.json. Guest post queue, SDL consumption and mouse-button mask remain unknown.

Closed own journals are gzip/SHA-roundtrip preserved, and byte-identical PNGs are hardlinked without changing any capture path/bytes; recovery order is in closed-evidence-compression.json. Direct cleanup finds all 97 recorded owned PIDs/groups absent, no matching helpers or new child temp directories. Native box ID/state/expiry are N/A; delegated own shutdown is complete. RECOVERY.txt retains the exact ordinary reload route and stop. Next work requires separately authorized runtime with sustained disk headroom, then ordinary Turret positioning/defense against the now identified Fighter. No FPS, audio or full-suite claim.

2026-11-01: safe headroom recovery remains below launch prerequisite

Run scratch/runs/20261101-battlezone-headroom-engagement/ stops during static preparation, with zero original game launches. Lossless compression of 31 released Icewind installed-data BIF files reclaims 51,884,563 logical bytes. Direct free space after compression is 2,198,523,904 bytes; the final check is 2,196,598,784 bytes, 203,401,216 bytes below the required 2,400,000,000-byte sustained launch margin. That prerequisite was not established. The direct stop 2,020,000,000 and floor 2,000,000,000 remain unchanged. No low-margin original or duplicate disk observer ran. The predecessor's transient writer and duration remain unknown.

Exact artifact ownership was appended before mutation. Each source has a full SHA-256 and size; gzip staging in tmpfs passes a complete decompression/hash roundtrip. The destination's compressed SHA, fsync and durable per-file recovery recipe precede original unlink. compression-plan.json, compression-NN.json and compression-result.json retain exact paths and recovery commands. The multiply linked cd2/data/mvefile2.bif is retained: the first preflight rejected its three links before mutation, then static compression skipped it. No useful native save, active log, current Urban/root observer artifact or coordinator compression receipt changed. Tmpfs staging is removed.

pins.json records current main and on-disk module/native executable/save hashes; there are no actual runtime-consumption pins in this static-only run. The predecessor's uncompressed native C:\m3 remains 66,050 bytes, SHA-256 e5506fc41132b395c3e7de455a070996d159890cf93b3e581f3343c93dd8d3da; the executable remains SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. The previously verified ordinary reload route and exact controls are retained in the predecessor and referenced by this run's RECOVERY.txt.

Urban worker 01a198ad-10a1-7960-b622-4a394f86645c runtime/source/build ownership was never consumed or claimed. No original, driver, controller, observer or lifecycle changes ran; original/driver OS waits and final key/queue samples are N/A, rather than inferred from predecessor evidence. Native box ID/state/expiry are N/A; no own runtime remains. Identified Turret combat, damage, kill and next objective remain unverified. The next concrete prerequisite is additional authorized recoverable disk savings sufficient to establish sustained 2.40 GB free space, followed by the exact Urban release gate. Only this note is tracked; no production/source/test/TODOS/STATUS edits.

2026-11-01: closed worker logs recovered; launch margin still insufficient

Run scratch/runs/20261101-battlezone-closed-log-headroom/ stops at the disk prerequisite with zero originals and zero drivers. One bounded census of immediate scratch/usage-watch/ output directories finds 1,685 with terminal result.json and final.txt, containing only 27,142,319 bytes of remaining events.jsonl/stderr.log files in total. Even their entire size cannot close the launch gap. Current workers, this output, coordinator rollout, font/laptop logs and uncertain owners are excluded from compression.

51 logs with explicit author release records and directly absent worker/ launcher processes are losslessly compressed, reclaiming 16,752,561 logical bytes. Exact board OWN records precede mutation. Streaming gzip staging uses tmpfs; each file has full original and compressed SHA-256, complete gzip roundtrip, durable destination and recovery recipe before original unlink. compression-plan.json, compression-NNN.json and compression-result.json retain the exact paths and recovery argv. A final independent read of all compressed bytes repeats both hashes and decompression checks. Protected final/result/notification hashes match; dedup, recovery data, current Urban and coordinator artifacts are untouched. No useful native save or Icewind data is compressed again, and no tmpfs staging remains.

The final direct free-space check is 2,202,288,128 bytes, still 197,711,872 bytes short of the required 2,400,000,000-byte sustained margin. This is the first blocker; no unchanged low-margin original is attempted. Stop 2,020,000,000 and floor 2,000,000,000 remain unchanged. The historical transient writer and duration remain unknown; no duplicate disk observer runs. The coordinator retains shutdown ownership of its observer.

pins.json records published main af500af4161fc1f36d591a9105fcb3f4eaa14453, verified directly over strict known-host SSH, and the on-disk module, native executable and save hashes. The module remains 7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0; the executable and uncompressed 66,050-byte native C:\m3 save retain the full hashes recorded above. Actual runtime consumption is N/A in this static-only run. The predecessor's verified ordinary reload route and exact actual-input controls remain preserved; no packaging checks or game tests are repeated and no predecessor input/queue evidence is presented as a fresh run.

Urban worker 01a198ad-10a1-7960-b622-4a394f86645c posts exact author DONE/PUSH RELEASE at 20:12:17 UTC; its terminal launcher final/result are observed. Runtime/source/build ownership is nevertheless never claimed or consumed, because the disk prerequisite fails first. Original/driver actual OS waits, final key/queue samples and native box ID/state/expiry are N/A. Direct process checks find no compressed worker/launcher matches; no owned runtime/helper remains. RECOVERY.txt preserves the stop and recovery recipe. Identified Turret combat, damage, kill and next objective remain unverified. Next requires additional authorized recoverable disk savings beyond this exhausted immediate-log scope, then sustained launch margin and a fresh runtime gate. Only this note is tracked; no source, test, TODOS or STATUS changes.

2026-11-02: inherited defense engagement closed; firing verified, outcome unproved

Runs scratch/runs/20261102-battlezone-native-defense-resume/ and scratch/runs/20261102-battlezone-existing-defense-handoff/ preserve one original engagement. After the predecessor worker unexpectedly terminated 143, its existing driver 10498 and original 10502 were recovered live at batch 10043. The successor launched no original and preserved the old worker events read-only. A pre-game bootstrap had exited 1 for a missing generic stdole2.tlb; the retained correction supplies the hash-verified local file before the actual original starts, rather than replaying a game.

The tested source snapshot is a8b85a5696143fc821d4de2cc018d2661e85ec1f. The actual compile observer records WASM SHA-256 68cccb08e53a59d84c0ec566ea0fa5056d5e30aab40ef9d89b55c4e6d3a2ae69, reusing COPY's accepted locked build and tests; no fresh build or source change was needed. The predecessor verifies 4,375 accepted source inputs before materialization. All 4,696 materialized runtime files match again after the run, as do 154 observed file reads. All 153 original fixture file pins match locally and in the runtime manifest. The observed original executable is bl_rel.exe, SHA-256 0554156663000fc1c0c0ad2275e460fa6fc9acda4abdd26c720b69019d04617a. The retained native save is 66,050 bytes, SHA-256 e5506fc41132b395c3e7de455a070996d159890cf93b3e581f3343c93dd8d3da. identity.json, runtime-inputs.json.gz, runtime.json, loaded.json and observer-events.json preserve exact identities and argv. Execution uses Node 24.22.1, test/run.js --app=battlezone_demo --no-build --batch-size=100000 --control-stdin --frozen --no-close --quiet-api --quiet-blocks --trace-input-dispatch --max-seconds=1800 --max-batches=100000, the pinned --wasm and exact-file native-save mount. This is the headless CLI route.

Ordinary Single Player → mission launch → Escape → Load Game → native Open dialog loads c:\m3; batch 2458, native-reloaded.png, shows the two Turrets, scrap 30/30, pilot 09/14, ammo 83, and the perimeter/base protection objective. Steering, pitch, approach and frontal targeting use ordinary keypad, F/C, Q and T inputs recorded in controls.ndjson and remote-controls.ndjson. identify-grey-tank-after.png visibly identifies a red Tank at 57 m. This run's acquired hostile is a Tank; the earlier Fighter observation belongs to the prior run.

Ordinary keypad-0/VK96 fire sequences at 8070–8191, 8191–8462, 8484–8905 and 8930–9351 consume ammo 83 → 48 → 29. Reviewed player-fire-before.png, aimed-defense-fire-held.png and aimed-defense-fire-after.png show the selected hostile, a player cannon plume/tracer during held fire, and retained ammo 48 after release. Player firing and weapon consumption are verified. Tank damage, a hit, a kill, Turret fire and defense completion are not established.

The predecessor's Space command at 9397–9443 and neutral 9743/10043 observations are retained without upgrading them to confirmed Turret placement. The successor first observes 10043/10343, then explicitly opens Defense and selects Turret 2 with two ordinary 2 presses. At 10435, reviewed handoff-select-turret2-after.png shows the selected Turret 39 m away and Space: Go To. Space at 10435–10481 is followed by small neutral steps to 10781/11081. These prove command input and subsequent observations, not arrival or an attack. Ordinary O at 11081–11127 displays the same instructions: build Turrets, establish a defensive perimeter, and defend the Command Tower and Solar Arrays. The Tank remains visible; no changed objective or mission result is proved. No injected state, progress patch or new save was used. The verified imported native-m3.bin is retained for recovery.

At final batch 11127, known scheduled/pending/cross-thread/renderer queues are empty; guest post queue, mouse-button mask and SDL consumption remain unknown. The default sampled key set covers only VK16/17/18/32, all physically and asynchronously released, with VK32's pressed latch still true. It does not prove every used key released. The ordinary controller quit yields original OS wait 0, runtime 1,122.510 seconds, no driver failure reason, and directly absent driver/original PIDs. A separate driver OS wait is unavailable.

All 498 archived file hashes match after retrieval, and 248 PNGs pass CRC and complete-deflate checks. evidence.tar.gz is 37,353,996 bytes, SHA-256 1270449562b60a70cc39972bad4c77865761f1d9e1e0ee0d056dae2c9b3ef5af; its extracted files and manifest are in the successor's retrieved/ directory. Only named captures above were visually reviewed; integrity is not gameplay acceptance. The inherited host disk guard 1090046 exits normally after its finish marker, with sampled local minimum 2,474,991,616 bytes. The remote driver records minimum 58,190,594,048 bytes; no disk-stop receipt occurs. These are finite observations, not an unsampled-space guarantee.

After retrieval and verification, explicitly delegated shutdown of known boat bx_kb6pdbx2 is confirmed STOPPED at 2026-11-02T12:50:56.118Z; its original expiry was 2026-11-02T13:28:54.998Z. Remote recorded bootstrap, driver and original PIDs are absent, and the scoped helper scan is empty. Host guard and own board watcher cleanup are recorded in cleanup.json. result.json, final.txt and RECOVERY.txt preserve the limited outcome and handoff. Only this note changes in git. No second original, compatibility edit, downloaded fixture, audio/FPS/full-game or wider-corpus claim follows from this run. Further combat work requires a separate assignment; this engagement ends with verified firing and an unproved defense outcome.

Not checked

Ad screen visibility (black headless), audio, the AVIs, multiplayer, the individual Scavenger harvesting/delivery or escort mechanics beyond the resource gains, Recycler deployment and Turret build above, mission 2 victory, WebKit.