Urban Chaos demo
Source https://archive.org/details/urbanchaos, urbanchaos.zip26,328,920bytes, SHA1 ba06b6ca040359ac07dd60ca75d680c9f2b5d15c verified after resumed download 2026-10-12 05:10Z. Public endpoint returned500; direct endpoint https://ia601406.us.archive.org/34/items/urbanchaos/urbanchaos.zip resumed the partial8.9MB file. Evidence0504Z preflight and0509Z recovery under scratch/runs/.
ZIP contains Demo.exe26,397,184bytes (WinZip SFX) and readme.rtf. 7z extracts 387files; unshield extracts752files from installer/data1.cab. Flatten the three groups Global_Media, Local_Media_-English, Program_Files into game/, preserving relative paths. Also copy ALL three installer directories clumps, data, talk2. Fallen.ex is already PE32, not compressed; copy as Fallen.exe without modifying bytes. Executable SHA256 8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229. Remote assets /home/user/urban-chaos-demo-preflight on bx_44dwy4jt snapshot.
Installer configuration is required
0512Z first run omitted clumps and config.ini, and displayed Cannot locate Urban Chaos CD-ROM. Trace reads D:/clumps/mib.txc after LocalInstall defaults0. The installer script explicitly writes config.ini sections Game, LocalInstall and TextureClumps: keys language, textures, sfx, speech, movies, enable_clumps. setup-config.log records original script offsets; enable_clumps is literal1. For a full local extraction, reconstructed config uses LocalInstall entries1, TextureClumps enable_clumps1, and Game language=text/lang_english.txt (use Windows backslash in actual INI). No guest or emulator code changed, no media fabricated.
0514Z with local files/settings passes media lookup and opens graphics/audio setup, but initial language=0 causes missing-language labels. The actual English file is already present at text/lang_english.txt; Game.language is a path, not a language number.0516Z corrected config renders legible settings: primary3Dcard, 16-bit colour,640x480, MicrosoftDirectSound3Dhardware by default. Screenshots reviewed. Setup Okay is ordinary graphics/audio selection, not a licence dialog.
Module SHA256 9e835f75dfe5d2fcfb91f592641e945f91baadcd20f6751460777f4dba888cd6; remote build PASS. This is the retained test branch build, not a claim of testing clean main. Main0304a8549 integration remains ops-owned. ole32.dll real-file warning retained; no ordinal trap seen so far. Readme documents arrow movement/turn, Space jump, Enter enter vehicles; further controls and gameplay/browser/touch/audio/FPS still unverified.
0517Z-settings: ordinary click275,445 at100 accepts the displayed graphics/audio settings and reaches the Urban Chaos Official Demo Press Any Key screen. Reviewed640x480 capture, exit0 at40-second bound/583batches. No licence question was displayed or answered. Ordinary menu navigation is the next test.
0518Z character-only keypress13 at200 leaves title unchanged.0519Z normal keydown13 at200/up230 advances to mission map (Urban Chaos Demo marker/Start). Reviewed screenshot, exit0 at40s/2441batches. This was input semantics, not a proven emulator input bug; use held keys for this game. Next Start on mission map.
0520Z: add Enter down500/up530 to reach demo mission briefing (objective: eliminate/arrest targets and save hostages). Reviewed image, exit0 at60s/2141batches. No gameplay qualification yet. Next held Enter after the briefing is visible.
0521Z: add Enter down800/up830 to enter a rendered3D mission opening/camera sequence. Reviewed rain/street/building scene with objective caption and speaker portrait, exit0 at80s/2131batches. This is level rendering, not yet player input. 0509Z prepared-files.json records1122files/40,299,202bytes including reconstructed config.ini, plus the original readme.rtf. Original game file bytes unchanged.
0523Z add Enter2200..2230: final2330/120seconds shows D'arci in the player scene, radar/health/4-minute countdown HUD visible. Planned2450 screenshot and2500 movement never execute before the time bound; before.png is therefore absent, not missing transfer evidence.0526Z earlier2250 capture is STILL cinematic; Up2260..2290 happens before the playable HUD. Final2350/124.6seconds shows the same starting position as0523Z, so no player movement is verified. These two runs establish transition to the playable scene, not successful input.
Next use control-stdin or browser HUD-gated input (not another guessed early batch), then reviewed displacement/control comparison. Registration, browser, touch, audio/FPS remain unqualified. No known missing payload after full1122-file layout; config is reconstructed from installer strings and successful behavior, not a captured completed installation. No source/emulator/binary changes.
2026-10-12 05:37Z live HUD-gated movement
Evidence scratch/runs/20261012T0537Z-urban-chaos-live. Same module/executable hashes as above; remote build passed. CLI --control=8123 --frozen reaches batch 2370 using the documented menu route. Reviewed before.png shows D'arci on the street with radar/health and3:58 timer. Only after reviewing that capture, queued keydown38, stepped110 batches, queued keyup38 and stepped10 more. after-up.png shows D'arci running into the car park past the alley, timer3:56: ordinary player movement verified. Control replies confirm batches2370->2490; explicit quit exits0. Frozen stepping is a diagnostic route, not a browser performance measurement. The module was available when the runner loaded it; build finished shortly after runner launch. The recorded post-build hash matches prior tested module9e835f75.
Added local-only urban_chaos_demo / urban-chaos-demo registration on the separate codex/dr2-registration branch. Remote --prepare passes1121 data files, including config.ini; executable is separate. Initial preparation used the wrong fixture root demo-installers and failed clearly; corrected root is test/binaries/candidates. Static syntax/manifest/local-only checks pass. Browser launch/touch and audio/FPS are still pending; no public desktop entry or deployment.
Exact reconstructed config.ini (Windows path separator):
[Game]
language=text\lang_english.txt
[LocalInstall]
textures=1
sfx=1
speech=1
movies=1
[TextureClumps]
enable_clumps=1
Registered Chromium gameplay and touch (2026-10-12 05:50Z)
0550Z-urban-chaos-browser: Chrome155.0.8059.39/WebGL/worker threads, module 9e835f75..8cd6, registered urban_chaos_demo. Ordinary setup Okay, held Enter through title/map/briefing reaches player HUD (step4). ArrowUp3seconds moves D'arci toward the alley confrontation (step5). At844x390, real CDP touch on the direction pad held upward7seconds advances D'arci farther into the alley (step6 before, step7 after). All images reviewed, stage complete05:54Z. No source fix required. Buttons fit at844x390; other action bindings, physical phone/Safari, audio and FPS are not qualified. AudioContext autoplay warnings remain in run.log; no claim of working sound.
Initial0549Z attempt never launched: server rejects out-of-checkout fixture symlinks. Original payload existed, but game/ linked to a custom remote directory. Copied original files into test/binaries/candidates/urban-chaos-demo/game and replayed as0550Z; do not weaken server containment checks. Keep real files inside the served checkout for browser preparation. Failed capture/log retained.
Integration reconciliation, 2026-10-13
Recovered from origin/codex/wsasocket-capability at 15b37458e during the
release merge audit. The local-only registration and its required runtime fixes
are now on main. Earlier statements that integration is pending describe the
historical run, not the current branch. This recovery adds no fresh gameplay,
audio, performance, Safari, or physical-device qualification; retain the
recorded run IDs, module hashes, limitations, and failed attempts above.
Automated recovery, 2026-10-14
Fresh ZIP acquisition from https://archive.org/download/urbanchaos/urbanchaos.zip verified26328920 bytes and SHA-1 ba06b6ca040359ac07dd60ca75d680c9f2b5d15c. Unpack Demo.exe with7-Zip23.01; extract data1.cab groups Global Media,Local Media - English,Program Files using the existing is5-extract fallback (unshield absent on this host). Copy component contents into game/ in that order, then installer clumps/,data/,talk2/ preserving relative paths. Fallen.ex_ is in installer/, NOT in Program Files; copy its unchanged bytes to game/Fallen.exe.
The exact config.ini shown above is written as documented reconstructed installation state. It is not original captured installer output. The1122-file reconstructed game tree matches the restored accepted fixture byte-for-byte, including that INI. Every other file also matches its original extracted source; EXE SHA256 remains8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229. The source archive retains readme.rtf outside game/.
Recovery: node tools/fetch-candidate-corpus.js --id=urban-chaos-demo; pristine verification: node test/test-urban-chaos-recovery.js. test/candidate-corpus/urban-chaos-demo-files.json pins all1122 paths,sizes and SHA256 values and labels config.ini reconstructed. First draft failed explicitly on a wrong Fallen.ex_ source path; corrected recipe was rerun from fresh forced acquisition. Mutable files may differ after playing. Private-fixture restrictions remain; no binary, public deployment or fresh gameplay acceptance is part of this recipe change.
2026-11-01 confrontation attempt: worker control timeout
Evidence: scratch/runs/20261101-urban-chaos-confrontation/. No new combat or
interaction success. One original launched; zero gameplay inputs. The accepted
street/alley movement above remains historical evidence and is not counted again.
Gate verified before tests/runtime: exact predecessor-authored
battlezone-target-hit-command DONE/PUSH RELEASE at 14:37:30 UTC, fresh launcher
final.txt/result.json with exitCode 0, and direct absence checks for all six
owned PIDs and four process groups. Actual main at the 14:43:28 gate was
2f5a350f043ace3d7db5791f09b10e8ada2e5644, advanced from dispatch
ad94d102f5b4a239e86d7b38970455dada45dabe; gate deadline was 15:08:28 UTC.
All 1,122 game files match the bundled file manifest, including the explicitly
reconstructed config.ini. Original Fallen.exe SHA256 remains
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229.
The complete accepted build comparison matched all 4,392 source pins and 141
recorded read pins from accepted build d86935e8274cd5d70a53bcc50c5d8bf5632d95c3,
so no fresh build was needed. Actual runtime WebAssembly.compile observation
recorded module SHA256
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Before/after source inventory, original tree, manifest and observed file-read
identity comparisons are retained separately. The observer records JavaScript
file reads and module identity, not compiler reads, every lazy read API, or
guest-memory contents. Offline validation passed actual sync/async execution,
whole 512 MB shared-memory read-only comparison, host/native clock identity,
and emitted late-event retention through event 10,001.
Native frozen CLI used registered urban_chaos_demo, software rendering,
--no-build --wasm=build/wine-assembly.wasm --batch-size=100000 --control-stdin --frozen --trace-input-dispatch. Ping, step25, input-state and
the batch25 PNG succeeded. Reviewed readiness-25.png is a plain teal 640x480
startup surface. The next request, step75 toward batch100, received no reply
within the worker helper's 120-second bound: this is the first prerequisite
failure. The final log nevertheless records 98 batches in 173.990 seconds;
this establishes progress, not a game freeze or emulator defect. No repair,
bypass, retry or second original was attempted. An automatically following
capture command queued a read-only host renderer key/button expression after
the timeout; it also received no reply and did not access guest memory.
Ordinary quit was requested, then SIGTERM requested orderly shutdown; the
runner exited 0 after 176.9995 seconds. Reviewed final.png is identical to
the readiness image. Setup Okay, title/map/briefing, playable HUD and alley
were not reached. The actual bundled English text names Kick, Punch, Action
and Controls, but supplies no binding values; the historical external
readme.rtf is absent from this local candidate tree. No combat binding was
invented or sent. The controls screen was not reached.
The batch25 input sample has zero scheduled/enqueued/dequeued/routed/dispatched gameplay events and empty known queues. Per-key physical/async samples and mouse buttons were unknown at that startup boundary. Final released per-key/button samples, before/held/released/settled combat captures and a matched neutral control were not obtained; no action was attempted. Guest post queues and SDL consumption remain unknown. The existing ole32.dll real-file warning remains in the log; no ordinal trap was observed. This run qualifies neither audio, FPS, browser/GPU/touch nor the full game.
Disk guards and controller sampled above 2,000,000,000 bytes free, with finite sampling gaps; no RAM gate or foreign cleanup. Local box ID/state/expiry are N/A. Owned runtime shutdown was explicitly delegated. Cleanup, publication and final/result records in the run directory document ownership handback. The concrete blocker is the worker's startup step timeout; any further attempt requires a separate assignment rather than continuing this stopped run.
2026-11-01 small-step confrontation attempt: startup readiness bound
Evidence: scratch/runs/20261101-urban-chaos-small-step-confrontation/.
No new confrontation or interaction outcome. One original, zero gameplay inputs;
historical movement is not counted again. Only scratch controllers changed.
The exact battlezone-target-hit-corrected DONE/PUSH RELEASE author line at
15:05:43 UTC, terminal launcher final/result (exitCode 0), and direct absence
checks for eight predecessor PIDs and five groups passed the gate at 15:06:13.
Actual main was 60500ae55fa28dc658402496afb6e36f88d4296f, advanced from dispatch
c6fba571254a85245166115a08880cec00dc911d; the gate deadline was 15:31:13 UTC.
All 1,122 manifest paths matched, including explicitly reconstructed config.ini;
Fallen.exe retained SHA256
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229.
Complete accepted build reuse matched 4,392 source pins and 141 recorded reads;
no new build was required. Actual runtime WebAssembly.compile observed module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Fresh offline observer validation passed actual sync/async execution, whole
512 MB shared-memory read-only comparison, host/native clock identity and late
event retention through 10,001. Before/after comparison passed 4,444 source
identities, 1,124 fixture/manifest identities and 132 actual observed reads.
Observation covers sync whole-file and fully covered bounded descriptor reads;
async file APIs, compiler reads and guest-memory content identity are not covered.
Native registered CLI retained software rendering, --no-build --control-stdin --frozen --trace-input-dispatch, batch-size 100000 and the accepted observer.
The actual stdin protocol was read before launch. All 31 sequential step requests
were exactly 25 batches, each with a 180-second controller bound; every reply
arrived. The earlier step75/120-second request was not repeated. Final batch775
advanced in 1,379.457 runtime seconds. All 32 PNGs have identical decoded pixels:
the reviewed plain teal 640x480 startup surface. PNG CRC/deflate and full gzip
validation passed. Thus the controller sizing fix permits continued stepping,
but Setup Okay, title/map/briefing, HUD, controls and alley remain unreached.
The first blocker is visible startup readiness not reached before the controller's shutdown reserve. The incremental loop stopped at 15:29:17; its 23-minute outer guard then requested ordinary quit, received acknowledgement, and requested orderly SIGTERM. Exit0 after 1,380.891 seconds was inside the 25-minute gate bound. No individual step timeout or crash occurred; batch progress does not establish a game freeze or its cause. The exit log reports one main-owned held critical section; its causal significance is unestablished. No second original or bypass.
Final batch775 input counters are all zero, with known scheduled/pending/ cross-thread/renderer queues empty. Explicit host per-key physical/async/pressed and mouse-button samples remain unknown; released state cannot be claimed. Guest post queues and SDL consumption are unknown. The bundled language file names combat actions without binding values; no binding was invented. Combat before/held/released/settled captures and matched gameplay neutral are unavailable because readiness failed. No audio, FPS, browser/GPU/touch or full-game claim.
Controller minimum free disk was 2,041,864,192 bytes, above the required floor; the separate guard and publication samples are retained with finite sampling-gap limitations. No RAM gate or foreign cleanup. Local box ID/state/expiry are N/A; owned shutdown was explicitly delegated. Cleanup/publication/final/result records document absent owned processes, groups, watchers and new child temps, and handback. Further work requires a fresh assignment to investigate the startup-readiness blocker; this attempt stops without source or configuration changes.
2026-11-01 startup diagnosis: RSX wait localized; trace host abort
Evidence: scratch/runs/20261101-urban-chaos-startup-readiness/.
The saved startup state localizes the pre-Setup wait to the Miles RSX provider;
the underlying reason it does not finish remains unproven. A fresh targeted
trace aborted in Node with JavaScript heap exhaustion before the first step25
reply. This is the first substantive failure of this attempt. One original,
zero gameplay inputs, no retry, no Setup/title/HUD or new gameplay acceptance.
The exact battlezone-scavenger-arrival DONE/PUSH RELEASE at 15:42:03 UTC,
fresh terminal launcher final/result (exitCode 0), and direct absence checks
for its six PIDs and five groups passed the gate at 15:42:09. Actual main was
a5981a6108e86462ba3379f9decf7e84dfa8e82d; runtime deadline 16:07:09 UTC.
All 1,122 manifest paths matched, including explicitly reconstructed config.ini;
Fallen.exe retained SHA256
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229.
This is reconstructed installation state, not a completed upstream installation.
Named hypothesis H1 was recorded before launch: distinguish an omitted native
dialog from a pre-dialog audio-provider wait. CLI controlPng copies only
renderer.canvas; scheduled dlg-png instead copies a visible dialog's
_backCanvas. Therefore teal captures alone cannot establish absent Setup.
The predecessor's actual batch25 through batch775 input-state records keep
main waiting indefinitely on handle 0x0e000042, EIP 0x023e9943, with an
RSX worker repeatedly near 0x023ea0f9. Its log records only the zero-size
Mucky Foot Shell window, audio-provider loading and an rsxtool.exe ShellExecute.
Original MSSRSX.M3D SHA256 is
8639a7812e87738040f487693baca9778689130e8fc9882de91bae2154edc827.
Recorded DllMain 0x02400f7c minus PE entry RVA 0x21f7c gives runtime base
0x023df000, preferred base 0x25000000. Saved main/worker EIPs translate
to 0x2500a943/0x2500b0f9. Aligned disassembly and original PE imports
identify the cleanup path: EnterCriticalSection, signal the object's event at
offset 0x88, then WaitForSingleObject(INFINITE) on its completion event at
offset 0x94. The worker loops over WaitForMultipleObjects on two events.
This identifies a wait chain; it does not prove why its signal is not completed,
that rsxtool.exe is the missing prerequisite, or an emulator deadlock.
Fresh native CLI retained --app=urban_chaos_demo --no-build --control-stdin --frozen, software rendering and batch-size 100000, adding selected window,
load/event/wait API tracing and register_dialog_frame tracing. Ping at batch0
succeeded. The only step request was 25 batches with a 180-second reply bound.
The retained log contains 13,351 RSX worker WaitForMultipleObjects entries,
the zero-size shell creation, rsxtool.exe request and SetEvent(0x0e000041)
at return 0x023e9943. Node then reports “Allocation failed - JavaScript heap
out of memory” and exits by SIGABRT (-6), after 45.524 seconds. Trace overhead
may contribute; it was not isolated, so no trace-allocation root cause is claimed.
No window inventory, dialog backing capture, final PNG or final input-state
sample completed. Capture omission remains unresolved; no blind Setup click,
licence input, guessed binding or memory write was attempted.
Complete build reuse matched 4,392 accepted source pins and 141 recorded reads;
no new build was needed. Canonical module before/after was
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Fresh offline observer validation, with explicit SWEEP_CASE_DIR, passed actual
sync/async execution, whole 512 MB shared-memory read-only comparison, clock
identity and emitted late-event retention through 10,001. The original's abort
prevented its exit observer from saving actual reads and compile events: those
runtime identities are unavailable. Offline records are explicitly renamed;
they are not original-runtime observations. Before/after inventories match
4,444 source identities and 1,124 fixture/manifest identities. These checks
do not establish compiler reads, all async file reads or guest-memory identity.
The original died before input sampling; released per-key/button state and consumed releases are unavailable. No action was sent; combat controls and before/held/released/settled or matched gameplay-neutral evidence were not reached. Disk samples stayed above 2,000,000,000 bytes, with finite sampling gaps; no RAM gate or foreign cleanup. Native box ID/state/expiry are N/A; owned shutdown was delegated. Cleanup and publication records document absent owned processes/groups/watchers/temps and ownership handback. Next work needs a fresh assignment with bounded diagnostic logging before native-window inspection; this stopped attempt makes no production fix or new acceptance.
2026-11-01 bounded RSX diagnosis: mixed named-event wait fails
Evidence: scratch/runs/20261101-urban-chaos-bounded-wait/.
The pre-Setup completion wait now has a precise generic API failure:
ThreadManager.waitMultiple rejects an infinite wait containing a valid
named-authority event before checking an already signaled unnamed event.
The original RSX worker receives 0xffffffff (WAIT_FAILED) repeatedly,
leaves the stop event unconsumed, and does not signal completion. No production
repair was attempted. One original, zero gameplay inputs, no second original;
Setup/title/gameplay remain unreached and historical movement is not recounted.
The exact coordinator D3D9-COM-GAP DONE RELEASE runtime/source/build author
line at 15:51:39 UTC, its terminal result (exitCode0), and direct absence of
the owned COM test process passed the gate at 15:52:36 UTC. Actual gate main
was 73e88746fc8672d921c6bf274937ba904db702ef, advanced from dispatch
a1c43c412f2406bb58081ad4fa654b36b2973af3; deadline was 16:17:36 UTC.
Hypothesis H2 disabled selected API/host tracing and retained only first64, last64, selected-tail128 and at most256 counter keys around real event/wait calls. Wrappers invoke the original method once and return its exact result; they sample synchronization state without writing it. Offline comparison exercised actual event/wait methods, identical results and whole512MB memory with/without wrappers, plus late10,001 calls and retention caps. The separate identity observer passed actual sync/async execution, whole512MB read-only comparison, clock identity and late10,001 retention with correct SWEEP_CASE_DIR. No heap increase, unbounded trace retention or tracked helper changes.
Native registered CLI used --app=urban_chaos_demo --no-build --control-stdin --frozen, software rendering and batch-size100000. Step1 and step24 both
replied within180 seconds; no request exceeded25 batches. Native WAT window
inventory, renderer inventory and dialog globals at batch1 and batch25 all
show only the invisible zero-size Mucky Foot Shell window, no active dialog
or dialog backing surface. Reviewed PNGs are plain teal640x480. No blind
Setup input was sent. Native inventory resolves the omitted-dialog hypothesis
at these sampled boundaries, without claiming every intermediate state.
At batch25, read-only evaluation retrieved the original main-module probe
directly as main-waits-batch25.json. Its64-record tail contains63 actual
worker waitMultiple(2, 0x08048bc8, false, INFINITE, 4) returns of 0xffffffff,
followed by main's completion-event poll returning258. The two handles are
0x0e000041 and 0x0e000040. Before and after every retained multiple wait,
the first is an auto-reset event with state1; the second is a valid issued
named-authority event (sync type4). Thus the ready event is not consumed.
Main is waiting indefinitely on unsignaled 0x0e000042, EIP 0x023e9943.
The sampled original object maps offsets0x88/0x8c/0x94 to those three handles.
Original aligned RSX code tests the wait result at preferred0x2500b0f9:
return0 takes the completion SetEvent path, return1 handles the other event,
and other returns loop. This links the actual failed returns and missing
consumption to the saved completion wait; it does not establish every possible
later startup prerequisite.
Exact implementation: lib/thread-manager.js:2034 checks all handles for
issued named events and returns 0xffffffff when timeout is nonzero, before
the wait-any readiness/consumption loop at2100. CLI test/run.js:4806 routes
this call there. src/09a0b-handlers-base-late.wat:1949 stores that result in
guest EAX; only internal0xffff yields. The next bounded repair should address
this generic mixed-event wait contract under the existing holds, then rerun
this reproducer and the original. No callback, concurrency, SEH or scheduler
algorithm was changed in this assignment.
Minimal real-API reproducer: flock /tmp/wine-build.lock timeout 120s node scratch/runs/20261101-urban-chaos-bounded-wait/reproduce.cjs. It compiles a
separate test module using production CreateEventA/SetEvent/
WaitForMultipleObjects/WaitForSingleObject handlers and the real named-event
authority. It creates an unnamed auto-reset event and an unsignaled named
event, signals the first, then waits-any on both. INFINITE returns0xffffffff;
a subsequent single poll returns0 then258, proving the failed wait did not
consume the signal. After a fresh SetEvent, the matched timeout0 control
returns0 and the following single poll returns258, proving consumption.
reproducer-result.json records this reproduced defect; terminal exit0 means
the diagnostic assertions passed, not that the API is repaired. Test fixture
writes are confined to its separate module memory; original game memory was
observed without writes.
The probe saw4,393,377 total calls with bounded storage. Its256 counter keys filled before RSX;4,391,874 overflow events are not an exact RSX failure count. The selected tail also retains repeated failures, not unique transitions. Inherited Node Worker preload contexts can overwrite timer snapshot files; the explicit main-context retrieval above is the authoritative original sample. Main's exit observer saved actual compile/read records separately from renamed offline records. The original exited0 after93.770 seconds by ordinary quit followed by delegated orderly SIGTERM; no heap OOM occurred in this bounded launch. This does not isolate the predecessor's heap-allocation cause. The first substantive API failure stopped further stepping and inputs.
All1,122 manifest paths matched, including reconstructed config.ini; Fallen.exe
retained SHA256 8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229.
Complete accepted build reuse matched4,392 source pins and141 recorded reads.
Actual original WebAssembly.compile observed module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Before/after4,444 source identities,1,124 fixture/manifest identities and132
actual runtime read identities match. Read coverage includes sync whole-file
and fully covered bounded descriptors, not compiler reads, all async OS reads
or guest-memory contents. The reproducer's separately compiled test module
is not the original runtime module.
Final batch25 input counters are zero and known queues empty. Host physical, async and pressed-key/button states remain unknown; consumed gameplay controls, sampled releases, before/held/released/settled and matched gameplay-neutral evidence are unavailable because no action was attempted. Guest post queues and SDL consumption remain unknown. No audio, FPS, browser/GPU or full27 claim. Disk guards sampled above2,000,000,000 bytes free with finite sampling gaps; no RAM gate or foreign cleanup. Native box ID/state/expiry are N/A. Own shutdown was delegated; cleanup/publication/final records document absent owned processes, groups, watchers and new child temps and ownership handback.
2026-11-01 ready-any repair: Setup, title and mission opening recovered
Evidence: scratch/runs/20261101-urban-chaos-ready-wait-repair/.
The generic mixed-event ready-any defect is repaired in waitMultiple.
The complete handle set is validated before acquisition, including closed and
tagged authority handles, local generations, thread aliases and process handles.
Bounded byte reads support unaligned x86 handle arrays. Ordered zero-time
acquisitions select the lowest ready index and consume only that object.
Valid named events no longer reject a nonzero-timeout call before checking
readiness. Named wait-all and nonzero-timeout multiple waits with no ready
object still return WAIT_FAILED: cross-process blocking multiple waits are
not implemented by this repair. Single-wait, scheduling, callback, concurrency,
SEH and CRT nonlocal-jump algorithms are unchanged.
The exact Battlezone black-dialog author DONE/PUSH RELEASE at 16:12:35 UTC,
fresh launcher result0/final, and direct absence of eight owned PIDs/seven
groups passed the gate at 16:12:43 UTC. Dispatch was 4d9446b48; actual gate
main was 6f4fa1d5ecb75f03aea8e0b354dbe7c051ca3c5f; original deadline
16:37:43 UTC. The first locked build failed the conservative whole-JS-file
callback census SHA check. first-build-failure.log is retained. Following
the documented inventory maintenance procedure, the complete producer delta
was reviewed, and only the matching thread-manager seal SHA/reason changed.
Ownership remains pending; no gate or held algorithm was bypassed. Locked
builds after review and after the final unaligned-read adjustment both passed.
Focused thread-manager, named-event authority/API/end-to-end, callback census
and census rejection-control tests passed. Maintained end-to-end regressions
invoke production WAT CreateEventA, SetEvent, WaitForMultipleObjects,
WaitForSingleObject and CloseHandle with a real authority in another process.
They exercise both ready indices, both-ready ordering, finite/INFINITE/zero
timeouts, auto-reset consumption, manual-reset persistence, partial/final close,
private creation tags, invalid later handles and unsupported no-ready/wait-all
cases. Unit cases additionally cover recycled local handles, pending/exited/
closed thread aliases, process validity, semaphore consumption and array bounds.
These are API/state-consumer tests, not native-Windows conformance certification.
The retained real-API reproducer now returns [0,258,258,0,258]: both INFINITE
and timeout0 return index0 and consume the signal. Its pass now means repaired
semantics, unlike the predecessor's diagnostic pass.
One original used the registered frozen native software CLI, batch-size100000,
bounded probes and sequential step requests of at most25 batches/180 seconds.
It ran 440.539 seconds and exited0 after ordinary quit and delegated orderly
SIGTERM. At batch25, native inventory and reviewed native backing show the
ordinary graphics/audio Setup dialog. main-waits-setup.json retains the
actual RSX wait on [0x0e000041,0x0e000040], INFINITE: result0, first event
state1 before/state0 after. Its worker subsequently exits. This directly
connects ready-event consumption to recovery of the previously blocked Setup.
After backing inspection, normal mouse-down/up on visible Okay at275,445 dismissed Setup. Title is reviewed at205; held Enter reaches mission map307, briefing409 and the loading image511. Rendered opening scenes at711,911,1111 and final1436 show rain, buildings, characters and objective captions. Playable HUD, player movement, combat and interaction success were not reached or claimed. No second original, alternate settings or game-specific workaround.
Every navigation action has before/held/released/settled samples and a matched neutral interval. Three held Enter pairs have six actual matching dequeue, route and DispatchMessageA records. The briefing release scheduled at485 is still pending in samples486/511 during loading; actual dequeue/route/dispatch occurs at658, and sample711 confirms its consumption. The first evidence-check assertion incorrectly required immediate consumption; its failure and corrected trace-based check are retained. No input was replayed. The native Setup mouse path records host actions, a visibly pressed button and dialog dismissal, but does not populate those generic guest dispatch counters; no equivalent mouse dispatch proof is claimed. Final1436 known queues are empty; sampled physical/async keys, including VK1/VK13, are false and numeric mouse mask is0. Their pressed latches remain true; guest post queues and SDL consumption are unknown.
Further progression stopped when the auxiliary disk-guard process was found terminal143, cause undetermined. Its last record is 16:22:23 UTC; discovery was 16:25:18 UTC. The independent primary driver remained live, sampling disk and enforcing disk/time/log guards. Its minimum was2,257,408,000 bytes; the auxiliary guard's earlier minimum was2,037,649,408. This is a helper-lifecycle blocker, not a game crash or a claim of continuous filesystem coverage between samples. The active advance controller was stopped, then final released-state and scene samples were obtained before shutdown. No original retry followed.
Actual observer prerequisites passed sync/async execution, whole512MB memory
read-only comparison, clock identity and late10,001 retention. The bounded
wait probe passed real-call result/whole-memory equivalence. Scratch observers
now write only from the controlling main context to prevent Worker snapshot
overwrites. Actual original compilation observed canonical module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
All4,444 source identities,1,124 fixture/manifest identities and289 observed
runtime reads match before/after. All1,122 manifest paths match, preserving
original Fallen.exe SHA256 and the reconstructed-config distinction above.
Read coverage excludes all-async/direct-OS/compiler-read and guest-memory
identity claims. All43 PNGs pass CRC/complete-deflate checks; full runtime gzip
passes. Primary milestone captures were visually reviewed. No FPS, audio,
browser/GPU or full27 claim. Native box ID/state/expiry are N/A; own shutdown
was delegated. Cleanup/publication records carry final ownership handback.
Next: investigate the auxiliary guard's unexplained termination before a fresh bounded original continuation toward playable HUD and held movement. The API correction does not authorize broader retained-context or blocking-wait work.
2026-11-01 guard lifecycle investigation and current-main HUD/movement
Evidence: scratch/runs/20261101-urban-chaos-guard-gameplay/.
The prior auxiliary guard's signal sender and reason remain UNKNOWN.
Retained raw worker events establish that its tool-owned launcher shell ended
143 at16:22:24.467 UTC,779.120 seconds after startup. The primary driver's
launcher shell ended143 at16:22:24.459, just8ms earlier. Its Python child
survived with PPID1 and the original remained live. The guard's last heartbeat
was16:22:23.486, free2,261,217,280 bytes; neither its disk-stop path nor a
normal guard-stop explains shell status143. This supports a correlated external
launcher/session lifecycle hypothesis, but no retained siginfo/audit record
identifies the sender or reason. It does not establish a game crash.
guard-diagnosis.json preserves the exact terminal events and limits.
Scratch-only prevention replaces the foreground tool-owned auxiliary guard with
a primary supervisor detached into its own session by a short, successful
launcher. An independent quarter-second watchdog writes atomic heartbeats and
enforces disk headroom and the absolute deadline; signal/exception records and
owned child shutdown are explicit. The primary driver independently checks
heartbeat freshness, process presence, disk, deadline, log bounds and original
exit. Requests require a live guard and have finite180-second timeouts.
guard-live-proof.json records advancing heartbeat counters/process identity
before original launch. This supervisor remained live through the entire new
attempt and ended normally:6,196 samples over its complete lifecycle, including
the gate wait, minimum2,145,710,080 bytes free.
These are sampled bounds with finite gaps, not continuous filesystem observation.
The exact battlezone-dialog-capture-harvest author DONE/PUSH RELEASE at16:53:27,
fresh launcher result0/final, and direct owned-process/group absence passed the
gate at16:53:41.915. Dispatch input wasa17c5a7c85aae38a97653b801da68de6147048df;
actual main was6ca46f8ffced6b6e4331e8fa4efa33e4d91c6872; deadline17:18:41.915.
All4,444 scoped source pins matched the accepted ready-any repair's final state,
so its exact locked build and seven focused checks were reused, with local
copies/hashes retained. No production source, build, test or fixture edits.
Fresh manifest verification passed all1,122 paths, original Fallen.exe SHA256
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229, and the
reconstructed-config distinction. Fresh observer validation passed actual
sync/async execution, whole512MB read-only comparison, unchanged clocks within
a180-second bound (560ms observed), and late10,001 retention. SWEEP_CASE_DIR
pointed to this run; only the controlling main context wrote observations.
One original used the registered native software CLI, frozen batch-size100000, step requests at most25, and no build. Immediate startup free space was 2,203,303,936 bytes. It exited0 after826.358 seconds, with no runtime/guard failure or second original. Native Setup backing was inspected at25; ordinary Okay, title205, map307, briefing409 and loading511 were reviewed. Opening711, 1111 and1711 show changing cinematic scenes. HUD appears naturally at2011: D'arci on the street, radar/health, fist icon and3:58 timer. No extra cinematic skip input was needed.
Current-main ordinary movement is demonstrated. Matched neutral samples at2011/2111/2112/2137 leave D'arci in the same starting street position. ArrowUp samples at2137/2237/2238/2263 show movement from the street into the passage toward the car park, then release and settlement at the new location. Each interval uses100 held/neutral batches,1 release-phase batch and25 settling batches, with individual requests capped25. All eight neutral/movement phase captures were inspected. VK38 is physically/asynchronously false before, true held, false released/settled. Actual dequeue/route/DispatchMessageA edges match: keydown2140, keyup2237. Three Enter pairs also match actual consumers; the briefing keyup is deferred until658 rather than claimed consumed at486/511. Final2263 known scheduled/pending/cross-thread/renderer queues are empty, sampled physical/async keys false and mouse mask0. Pressed latches remain set; guest post queues and SDL sampling remain unknown. The104 step replies are all present, maximum25; all recorded request latencies are within180 seconds (max74.611s). Native Setup mouse evidence is ordinary host input/dialog dismissal, not equivalent generic guest-dispatch proof.
Actual runtime compilation observed canonical module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
All4,444 source and1,124 fixture/manifest pins match after runtime. All304
observed read hashes match after-run files:256 also match fresh before pins;
48 font/PNG dependency reads lack fresh current-run before hashes and instead
match retained accepted prior observations. That narrower coverage is explicit
inidentity-comparison.json andaccepted-prior-loaded.json; no fresh
before/after claim is made for those48. Observation excludes unhooked async,
direct-OS and compiler reads and guest-memory identity. All52 PNGs pass CRC and
complete-deflate validation; the full runtime gzip passes. Selected milestones
and every neutral/movement phase were visually reviewed, not every remaining
PNG. No FPS, audio, full27, browser/GPU or combat qualification.
Own original/driver/supervisor completed, watchers were explicitly stopped, and cleanup records verify listed own PIDs/groups/helpers and new child temps absent. Native box ID/state/expiry are N/A; own shutdown was explicitly delegated and ownership handed back. Remaining diagnostic gap: historical SIGTERM sender and reason cannot be recovered from these records. The scoped HUD/movement goal has no remaining gameplay gap; no broader source repair is authorized here.
2026-11-01 native encounter attempt: guard-check failure before gameplay
Evidence: scratch/runs/20261101-urban-chaos-native-encounter/.
No encounter or objective interaction was reached. The first guard failure
stopped the single original at batch26 before any gameplay input; no retry.
Historical HUD/movement evidence above remains historical.
Exact battlezone-renderer-dialog-freshness DONE/PUSH RELEASE at17:23:38,
fresh launcher final/result0 and independent absence of its21 listed PIDs/
20 groups passed the gate at17:23:46.034 UTC. Dispatch input was
f20e19b24651f2e37fe53b422378205328338270; actual main was
ebd313fd0b3415fe1dc0c6e7154317f855dbbfea, including renderer fix
feb4934388f06da990a750d0073fd38776fa0c41. Gate deadline was17:48:46.034,
inside the18:15 overall bound. The renderer worker's passing locked build
and relevant checks were reused after exact comparison of its4,063 pins and
the prior Urban run's pins (4,445 unique accepted inputs, zero differences).
Fresh renderer freshness, thread-manager and real-WAT named-event end-to-end
checks also passed. No production source, test, build or fixture edit.
All1,122 manifest paths matched; Fallen.exe SHA256 remains
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229.
Config.ini remains explicitly reconstructed installation state. Original
readme.rtf was recovered read-only from the retained recovery ZIP into this
run, without a download or rewriting archived evidence. Its controls tables
specify arrows for movement/turning, C for action, Z punch, X kick and Space
jump. It describes talking to a person with action and arresting a prone
suspect. The vehicle hint conflicts with the table (Enter versus C); no
vehicle input was selected. Bundled config and English instructions were
also read. No combat/action binding was guessed or sent.
The prior guard controller and diagnosis were read, and its proven detached
supervisor/independent heartbeat-thread mechanism was reused. Advancing
before-launch heartbeat samples and the final during-original heartbeat are
retained. Original startup free space was2,182,414,336 bytes, above the
2,150,000,000-byte immediate-launch minimum. Registered native software CLI
used --no-build --control-stdin --frozen, batch-size100000, step requests
1/24/1 and180-second finite request bounds. All recorded requests were
acknowledged, maximum latency0.627 seconds. Reviewed Setup at25 is readable
in the standard640x480 capture and the332x470 native dialog backing; final
capture shows the same ordinary graphics/audio Setup. Okay was never pressed.
The driver then raised RuntimeError('primary guard lost') in its combined
heartbeat-age>1second / missing-supervisor / guard-failure-file check. The
failing operand and age were not retained. The supervisor has no signal or
failure file and ended normally with reason null after823 samples, minimum
free2,180,182,016 bytes. Its last heartbeat was17:25:14.888; driver terminal
was17:25:15.139 and supervisor terminal17:25:15.156. This does not establish
supervisor death, a game crash or the cause of the rejected check. It differs
from the prior foreground launchers' unexplained143. The driver requested
ordinary failure-quit (acknowledged); original exit0 after20.163 seconds,
then driver0 and normal supervisor completion. The Setup-neutral helper
subsequently rejected the now-stopped guard before sending its first request.
No unchanged retry, second original, guard bypass or source repair followed.
Fresh observer validation passed actual sync/async execution, whole512MB
read-only comparison, unchanged host/native clocks bounded456..457ms, and
late-event10,001 retention. Actual original compile observed canonical module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
All4,678 source/dependency pins and1,124 fixture/manifest pins matched after;
all133 actual observed read hashes matched fresh before and after pins.
Coverage excludes unhooked async/direct-OS/compiler reads and guest-memory
identity. All four PNGs passed CRC/complete-deflate; full runtime gzip passed.
Recover the log with gzip -dc runtime.log.gz inside the run directory.
The batch25 state has zero gameplay counters and empty known queues. Sampled physical/async/pressed keys and mouse mask were unknown; final released-state sampling was not obtained because shutdown had already completed. No held keys were injected. Before/held/released/settled gameplay and matched neutral captures are unavailable. No HUD, movement, combat, objective, FPS, audio, browser/GPU or full27 claim. Disk evidence has finite sampling gaps; continuous coverage is not claimed. Local native box ID/state/expiry are N/A; delegated own shutdown completed. Cleanup records verify known own PIDs/groups/watchers/helpers and new child temps absent, with ownership handed back.
Next: retain the rejected guard predicate's exact operand, monotonic age and heartbeat history before a separately assigned original continuation. The ordinary encounter/objective goal remains open; recovered original controls and the prior HUD-to-carpark route are ready for that continuation.
2026-11-01 guard diagnostics: controlled checks pass; helper terminal143 stops encounter attempt
Evidence: scratch/runs/20261101-urban-chaos-guard-diagnostic-encounter/.
The encounter/objective goal remains unachieved. One original reached the
loading screen at batch386; no gameplay movement or encounter input followed.
The first new lifecycle blocker ended this attempt without a second original
or replaying the interrupted request. Historical HUD/ArrowUp evidence above
is not new encounter evidence.
The predecessor Urban driver/supervisor and retained diagnosis were read.
Its combined age/process/failure-file predicate still has an unknown historical
failed operand. Separately, retained Battlezone evidence measured1.119649s
heartbeat age before its primary advanced454→455; that does not identify the
cause of either historical delay. Cheap no-game controls now separately retain
healthy advancement, a healthy measured1.500946s pause, stale3.152561s age,
dead process, replaced process-start identity, and explicit failure-file cases.
Each control includes wall/monotonic timestamps and ages, process start ticks,
failed operands and heartbeat history. Atomic scratch writes plus gzip
encode/read/decode produced52 healthy I/O samples, maximum observed check
interval0.930804s. no-game-delay-analysis.json corrects an overly broad
observedIoMaxEmitGap label in the retained raw controls: that statistic includes
the deliberately injected pauses and must not be called an I/O-only delay.
Scratch freshness policy is now a bounded3-second monotonic tolerance, informed by those measured healthy delays; stale3.15s is still rejected. This is an engineering margin, not proof of the historical cause or a measured need for exactly3seconds. Detached primary supervision and a separate independent process retain direct disk/deadline enforcement, reciprocal heartbeat/process checks and process start identity checks. The driver independently checks original liveness, both heartbeats, disk, deadlines, request/log/capture bounds. The independent process detects primary absence and owned-process identity replacement; missing-original detection specifically remains the driver's poll, not a separately claimed independent missing-original predicate. Guard rejections retain operand details and bounded heartbeat histories. An exclusive-create initiating guard-failure file cannot be overwritten by later signal/cleanup records. No safeguards, production code or tests changed.
Exact Battlezone battlezone-fixed-capture-harvest DONE/PUSH RELEASE at17:31:26,
fresh launcher result0/final and direct absence of its listed owned PIDs/groups
passed the gate at17:41:41.263 UTC. Dispatch input was7d47a40e5;
actual main was98c2a1d5e2aa1fef348a0697aed7de58c59fc1d7.
The25-minute gate deadline was18:06:41.263, within the19:15 overall bound.
All4,445 accepted build inputs matched, permitting exact renderer locked-build
and focused-check reuse. Fresh manifest, observer whole512MB read-only
sync/async/clock/late10,001, renderer, thread-manager and real-WAT named-event
checks passed. All1,122 manifest paths matched, including original Fallen.exe
SHA256 8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229;
config.ini remains reconstructed installation state. Original controls were
read: C action, Z punch, X kick, arrows movement/turn. No combat binding was
sent or guessed.
The registered native software original used --no-build --control-stdin --frozen, batch-size100000, requests at most25 and finite180-second bounds.
Immediate startup free space was2,169,905,152 bytes, above2,150,000,000.
Reviewed standard/native Setup, ordinary Okay, held Enter navigation, mission
map182, briefing284/335 and loading360/386 were retained. Menu actions have
before/held/released/settled and matched neutral captures. Five actual key
edges match dequeue/route/DispatchMessageA; the third Enter key-up remains
queued at386 and is not claimed consumed. Final sampled physical/async keys
are false and mouse mask0; scheduled/pending/cross-thread queues are empty,
renderer queue contains that one key-up, pressed latches remain set, and guest
post/SDL consumption is unknown. Native Setup mouse evidence is host input
and visible dismissal, not generic guest-dispatch proof.
The briefing helper's tool session79606 returned143 after printing
NAVIGATED briefing. Helper3968062 was subsequently absent. Its final PNG
request was acknowledged in the original runtime log, but the helper did not
record receipt. That request was not repeated; its helper receive time/latency
is unknown. All other recorded request latencies are below180seconds,
maximum2.246387s;27 step requests have maximum25. Sender/reason for the helper
terminal status remain UNKNOWN. At discovery, original3966982 and both
detached guards3966061/3966066 remained live; no guard rejection/failure file
was present. This is not evidence of a game crash, a failed heartbeat check,
or signal143 delivered to the original. initiating-failure.json preserves
the helper event separately from orderly shutdown. No helper/request/original
retry followed; only final read-only sampling and ordinary quit were performed.
The original exited0 after170.797s, driver0, primary/independent normal with null reasons. Primary861 samples had minimum2,163,261,440 bytes free and maximum observed gap0.529202s; independent860 samples had minimum2,163,302,400 and maximum gap0.399494s. Direct disk stop remained2,020,000,000, before the required 2,000,000,000 floor. These are finite sampled bounds, not continuous disk observation. Guard freshness worked during this short run; it is not qualified for a complete encounter flight. No unchanged1-second monitor was reused.
Actual original compilation observed canonical module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
Before/after4,678 source/dependency and1,124 fixture identities match; all269
observed runtime reads match fresh before/after pins. Coverage excludes
unhooked async/direct-OS/compiler reads and guest-memory identities. All40
PNGs pass CRC/complete-deflate; full runtime gzip passes. Selected useful
images were visually inspected, not every PNG. No FPS/audio/full27,
browser/GPU, deployment, guest patch or replacement claim.
Own shutdown was explicitly delegated. Direct cleanup verified19 known PIDs
and17 groups absent, no matching helpers/watchers and no new child temps.
Native box ID/state/expiry are N/A; ownership is handed back. Recover the log
with gzip -dc scratch/runs/20261101-urban-chaos-guard-diagnostic-encounter/runtime.log.gz.
The directory retains executable scratch guard/control/observer scripts,
accepted build hashes/logs, original-command and gate records, immutable
initiating failure, final state and cleanup records; do not treat recovery
recipes as permission to launch another original in this assignment.
Next: separately diagnose the controller/tool terminal143 lifecycle before
another assigned continuation toward an ordinary encounter. This assignment
stops with that concrete blocker and the gameplay objective gap disclosed.
2026-11-01 detached commands: receipts survive; heap failure and stale heartbeat stop continuation
Evidence: scratch/runs/20261101-urban-chaos-detached-controller-encounter/.
The ordinary encounter/objective goal remains unachieved. One original
completed step1936; the next25-batch request has no retained ACK and was not
replayed. Useful actual PNGs show standard/native Setup, mission map, briefing,
loading and changing cinematic scenes through1586. No new HUD capture,
gameplay movement, combat or contextual interaction is verified. The accepted
HUD2011/neutral2137/ArrowUp2263 evidence above remains historical.
Scratch detached.py launches each named command in an owned detached session,
records runner/child PID and start identity, uses explicit interpreter/argument
vectors, serializes commands with a lock, and bounds each command. Terminal
receipts use fsynced temporary files, atomic publication and directory fsync;
exclusive atomic links preserve complete immutable first-failure records.
Unique job directories reject command replay. Individual control requests also
retain durable sent/ACK records and finite180-second waits, with steps at most25.
The no-game harness killed its initiating clients with SIGTERM: the detached
success command still completed, exit7 retained a terminal failure, and a finite
timeout retained its first cause through a secondary cleanup write. All three
passed; their owned processes subsequently disappeared. A detached Node dispatch
control passed too. After Battlezone release, its Python-only runpy helper was
consulted; this wrapper uses the correct explicit Node/Python interpreter.
The predecessor's immutable failure is copied without alteration. Its historical
status143 sender/reason remains UNKNOWN; identification was not a prerequisite.
Exact battlezone-guarded-capture-harvest author DONE/PUSH RELEASE, fresh external
launcher result0/final and direct owned PID/group absence passed the gate at
17:54:20.945 UTC. Dispatch input wasc32bdc566; actual main was
19f8b03404cbb6a79bdedf02c33f9e39255610dc, deadline18:19:20.945 within20:00.
All4,678 accepted source/dependency pins and compiled-module hashes matched,
allowing locked-build/focused-test reuse. Fresh fixture verification passed all
1,122 manifest paths and original Fallen.exe SHA256
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229;
config.ini remains reconstructed. Fresh observer sync/async execution,
whole512MB read-only comparison, clock identity/bounds and late10,001 retention
passed. No source, build, production-test or fixture edits were made.
The single registered software original used --no-build --control-stdin --frozen, batch100000 and finite bounds. Immediate launch free space was
2,155,229,184 bytes, above2,150,000,000. The detached menu and advance commands
passed the predecessor's interrupted386 point with complete receipts. Eleven
command terminal receipts exist: ten succeeded and the last retained exit1
with its traceback. This demonstrates the new command receipt path on an actual
failure, not complete runtime-lifecycle qualification.
At18:01:26.521 the driver rejected primary heartbeat1449 at monotonic
age3.012139s against the3-second tolerance. The primary PID/start identity still
matched and its sampled state was S. The immutable guard-failure.json retains
the exact operands and128 preceding observations. Later primary/independent
rejections and the detached command's exit1 record are separate. The3-second
monotonic/identity predicates were retained; scratch persistence additionally
changed heartbeat/failure writes to fsynced atomic publication. That persistence
change was not independently qualified under game load. Stored heartbeat
timestamps do not measure publication completion; the cause of the stale
published observation remains unknown. No guard relaxation or retry followed.
The recovered runtime stream also contains133 guest heap allocation failures,
beginning after the1936 ACK: sparse commit and bump/low/sparse exhaustion at
EIP0x00494d9f, caller0x0048c76e. This is a concrete changed runtime blocker,
not a host-RAM admission measurement. Exact OOM onset versus guard timing is
unknown because those log lines lack wall timestamps. No guest crash/trap cause
is established. The Node exit hook reports1; the original OS wait receipt and
driver terminal receipt are missing. Supervisors entered owned SIGTERM cleanup;
no ordinary original exit0 is claimed. The driver gzip footer is incomplete.
Its original bytes are preserved; runtime-recovered.log.gz contains all207,005
decodable bytes, with no claim that an unflushed tail was recovered.
Of218 manual requests,217 have unique durable successful ACK receipts; the one unresolved step was not repeated. There are89 bounded step requests, maximum25; maximum recorded successful reply latency13.365361s. All six Enter edges match actual dequeue/route/DispatchMessageA records. No gameplay key was sent. Last sample1586 has physical/async keys false, mouse mask0 and empty known queues; there is no fresh post-failure released-state sample. Pressed latches and guest post/SDL consumption remain distinct/unknown. Before/held/released/settled and matched neutral menu captures exist; encounter phase captures do not.
Before/after4,678 source and1,124 fixture identities match, as do300 observed
runtime reads. Actual compilation observed canonical module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
All42 actual PNGs pass CRC, complete deflate and raster-length checks; selected
milestones were visually reviewed. Primary1,450 history samples had minimum
2,108,903,424 bytes free; independent1,449 had minimum2,110,160,896. These are
finite sampled bounds, not continuous disk/freshness coverage. Direct stop
remained2,020,000,000 before the2,000,000,000 floor. No FPS/audio/full27 claim.
Direct cleanup verifies40 recorded owned PIDs and38 groups absent, no matching
helpers/watchers and no new child temps. Native box ID/state/expiry are N/A;
explicitly delegated own shutdown is complete and ownership handed back.
RECOVERY.txt preserves executable scratch recipes and the log limitations.
Next requires a separately assigned investigation of the guest heap failure
and heartbeat publication/cleanup-receipt gap before another encounter attempt.
No second original, unchanged replay, source repair or broader qualification
was performed in this assignment.
2026-11-01 offline heap diagnosis: vertex-pack refusal; capacity versus fragmentation still unmeasured
Evidence: scratch/runs/20261101-urban-chaos-heap-cause/ (analysis.json,
failure-events.json, three disassemblies and RECOVERY.txt). Static analysis
only: no original, runtime, build or tests ran, and no production files changed.
Dispatch input was 7699f3fb00184b5d46c2cd98ad03569c5017318d; actual main
repinned to 424b9e1b1813da28543fe19c5770016227f34b3c.
Original Fallen.exe SHA256 is
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229.
Ten relevant allocator/dispatch/layout/host source hashes match both retained
runs and this analysis tree; both runs loaded canonical WASM
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
The failing operation is runtime vertex packing, not an oversized game malloc.
At 0x00494d60, Fallen builds a triangle fan's WORD indices at 0x00ee4438.
The loop exits at 0x00494d9f (pop ebp); 0x00494dbf calls vtable slot35
(+0x8c) with primitive4/TRIANGLELIST, VB=[ecx+0x80], generated indices/count
and flags0x1c: IDirect3DDevice3::DrawIndexedPrimitiveVB. The EBP-chain return
0x0048c76e follows call 0x00487020 at 0x0048c769; it is an outer frame,
not the allocating function. The interpreter's $th_call_ind dispatches a thunk
before updating EIP to the return address, so the logged EIP can be this block's
entry rather than the actual COM call instruction. There is no malloc at that EIP.
The handler calls $d3dim_vb_draw_indexed_primitive, clamps the VB capacity and
packs only through the highest referenced index. $d3dim_pack_fvf_vertices
requests count * 32 bytes. The66 complete small failures report aligned
allocator extents25,960..44,072 bytes, all 32*N+8, consistent with811..1,377
vertices; the first32,296-byte extent corresponds to a32,288-byte/1,009-vertex
pack on this path. These are inferred pack operands from code and logged
extents, not captured VB/index measurements. Successful packs are freed after
the draw. Failed packs skip the draw and still return S_OK: visible HUD and
movement do not prove all geometry drew or all allocations succeeded.
The67 separate1,048,576-byte messages are $heap_sparse_alloc's minimum new
arena commit, not67 independent1MiB game allocations. After a free-list miss
and low-arena refusal, it reserves guest address space, then
$virtual_map_commit returns0; reason3 is emitted after its reclaim/retry path.
Reason1 subsequently reports the original aligned small extent. The recovered
stream ends on an extra reason3 with no following small failure/site retained.
It establishes new arena commit refusal, not which commit check failed.
The nominal sparse backing window is288MiB inside the fixed512MiB memory;
there is no extension at that size. Actual live committed bytes, map count
(limit8,192), largest gaps, arena live-byte counts and free lists were not saved.
Reason3 can also cover metadata/lease/page-publication refusal. It does not
uniquely mean the288MiB pool was fully occupied.
Fragmentation is a specific candidate: ordinary heap free prepends blocks, allocation splits them, and this path does not coalesce adjacent free blocks. Reclaim deliberately releases only whole-block sparse arenas, so carved arenas can retain backing even when their live-byte count is0. The render coalescer exists but is called by the D3D9 software backend, not this legacy VB pack path. That mechanism is present in the pinned code; its occurrence/byte cost in this game is unmeasured. The pack's balanced success/free path does not establish a leak, and no allocation history identifies the owner of the retained backing. Expected live capacity, fragmentation, another owner's leak and a commit bookkeeping failure therefore remain distinguishable hypotheses. No bad-size reason5 or absent VB implementation explains the observed refusal.
Correction to interpreting the successful HUD route: its retained
20261101-urban-chaos-guard-gameplay/runtime.log.gz contains20,432 OOM lines
(10,216 reason3/reason1 pairs), first after ACK2061, including19,468 site lines
at the same EIP/caller. It subsequently reached2263 and ordinary exit0 with the
documented HUD/movement evidence. Its lack of a terminal runtime failure must
not be read as absence of heap failures. The detached run first logs OOM after
ACK1936; differing batch onset is not a controlled memory comparison. Among
shared before pins only renderer.js and the callback-census JSON differ; the
allocator/WASM identity is unchanged. Neither log supplies OOM wall timestamps.
No OOM-to-guard causality, guest crash or host RAM failure is established.
The smallest next discriminator is a single first-refusal allocator snapshot,
before another encounter flight: retain wall/monotonic time, batch, instance,
size/reason and WASM stack at the existing heap_oom_trace import; copy the
bounded map/reservation/heap-arena tables using the matched generated region
map, plus pure heap cursor/stat getters and the passive
d3d_render_free_list_diagnostic fields. Read free headers through a host-side
translator; calculate live backing union/free gaps, per-arena live bytes,
free-byte total/largest fit and validated adjacent free extents offline. Capture
the active Win16 lease ledger only if its arena is nonzero. Do not call
get_free_list or heap_free_census as passive probes: both flush bins.
Reason3 alone needs a branch-specific commit refusal code if these data leave
more than one failing predicate possible. Capture before expensive formatting,
publish once, bound output, and retain the original guard predicates.
First use a bounded no-game pack/free sequence varying811..1,377 vertices on
the pinned allocator: a refusal with low live bytes and enough adjacent free
space identifies fragmentation without spending another original attempt.
If live backing instead approaches capacity, collect allocation/free ownership
before choosing a memory-size experiment; if maps are full or publication/lease
checks fail with space available, target that exact runtime behavior. No memory
increase, coalescer mutation or guard change is justified by the present logs
alone. This assignment stops at that precise missing measurement. Recovery is
offline via python3 scratch/runs/20261101-urban-chaos-heap-cause/analyze.py;
the predecessor's truncated original gzip and completed recovered gzip remain
untouched. Native box ID/state/expiry N/A; no boxes were created or altered.
2026-11-01 no-game allocator probe: balanced pack sizes reproduce fragmentation-driven retention
Evidence: scratch/runs/20261101-urban-chaos-allocator-probe/. Dispatch main
was a34f3648016a508b5fd268297ff84e7c00a5fe8e; actual measurement main was
39242f4b76421b468157ee51d7593037fcb527de. After Battlezone's explicit release,
fresh launcher result/final and direct absence checks for its55 PIDs/groups,
136 source/module/interface hashes matched the accepted build. The unchanged
canonical WASM SHA256 was
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4.
One Node24.18.1 instance used fixed512MiB memory, public init_thread,
heap_init, guest_alloc and guest_free; no PE, vertex payload, rendering,
game replay, source edits or build. Host imports were stubs except the OOM
capture. Heap initialization used the matched generated GUEST_HEAP_BASE.
Balanced varying pack sizes alone reproduce a refusal. The exact sequence
was2,000 alloc(32*1009)/free pairs, then ascending
alloc(32*(811+(i%567)))/free pairs, with at most one live request and each
success freed immediately. The constant control retained one low arena and no
sparse backing. Ascending iteration15,051 (zero-based) refused1,120 vertices:
35,840 requested bytes, aligned need35,848. It emitted reason3/1MiB followed
by reason1/35,848 bytes. There were17,051 completed balanced pairs in total;
measurement elapsed5.15s, supervisor exited0/reaped its child in5.27s. Limits
were100,000 ascending pairs,540s internally/570s externally, and a direct disk
stop at2,020,000,000 bytes; minimum observed free disk was2,094,579,712 bytes.
| First-refusal measurement | Bytes/count |
|---|---|
| Sparse backing union / capacity | 301,989,888 / 301,989,888 |
| Free backing / largest backing gap | 0 / 0 |
| User live requested bytes / summed arena live bytes | 0 / 0 |
| Free carved headers, including off-list blocks | 15,320 / 305,107,584 bytes |
| Largest free header / required extent | 35,816 / 35,848 |
| Largest adjacent free run within one arena | 1,048,576 |
| Maps / limit; arenas / limit | 288 / 8,192; 291 / 1,024 |
All288 sparse arenas retained their backing. Of these,287 were retired, zero-live and carved:300,941,312 bytes could not satisfy the whole-block reclaim predicate because the first header did not span its arena. The active sparse arena also had zero live bytes. Three low arenas brought retained arena space to305,135,616 bytes;28,032 bytes remained unused tails. The validated ordinary list held15,216 blocks/305,092,608 bytes;14,976 header bytes were off-list. Their free status follows from this balanced sequence and zero arena live counts, and is not an inference to apply to an arbitrary game snapshot.
The first OOM import copied bounded map/reservation/hole/arena/shared/Win16
tables, cursors, counters and headers once, before formatting. Passive
d3d_render_free_list_diagnostic reported valid; no bin-flushing getter,
retirement, coalescer or allocator mutation was used for measurement. Offline
walks validated all headers, adjacency and table hashes. First-import and final
tables, headers, list, cursors and counters were identical. The Win16 arena was0;
map/reservation/arena limits were not full. Fully occupied backing leaves no
fit for even the recursive64KiB commit floor. The matched source therefore
explains the arena commit refusal through backing no-fit/split exhaustion;
the exact internal branch was inferred, not instrumented. The measured
underlying mechanism is fragmented, empty carved arenas retaining backing.
This establishes a sufficient allocator-only mechanism, not the game's actual first-refusal cause or OOM-to-guard causality. The successful HUD route also emitted heap OOMs. The remaining game-only discriminator is the same passive first-refusal snapshot on the allocating instance, with size/reason, wall/monotonic time, batch and allocation ownership/bin information if live counts are nonzero. A branch-specific refusal code is needed only if that snapshot leaves multiple predicates possible. No memory increase, coalescer change or guard rewrite is justified by this probe alone.
Recovery is offline:
python3 scratch/runs/20261101-urban-chaos-allocator-probe/analyze.py.
RECOVERY.txt records the exact sequence, bounds and module recovery recipe;
pins.json, snapshots, measurement.json, runtime-result.json and
verification.json preserve the evidence. No further runtime is needed to
reanalyze it. Own local helpers were stopped; native box ID/state/expiry N/A.
2026-11-01 generic allocator repair: reuse a provably local empty sparse arena
Evidence: scratch/runs/20261101-urban-chaos-empty-arena-repair/. This is an
allocator regression repair; Urban Chaos's actual first-refusal cause and
gameplay failure cause remain unmeasured. No game was replayed. Runtime work
started only after exact Battlezone worker 01a19867-ff52-75a3-8cf4-57531732e92d
released runtime/source/build, its launcher result/final arrived, and all22
recorded PIDs/groups were directly absent. Measurement input main was
02b44b8944e97cd518c7349119037eb43bedccb5; all136 probe source/module/interface
pins matched before editing.
Simply broadening $heap_arena_release_free is unsafe. Arena records/maps are
shared, but bins and free lists are instance-private. Another instance may own
freed carved headers even when the shared live-byte count is0. Releasing and
reserving that guest range again can make its surviving links validate against
new live allocations. The existing whole-block release guard prevents this
aliasing and remains in force.
The repair instead rewinds only this instance's current sparse bump arena on an exhausted bump request. It requires the exact current base/high/end record, zero live bytes, a fitting arena, and a carved first header. It flushes this instance's bins, validates the complete local free list, then walks actual carved header boundaries and proves every header is on that list. Matching header/list-entry counts reject extra interior entries. The proof is bounded to65,536 list entries and1,000,000 combined search steps; failure retains the existing allocation path. Success purges local links before clearing the old prefix through page-aware guest memset and rewinding cursor/high. The arena's mapping, reservation and record base stay intact. New bump allocations retain zero-filled semantics. No memory increase, global coalescing, CRT, callback, guard, concurrency or SEH change is involved.
test/test-heap-empty-arena-reuse.js calls the actual production WASM public
allocation/free exports in fixed512MiB memory. The original pinned module
dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4
again failed at ascending iteration15,051/request35,840 after17,051 completed
pairs, with288 maps and reason3/1MiB followed by reason1/35,848. The repaired
module
7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0
completed2,000 constant plus100,000 ascending pairs with one1MiB backing
map and no OOM events, then20,000 varying-size pairs across bins, page sizes
and the large-allocation threshold. Live peer/large allocation contents were
checked byte-for-byte, and reused allocations were zeroed. Separate ownership
cases prove a locally owned binned header permits rewind with identical map
records; a live header or peer-owned binned header prevents it. Subsequent
small allocations and peer reuse preserve the new live allocation. Surviving
arena headers, map translations and passive free-list diagnostics validate.
The mandatory locked tools/build.sh passed. Existing arena-tail reuse
(11 cases), cross-instance release, randomized bins, large-free map pressure
and large-block arena tests all passed on the repaired source; tail reuse also
used the production module directly. The new regression took28.45s; the old
exact reproduction took2.25s including its supervisor. Logs, module/source pins
and command results are retained in the evidence directory.
This prevents the demonstrated balanced-pack accumulation; it does not reclaim arbitrary retired carved arenas or fragments owned by another instance. Those still need an explicit cross-instance ownership/generation solution before releasing their guest ranges. This scope is deliberate: zero live bytes alone does not make surviving free-list pointers safe to invalidate.
2026-11-01 repaired-allocator encounter attempt: disk stop at Setup
Evidence: scratch/runs/20261101-urban-chaos-repaired-encounter/, with the
actual original under ready-runtime/. No new HUD, movement or encounter
was reached. The first real run failure was the disk guard at
19:33:53.365540 UTC: free space sampled at2,003,222,528 bytes, below the
unchanged2,020,000,000-byte stop threshold. Launch free space was
2,152,325,120 bytes, above the2,100,000,000-byte requirement. Free space had
rebounded to2,151,092,224 bytes at19:34:03. The responsible disk writer and
duration of the transient are unmeasured; none of these finite samples was
below2GB. No RAM gate, threshold relaxation, second original or replay followed.
The exact Battlezone worker 01a19875-ed1d-7551-84ea-e74fb416adc1 author
DONE/PUSH RELEASE, fresh launcher final/result0 and direct owned PID/group
absence passed at19:32:26.191 UTC. Actual input main was
732f271c20da0df9561d73293576e17fa847f795; the25-minute gate ended19:57:26,
inside the20:19:28 outer deadline measured from this worker's launch. The
coordinator reclaimed a SHA-verified duplicate before the gate to restore
launch headroom. Runtime/source/build ownership was serialized; this worker
made no production source, test, build, fixture, TODOS or STATUS changes.
The corrected scratch guard reads/parses heartbeat before sampling its comparison clocks. Deterministic controls reproduce the old false future rejection and pass the corrected order while still rejecting genuine future, stale6.15s, dead/replaced processes and explicit failure records. Healthy3.5s delay, immutable first failure and direct disk/deadline controls pass. Detached Node and Python exit0/7/timeout controls retain durable terminal records. The exact accepted readonly observer was reused, including its whole512MiB, clock-identity and late10,001 retention validation.
One preparation error occurred before any original: the prerequisite helper
checked the independent guard before its first heartbeat existed. The guard
correctly rejected the absent file and both preparation supervisors stopped.
Root guard-failure.json, initiating-failure.json and terminal records remain
unchanged. preparation-race.json records the error, hashes and direct cleanup.
The scratch scheduling correction waits for independent-ready and two
advancing healthy guards in a fresh ready-runtime/ directory. It preserves
every guard predicate and the original gate/deadline; no game was replayed.
Fresh checks matched4,294 source/module/dependency pins and the1,122-file game
manifest. Fallen.exe SHA256 remains
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229;
config.ini remains reconstructed installation state. The published allocator
repair's exact locked build, regression and five focused tests were reused.
Actual runtime compilation observed repaired WASM SHA256
7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0.
After shutdown all4,294 pins,1,124 fixture/manifest pins and133 observed read
hashes match. Observation covers main-context synchronous whole-file and
complete bounded descriptor reads, excluding unhooked async/direct-OS reads
and guest-memory identity.
The sole original PID4084652 used the registered native software CLI,
--no-build --control-stdin --frozen --batch-size=100000, with steps at most25
and finite180s reply waits. Standard and native Setup PNGs at25 were visibly
reviewed. Last acknowledged batch26 processed the ordinary dlg-png request;
no mouse/key gameplay edge or Setup Okay was sent. The subsequent setup helper
rejected stopped guards before sending its first request. All four actual PNGs
pass CRC/complete-deflate checks; final.png has the same filtered raster as
the reviewed standard Setup image. The full runtime gzip is complete. No heap
refusal appears in this short startup log; it does not measure the game's
allocator behavior during gameplay or historic OOM/guard causality. The reused
observer has no first-refusal import sampler; no new production hook or
bin-flushing diagnostic was introduced.
Cleanup has a specific remaining evidence gap: the original Node exit hook records0, and the supervisor's actual OS wait records driver exit1, but there is no durable original OS-wait receipt. Primary and independent shutdown both sent SIGTERM; repeated driver signals interrupted its exception handling and finally block before it could publish that receipt. Its lost wait cannot be reconstructed from process absence or the Node hook. No fresh final key release/settled sample exists; no physical input was sent in this original. All recorded owned original/helper/watch PIDs and groups are directly absent, with no remaining helpers or new child temps. Own shutdown was delegated; native box ID/state/expiry are N/A.
The concrete blocker is the transient disk stop before gameplay, accompanied
by the shutdown receipt gap. A separately authorized attempt needs sustained
disk headroom and a single-owner, idempotent shutdown path that waits the
original and records its result despite repeated shutdown requests. Preserve
these first failures; RECOVERY.txt, control terminal records, identity and
capture validation permit offline recovery without another game launch.
Historical HUD2011/neutral2137/ArrowUp2263 remains the latest movement evidence.
The synthetic allocator repair is not demonstrated encounter success. No FPS,
audio, full27, browser or GPU qualification is claimed.
2026-11-01 scratch shutdown repair: repeated stops retain direct waits
Evidence: scratch/runs/20261101-urban-chaos-shutdown-repair/. This repairs the
demonstrated helper shutdown boundary using tiny native dummy processes only.
No game, guest, WASM, build, browser, GPU, boat, download or original replay ran.
Battlezone retained runtime/source/build ownership. The predecessor helpers,
historical receipts and first disk failure remain byte-identical. The original
game's missing OS-wait status remains unknown; encounter success is unverified.
The scratch primary now holds an exclusive shutdown-owner lock. The independent
guard preserves the immutable first failure and requests shutdown without sending
signals. The primary requests driver shutdown once. Driver signal handlers latch
requests without raising, including during finally. Cleanup avoids failed guard
checks, sends an ordinary nonblocking quit, drains output without waiting for EOF,
allows one second for cooperation and two for SIGTERM, then uses bounded SIGKILL
and a direct wait. It fsyncs original-wait.json, exit.json and driver cleanup
completion before exiting. The supervisor directly waits the driver and independent
guard. Repeated requests cannot reset cleanup deadlines or interrupt these writes.
The primary's driver budget is eight seconds. Only after the driver's actual wait
may the primary take over orphan cleanup as a Linux subreaper. It records genuine
waitpid status and bounds orphan TERM/KILL cleanup to three seconds. The deliberately
stopped driver required primary escalation; its missing driver cleanup receipt is
reported explicitly, and the adopted original has a genuine supervisor wait. An
external SIGKILL between reaping and durable publication can still lose a receipt;
the before-wait driver-death control does not establish recovery of that case.
All 15 accepted lifecycle cases pass using the actual corrected helpers. The
old driver fixture retains the predecessor's exact raising handler and finally;
repeated SIGTERM demonstrably kills it before its terminal receipt. The test owner
directly waits that dummy original. Corrected cases cover natural exit0/7,
simultaneous independent/primary shutdown plus30 repeated signals during cleanup,
slow cooperation/TERM, ignored TERM requiring KILL, pipe pressure, an owned
grandchild, driver death, stalled-driver escalation, supervisor shutdown, real
deadline expiry, independent death, duplicate ownership and shutdown before launch.
Normal original-wait and cleanup receipts precede the supervisor's driver-terminal
receipt without primary escalation. The slowest accepted case took10.53 seconds.
The exact predecessor guard, detached controller and control scripts were reused.
Read-before-clock ordering, healthy3.5s/stale6.15s, true future, dead/replaced PID,
explicit failure, immutable first failure and direct disk/deadline controls pass.
Detached Python/Node exit0/7/timeout and initiating-client death pass; the accepted
ledger directly waits all nine detached runners and three clients, with child
waits retained by their runners. Rejected preparation ledger/import errors remain
documented in preparation-notes.json; they are not accepted test results.
A later preparation correctly rejected the copied2.1GB launch requirement at 2,090,835,968 bytes before creating a dummy child. That rejection is retained. The real original launch requirement remains2,100,000,000 bytes. Explicit tiny fixtures use2,020,000,001 bytes and still run the unchanged direct guard stopping at or below2,020,000,000 before the2GB floor. No RAM gate was added. Finite sampled headroom is not proof of sustained disk safety or the historical writer's identity.
Offline recovery: python3 scratch/runs/20261101-urban-chaos-shutdown-repair/verify.py.
RECOVERY.txt, helper-pins.json, source diffs and SHA-pinned helpers.tar.gz
preserve the scratch implementation and exact test commands. The final direct
cleanup found all319 recorded owned PIDs and303 groups absent, with no own helpers
remaining; accepted lifecycle children were waited/reaped. Native local box
ID/state/expiry are N/A; delegated own helper shutdown is complete. Only this note
is committed. Next is a separately authorized original with sustained disk
headroom and a fresh gate integrating this boundary. Historical movement remains
HUD2011/neutral2137/ArrowUp2263; no new FPS, audio or full27 claim is made.
2026-11-01 durable encounter attempt: HUD reached, heap refusal blocks the next step
Evidence: scratch/runs/20261101-urban-chaos-durable-encounter/. The ordinary
encounter goal remains unachieved. One original reached the street HUD at2011
and acknowledged2036. During the next25-batch step, heap refusal recurred and
no reply arrived within180 seconds. No ArrowUp or other gameplay key was sent;
there is no movement, guard response, encounter or objective-progress claim.
No request or original was replayed after this failure.
The coordinator reviewed published scratch shutdown repair079959006 and its15
accepted lifecycle cases before dispatch. This worker reused the exact tested
drive.py, supervisor.py, independent.py, lifecycle.py, runner.py and
detached.py; only the copied guard's task absolute outer deadline changed.
Accepted lifecycle, guard and detached-controller controls were pinned/reused,
not rerun or presented as fresh gameplay proof. Guard/disk predicates and the
immutable initiating failure were preserved. No production/source/test edits.
Exact Battlezone worker01a1989a author DONE/PUSH RELEASE, fresh launcher/run
final/result and directly absent owned processes/groups passed the gate at
20:00:34.870 UTC on published main346788a42a40e77361c53fa0d2db45f59051eab7.
The25-minute deadline was20:25:34.870; the75-minute task limit stayed absolute.
All4,294 accepted source/dependency pins matched. The accepted allocator's
locked build, production/source regressions and five focused checks were
reused on exact unchanged inputs. The selected module's before/after SHA256 is
7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0.
All1,122 manifest paths match, with original Fallen.exe SHA256
8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229;
Config.ini remains explicitly reconstructed. All1,124 fixture/manifest files
and source pins still match after shutdown.
A41-sample direct20-second headroom check passed, minimum2,158,387,200 bytes.
Original PID4117224 launched with2,158,342,144 bytes free, using the ordinary
registered native software CLI, --no-build --control-stdin --frozen,
batch-size100000, and step requests at most25 with180-second reply bounds.
Launch2100000000, stop2020000000 and floor2000000000 were unchanged; no RAM gate.
The primary/independent histories contain1,812/1,810 samples, maximum gaps
0.636/0.599 seconds, with respective minima2,137,595,904/2,092,912,640 bytes.
The driver's observed minimum was2,135,396,352. There was no disk stop here.
These are finite samples, not continuous safety or an identification of the
historical transient writer. The coordinator owns its separate passive disk
observer; this worker neither changed nor stopped it.
Actual standard/native Setup25, ordinary Okay dismissal, title205, map307, briefing409, opening711/1111/1711 and HUD2011 were visually reviewed. Standard before/held/released/settled captures exist for Setup and three Enter phases. At2011 D'arci stands on the street with radar/health, fist icon and3:58 timer. The neutral interval has its actual before PNG and one successful25-batch step to2036; its later phase captures and subsequent ArrowUp route were never reached. All44 actual PNGs pass CRC/complete-deflate checks. No synthetic image, state/progress injection, new production hook or bin-flushing probe was used.
The log contains8,016 heap OOM messages. The immutable first passive observation retains the first1,048,576-byte sparse-commit refusal, then32,296 bytes refused by bump/low-reserve/sparse allocation, EIP00494d9f and caller0048c76e. It follows ACK2036; no subsequent ACK exists. The pending request was sent at20:05:16.251 UTC and its180-second wait expired around20:08:16. There is no per-line wall clock or first-refusal import sampler, so exact onset and OOM-to-stall causality remain unproven. The earlier zero-refusal observation through2011 was finite; it does not establish allocator exhaustion was fixed in real gameplay.
initiating-failure.json first records the detached neutral helper's exit1;
its output preserves the exact timed-out step. The driver then published the
secondary guard failure and entered the reviewed shutdown path. Ordinary quit
and SIGTERM could not stop the unresponsive original; SIGKILL was required.
Direct original OS wait records-9, durable after3.074 seconds of cleanup,
before driver cleanup completion and the supervisor's driver OS wait0.
Independent and supervisor direct waits also record0. The primary did not
escalate the driver. Node exit hooks are not used as OS-wait evidence. This
actual failure path retains the waits missing from predecessorc157668dc.
SIGKILL prevented the unchanged observer's exit-only loaded.json, compile-event
record and Node exit hook from being published. Actual consumed-byte and
compiled-byte observer identity are unavailable, not inferred from previous
runs. The direct frozen inventory/startup log names Fallen, binkw32, mss32 and
seven Miles modules; actual-file-pins.json maps those ten loaded module names
and bases to unchanged before/after file hashes. That narrower evidence does
not enumerate every on-demand game-data read or hash in-memory consumed bytes.
The selected WASM path and disk hash match; a fresh actual compilation observer
hash is not claimed.
The audit retains234 requests,233 successful replies and the one missing reply. No final neutral release/settled sample could be obtained after the timeout. The last complete input sample is2011: known queues empty, sampled physical and async keys released, numeric mouse mask0. Only three balanced Enter pairs and Setup mouse input were sent; no held movement key existed. That sample does not prove final2036 guest post-queue or SDL state. No final image was fabricated.
Direct cleanup found all31 then-recorded owned PIDs and30 groups absent, no own
helpers and no new child temp directories; publication processes are checked
again in final cleanup.json. Original/helper shutdown was delegated and is
complete. Native local box ID/state/expiry are N/A. verification.json passes
scoped evidence checks while explicitly recording the unsuccessful gameplay
goal and observer/final-release gaps. RECOVERY.txt, exact helper pins, complete
runtime gzip, request receipts and PNGs support offline inspection.
Next: investigate the remaining real-game sparse-commit/allocation refusal and unreturned2036→2061 step before another separately authorized original. The synthetic allocator repair does not establish encounter success. No FPS, audio, full27, browser/GPU or broader gameplay qualification is claimed.
2026-11-01 post-repair static finding: rewind eligibility is the missing discriminator
Evidence: scratch/runs/20261101-urban-chaos-postrepair-heap-cause/analysis.json,
reproducible with python3 scratch/runs/20261101-urban-chaos-postrepair-heap-cause/analyze.py.
Static inspection only; no original, WASM execution, build, synthetic allocation
test, browser or box operation. Eight relevant source/interface files match
published af500af41 byte-for-byte at analysis head 71671ee45; their exact
SHA256 values are retained. Allocator repair is 0f9dfbb45eff6ded35a15bcf3a316c700594d831.
The accepted selected module remains
7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0;
the durable original's actual compile/consumed-byte observer records remain absent.
The narrow production-path conclusion is fresh sparse-arena commit refusal
after current-arena reuse did not serve the request. The particular rewind
rejection and underlying commit predicate are unknown. In
src/10-helpers.wat:2644, $heap_sparse_alloc tries $heap_sparse_rewind only
on current sparse bump overflow. Its subsequent reason3 proves that it went on
to obtain a guest reservation and that commit/reclaim logic still returned0.
This is not the reason2 guest-VA-reservation failure, reason4 heap-arena-record
registration failure, or reason5 bad-size path. Rewind might have had no current
record; the log does not establish that a populated arena was examined.
$heap_sparse_rewind (src/10-helpers.wat:2573) considers one current arena,
not all zero-live arenas. Any live allocation in that arena rejects rewind
before bin flushing, even if its other bytes are free. It also rejects cursor/
record mismatch, insufficient arena size, a whole-block first header, invalid
local links, missing local ownership of any carved header, extra interior links,
or the65,536-list/1,000,000-search bounds. A peer-held free header can therefore
prevent reuse without any live bytes. Retired carved arenas are still excluded
from $heap_arena_release_free (:2016). These are exact production conditions,
not measured explanations for this game's failure. One persistent allocation
sharing the pack arena is a sufficient way to defeat the repair's empty-arena
precondition; its occurrence here is unmeasured.
The balanced102,000-pair regression established reuse with one map when temporary
requests were freed before the next request. It did not establish that actual
game arenas become empty, remain locally owned, or pass the proof bounds. The
legacy VB path allocates the pack in $d3dim_pack_fvf_vertices
(src/09ab-handlers-d3dim-core.wat:1693), then frees a successful pack after
the draw (:2194). It has no private pack arena: other $heap_alloc users in
the same instance can share its bump arena. A balanced pack/free call pair
therefore does not imply an empty arena at the next bump miss.
Offline full-gzip parsing confirms8,016 messages in4,008 reason3/reason1 pairs,
first at one-based line560 (the predecessor's559 is a zero-based index).
First extent32,296 follows ACK2036; the sole pending request is
route-1793563516251341528, step25 sent at20:05:16.251 UTC, with no terminal
ACK receipt and the retained180s timeout. Small extents span296..44,264 bytes,
all 32*N+8 (arithmetically9..1,383 vertices, not captured pack operands).
All exceed HEAP_BIN_MAX=256 and are below HEAP_LARGE=0x7f000: a fitting block
hidden directly in a small bin cannot serve these requests. Bins can still
affect ownership/adjacency. There are7,260 site lines at00494d9f/0048c76e,
724 at00494d16/0048c76e and32 at00494d16/401ac090; only the first site's prior
disassembly is used here. Do not attribute every refusal to one captured caller.
Runtime gzip SHA256 is
8b158c5e316a05ccf26ac051f8fce59ac5efdf497967dea70db0920ebcba496d.
Absent actual state: map/reservation/hole/arena tables, allocated high-water and
live-byte words, current sparse cursors, free-list headers, instance bin heads,
lease state, and allocation ownership/history. Host input/scheduler samples do
not contain these. Reason3 is emitted after failed-range release/possible
rollback; it cannot identify which $virtual_map_commit_locked check failed
(src/10-helpers.wat:1197): backing placement/split exhaustion, map slots,
lease exclusion or page publication remain candidates. The288MiB capacity is
not a measurement of occupied or live bytes. Disk did not stop the original;
direct OS wait-9 precedes driver wait0. Neither refusal order nor that timeout
proves OOM-to-stall causality.
One next discriminator, on a separately assigned original: take one bounded
first-reason3 snapshot at the existing heap_oom_trace callback, before its
formatting. A source breakpoint through Node's existing inspector at
lib/host-imports.js:2034, conditional on reason===3, permits read-only
inspection without adding a hook or editing source. This is a proposed capture,
not an executed/validated inspector procedure. Ordinary control-stdin eval
(test/run.js:7372) runs between slices and cannot promise this boundary while
a step is unreturned; --trace-host=heap_oom_trace logs arguments only.
At that stop retain size/reason, wall/monotonic time, last ACK/pending request,
allocating instance and WASM stack. Copy matched-region VIRTUAL_MAP_STATE,
VIRTUAL_MAP_TABLE, VIRTUAL_RESERVE_TABLE, VIRTUAL_HOLE_TABLE, HEAP_ARENAS,
HEAP_SHARED, PTEs and active Win16 lease ledger; read pure heap cursor getters
and heap_stat. Obtain the current record by matching sparse end/high in the
copied arena table. Read validated8-byte headers through a host translator
using the copied PTEs; bound to65,536 headers/list nodes,16MiB output and5s
capture work, mark truncation/invalid translations explicitly, and remove the
breakpoint after the first hit. Keep guard/deadline/shutdown ownership intact;
abort capture on its bounds rather than extending runtime guards. Require
allocator writers quiescent, or label the copy non-atomic. Passive
d3d_render_free_list_diagnostic (src/09ai-d3d-render-lifetime.wat:44) exposes
this instance's head/reason but omits bins. Never use get_free_list,
heap_free_census, retire, adopt or coalesce as read-only probes: they mutate.
Read private bins from debugger state only if actually exposed; otherwise
record them unavailable and leave unlisted headers' ownership unknown.
Offline distinguish reserved VA (reservation intervals), occupied backing
(union of published map backing intervals, including non-heap owners), arena
live charged extents (+12, not user payload), and validated locally free
extents (list/bin membership). Compare total/largest backing gaps with the
1MiB commit and recursive64KiB floor; total free bytes alone cannot prove a
fit. Within the current arena, compare live count, cursor/high/end, free
header coverage and adjacent runs against need32,296 and the exact rewind
proof/bounds. Intended hypothesis: live co-residency or incomplete local
ownership prevents current rewind while retired carved backing remains held.
High live occupancy supports capacity pressure; zero-live stranded adjacent
space supports retention/fragmentation only where free ownership is validated.
Ample fitting backing with metadata/lease/publication constraints unresolved
requires the exact failing predicate next. Missing bin/peer ownership cannot be
inferred from aggregate zero live bytes. No source repair or memory increase
is justified by this analysis, and this capture plan proves no cause.
Only the note is committed; the offline script/report remain in the run directory. Own board watcher was stopped and directly waited; no other helpers were launched. Local box ID/state/expiry and shutdown are N/A; Battlezone's runtime/source/build slot and coordinator-owned processes were untouched.
2026-11-01 first-refusal debugger: dummy capture passes; original fails at batch0
Evidence: scratch/runs/20261101-urban-chaos-first-refusal-debugger/.
The allocator discriminator remains missing. The single authorized original
crashed at batch0 in NamedEventAuthority initialization, before the retained
Setup/Enter route or any recorded first-reason3 breakpoint hit. No original was
repeated. This is a concrete runtime blocker, not evidence about rewind
eligibility, sparse backing pressure, or the cause of the earlier post-ACK2036
stall.
The existing-source inspector proposal was first exercised on a tiny scratch
WASM module importing the unchanged lib/host-imports.js callback. Accepted
controls are prep-final/ and prep-final-watchdog/; preparation-audit.json
distinguishes earlier scratch failures. A conditional breakpoint at line2034
skipped reason1, paused on the first reason3, and retained actual paused callback
arguments, a WASM caller/local42, a private global512, matched shared memory,
an audited pure cursor getter and a validated24-byte synthetic arena header.
The same bounded host-copy reader captured synthetic table/current-arena data.
The accepted snapshot is9,381 bytes and took109.38ms; the limits were16MiB and5s.
Breakpoint removal and resume were acknowledged, the later reason3 did not
pause again, and both processes were directly waited with exit0. Startup
--inspect-brk/pending-source-breakpoint resolution was also exercised. These
are synthetic preparation controls, never game allocator evidence.
The independent watchdog control deliberately left the dummy paused after its read-only snapshot. Its parent continued sampling independently of the paused Node event loop, stopped the dummy after about2s, and durably recorded direct OS waits. No allocator diagnostic, guest translator or mutating free-list/bin probe was called. The reader labels shared-memory copies non-atomic and missing bin ownership unknown. It caps header/list work at65,536 nodes and validates arena bounds, extents and copied-PTE translations before header reads. Private bin indices were not validated against the selected production module; the scratch reader does not infer ownership from absent links or zero live bytes.
Two preparation corrections are preserved: the first scratch wrapper falsely reported pass after its error path exited0, while its terminal file actually reported a missing WASM caller. Node's paused frame URLs were empty; selecting the actual module scope fixed the matcher, and subsequent acceptance also checked terminal error and snapshot contents. A later scratch expression had a syntax error, corrected before the accepted controls and original. Neither failed attempt was accepted as preparation or represented as game evidence.
At20:38:12.637 UTC, exact Urban worker01a198c0 and Battlezone worker01a198bc
author releases, terminal launcher final/results, published cleanup evidence
and direct process/path checks passed. Runtime/source/build were then claimed
on main e48dcc193797bff09035a5b05142e92383b07d88. All4,294 accepted source/
dependency paths and1,124 fixture/manifest paths matched before and after.
The selected actual path is
/home/user/wine-assembly/build/wine-assembly.wasm, SHA256
7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0.
Original Fallen.exe and the reconstructed Config.ini retain the accepted
fixture identities. Exact paths and file hashes are in the pin reports.
No source, test or build changes were made and no build was run.
The exact reviewed079959006 drive.py, supervisor.py, independent.py,
lifecycle.py, runner.py and detached.py were reused byte-for-byte. Only
the copied guard's absolute task deadline changed. Accepted build/regression/
lifecycle receipts were reused on matching inputs, not rerun or claimed fresh.
NODE_OPTIONS=--inspect-brk=127.0.0.1:0 enabled the existing Node inspector
without changing the driver or production code. The command remained ordinary
registered urban_chaos_demo --no-build --control-stdin --frozen, software,
batch-size100000, with the retained observer and control route. Both independent
guards remained active; the runtime deadline stayed21:03:12.637 UTC, within
the original25-minute and overall60-minute limits. It was never extended.
The41-sample20-second headroom check had minimum2,182,795,264 bytes; original PID4156945 launched at20:38:35.957 with2,182,770,688 bytes free. Launch2100000000, stop2020000000 and floor2000000000 were unchanged, with no RAM gate. Primary/ independent histories have289/290 samples, minima2,181,300,224/2,181,341,184 and maximum gaps0.600/0.602s. The driver's observed minimum was2,177,409,024. No disk stop occurred; finite samples do not establish continuous disk safety.
The source breakpoint resolved before the first control ACK. Ping acknowledged
batch0 at20:38:37.948; the pending request was exactly
{"action":"step","n":1,"id":"2"}, sent immediately afterward. No ACK2
exists. The actual log reports:
CRASH at batch 0: Authority deadline exceeded; operation outcome unknown; client is unusable.
Its stack identifies receive and new NamedEventAuthority in
lib/named-event-authority.js, through cliEventEndpoint, openEvent and
h.open_event; guest EIP is004c0fc0. The constructor waits for its Node Worker
with a3000ms bound. Inspector/Node Worker startup interaction is a candidate,
not isolated causality. The tiny dummy had not validated named-event Worker
coexistence. There is no recorded reason3 hit, allocator snapshot or heap OOM
line. The controller never passed driver readiness: no Setup/Enter/gameplay
input, route request or PNG was produced. No allocator state was reconstructed.
The genuine batch0 failure is retained in initiating-runtime-failure.json.
Own debugger PID4156953 was explicitly terminated/disconnected and directly
waited-15, permitting the already-failed Node process to leave its debugger
disconnect wait. There was no explicit remove-breakpoint command on the
original; removal/resume were validated on the dummy only. The reviewed helper
then durably recorded original OS wait1 at20:39:32.677, before driver OS
wait0, with no driver escalation. Independent/supervisor direct waits are0.
The detached controller and its runner exited1 on shutdown before readiness;
their actual waits remain distinct from the original failure. The parent's
secondary “debugger terminal before reason3” error followed the deliberate
disconnect and does not replace the initiating crash. No request was replayed.
Unlike the previous SIGKILL attempt, the unchanged observer published its exit
records: the actual WebAssembly.compile input hash matches the selected
module, and122 observed main-context file reads match retained before/after
pins. These are the observer's sync whole-file/complete bounded-descriptor
observations, not an identity for every consumed game byte or guest memory.
The Node exit hook is separate from the actual OS wait receipt.
Direct cleanup found all40 recorded runtime/preparation PIDs and30 groups
absent. One exact empty authority temp directory was removed with rmdir;
temp-cleanup.json states that its ownership was inferred from the exclusive
runtime, constructor timing and sole matching new empty directory, rather than
directly instrumented. Foreign directories were untouched. Append-only board
watchers are separately owned/waited; final cleanup and publication receipts
retain those results. Native local box ID/state/expiry are N/A. Delegated own
debugger/game shutdown is complete, runtime/source/build are author-released,
and no runtime or follow-up task is held.
Next prerequisite is synthetic inspector/named-event Worker coexistence preparation before any separately authorized original. This run authorizes no unchanged retry. Actual current-arena live co-residency/local ownership/bounds, fresh backing/metadata refusal and the precise commit predicate remain unknown. No encounter, FPS, audio or full27 claim is made. Only this note is committed; evidence, scripts, exact commands and recovery limits remain in the run directory.
2026-11-01 inspector/authority coexistence: startup pause isolated synthetically
Evidence: scratch/runs/20261101-urban-inspector-worker-coexistence/.
The synthetic preparation blocker is resolved; the game allocator discriminator
is still unknown. No original, guest runtime or production WASM was executed.
Constructor owned the heavy runtime lane during these controls; this worker
never claimed runtime/source/build ownership.
Only this note is committed. The controls use Node24.18.1/libuv1.52.1 and the
unchanged lib/named-event-authority.js API with its default3000ms deadline.
The observed cause in the small reproducer is an inherited Worker startup
break, rather than a slow event server or a broken event protocol. Ordinary
Node completes authority construction and event operations. With
NODE_OPTIONS=--inspect-brk=127.0.0.1:0, the same constructor times out around
3000ms under both the existing historical inspector controller and a minimal
Runtime-only controller. Enabling NodeWorker with
waitForDebuggerOnStart:false alone does not cure it. The nested Worker
protocol actually reports Debugger.paused, reason Break on start, at the
start of the unchanged authority module. Sending the Worker
Runtime.runIfWaitingForDebugger and resuming its startup pause makes authority
construction and normal client operations pass, without extending the deadline.
The client bridge independently exhibits the same startup break and needs its
own resume under that launch mode.
brk-worker-state/protocol.ndjson retains paused frames and direct Worker-scope
reads: parent and both Workers have process.execArgv=[], while their
NODE_OPTIONS contains the exact inspect-brk option. inspector.url() is
undefined in those Workers; that does not mean they are not paused through the
parent's Worker inspector transport. The attachment event even says
waitingForDebugger:false before the subsequent Break on start event.
Thus neither execArgv alone nor that attachment field is a sufficient readiness
check. Worker resumes change the outcome in this controlled test. This isolates
a concrete mechanism matching the earlier batch0 failure; the historical
original did not record its Worker debugger state, so this does not retroactively
supply an original Worker trace or any allocator evidence.
The preferred scratch strategy avoids Worker startup breaks entirely:
main-wait.cjs opens an ephemeral loopback inspector and calls
inspector.waitForDebugger() only when isMainThread is true. Launch the
intended target with a per-process environment containing
NODE_OPTIONS=--require=/home/user/wine-assembly/scratch/runs/20261101-urban-inspector-worker-coexistence/main-wait.cjs,
CASE_DIR=<owned evidence directory> and STALL_AUTH=0; retain any separately
required observation preloads and their exact hashes. Do not also pass
--inspect-brk through options or argv. Clear this target-only NODE_OPTIONS
from the inspector helper's environment. Arm the existing conditional source
breakpoint, then acknowledge Runtime.runIfWaitingForDebugger; require actual
source breakpoint resolution before issuing the diagnostic trigger. The preload
is inherited, but Workers skip the main-thread wait and have no inherited
inspect-brk launch option. It changes no Worker constructor, production module,
allocator state or authority timeout. The STALL_AUTH=1 branch is exclusively a
synthetic negative-control injection before the authority module, never a
future-original setting. The external parent bounds startup/attach and total
synthetic work, and directly owns shutdown and OS waits.
The final accepted suite is final-validation.py, with per-case results:
| Control | Actual result |
|---|---|
final-normal |
Authority ready33.07ms; create/acquire/set/reset, finite registered wait/status/forget/close pass; ENOENT and EBADHANDLE preserved. |
| Same process, tiny existing-source callback | Reason1 skipped; first reason3 paused at lib/host-imports.js:2034; synthetic memory/local/WASM caller snapshot9,377 bytes in90.55ms; breakpoint removal and resume acknowledged; later reason3 runs without another pause. Target/controller direct waits0/0. |
final-watchdog |
Event API passes, then controller deliberately leaves the tiny callback paused. Independent Python parent triggers at about2s, sends own SIGTERM and directly waits target/controller-15/-15. Snapshot9,365 bytes; no dependence on paused Node timers. |
final-stall |
Scratch preload truly stalls the authority Worker with Atomics.wait before module entry. Same strategy still reports ETIMEDOUT at3003.40ms; target direct wait1; controller explicitly stopped/direct wait-15. No snapshot or false pass. |
final-dead-client |
After normal event operations, own authority Worker is explicitly terminated and awaited. Client reports ECONNECTION, becomes closed, and next call reports ECLOSED; tiny snapshot/removal/resume still pass; direct target/controller waits0/0. |
dummy.cjs and capture.js are byte-for-byte copies of the prior tiny controls;
no production compiled module is loaded. The controller is a scratch derivative
of the previous inspect.cjs: error-path cleanup failures are recorded, and its
successful stop no longer issues an unnecessary resume on an already running
target. RPC and paused evaluation errors still fail the control. The exact
historical controller is also run directly in historical-exact-controller/ to
reproduce the inspect-brk timeout, separately from this cleanup correction.
The accepted reader retains its5s/16MiB capture bounds, checked actual snapshot
contents, pure getter use, non-atomic-copy label and unknown bin ownership.
The scratch event harness calls real API methods and asserts both successful
results and expected protocol errors; none are converted into success.
pins-before.json and pins-after.json match156 files: all selected lib JS,
Node executable and then-current scratch inputs. Additional historical-control
inputs and final script identities are retained separately. No production
source/test/build edit or build occurred. Evidence remains below8MB; sampled
/home free space remains above2,000,000,000 bytes. Every own target/controller
is directly waited and subsequently checked absent; failed constructor namespace
cleanup is checked for the deliberately stalled Workers. The board watcher has
its own direct wait receipt. Box ID/state/expiry are N/A; no box was created,
shutdown was explicitly delegated for these own synthetic helpers only, and no
runtime ownership or held follow-up remains.
Next is a separately authorized original using this bounded main-thread attach strategy and freshly verified source/fixture/module identity, with the existing independent guards unchanged. This assignment authorizes no original retry. The current-arena/ownership/backing/metadata discriminator and exact commit refusal predicate remain missing; synthetic success is not game, encounter, FPS, audio or allocator-cause proof.
2026-11-01 main-thread first-refusal assignment: launch headroom fails
Evidence: scratch/runs/20261101-urban-main-thread-first-refusal/.
No original was launched; the allocator discriminator remains unknown.
The required sustained launch check stopped at its first sample, at
21:05:19.887 UTC: /home had 2,096,635,904 bytes available, which is
3,364,096 bytes below the unchanged 2,100,000,000-byte launch threshold.
The intended 41-sample/20-second check therefore did not pass. The
2,020,000,000-byte stop and 2,000,000,000-byte floor were retained; no RAM gate
was used. No disk recovery, lowered guard, or repeated admission attempt was
performed after this prerequisite failed.
The earlier prerequisites passed: exact Constructor bf77b2c885 author RELEASE,
the coordinator's terminal/direct-cleanup verification, fresh absence checks
for its recorded runtime PIDs, and ancestry of both Constructor and f3bb8e295
on main. Constructor's successor is static only. Runtime/source/build were
never claimed by this worker. All 4,294 source/dependency paths and 1,124
fixture/manifest paths matched before and after. The selected, unexecuted
module remains /home/user/wine-assembly/build/wine-assembly.wasm, SHA256
7278c1018fa8208731a8e285f0bb2a5c35aa8e299592750ce3cedf6d354caef0.
There is no actual compile-input or consumed-game-byte observation in this run.
The exact reviewed 079959006 lifecycle helpers, observation preload, bounded
reader and f3bb8e295 main-thread preload/control receipts were pinned for
reuse. Unchanged synthetic controls were not repeated. Scratch helper copies
are preparation inputs, not evidence that the attachment or game ran. No
debugger, driver, supervisor, independent runtime guards, Setup/Enter input,
encounter input, breakpoint trigger, or snapshot was started. Consequently
breakpoint resolution, run-if-waiting acknowledgement and removal/resume are
unexecuted here. Current-arena ownership, live co-residency and bounds versus
fresh backing or metadata refusal remain unknown; no predicate is inferred.
The headroom sampler PID4185307 was directly waited with exit2 and checked
absent. The append-only board watcher has its separate terminal wait receipt.
No game/debugger cleanup ordering was exercised because neither existed.
Local native box ID/state/expiry are N/A; no foreign process was stopped.
Only this note is committed; no source/test/build changes or build occurred.
RECOVERY.txt, pins and the exact failed sample remain in the run directory.
Next requires sustained launch headroom and a fresh assignment; this stopped
run holds no runtime or follow-up and supplies no gameplay/FPS/audio/full27 proof.
2026-11-01 released-log recovery: authorized scope still insufficient
Evidence: scratch/runs/20261101-urban-released-log-headroom-diagnostic/.
Recovery completed losslessly, but no original was launched. The seven
explicitly authorized, completed worker events.jsonl files totaled 3,851,546
bytes. Their gzip replacements total 1,084,440 bytes, recovering 2,767,106
logical bytes. At recovery completion /home had 2,096,902,144 bytes
available: 3,097,856 bytes short of the unchanged 2,100,000,000-byte launch
requirement. This is the first blocker. The sustained launch check was not
started; stop2,020,000,000/floor2,000,000,000 and no RAM gate remain unchanged.
No additional recovery scope, original, or launch retry was attempted.
Only events.jsonl in these scratch/usage-watch/ output directories changed:
urban-main-thread-first-refusal-nov01,
constructor-uvconfig-probe-branches-nov01,
urban-inspector-worker-coexistence-nov01,
constructor-original-startup-nov01,
constructor-original-fixture-recovery-nov01,
new-local-game-replacement-nov01, and
urban-chaos-first-refusal-debugger-nov01. Each was claimed individually after
reviewing its author release, terminal exit0 result and final. Size, inode,
mtime and full SHA were stable across a one-second interval. Compression was
staged in tmpfs. The durable destination gzip and its full decompressed SHA
were verified, and a recovery receipt fsynced, before unlinking each original.
Final verification repeated all seven archive SHA/full round-trip checks and
confirmed preserved result/final/notification hashes. The receipts contain
exact paths, full hashes and restore commands; recipes, pins, fixtures, saves,
active logs and coordinator logs were untouched.
Constructor's sequencer-contract successor owned runtime/source/build first; this worker stopped on insufficient recovery without awaiting or claiming its release. There was no Urban runtime/source/build claim or production edit. Post-release source/module/fixture verification was therefore unexecuted; this run does not assert unchanged source while Constructor edits ToyVM. The historical selected module identity remains a requirement, not a freshly tested compile identity here. Reviewed lifecycle helpers and main-thread inspector controls were not rerun or exercised. No debugger, game, guard, route input, breakpoint acknowledgement or allocator snapshot exists for this run. Current-arena ownership, live co-residency and bounds versus backing or metadata refusal remain unknown.
Only the append-only board watcher required shutdown; its actual wait and direct absence are recorded separately. There was no original OS wait or debugger removal/resume sequence to exercise. Native box ID/state/expiry are N/A; Constructor and coordinator lifecycle ownership remain untouched. Only this note is committed. The assignment stops without held work; next requires separately restored sustained headroom and a fresh assignment. No gameplay, FPS, audio, full27 or allocator-causality claim follows.
2026-11-11 current browser: player movement and melee damage
Evidence: scratch/runs/20261111-urban-chaos-current-player-consequence-1010/.
The original installed demo now reaches player control and a visible combat
consequence through the normal urban_chaos_demo application configuration.
No production source change was needed. Historical allocator and held-input
failures did not reproduce on this route; this does not identify their cause.
The tested source was a603418dbf281e1f7844079a3b151b8b309d0d9e.
The actually loaded 2,189,363-byte module was SHA256
75ecda5701c15f6fe2c269455ead4e15457b90653733df15d35674b08e8849d5.
The original test/binaries/candidates/urban-chaos-demo/game/Fallen.exe
was SHA256 8cc73044e19b423ef0f13858d8cb3bb3025ca023114cf87ff46c57838f4b2229;
the complete companion manifest passed test/test-urban-chaos-recovery.js
(1,122 files). The matching locked build was reused with unchanged source and
module pins (reused-build-proof.json), rather than rebuilt for this notes-only
change. Chrome 151 used the live owning legacy/WebGL endpoint and a WebGL2
render-worker context. The viewport was 1440×1000, game backing 640×480;
observer attachment preserved the viewport with defaultViewport: null.
Native Setup, Options/Keyboard, Start, map selection and briefing were navigated
from reviewed full images. The shipped Keyboard screen is retry/retrieved/ original-output/frame-28.png: numpad movement, Z punch, X kick, C action.
Opening camera/loading images were excluded from player-motion evidence.
HUD frames 50–51 establish an idle player beside the curb, with the police car
and phone booths across the street. Balanced ArrowUp holds then move D’arci
relative to the curb and scenery, with the follow camera moving as she advances
into close melee (frames 53, 55, 60 and 67). Numpad8 was also tried; it is not
claimed as independently conclusive movement. Z/X produce combat poses against
an approaching hostile NPC. Frame 69, compared with idle frame 51, shows the
fixed HUD health arc visibly depleted after the exchange. Enemy health-bar
changes are also visible, but perspective and occlusion prevent exact HP claims.
No arrest, kill, training completion or mission completion is claimed.
The same sandbox passed harmless trusted, balanced input controls, including
Z/X/C and numpad down/up, before their use in the game. Each action retained a
fresh reviewed full-frame hash and live owner/module/backend checks. The first
attempt refused Enter before dispatch because the scratch controller omitted
legacy from its backend allow-list. Its normal exit, actual waits and absence
checks were retained; the concrete allow-list correction and fresh same-box
controls preceded the retry within the original fixed allocation. An initial
extra-key qualification failed before input because its observer import was
missing; the corrected direct connection retained viewport and identity guards.
No uncertain command was replayed, guest state edited, or existing guard removed.
Escape offered Resume/Restart/Abandon, with no save entry. Abandon and the normal
main-menu Exit sequence returned the guest with ExitProcess code 0 (frame 93).
Chrome, controller, owner and transport have actual waits and separate stdout/
stderr EOF evidence. Independent checks found their PIDs, tmpfs profile and
loopback server absent. All unique captures/logs and extra-key dispatch receipts
were recovered and raw-SHA verified before shutdown; compressed JSON has a
round-trip hash map and restore recipe in the run. Observer tests passed (2).
Private box bx_2m93xu6a was independently confirmed archived, expiry null,
with nonforced completed stop stop_e1eefccb1f4c2c2b at
2026-11-11T06:39:25.780Z. Its original expiry was
2026-11-11T06:54:50.969Z; the prearmed fixed watchdog was never extended and
exited unfired with actual wait 0 and both EOFs. No box remains running or held.
The assigned player-control-plus-consequence goal is complete; a later arrest
or mission-completion route would be additional work.