Deus Ex demo — local candidate

The official 1.002f demo installer is pinned in test/candidate-corpus/manifest.json from Archive item DeusExDemo. Its bundled license permits private/domestic use but does not grant redistribution, so the installer and installed fixture stay ignored and must not be deployed or rehosted.

Fetch and test:

node tools/fetch-candidate-corpus.js --id=deus-ex-demo
node test/test-deus-ex-demo.js

The fetcher retains DeusExDemo.exe as the original distribution. Wine Assembly runs that self-extractor, captures the Setup.exe it launches, and resumes setup from the captured guest VFS. The wizard installs to C:\DeusExDemo. The frozen driver exports the VFS after the complete game payload is written and before setup chains into its obsolete bundled DirectX updater; Wine Assembly already provides that runtime. No host archive or cabinet extractor prepares the playable tree.

The test uses the prepared installed tree, verifies the software-renderer INI, and launches the same app manifest as the browser. It requires native UE1 DLLs, live audio, and a rendered animated intro without spending the normal suite's time budget on a full map load. A separate frozen controlled acceptance loaded the Training map, rendered its 3D HUD and transmission sequence, and moved JC through the world with ordinary Win32 keyboard input.

Historical browser result (superseded by the 2026-10-31 browser acceptance below): the browser's experimental Worker backend is functional but substantially slower during UE1 startup. An exact isolated Chrome run stayed live (changing main-thread EIPs, no trapped Worker), created its third Worker and 648x508 game window after roughly 45 seconds, and progressed from the gray allocation frame through the animated Deus Ex title. It was still in that title sequence around 90–110 seconds. Keyboard events were routed with hwnd=0 in this mode even after a scripted viewport click, so the interactive acceptance remains the cooperative backend; leave Threads off for the responsive demo path.

Headless route to Training, uop tier on (2026-09-29)

With the uop tier on (the default) and 553db124 (never enter a freed uop program), the demo runs from boot through the main menu into the Training map and walks. Nothing else needed fixing. Driven frozen, one step at a time:

node test/run.js --app=deus_ex_demo --quiet-api --quiet-blocks --no-close \
  --control --frozen --max-seconds=3000 --max-batches=100000 \
  --batch-size=200000 --tick-ms-per-batch=25
# then with tools/ctl.js (screen coordinates; the window's client centre is 345,284):
step 3000                     # rotating 3D logo (the menu's background level)
key VK_ESCAPE; step 200       # "Welcome to DEUS EX" main menu
cmd mousemove:502:396; step 30     # cursor onto Training (see mouse notes)
cmd mousedown:345:284; step 40; cmd mouseup:345:284; step 100   # LOADING
step 400                      # UNATCO Training Facilities, HUD, Jaime Reyes transmission
cmd keydown:38; step 150; cmd keyup:38   # JC walks to the first door

About 7 minutes of wall clock to the logo on an idle bench box (cooperative threads). Every capture above was checked by eye.

The mouse is relative, not absolute. Once the viewport captures the mouse, WinDrv's WM_MOUSEMOVE handler (windrv+0x1110ad8f) takes lParam minus the client centre as a delta, feeds MouseDelta/IK_MouseX/Y and recentres the cursor with SetCursorPos (windrv+0x1110af82). So an absolute mousemove:X:Y moves the game's own cursor by about 2 x (X,Y - 345,284), and a move up or left of centre from the start position (top-left corner) is clamped and looks like nothing happened. click alone does not press a menu button: the button only takes focus from a move event over it, and the press must span game frames — mousemove onto the button, then mousedown, step 40, mouseup. WinDrv's button handlers (0x1110b4a4 down, 0x1110b302 up) were confirmed to fire with --count; the earlier "click does nothing" was only the missing hover move, not an input bug.

Compatibility work required for this route:

Call-form census (2026-09-29)

Route. Box1 ran --app=deus_ex_demo --batch-size=200000 --tick-ms-per-batch=25 --repaint-every=10.

Census over 450..800 (docs/uop-tier-design.md §15.1):

MMX qword-fill fold (2026-09-29)

The fill loop is the SoftDrv frame clear: movq [edi],mm0; add edi,8; dec ecx; jnz. It is at softdrv+0x10d3ed70, and a 16-bit-path copy sits at +0x10d3eec0. Its first entry falls through from shr ecx,1; nop at +0x10d3ed6d. It is now one H419 0x80000005 super-op (docs/loop-idiom-superops-design.md §23; gate --no-mmx-fill-superops).

Box1 run. Route: --app=deus_ex_demo --batch-size=200000 --tick-ms-per-batch=25 --repaint-every=10 --branch-clock --wall-clock-ms=1790673326000 --max-batches=800.

Next hottest. All SoftDrv compute, declined head-unsupported by the uop tier:

Covered (2026-09-29, docs/uop-tier-design.md §16). The uop tier now lowers MMX, and it retries a scan-limit head with a halved span. Every head above is live. +0x10d2b180/+0x10d2b640 run 86 blocks per entry, at 276 instructions each. +0x10d05dbb is reached inside another program. Threaded block entries over batches 450..800 went from 177.5M to 100.4M. User CPU for 800 batches went from 55.83/55.94 s to 27.69/27.75 s (−50.4%), and all five frames are md5-identical. What remains hot is x87 triangle setup: +0x10d2759d/+0x10d27656 are fld heads, 4% each.

2026-09-29: the uop tier ran a freed program (call through NULL in the 3D intro)

Symptom. With the uop tier on (the default since 2026-09-28), the 3D logo intro died around batch 1163–1174 on a guest call through NULL: EIP 0x10000, then an AV. A pinned calendar makes it deterministic. It crashes at batch 1163 on every run:

node test/run.js --app=deus_ex_demo --batch-size=200000 --tick-ms-per-batch=25 \
  --repaint-every=10 --wall-clock-ms=1790673326000 --uop-census --quiet-api \
  --no-close --max-batches=1200 --max-seconds=1400

This is a heavy route (about 5 min for 3000 batches), so run it on a bench box, not the Mac.

Cause. This is not a slot-indexing bug and not a Deus Ex bug. The tier entered a program the arena had already freed:

The census (instrumented run, box 4) shows the exact coincidence:

Fix (src/07d-uop-engine.wat). Each arena now keeps a program-starts bitmap, one bit per 16-byte granule of code bytes:

A freed header's bit stays clear until something is installed at that exact address, so no leftover arena bytes can make it look live again. Space for the bitmap came from shrinking the per-arena code tail to 0x22000.

Regression test: the stale-enter case in test/test-uop-compiler.js. It forges a freed header with a current gen word and a matching head, then enters it:

After the fix. The same pinned run reaches batch 3000 without the crash. The capture shows the rotating 3D Deus Ex logo intro rendering in software.

2026-10-06: renderers, fullscreen and the mouse

The demo ships SoftDrv, D3DDrv, GlideDrv, OpenGlDrv (plus MetalDrv/SGLDrv). The installer and Default.ini select SoftDrv. lib/apps.js now edits DeusEx.ini at mount time (iniSet, applied by both hosts through lib/app-files.js applyIniSet): renderer keys and StartupFullscreen=True. To try another renderer, change deusExRenderer there; on the CLI an ad-hoc INI=<file> argument plus --vfs-include=<file> works too (UE1 Core parses INI=; there is no command-line renderer override, and a map name on the command line is ignored -- the demo always boots DX.dx).

Renderer status:

Mouse. WinDrv reads WM_MOUSEMOVE as an offset from the client centre and re-centres with SetCursorPos (windrv+0x1110af82). Pointer lock only engages under exclusive presentation, so with -windowed the page fed absolute positions into that relative protocol -- the erratic mouse. Fullscreen fixes it; a CLI relmousemove 40,30 arrives as exactly (+40,+30) from (320,240). Pointer-lock counts now reach the guest 1:1 (cc69ac8e) instead of being scaled by the picture stretch (~46% at 640 wide on a 1380-px page).

Keyboard menu route (browser, where the relative mouse makes clicks awkward): Esc after ~250 presents, then Down x3 and Enter reaches Training; use 0.1 s key holds -- 0.4 s auto-repeats and lands on Credits.

Evidence: scratch/runs/20261006T1115Z-deusex-renderer-bench (windowed), 20261006T1215Z-deusex-fullscreen-bench (fullscreen), 20261006T1100Z-deusex-d3d-cli and 20261006T1110Z-deusex-soft-cli (CLI routes).

Additional dated branch findings

These historical observations supplement the current facts above; they do not establish a new current-main runtime result.

2026-10-22: Training runtime blocked by incomplete support staging

Verdict: BLOCKED before menu; gameplay acceptance remains open. The bounded run used accepted main c26252c4a95babe583ade1d1c6e2da0c050be45c and the exact previously qualified shipping module, SHA-256 f644bbb26edc077239dac17b8b2cfda482aafd9eb44af4c4776d4fe153d5a1a7. No build, engine change, source-gate bypass, or main merge occurred.

Evidence is in scratch/runs/20261022-deus-ex-training-runtime/, especially run.log, first-run.log, source-mapping.json, assets.json, missing-native-support.json, input-log.jsonl, and RECOVERY.md. The original remote logs/exit files/timestamps/identity manifests are also preserved in evidence-artifacts.tar.gz (SHA-256 6f4928e32bb6421e6f24b709cea0c12dda292796dce8b4fc96063bd4dbb12cfd).

The registered app's current 112 paths were individually hashed locally and after remote staging: 233,696,463 bytes, including the original 253,952-byte EXE. EXE/Core/Entry match the three pins in test/test-deus-ex-demo.js; every declared registry file size and the prepared INI settings pass. This verifies the game's declared asset closure, but the staging omitted native system DLLs outside that list. Original files remained unchanged. The app's existing mount-time INI transformation selects SoftDrv and fullscreen.

At 01:54:16 UTC the first launch exited 1 before guest execution because src/api_table.json was absent from the sparse staging list. The unchanged accepted-main file was added, and all 192 selected source/support files and 112 assets were reverified. The single controlled retest at 01:54:47 UTC also exited 1, within that same second, during native DLL initialization. run.log warns that msvcrt.dll, comctl32.dll, shell32.dll, and ole32.dll were unavailable on disk and their imports fell back to WAT stubs. Render.dll initialization (module=0x11652000, DllMain 0x1167997a) then hit UNIMPLEMENTED API: _vsnwprintf at EIP 0x10bed970, trapped with unreachable, and startup rollback also failed while detaching D3DDrv. This is an observed failure with an incomplete support closure; it does not establish a regression in the accepted engine or the cause of the native code's formatting call.

Node was v24.18.1. launch.sh records the complete command: registered deus_ex_demo, exact pinned Wasm, explicit --no-threads, frozen control, 200,000-block batches and 25 ms per batch, with a 3,000-second wall limit. Both control snapshot requests failed to connect because startup had already exited. No key/mouse input reached the guest. There are no game frames to review, no intro/menu observation, no Training world/HUD, no before/after movement evidence, and no successful bounded live interval. Historical screenshots were not substituted. Windows parity, browser/FPS/audio quality, and overall game completion remain unqualified.

2026-10-31: original Training and native movement verified locally

Training and ordinary forward/backward movement passed. The registered 1.002f demo reached the readable UNATCO Training Facilities corridor, HUD, unlocked door, and Jaime Reyes transmission. I personally inspected the before/after captures: Up-arrow moved JC from the corridor to the door; Down-arrow moved him back, with the door and corridor visibly receding. No game assets, guest state, cheats, or engine sources were changed.

Evidence: scratch/runs/20261031T0416Z-deusex-training/:

Preparation began on shared main 59a615fc5. No runtime overlapped the live Diablo II browser: PID1493231 was absent before launch at04:21:55Z, and /tmp/wine-browser.lock was held for the full CLI/controller lifetime. The run quit normally at04:23:56Z, exit0, 1745 batches/120.254 seconds reported by the harness. No browser or test box was launched; ID/state/expiry N/A. Runtime/controller/watcher are absent, localhost18126 is closed, the lock is released, and cleanup recorded4.93GB free disk.

The retained module was SHA-256 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, byte-identical to Re-Volt's recorded locked after-build module at scratch/runs/20261031T0403Z-revolt-backbuffer/after.wasm. Its source identity is that run's saved dirty source-after snapshot/manifest, including concurrent font changes; it is not inferred from checkout HEAD. All files in that manifest matched this checkout during preflight. This task did not rebuild.

Support closure matters. The original112 registered paths and EXE/Core/Entry pins passed. The native graph contains19 DLLs and no missing native files when using the exact existing support fixtures. Static checks resolved all named imports over48 native edges and all62 native ordinal imports. The extra native system files include existing comctl32/msvcrt and these Explorer fixtures:

File Microsoft version resource SHA-256
shell32.dll 4.72.3612.1700 533c5809a9b6bdb236cce17a14c03ff251c4ca4d00790ea8e8034ee5d14d7f61
ole32.dll 4.71.2900 bb487eb6c5ade54d5ae0352f75cec84b4f1cbbb8d9fca5b3718d813b4a0c938d

shlwapi.dll5.00.2614.3500 is also required transitively by shell32; its exact hash and version are in the identity/version evidence. These are the already recorded Microsoft fixtures used by the repository's recovery notes, verified against PE version resources and prior hashes. An independent chain back to original Windows media was unavailable. They were supplied with an explicit --dlls list generated by the existing dependency walker, using their existing paths under test/binaries/explorer98/dlls/; nothing was copied or linked into the game tree or shared system directory. Every native DllMain returned1. The log contains live44100Hz stereo waveOut opening and no missing-on-disk, unimplemented-API, unhandled-exception, or critical-error messages. All131 before/after asset/native hash checks passed, and the pinned module stayed unchanged. Live waveOut opening does not establish audible playback quality.

The CLI used the registered app's existing SoftDrv/fullscreen INI transform, cooperative threads, frozen control on localhost18126, 200000 blocks/batch, 25ms/batch, repaint10 and a600-second guard. The launch arguments are recorded in launch.json; dll-args.txt contains the exact19-file list. After Escape, a relative mouse move hovered Training, followed by mouse-down/20 batches/ mouse-up/100 batches and another400 to the initial world capture. Up-arrow (VK38) was held for150 batches and released for5; Down-arrow (VK40) for100 and released for5. The input log records the earlier overlapping step request that was rejected; subsequent menu and movement commands were sequential.

The previously observed Thread1 NULL exit still appears, at prev_eip=0x10901a29, while the main engine renders, accepts input, loads Training and walks. Its cause is unestablished, and it did not prevent this acceptance. The earlier incomplete remote support failure is not evidence of a current source regression. The coordinator's unchanged long smoke was not repeated. Browser/default-worker gameplay, Liberty Island, sustained play, audio quality and game completion remain unverified by this CLI result. Next useful check is the same ordinary Training movement route on a separate idle test host with the complete recorded support closure; no new task or worker was launched here.

2026-10-31: default guestWorker browser Training and first-door traversal

Passed: original Training world/HUD, ordinary movement, door opening and traversal in Chrome's default guestWorker transport. Evidence is in scratch/runs/20261031T0447Z-deusex-browser/. I personally inspected the location text, corridor, health silhouette, compass, inventory belt, Jaime Reyes transmissions, closed unlocked door, open doorway and red-floored room beyond it. This supersedes the older browser-input limitation above for this specific retained baseline; it does not qualify current HEAD or completion.

The private local page used Chrome151.0.7922.108, a fixed1024×768 viewport, index.html?debug&no-log, default Threads checked and actual wine.guestWorker === true. Read-only telemetry identifies three lib/guest-worker.js instances, main slot0/thread1, main HWND65539, and increasing presentation counters. No cooperative fallback, clock/state patch, renderer switch, source repair, build or public deployment was used. The game's registered mount-time INI selected SoftDrv.SoftwareRenderDevice and fullscreen. The page's GPU selector read WebGL; that does not change the game's SoftDrv selection or establish a renderer throughput result.

The module remains SHA-256 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676. The browser server supplied the matching saved292-file ReVolt source snapshot, whose manifest hash is 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Every source byte was checked against that manifest before serving; requested runtime bytes are retained by hash in served-snapshots/ and mapped in snapshots.json. Checkout HEAD was observed separately and is not claimed as the compiled source. The optional /build-info.js request returned404 because it was absent from the saved source; launch and gameplay proceeded. Font substitutions used the separately retained prior browser response.

The registered app and mount-time transform were reused, with only the launch configuration's DLL URL list replaced by the exact19 existing native paths from the successful CLI recovery. The original112 registered inputs and all 131 asset/support hash checks passed before and after. Named and ordinal closure evidence is preserved; no original asset or support DLL was altered.

Native browser input is recorded with trusted key/button events in native-input.json and commands in commands.jsonl. Escape opened the visible menu. Initial short Down presses did not establish a usable selection, so the actual game cursor was moved onto the visible Training button, inspected in hover-training-settled.png, then pressed for180ms. Loading led to training-world-first.png, showing UNATCO Training Facilities and the HUD. Up-arrow1800ms approached the door (forward-settled.png); Down-arrow900ms retreated (retreat-settled.png). Up-arrow3200ms reached the closed door with its “Unlocked” indicator (door-before-interact.png). The shipped User.ini binds RightMouse to ParseRightClick: a native180ms right-button press opened it (door-after-right-click.png). Up-arrow2100ms then crossed into the red-floored room (beyond-first-door.png); the next Jaime transmission described the key on the desk. mouse-look-settled.png also shows that room, wall panel and plants. Later captures named look-back-turn* were attempts: Left is actually StrafeLeft in the shipped bindings, and those captures do not prove a view back through the doorway.

The controller ran04:47:38–04:51:16UTC, well inside the25-minute investigation bound. It acquired /tmp/wine-browser.lock after triage's GLTron CLI closed, checked processes before launching and held the lock through browser/server cleanup. Chrome1551364 exited0; controller1551290 and lock holder1551289 are absent, HTTP38133/CDP41333 are closed, the profile is removed and the board watcher is stopped. Cleanup and focused evidence/helper checks are retained. No Boat/test box was allocated; ID/state/expiry are N/A. Audio quality, sustained play, collecting the key, opening the next locked door, Liberty Island and Training/game completion remain unverified. The next narrow gameplay check is the visible desk key and next locked door, if requested; this single-task assignment ends here.

2026-10-31: desk NanoKey and next locked door traversed

Passed: collect the desk key, unlock the next door, open it and walk through. Evidence: scratch/runs/20261031T0606Z-deusex-key-door/. I personally inspected key-center-before.png showing the targeted NanoKey on the desk and key-after-pickup.png showing the empty desktop and “Trainer Door added to Nano Key Ring”. locked-door-target.png shows Locked, and ordinary right-click produces “It's locked” in locked-door-interaction.png. Slot0 equips the ring (key-ring-equipped.png); a180ms native left-button press produces “Your NanoKey Ring unlocked it” and Unlocked in door-after-key-use.png. A180ms right-button press opens the doorway (next-door-open.png), and Up1000ms crosses into the gray corridor (beyond-next-locked-door.png). doorway-behind.png, viewed from that corridor, shows the red-floored desk room through the doorway and the next Jaime transmission about goals/notes. This establishes passage beyond the locked door, not Training completion.

The unchanged original User.ini binds Z/C to TurnLeft/TurnRight, arrows to movement/strafe, PageDown/Delete to LookDown/LookUp, End to CenterView, RightMouse to ParseRightClick, 0 to ActivateBelt0 and LeftMouse to ParseLeftClick|Fire. Native C700ms revealed the desk after replaying the first door. Further native turns/movement/look and small mouse deltas targeted the visible key; RightMouse180ms collected it. Commands and74 trusted native key/button records are retained. Short menu/aim attempts are recorded as attempts; successful actions are identified by the inspected captures above.

The private Chrome151.0.7922.108 page used a1024×768 browser viewport, default Threads checked, actual guestWorker === true, three guest workers and registered mount-time SoftDrv/fullscreen INI. Module SHA-256 remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676; the exact292-file saved source manifest remains 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. The server supplied matching saved source bytes, not the separately observed checkout HEAD. All131 original/support pins,292 saved source hashes and105 served runtime snapshots passed, with the original19-DLL closure preserved. The launch registry differs only in its explicit existing native DLL URLs. No build, source repair, asset change, guest memory/state/clock patch, cheat, input bypass or public deployment occurred. No FPS or audio-quality claim is made. Resolution limitation: the preserved INI requests640×480 and the runtime console records canvas 684x570 -> 640x480. Thus1024×768 describes the browser capture, not the guest render resolution; this run did not meet the requested1024×768 guest-render qualification. The runtime was already closed when this distinction was checked; no second startup was run.

Preparation started06:06UTC; the browser/controller ran06:07:16–06:12:14UTC, within the30-minute assignment. The full-lifetime /tmp/wine-browser.lock and process/profile gate serialized execution after Forsaken. Chrome1639004 exited0; controller1638968, holder1638050, watcher1638049, tracked children and crashpads1639020/1639022 are absent. HTTP41963/CDP40803 are closed and the profile is removed; cleanup verified4.00GB free disk. No Boat allocated: ID/state/expiry N/A. RECOVERY.md, verification.json, gameplay-check.json and cleanup-verification.json record recovery and focused checks. The next bounded gameplay check, if assigned, is to follow the corridor into the next Training exercise with ordinary native input. This assignment stops here.

2026-10-31: next Training exercise — crowbar and crate breaking

Passed: follow the corridor beyond the NanoKey door, collect/equip the crowbar and break a wooden supply crate with native attacks. Evidence is in scratch/runs/20261031T0617Z-deusex-next-exercise/. This was a fresh launch, replaying the accepted first-door/key/unlock route rather than restoring old guest state. I personally inspected door-unlocked.png ("Your NanoKey Ring unlocked it"), next-door-open.png and beyond-locked-door.png, then followed the corridor into the next red-floored exercise room. next-exercise-instructions.png shows Jaime's request to pick up a weapon and break open the crates.

crowbar-before-pickup.png shows the targeted Crowbar on the striped bench. RightMouse180ms collected it: crowbar-collected.png shows its absence, "You found a Crowbar" and belt slot1. Native1 equipped it. crate-target-before.png shows the intact upper General Supply Crate within crowbar range. Native LeftMouse180ms, then LeftMouse1800ms broke that crate. crate-after-strikes.png shows the upper crate gone; crate-broken-result.png and exercise-completed-view.png show the lower wooden crate and adjacent metal crate still present, with an exposed item above the lower crate. exercise-response.png shows Jaime's follow-up lock-picking instructions. The exposed item was not collected or identified by interaction. This demonstrates the requested next exercise interaction; lock-picking, subsequent progression and full Training completion remain open.

The replay preserves module SHA-256 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676 and the matching saved292-source manifest SHA-256 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. All131 original/support pins,292 saved source files and105 retained runtime responses passed verification. The original19-native-DLL route and registered SoftDrv/fullscreen mount-time INI are unchanged; only the explicit DLL URLs are supplied as in the predecessor. Chrome151.0.7922.108 used default guestWorker === true with three workers. The browser captures are1024×768; the preserved guest canvas is640×480, confirmed by this run's resize log. No resolution change, build, engine/source repair, asset/state/clock patch, cheat, input bypass, FPS qualification or public deployment occurred. Commands and102 trusted native key/button records are retained. Input timings are scene-dependent: inspect the fresh view rather than blindly replaying turn durations.

The browser/controller ran06:17:07–06:24:02UTC under the full-lifetime /tmp/wine-browser.lock after the live-process/profile gate passed. Chrome1649556 exited0; controller1649529, holder1649528, all tracked children and crashpads1649572/1649574 are absent. HTTP35159/CDP35299 are closed, the profile is removed, all held keys were released and no watcher was started. Cleanup verified3.58GiB free disk. No Boat allocated: ID/state/expiry N/A. RECOVERY.md, verification.json, gameplay-check.json and cleanup-verification.json preserve recovery and focused checks. Gameplay stopped on demonstrated progression; no successor was launched.

2026-10-31: native lock-picking and traversal into the next room

Passed: collect the exposed Lockpick, pick the exercise door, open it and walk through into the next room. Fresh replay evidence is in scratch/runs/20261031T0648Z-deusex-lockpick/. I personally inspected the original visible Training menu, first-door opening, desk NanoKey pickup ("Trainer Door added to Nano Key Ring"), slot0/LeftMouse unlock receipt and RightMouse opening, then replayed the corridor and crowbar/crate exercise. Recorded turn durations varied enough to require visible navigation corrections; the command log preserves missed pickup/attack attempts as well as successes.

object-target-before.png identifies the exposed object as Lockpick above the remaining lower wooden crate. Native RightMouse180ms collected it: lockpick-receipt.png reads "You found a Lockpick" and shows belt slot2. Original F1 inventory, then native selection of the item, produced lockpick-description.png: Lockpick, Uses: 1. No item or slot was assumed. locked-door-before.png shows Locked, Lock Str: 5%, 1 pick, Door Str: INF. Native2 equipped the observed belt slot; native LeftMouse180ms began use. lockpick-in-progress.png reads Picking the lock..., with strength reduced to2%. lockpick-door-after-use.png shows Unlocked and slot2 empty. These world/inventory changes establish success beyond elapsed time.

Native RightMouse180ms opened the picked door (picked-door-open.png), and ArrowUp1200ms crossed it. corridor-south-view.png looks back through the doorway into the red crate room. Following the blue corridor east reached a distinct room with a wall map, desk item and observation window, personally inspected in next-scene-after-lockpick.png and final-next-room-view.png. That desk item was not identified or collected; no next exercise interaction or Training completion is claimed. F2 displayed no listed goals, which is not completion evidence.

Module SHA-256 remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, with matching292-file source manifest SHA-256 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. The retained ReVolt0403 source/build pair supplied the runtime, independently of observed checkout HEAD939ff7db0. Before/after checks passed131 original and support files,292 saved source files and105 served responses; the19-native-DLL closure and mount-time SoftDrv/fullscreen INI remain unchanged. Default guestWorker/three workers, guest640×480 and browser1024×768 were observed. No source/build change, asset/INI/state/clock patch, cheat, input bypass, resolution/FPS/audio diversion or deployment occurred. gameplay-check.json records212 trusted native events, eight interaction deltas and32 personally inspected frames.

Preparation began06:47UTC; the browser lock was acquired06:49:19UTC and the browser closed07:06:13UTC. Chrome1685329 exited0. Controller1685276, holder1683955, wrapper1683945, all tracked descendants/crashpads and owned watcher1684069 are absent; HTTP36975/CDP41383 are closed, profile removed and no keys held. The outer tool session returned143 despite browser exit0; this is retained in cleanup evidence without attributing a cause. Cleanup verified2.42GiB free disk. No Boat allocated: ID/state/expiry N/A. Recovery, native commands, source/build identities and focused checks are retained in the run directory. Stopped on the demonstrated bounded outcome; no successor.

2026-10-31: DataCube code, security keypad and next multitool room

Passed: identify/read the desk DataCube, enter its code at the security keypad, and traverse the opened door into the next exercise. Evidence: scratch/runs/20261031T0713Z-deusex-lockpick/. This fresh original Training launch replayed the visible menu, NanoKey pickup/unlock, crowbar/crate and native lockpick route. Receipts and world changes were personally inspected; turn durations required fresh corrections. The first crate attacks and first lockpick use missed; successful later actions are preserved separately.

desk-item-identified.png labels the item DataCube and shows Jaime's instructions: right-click the cube, read its contents, type the code into the keypad, and activate the keypad with the right mouse button. desk-item-close-view.png says the door code is stored in that cube. Native RightMouse180ms produced datacube-contents.png: Note Received — Check DataVault For Details, with Jaime's message giving code 0012. This item is read in place; no belt slot or inventory pickup was assumed.

keypad-label.png identifies Security Keypad, Bypass Str: 20%, 2 tools. Native RightMouse180ms opened its interface. Four ordinary keyboard presses, 0/0/1/2, populated the masked display (keypad-first-digit.png, keypad-second-digit.png, keypad-third-digit.png); keypad-two.png shows green GRANTED. The interface then closed automatically. The later keypad-code-result.png notification “It's secure” does not negate the captured GRANTED or subsequent open door. code-door-after-granted.png shows the automatically opened passage and a different closed door beyond. Native strafe corrections and forward movement crossed the open passage. next-exercise-arrival.png, next-exercise-instructions.png and next-exercise-room.png show the next room and Jaime's new transmission about disposable multitools, finite resources and hacking devices. code-door-behind.png looks back through the open door at the DataCube desk. These are distinct visible progression states. The multitool exercise was not performed, and Training completion is not claimed.

Runtime SHA-256 remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676; the exact292-file saved-source manifest remains 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. All131 original/support files and292 saved source files passed before/after verification. The105 served snapshots exactly match the predecessor's URL/hash closure; current-host/old-module mixing was not used. Original 19-DLL closure, mount-time SoftDrv/fullscreen INI, default guestWorker with three workers, guest640×480 and browser1024×768 are preserved. Observed checkout HEADdf174b3cf is metadata, not the runtime build identity. No source repair/build, asset/INI/state/clock patch, cheat, input bypass, resolution/FPS/ audio diversion or public deployment occurred. Focused checks passed224 trusted native events, ten interaction deltas, four keypad digits,196 PNG signatures and35 personally inspected captures.

Preparation began07:12UTC; the browser lock was acquired07:13:06UTC and cleanup completed07:24:00UTC. Chrome1713772 exited0; controller1713762, holder1713750, all tracked children/crashpads, own quit wrapper1725079 and watcher1713955 are absent. HTTP42491/CDP32957 are closed, profile removed, no keys held, and2.26GiB remained free. Full browser lifetime and shutdown were under /tmp/wine-browser.lock; the post-lock live-process gate passed. The quit helper waited30 seconds for an acknowledgement that the controller does not emit on quit, then exited normally; this was not a runtime failure. RECOVERY.md, identity, gameplay, verification and cleanup records are retained. No Boat allocated: ID/state/expiry N/A. The next work, if assigned, is the visible multitool exercise; no successor was launched.

2026-10-31: Multitool bypass and next cadet exercise

Passed: identify and collect the multitools, bypass the intended keypad, traverse into the next exercise, and interact with its unconscious cadet. Fresh evidence: scratch/runs/20261031T0736Z-deusex-multitool/. A new original Training launch replayed the visible menu, NanoKey, crowbar/crate, lockpick, DataCube0012 and keypad route. Labels and world changes guided corrections; missed menu presses, attacks, aiming, pickups and door activations remain in the evidence. The fresh keypad-granted.png filename describes the intended checkpoint but captures four masked dots, not GRANTED. The subsequent open door and crossing into the multitool room prove code progression; the prior 0713Z run retains the actual green GRANTED capture.

multitool-room-view.png and multitool-target-instructions.png show Jaime instructing use of the table's disposable multitools on the keypad ahead, explaining finite resources and reduced requirements at higher skill levels. multitool-label.png identifies the yellow table object as Multitool. Three individually targeted native right-click pickups remove the three devices; multitool-belt-one.png, multitool-belt-two.png and multitool-stack-three.png show the receipt and growing stack. F1, then an ordinary mouse selection, produced multitool-description-confirmed.png: Multitool, Count:3, Uses:3, inventory top row/third column after the two-column crowbar, and belt slot2. Its description identifies an electronics tool for code locks and other security systems. The used-up lockpick's earlier slot was observed empty before collecting these tools.

hack-keypad-label-confirmed.png identifies the actual target as Security Keypad, Bypass Str:5%, 1 tool, distinct from the preceding code keypad's 20%/2 tools. Native 2180ms, wait1000ms, then LeftMouse180ms starts the hack. native-hacking-progress.png shows Bypassing the device..., a red scan and strength2%. After4000ms, native-hack-result.png shows You bypassed the device, Security Keypad: Bypassed, and belt Uses:2. The door has opened automatically; hacked-door-open.png shows its blue corridor.

Native movement crosses the passage. next-door-before.png shows a UNATCO troop collecting equipment and the belt cleared except for the NanoKey ring. Opening the next door and crossing produces Jaime's new cadet transmission (next-exercise-world.png) and a different observation room with a body. cadet-label.png labels that body Unconscious. Native RightMouse180ms produces cadet-interaction-result.png: Medical Room added to Nano Key Ring, You found: and Received: NANO. This proves next-exercise interaction as well as traversal. Body carrying, medical-room unlocking, healing and Training completion are not demonstrated. No unresolved runtime/control dependency blocked the assigned multitool outcome.

Runtime SHA-256 remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, paired with exact292-file saved-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after checks pass131 original/support files,292 saved source files and105 served snapshots matching the prior URL/hash closure exactly. The 19-native-DLL mount closure, original SoftDrv/fullscreen configuration, default guestWorker/three workers, guest640×480 and browser1024×768 remain preserved. Checkout HEAD501f52cf1 at startup is metadata, not the runtime build identity. No source/build change, asset/INI/state/clock patch, cheat, input bypass, FPS/resolution/audio diversion or deployment occurred. Focused checks pass366 trusted native events,13 interaction deltas, four code digits, 291 PNG signatures/dimensions and40 personally inspected captures.

Preparation began07:36UTC; the browser lock was acquired07:36:27UTC and the browser closed08:00:17UTC. Chrome1738864 exited0; controller1738854, holder1738845, all tracked descendants/crashpads and own watcher1739217 are absent (19 recorded PIDs including watcher wrapper). HTTP33489/CDP40745 are closed, profile removed and no keys held. The full browser/controller lifetime used /tmp/wine-browser.lock, with a post-lock foreign-runtime gate; cleanup left2.07GiB free. Helpers, commands, recovery, identities and focused evidence checks are retained in the fresh run; prior evidence was read-only. No Boat allocated: ID/state/expiry N/A. Next work, if assigned, is the Medical Room NanoKey/cadet exercise. No successor was launched.

2026-10-31: Medical Room attempt stopped before replay completion

Medical Room progression was not demonstrated. The accepted multitool/ cadet result above remains the furthest proved state. Fresh evidence is scratch/runs/20261031T0817Z-deusex-medical/, with a separate retry/. Both attempts launched original Training afresh; neither restored prior state.

The first replay confirmed the Trainer Door NanoKey receipt, collected the crowbar, broke crates to expose a lockpick, and used the lockpick on the labelled5%/1pick door. Personally inspected lock-door-five-percent.png, native-picking-progress.png and native-picked-result.png show5%, then2%, then Unlocked, with slot2 emptied. Native LeftMouse input precedes the progress; the final transition completes while waiting. datacube-contents.png shows code 0012, and code-keypad-identified.png labels the next target Security Keypad, Bypass Str:20%, 2 tools. No code digits, multitool pickup, cadet search or Medical Room action occurred in this fresh replay. Several intended-checkpoint filenames capture misses; the inspected labels and actual receipts, not filenames, determine these claims.

The first precise failure was in the local harness: controller1782534 and lock holder1782532 disappeared after the last capture at08:32:29UTC, leaving own Chrome1782571 orphaned. The cause is unproved; no guest/runtime defect is inferred. Consequently this attempt does not qualify as a continuously locked browser lifetime. A recovery helper reacquired the lock, checked the exact own profile, captured the unchanged keypad, and closed only that browser at08:34:03UTC. A second fresh launch used a detached lock/controller lifetime from08:34:58UTC. It confirmed the NanoKey route and reached the crowbar/crate area, then stopped at08:43:36UTC to release the shared lane. Its final capture shows the crowbar equipped against the wall beside the crate area; the lockpick had not been collected. No Medical Room control dependency was isolated, because the fresh route never reached that exercise.

Read-only examination of original system/deusexcontext.u, included in the 131-file hash closure, identifies the cadet instruction at byte2908993: search the unconscious cadet for the Medical Room key, then pick up his body and place it on the medical table so an aide can revive him after the exercise. A separate line at byte2905684 acknowledges Private Winslow and directs the player onward. These are static original-text findings, not observed trigger order or fresh gameplay proof. original-training-instructions.json retains offsets and text. Body pickup, key use on the medical door, table placement, any medical-device interaction and traversal onward remain untested.

Both attempts used retained runtime SHA-256 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676 with exact292-file saved-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after checks pass all131 original/support files and292 saved source files. Each105-response URL/hash closure matches the accepted predecessor; 19-DLL mount lists match, with only those DLL URLs changed from the saved registry. Saved sources were served with their matched module, never current host sources. Original SoftDrv/default guestWorker/three workers, guest640×480 and browser1024×768 were retained. There were no source/build, asset/INI/state/clock changes, cheats or input bypasses, FPS/audio diversion, or public deployment. Focused checks pass326+118 trusted native input events, empty held-input sets and90+29 PNG signatures/dimensions. Historical runs were read-only; duplicate served bytes refer to verified predecessor copies.

Own browser/controller/holder and tracked descendants are absent; both HTTP ports35169/45489 and CDP ports36395/39095 are closed. Profiles are removed, retry Chrome exited0, and cleanup left more than2GB free. The first profile's small post-close files were removed after another live-profile check. focused-checks.json, inspected-proof.json, recovery helpers and commands preserve the evidence and limitations. No Boat was allocated: ID/state/expiry N/A. No worker or successor was launched. The next dependency is a durable, continuously locked fresh replay reaching the already accepted cadet state, then native key/body/table interaction with personally inspected final proof.

2026-10-31 0849Z — MedicalRoom onward to light exercise; native save restored

Run scratch/runs/20261031T0849Z-deusex-medical-durable/ extends the accepted e8e86b767 cadet NanoKey receipt with personally inspected native gameplay. A fresh original Training replay reached the cadet, searched him for the Medical Room key (96-medical-key-receipt.png) and picked up his body (97-native-body-carry.png: held cadet, cleared floor, “You found a body”). After temporarily putting him down, the equipped keyring's native left-click unlocked the signed MedicalRoom door: 119-medical-key-receipt.png shows “Your NanoKey Ring unlocked it” and “Unlocked.” The cadet was retrieved and carried through the open door (124, 130) to the medical table (131).

Native right-click toward the table cleared the held cadet and started an incoming transmission (132-table-placement.png). No retained frame shows a body resting on the table, and no readable Winslow acknowledgement or healing is claimed. Table acceptance is inferred from that trigger, the opened onward route, and the next distinct exercise. 144–146 personally show Jaime's instruction to use the light augmentation with F12 and find the exit from a dark room; 147-light-exercise.png shows “Light activated.” The next gameplay is that light/dark-room exit exercise.

At the first new body-carrying progress, F6 proved to be an empty augmentation slot. The ordinary Escape → Save Game → Enter UI then wrote actual native save bytes. first-native-checkpoint/ preserves the completed first export: c:\save\save0001\saveinfo.dxs (34,576 bytes) and c:\save\save0001\00_training.dxs (3,227,383 bytes), with guest paths and SHA-256 in its manifest. light-native-checkpoint/ separately preserves the later save count 2 light-room checkpoint (34,577 and 3,229,336 bytes). No save bytes were fabricated or edited. Same-application native Load Game restored it, then a page reload created a fresh native application in the same continuously locked browser. Only the two exact exported save files were added through the normal file loader. 163-fresh-save-list.png recognizes save count 2; native Enter loads it; 165-fresh-load-result.png restores the lit room; 167-fresh-native-light-response.png shows native F12 “Light deactivated.” Early teal/loading frames were transient, not guest failures. RECOVERY.md, make-restore.py (prepares a fresh bounded run without launching), fresh-load.cjs, both checkpoint manifests and inspected-proof.json retain recovery steps, hashes, observations and limits. The earlier cadet checkpoint's fresh load was not separately tested; the later checkpoint's was.

Exact retained WASM SHA-256 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676 and 292-file saved-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379 were used throughout, never current host sources with an old module. Before/after 292-source and 131-original/support hashes pass; all 105 served source URL/hash pairs and 19 DLL paths match the accepted predecessor. Duplicate response bytes reference verified 0736Z canonical snapshots. The restore adds only two authentic user-save data responses to the 97 original native data files. SoftDrv/default guestWorker/three workers, guest 640×480/browser 1024×768 remain the original configuration. There were no production-source, asset/INI/clock, guest-memory or executable changes, cheats, input bypasses, FPS/audio diversion or public deployment.

The detached setsid wrapper, controller and independent cleanup monitor kept fd9 on /tmp/wine-browser.lock across the entire runtime; proof at 09:25Z shows all three alive after 35min. This resolves the prior 15min harness-lifetime failure without treating it as guest behavior. The process/profile orphan gate passed before launch. Final checks pass 598 trusted native input events, empty held inputs, 167 valid 1024×768 PNGs, both save hashes and fresh-restore registry scope. All 19 tracked owned PIDs are absent, HTTP 34127/CDP 38929 closed, profile removed, browser exit 0, wrapper lifetime exit 0 and monitor ended. The lock was verified available and the lane explicitly released at 09:30Z; free disk stayed above 2GB. No Boat: ID/state/expiry N/A. No worker or successor was launched. focused-checks.json and check.py retain the exact checks.

2026-10-31 0940Z — Native checkpoint to dark-room exit and repair-bot briefing

Passed: restore the accepted light checkpoint, exit the dark storage room, and reach the next distinct exercise. New evidence: scratch/runs/20261031T0940Z-deusex-light-exit/. Only the two unedited 5781cc8cf light checkpoint files were added through ordinary app.files/vfsPaths; native Load Game recognized save count 2 and restored the world. No Training replay occurred. 12-native-light-response.png and 13-room-view.png show native F12 deactivation/activation receipts.

The original inspected controls use arrows for movement/strafe, Z/C for turning, Space for jumping and right mouse for interaction. Navigation included backtracking and missed crate routes. 40-north-perimeter-end.png labels the final door Unlocked; native RightMouse180ms opens it in 41. 42 crosses into a distinct red-lit passage and grated ramp. Personally inspected 44-next-instructions.png shows Jaime explaining that repair bots can recharge bioelectric energy spent crossing the dark area. The later 57-native-quicksave-result.png also contains his readable security-camera and security-grid alarm briefing. Repair-bot use/recharge, camera interaction and Training completion are not demonstrated. No previous-table healing or resting-body claim is added.

Save Game could not be activated while the transmissions were active: keyboard navigation skipped it, mouse clicks did nothing, and the inspected keypad-plus QuickSave binding left exported hashes unchanged. These are retained failed save attempts, not new checkpoints. After speech ended, the same normal mouse control opened Save Game (61); Enter180ms and a six-second wait wrote actual new bytes. The native save replaced slot 1 in this new guest VFS; frozen predecessor files remain unchanged. repair-native-checkpoint/ preserves immutable indexed copies and a manifest:

A fresh native application, created by page reload in the same continuously locked browser, mounted only those two new save files. 66-fresh-load-state.png recognizes save count 3 and the ramp preview; native Enter loads it; 67-fresh-repair-load-result.png restores the ramp boundary; 69 shows native F12 Light deactivated. RECOVERY.md, make-restore.py, fresh-load.cjs, inspected-proof.json and manifests retain recovery and the exact limits. The next gameplay boundary is the repair-bot/camera exercise beyond the ramp.

Runtime remains exact retained WASM 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676 with 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after checks pass 292 sources and 131 original/support files; all 105 served source URL/hash pairs match verified canonical snapshots, and the original 19 DLL paths match. There is no retained-module/current-host mix. SoftDrv/default guestWorker/three workers, guest640×480/browser1024×768 remain unchanged. No production-source, asset/INI/clock, save-byte or guest-memory edits, cheats, input bypasses, FPS/audio diversion or public deployment. Focused checks pass 196 trusted events, empty held inputs and 69 valid PNGs.

One detached browser lifetime ran09:39:43–09:56:23UTC, within the18min bound; cleanup ended09:56:29. Wrapper1870426, controller1870429 and monitor1870428 held the same fd9 flock throughout, with initial/final proofs. The live process gate and bounded profile matcher handle collapsed Chrome cmdlines. Crashpads used a global configuration directory, so they were additionally identified through the own Chrome handler-PID reference and matching launch start identities, then stopped under the full flock before browser shutdown. All20 owned runtime PID/start identities are absent, HTTP/CDP ports closed, profile removed, browser/lifetime exit0 and monitor ended; full flock release was verified before lane handoff. Free disk stayed above2GB. No Boat: ID/state/expiry N/A. No worker/subagent/successor was launched. check.py, focused-checks.json and cleanup-audit.json retain the checks.

2026-10-31 1007Z — RepairBot recharge and fresh native checkpoint validation

Passed: native RepairBot recharge from 0% to 75%, actual native save export, and fresh-application native Load Game validation. Evidence is in scratch/runs/20261031T1007Z-deusex-repair-camera/. The accepted 173d88a99 repair checkpoint was mounted only as its two immutable, unedited files through ordinary app.files/vfsPaths. Native Load Game recognized count3 (18), restored the ramp (20), and F12 produced “Light deactivated” (22). There was no Training replay.

The Repair Bot is in the red-lit alcove beside the ramp. Navigation included backtracking, missed alignment and travel beneath the raised walkway. The inspected original controls also bind PageDown to LookDown and F to CenterView. 57-target-bot.png labels the bot; native right mouse opens its actual interface in 63-bot-ui.png, showing Bio Energy 0% and Ready. A mouse click on Recharge did not change that state (65). Native Enter then produced Bio Energy 75% and Recharging: 58 seconds (67-energy-result.png). This is an inspected recharge result, beyond approach or instructions.

Escape → Save Game → Enter wrote actual new bytes, overwriting slot1 only in this new guest VFS. recharge-native-checkpoint/ preserves immutable indexed copies, guest paths and a manifest:

A page reload created a fresh native application in the same locked browser, mounting only these two exported files. 126-fresh-list.png recognizes save count4 and the bot preview. Native Enter restores the bot (128); native right mouse opens its interface with 75% Bio Energy (130). RECOVERY.md, make-restore.py, fresh-load.cjs, inspected-proof.json and the checkpoint manifest preserve recovery and the exact boundary.

The camera corridor beyond the ramp was approached (163-camera-area.png), including a native jump attempt. Subsequent frames show movement below the raised walkway, with a green camera visible (168). There is no inspected camera activation/detection, alarm receipt, camera-exercise completion or next distinct instruction/area. The first unresolved gameplay dependency is navigating the raised camera route and obtaining an actual camera response or onward trigger. No Training completion, healing or resting-body claim is added. Transient teal captures recovered without a restart. Console evidence includes caught native C++ throws during menus and blocked optional /build-info.js requests; no page error was recorded.

Exact retained WASM remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, with 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after checks pass all292 saved sources and131 original/support files; 105 served source URL/hash pairs match the verified canonical closure, and the19 original DLL paths match. SoftDrv/default guestWorker/three workers, guest640×480/browser1024×768 remain unchanged. No production, asset/INI/clock, save-byte or memory edits, cheats, input bypass, FPS/audio diversion or public deployment occurred. Frozen predecessor evidence stayed read-only.

Pirates explicitly released the lane at10:11:20UTC. This single detached wrapper/controller/monitor lifetime acquired the full browser flock and passed the actual live-process gate before launch at10:12:16; browser exit was 10:26:57, cleanup10:27:03, within the18/19min bounds. Initial/final fd9 proofs cover all three holders. Both global-database crashpads were identified through own profile Chrome child handler references; the secondary started one tick after the primary. Their exact PID/start identities were recorded and stopped under the full flock before browser closure. No foreign process was signalled. All21 owned runtime PID/start identities are absent; HTTP34857/CDP36113 closed, profile removed, browser/lifetime exit0, monitor ended and full flock release verified before explicit lane handoff. Focused checks pass266 trusted events, empty held inputs and106 PNGs. Free disk stayed above2GB. No Boat: ID/state/expiry N/A. No worker/subagent/successor was launched.

2026-10-31 1034Z — Raised-route geometry, native recovery stairs, unresolved camera boundary

Camera detection/alarm and the next distinct instruction/area remain unproven. Evidence: scratch/runs/20261031T1034Z-deusex-camera-route/. Only the accepted f13fa567a recharge checkpoint's two immutable, unedited files were mounted through ordinary app.files/vfsPaths. Native Load Game recognized count4 (33-save-list.png) and restored beside the bot (36). No Training replay or repeat recharge validation occurred.

Short movements and inspected PageDown/F views established the grated ramp, upper bend, overhead pipe and separate raised platforms. The original user.ini also binds X= Duck and Delete= LookUp; both produced native view changes. X+Up passed the first pipe (128, 133, 136). The first concrete failure is the first inter-platform gap: 142-near-edge.png shows the near platform and the next platform's supports; ordinary Up500ms (143), without a jump, leads to the lower-floor view (145). The subsequent jump does not establish a stable landing on the next platform; 152 explicitly shows its underside. This is a navigation failure, not an established emulator geometry/input defect.

The lower floor has a native recovery stairway: 187-rungs-near.png shows actual steps, 191 shows ascent and 197 shows the upper landing. A further side fall and stair recovery are retained. Later centered pipe/gap attempts (246–270) end with grating below the view and another pipe ahead, but do not establish which gap was crossed or an onward trigger. The camera remains green/visible; visibility alone is not camera interaction success. There is no alarm receipt, Training completion or healing claim. A colored HUD bar was initially mistaken for leg injury during live narration; injury is unproven.

Native Save Game (305) then Enter (306) wrote changed bytes. Immutable route-native-checkpoint/ exports contain:

Menu navigation accidentally opened exit/new-game setup dialogs; they were backed out without starting a new world. A page reload mounted only the two new exported files in a fresh native application. Startup reached the title menu (317) before the fixed deadline. Fresh native Load Game validation and the new save count are not established. The exported route checkpoint is experimental; the predecessor's count4 remains the validated recovery source. RECOVERY.md, manifests, inspected-proof.json and check.py preserve these limits. No second browser attempt was made.

Exact retained module remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, with 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after closure passes292 sources and131 original/support files; 105 served source URL/hash pairs and19 DLL paths match the accepted closure. SoftDrv/default guestWorker/three workers, guest640×480/browser1024×768 remain unchanged. Focused checks pass288 trusted events, empty held inputs and96 PNGs. No production, assets/INI/clock, save-byte or memory edits, cheats, input bypass, FPS/audio diversion or public deployment occurred. Frozen evidence stayed read-only; disk remained above2GB.

One detached wrapper/controller/monitor lifetime acquired the full browser flock at10:35:17UTC after the previous explicit release and a clear actual Chrome/crashpad process gate. All three fd9 holders have initial/final proofs. The18min command deadline ended at10:53:17; browser close finished10:53:20, monitor10:53:26, within the19min cleanup bound. Both global-config crashpad PID/start identities were recorded through own profile child references and matching launch identities. All20 unique runtime identities and3 remaining owned helper/watcher identities are absent, HTTP36805/CDP43479 closed, profile removed, held inputs empty, browser/lifetime exit0, and actual full flock release verified before explicit handoff at10:54:06. No foreign process was signalled. No Boat: ID/state/expiry N/A; no subagent, worker or successor.

2026-10-31 1103Z — Route save count5 validated; native controls respond, gap still unresolved

The immutable route checkpoint now passes fresh native Load Game validation. Stable gap landing and camera detection/alarm/onward instruction remain unproven. Evidence: scratch/runs/20261031T1103Z-deusex-gap-jump/. No production change or emulator defect finding resulted.

Static prep preceded browser use. Original system/user.ini binds Space=Jump, X=Duck, Up=MoveForward, Shift=Walking and CapsLock=ToggleWalk, with bAlwaysRun=True. Jump aliases Jump | Axis aUp Speed=+300.0; Duck aliases Button bDuck | Axis aUp Speed=-300.0. The predecessor's seven Space-down events all occurred with X released. Its first142→145 fall used Up500ms without Space; a stuck browser Duck key does not explain the later attempts. Trusted browser events establish delivery to the page, not successful guest physics or landing. No applicable existing route-map parser was found in tools/lib; no map, INI or memory was changed.

Only the predecessor's two authentic immutable route-native-checkpoint/ files were mounted through ordinary app.files/vfsPaths. Their sizes and hashes remain exactly those recorded in the1034Z section. A fresh application listed Save Count5, pipe-corridor preview and play time00:53:38 (23-load-dialog.png). Native Enter restored the corridor (26) matching the predecessor's270 view. Native LookDown and stationary captures inspected the near rim, grating and lower floor (29, 31, 33). This validates the existing export; no new save was written. The older count4 recharge save remains an unchanged validated fallback. No Training or recharge replay ran.

Space alone120ms produced a brief upward viewpoint excursion (35→36) and returned after1000ms (38). On a backed-away platform position, X350ms lowered the viewpoint (80→81); native release plus1000ms returned the standing view (83). Space150ms again raised the view (84). These are visible native control responses; they do not establish jump height/range or rule out a gameplay/timing constraint.

Native comparison Inspected result
Initial Space+Up450ms, stop and inspect; then1200ms from that changed position No landing proved; longer attempt reaches below the grating (45, 47).
Count5 reload, back away, inspect grating, short approach, Space+Up850ms Still on the near side after1500ms (75, 77); no next-platform footing.
Explicit Duck press/release, inspected edge, Space+Up120ms tap then Up900ms Intermediate grating view (92) becomes an underside view after1800ms (94): actual fall.
Exact count5 reload, uninterrupted Space+Up1200ms After2000ms, grating overhead (108); LookDown confirms lower floor (110).

The concrete remaining dependency is a grounded launch position/trajectory that reaches the next raised grate and remains there after release and a stationary wait. The inspected inputs above did not achieve it. No camera detection, alarm receipt or next instruction was observed; green camera visibility proves none of these. No new checkpoint was created to preserve a failed landing. inspected-proof.json, controls-analysis.json, native input and command logs retain the comparisons and limits.

The matched module remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after292 sources,131 original/support files,105 served source URL/hash pairs and19 DLL mounts are checked. Original SoftDrv/default guestWorker, three workers, guest640×480/browser1024×768 were retained. No asset/INI/clock, save-byte or memory edits, cheats, input bypass, FPS/audio diversion or public deployment occurred; frozen evidence and protected laptop assets stayed untouched.

Focused checks pass110 trusted events with empty held inputs, seven current Space presses with Duck released,45 PNGs, unchanged two-file save identity, and run-local evidence containment/completeness. Documentation diffcheck passes; no full build is needed for this documentation-only change.

Pirates'11:08:52 explicit FULLRELEASE preceded the actual process gate and full-lifetime flock acquisition at11:09:53. Wrapper1960914, controller1960917 and monitor1960916 each retained fd9; browser1960934 and both global crashpads 1960950/1960952 were tracked by PID/start. Native commands stopped early; browser cleanup finished11:18:09 and monitor11:18:15. All19 tracked runtime/ helper identities were absent, HTTP39795/CDP32987 closed, profile removed, held inputs empty and actual full flock release verified before11:18:40 handoff. Later documentation-watcher cleanup is recorded separately. Disk stayed above2GB. No foreign kills, Boat (ID/state/expiry N/A), subagents, additional guest launch or successor.

2026-10-31 11:34Z — first gap resolved with grounded running launch

Run scratch/runs/20261031T1134Z-deusex-grounded/ resolves the first raised grate-to-grate jump. This is actual stable footing, not camera visibility. The remaining pipe, farther gap and next instruction were not tested. No new save was created; the two immutable1034Z count5 files remain recovery.

Read-only analysis preceded the single native session. Scratch decode.py parses original UE1 v68 package tables, state headers and tagged properties; geometry.py follows compiled Model36 BSP references through its exact object end. Bounds, names, object references, compact-index cases and truncation rejection are checked. Original map and saved Model36 both have 16184 points,9376 nodes and5300 surfaces. Selected corridor faces place the near platform end at Y=-274.0007 and next platform beginning at Y=-429.9999: a 156-unit gap. The next grate interior is Z=-36. The lower floor is Z=-192; the pipe ahead spans approximately Y=-630..-666, Z=16..52. The farther gap is about178 units, Y=-830..-1008. These measurements describe compiled faces, not a full collision simulator; some original faces are nonplanar, so no global zero-residual claim is made.

Saved JCDentonMale0 is PlayerWalking/Physics1 at (6953.8433,-272.4012,16.9320), yaw114962, radius20 and standing height43: its center starts only1.60 units from the near edge. Structurally decoded class defaults give JumpZ300, GroundSpeed320 and Human AccelRate1000. Original DeusExPlayer.DoJump requires Walking and sets vertical velocity before switching to Falling; it supplies no forward velocity. This made a running launch from a backed-away grounded position a distinct test, rather than another guessed simultaneous Space/Up duration from the saved edge. It does not retrospectively assign every earlier failed comparison a cause.

A narrow scratch observer established state mapping before movement. It uses original Core exports' GObjObjects/FName arrays, enumerates only that declared object array, and requires unique JCDentonMale0/class identity. Original native disassembly supplies Physics+0x30, Location+0x114, Rotation+0x120, Velocity+0x144, Acceleration+0x150 and cylinder+0x1d4/+0x1d8. The initial values independently match the immutable save's exact position, rotation, Physics1, radius20 and height43, with zero velocity. Reads run in the owning slot0/thread1 worker's timer callbacks after JS/Wasm returns; there are no guest writes or guest gameplay calls. Wrong position, physics, array capacity, duplicate actor and radius fixtures are rejected. An initial FName-count ceiling rejected the real74583-name array; the recorded rejection was resolved with a bounded200000 ceiling before gameplay. The initial URL-only worker selector also refused the three-worker ambiguity; explicit slot/thread identification resolved it.

Native Load Game showed count5/time00:53:38 (12-count5-dialog.png) and restored the original corridor (15-restored-world.png). Only the two immutable files were mounted through app.files/vfsPaths. Native Back moved to a stationary grounded run-up at Y=-164.9484. Up then accelerated along the original grate. The Space-down gate required Physics1, correct corridor X and measured forward speed; standing height43 was observed at launch.

Observed boundary Native state / inspected evidence
Grounded launch (22-grounded-launch.json) (6953.1431,-245.6348,16.9320), Physics1, velocity(8.3388,-319.8913,0), height43.
Airborne (helper-jump-observer-final.json) Physics2, first observed upward velocity226.85; peak observed center Z63.9063. Up remained held across the gap.
Release and stationary wait (25–28) (6959.7808,-500.5729,15.8000), Physics1, velocity/acceleration zero, height43.
Inspected feet (26-landing-feet.png, 27-landing-feet-still.png) Next platform's grate fills the view beneath the player and remains unchanged after an additional2-second wait. 28 shows the pipe ahead.

All38 browser events were trusted and balanced, with exactly one Space-down while Up was held and Duck released. The observer confirms trajectory and stable contact; screenshots independently confirm raised grate underfoot. No emulator defect is established.

Original map and save mark all four SecurityCamera actors bNoAlarm=True, including UN_obst_jump. Original bundled SecurityCamera.uc gates its normal visibility/detection block on !bNoAlarm. Thus green camera visibility here cannot establish an alarm milestone, and a normal camera alarm is not an appropriate acceptance condition for these actors. Original Mission00 script provides no jump-camera gate; the later dl_cratestack trigger lies near (7063.58,-1632.38,8.1). Reaching that instruction remains unproven.

The retained module8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379, 131 original/support hashes,105 canonical served sources and19 DLL mounts pass before/after checks. Original SoftDrv/default guestWorker/three workers and guest640×480 were retained. A helper's initial Puppeteer connection briefly set browser800×600 (20/21); this was corrected to1024×768 before the jump. All launch/trajectory/landing captures use1024×768. Reused 17/18 filenames show the backed-away run-up after that correction, not the exact initial save. No claim of uninterrupted browser viewport identity is made.

Pirates'11:40:16 explicit FULLRELEASE and the actual global Chrome/crashpad gate preceded the11:43:52 launch. Wrapper1997135, controller1997138 and monitor1997137 held the full-lifetime lock. At11:48:56 the disk guard closed the browser successfully at1989124096 free bytes, below the2GB floor, after the landing proof. A later pipe helper could not attach and sent no inputs; there was no second launch or onward test. No fresh save/Load validation was attempted. Cache was disabled; the guard acted as required, but continuous free space above2GB was not maintained.

Cleanup verified25 owned runtime/helper/watcher PID/start identities absent, both global crashpads gone, HTTP34907/CDP35571 closed, profile removed, held inputs empty, actual whole browser gate clear and full flock released before 11:50:27 FULLRELEASE. No foreign signals or Boat (ID/state/expiry N/A). Focused identity/save/input/observer/decoder/landing/cleanup checks and doc diffcheck pass. No production edits, full build, Training/recharge replay, asset/INI/clock/save-byte/memory mutation, cheats, input bypass, FPS/audio diversion, public deployment, subagent or successor.

2026-10-31 12:05Z — pipes, farther gap and crate instruction reached

The onward obstacle passage passed, and a genuine count6 save passed fresh native Load Game. Run scratch/runs/20261031T1205Z-deusex-pipe/ reused the accepted original package/geometry/control analysis and restored only the two immutable1034Z count5 files through app.files/vfsPaths and native Load Game. No Training or recharge replay occurred. I inspected the platform footing, crate-room transmission and fresh restored world captures.

The original X binding is Duck; original bToggleCrouch=False makes release restore standing when clear. Original player code lowers collision height 43→16 while crouching, lowering the center27 units. The scratch pipe guard therefore permits the expected centerZ=-11.2 while requiring Walking contact; it does not mistake crouching on the grate for a fall. The farther jump uses its own grounded launch gate before the measured178-unit gap, Y=-830..-1008. The narrow observer retains the checked original exports, actor offsets, unique actor/class, owning slot0/thread1 and JS/Wasm return-boundary reads. Its corridor envelope alone extends fromY=-1600 to-1800 to include the already decoded crate trigger; fresh validation changes only the expected position to the newly decoded save. There are no guest writes or gameplay calls.

Native passage Observed stable result / evidence
Reproduce first gap Grounded launchY=-251.29, velocityY=-319.89, height43; Physics2 then stationary Physics1 atY=-506.78 (22, 25–28).
X+Up under pipeY=-630..-666 Height16 throughout movement; release restores43 atY=-718.68, Z15.8, zero velocity (29–31).
Running Space across178-unit gap Grounded launchY=-816.80, velocityY=-319.89, height43; Physics2 peak observedZ62.99, then stationary Physics1 atY=-1055.49 (32–37). Feet captures two seconds apart show the farther grate.
Last pipe nearY=-1090 Native X+Up and release reach standing Physics1 atY=-1145.18 (38–40).
Door and crate room Native Up reaches the unlocked door; a bounded attempt stops there. Original RightMouse opens it (43), then Up reaches stationary(7003.7866,-1550.2075,15.8).
Next instruction instruction2-centered, 44 and 46 show Jaime Reyes explaining that the wooden crates are too large/heavy and metal crates should build steps; RightMouse picks up and Tab drops. This is the reached dl_cratestack instruction, not completion of crate stacking.

The save used the ordinary Escape → Save Game → Enter path (49–58). Only the newly written save0001/saveinfo.dxs and 00_training.dxs were copied to immutable progress-native-checkpoint/0.bin and 1.bin, with hashes and VFS destinations in its manifest. Original count5 recovery remains unchanged. Original-format decoding gives count6/time00:55:58 and the crate-room actor position above. A fresh page/game instance mounted only those two count6 files, then ordinary native Load Game displayed count6/time00:55:58 (70) and restored the room (73). The observer independently matched the saved position, rotation, Physics1, radius20, height43 and zero velocity. Inspected feet frames 75/76, separated by2200ms, remain stable; 77 restores the forward view. This validates recovery beyond both gaps. Crate manipulation remains next.

The matched module remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after292 sources,131 original/support files,105 canonical served URL/hash pairs and19 DLL mounts pass. Original SoftDrv/default guestWorker, three workers, guest640×480 and browser1024×768 were retained. All49 PNGs are1024×768. All102 native events are trusted and balanced, including two Space presses with Up held/Duck released and two Duck holds. Observer positive, five rejection fixtures, process-identity guards and focused checks pass. RECOVERY.md gives the restore recipe; check.py reproduces the checks. No production edit or full build was needed.

Pirates'12:07:47 FULLRELEASE and12:08:00 cleanup correction preceded the actual whole Chrome/crashpad process gate and full-lifetime flock. The sole browser session ran12:08:09–12:13:15 UTC, about305 seconds, including the fresh page restore. Cache was disabled and console evidence capped; prelaunch free disk exceeded2.3GB and the one-second monitor never reached its2.1GB stop threshold. Wrapper2022756/controller2022759/monitor2022758 retained fd9; browser2022769 closed with exit0. Cleanup checked30 owned PID/start identities absent, both global crashpads gone, HTTP42785/CDP37541 closed, profile removed, empty held inputs, actual global gate clear and full flock release before 12:13:43 FULLRELEASE. No foreign signals, Boat (ID/state/expiry N/A), subagents, successor, additional browser, heavy CLI, build, old-module/current-host mix, asset/INI/clock/save-byte/memory mutation, cheats, input bypass, FPS/audio diversion or public deployment. Camera alarm remains an unsuitable milestone because the original Training cameras have bNoAlarm=True.

2026-10-31 12:20Z — native crate manipulation and recoverable placed steps

Pickup, drop, stacking and climbing passed; the blocked exit remains untraversed. Run scratch/runs/20261031T1220Z-deusex-crates/ restored only the immutable count6/time00:55:58 two-file checkpoint from1205Z through app.files/vfsPaths and ordinary native Load Game. The original actor observer matched (7003.7866,-1550.2075,15.8), rotation114962 and standing Physics1 before movement. No Training, pipe or recharge replay occurred.

Cheap original-package analysis preceded the browser. Original small metal crate defaults have mass40, below the player's default lift limit50; the medium crate has mass80. MaxFrobDistance=112 explains why several distant aims did not highlight. Original DropDecoration places a carried object on a highlighted target; without one it throws it. Original bindings are RightMouse pickup, Tab drop, C/Z turn, arrows move/strafe, PageDown look down and F center view. The wooden barricade is original BSP geometry, with low topZ32, adjacent topZ96 and taller topZ160 surfaces. It cannot be highlighted as a decoration for the target-placement branch.

Inspected ordinary manipulation Result
Small crate pickup (33) RightMouse produces the translucent carried crate.
Targeted Tab drop (40,42) Small metal crate rests on the medium metal crate; it is then picked up again.
Relocation (57,77) The crate is dropped near the barricade and a continuous native run/Space sequence lands on it at standing Z47.57.
Second small crate (98,110) Pickup succeeds. The attempted two-small stack lands beside the first crate; 110-two-small-stack.png is a label, not a successful stack.
Useful steps (127,130,135,140) A level Tab throw places one small crate on the low wooden crate. JC climbs the floor crate, low wooden crate and upper metal crate, reaching stationary Physics1 at Z47.23,79.8 and109.73.
Upper jumps (142,173) Both return to the floor. No doorway crossing or next exercise is claimed.

The final upper jump peaks at centerZ156.913. Its last sampled descending center above143.8 is (7026.23,-1749.40,147.58), still short of the topZ96 wooden surface beginning nearY=-1808. It finishes stationary on the floor at (7058.3203,-1773.3511,15.9). The concrete next dependency is a closer upper crate placement/run-up that reaches that surface at sufficient height. This attempt does not establish an emulator defect. Timed turns vary enough that screenshots and observed state are needed between approaches.

Ordinary Escape → Save Game → Enter created a genuine count7/time01:09:32 checkpoint (181-native-save). Whole native output bytes were exported to immutable progress-native-checkpoint/0.bin and 1.bin; its manifest records the two VFS paths, sizes and SHA256 hashes. Count6 remains immutable and intact. Original-format decoding records the new player position above and relocated small crates at (7061.3101,-1701.6576,-13.1472) and (6997.2246,-1707.7416,49.3502). The medium crate remains in its original place.

A fresh page/game instance mounted only those two count7 files. Native Load Game showed count7/time01:09:32 (208) and restored the world. The owning slot0/thread1 observer independently matched decoded location, rotation184635, Physics1, height43, radius20 and zero velocity. Feet frames two seconds apart remain stable. The copied helper retained 75/76-fresh-count6-* filenames; these files actually depict count7. Inspected 214/216, separated by2200ms, show the restored upper metal crate still resting on the low wooden crate. RECOVERY.md supplies the exact restore recipe and focused checks.

The exact module remains 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, with292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379. Before/after292 sources,131 original/support hashes,105 canonical served responses and19 DLL mounts match. Original SoftDrv/default guestWorker, three workers, guest640×480 and browser1024×768 were retained. All68 PNGs are 1024×768. All294 native events are trusted and balanced, including seven RightMouse presses, four Tab presses and16 Space presses. The accepted read-only observer changes only its route envelope toX<=7450,Y>=-2500, justified by decoded room actors/geometry; exact initial save crosschecks, original exports/offsets, unique actor/class and return-boundary reads remain. The observer positive case and five rejection fixtures pass.

Harness limits are preserved: the first fresh-load command reloaded before the checkpoint manifest existed and rejected ENOENT before guest launch. After the manifest was prepared, fresh restore succeeded in the same browser. The second reload overwrote pre-fresh-native-input.json; complete first-page events survive in 183-after-save.json, combined with the fresh page's final log by check.py. The controller's final observer-stop sees the detached old worker; the initial observer had already stopped before reload and the fresh helper separately stopped its observer. No guest failure is inferred.

Prior Deus/Pirates explicit FULLRELEASE and the actual global process gate preceded the sole12:22:10–12:38:45 browser session (994.9 seconds). The durable wrapper/controller/monitor held fd9 for the browser lock. Cache was disabled; prelaunch free disk exceeded2.3GB and the one-second2.1GB guard never fired. Cleanup verified42 tracked PID/start identities absent, including all helpers, watcher and both global crashpads; both ports closed, profile removed, inputs released and full flock available before12:39:18 FULLRELEASE. Browser exit0 and controller wait0 are recorded. The wrapper's raw exit code was unavailable after its process disappeared; no raw wrapper-exit0 claim is made. No foreign signals, production edits, full build, arbitrary memory scans or writes, asset/INI/clock/save-byte edits, cheats, input bypass, guestWorker disable, public deployment, Boat (ID/state/expiry N/A), subagents or successor.

2026-10-31 crate exit: repositioned launch reaches a transient wood contact; exit remains blocked

Run 20261031T1301Z-deusex-exit restored only the immutable count7/time01:09:32 two-file manifest through app.files/vfsPaths and ordinary native Load Game. The accepted observer matched player (7058.3203,-1773.3511,15.9),yaw184635, Physics1,height43,radius20 and zero velocity before movement. Count6 and count7 files remain unchanged. Training, pipes and recharge were not replayed.

Original BSP analysis identifies the intended lower wood at X7024–7088, Y−1872–−1808,topZ96. The adjacent tall wood at X6960–7024,Y−1856–−1792, topZ160 intercepts an approach from the left. The actual next exercise is the original dl_ladder trigger near (6991.97,−2362.70,5.55), beyond the crate-room doors near Y−2000. Neither those doors nor that exercise was reached.

Ordinary climbs again reached the floor small crate, low wood and upper metal. Attempts with insufficient running speed fell short; longer approaches could leave the upper crate before Space took effect. RightMouse then lifted the upper crate (63/65); ordinary movement while carrying and a downward-aimed Tab drop repositioned it (69/74). Subsequent stable standing centerZ111.427 is recorded. Carrying mass40 prevents jumping in the original DoJump script; the crate was dropped before every subsequent jump.

The repositioned diagonal jump 78 reached centerZ158.77, but its forward velocity collapsed near (7019.78,−1770.71,157.87). Original geometry explains this as contact with the front of the topZ160 wood, before the player cleared its right edge. A more easterly ordinary running jump (90:Up220ms,then Space650ms,then both released) cleared that obstruction. Its read-only trace records Physics1 and zero vertical velocity at (7102.9736,−1797.7441,143.4226), consistent with brief contact at the topZ96 wood's east edge. Horizontal velocity was still (206.73,−106.78). This contact lasted only two observer samples,25ms apart; no image captured that instant. The player continued into the east floor area, where the inspected final image shows no new exercise. This is not a stable landing or exit traversal.

The precise remaining dependency is to retain the topZ96 landing after clearing the topZ160 wood, then cross its far edge toward the actual doors. Releasing Up earlier is the next control hypothesis, not a verified solution. The attempted repeat failed to regain the same launch state. Native save decoding later showed that contact had shifted the floor small crate north to (7061.5371,−1666.8571,−13.1472), and the repositioned upper crate to (6968.7236,−1743.5118,51.05). Reusing earlier approach coordinates was therefore unreliable. No emulator defect is established.

An ordinary native save produced count8/time01:24:25. Whole immutable files are retained in this run's progress-native-checkpoint/: saveinfo34576B,SHA256ba471693be27576e9dd30a4b901415ad00b916a2d828a8443eb37d7f46e7dc68; Training3232653B,SHA256b25220987668c04f8346f21cb925d8151d1e55d286b6787f1e618331abb4bd4b. A reserved fresh page mounted only those two files, showed native count8 in the Load dialog (141), and ordinary native Load restored the decoded player (6948.3442,−1755.5117,111.427),yaw335104,Physics1,BaseCrateUnbreakableSmall0. The observer independently matched location/rotation and zero velocity, then remained stable2908ms. The final screenshot (145) still shows LOADING: restored gameplay rendering/layout was not visually validated before the bounded deadline. A requested second capture was never executed. Count8 is a diagnostic save with this explicit limit; accepted count7 remains preferred.

Exact module8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379, 131 original/support hashes,105 canonical served sources and19 DLL mounts pass before/after checks. Frozen0403Z source/module and0736 canonical support, SoftDrv/default guestWorker,three workers,guest640×480,browser1024×768 remain. All528 native events are trusted and balanced, including30 Space,one RightMouse and one Tab; all58 PNGs are1024×768. The initial observer is byte identical to the accepted count7 observer; the fresh observer changes only decoded-save initial location/yaw. Positive and five negative fixtures pass.

Fresh Pirates FULLRELEASE13:09:15 and the actual process gate preceded the single bounded session. Controller start13:09:32.735 to cleanup13:26:34.562 was1021.827s. A detached durable supervisor retained actual Popen wait results: wrapper0,controller0,monitor0; browser exit0. Full-lifetime fd9 history covers wrapper/controller/monitor while the profile existed. Disk started>=2.3GB, minimum2227294208B; the2.1GB early stop did not fire. The timed final wait crossed the17-minute controller deadline by less than one second; browser cleanup still finished before18 minutes,all cleanup before19 minutes. Historical audit verifies21 runtime PID/start identities absent,including both global crashpads,two ports closed,profile absent,held inputs empty and full flock available. The board watcher was separately qualified/terminated; the pending capture helper timed out and exited. FULLRELEASE was announced 13:27:50 after actual absence checks. No further browser was launched; no global gate is rerun during the next worker's session. No foreign signals, production edits,build,asset/INI/clock/save-byte/memory edits,cheats,input bypass,guestWorker disable,public deployment,Boat or subagents.

2026-10-31 13:40 UTC: corrected bearing and earlier release, takeoff dependency

Run scratch/runs/20261031T1340Z-deusex-wood/ restored accepted count7 through app.files/vfsPaths and ordinary native Load. The count7/time01:09:32 dialog (13), actual restored crate room (17), and original owning-worker observer save crosscheck were inspected. Stable small-crate/wood/upper-crate climbs reached centers Z47.23/79.8/109.73. No stable Z96 landing, exit traversal or ladder exercise was reached; no new save was created. Count7 remains preferred, count6 untouched, and count8's previous LOADING limitation remains unresolved.

Cheap original analysis first reread the accepted recovery and trajectory90, original decoded BSP nodes8270/8276 and original DoJump. Radius20 puts the prior 7102.9736,−1797.7441 contact18.1492 units from lowerwood's southeast corner; it lies outside both face bounds. A conservative full-footprint target is X7044..7068,Y−1852..−1828, not a demonstrated reachable landing. The prior trace also retains substantial airborne velocity after acceleration becomes zero. Releasing Up cannot be assumed to stop airborne motion immediately.

The original collision remains authoritative. The rotated top32 base's actual BSP top is a diamond centered near6992,−1684, with vertices extending to X6946.75..7037.25 andY−1729.26..−1638.75. Crates drift during native collisions; visual layout and observed footing must be checked after each approach. Native RightMouse pickup succeeded. Forward Tab drops first returned "Can't drop that here"; a backward offset allowed a drop to the floor, then ordinary count7 Load restored the accepted layout. There was no Training/pipes replay.

Checked attempt Actual result
97: target yaw52500, Up190ms then Space190ms, release Physics changed to falling125ms into the trace, before Space; no upward velocity. Player walked off the upper crate.
127: back to7015.20,−1691.53,109.58; target yaw52500; Up220ms then Space220ms, release Actual jump, first observed upward velocity236.4. Tallwood collision reduced northward speed near7041.92,−1768.63,155.96; after early release, northward velocity fell to about−10.3. Fell to floor.
143: back to7013.73,−1692.47,109.73; target yaw53500; Up220ms then Space240ms, release Wider bearing, but support lost before Space. No upward jump or landing.
161: upper footing7012.53,−1693.83,109.60; target yaw53500; Up140ms then Space500ms, release Support was already lost93ms into the trace; no upward jump. Final inspected gameplay (162) is still the crate room floor.

The first dependency is grounded takeoff before the moving upper crate loses support, followed by sufficient east clearance of tallwood. Earlier Up release and a corrected bearing have now been tried together, but these attempts do not establish whether they retain a Z96 landing after a valid clear jump. A stationary Space-first takeoff from rechecked footing is untested; its horizontal range must be measured, not assumed. No emulator defect established.

Prelaunch initially stopped at2.130GB free, below required2.3GB. The coordinator reclaimed released historical diagnostics; this worker deleted no foreign files. Pirates13:40:11 FULLRELEASE and a fresh actual whole Chrome/global-crashpad gate preceded sole browser launch13:42:36. New deadlines were generated, not reused. Original frozen module8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest,131 original/support hashes before/after, original SoftDrv, default guestWorker,three workers,guest640/browser1024 were retained.

check.py passes identity, accepted observer byte equality, positive plus five rejection fixtures, native save crosscheck,540 trusted balanced input events, 54 PNG dimensions, actual trajectory classifications, and cleanup. One repeated observer install after native reload returned "observer already installed"; the original observer continued sampling its checked actor. The final quit command deliberately exits the controller before command acknowledgement; its sender timed out and was confirmed absent. Neither is a guest failure.

Actual browser cleanup13:56:56 completed858.995s after controller start, within 18min; raw browser/controller/wrapper/monitor exits0/0/0/0 are retained by the durable Python supervisor. All1002 active-profile monitor samples show wrapper, controller and monitor holding inherited fd9; minimum free disk2228436992B. Historical cleanup audit confirms22 owned PID/start identities absent, including global crashpads and the final command helper, both ports closed, profile gone, held inputs empty and full flock available. Monitor used no signals. FULLRELEASE was announced after actual absence. No further browser, production edits, build, heavy CLI, asset/INI/clock/save-byte/memory edits, cheats, input bypass, guestWorker disable, public deployment, Boat(ID/state/expiry N/A), agents or successor.

2026-10-31 grounded takeoff: Space-first is genuine; wood96 landing remains unreached

Run scratch/runs/20261031T1403Z-deusex-grounded/ establishes a genuine stationary upper-crate jump. No stable topZ96 landing, exit or ladder exercise was reached. No manual checkpoint was created or exported. Count7 remains preferred, count6 untouched and diagnostic count8 unpromoted. No emulator defect is established.

Cheap preparation checked original DoJump, the earlier trajectories and original BSP bounds before launch. DoJump supplies vertical velocity only; the approximately154-unit range from the earlier upper footing to (7056,−1840) cannot be assumed reachable from rest. Tallwood occupies X6960–7024, Y−1856–−1792,top160; lower wood occupies X7024–7088,Y−1872–−1808,top96. The radius20 full-footprint inset remains a planning hypothesis, with native collision authoritative. No Up140–220ms blind upper-crate lead-in was repeated.

Ordinary native Load showed count7/time01:09:32 (13); inspected actual gameplay (17) and the unchanged original observer independently matched the immutable save. Short native climbs reached upper-crate centerZ109.55. The new Space-first action required two Physics1/zero-velocity samples120ms apart before input, then confirmed positive vertical velocity before applying Up.

Actual checked action Result
28, stationary Space-first, actual yaw54272 Grounded at (7014.9941,−1714.2648,109.5548); native Space followed by observed Vz+239.2064 while horizontal velocity was zero; only then native Up. Apex centerZ156.6893. Returned to upper support at (7021.3589,−1726.1774,109.6133), confirmed stationary by 30. Horizontal range13.5063 units.
32, observed approach Last supported sample near (7019.71,−1723.09), speed142; next sample fell. Guard released Up without pressing Space. This was a canceled approach, not a jump.
50, earlier observed Space, actual yaw53308 Physics1 at (7018.5864,−1722.1311), velocity(55.87,−132.71,0) before Space; actual Vz300 confirmed. Near (7039.80,−1769.65,156.13), north speed collapsed to−52.19, then−19.31: another tallwood collision. Up was already released; airborne momentum persisted. Fell to floor.
76, eastward preparation Initial stable-upper gate refused because footing had dropped to the low wood. No jump input.
115, more easterly actual yaw54983 The last Space-decision sample was Physics1 near (7030.65,−1719.51), but the next state was Falling/Vz0. Support was lost during input delivery; stopped immediately without claiming takeoff.

This separates two remaining requirements: a supported approach that survives input delivery, and enough east clearance plus range to reach wood96. A Physics1 decision sample alone is insufficient; positive Vz is necessary evidence of takeoff. The stationary jump solves the original false-takeoff dependency, but supplies far too little range. Ordinary climbs also shifted the useful footing: an eastward step later dropped from centerZ109.73 to79.8. Reusing an earlier coordinate as current support is unsafe. A changed native crate layout or measured longer supported approach remains unverified.

Limits and recovery are explicit. The attached helper initially used Puppeteer's default800×600 viewport for 32 and 50; it was corrected to preserve the owned page and1024×768 was restored at 51. Those two PNGs are 800×600; the other41 PNGs are1024×768. The accepted observer's background timer reached its20000-sample cap; subsequent on-demand samples still used the same original function, checked exports/offsets, unique actor/class and owning-worker return boundary. No guest memory or actor writes were introduced.

An attempted native menu click unexpectedly started Training instead of Load. Its delayed LOADING view (87) was followed by the start hall (90), where the observer rejected the route envelope. No Training exercise was continued. Keyboard navigation first opened Save Game (95), which was canceled without saving, then actual Load Game (100) showed count7/time01:09:32. Native Load restored the crate room (103); clearing only the old JavaScript observer state and reinstalling the unchanged observer passed the exact save crosscheck on the new actor. Neither the misleading intermediate labels nor the start hall are treated as progress. Final inspected gameplay (117) is the crate-room floor. No fresh-page checkpoint validation was attempted without a meaningful landing, and no new save is promoted.

Identity remains module 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379, 131 original/support files before/after,105 canonical served responses and19 DLL mounts. Original SoftDrv/default guestWorker,three workers and original640×480 guest mode were retained; the browser viewport exception is stated above. An unpinned /build-info.js request was refused404, not served from current host. check.py and analyze.py pass immutable count7/source/module checks, accepted observer positive/five rejection fixtures,392 trusted balanced input events (18 Space presses), actual trajectory classifications and retained cleanup. All26 references in this run pass containment, existence and schema checks. The global evidence checker still reports9 historical outside-run errors, 10 missing references and103 invalid records; these were not changed.

Pirates14:16:30 explicit FULLRELEASE and a fresh whole Chrome/global-crashpad gate preceded the sole14:17:03–14:30:17 browser session (792.692s from controller start). The durable supervisor retained actual browser/controller/wrapper/monitor exits0/0/0/0. All1018 active-profile samples show wrapper/controller/monitor holding inherited fd9; minimum free disk2392264704B. Cleanup confirmed30 tracked PID/start identities absent, including guarded helpers, global crashpads and the preparatory watcher; both ports closed, profile gone, inputs empty and full flock available before14:30:36 FULLRELEASE. No monitor signals were needed. The initial tail watcher's raw exit was unavailable after its parent exited; its own PID/start absence was checked, with no raw-exit0 claim. No production edits, build, heavy CLI, asset/INI/clock/save-byte edits, cheats, input bypass, guestWorker disable, public deployment, Boat(ID/state/expiry N/A), agents or successor.

2026-10-31 native crate reposition: genuine raised stack; launch access unresolved

Run scratch/runs/20261031T1450Z-deusex-reposition/ changed the native crate layout and produced a genuine two-small stack. No stable96 landing, exit or ladder exercise was reached. The new diagnostic save is NOT accepted: fresh-page nativeLoad/gameplay validation did not execute before the disk guard shut down. Original count7/time01:09:32 remains preferred, prior1301Z count8 remains unpromoted and count6 remains untouched. No emulator defect is established.

Recovery starts with the immutable two-file count7 manifest in ../20261031T1220Z-deusex-crates/progress-native-checkpoint/manifest.json. Use the exact0403Z after.wasm/source-after/ and0736Z canonical support, original SoftDrv/default guestWorker/three workers,640×480 guest and1024×768 browser. This run's keyboard nativeLoad showed count7/time01:09:32 (18), then actual crate-room gameplay (22) and the unchanged original observer passed its exact unique-player/save crosscheck (21-exact-count7-crosscheck-observer.json). Early startup labels 08/12 were premature:12 is the actual main menu, not a Load dialog. No Training selection or exercise replay occurred.

Original DropDecoration was read before launch: a highlighted actor gets a zero-velocity placement with randomized XY above its collision height; without a target, Tab throws forward with view-based velocity. Native traces and collision decide acceptance. Mass40 small crates are liftable under limit50, but original DoJump refuses jumping while carrying mass>20. The BSP wooden support cannot be highlighted. Original User.ini confirms Delete/R LookUp, PageDown/V LookDown, F/End CenterView, X Duck; PageUp is unbound. Two PageUp holds had no effect; subsequent Delete holds were ordinary bound input.

Native action and inspected evidence Actual result
RightMouse34 from the floor north side,36 “There's no room to lift that”; no carried floor crate. Tab41 therefore did not create a stack.
RightMouse44,46; untargeted Tab54,56 Lifted the original upper small crate, then threw it back onto the wooden support. The throw shifted it west rather than toward the exit.
RightMouse67 from clear east side,69; highlighted-target Tab84/90,83/86/89/92 Floor small was lifted. Both target drops were refused with “Can't drop that here”; crate remained carried. No stack was inferred.
Untargeted Tab96,98; climbs107/113/121 Floor small settled near the wooden support. Genuine positive-Vz climbs reached stable centerZ48.09,79.8 and109.66. The moved upper footing was (7015.09,−1709.86,109.66), not a closer launch than count7.
Crouched raised RightMouse151 From wood centerZ79.8 at (7017.76,−1727.91), native X crouch brought the floor crate within reach. RightMouse lifted it; X was released in finally. Attached Puppeteer used defaultViewport:null and explicitly1024×768.
Highlighted-target Tab161,160/163 Raised target drop succeeded. The carried crate became opaque above the other small crate. Later decoded native save independently confirms small1.Base=small0, with centers32.515 units apart vertically and11.22 horizontally.
Stack climb165 Started stable Physics1, observed Vz+159.40, apex centerZ125.87, then returned to wood at (7019.12,−1725.67,79.8), zero velocity. The raised stack's upper surface was not reached.

The saved actual dynamic positions supersede any reused checkpoint coordinates:

Actor Position Saved base
Small0, lower (6984.11865,−1699.88647,49.28199) LevelInfo0
Small1, upper (6989.43408,−1709.77527,81.79706) CrateUnbreakableSmall0
Player (7019.12061,−1725.66577,79.80000),yaw286431,Physics1,height43 LevelInfo0

This is a genuine raised stack but still leaves access to its upper surface unresolved. Its nominal upper top is about97.05; from wood standing center79.8, the tested jump apex stayed below the approximately144.85 standing center needed there. Native collision remains authoritative. A usable lower intermediate step or a different supported arrangement is the next dependency; repeating exit-jump timing from the old upper location is not supported by this result. The tallwood/wood96 bounds and radius20 hypothesis remain as in the1403Z recovery.

A native Escape→Save Game→Enter created diagnostic count8/time01:22:17. The first export174 caught an in-progress save.tmp and unchanged map bytes; only completed export177 was used. Its whole, unedited files are chmod444 in progress-native-checkpoint/, with exact pins:

File Bytes SHA256
0.bin,saveinfo.dxs 34576 b3f199bdc14d67e3e518dc7e19b7145e2a541e101b44cc12df3f084059f26884
1.bin,00_training.dxs 3232834 18799fe3c57787ed97291c231be8ce8e403dd2b1a0b44be3171ae2581ab4a48f

These pins identify this new diagnostic count8, not the earlier1301Z count8. prepare-save.py decoded only named player/crate exports from the native save; observer-new-save.cjs changes only the decoded expected position/yaw in the original observer. It was prepared but never installed. Do not promote or choose this save as preferred without an actual fresh-page nativeLoad, inspected gameplay and original observer/save crosscheck.

The monitor submitted quit after completed-export177, before pending fresh-load178 executed. No fresh launch/reload/nativeLoad occurred; the fresh request timed out with tool raw exit1. All runtime processes exited cleanly; this request timeout is not a guest crash. The original observer timer had reached20000 samples; later direct reads used the unchanged checked function, original exports/offsets, unique actor and owning-worker return boundary. No arbitrary live actor reads, scans, writes or input bypass were introduced.

Identity checks retain module 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676, 292-source manifest 9dee97b3f74481ba14d6808ecce1d4b665099d8b03c58235e0b02812b234e379, 131 original/support files before/after,105 canonical served responses and19 DLL mounts. Unpinned build-info.js was refused404. check.py passes immutable count7/source/module/native-input/stack-base/trajectory/cleanup checks and the original observer's positive/five-rejection fixtures.450 trusted input events are balanced, including5 RightMouse,6 Tab,10 Space and1 X presses. All56 PNGs are1024×768; the exact named inspected-image hashes are in focused-checks.json. No stable96/exit/new-save-acceptance claim is made.

Pirates explicit FULLRELEASE14:49:01 plus an actual whole Chrome/global-crashpad/ profile PID-start gate preceded the sole14:50:41–15:04:52 browser session (850.202s from controller start). The2.1GB guard fired at2075869184B; minimum sampled free disk2068295680B remained above2GB. All907 active-profile samples show inherited fd9 in wrapper/controller/monitor. Actual browser/controller/ wrapper/monitor exits0/0/0/0 were retained by the durable Python supervisor; monitor used no signals. Cleanup confirmed23 tracked PID/start identities absent, plus the guarded helper and qualified own request processes absent, both ports closed, profile gone, held inputs empty and full flock available. FULLRELEASE was announced15:06:11 after actual cleanup. No further browser, heavy CLI, production edits/build, asset/INI/clock/save-byte/memory edits, cheats, guestWorker disable, public deployment, Boat(ID/state/expiry N/A), agents or successors. Only this game document and new scratch evidence were changed.

The run evidence checker reports no outside-run, missing or invalid references for this run. Its global report still contains9 historical outside-run errors, 13 missing references and103 invalid records; unrelated evidence was unchanged.

2026-10-31 interrupted stack-access recovery: exact count8 restored; access still blocked

The restore limitation in the preceding 14:50 investigation is now resolved by scratch/runs/20261031T1514Z-deusex-stack-access/, recovered in scratch/runs/20261031T1521Z-deusex-stack-recovery/. The coordinator transferred the absent worker's documentation and live runtime to this recovery worker. The original launcher reportedly returned143 without result/final; its cause is UNKNOWN. No missing original result or exit receipt was reconstructed. The existing browser/controller/supervisor were reused without a new browser, reload, controller, deadline extension or production change.

The 14:50 whole unedited diagnostic count8/time01:22:17 is accepted as a restorable raised-stack checkpoint. Count7/time01:09:32 remains preferred. This does not promote the earlier13:01 count8 or establish stack-top access. The inherited fresh page mounted exactly the preceding two SHA256-pinned files. Inspected18-load-dialog shows count8/time01:22:17; ordinary Enter19 performed nativeLoad. Checked21-exact-count8-crosscheck-observer matches the decoded unique JCDentonMale0 at (7019.12061,−1725.66577,79.80000),yaw286431, Physics1,height43 and zero velocity.22 still shows LOADING and is insufficient alone;24 shows actual gameplay and27 shows the highlighted raised metal crate. The named saved exports independently retain small1.Base=small0 and the exact two crate positions documented above. The runtime observer reads only the original checked player fields, not live crate transforms. All save files and the preferred count7 remain immutable and hash-identical after shutdown.

Recovery restarted the existing checked observer, centered the view with End, and made one bounded ordinary Space→X+Up attempt from the stable wood position. 31-native-air-crouch-stack-trajectory records positive Vz300 and sampled center apex126.32271. Airborne crouch clearance was not established: height remained43 through the sampled flight and return to Physics1; the native height16 change appears only afterward. Trusted key delivery does not prove when the guest consumed X. Horizontal displacement was only0.19 units. Inspected32 and its actor record settle at (7019.21045,−1725.49902,52.92253),Physics1,height16, zero velocity: feet36.92253 remain at the previous wooden-support level. The first failed access dependency is therefore an actual usable intermediate support or demonstrated airborne crouch clearance. No new stack footing, wood96 landing, exit, save or emulator defect is claimed. Gameplay stopped there; no previous crate work or exit-jump sequence was replayed.

The original named-export/offset/unique-player/owning-worker return-boundary observer is byte-identical to the inspected14:50 observer-new-save.cjs. Its positive and five rejection fixtures pass. Recovery check.py verifies the exact8459597c8d module,292 source files,131 original files before/after, 105 canonical served responses,19 DLL mounts, both immutable checkpoints, 30 balanced trusted input events,12 images all1024×768, trajectory and cleanup. Original SoftDrv/default guestWorker/three guest workers remain unchanged. No arbitrary memory read/write, cheats, input bypass or save-byte edit occurred.

Cooperative quit33 was submitted before the fixed15:31:55 runtime deadline; cleanup completed15:22:16 and the inherited supervisor retained actual controller/monitor/wrapper waits0/0/0, with browser exit0. The monitor used no signals. At release,19 tracked PID/start identities plus12 recorded request helpers were absent; both ports were closed, profile absent, held inputs empty, global Chrome/crashpad gate clear and full flock available. All450 active-profile samples show the original three inherited fd9 holders. Minimum sampled free disk2243272704B remained above2GB. FULLRELEASE was announced15:22:33 only after the audit, preserving both original deadlines. These runtime exits do not explain or replace the original worker's reported143.

The recovery's focused-checks.json and takeover.json preserve inspected-image hashes, actual limits and original identity/deadlines; full-release-audit.json retains the lifecycle evidence. Boat ID/state/expiry are N/A. No further browser, heavy CLI/build, public deployment, laptop work, foreign signals, agents or successors were used. Next work would need a genuinely different supported access route; count7 remains the preferred fallback.

2026-11-04 supported-stack preparation: crouch consumed before takeoff; first step still unresolved

Run scratch/runs/20261104-deusex-supported-stack/ restored the preferred count7/time01:09:32 through ordinary native Load. It established grounded and airborne crouch consumption, but no new supported footing, stack access, exit, ladder exercise or measured FPS. Gameplay stopped after the first new access attempt returned to the floor. No crate pickup/drop or new save occurred; both accepted checkpoint pairs remain unchanged. The failed prior Space→X+Up sequence was not repeated.

Historical identity was recovered before execution. The0403Z manifest records base a504fa248caaecbbb042b65e9ae5384a340ade26, with dirty source including concurrent font work, rather than a clean current-main build. All292 exact source hashes now also reproduce from that Git base plus this run's historical-source.patch: git diff --binary over nine changed files, independently applied and checked using a temporary Git index, without a source snapshot, checkout copy or branch. historical-source-recovery.json records the patch hash. Module SHA256 is 8459597c8d66e098d727fb4ab53e1cf46332241dc7eef2e158f3c475a78b1676. All131 original/support pins,105 canonical served-response identities and19 native DLL paths passed. Original SoftDrv, default guestWorker, three original live guest workers and1024×768 captures were retained. Current host modules were not used.

The initial launch failed before a guest existed because the run's restricted server lacked the five global .fon files. The original failure frame and receipts remain in recovered/startup-failure/. The accepted14:50 response log supplied exact hashes for40 additional global assets; matching existing files were provisioned directly into the test runtime. startup-dependency-check.json and startup-repair-verification.json record this bounded preparation repair. No media was copied into evidence. The browser retest retained the first controller's original18-minute deadline. There was no guest-route replay.

Inspected 05-native-load.png shows count7/time01:09:32, and 07-exact-count7-observer.json passes the unchanged observer's unique JCDentonMale0/save crosscheck at (7058.32031,−1773.35107,15.90000),yaw184635, Physics1,height43,zero velocity. 08 still says LOADING; only inspected 09-restored-scene.png establishes actual crate-room gameplay before movement. The original observer and its positive/five rejection fixtures were retained byte-for-byte. Reads remained at the owning-worker return boundary; there were no arbitrary live crate reads, memory scans/writes or injected game calls.

Cheap preparation reread original DoJump, PlayerWalking.ProcessMove, DropDecoration, bindings and accepted saved actor geometry. ProcessMove performs the height16 change while walking; DoJump allows crouching but refuses a carried decoration heavier than20. A nominal offset upper-crate layout was considered, but highlighted-target drops randomize their XY placement rather than accepting a chosen offset. It was not treated as a reachable new support. The selected bounded hypothesis was instead to consume X on the ground before Space, then test the original floor small crate as the first support, whose saved top is about2.103. Its saved transform remains planning evidence, not a new measurement of its live transform after collision.

Inspected action/evidence Actual result
Grounded X; 10 and 11,461ms apart Both Physics1,zero velocity,height16,centerZ−11.10000. Feet remain−27.10000. Crouch consumption is demonstrated before any Space.
ArrowRight400ms; 12 This strafed, rather than turned: player moved to (7084.43994,−1761.60461,−11.10000),yaw unchanged. Original User.ini was reread; C is TurnRight. This control error is retained.
C600ms then1050ms; 13,14 Actual yaw216816, stationary,crouched, facing the highlighted floor small crate. No crate was lifted or moved by a pickup/drop action.
Space+Up with X already consumed; 15–18 129 checked trajectory samples retain height16 throughout. Positive sampled Vz204.05661 establishes takeoff; sampled apex centerZ36.18572. 16 is airborne, not a landing.
Settled 17,18,777ms apart Identical (7076.08545,−1739.81909,−11.20000),Physics1,height16,zero velocity. Feet−27.20000 and inspected crate-ahead view establish return to floor, not new support.
Release X; 19,20 Standing height43,centerZ15.80000 at the same XY,zero velocity; all native inputs released.

The tested crouch-first hypothesis resolves input consumption, but not access. Horizontal displacement during the jump was small: at the last sampled descent state above the saved crate-top height, player XY was (7081.22363,−1753.20996); by (7080.22559,−1750.60718) the feet were already below that nominal top. Those positions are55.26 and52.48 units from the saved crate center. No live crate transform or precise collision margin is inferred from those distances. The first unresolved dependency is reaching usable horizontal support before descending below it. A closer, rechecked stationary launch is the next hypothesis; it has not been attempted or qualified. No exit jump, repeated crate route, new checkpoint or emulator fix followed this failure.

check.py passes the source/Git-patch identity, both immutable checkpoint pairs, observer fixtures/native crosscheck, grounded/airborne height16, positive trajectory and stable floor result,30 balanced trusted native events,12 completed receipts (maximum12 steps, limit64),18 PNG dimensions,458 sealed runtime inputs with the explicitly recorded helper repair, and95 recovered unique evidence files/full SHA256. Both browser sessions together lasted329.236s through final cleanup, within the original18-minute bound. Controller exit0, held inputs empty and port8123 closed are retained. No production source was changed, so no production build or defect claim is made.

Fresh authorized box bx_z9s6mjdr,TTL3600,expiry 2026-11-04T01:28:40.864Z, was recorded before execution. It was handed back RUNNING/idle, with coordinator shutdown ownership; the worker did not stop, delete or resume any box. One module was executed and recovered locally with its full SHA256. The accepted historical closure supplies only the canonical845 module, with no second compatibility-module identity; therefore both-module recovery cannot be claimed, and no current compatibility module was substituted. remote-recovery-manifest.json, local-recovery-verified.json and boat-handoff-info.json retain these evidence and lifecycle limits.

Subsequent coordinator board entry00:41:33Z records the verified named snapshot deusex-supported-stack-20261104 and actual normal STOPPED at 2026-11-04T00:41:10.755Z,forcedfalse/errornull. This was coordinator action after the worker's RUNNING/idle handback; the worker performed no box lifecycle mutation. The original expiry is now inactive.