Braveheart demo (Red Lemon Studios / Eidos 1999)

Registry id braveheart_demo. Real-time tactics over a 3D battlefield. The "Covermount Demo" ships three renderers side by side: brave.exe (Glide), bhd3d.exe (Direct3D) and bhsoft.exe (software, into DirectDraw). The registry entry runs bhsoft.exe. Fixture: test/binaries/win98-games-a-d/braveheart-demo-Glide (InstallShield 5 Disk1); the installed tree lives beside it in braveheart-demo-Glide-installed/braveheart covermount demo (199 MB, dad.io is 133 MB of it).

Install (headless)

The Drakan/Driver recipe (docs/re-notes/drakan-demo.md) unchanged: SETUP.EXE --vfs-include='*' --tick-ms-per-batch=5 --capture-launch=cap, then cap/windows/temp/_istmp1.dir/_ins5576._mp with --vfs-tree=cap --tick-ms-per-batch=5 --control=PORT --frozen --save-vfs=inst --save-vfs-prefix='c:\program files': Next (413,394) on Welcome, Destination and Program Folder, ~1.2M batches of copying (about 3 minutes), Finish. Target: C:\Program Files\Red Lemon Studios\Braveheart Covermount Demo. No emulator fixes were needed for the install or the game.

Running it

Headless clock spiral (read before calling the battle frozen)

At the default --tick-ms-per-batch=200 the battle freezes on one frame for good. It is not a hang: the sim is a fixed 16 ms tick fed by a sliding window of the last eight frame deltas, and every frame costs more guest time than it simulates, so the window only grows.

Measured with --input/ctl dump-mem:0xc5aef4:4: 180,176 steps per frame at 200 ms/batch; at 5 ms/batch it starts at 551 and creeps up (~1,650 after 100k batches, still diverging slowly); dropping to 1 ms/batch (ctl POST {"action":"tick","ms":1}) converges to under 200 within ~30k batches. So run the battle at --tick-ms-per-batch=1 (or 5 to reach it faster, then 1). The browser uses real time. Convergence was a hypothesis here, not an observation; the 2026-11-02 browser run below does not establish convergence.

Route (CLI): --tick-ms-per-batch=5, Esc every 40k batches x6 -> Mission Brief by ~160k; Take the Field; battle frames from ~390k. Evidence: scratch/runs/20261006T1640Z-braveheart_demo-gameplay.

Native browser observation — 2026-11-02

Run: scratch/runs/20261102-braveheart-native-browser/. This is an honest first-blocker result, not accepted battlefield gameplay. One original ran in headful Chrome 151.0.7922.108 with the default host renderer (webgl), real clock, guest worker and threads. No guest patches, injected state, renderer selection, clock changes, download or public promotion were used.

The prelaunch registry check found braveheart_demo in APPS and LOCAL_CANDIDATE_APPS, absent from DESKTOP_APPS. The registered bhsoft.exe and six real movie/audio DLLs were retained. All 145 installed files (207704346 bytes) matched the coordinator's locally observed hashes; those hashes are not upstream provenance. All 143 browser-manifest entries matched actual file sizes. Exact per-file identities are in original-inputs.json; bhsoft.exe SHA-256 is 70ae3feedc1f977ca8b8ed32560ebdaa381a25b6261f2cbc77396c887e2a8955 and browser-manifest SHA-256 is 2b4fd55edf1b74adb3a102893b12e052283130cbb22aeaca5d048b23dd40b77f.

After the ordered predecessor releases, source snapshot 345270459395baf28e88d68ace9c3d311f1f4e70 matched all 3936 accepted inputs of 20261102-gl-stipple-implementation-recover and its locked full build (exit 0). Tested module: 2071475 bytes, SHA-256 307168f9e0475fd3c10f87cbc5d98f05006c6d4494d2f564381a9a0ebe289fe6. No production code changed for this investigation. snapshot-inputs.json, accepted-build-inputs.json and build-identity.json pin runtime/compiler inputs. The immutable transfer included src/api_table.json; all 3010 members were verified before and after extraction. Archive SHA-256: 4ec58adf4645baf981a176e4b68b3ac9908cef2bb8627cfa2714b72ae209e85a.

Actions were chosen after viewing each screen: interlaced Eidos movie → balanced native Escape → Game Features → visible Back → main menu → visible Raid Town → Stephen Feherty mission brief → visible Conduct Raid → lazy asset loading. The shipped readme names renderer executables but supplies no battlefield control instructions. The older CLI Take the Field route above was not imposed on the browser's different visible screens. Native event records contain 11 trusted events and four balanced press/release pairs; actual holds were 4451, 2368, 1443 and 4312 ms, including capture overhead. 04-raid-town-settled provides the before image for Conduct Raid; 05-conduct-raid-held, -released, -settled preserve its input sequence.

The first blocker was persistent unreadable software presentation: a noisy green/brown strip across the upper canvas with the lower area black, seen through 07-loading-progress, both 08-neutral-render-check captures and 09-final-neutral-blocker. The latter was personally reviewed after retrieval. Its cause is unlocalized. No readable Take the Field button, unit selection or ordinary battlefield command was reached. Inputs stopped at that blocker; no unchanged original was replayed.

Read-only simulation samples were 10, 14, 13, 21, 39, then 12 steps, with changing delta rings, guest registers, work statistics and real tick values. The final ring was [168,172,144,377,170,152,151,212]; final guest tick 261836, host stepTicks 28615. The recorded frame-end/pump counters remained 623/620 across the malformed presentation samples. These counters do not establish displayed FPS, completed game frames, a browser stall or sustained simulation convergence. Preserve retrieved/input-state.jsonl rather than substituting the historical fixed-clock CLI spiral for this observation.

The controller and Chrome exited 0 with input released, server closed and profile removed. All 111 evidence files were retrieved and SHA-verified before the explicitly delegated shutdown of own boat bx_d5ebwy9z (original expiry 2026-11-02T19:12:47.558Z); box-final.json records its final state. Forty-five PNGs passed chunk CRC verification. The guarded boat/transfer/run phase kept owner /home free above 2000000000 bytes (minimum 2034028544). During the earlier scratch-only predecessor wait, unrelated staging reduced observed free disk below that floor; preparation-disk-limit.json preserves this limitation. Continuous whole-task floor compliance is therefore not claimed.

Recovery: use this run's RECOVERY.txt, manifests and retrieved captures. Local ignored fixture/module bytes require the pinned local archive or existing installed files; Git alone does not recover them. A future task should localize the unreadable presentation from these observations before another original run. Browser battlefield interaction and convergence remain unverified. No public registry change is warranted.

Restored display format repair — 2026-11-02

Run: scratch/runs/20261102-braveheart-presentation-cause-next/. This localizes and repairs a generic format disagreement observed during the menu to battle transition. It does not establish that the unreadable battlefield is repaired, nor verify unit commands, frame completion or convergence.

The saved baseline image and mission brief were reviewed first. Its trace lacks surface descriptors and Unlock counts, so neither a pitch defect nor absent completion follows from its 623/620 counters. Guest disassembly identified the retained offscreen pointer setup at 0x48d6fc..0x48d734, presentation selector [0xda3108], and descriptors at 0xda49fc / 0xda4a68. A primary Lock followed by an offscreen Lock clears the emulator's scalar present_lock_whole; that is a static counter limitation, not a diagnosis of the baseline battlefield.

After GL69793 and Oni96725 terminal results, final files, exact AUTHOR RELEASEs and scoped helper absence, one bounded read-only native diagnostic used the registered original, default WebGL renderer, real clock and threads. Source snapshot 3f8e6102c613c6fcc07fc58fc5661e40605faa1a matched all 3936 accepted inputs and all 145 fixtures. The browser used the baseline module 307168f9e0475fd3c10f87cbc5d98f05006c6d4494d2f564381a9a0ebe289fe6. No guest writes, patches, injected state, forced renderer or clock changes were used. This diagnostic added surface descriptors, per-row byte hashes, current mode and shared dirty/lock samples to the original route.

The reviewed menu/mission surfaces were 800×600 RGB565, pitch 1600. During Conduct Raid, mode-active became zero, the guest's depth variable [0xda4b44] became 32, and the selected presentation function became 0x4fe980 (Flip). Its newly retained offscreen pointer instead referred to a 640×480 RGB565 surface with pitch 1280 and masks f800/07e0/001f. presentation-state.jsonl retains the literal descriptors and surface table. The guest's windowed initialization at 0x4fe1cc..0x4fe203 obtains depth from GetDeviceCaps(BITSPIXEL) × GetDeviceCaps(PLANES). The emulator's GDI current-depth query correctly returns the restored 32-bit desktop, while DirectDraw's default allocation used the remembered 16-bit request. The 16-bit descriptor's pitch is internally consistent: this is a demonstrated current-format disagreement, not a pitch diagnosis from a screenshot.

The controller closed during asset loading, before the noisy battlefield was re-observed. Its neutral captures were black with zero simulation steps, so this run supplies no new evidence about battlefield completion or a stall. The original was not replayed after the repair.

The repair makes GetDisplayMode's pixel format/pitch and default CreateSurface formats use the existing current-display depth helper, also used by GDI. Remembered explicit mode values remain intact for repeat-mode notification suppression; existing surface formats and explicit offscreen formats remain intact. Display dimensions, presentation selection, renderer, clock, route, original files and api_table.json were not changed.

test/test-directdraw-restored-display-format.js emits literal expected/actual descriptors for six transitions (8/16/32-bit selected modes, restored 8/32-bit desktops), including two shared-memory instances, odd-width row alignment, primary/offscreen Locks, explicit and existing formats, and repeat-mode notifications. The exact final test failed 12 format checks on old source (exit 1) and passed on repaired source (exit 0). Four existing regressions also passed: GetDisplayMode depth, palette format, GDI 16-bit layout and shared fullscreen state. flock /tmp/wine-build.lock bash tools/build.sh passed on the own boat. Fixed module: 2071475 bytes, SHA-256 6fbd12aef206e346c0bde3688fc8b2ee94d410810613f2291410f388cbf9bda8. The first build attempt lacked two generated ToyVM bundles; regeneration matched their local hashes before the successful full build. The boat had no Git HEAD for the append-only API-table comparison; the local HEAD-aware API gate separately passed, and the unchanged table hash was checked on the boat. Preliminary test/transfer mistakes are retained in the run, including an invalid "new" trial that actually ran old source after a transfer argument limit. Only the SHA-verified final old/new trials support the repair result.

The original browser evidence comprises 86 retrieved and SHA-verified files and 27 CRC-checked PNGs. The final build/test evidence comprises 51 further verified members, including the fixed module. Recovery inputs and manifests remain in this ignored run directory. Own boat bx_ujxkt2u2 had original expiry 2026-11-02T19:30:40.054Z, recorded before use; cleanup and final state are recorded in cleanup.json and box-final.json. Evidence verification preceded the explicitly delegated shutdown. The transfer/runtime/retrieval guard enforced owner /home free >2000000000 bytes; its minimum was 2030542848 bytes (local-disk-final.json). No foreign fixture cleanup, downloads, workers or public deployment occurred.

Next: a separately dispatched changed-module browser observation can determine whether this generic repair makes the battlefield readable. A harness pass is not gameplay acceptance; the private lane still needs that observation.

Repaired-module native observation — 2026-11-02

Run: scratch/runs/20261102-braveheart-restored-format-gameplay/. The restored-format repair is active, but this bounded original run still reaches an unreadable post-loading image. Readable battlefield gameplay, Take the Field, unit selection and ordinary commands remain unverified. The remaining presentation cause is not localized; no further repair or original replay was attempted.

Admission followed GL22106, then Oni71461 actual terminal results, final files, exact AUTHOR RELEASEs and named/scoped helper absence. Their 291 and 168 final artifact hashes were independently checked. The predecessor repair's 272 final hashes and published commit were also checked. Scratch-only gate polling remained live at 41-second intervals before admission.

The browser used immutable source from published repair commit ab3596819df988f9eea258604f66a0104eea7d36, paired with its tested 2071475-byte module SHA-256 6fbd12aef206e346c0bde3688fc8b2ee94d410810613f2291410f388cbf9bda8. All 396 runtime-source and 1005 compiler/source records matched the predecessor identities with the published repair applied. The current checkout's 3936 accepted-baseline inputs were separately rehashed: the repair and subsequent GL changes are explicitly recorded in accepted-inputs-current-refresh.json.gz. Those later GL changes were not mixed into this observation. The shared build was left unchanged. No new production code or build was needed for this documentation-only task.

All 145 installed original files matched before and after use. The registered bhsoft.exe, six real DLLs, default WebGL host renderer, real clock and guest threads were retained. API-table SHA-256 remained 8319e5f866cf102879f4debfff34a421924a8a71d87cacf8d56c5a66afaaead5. Local and public DESKTOP_APPS refreshes both excluded braveheart_demo (46 local / 44 public entries). Source was streamed directly from immutable Git objects and fixtures from their installed paths; no local archive or fixture copy was created. All 4015 streamed members matched exact hashes on the own boat before launch.

Reviewed route: visible movie → balanced Escape → Game Features → Back → main menu → Raid Town → Stephen Feherty mission brief → Conduct Raid. One Back press initially missed because canvas coordinates omitted the 34-pixel page header; its before/held/released/settled captures and trusted events are retained. The corrected native page-coordinate press reached the menu. Menu entry is historical capability, not new gameplay progress.

06-conduct-raid-before, -held, -released, -settled preserve the entry sequence. At settled entry, mode-active was zero, guest depth was 32, and the new offscreen descriptor was 640×480, 32-bit, pitch 2560, RGB masks 00ff0000/0000ff00/000000ff; the primary surface was also 32-bit. Thus the old restored-depth disagreement was absent in this observed allocation.

Eighteen seconds of neutral waiting then produced the first concrete blocker: a noisy magenta strip across the top with the lower area black (07-loading-neutral). No gameplay input followed. After another six seconds of neutral settling, personally reviewed 08-final-neutral-blocker still had that unreadable presentation. The final retained descriptor and both surface records were 1024×768, 32-bit, pitch 4096, with the same RGB masks. The literal guest dimensions/stride/depth tuple was [1024,768,3072,32]; this observation alone does not diagnose the remaining encoding or presentation defect. The older second guest descriptor still described its existing 800×600 RGB565 surface; the repair intentionally preserves existing surface formats.

Read-only simulation samples changed from zero during loading to 6 and 8; the final delta ring was [158,150,141,141,157,140,116,126]. Shared dirty samples changed from 11321 to 11799 to 12579, with differing offscreen row hashes. Frame-end/pump counters stayed 934/931. The final shared lock count was 1. These are retained observations, not evidence of readable completed battlefield frames, displayed FPS, a browser stall or sustained convergence. The preserved scalar completion-counter limitation still applies.

Four mouse pairs and one Escape pair produced 14 trusted native events. Actual holds were 574/809/848/947 ms for the mouse and 559 ms for Escape, including capture overhead. Each pair has reviewed before/held/released/ settled page captures. Final input had an empty queue, no active event, zero mouse buttons, and Escape false in both physical and async key state. Controller and Chrome exited 0; server/profile and named/scoped remote helpers were absent.

All 120 evidence files (11574746 bytes) passed remote/local SHA comparison and an independent local reread; all 49 PNGs passed chunk CRC checks. This verification completed before the explicitly delegated shutdown of own boat bx_qcyyf29j. Its original expiry 2026-11-02T21:03:15.325Z was recorded before use; STOPPED was confirmed at 2026-11-02T20:11:28.467Z. shutdown-admission.json, retrieval-verification.json, retrieved/remote-helper-cleanup.json and boat-state.json preserve ordering. The guarded transfer minimum was 2043191296 bytes and retrieval minimum 2020298752 bytes, both above the required 2000000000-byte owner disk floor. A separate shared-disk excursion during the earlier scratch-only wait is retained in preparation-disk-limit.json; whole-task floor compliance is not claimed.

Recovery and exact member identities are in this run's RECOVERY.md, compressed input manifests and retrieved evidence. Git recovers the published runtime source, but not ignored fixture/module bytes. The next bounded task should localize the remaining unreadable presentation from these descriptors, rows and pixels before another original run. No public promotion is warranted.

2026-11-02: current-mode geometry repaired; producer contract localized

Run scratch/runs/20261102-braveheart-restored-pixels-cause/ adds a bounded startup/Conduct Raid diagnostic to the preceding restored-format observation. It does not establish a readable battlefield or unit selection/command. The single subsequent changed-source native observation remains a magenta strip over black. Braveheart remains excluded from DESKTOP_APPS.

Ordered GL64778 then Oni93804 gates were independently checked at live 41-second intervals: actual terminals zero, result/final, exact AUTHOR RELEASE, matching 209/70-file final manifests and absent named/scoped helpers. Only fresh scratch was owned until both released. Main refreshed to 2317cfdec69bfe4e19b8ba34bbc016dcab80b1a0; foreign/laptop WIP and the untracked test/test-directdraw-current-display-size.js were preserved.

The new diagnostic resolves these previously missing facts:

The tuple at 0xda4b38, finally [1024,768,3072,32], is width, height, height×bytes/pixel, depth. Original 4fe2de..4fe2e9 proves that 3072 is not row pitch. Actual row pitch comes from the descriptor and producer context. The observer reads direct/DIB/sparse mappings in JavaScript, without invoking translation exports that can mutate guest fault/lazy-fence state. It forwards all original imports and API results, bounds its event collection to 256 records, and resets only its own records before Conduct Raid. Entry/exit stacks, descriptors, mode entries, pointers, rows and HRESULTs are preserved; EnumDisplayModes' dispatch exit is explicitly marked a deferred continuation, not misreported as its final API return.

The runtime repair is narrowly the demonstrated current-display query error. GetDisplayMode now uses the same active/restored screen dimensions as GDI screen metrics and computes its DWORD-aligned pitch from that width. It retains the remembered SetDisplayMode request, existing surface formats and current depth/masks. This follows the documented GetDisplayMode current-mode contract. It does not claim to repair the independently observed 16-bit producer selection or normal-primary resize behavior.

test/test-directdraw-current-mode-geometry.js runs the real compiled handlers in two instances sharing process memory. Unchanged source fails ten inactive queries; all three active-mode semantic controls pass. Changed source passes all thirteen initial/active/restored/resized cases, including 8/16/32-bit selection, restored 8/32-bit desktops, odd-width DWORD rows, descriptor bounds, invalid input without host calls, unchanged remembered requests and active modes ignoring a host resize. The restored-format regression's obsolete 800-wide pitch expectation now matches its actual 1024-wide host desktop. Seven focused regressions pass: restored display format, GetDisplayMode depth, windowed primary size, primary mode change, palette format, Surface3 descriptor and Surface4/7 ABI. The Surface3 first run lacked its existing calc.exe test fixture; that failure is retained, then the exact 94208-byte fixture was streamed and both ABI tests passed. Locked full flock /tmp/wine-build.lock bash tools/build.sh passes on the own boat, compiling module 2076989 bytes, SHA-256 17e683cb60cd7d067753b1c988b5f6c5faa7066898aee82e87e6e6b9e9914496. The API table remains 4602 entries, SHA-256 8319e5f866cf102879f4debfff34a421924a8a71d87cacf8d56c5a66afaaead5.

Exactly one changed-source original registered bhsoft/default-WebGL/threads/ real-clock native observation followed that build. It used the unchanged 145 installed files (207704346 bytes), bhsoft.exe SHA-256 70ae3feedc1f977ca8b8ed32560ebdaa381a25b6261f2cbc77396c887e2a8955. Native Escape → Game Features → Back → Raid Town → Stephen Feherty mission brief → Conduct Raid was captured before/held/released/settled. The corrected initial GetDisplayMode now reports 1024×768×32, pitch 4096. The neutral field capture at 21:47:00Z remains unreadable; no gameplay inputs followed. Its producer/surface disagreement and NOTFLIPPABLE returns remain captured. Three mouse pairs and one Escape pair were trusted native events; final queue was empty, no active event, mouse buttons zero and Escape false in physical and async state. No FPS, audio, completed battlefield, unit command, stall or convergence claim is made.

The next causal question is why the original chooses its stale 16-bit windowed entry before restoration and retains the fullscreen presenter on that path. Capture the exact GDI return/cache-write and relevant window-message order before any further repair; the current bounded observer did not record those GDI returns. A native Windows reference may be needed to distinguish an application limitation on a 32-bit desktop from another runtime ordering error. Do not replay an unchanged original merely to rediscover the magenta strip.

Own boat bx_mx6rktxr, original expiry 2026-11-02T23:29:25.950Z, was recorded before use. Source/fixture transfers were direct hashed streams with no local archives/copies. The initial locked build omitted two existing browser-bundle inputs; its failure was retained, both exact inputs streamed, and baseline and changed locked builds passed. All 200 retained remote evidence files passed SHA comparison and independent reread before the explicitly delegated own-boat shutdown; 46 retained PNGs passed chunk CRCs. All native canvas before/held/released/settled captures are retained. A first retrieval was intentionally stopped at its actual 143 result after inventory showed that retrieving every redundant page PNG would consume the floor reserve. The selected manifest records 30 omitted page PNGs; retained canvas captures and all causal/input/build/identity evidence were verified. Identical immutable files within this own run were hardlinked only after byte/SHA equality, preserving every retained path. No disk-floor latch occurred.

Both controllers and Chrome exited zero; eight named helper references and all scoped helpers were absent before shutdown, both profiles removed, and transfer chunks/archive absent. Own boat STOPPED was verified at 2026-11-02T21:50:54.272Z, error null, after retrieval verification. Exact inputs, hashes, bounded hooks, failed/pass logs, selected evidence manifest, shutdown ordering and recovery recipe are in this run's RECOVERY.md and result artifacts. No public deployment or guest/renderer/clock patch occurred.

2026-11-02: causal observer preparation; runtime dependency remains

Run scratch/runs/20261102-braveheart-causal-return-trace/ audits the released 20261102-braveheart-windowed-producer-contract observer draft. No original, browser, compiled handler, boat or transfer ran. Actual GDI cache-update returns and window-message order remain unobserved; the first established mismatch remains the earlier 640×480 producer16/pitch1280 versus surface32/ pitch2560, before the later resize. No repair or readable battlefield is claimed.

The audit found concrete observation limits: the draft retains only twelve message records per call site; repeated names can exhaust its name inventory; and dispatch-handler exit can precede a redirected guest callback's completion. ShowWindow and positioning callbacks can enter a wndproc without a guest DispatchMessage/SendMessage API call, so logging those APIs alone cannot prove complete message order. These gaps do not establish why the older trace missed GetDeviceCaps.

A fresh scratch draft records distinct names, invocation IDs, raw exit stacks, explicit handler-boundary labels and dropped-record counts, with a 1 MiB payload cap. Eight synthetic scenarios pass, including supplied BITSPIXEL/PLANES values, COM-name recovery, deferred enumeration labeling, memory-read immutability and saturation. These are observer tests, not actual guest/API observations. The draft still needs compiled-handler/callback validation, a proven wndproc-entry observation boundary and protection against losing critical events at the cap. See OBSERVER-PREPARATION.md, static-inputs.json and observer-test.json.

The serialized Oni successor's fresh disk observation again breached 2 GB before transfer and identified temporary plugin-catalog writes; this is a new dependency, not a replay of an old guard latch. Its terminal/release verification is retained in this run's gate records. A successor needs enough stable reserve for those transient writes, then refreshed original/compiler/API/module pins and local/public exclusion before one bounded causal observation. No unchanged visual attempt was repeated, and foreign/laptop WIP was preserved.

2026-11-02: compiled observer validated; original blocked in DLL initialization

Run scratch/runs/20261102-braveheart-causal-observer-resume/ validates the observer against actual compiled handlers, then records one new original startup failure. Actual Braveheart GDI cache-update returns and window-message order remain unobserved. The established first mismatch remains the earlier 640×480 producer16/pitch1280 versus surface32/pitch2560 observation.

After the Oni CLI dependency task's actual terminal zero, result/final, exact AUTHOR RELEASE and scoped-helper absence, the worker refreshed 145 original files (207704346 bytes), 396 runtime inputs and 1005 compiler inputs. The snapshot was main 3fb0103a832464ceabcc8fed243499644d229cb5; all 10182 streamed members passed remote SHA verification. An initial stream stopped on concurrent TODOS.md drift; that unrelated coordinator-owned status file was excluded before restarting the transfer. No runtime or fixture substitution occurred. The locked full build passed, producing the same 2076989-byte module SHA 17e683cb60cd7d067753b1c988b5f6c5faa7066898aee82e87e6e6b9e9914496 as the earlier geometry repair. The shared local module was only hashed and was not selected or replaced. The API table remained unchanged.

The generated-PE test runs real GetDC/GetDeviceCaps, class registration, CreateWindowExA, ShowWindow, MoveWindow and SendMessageA. Immediate BITSPIXEL32 and PLANES1 returns reach independent guest stores. Its wndproc independently records every received message; all ten entries exactly match the observer's sequence: 81,1,1c,6,7,3,5,46,47,8001 (hex). This includes synchronous MoveWindow callbacks before handler exit and SendMessage's redirected callback after handler exit, followed by the caller's actual EAX42. All 29 observer records survive, with zero dropped records or unpaired exits. These are compiled generated-fixture results, not original-game GDI values. The collector aborts on saturation or observation errors; API-boundary cache samples still do not identify an exact x86 store instruction.

The single registered bhsoft/default-WebGL/threads/real-clock native attempt then failed at 2026-11-02T22:41:19.071Z. Worker slot 0 reported stage loadDlls, sequence 21, DLL initializer did not return: module=0x102e000. The initial isolation/worker/configuration gate passed, but the worker retired before the controller's first observer snapshot, which contains no workers. No menu, GDI returns, cache transition, complete original message order or producer mismatch was captured. The canvas personally reviewed is the launch failure report, not gameplay. No native inputs or unchanged retry followed. The page also reports unavailable stdole2.tlb and missing ole32.dll; neither warning establishes the initializer's cause. The module name, initializer outcome/cause and original observer counters were not retained before worker retirement, so this is not a diagnosed runtime or observer defect.

The next task should preserve the failing initializer's identity, outcome/cause and observer state before worker teardown, without repairing shared runtime or repeating the visual route. Compiled callback validation now has evidence; complete original startup/causal coverage still does not.

All 47 selected remote evidence files (2037416 bytes) passed remote/local SHA comparison and an independent reread before own-boat shutdown; four PNGs passed chunk CRC checks. Controller exited 1, Chrome exited 0, and server/profile, scoped helpers and transfer chunks/archive were absent. The operation wrapper's zero exit is not a successful browser result. Own boat bx_dp74u33t, original expiry 2026-11-02T23:06:52.933Z recorded before use, was verified STOPPED at 2026-11-02T22:44:00.637Z, error null. Guarded operations stayed above 2000000000 bytes; the minimum was 2028453888 bytes during transfer. Recovery, compiled events, startup error, exact inputs and lifecycle ordering are retained in this run. No shared runtime repair or public deployment occurred.

Initializer-boundary capture preparation stopped by disk guard (2026-11-02)

Run: scratch/runs/20261102-braveheart-initializer-failure-capture/. The original DLL identity and initializer cause remain unobserved. This run made zero build, generated-control, browser or original-game invocations. Its new blocker is a retained disk-floor breach during source/fixture transfer, before a remote checkout existed.

Static inspection found an existing evidence-loss boundary: loadDllsOwned already logs each DLL name, mapped base, DllMain address and original base; dllMainSteps logs entry/result/halt and retains an outcome including faults. finishDllInitialization attaches that outcome and its fault as the error cause. However, the worker's loadDlls handler supplies a no-op logger, and its command catch sends only error.message. Those paths explain why the predecessor's module 0x102e000 report lacks the available loader identity/outcome. They do not establish the DLL's identity or diagnose its original initializer failure.

Scratch-only capture code was prepared to forward the real loader call while retaining module/initializer boundaries, live registers, exception/outcome chains and bounded observer records/counters before the worker error. Generated compiled success, false-return, bad-stdcall, NULL-jump, real trap and deliberate observer-failure controls, plus callback and browser-transport controls, were prepared but not executed. This diagnostic is unvalidated and must not be used for an original invocation until those controls actually pass. No shared runtime source or API behavior was changed.

The GL boundary task and then Oni residency task had actual terminal zero, result/final, exact AUTHOR RELEASE and absent scoped helpers before transfer. Their 208 and 197 final member hashes, and the predecessor's 108, were reread successfully. The pinned snapshot was main 3785da2fafee25f8de748277051ab7e59646ad76: 396 runtime and 1005 compiler inputs matched Git, and 145 local original fixtures (207704346 bytes) were hashed. API-table SHA was 8319e5f866cf102879f4debfff34a421924a8a71d87cacf8d56c5a66afaaead5. The existing shared WASM was identified but neither selected nor replaced; there is no newly compiled module or completed remote 145-file verification.

At 2026-11-02T23:25:29.179Z, the foreground transfer guard observed 1991254016 free bytes on /home, below the required 2000000000-byte floor, and killed its own transfer process group (exit -15). The guard latched, blocking subsequent heavy phases. Free space recovered to about 2.045 GB, but the roughly 55 MB transient loss was unexplained and exceeded the available headroom. Repeating under that headroom would not preserve the required floor. No foreign files were cleaned. Source and fixture bytes were streamed directly to the boat; no local archive or fixture copy was created. A separate 7.218-second monitor rotation gap during earlier read-only preparation is also retained. The 250 ms passive watcher missed the short breach; the 100 ms operation guard retained it. Continuous floor compliance is not claimed.

Own finite-TTL boat bx_amntdhxb, original expiry 2026-11-02T23:54:13.677Z, was recorded before use. Its 610 partial chunks (48800000 bytes) had not been extracted. Chunk identities and absent helpers were retrieved, SHA-compared and independently reread before cleanup. Cleanup metadata was likewise retrieved and reread before the explicitly delegated shutdown. The boat was verified STOPPED at 2026-11-02T23:29:04.071Z, error null, with its own chunks/archive and scoped remote helpers absent.

Next: obtain stable additional coordinator-owned disk headroom, make a fresh guarded transfer, compile the pinned source and validate the diagnostic with actual generated controls. Only then admit at most one changed-diagnostic original startup. No original GDI mismatch, initializer cause, observer-fault classification or gameplay progress was established here.

Recovered-headroom follow-up stopped before transfer (2026-11-03)

Run: scratch/runs/20261102-braveheart-recovered-headroom-initializer/. The original initializer identity and cause remain unobserved. No source or fixture transfer, build, generated DLL control, browser or original startup ran. The prepared initializer/callback/browser diagnostics remain unvalidated; no runtime repair or gameplay result follows from this run.

After the ordered GL and replacement Oni terminal results, final files, exact author releases and absent scoped helpers, their 587 and 478 final artifact hashes matched. The predecessor's 73 also matched. Fresh pins cover 396 runtime and 1005 compiler inputs at 024c732b07dd7a44c1822f5ad3eb9c4c94e12c77, plus all 145 local original files (207704346 bytes). The shared WASM was identified only; no newly compiled module was produced or selected. The worker still supplies the loader's no-op logger and forwards only the shallow error message.

Four small process controls passed: normal completion, supervisor death, supervisor stall, and child-watchdog death. The negative cases stopped only their synthetic commands. These validate monitor termination behavior, not DLL initializers or browser capture. A foreground supervisor with a separate child watchdog then owned the actual preparation phase, persisting periodic heartbeats, minima and a retained abort latch.

The proposed single SSH stdin stream failed its small preflight with HTTP 403, api_key_action_forbidden / This API key cannot perform ssh. No fixture bytes were sent. A signed-HTTP-upload alternative was prepared but never executed. At 2026-11-03T00:57:46.656994Z, the foreground guard latched a disk-floor breach and terminated its own runner (exit -15). Its minimum was 1999269888 bytes; the independent child recorded 1985236992 bytes, both below 2000000000. Maximum sampling gaps were 0.949 and 0.246 seconds. The transient's cause is unproven. Recovered headroom did not establish stability; neither this latch nor the prior 1991254016-byte breach was cleared or retried. Earlier passive preparation gaps are retained separately; whole-task continuous floor compliance is not claimed.

A separate bounded recovery-only supervisor retrieved and independently SHA-verified metadata proving the own boat had no checkout, transfer files or scoped helpers. This preceded the explicitly delegated shutdown of bx_59uhh53n, confirmed STOPPED/error null at 2026-11-03T01:00:20.124Z. Its original expiry 2026-11-03T01:55:23.978Z was recorded before first use. RECOVERY.md, the abort records and the recovered metadata preserve the restart boundary. The raw creation-output filename collision is also retained; the parsed creation-ready record and original expiry were fsynced before use.

Next: stable additional headroom and a validated permitted transfer route, then a fresh pinned build and actual initializer/callback/browser controls before admitting any original startup. No foreign files, active catalogs or caches were cleaned. Git alone does not recover the ignored original fixture.