OpenGL / Direct3D corpus status, updated 2026-11-18

November 18 capture, input and shutdown prerequisites

Browser startup fix 8fd08ea3b adds --bw-timezone=Etc/UTC to the private browser helper. A traced BetterChromium timezone network probe had consumed the 15-second launch bound; explicit UTC passed with that bound unchanged. The same author's real production browser Worker passed malformed-record, queued CLOSE, same-ID replacement and STOP checks, including pixels, canaries, barriers and target destruction. This is browser evidence, distinct from the earlier Node Worker result. The coordinator reviewed two original software Tunnel frames with changed geometry. Ordinary title close was observed, but native-exit logging and individual Chrome/display waits were lost after a harness failure; absent PIDs and supervisor cleanup do not fill those gaps. Initial black output changed after an inspection resized the viewport and an ordinary focus gesture, so no isolated renderer fix is established. The run also records disk-floor breaches. Neither full-27 nor crash freedom follows. Evidence: scratch/runs/20261118-d3dim-browser-startup-cause-shutdown-original/result.json.

Follow-up c7ac59e41 adds the actual-browser regression to tracked tests; report 0a55fecd6 also closes the software Tunnel's native-exit and host-wait gaps in a new scope. The coordinator inspected the native ExitProcess(0) record before host stop, the exited process state, guest/render retirement and STOP acknowledgement. Chrome, Xvfb and openbox have individual actual waits with both EOFs; the original driver and enclosing command completed 0 without leaks. This does not repair the predecessor's missing receipts. The new original had early black output before a later textured client, so no new motion or isolated rendering-fix claim follows. The global disk floor crossed after driver cleanup; whole-task disk compliance failed. The WebGL counterpart was still open in that run; arbitrary browser/OS crash freedom is not established. Evidence: scratch/runs/20261118-d3dim-browser-crash-regression-original-close/result.json.

The WebGL follow-up (dd9b09da5, report 72cd36556) passes the tracked production-browser regression on both executors. The original's actual GpuStreamWebGL endpoint records WebGL 1 context activity, 810 draws and 228 presents in the retained later observation. The coordinator reviewed the changed textured Tunnel geometry and inspected native exit 0 before host stop, exited process state, retirement and individual Chrome/Xvfb/openbox waits with both EOFs. These are backend and shutdown observations, not a physical-GPU or performance claim. Original-run samples stayed above the floor; whole-task disk compliance failed before/after that run. The retained original module and authored regression module have separate source/hash records. Full-27 coverage and arbitrary crash freedom remain open. Evidence: scratch/runs/20261118-d3dim-webgl-browser-crash-close-tunnel/result.json.

Globe ordinary close is repaired by 1ad9afd92 (report fa13214c8). The fifth COM release yielded through FreeLibrary to the idle page CPU, abandoning the native WM_DESTROY callback. Servicing that unload on the owning Worker completes all six releases and permits native exit. The native old-fail/new-pass fixture, browser unload controls, API batching and locked build passed. For both software and WebGL originals, the coordinator inspected ExitProcess 0 before host stop, exited process state, guest/render retirement and individual Chrome/Xvfb/openbox waits with both EOFs. The repaired scopes do not erase the earlier UNKNOWN closes. They use the retained a6002d5d… module with the recorded RPC/source graph; sampled original minima above the floor do not establish continuous disk compliance. Full-27 and arbitrary crash freedom remain open. Evidence: scratch/runs/20261118-dx-globe-native-close-cause-repair/result.json.

Viewer report ee3cf7db8 records the original supplied Microsoft model moving on actual GpuStreamSoftware and GpuStreamWebGL executors at stable 800×600. The coordinator reviewed the software whole-frame pair showing rotation; the author reviewed both pairs. Completed draw/present records and a live WebGL context support the backend observations. Ordinary title close reaches native ExitProcess 0 before cleanup on both routes, with guest/render retirement and individual driver, Chrome, Xvfb and openbox waits 0 with both EOFs. The first WebGL observer scope remains UNKNOWN and closed; a corrected active executor observer supplied the passing scope. This reuses the recorded a6002d5d… module and 1ad9afd92 RPC graph without a product or media change. Sampled original disk minima do not prove continuous compliance. Hardware, performance, Windows parity, full-27 and arbitrary crash freedom remain open. Evidence: scratch/runs/20261118-dx-viewer-current-browser-paired-motion-close/result.json.

Twist report 671bb401e adds author-reviewed moving textured frame pairs on actual software and WebGL executors with stable geometry. Software's initial black frame is excluded: its visible pair is later/settled, while retained executor counters cover the earlier observation interval. Both ordinary title closes reach native PostQuitMessage and ExitProcess 0 before host stop, followed by guest/render retirement and individual driver/Chrome/Xvfb/openbox waits 0 with both EOFs. No product or media change was needed. This is named-app motion and shutdown evidence; full-27, parity and performance remain open. Evidence: scratch/runs/20261118-dx-twist-current-browser-motion-native-close/result.json.

Boids/Flip3DTL reports dcec6ff5e and 685d5f5a6 add four author-reviewed settled moving scene pairs: each registered original on actual software and WebGL executors. Completed draw/present records identify the active endpoints; ordinary title close reaches native ExitProcess 0 before host cleanup in all four scopes. Guest/render retirement and all sixteen individual driver/browser/ display-manager waits complete 0 with both EOFs. Two earlier preparation scopes remain UNKNOWN and closed; failed readback-count assumptions are observer defects, not demonstrated product failures. No product/media change or recompile was needed. The disk floor was breached during finalization after the originals closed, so whole-task disk compliance failed. Full-27, Windows/backend parity, hardware, performance and arbitrary crash freedom remain open. Evidence: scratch/runs/20261118-dx-boids-flip-current-paired-motion-native-close/result.json.

Architecture report 1960ec53d adds author-reviewed moving scene pairs on actual software and WebGL executors. Ordinary Escape reaches native ExitProcess 123 before host stop, followed by guest/render retirement and individual driver/browser/display waits 0 with both EOFs. This records the exact guest status; it does not classify 123 as a normal Windows exit. Wormhole also shows moving pairs and ordinary Escape/native exit 0 under both selectors, but both execute native DirectDraw/Canvas2D with no D3D or WebGL executor. Those Wormhole observations do not fill a D3D/WebGL coverage cell. Four earlier UNKNOWN scopes remain closed, including an external exit 143 with unknown sender and missing waits. No product change or recompile was justified. Evidence: Architecture/Wormhole report, run 20261118-wormhole-architec-current-paired-motion-native-close.

Hexen II reports 4a78e3fe3 and ae6351986 verify ordinary native Pause/resume and menu Quit with guest exit 0 before host cleanup on the retained WebGL graph. Three finite Ctrl gestures leave both spiders upright; enemy damage and defeat remain unproved. Receipt delivery alone does not establish that the active game binding performed an attack. Unique selected images/logs were preserved in tmpfs when disk space prevented full durable retrieval; minimal receipts retain the successful exit and cleanup evidence. The actual module object was retained, but executing bytecode was not independently recovered. This adds a shutdown observation without completing combat or full-27 coverage. Evidence: Hexen II Pause/combat report, run 20261118-hexen2-native-pause-living-enemy-combat.

ReVolt report 3026b89f9 advances beyond menu navigation: personally chosen Up and Up+Left holds produce road-to-curb movement and changed heading at 8 mph, which the coordinator reviewed in the retained whole frame. Review ages are 3904/4438 ms within the unchanged five-second limit, with independent neutral releases. A separate promotion-aware Quit run observes the guest exited with code 4097. Driving and exit are separate scopes, and that nonzero code is not classified as Windows-normal. Executing-module/Worker-role attestation and outer invoker/EOF receipts remain missing; earlier UNKNOWNs and disk breaches remain recorded. No checkpoint or lap completion is proved. Published helper fixes 93ee0d25c/cce80c7c4 passed focused tests and locked builds. Evidence: scratch/runs/20261118-revolt-capture-review-handoff-driving/result.json.

ReVolt report e8feff64f reached manually selected Single Race, RC Bandit and Toys in the Hood 1. The coordinator reviewed the actual car and race HUD: lap 1/3, eighth place and 2 mph. Nine navigation gestures had matching independent neutral releases. Disk reserve permanently closed the scope before the first acceleration command, so 2 mph is not credited as player driving. Acceleration, steering/checkpoint consequence and ordinary Quit remain open. The earlier unnamed overlapping DIV did not recur and its identity remains unknown; no production visibility repair is claimed. Runtime samples stayed above the floor, but prelaunch and lifetime compliance gaps remain recorded. Evidence: scratch/runs/20261118-revolt-startup-dom-obstruction-manual-race/result.json.

Published f6962ba5b corrects capture-request freshness for a retained static display. A real browser receipt with a 32.761-second-old publication and a 43-ms-old request fails the old validator and passes the new one. Old requests and owner/output replacements remain rejected. Native tests and the locked build passed. The affected GLTron original stopped before input because two operator outputs used the same filename. A subsequent operator qualification used a GL startup predicate on a GDI fixture and also stopped before input. Neither run demonstrates original turns, a round outcome or Quit. Evidence: scratch/runs/20261118-static-display-fresh-capture-gltron-gameplay/result.json and scratch/runs/20261118-gltron-unique-capture-output-live-turns/result.json.

The later matching-fixture operator qualified actual native GL capture, trusted input, independent release and Quit before one unchanged GLTron original. That original reached the arena and reported WINNER 3, PLAYER 4 WINS and WINNER 2. The coordinator reviewed a whole arena frame showing PLAYER 4 WINS. Ordinary QUIT selection, credits and Escape produced guest exit code 0 before host cleanup; the coordinator inspected the retained process-exit state. Two visible input-caused turn effects and successful pause remain unproved, despite trusted A/S events and physical-UP receipts. A player-selection menu alone is not key-binding evidence. This run used f6962ba5b, module a6002d5d…, with all 92 media pins unchanged; independently executing module/roles remain unknown. Its 59 retained waits completed with both EOFs and 219 recorded PIDs were absent, but the outer tool143/enclosing-wait gap remains unknown. Evidence: scratch/runs/20261118-gltron-matching-fixture-operator-live-round/result.json.

The normal-pause follow-up (bc3af94e4, corrected report 06686105d) establishes the two manual WebGL turn effects. Space paused a live arena; Enter resumed it. The coordinator reviewed whole frames showing the same intact yellow rider's trail turn from rightward to upward after A, then add a second rightward corner after the independently chosen S. The pause held for 6.732 seconds. These were separate resume/turn/pause intervals, not a simultaneous Enter+A event or autonomous play. Quit Enter was NOT_SENT after disk-margin closure at 1,535,164,416 free bytes. This is separate from the earlier original's Quit 0, and the first authored scope's actual floor breach remains recorded. Both tasks used the unchanged 92 media and compiled/ served f6962ba5b/a6002d5d… graph; independent executing-bytecode provenance remains unknown. Software-backend gameplay is still a separate open outcome. Evidence: scratch/runs/20261118-gltron-normal-pause-manual-turn-effects/result.json.

The later actual software run (66b475a05) reached a textured arena with gl-gps=software, glRenderer=null and 4,430 native draws recorded for the original itself. The coordinator reviewed the intact rider and GAME IS PAUSED frame; the worker also records ordinary Enter resume. This initial launch pause does not establish normal Space pause or either manual turn. Disk closure preceded those actions and Quit, and later sampled free space fell below 1.5 GB. The earlier changed/cropped display refusal did not recur; the uncommitted readiness candidate remains unproved. No game exit or full-duration disk compliance is claimed. Evidence: scratch/runs/20261118-gltron-headroom-restored-software-gameplay/result.json.

Quake II report d0ed22840 records Game -> Easy and four manual world moves on unchanged 67 media/control rows and f6962ba5b/a6002d5d…. The coordinator reviewed the initial room and a later doorway view beside an existing corpse. There was no firing, identified living-enemy damage/defeat or ordinary Quit; the corpse is not credited as a kill. Shared free space reached 1,498,701,824 bytes, permanently closing the run with zero later requests. Actual driver, browser, supervisor and watchdog waits are retained; the outer-tool143/ enclosing-wait gap and independently unknown executing modules remain limits. The later audit confirmed five-second world-input freshness; the misleading 30-second refusal diagnostic was corrected. No combat or whole-corpus result is inferred from these movement frames. Evidence: scratch/runs/20261118-quake2-current-reviewed-browser-enemy-progress/result.json and scratch/runs/20261118-quake2-tmpfs-world-enemy-consequence/result.json.

A further original (c5d5bbc33) used the published 06f27181f renderer and admission changes with the unchanged a6002d5d… module and 67 media pins. Nine guest gestures, including seven world gestures, reached the damaged lower doorway; the coordinator reviewed the doorway frame with health 100. There was no fire, identified living enemy, combat consequence or ordinary Quit. Native readbacks validated, but the new completion-admission branch was not exercised, so this is not causal qualification of that branch. Disk UNKNOWN permanently closed the run with no later requests; the sampled minimum was 974,835,712 bytes. The original invoking tool's exit143 and missing enclosing wait remain explicit limits despite 23 retained runtime waits. Evidence: scratch/runs/20261118-quake2-headroom-restored-enemy-gameplay/result.json.

LTris startup helper 6fe4afd9e distinguishes an initial empty desktop publication from the first completed SDL window publication. Authored browser controls and native Quit passed; the original reached only its menu before disk preemption. That controller also sent its final DOWN with a 15.493-second-old readback despite a 15-second freshness limit. The subsequent send-boundary guard 75f495225 rejects stale DOWN locally and requires a new reviewed intent; its original-game result remains pending. No line clear, score consequence or original Quit is established by these helper tests. Evidence: scratch/runs/20261118-ltris-display-startup-readiness-line-clear/result.json.

ReVolt's qualified operator reached Select Race, with Single Race selected; the coordinator reviewed the whole host image. Enter followed release arming by 126 ms, and independent UP confirmed physical state zero. Disk closure preceded the next DOWN. Manual racing, driving, checkpoint and ordinary Quit remain unverified. Later post-cleanup free-space observations fell below the required 1.5 GB floor; the writer is unknown. Cleanup receipts do not prove that the floor held through all drains and later writes. Evidence: scratch/runs/20261118-revolt-qualified-operator-manual-race/result.json and scratch/runs/20261118-coordinator-revolt-select-race-review/result.json.

The coordinator reran the existing current-stream shutdown regression on the shared working tree: actual software WAT CLEAR/READBACK/PRESENT, ordered pending CLOSE/STOP, retired callbacks and cancellation of a real SYNC producer passed. The exact module was fbdaaad01060…, with unchanged relevant working diff before/after. Foreign WIP was included; this is not clean-published-module, browser/WebGL, original-game or whole-corpus crash-freedom evidence. Renderer exit 1 was explicit harness termination, not a spontaneous crash. Evidence: scratch/runs/20261118-coordinator-current-stream-crash-sweep/result.json.

November 17 NFS3 driving and paired SDK rendering

NFS3 inherited-timing report, published c9ff55052, records released acceleration and steering in the original: speed rose from 0 to 3 and 4 MPH, the wheel turned left, and road alignment changed. The first hold overlapped the countdown. The coordinator reviewed the whole after-steering frame at 4 MPH and delivered it as Telegram message3008. This is evidence of manual driving on the pinned Direct3D route, not a high-speed race or lap completion. The tested runtime remains 013d916c6, module cf674e8d…, with all135 media hashes unchanged. Escape expired unapplied at its unchanged release-bound deadline; Pause, Continue and native Quit remain unproved. Native and enclosing process waits and both pipe EOFs completed. The terminal input attempt stays closed. Evidence: scratch/runs/20261117-nfs3-inherited-timing-visible-driving/result.json.

A later NFS3 supervision and timing run, report a0a3cbaa8, observed released driving to16 MPH with steering and an off-road view. The coordinator reviewed the complete16-MPH frame; it was delivered as Telegram message3014. The original now supplies the missing timing: one synchronous gpu_stream_kick took2916.589 ms and blocked HTTP entry until2463.302 ms after authorship, beyond the unchanged800-ms Escape deadline. The handler did not apply and no native consumption occurred. All keys were subsequently observed UP with an empty queue, but that observation was1938 ms late. Pause/Continue/native Quit remain open. No scheduling repair was established by that run; the endpoint remains closed. The earlier unexplained SIGTERM sender remains unknown, separately from this measured render-starvation cause. Native/inner lifecycle waits completed; the final advisory watcher wait remains unverified. Evidence: scratch/runs/20261117-nfs3-supervision-signal-source-driving/result.json.

The asynchronous Node execution repair, code 9191d4afb and report 9b6594d8c, adds an opt-in render owner with separate native memory and a single consumer per shared ring. A real native 64-draw test reduces independent event-turn delay2642.46→0.876 ms while preserving pixels, order, sequence and arena retirement. Twelve focused suites, shared-consumer/producer controls and the locked build passed. The later native fixture still missed800-ms consumption because USER polling occurs once per 32-draw frame; that UNKNOWN remains closed despite prompt host application. The single affected NFS3 original exited during ProcessDesktop hello with ETIMEDOUT, before game execution or GPU kicks. Driving/Pause/Continue/native Quit remain unverified on this change. Native/supervisor/watchdog waits and separate EOFs completed with no leaks. A fresh worker investigates the actual startup cause; neither the event-turn control nor the earlier authored PAUSED caption proves original-game Pause. Evidence: scratch/runs/20261118-d3dim-async-execution-boundary-nfs3/result.json.

The rendered NFS3 follow-up, code 70062e2eb and report ddb0c62d9, fixes a CLI window-creation override that omitted process/lifetime metadata. Frame admission correctly rejected605 fixture presents despite matching native target/sequence/stamp. Publishing the same ownership metadata as the shared host-window path admits479/479 fixture presents without relaxing checks. Native rendered Pause/Continue/ordinary Quit, focused async/Flip tests and the locked build passed.

One original on the software/cooperative async-owner route rendered the race with572 accepted presents and44696 draws. The coordinator reviewed whole cockpit/race frames. Up/Left/Right/Escape received native USER-return receipts in87–258 ms, within the unchanged800-ms confirmation bound, and independent UPs were acknowledged. Acceleration/steering effects and Pause were not demonstrated; later1 MPH remains unattributed. Actual game control/DirectInput polls were not measured, so USER receipts do not prove game actions. Original Continue/Quit remain open. Owned cancellation and native/supervisor/watchdog waits with both EOFs completed without leaks. A fresh worker investigates the actual game-consumer route; earlier driving evidence remains separate. Evidence: scratch/runs/20261118-nfs3-rendered-continue-quit-driving/result.json.

Four SDK examples (112c6ef43) and the software Tunnel/Flip follow-up (53bce46ea) together record moving scenes for Boids, Flip3DTL, Tunnel and Twist on both actual software and WebGL GPS executors. These runs use pinned 70a680b54, module cf674e8d…; they do not establish current-main parity. The coordinator reviewed Boids bird/terrain motion, Flip3DTL textured-cube rotation and the fresh software Tunnel's changing textured scene. The fresh software Tunnel and Flip runs include complete enclosing waits and EOFs. The earlier software Tunnel capture refusal and earlier Flip software missing enclosing waits remain separate UNKNOWN results. These observations do not establish visual parity, FPS, audio, native Quit or the complete historical 27-app goal. Later Globe and Viewer results are recorded below. Evidence: scratch/runs/20261117-dxsdk-four-examples-paired-rendering/result.json and scratch/runs/20261117-dxsdk-software-tunnel-flip-rendering-finish/result.json.

November 17 Globe/Viewer window creation repair

Globe/Viewer report records published repair 3ab982f2d and report 58bca5d7f. A native counterexample showed WM_SIZE reaching a hidden overlapped window before CreateWindow returned and before Viewer initialized its device. The repair defers that size notification; popup, creation rejection, positioning, showing, child/startup neighbors and callback-review refusal controls passed, along with the canonical locked build.

Four serial affected originals used the published source graph and module 06eaa397…. Globe software and WebGL produced textured spheres with natural motion; Viewer WebGL produced a rotating extruded Microsoft model. Actual native D3DRM/Device2 and software/WebGL GPS execution were recorded. The coordinator reviewed the whole software Globe and WebGL Viewer before/after pairs. Viewer software produced one meaningful model image, but its later pair was blank after an attached inspector implicitly changed the viewport to 800×600. Its motion remains unproved. The scope is closed; a corrected non-resizing inspector requires a fresh run. Initial failures remain recorded separately. All nine scopes, including the authored viewport control, retained actual native/browser/helper/watchdog/enclosing waits and EOFs. No native Quit, visual parity, FPS, audio or full27 result is inferred.

Evidence: scratch/runs/20261117-dxsdk-globe-viewer-paired-retained-rendering/result.json.

Fresh software Viewer and ARCHITEC follow-up

Non-resizing inspection report, helper ae10d9c42 and report 9882330c2, adds a fresh software Viewer pair: the supplied Microsoft model rotates from its front toward its edge. A fresh ARCHITEC software pair shows textured architecture and a moving green mechanism. The coordinator reviewed both complete image pairs; ARCHITEC's later frame was delivered as Telegram message3012. The helper's focused geometry/generation/refusal tests, actual native GetPixel agreement and locked build passed. Both originals use 3ab982f2d / module 06eaa397….

ARCHITEC WebGL failed its pre-capture geometry guard: backing dimensions changed from 1280×960 to 640×480 and presentation became 1280×926 while the viewport remained 1280×960. The evidence does not establish whether ordinary startup or inspection caused that transition. No image or motion pass is claimed, and the scope closed without later endpoint queries. All four scopes retained native/browser/helper/watchdog/enclosing waits and EOFs. Original per-call Device2 traces were not captured; matching authored Device2 controls remain separate from original native-D3DRM/GPS evidence. Earlier software Viewer's blank captures remain separate from this successful fresh pair.

The startup geometry follow-up, helper 0aba2d17a and report fa6f4ae83, adds finite stabilization before inspection binding. An authored native startup reproduced the backing and presentation transition with a fixed viewport before secondary inspection. Completed presentation and stable geometry establish one baseline; bound geometry/generation guards remain strict, with never-ready/stable, stalled, post-bind mutation and rebind-refusal controls. Focused tests and the locked build passed. No dimensions are forced and no bound scope is rebaselined.

Three fresh originals on the same 3ab982f2d / 06eaa397… graph show ARCHITEC WebGL's rotating mechanism and GEOMETRY's gold shapes moving on both executors. The coordinator reviewed all three complete pairs; GEOMETRY WebGL's later frame was delivered as Telegram message3013. The historical closed ARCHITEC transition remains separate and its cause remains unassigned. Native, browser, helper, watchdog and durable enclosing waits/EOFs completed; the final caller shell's143 is distinguished from the durable owner's actual command0. Evidence: scratch/runs/20261117-rm-startup-geometry-architec-geometry-paired/result.json.

The JAZZ/ROCKROLL report, published 501b739688, adds JAZZ software and WebGL pairs with naturally rotating musical shapes on the same 3ab982f2d / 06eaa397… runtime graph. The coordinator reviewed all four whole images. Actual software/WebGL GPS draw/present execution and native RM/DirectDraw loading were recorded, with matching source/module/media hashes. No audio quality or parity was assessed. ROCKROLL on both backends reached the unchanged15-second prebinding timeout without a baseline or frame. Those cases closed UNKNOWN without later queries or replay; this does not establish a renderer/API fault. All four scopes have actual native/browser/watchdog/enclosing waits and separate EOFs. No production repair was made in that task. The author adopted JavaScript-only policy after the four launches; the follow-up investigation uses JavaScript supervision. Evidence: scratch/runs/20261118-rm-jazz-rockroll-paired-rendering/result.json.

The JavaScript startup follow-up, code a373de9ba and report d88c6963e, measured early debugger attachment slowing an authored native startup. Deferring attachment until after capture preserves the same finite startup and geometry guards; focused native controls and the locked build passed. Both fresh ROCKROLL originals admitted on the same 3ab982f2d / 06eaa397… runtime. Software produced two whole frames showing changing guitars and picks, personally reviewed by the coordinator; actual software DRAW counts rose388→895 and PRESENT counts12→31. WebGL received an external supervisor SIGTERM before its first capture. Its sender is unknown, the scope is closed, and WebGL motion remains unproved. No native Quit or complete separate original pipe EOFs are claimed. Evidence: scratch/runs/20261118-rockroll-js-startup-rendering-repair/result.json.

The named-scene follow-up, report d998c75f5, reused four historical Wormhole/Sci-Fi motion pairs and used its authorized fallback for one fresh ROCKROLL WebGL run. ROCKROLL bound in2786 ms and produced a changing guitar/pick pair on actual WebGL; completed DRAW counts rose4194→8309 and PRESENT counts188→295. The coordinator reviewed both whole1280×960 frames. All three loaded modules match 06eaa397…; native/browser/watchdog waits and separate EOFs completed with no survivors. This supplies ROCKROLL's missing WebGL motion pair without changing the earlier interrupted result. Historical Wormhole used DirectDraw/Canvas2D under both selectors, with zero active WebGL draws; Sci-Fi's historical pair exercised actual software/WebGL. Those older pins do not prove current-main behavior or complete27 membership. Evidence: scratch/runs/20261118-dx-wormhole-scifi-paired-motion/result.json.

The OrganicArt startup investigation, observer 47095dcb6 and report 1cd05040d, separates completed presentation from display composition and records bounded native progress. Native controls and the locked build passed. One software original first sampled a completed PRESENT at3568 ms and bound at4108 ms; earlier samples had PRESENT0 with composed UI and advancing native progress. Actual software execution recorded 10 PRESENTs and152 DRAW_UP calls, with all three loaded modules matching eb2395e1… and88 media pins unchanged. The coordinator reviewed the whole colored scene, including the unresolved lower-right white rectangle. Initial black-to-scene captures do not establish a motion pair. The earlier15-second failure did not reproduce and its cause remains unknown; this is an observer change, not a demonstrated emulator repair. Actual original native/command/ watchdog waits and separate EOFs completed without leaks. A fresh worker now targets software/WebGL natural-motion pairs. Evidence: scratch/runs/20261118-oasaver-bounded-startup-progress-cause/result.json.

The paired-scene follow-up, report 37118e0f0, adds a reviewed WebGL motion pair on that same pinned graph: silver objects move and rotate against an unchanged cloud background. The coordinator reviewed both whole frames. Actual WebGL PRESENT counts rose43→101 and DRAW_UP counts1117→2636; all three module identities and88 media pins match. Software hit the unchanged15-second prebinding limit without an admitted frame. That software request disabled startup observations, so independent presentation/composition predicates and actual instance inventory are unknown; zero PRESENTs cannot be inferred. A fresh worker uses the exact previously successful observed software route to target its missing motion pair. Both completed scopes retain actual native/command/watchdog waits and EOFs with no leaks. Software motion, scene parity, the white rectangle/logo issue, FPS, audio, native Quit and full27 remain open. Evidence: scratch/runs/20261118-oasaver-paired-natural-motion/result.json.

The software presentation-path observation, observer 054baf9e5 and report d5371571f, now adds the missing software motion pair. The coordinator reviewed both complete cloud-backed frames: left silver/blue and right dark articulated objects change orientation while the background and lower-right white rectangle remain unchanged. Actual GpuStreamSoftware completed PRESENT21→33 and DRAW_UP124→196 on the same owner/window/epoch; three actual module-byte inventories matcheb2395e1… and all88 media pins remain unchanged. Before stream drawing, this execution also records canonical RGB565 DirectDraw upload/publication; zero stream PRESENT therefore does not mean zero presentation of every kind. The old colored-geometry frame's exact origin and15-second startup failure remain unknown and did not reproduce. No behavior repair, same-frame parity, FPS, audio, logo correctness, native Quit or full27 completion is claimed. Original native/command/browser/watchdog actual waits were0, both EOFs completed and all known owned PIDs/listeners were absent after cleanup. Evidence: scratch/runs/20261118-oasaver-software-presentation-path-cause/result.json.

The following table covers the named November17–18 moving-scene results only; it neither reconstructs nor completes the historical 27-app set. Each cell means reviewed natural scene motion on that actual executor, not visual parity, FPS, audio, gameplay completion or native Quit.

Registered app Software WebGL Evidence runs under scratch/runs/
dx_boids Motion and native close observed Motion and native close observed 20261118-dx-boids-flip-current-paired-motion-native-close; preparation UNKNOWNs and finalization disk breach retained
dx_flip3dtl Motion and native close observed Motion and native close observed 20261118-dx-boids-flip-current-paired-motion-native-close; preparation UNKNOWNs and finalization disk breach retained
dx_tunnel Motion observed Motion observed Same two runs as Flip3DTL; software uses the fresh follow-up
dx_twist Motion and native close observed Motion and native close observed 20261118-dx-twist-current-browser-motion-native-close; software startup black excluded from motion pair
dx_globe Motion observed; native close repaired Motion observed; native close repaired Motion: 20261118-dx-globe-current-browser-paired-rendering-close; repaired exit: 20261118-dx-globe-native-close-cause-repair
dx_viewer Motion and native close observed Motion and native close observed 20261118-dx-viewer-current-browser-paired-motion-close; first WebGL observer UNKNOWN retained
scr_architec Motion observed Motion observed Software: 20261117-dxsdk-viewer-viewport-fix-architec-paired; WebGL: 20261117-rm-startup-geometry-architec-geometry-paired
scr_geometry Motion observed Motion observed 20261117-rm-startup-geometry-architec-geometry-paired
scr_jazz Motion observed Motion observed 20261118-rm-jazz-rockroll-paired-rendering
scr_rockroll Motion observed Motion observed Software: 20261118-rockroll-js-startup-rendering-repair; WebGL: 20261118-dx-wormhole-scifi-paired-motion; earlier UNKNOWNs retained
scr_oasaver Motion observed Motion observed Software: 20261118-oasaver-software-presentation-path-cause; WebGL: 20261118-oasaver-paired-natural-motion; earlier startup failures remain unexplained
scr_scifi Motion observed Motion observed 20261118-scifi-fallingl-current-paired-motion
scr_fallingl Motion observed Motion observed 20261118-scifi-fallingl-current-paired-motion

The fresh SciFi/Falling Leaves run, report 7268cc780, uses published source6f25a1ea0 and moduleeb2395e1…. The coordinator reviewed all eight whole frames: SciFi creatures and ground shadows move over fixed dunes on both executors; Falling Leaves objects turn and change position over a fixed printed backdrop on both. All twelve actual page/guest/render module-byte inventories match. Completed PRESENT counts advance87→161/276→462 for SciFi software/WebGL and60→104/393→814 for Falling Leaves. Each original has actual native/browser/command/watchdog wait0, both EOFs and no reported leak; all107 known owned PIDs and four listeners are gone. No runtime repair was needed. Dark shapes alone do not establish an alpha defect, and these pairs establish neither same-frame parity nor FPS, audio, native Quit, real-Windows correctness or historical27 membership. Evidence: scratch/runs/20261118-scifi-fallingl-current-paired-motion/result.json.

November 17 Quake III arena, depth repair and delayed input

The registered original on pinned e7c13db29 / module f40fc54c… reached Q3DM1 with an actual guest Worker and owning software GPS executor. The coordinator reviewed the complete before/after images and verified a turn; the arena image was delivered on Telegram as message2995.

The guest unreachable at 0x4629e8 was resolved by a fresh diagnostic to the actual glEnd thunk and the lost-depth guard during CPU fallback. Published 337b6205c preserves GPU-owned D16 depth before that fallback; the guard remains in place. Native old-fail/new-occlusion controls, all65,536 D16 roundtrips, five focused suites and the locked build pass. The affected original reached the main menu, then stopped on DOWN acknowledgement887ms against a200ms bound. Forward movement, visible firing/enemy consequence, affected-original arena verification and native Quit remain unproved. Health125→124→123 is excess-health decay, not evidence of enemy damage. All21 gestures completed independent physical release; host dequeue is separate from the uninstrumented native USER return boundary. The terminal UNKNOWN run stayed closed without further guest queries or input.

Findings f44b5577c and run 20261117-quake3-known-menu-focus-arena-gameplay retain the accepted images, 240 focused evidence checks and lifecycle limits, including the missing outer helper wait. Detailed result and next investigation. This updates that software original only; earlier backend results below and the full27, parity, FPS and audio requirements remain separate and open.

The subsequent diagnostic, published as 4ce3e278f, establishes that the original already used a software render Worker. A representative native render load did not reproduce the input refusal. One fresh original instead had prompt DOWN acknowledgement59ms and handler9ms, followed by host dequeue about3,680ms after DOM DOWN; UP also dequeued later. Native USER/game consumption was not proved, and the selected image remained at the main menu. This identifies a delayed boundary, not its production cause or an input fix. The fresh diagnostic's browser, driver, helper, supervisor and watchdog have normal actual waits and EOFs; browser closure is not game Quit. Earlier UNKNOWN runs remain closed. Depth repair and affected-run limits; timing, ownership and unresolved cause.

November 17 Forsaken menu follow-up on both executors

Published 9c61a9092 follows the six-title startup matrix on its exact G1 source 7e3028380 / module eda12ced…. Both threaded originals advanced naturally beyond the logo to the textured 3D main menu. The coordinator reviewed both whole images and delivered the WebGL image as Telegram3001.

Actual executor Observed progress Remaining boundary
Software D3DIM GPS Rendered main menu;534 draws/30 presents Enter was refused as stale before effects; subsequent permitted capture hit the unchanged20s deadline.
WebGL D3DIM GPS Rendered main menu;17,772 draws/1,042 presents Enter DOWN/UP completed; subsequent capture hit the unchanged20s deadline. No level transition was observed.

Both terminal UNKNOWN runs stayed closed without replay. Individual browser, driver and watchdog waits/EOFs are retained; an advisory outer-tail wait gap remains separate. No runtime defect or production fix was established. Movement, firing, enemy consequence and native Quit remain unproved. These two follow-ups do not replace the accepted24 startup cells, whose four Forsaken endpoints were logo-only, or qualify cooperative mode, latest main, full-scene parity, FPS/audio or the historical27. Earlier combat runs below retain their own generations and scope. Evidence, exact pins and lifecycle limits.

November 17 renderer regression results

These component results do not update the historical per-game/backend rows below. Current original-game checks and the full 27-app requirement remain open.

Path Result Scope and evidence
Repeatable browser palette oracle c2a980c5f: all32 observations pass on each software/WebGL executor Permanent test/test-d3dim-gps-palette-web.js runs the unchanged native oracle in Chrome, with26 draws/32 readbacks per backend and zero fallbacks/declines. Actual driver/Chrome exit and EOF checked; cooperative component fixture, no original-game or Worker-lifetime claim. Regression and provenance; run 20261117-coordinator-palette-browser-regression-final.
Browser Worker palette lifetime b06190e7b: current GPS fixture passes23 exact pixel observations and26 producer draws per backend Real guest/shared render Workers on software and Chrome WebGL cover caller Release, SetEntries generations, final detach, natural COM-slot reuse, queued state snapshots, alpha/key/depth pixels, pending CPU fences and retirement. Actual scheduled callbacks are held at the supported boundary; no decoder or Release substitute. Run 20261117-d3dim-current-browser-worker-palette-lifetime.
D3DIM palette color-key rejection Fixed in 280503c1a; mandatory palette oracle passes all 32 observations The producer saturated alpha-minus-threshold before TEXKILL, so keyed fragments survived. The fix preserves the negative intermediate. Real WAT palette handlers/stream production into the software executor, shader canonicalization, keyed-mipmap regression and locked build pass. No browser/game claim. Cause and tests; run 20261117-coordinator-gps-palette-fixed.
Browser D3DIM palette pixels All 32 existing observations pass on each of software and WebGL, with identical pixel values Real WAT producer and READBACK into Chrome, 26 draws/32 readbacks per backend, zero producer fallbacks/declines. Retained module ad38e26f…; cooperative executor fixture, no Worker lifetime or original-game claim. Both driver and Chrome actual exit/EOF observed. Route and limits; run 20261117-coordinator-browser-gps-palette.
Shared browser render Worker RGBA8 wire test corrected in 5cf975062; actual Chrome WebGL and native software arms pass One physical shared Worker, four readbacks, 4,096 channel comparisons, 576 padding/end-canary comparisons and four presents. Chrome process exit/close and both EOFs observed. Raw RGB565 requests correctly refused; native Glide conversion tested separately. ANGLE SwiftShader, no physical-GPU claim. Run 20261117-coordinator-browser-shared-graphics.
D3DIM surface fences Native regression passes Real WAT stream producer/software executor checks alias-specific/global readback, RGB565 pixels, dirty state, lazy Lock/Unlock and old/new backing replacement. Native process and watchdog waits/EOFs complete. Instrumented authored fixture, no original app or WebGL claim. Run 20261117-coordinator-d3dim-surface-fences.

The earlier browser palette-lifetime failure is retained in 20261117-coordinator-browser-palette-lifetime: its retired legacy transport rejected gpu_stream_kick before rendering. The fixture migration in b06190e7b resolves that dependency through the current production GPS path. Queued records own their palette snapshot, so palette API completion may precede drain; a real pending CPU fence and unchanged old-generation pixels verify the current contract. Pending readback completes on CLOSE, and a reopened completed empty ring rejects an old scheduled callback. This does not claim that destroyed resource IDs can resume. All four actual browser Worker targets retire; Chrome and driver exit/close and both EOFs are recorded. These authored component results establish no original-game, historical27 or physical-GPU coverage. Original GLTron gameplay remains assigned for native-input/backend evidence.

The later original checks in 20261117-gltron-continuous-capture-release-arm-gameplay use published b29ef843a: continuous software arena capture completes in655 ms and independent-UP arms in6–7 ms. Subsequent input cancels with reported context ages293 ms software and259 ms WebGL, beyond the unchanged250 ms fence. Diagnostic report 60371e9f3 establishes that these token-zero receipts do not prove native USER dequeue. Its authored software workload reproduces cancellation at255 ms while the guest waits492 ms for rendering, but equivalence to either original workload remains unproved. No production fix follows. Physical release succeeds; no two-turn, round or native Quit result follows. A fresh original diagnostic targets the missing command-mix and phase evidence. Repairs and original limits; phase evidence and attribution limits.

November 18 Donuts native controls and ordinary Quit

The Donuts boundary correction, code 95997a6a5 and report e400e2cb4, separates fresh native consumption from delivery-relative confirmation. Targeting remains5 seconds; confirmation 500 ms, independent UP1750 ms and operation3000 ms remain unchanged. Eight compiled-native boundary controls and the locked build passed. One original reached LEVEL001 and consumed five manual gameplay gestures. The coordinator reviewed whole frames showing the ship turn right. Ordinary Escape was consumed and the exact native owner exited0 with actual wait/EOFs and no host cancellation. Final physical UP was unacknowledged after owner exit, not an ACK. LEVEL002 and isolated shot/score causality remain unproved; outer launcher143 limitations remain recorded. This is DirectDraw/Canvas2D evidence, not a new GL backend case. Evidence: scratch/runs/20261118-donuts-targeting-consumption-confirmation-boundary/result.json.

The subsequent numbered-level run, report f49ef729e, directly confirms LEVEL002 in whole frame u43.png, personally reviewed by the coordinator. The same original consumed36 finite gameplay gestures, including six after that checkpoint, and exited through ordinary Escape with native actual wait0, both EOFs and no host cancellation. LEVEL003 remains unconfirmed; no individual shot or ambient score is credited with the transition. Final UP remained OWNER_EXITED/unacknowledged. This run used the same immutable source37118/moduleeb2395e1…, original EXE+37 files and unchanged input guards. All44 known owned PIDs were nonlive after cleanup. Evidence: scratch/runs/20261118-donuts-confirmed-level-progression/result.json.

November 18 FoxBear visible controls and ordinary Quit

The corrected-boot original, report c712f0790, establishes visible right/left movement with forest scrolling, an airborne jump and a red apple appearing at the fox's mouth/hand. The coordinator reviewed the complete initial, right, jump, apple and left frames. Projectile travel/impact and separate crouch/stop consequences remain unproved. Six finite gameplay DOWNs were consumed and their independent UPs acknowledged. Ordinary F12 was then consumed; the native owner actually exited0, signal null, with both EOFs and no forced cancellation. Final UP was OWNER_EXITED/unacknowledged, not a physical release ACK.

Published preparation 084802a67/6308e7a3b supplies retained-frame native context binding, actual main/Worker media aliases and mandatory boot data. The corrected recipe uses source6308e7a3b, moduleeb2395e1… and unchanged original EXE/ART pins. Native oldfail/newpass, refusal, queued-expiry, independent UP, durable TERM/KILL and locked-build proofs were reused; no runtime repair or additional original was needed. All11 known owned PIDs were nonlive after cleanup, with no reported leak. Earlier failed scopes retain their limitations. This game-control result does not add a GL backend or measured FPS claim. Evidence: scratch/runs/20261118-foxbear-corrected-boot-visible-controls/result.json.

November 18 Blood II preparation and level-entry limit

The frozen preparation repair, code 2ffa7f555 and report c277a06fb, restores the missing pngjs dependency. Same-loader native main/Worker tests construct the renderer and capture whole PNGs; missing-dependency, startup/cancel, terminal closure and locked-build checks passed. One original on module 7d347a6a… rendered menus through the Demo Level1 card. The coordinator reviewed the whole Homicide and level-card images. Three Enter gestures were consumed; the fourth missed its unchanged 500-ms confirmation deadline and permanently closed the scope. Independent UP was acknowledged. In-level movement/fire, enemy consequence and ordinary Quit remain unproved on this run. The launch tool exited143; native/watchdog wait records and buffered logs are missing, with sender unknown. Known PIDs and the listener were absent; this does not reconstruct the missing waits. All98 media files and the manifest matched. A fresh worker investigates level-entry input and durable supervision; earlier gameplay below remains separate evidence. Evidence: scratch/runs/20261118-blood2-frozen-renderer-dependency-repair-combat/result.json.

November 12 current original-game progress

App Software WebGL Native outcome Evidence
GLTron Moving trails and round outcome Moving trails Browser gameplay on module 1165f15c…. Cooperative and actual Worker CLI gameplay also works with --real-ticks. Current scenes, run 20261112-gltron-current-device-scene-repair-0015
Forsaken Earlier actual Worker combat and white interval Actual Worker combat; fresh sampled whitening explained Current module04 shows player damage/respawns. Joined raw state for a sampled white draw is already additive before enqueue and agrees through packet/replay/description. Original raw commands and620 observed guest selections agree with additive state;1142 equal restores also match CPU flags. Sixteen retained explosion texels and atlas rectangles match the original asset; immediate prior target color also joins. A later complete14-triangle draw has an uploaded texture equal to cache and an actual GL/native565 white bound for all16384 depth outcomes, but its incoming pixel is already white. A fresh complete28-triangle unsaturated-to-white transition then matches actual GL/native565 at every step; color changes prove two depth passes. Historical missing56 triangles and the broader old white interval remain unresolved. No rendering repair is proved. Fresh unsaturated transition, run 20261112-forsaken-unsaturated-whitening-1845-retry; complete draw and limits; retained texture and target order; original selector; joined state trace; paired combat
Darkstone Actual Worker locomotion and combat Actual Worker locomotion, combat and Town healing Victories11 retained; native food purchase/eating, rest to35/35 and sword purchase. New Save/Load restores Food82, gold393, equipped sword/AC6 and accepts fresh movement; Town bank0/service checks yield no gain, planned Ardyl quest/recovery next. Current gameplay, healing and saves, run 20261112-darkstone-town-food-equipment-0925
Quake III demo Earlier module1165 Worker arena movement and firing Current04 Worker arena movement, native I Can Win and shotgun firing Current04 verifies I Can Win with the cursor away before Fight, actual owning WebGL and released movement/aim/fire. Shotgun ammo10→8; score0–0 and no proved hit or frag. The old capture-timeout cause remains unknown. A local command-publication race is reproduced and repaired; historical Quit21's exact cause remains unknown. Current encounter, run 20261112-quake3-major-frag-2055; atomic command publication; reviewed target and pickup; earlier paired arena
Hexen II Current courtyard/chamber traversal and living attack at50 HP Current courtyard/chamber traversal and living attack at69 HP Incoming player damage on both; SoftwareGL death/native restart retained. Enemy damage/kill unproved. Current paired traversal, run 20261112-hexen2-api-envelope-paired-progress-1040
Re-Volt demo Actual Worker driving, steering and recovery Actual Worker driving and steering Software21/26 mph and WebGL Time Trial16 mph; no checkpoint/lap/finish. Unsupported helper action stopped later steering before dispatch. Current paired driving, run 20261112-revolt-current-paired-driving-checkpoint-0911
NFS3 demo Actual Worker acceleration, steering and recovery Actual Worker acceleration, steering and recovery Earlier software82 MPH and WebGL/SwiftShader48 MPH road progress remain verified. The1705 attempt aborts on native INITTIMER;1725 reaches rendered Single Race and Hot Pursuit with actual SoftwareD3D and successful traced initialization. The earlier abort does not recur, so its cause remains unknown. No driving was attempted in1725. The1740 road attempt reaches a winding approach but has null CPU epoch binding and a missing local manager wait; it is not fully qualified driving. The1815 run qualifies full private22-field binding and independent release, reaches82 MPH, but repeatedly leaves the road; seven holds have full held frames and nine short commands do not. Successive bends, lap/finish and FPS remain open. Paired driving, run 20261112-nfs3-provision-gated-paired-driving-0840-fresh; later startup abort; real Worker observer and race startup; bounded road attempt and missing binding; full binding and remaining cadence limit
MechWarrior 3 Qualified moving scene publications Qualified moving scene publications Released Digit5 motion at 40.5 km/h; native Digit1 then 24.9→8.5→0.0 km/h. Native i supplies a working menu route; historical cursor-freeze cause remains unknown. Native controls, run 20261112-mw3-native-cursor-throttle-stop-0345
Blood II BUILD109 Actual Worker movement and firing Actual Worker movement and firing Original local demo, ammo 50→48 and health 100 on both routes; enemy hits remain unproved. Original pin and route, run 20261112-native16-original-crossapp-crash-sweep-0502
Tomb Raider II demo Actual Worker Venice movement and turning Actual Worker Venice movement and turning Native options and original world. WebGL pistols drawn; firing remains unproved. Original route, run 20261112-native16-original-crossapp-crash-sweep-0502
Heretic II demo Current8a834/04 whole-four crate0400 Load at31 HP; bounded movement/released intervals with native pause cycles Current Easy docks; crate Load exits natively with code8 Original completed640×480 owning-WebGL pixels are recovered5 ms before ExitProcess(8): loading console/Paused over black, no visible diagnostic. Exit cause remains unknown. Earlier post-handler logs do not establish guest-visible read failures. Later payload-free startup park/retry traces cover no executed Load. No paired onward proof. Original exit frame, run 20261112-heretic2-webgl-exit-witness-2055; API interpretation; living SoftwareGL control; later observer limits and cleanup
Half-Life Uplink Actual SoftwareGL rear-panel destruction, corridor/headcrab encounter and repaired native Pause/resume Current actual WebGL living controls and elevator onward Complete thirteen-file03:18 checkpoint has fresh-browser Load/movement proof. General Pause scan-code repair now produces native PAUSED35 HP stable7.916 s, then ordinary resume/PageDown at30 HP and native PAUSED30 HP. Whole13 remains unchanged. End centering, enemy defeat, safe onward and paired save/pause parity remain open. Native Pause repair, run 20261112-halflife-native-pause-2000; fresh whole13 and guarded doorway; passage and checkpoint; paired controls

Half-Life uses original hldemo/PAK bytes on module04c53419…. The coordinator verified1,368 references/1,122 paths,870 decoded raw records,237 balanced trusted events and86 captures retaining all22 bound identity fields, and inspected both living scenes and the WebGL elevator/loading/bay consequence. The actual GL context proves software ownership despite the page dropdown. Both browsers, outer parents, manager and watchdog joined normally before normal box stop. No historical floor export was available; this run starts New Game Easy and claims no save continuity, old SIGKILL cause, parity, FPS or audio result. The later1430 run qualifies its own earned four-file chamber checkpoint: 324084 unchanged bytes restore through native Load in a fresh browser, followed by fresh movement. Coordinator review verified634 references,22 identity fields, 111 trusted balanced events and the full persistence sequence. Both save cases use SoftwareGL; this adds no WebGL save-parity, FPS or audio claim. The1500 onward attempt stopped on a strict renderer transition during capture, leaving that checkpoint intact and onward progress unqualified. The1510 continuation now reaches the scientist/guard upper deck and qualifies a five-file483845-byte native successor through Save/export/move-away/Load/fresh movement in one browser. The1535 continuation verifies that set in a fresh browser, reaches the red bay and qualifies its nine-file997089-byte native successor. The1600 run restores that entire set in a fresh browser, selects UPLINK00:20 and verifies fresh movement on actual SoftwareGL. Root review checked422 references/36 full PNGs and the full load/movement frames. Play ended at an actual disk-floor crossing, without a new weapon, encounter, onward room or successor Save. See the upper-deck checkpoint and fresh bay recovery.

The1615 continuation then opens the maintenance supply room and collects the crowbar at health100/armour0. Its complete eleven-file1670445-byte save passes native Save00:56/export/move-away/Load/fresh movement in the same browser. Root checked526 references/46 PNGs and six full opening/pickup/persistence frames. Crowbar checkpoint, run 20261112-halflife-bay-clearance-1615. Native shutdown and normal box stop passed; missing local starter143 manager/watchdog joins remain explicit. The1635 run restores that eleven-file set in a fresh browser, destroys the small crate with native attacks and moves into the cleared space. Its new twelve-file2006166-byte save passes normal Load and fresh movement in that browser. Root checked627 references/56 PNGs and nine full sequence frames. The taller structure remains and no encounter is claimed. The1655 continuation stops before CDP/guest launch on a missing Chrome endpoint; the whole save remains unchanged. The1710 continuation first passes an actual blank Chrome endpoint/CDP/capture/normal-close control, then restores the complete twelve files through native Load01:48 and fresh movement on actual SoftwareGL. Root checks474 references,41 PNGs and all22 within-capture identity fields. The old timeout remains causally unexplained; no production repair follows. Tall-side/jump inspection proves no passage, and crouch is not dispatched because the frozen key helper lacks Control/chord support. The close export retains all twelve save files unchanged. Fresh whole12 recovery and limits. The1730 continuation then validates emitted native Control/chords and independently released keys. Eight full held captures and the released pair confirm crouch; central, side and crouch-jump approaches still meet wooden faces. Root checked 365 references,36 PNGs and22 binding fields; the whole12 set remains unchanged. Actual Chrome/outer and both local manager/watchdog waits pass,364 processes are reaped and both boxes stop normally. A disk-floor crossing after quit is retained explicitly. No passage, encounter or successor save is claimed. Native crouch and remaining clearance. Crate destruction and complete recovery, run 20261112-halflife-crowbar-obstruction-1635.

The1825 continuation restores whole12 in a fresh browser and uses native attacks against the separate rear panel. Its removal, forward passage, south/east route and new corridor are visible; a headcrab later lunges at the player. Native Save03:18 produces a complete thirteen-file2342063-byte successor, including both maps' auxiliaries. Export/move-away/Load/fresh movement passes in that browser; the close export preserves identical members/data despite order changes. Root checked586 references,52 full PNGs,22 within-capture fields and seven full route/persistence frames. Saved health40/0 falls to30/0 in the restored frame and 20/0 at the fresh lunge; no frozen post-Load health or enemy kill is claimed. All actual parent waits/EOFs and normal stop pass,367 processes reaped. The1850 continuation now verifies all thirteen files through fresh-browser VFS, native Load03:18 and movement. It exposes the guard-room doorway with headcrabs, preserves player death and restores40/0 through normal Load; no hit or new save. The1905 continuation visibly lowers aim with PageDown, then dies before attack. Its caller selects an old capture after local compression changes mtimes; the remote exact-reference guard correctly refuses input. End centering remains unproved; the later native Pause repair is qualified below. Safe onward, enemy defeat and paired save parity remain open. Fresh whole13 recovery and limits, native look and selection defect. Native passage, encounter and checkpoint.

The2000 native Pause repair, code ae83f8d55, maps Pause to scan code0x45 without the extended bit and keeps NumLock extended. Actual Chrome-to-owning-Worker Win32 messages and original SoftwareGL before/after frames verify the change. Native PAUSED35 HP and the scene stay stable7.916 seconds; ordinary resume/PageDown changes the view at30 HP, then native PAUSED30 HP is verified. Root checked555 references, 46 full PNGs and six original frames. Both closing exports preserve whole13 members/data. No attack, enemy hit/kill, new save or WebGL Pause parity follows. Normal stop passed; the original local manager143/watchdog joins remain unknown. The later bounded-transfer caller passes local chunk/whole-hash and corruption-refusal controls after a whole-PAK PUT timeout. That old operation remains UNKNOWN/unreplayed; no later remote transfer or new gameplay is claimed by those local controls.

Diablo II's current original demo now heals through Akara to55/55 and earns outdoor XP380→401. Native Save & Exit, complete four-file export, normal Load, retained XP401/Den quest and fresh movement pass in the same browser. Root review checked811 references/75 full PNGs and six persistence-sequence frames. The1615 run then verifies that unchanged four-file set through normal Load, native XP401 stats/Den objective and fresh movement in a new browser. Root checked874 references/84 PNGs and distinct document/guest/renderer identities. Later navigation stays in camp; a stale reviewed frame refuses command83 before input, with no replay or new save. The Den remains uncleared, and no exact species/count or runtime fix is claimed. Fresh XP401 qualification and earned XP401 save, run 20261112-diablo2-native-menu-short-combat-1540.

The GLTron software proof binds its actual GL software owner and completed raster work, with a legacy D3D WebGL transport. Its default cooperative CLI still has a black scene and zero draws under the abstract CPU/device service model. The real-ticks results preserve that remaining limitation. Forsaken's later look capture shows a respawn, so controlled turning and enemy defeat remain unverified. Darkstone's latest native save restores full health, food and an equipped sword. Its current character index and all five map files are preserved as a complete six-file continuation set.

MW3 run 20261112-mw3-moving-scene-frame-cost-0410 measures distinct published terrain updates: software 2.58–3.83 Hz and WebGL 1.75–2.08 Hz. These are not strict completed-game-frame FPS: HUD uploads lose the required present-request identity, and observer self-time is only a partial overhead accounting. Driver-level diagnostics identify SwiftShader despite the page's Intel string; this is not a hardware GPU benchmark.

Published b54b1cbf8 keeps every native16 storage conversion and checks sticky GL errors once before each submitted draw can succeed. Exact DIB/RGBA controls on WebGL1/2 and early-error rejection pass, as does the locked build. Balanced moving A/B windows on the same guest, Workers, module, context and epoch show 77.5% fewer storage error queries and 10.8% less submission time per conversion. Terrain updates rose 1.763→1.874 Hz; naturally evolving scenes limit attribution. The coordinator checked 822 recovered references and the full candidate scene. Measurement, correctness controls and limits.

The additional Blood II/TR2 sweep tested runtime b54b1cbf8 / module 1165f15c…. Both WebGL arms exercised RGB565 native16 storage: 297,917 completed conversions for Blood II and 108,823 for TR2, with no reported GPU error or fallback. Software ownership was separately bound to the original RGB565 surface and completed software raster work. The coordinator checked 480 recovered references / 388 unique raw hashes and four full paired scenes. All four runtime parents exited normally, and the box stopped before expiry. These bounded scenes found no runtime defect; they do not prove general crash freedom, native-reference parity, strict FPS, audio or historical27 completion.

Re-Volt now has current8a834/module04c53419 active driving on both owning backends. Software Single Race drove, steered and recovered, while WebGL Time Trial cleared the early fence and drove along the street. Wrong-way drain travel, native positions and authored route proximity are not checkpoint acceptance. No checkpoint/lap transition, finish, save or FPS was proved. The coordinator checked519 references/414 unique compressed and raw hashes, executing module/backend bindings, native SwiftShader reporting and both full driving scenes. Forty-four bound scenes and37 balanced pulses are retained. A caller used unsupported action:"keys" instead of action:"key"; it failed before dispatch and ended control without replay. Software owner wait0 and WebGL owner wait1/Chrome0 have separate EOFs. Local wrapper143 and unavailable actual parent wait for the reparented watchdog remain explicit; the box is independently stopped normally. No production change or build was needed. Hexen II now has current8a834/module04c53419 courtyard, ramp, skull-door and chamber traversal on both actual owning GL backends. Reviewed living attack poses show69 HP on WebGL and50 HP on SoftwareGL. Incoming player damage is proved; SoftwareGL death and native restart are retained. Enemy hits, kills and kill-counter changes remain unproved. Root checked566 remote references /440 unique compressed/raw hashes and both full attack frames. All48 commands have exact ACKs;40 pulses and86 trusted events are balanced. The run-local caller now separates API execution1..600 seconds from the longer lease/observation deadline, with one-publication/delayed exact-ACK controls. No production emulator source changed. Both actual runtime/Chrome parents, local manager and watchdog joined0/EOF and the box stopped normally. Documentation54d1c903a is published. The fresh graphics worker targets current Heretic II's paired crate0400 checkpoint/traversal and onward living consequence.

Heretic II current8a834/module04c53419 now restores the exact living crate0400 checkpoint on its actual SoftwareGL owner. Full upright loaded, fresh backward and released-interval frames verify native progress. WebGL reproduces ExitProcess(8) twice; all four source files are fully read in both arms. Current/game.ssv consumption is20480/62728 bytes in both, so no mismatch is established there. The512-call tail has successful swaps, PeekMessage0 and GetLastError0, without native error text or a demonstrated contract violation. Root checked761 references /538 unique compressed/raw hashes,61 bindings, three backend/process identities and full loaded/moved/released frames. No production fix, new save or current paired onward proof follows. All five browser parents/Chrome joined; the box stopped normally. Original manager143 and absent original manager/watchdog parent joins remain explicit. b562925d8 publishes the comparison. A permitted native diagnostic or reproducible API mismatch is needed before another WebGL retry.

The later runtime-boundary comparison, published 8f68731a1, repeats the living SoftwareGL load at 31 HP and WebGL native exit 8 on the same current04 module and complete four-file save. The SoftwareGL 600 ms backward/3000 ms released and separate 4500 ms neutral checks use native pause/resume cycles; their final frames are paused, not evidence of sustained gameplay. Root checked 364 references, 30 full identities/PNG CRCs and four full frames. Existing error queries return 0 only at startup; neither arm queries the backend error flag during loading. No context loss, renderer exception or runtime violation was proved. The final successful swap precedes native exit by five milliseconds; that comparison did not retain its completed pixels. The later original exit-frame witness below fills that gap through permitted runtime/renderer observations, without guest reverse engineering. Both runtime parents have actual waits/EOF; original manager/watchdog parent waits remain unknown. The box stopped normally. No production fix or new save follows from this comparison.

The2055 original exit-frame witness, published 6e8c88fa2, recovers the final completed640×480 owning-WebGL image. Its decoded RGBA matches the independently retained pixel hash and all22 owner fields match. Completion precedes native ExitProcess(8) entry by5 ms and host exit selection by6 ms. The full original image shows normal loading-console text and Paused over black, without a visible error diagnostic. Root verified 136 encoded/raw blobs mapping161 paths,11 original PNGs,42 balanced trusted edges and seven independent releases. The original final-frame objective is met; exit cause, loaded-world survival and a new paired comparison remain open. The box stopped normally; browser parent1 after the pending post-exit capture is distinct from Chrome0 and stop/manager/watchdog actual0 waits/EOFs.

The retained API audit checks 27,236 selected API rows. All54 zero-EAX ReadFile post-handler records have matching immediate re-entry/success within the ReadFile stream. The source hook also runs during internal lazy-I/O parking, and those original records omit yield state. They therefore do not prove54 guest-visible failures or an exit cause. Passive park/completion correlation is the next missing observation; no speculative runtime repair or guest reverse engineering is justified.

Darkstone's latest WebGL continuation uses current8a834/module04c53419. Native food purchase/eating raised Food0→99 for120 gold; rest healed21→35. A native1200-gold sword purchase equipped the weapon and raised displayed Damage2–2→2–6. New native Save/Load restores35/35, Food82, gold393, 11 victories and equipped sword/AC6, followed by fresh movement. Shield1350 and rags900 remain unaffordable; no quest completion or further combat follows. The coordinator checked903 references /819 unique raw hashes, all six complete native files and full reload/fresh movement images. Actual runtime/Chrome waits reached0/EOF and the box stopped normally. Run 20261112-darkstone-town-food-equipment-0925 retains the exact recovery set; the0959 Town checks found bank0, skills800>393, empty Repair and no visible Upgrade response even with the bagged sword. No new reward/save follows; all six files remain unchanged. The helmet graphic is not independently proved as an item. Root checked1101 references /998 raw hashes and full bank/final panel. The1045 Ardyl route found no quest/dungeon/recovered kit before an unsupported Tab request failed the helper allowlist before native dispatch. No new save or final live-file readback follows; the six-file supplied fallback remains valid. Root checked996 references /905 raw hashes and the full visited-list/outdoor frames. Documentation4d5d3c884 records the route and runtime1/Chrome0 EOF, normal stopped box and unresolved outer143. This is a helper/operator failure, not a demonstrated emulator defect. The gameplay slot now targets SimGolf's unfinished native hole after its documented caller cancellation prerequisite.

Outlaws' covered controls now use the original14-bit angle period16384. Eight native release/idle tests are stable, but relative mouse behavior and the prior turn drift remain unisolated. A changed eastern-window jump did not enter the store; no new health/consequence/save follows. Native14bit/helper correction is not an emulator fix. The safe-entry dependency and authentic checkpoints are preserved. Caesar III completed the native population150 objective: Victory, promotion to Clerk, and Assignment2—Esquiline are captured in 20261112-caesar3-compact-senate-jobs-1135 (published ed4a50e8a). Root verified1,671 outputs and241 exact captures and inspected the transition and Senate30/30 staff. Exact final population was not captured. The fresh Assignment2 worker targets granary/wheat progress. Caesar III is2D and adds no GL/D3D coverage.

Disciples II's postbattle Load stall is now traced to original demo AI serialization, documented in 281713c5f. The original save header writes version2 while AI saving requests the newest layout; loading that data as2 reads0x80630000 as a list count and continues allocating after EOF until a native SmartHeap exhaustion dialog. The matched known-good a save reaches the map; its complete AI-state fidelity is not independently validated. Authentic f remains unqualified for Load/fresh movement, with no emulator fix or save rewrite claimed. The coordinator checked398 references /357 unique raw hashes, original opcode/blob audit and full SmartHeap/control-map images. That gameplay lane shifts to Darkstone native healing/quest progress while the compatibility dependency is preserved. This Canvas2D result adds no GL/D3D app coverage. Evidence: 20261112-disciples2-postbattle-native-load-cause-0836/RESULT.txt.

Published 8a834a3dc resolves the remaining MW3 scene-admission cause: allocation of a shared-child canvas was mistaken for actual contributing content. The compositor now supplies an empty-region proof for that exact composition, window, memory, canvas, DX layer and source version. Nonempty, missing, replaced and stale proofs remain rejected; a blocked cache cannot be rehabilitated. Actual WAT/Worker/decoder/compositor fail-before/pass-after, source-lifetime/transient/scheduler tests and the locked build pass.

Original software and WebGL each produced10 qualified exact-source moving scene rows in one instrumented window. Nine publication intervals averaged 619ms software and883ms WebGL/SwiftShader, equivalent to1.615 and1.132 qualified publications/s over first-to-last spans. These are bounded samples, not a broad benchmark or physical GPU result; observer costs and scene variation remain. Executing module stays 04c53419…. The coordinator checked 180 baseline and395 final references, source/test changes, both full moving scenes and timing JSON. All three boxes stopped normally. Refused terminal read15332 and outer transport143 remain disclosed; actual runtime/Chrome waits are0/EOF. Full27/parity/audio/crash work remains open. The fresh graphics lane now targets current original NFS3 driving on both backends. Evidence: 20261112-mw3-native-source-composition-admission-0755/report.txt.

Published 2aa8833a4 adds a completed synchronous native Flip source witness on the owning guest Worker, after association rotation and before present parking. Actual WAT/Worker/decoder/compositor controls fail on the old source and pass with the fix, including partial-HUD/stale/owner/concurrent rejection; the locked build passes. Original WebGL yielded11 accepted complete sources, each bound to its exact full front upload and fallback copy, but zero qualified compositor rows. Strict FPS remains unknown, and software original was not retested. The next dependency is the exact ordinary decoder/cache and composition admission failure, not a speculative guard relaxation. Executing module is 04c53419…. The coordinator reviewed source/test changes,68 initial and184 final recovered references plus the full moving cockpit image. Both boxes stopped normally; the report retains missing initial runtime wait and deliberate retest cleanup exit1 separately from Chrome0 and outer transport143. Evidence: 20261112-mw3-synchronous-flip-producer-witness-0708/report.txt.

Published 29f6295a2 repairs writer-observer attachment at an actual completed boundary. Pending secondary Worker calls drain naturally while recorded main continuations wait; participant identity is checked through renderer acknowledgement, and failure releases the holds. Actual Worker controls, scheduler/lifecycle regressions and the locked build pass. Original MW3 arms once before both backend windows, but its synchronous native Flip bypasses the queued producer observer, leaving zero qualified completed frames. Strict FPS remains unknown. The coordinator checked410 references /344 unique raw hashes and both full moving scenes. The next worker targets the missing synchronous Flip witness. Evidence: 20261112-mw3-all-writer-observer-boundary-0625.

Published d55e326c5 fixes Stop All → relaunch window lookup after process exit. Retained import contexts and window/native-dialog publications now end their lookup lifetime with the process; live foreign owners and reused HWNDs remain valid. The actual-host regression fails on the genuine old source, and focused lifecycle tests plus the locked build pass. Original TR2 stopped PID1000 and relaunched PID1001 on the same page, then rendered Venice and responded visibly to movement/turning. The coordinator checked154 recovered references /131 unique raw hashes and the final full scenes. This repairs a host lifecycle crash; it adds no general crash-freedom or FPS claim. Evidence: 20261112-host-window-retired-context-relaunch-0555/report.txt.

Published 703015ea5 repairs a generic DirectSound allocation leak exposed by original Disciples II. Suspended long-ring refreshes retained decoded buffers; coalescing them preserves latest PCM and the guest cursor on resume. The old source fails the allocation regression; six audio suites and the locked build pass. The repaired original page completed native Save and End Turn to day2, with a sampled renderer peak of781 MiB over almost20 minutes. The coordinator checked837 recovered references, the complete208,644-byte native save and full Save/day2 images. Native Load and fresh movement remain the next game action. This is Canvas2D gameplay; audible playback and GL backend coverage are separate. Cause, repair and native save.

The coordinator reviewed the full scenes and checked the raw hashes in the recovered-artifact inventories for the six earlier runs and the additional paired Blood II/TR2 sweep. The Forsaken repair has a failing old-source regression, passing focused lifecycle/publication/parity tests and a locked build. These app outcomes advance the corpus. Historical 27-app membership, remaining backend cells, reference parity, FPS and audio remain open.

November 10: current-source paired Twist scene motion

The registered original dx_twist now has personally reviewed software D3D and WebGL guest-Worker scenes from current main e2063597c7, built under the lock and observed on live module 79f0842e…. All six full 15/30/45-second captures show substantial textured-body orientation and silhouette changes. Actual guest/backend and render-Worker ownership, compiled bytecode and served source bodies are pinned. This closes the current paired software/source motion gap left by November 7; no renderer defect was demonstrated. Complete scene correctness, reference parity, measured FPS and historical27 completion remain unqualified. Scenes, controls and recovery.

November 8: qualified consequences remain separate from route attempts

Published 19d2fb24a ends Quake II local route qualification without a shot, box or new consequence. Closer captures correct flame/wall-feature guard identifications; the corpse predates input. The coordinator inspected the closer view and verified472 files/18,289,946 bytes. Twenty-one ACKs,18 independent releases, currentcc438 owning WebGL1, PNG/pin checks and local drain pass; physical GPU and an absent numeric fallback counter remain unknown. Fresh Drakan poll/flight work replaces this slot. Corrected route observations.

Published 1e2bdf191 exposes Recoil's uphill gated ramp but the tank dies before power-control identification. No objective increment, access or restart is claimed. The coordinator inspected ramp/death frames, verified307 references, 26,957,532 bytes, and independently checked108 fixture/manifest pins totaling 76,019,820 bytes against staged pins. The empty-file fixture report is withdrawn; its nonempty manifest replacement and negative checks pass. Normal archived stop/full drain pass. A fresh worker tests native menu-assisted review against the mission clock and uses justified defense while pursuing the objective. Software/Canvas2D is explicit; toolbar WebGL supplies no GL coverage. Death boundary, corrected closure and source pins.

The protocol remains active uncommitted work. Positive production-path tests and the receiver-preservation fix still require final regression/build review; no live pacing, FPS, full27, parity or audio completion follows from them.

November 8: Asghan same-plane contact; outgoing consequence open

Published 6b38c6987 establishes broad same-ground frontal preparation and actual Ctrl+Up/Ctrl+Left consumption with independent release. The orc remains standing; the shorter hero health bar does not prove outgoing damage. The coordinator inspected frontal/swipe captures and verified198 references, 14,737,528 bytes. Fifteen exact ACKs, zero UNKNOWN, currentcc438 module, software/Canvas2D and normal archived stop/settled drain pass. No emulator cause, outgoing hit/pickup or GL acceptance. Fresh Recoil force-field objective work replaces this lane; Quake II guard/door work and protocol implementation remain active. Full27, parity, audio and FPS stay open. Same-plane evidence and remaining attribution.

November 8: qualification limits and active replacements

Published 2ea85dd91 fixes NFS3 local staging/attachment but the full native route aborts at INITTIMER. A debugger-order variation also aborts, then ends UNKNOWN. No box, control trace or emulator-cause acceptance. The coordinator verified391 files/13,438,371 bytes. Exact timer predicate/API-return evidence and a matched observer-off control remain required; per-case helper recovery and changed-case game-module qualification are explicitly limited. Fresh Quake II guard/door progress replaces this gameplay lane. Local timer boundary.

Published 707b0f5da finds no Forsaken latency remedy: serial21.839s and fresh-byte overlap21.373s both hit UNKNOWN under an optimistic original timing prefix. No original/box followed. The coordinator verified165 files/9,310,497 bytes. Negative identity/replacement/release tests pass; the local module differs from the original, and Evaluate575's internal wait remains unattributed. Fresh Asghan frontal-target work takes the slot. Existing movement/ammo proof stands; look/navigation and rates remain unproved. Exact requests and failed comparison.

The second protocol handoff remains uncommitted WIP. It implements exact queued request/owed-present/RPC transfer and optional free-run/shared-render writer exclusion, tested with real parallel WAT stores. The coordinator verified64 hashes/154,472 bytes. Build and14 focused suites pass; two failures reproduce baseline. A fresh worker continues complete writer coverage, per-surface canonical versions, canvas provenance, production Join and discard accounting. No installed association, live pacing or FPS is qualified. Protocol implementation and remaining boundaries.

November 8: fresh control attempts stop on harness boundaries

Published 17d96f983 adds static NFS3 binding/packet/queue/simulation/car boundaries but obtains no live consumption trace. An absolute observer path was staged at the wrong location; attachment failed before releasing the guest, then startup timed out UNKNOWN. Corrected path tests pass locally; a fresh worker must rehearse the exact staged attachment before allocation. The coordinator verified27 artifact references/6,659,010 bytes. Normal archived stop passes; settled drain stdout is retained, standalone JSON was not fetched. Exact staging failure and next trace.

Published 4ec60dbc0 obtains no Forsaken player HUD or look. Menu command5 stops UNKNOWN during owning-guest capture-before identity: admission took8.085s, then that read remained pending9.092s at deadline. Enter independently released at2s. A fresh worker investigates this measured observation path before another original. The coordinator verified103 references/10,223,468 bytes. Normal archived stop/full drain pass; an initial supervision gap remains recorded, with replacement delegates retaining the original deadline. Earlier current movement/ammo acceptance stands. No emulator defect, navigation, rates or full27 completion follows from these attempts. Observation timing and lifecycle limitation.

November 8: race rendering, corridor progress and protocol WIP

Published 14827978c reaches an ordinary NFS3 cockpit race through native D3D selection and numeric renderer binding. Actual7ec49 WebGL2 reports122742 draws, zero errors/fallbacks. Up reaches the guest scan array and clears after release; acceleration, steering and input-caused road displacement remain unproved. The coordinator viewed the cockpit and verified328 references/18,759,648 bytes. Normal archived stop and settled drain pass. A fresh worker connects active bindings and simulation consumption before another control claim. Race and key-array evidence.

Published 9ef46f315 verifies current Forsaken WebGL thrust into the initial corridor and PULS2000 to1998. Eleven thrust pulses left look too late; the reserve guard refused it before publication. No look or beyond-corridor acceptance. The coordinator viewed the corridor and verified271 references/16,930,434 bytes. Twenty exact ACKs, zero UNKNOWN, actual7ec49 owner/context and normal archived stop/full drain pass. A fresh worker tests turn/look earlier, then navigation. Composition qualification and current progress.

The D3DIM protocol task leaves tested partial source/test WIP uncommitted. Completed-producer identity exists, but owed-present and stable consumed-version witnesses remain missing. Actual WAT/full composition regressions demonstrate front-DIB rebinding and in-place/ABA changes with unchanged upload version and pointer. Build and focused tests pass; the coordinator verified25 evidence files/71,686 bytes. No box, production join or pacing qualification. A fresh worker reclaims this WIP to finish the protocol; no completion-only commit is accepted. Full27, parity, audio and unmeasured FPS remain open. Uncommitted implementation and exact missing boundary.

November 8: presentation identity fails local qualification

Published 6eb7a8f35 identifies the missing association between a completed owning Flip and the canonical surface version consumed by page publication. Production-boundary tests show duplicate notifications increase counters, delayed arrival changes receiving intervals, and repeated bytes receive new composition sequences. The tested Join contract is a model, not a production fix. The coordinator independently verified21 files/214,361 bytes. No box was allocated and no rates are reported. The296ns/call microbenchmark measures only telemetry-seam timestamp logging, not gameplay overhead.

A fresh worker now implements the version-carried production association and adversarial regressions before a locked build and any live pacing task. Native TR2 controls remain accepted; logical FPS, physical scanout, observer slowdown, full27, parity and audio remain unproved. Boundary evidence and concrete remedy.

November 8: current TR2 controls and TR3 WebGL translation verified

Published 811bca5b6 qualifies native Enter/default-OK locally and remotely, then persistent Venice Up displacement and a separate Left heading change on actual owning WebGL2 module7ec49cbb. The demo enters Venice directly. The coordinator inspected paired terrain frames and verified106 references, 9,641,848 bytes. Eight exact ACKs, zero UNKNOWN, full settled process drain and normal archived stop pass. Rates remain unmeasured; a fresh task qualifies frame identity and overhead before live pacing windows. Native OK and current Venice controls.

Published 767dc5774 establishes current WebGL terrain-relative Jungle translation retained through twelve- and ten-second idle captures. The coordinator inspected the same rock/boulder landmarks and verified204 references/14,528,997 bytes. Eighteen exact ACKs, zero UNKNOWN,48 actual module7ec49cbb checks,189 HTTP reads and normal archived stop/full drain pass. Together with 78df2e4d9, bounded current translation and separate heading are verified on WebGL and Software. Firing, level completion, FPS, audio, parity and full27 remain open. Fresh Forsaken progress takes this slot. WebGL translation.

Published 369e2613b loads NFS3 with the complete executable closure after native D3D selection and reload. The owning WebGL2 D3DIM endpoint records805 draws and zero errors/fallbacks on7ec49cbb. Race input remains untested: the run-local helper incorrectly required a string endpoint rather than numeric1. The prepared correction passes local regression tests; a fresh worker now qualifies its production binding and race controls. The coordinator verified130 references/9,155,030 bytes; the prior box is normally stopped and archived. Native D3DIM consumption and helper boundary.

November 8: native NFS3 selection and current TR3 movement

Published 4c1d0352a qualifies both NFS3 thread-owner repairs through the native chooser and ordinary Direct3D Device 1 selection/OK. Native registry writes are observed. The game did not load: run preparation omitted its registered executable and returned HTTP404. The corrected complete closure passes the local HTTP/size/SHA regression but awaits fresh race execution. The coordinator verified115 artifact references/8,753,946 bytes and inspected the selected chooser. The box is normally stopped and archived. Chooser, staging failure and corrected preparation.

Published 78df2e4d9 establishes current Software D3DIM persistent Jungle translation and separate turning; current WebGL turning also passes. WebGL translation remains ambiguous because camera follow confounds the captures. The coordinator verified415 references/25,865,913 bytes and inspected paired software movement captures.37 exact ACKs, zero UNKNOWN, actual module7ec49cbb, both backend owners and normal stopped-box cleanup are recorded. Firing, parity, FPS, audio and full27 remain open. Fresh work isolates WebGL translation. Current movement and evidence limits.

Published c007b3b75 is another failed TR2 OK attempt: the worker misread the visible Graphics sheet and missed the activation window. No emulator defect is established. The coordinator verified62 references/7,090,612 bytes and personally confirmed the visible sheet; startup/fresh/refresh hashes match. The box stopped normally after settled process drain. A fresh worker first qualifies native default-OK activation locally before allocating another box. Current controls and presentation rates remain unqualified. Corrected attempt.

November 8: TR2 controls/rates blocked; image assessment corrected

Published 957b861cd collects no original gameplay input or rates. A pixel- coordinate request was rejected locally before queue publication. The worker's persistent Graphics-absence assessment is withdrawn after coordinator raw review: startup, fresh-graphics and refresh-graphics have identicalee913bc1 hashes and show the property sheet; only the final idle frame shows desktop. No renderer endpoint exists, and no compositor cause is established. The coordinator verified63 references/56 hashes/7,092,663 bytes. Four exact ACKs, module7ec49cbb/596pins/173HTTP checks and normal archived stop pass; controller PID remains in both prestop checks, so full prestop drain is unproved. Fresh work qualifies correct native OK and current controls before any collector. Queue/notification counter examples are not unique-frame or logical FPS proof. Full27, control/rate acceptance, parity and audio remain open. Corrected raw-image assessment and collector gap.

November 8: NFS3 thread ownership fixes; chooser qualification pending

Published e6f4672ff fixes two demonstrated wrong-thread dispatches. Shared NC retrieval delivered a main-window WM_NCPAINT to the secondary thread; DestroyWindow later promoted a foreign window and synchronously called its WM_SETFOCUS on the destroying thread. Both paths now restrict ownership. Failing-before/passing-after regressions, focused neighbors and locked build pass; an unrelated BUTTON expectation fails identically on baseline. The coordinator reviewed code and64 artifacts/8,947,608 bytes. Final native qualification of both fixes stopped on disk4.46GB, with no box allocated. A fresh worker now qualifies the chooser after headroom recovery. Race/context, three-renderer timing, full27, audio and parity remain open. Wrong-context chains and regressions.

November 8: current TR3 WebGL Jungle and pistol draw

Published 0a38cf484 reaches Jungle through ordinary native welcome/setup, then binds an actual owning WebGL2 producer using modulef5cfb1c6. Space2000 shows raised pistols. Short Up/Left attempts do not prove persistent translation or heading. Disk4.48GB stops further input; software remains unexecuted. The coordinator inspected the armed frame and verified215 references/193 hashes, 14,085,840 bytes.20 exact ACKs include three zero-effect stale refusals; no UNKNOWN.602 pins/189 HTTP reads/45 module references and normal archived stop pass. Intel is the reported provider string, not proved physical hardware. A fresh worker continues movement/turning after headroom restoration. Full27, parity, FPS, audio and level completion remain open. Current route, actual owner and resource boundary.

November 8: additional startup coverage stops at native dialogs

Published e6d667fa8 adds12 corrected browser cells across TR2, TR3 and Drakan, cooperative/threaded modes and Software/WebGL selectors. All complete without classified crash/deadline/UNKNOWN. TR2 remains in Graphics setup, TR3 at welcome and Drakan at its beta warning; no owning graphics backend is created. Three initial optional-type-library staging gaps remain separate. The15 executions are three titles, not15 or27 apps, and do not prove active backend rendering/gameplay/parity. Published78299/f965 runtime excludes peer WIP.

The coordinator reviewed three four-cell sheets and verified521 references, 424 hashes and38,459,391 bytes.15 PNGs,2327 served reads/53 loaded sources, serial cleanup and normal archived stop pass. Fresh TR3 gameplay work handles the ordinary dialogs before binding an actual WebGL producer. Cells, staging correction and shutdown.

November 8: NFS3 loader repair reaches a null chooser-object boundary

Published 0df0f1667 makes absent vendor LoadLibrary requests fail with NULL and error126 while retaining implemented builtins and native module identity. Failing-before regression, A/W/Ex checks, lifecycle neighbors and locked build pass. A clean production Worker using modulef5cfb1c6 now executes the native d3da suffix retry and export lookup. The chooser remains blank; its secondary thread reaches a null virtual call during WM_NCPAINT. The object-lookup cause is unproved. Mixed-module and exploratory observer runs are not acceptance.

The coordinator reviewed the blank chooser and verified81 artifacts/9,051,214 bytes and133 media hashes. No remote box or chosen-D3D/race/context proof. A fresh worker traces the first causal mismatch; earlier registry/bootstrap progress stands. Full27, benchmarks, parity and audio remain open. Loader contract, clean retry and null-object failure.

November 8: NFS3 native device enumeration advances setup

Published b1b631f12 implements the live Win9x Config Manager registry tree for the existing virtual display and monitor. Focused failing-before/passing- after registry/display tests and the locked build pass. Native catalogue and initial software-default setup writes transfer through ordinary shared storage. The second setup loads d3dtest.dll, then an absent d3d.dll incorrectly yields the EXE handle, suppressing its authentic d3da.dll suffix retry. The chooser and actual D3D race remain unqualified; no box was allocated. A fresh worker owns the general loader mismatch. No alias, guest patch or forced choice.

The coordinator verified57 artifacts/2,427,119 bytes and133 media hashes; executing module and260 served-source reads are retained. This is a platform repair, not a renderer/gameplay/FPS/parity or full27 acceptance. Live tree, native bootstrap and loader boundary.

November 8: Warcraft cause remains unproved; audio fixture passes

Published dc07cd57a, run 20261108-warcraft3-renderer-process-loss, finds no retained mapping between vanished OS PID11520 and Warcraft's page, and no original exit reason. Its matching production Worker/audio refresh fixture passes lazy fallback, promotion, loop2 refresh, notification, stop and replay. The coordinator independently parsed77682 raw records into38841 balanced entry/return pairs and checked17 compressed hashes. No missing or unbalanced fixture call was found. Probes add182.5µs per refresh:25.48× in this synthetic burst, with original overhead unmeasured.

A deliberate local Page.crash yields exact-target error134, native SIGABRT and crash artifacts; those explain only the control. Prelaunch PID mapping remains absent. No fresh original or box was launched, and no emulator fix is justified. Future original diagnostics need independent target-exit and process evidence with qualified remote service and explicit observer overhead. The current additional-title startup sweep proceeds separately. Player control, historical27, parity, FPS, audio and campaign remain open. Raw-call checks, source path and mapping gap.

November 8: Quake II pickups on both actual GL paths

Published 327e830da, run 20261108-quake2-native-encounter-progress, adds native HEALTH pickup consequences: WebGL100→104 and SoftwareGL100→102. Ordinary Game/Easy, original controls and media are unchanged. The owning WebGL context has software GL disabled; the owning SoftwareGL context has it enabled, independently of its WebGL presentation surface. This run reports SwiftShader and Chrome151.0.7922.71, unlike the predecessor's Intel report; no hardware acceleration or matched-renderer parity claim follows.

The coordinator reviewed pickup pairs, delivered Telegram2301/2302/2305 and independently verified436 references/394 hashes.33 exact ACKs,27 independent releases,46 PNG checks,707 unchanged pins and359 served reads pass. Both browsers drained serially and the box stopped normally after retrieval. Guard defeat, door/onward transition, full27, FPS, audio and level completion remain open. No emulator repair was needed. Native route, actual backends and shutdown.

November 8: Drakan native bootstrap and both current D3DIM worlds

Published ed993c025, run 20261108-drakan-native-warning-gameplay, qualifies ordinary native MessageBox/GDI Enter before renderer creation, then a one-way handoff to strict actual D3DIM ownership. Production software/WebGL triangle, native input and zero-input rejection controls pass. Both original arms reach the rainy dragon level through native New Game/Normal/story input. Software D3DIM and actual WebGL2 producer work and both owner modules are verified; WebGL reports Intel/Mesa ANGLE. Native relative mouse input changes terrain-relative heading. Sustained forward flight/altitude is unproved.

Software primary/secondary attacks show no outgoing target consequence and incoming enemies drain health; WebGL does not establish a target engagement. The coordinator inspected both world/heading views and the software attack, independently verified459 references, and delivered Telegram2296.39 exact ACKs, zero UNKNOWN, unchanged681pins/454HTTP checks and normal archived stop pass. No emulator repair is claimed. Actual attack polling/timing needs local qualification before another original attack. Parity, FPS, audio, full game and full27 remain open; Asghan takes the released gameplay slot. Bootstrap controls, action limits and cleanup.

November 8: recorder repaired; renderer disappears after Prologue scene

Published 14057deaa, run 20261108-browser-journal-remote-service, repairs diagnostic recorder overload. Corrected remote production controls pass with75ms maximum delivery lag, compared with56494ms in the baseline. Current SoftwareGL reaches a reviewed rendered Prologue scene. Final passive capture times out after the journal's nested voice_play_ring entry without return/error; page and observer emissions stop and a renderer PID vanishes from later browser samples. OS PID-to-Warcraft-target mapping was not retained, so the vanished process is not yet proved to own that page. No exit reason or inner audio boundary is captured; audio deadlock, OOM and emulator fault remain unproved.

All32552 Node-admitted batches became durable. Earlier8528 lost observer slots remain explicit; the final218625-slot tail is contiguous. The coordinator reviewed the scene and independently verified340 artifacts/296 hashes plus the2,582,639,811-byte decompressed journal hash.15 original exact ACKs precede UNKNOWN; player control is unproved and WebGL unexecuted. The box stopped normally after retrieval. Fresh work qualifies full-load bursts and captures process/target exit and inner call evidence before another original. Full27, parity, FPS, audio quality and campaign completion remain open. Transport comparison, final call and process evidence.

November 8: remote journal delivery fails before startup readiness

Published 7fa820727, run 20261108-browser-page-nonreturn-cause, qualifies local production observer/busy-call controls. The fresh remote recorder receives idle journal batches with median33s/max53.7s lag; guest attachment is processed after Chrome exits. Seven page heartbeats after the readiness request exclude uninterrupted synchronous execution during that interval. No production call entry, startup owner proof, original input or image is recovered. This does not identify the historical stall or justify an emulator change. The coordinator independently verified32 references (4,787,418 bytes) and the normal archived shutdown. A fresh worker must qualify remote recorder service and attachment before launching the original. Thrall control, paired gameplay, full27, parity, FPS and audio remain open. Diagnostic coverage and delivery failure.

November 8: optional capture-state change does not resolve the page stall

Published a7581bcd5, run 20261108-browser-page-capture-cause, removes an avoidable supplemental-state dependency in controlled production Worker tests, retaining all strict identity/module/image/admission checks. It does not fix the fresh original. SoftwareGL reaches Prologue loading, then page heartbeats stop before the final owner evaluation is sent. No entry probe, stack or profile is recovered for that request before the unchanged deadline. Input closes on UNKNOWN; WebGL is skipped. No active operation or emulator defect is identified, and lost object probe payloads limit interpretation.

The coordinator inspected the last loading frame and independently verified all208 artifact references. Both modules match f965;627pins/399HTTP checks, 18 preceding ACKs and normal archived shutdown pass. New work qualifies an independently readable, durably streamed journal/trace before another run. The prior current-module cinematic remains valid; Thrall control/tutorial, paired gameplay, parity, FPS, audio, campaign completion and full27 are open. Composition controls, original stall and diagnostic gaps.

November 8: Warcraft producer readiness and current Prologue introduction

Published 9fd4b4a5c, run 20261108-warcraft3-producer-ready-gameplay, qualifies the actual software endpoint1→transport2 transition before binding, and rejects postbinding replacement/invalid producers with zero input. Production Worker controls pass on softwareGL/WebGL. The original software arm retains exact ownership through native profile, Campaign and Prologue loading, then renders two advancing introduction frames. Both owning modules match f965 and the actual context is SoftwareGL. These cinematic frames do not establish Thrall selection, displacement or tutorial progress.

The final passive capture stalls20s in page-state Runtime.evaluate before the strict binding/screenshot stage; its ACK is absent/UNKNOWN, input closes and original WebGL is skipped. The coordinator reviewed the prompt/cinematic pair, independently verified231 references and delivered Telegram2293. The box normally stopped after retrieval. Its watchdog was armed after allocation but before browser launch; the lifecycle deviation is retained. Fresh work addresses this page-state boundary before another original attempt. Current paired gameplay, parity, FPS, audio, campaign completion and full27 stay open. Controls, original route, UNKNOWN and lifecycle.

November 8: Drakan native warning precedes graphics creation

Published 6b486344a, run 20261108-drakan-current-browser-action, reaches the reviewed beta-warning dialog with a running owning guest Worker, but no graphics endpoint, renderer Worker/context or D3D stats. The driver requires a graphics producer before admitting OK, so readiness exhausts with zero input/commands/ACKs. WebGL is unexecuted. This identifies a native-dialog bootstrap gap, not a D3D rendering failure or current gameplay result.

The coordinator inspected the warning and independently verified all31 artifact references. Focused T/G input/release controls pass through a real guest Worker. All64 current fixture files remain unchanged; unavailable historical per-file hashes are explicitly unknown. The box normally stopped after retrieval. Fresh work qualifies pre-graphics native-dialog input and then positive D3DIM producer readiness before another original route. Movement, flight, attack, actual backend coverage, parity, FPS, audio and full27 remain open. Warning, exact gate, provenance and cleanup.

November 8: owning GL readiness corrected; Q3 actions remain open

Published c142f0785, run 20261108-q3-gl-readiness-gameplay, corrects the run-local observer with actual owning GL draws/presentation, context and module proofs, retaining software legacy-transport attachment. Both production Worker positive controls and all zero-input negative controls pass. The fresh original WebGL path opens input despite empty D3D statistics.

Current SoftwareGL reaches Q3DM1 and a burning torch. Ammo remains100 and intervening deaths/respawns prevent movement/fire attribution. WebGL reaches level selection; a later Enter command releases independently, then its post-input page-owner capture exceeds the unchanged work deadline. Its final ACK is absent/UNKNOWN, with no replay. The coordinator inspected courtyard, torch and menu, and independently verified295 artifact references. Both actual owner modules match f965; this WebGL1 provider is Intel/Mesa ANGLE. The box normally stopped after retrieval. Further Q3 runtime needs a locally validated capture-latency remedy; paired gameplay/torch, parity, FPS, audio and full27 remain open. Warcraft III producer-readiness work takes the coverage slot. Controls, original actions, UNKNOWN and cleanup.

November 8: Warcraft III current-browser readiness blocker

Published 02129c6b0, run 20261108-warcraft3-current-browser-objective, stops before player input. The SoftwareGL driver binds endpoint1 on a black startup frame with presentation writeSeq0. The next passive capture rejects endpoint2 after the software legacy transport attaches; read-only state then records 172,189 triangles and writeSeq27. This establishes premature driver binding, not an emulator crash or current gameplay. WebGL was skipped.

The coordinator inspected the black startup capture and independently verified all42 artifact references. Both live modules match f965;624pins and400HTTP response checks pass. Four extracted-media hashes match historical notes; the other eight retain current pins without historical-hash equivalence. The box normally stopped after retrieval. Readiness must be qualified after actual presentation/transport attachment before another original attempt. October8 movement remains historical; current Prologue, objective/combat, paired gameplay, parity, FPS, audio and full27 remain open. Observed ownership transition and cleanup.

November 8: current Quake III world and GL observer boundary

Published 702535909, run 20261108-q3-current-browser-world, resolves the earlier gray/white startup observation through natural loading: both actual GL backends reach the native menu, and SoftwareGL reaches textured Q3DM1. The coordinator reviewed the courtyard/HUD and WebGL menu and independently verified all 388 retained artifact references. The current f965 module is recorded; WebGL's completed per-owner binding remains absent.

SoftwareGL's first mouse hold leaves ammo100, followed by death during a W hold; neither firing nor movement is established. A later respawn request was refused before queueing with zero effects. WebGL input never opens because the run observer requires d3dStats.glTriangles on a native GL endpoint whose d3dStats is empty despite advancing GL batches and presentation. This is a harness predicate mismatch, not an emulator startup defect. Its post-close observer timeout remains recorded. All browsers drained and the box normally stopped after verified retrieval. Fresh GL-readiness work tests that exact predicate before another original action attempt. Paired gameplay, torch, parity, FPS, audio and full27 remain open. Exact route, identities, observer failure and cleanup.

November 8: Forsaken initial world and capture-cost diagnostic

Published c1ab23eec, run 20261108-forsaken-capture-composition-cause, reaches the current browser's native Volcano loading screen and initial textured world/HUD: Lives 5, shield/hull 128, PULS 2000 and MUG 5. The coordinator inspected both frames and independently checked 143 artifact references. The actual owning legacy WebGL2 context and guest/render modules match the recorded f965 runtime. No movement, firing or combat was tested.

The earlier command deadline did not recur: ten commands, including Start, received exact successful ACKs under unchanged budgets. The measured startup capture cost 16.472 seconds, dominated by owning-guest inspector evaluations up to 7.991 seconds; unrelated-thread reads and module transfers were short. The locally tested unrelated-thread-selection candidate was never enabled in the original because those timings do not support it. No causal repair was validated, and the previous failed Start remains failed with its exact pending await unknown. Normal shutdown followed verified retrieval. Actual phase timings, controls, original frames and limits. Full27, paired gameplay, FPS and audio remain open.

November 8: browser startup sweep after the COM boot repair

Published 29027f029, run 20261108-graphics-post-renderer-repair-sweep, checks the actual browser boot changed by 37ce9588e. Twenty-four intended cells cover six titles in cooperative/threaded mode and requested software/ WebGL paths. Three additional Quake II controls make 27 executions, not 27 apps. The initial software-labelled Quake II cases actually used WebGL; corrected cases establish its software path. The twelve CLI cases were deliberately unexecuted because they bypass the changed browser sequence.

Original Reviewed endpoint in all four intended cells
Quake II Main menu with console text behind
Quake III Three gray frames and one white frame; menu unreached
GLTron Colored main menu
Hexen II Undismissed notice; GL startup untested
Heretic II Textured book menu
Forsaken Logo only; probes do not establish active WebGL drawing

Raw diagnostics record no classified escaped trap, unimplemented API, guest exit, setup failure or case deadline. Quake III's sampled EIPs progress; gray/white at the 20-second observation endpoint is not a crash diagnosis. Capture and diagnostic overhead have separate timestamps. The coordinator inspected all six contact sheets and independently checked all 816 artifact references. Twenty-seven module reads, 96 loaded-host reads and applicable COM registry observations match the staged runtime. Two launcher-143 interruptions and missing outer reports are preserved; the serial run and evidence survived. The box normally stopped after verified retrieval. No emulator repair was introduced by this sweep, and these startup windows establish no gameplay, parity, FPS, audio or full historical27 completion. Exact cases, identities, failures, recovery and shutdown.

November 8: Tachyon browser flight controls on both backends

Published aba977513, run 20261108-tachyon-current-browser-backends, verifies original Single Player → pilot A → Play → base door 01 → Ship Bay → Launch through ordinary controls on serial browser sessions:

App Software Direct3D WebGL
Original Tachyon demo Owning legacy endpoint 1, positive software draw completions, no WebGL context or adapter. Textured cockpit, throttle changes, turning and blue laser bolts. Owning legacy endpoint 1, actual WebGL2 context on Intel/Mesa ANGLE. Textured cockpit, throttle changes, turning and blue laser bolts.

The coordinator inspected both blue-laser cockpit frames and independently checked all 453 artifact references. Live CPU/render modules match f9650607…; the recorded host includes COM boot repair 37ce9588e. All 15 original/manifest files remain unchanged. Thirty-nine exact ACKs complete without UNKNOWN, with independent key release. One software pilot-menu click preceded review of its immediately preceding capture; this review-order limit is preserved. Flight captures were reviewed before dependent control requests. Both browsers and their controller processes drained before normal nonforced box shutdown after verified retrieval.

Gate animations and changed instructions do not establish a completed jump: the final recorded objective remains Jumps 2 / Head to Luna. No target hit, Luna arrival, mission completion, parity, FPS, audio or full playability is claimed. Cinematic-band artifacts remain disclosed. Originals, backend identities, controls, captures and limits. Full historical27 and remaining per-app coverage remain open.

November 8: Redline browser texture and HUD repair

Published 37ce9588e, run 20261108-redline-texture-realization, fixes browser boot initializing the shared DirectX COM registry twice: once on the owning guest Worker, then on the idle page instance. Later guest threads read the replacement registry and misidentify the original DirectDraw4 creator, returning Surface2 interfaces that correctly reject its 124-byte descriptors. The repair preserves the owner's registry; local boot still initializes it. The actual owner/page/later-thread regression fails before and passes after, with strict descriptor checks and the locked build passing.

The unchanged original Stadium City route now completes all 467 gameplay texture realizations. The coordinator inspected retained before/after captures: flat surfaces and HUD blocks become textured scenery and weapon, with readable ammo 200 and health 100. The owning legacy renderer is WebGL2 on Intel/Mesa, with matching exports/memory, 204 cached textures and 374 uploads. The WAT module remains f9650607…; the host boot code changed. Twenty-six commands complete with exact ACKs and no UNKNOWN. A single Control action produces no demonstrated ammo decrement or distinct shot consequence, so firing and vehicle entry remain open. This run does not establish software-backend coverage, parity, FPS, audio or full playability. Evidence was verified before normal nonforced shutdown. Cause, regression, original captures and cleanup. The prepared follow-up crash sweep must exercise actual browser Worker boot; a CLI-only run would bypass the repaired sequence. Full historical27 and remaining per-app coverage remain open.

Subsequent original verification 56d1ede3a, run 20261108-redline-ready-firstperson-action, establishes firing: visible muzzle/ projectile effects and ammo 200→150 persisting after independent Control release. The coordinator inspected those frames and checked all 211 artifact references. The requested held capture completed after key-up, so no sampled-held or input polling claim follows. This later run used actual legacy WebGL2 SwiftShader; it is not hardware-GPU or paired-backend proof. Vehicle entry, mission progress, FPS and audio remain open. Firing, readiness correction, interrupted custody and shutdown.

November 8: Tux Racer completes current browser races on both backends

Published d58dffae2, run 20261108-tuxracer-current-browser-backends, completes original Practice → Frozen River through ordinary native controls:

App Software GL WebGL
Original Tux Racer 0.61a Owning software OpenGLHostBridge context/shared memory; main guest/render CPU flags 1. Finish and matching best result: 01:50.50, 15 herring, score 0. Owning live WebGLBackend v1, Intel/Mesa ANGLE; guest/render CPU flags 0. Finish and matching best result: 02:02.96, 19 herring, score 0.

The coordinator inspected both returned best-result screens, the WebGL finish and software steering frame. Textured terrain, transparent tree silhouettes, Tux and readable changing HUD accompany acceleration, steering and release. Both actual guest/render modules match f9650607…. Twenty-four exact ACKs and sixteen independent key releases complete without UNKNOWN or replay. The black backing behind the speed readout remains disclosed; these are bounded visual checks, not flawless rendering or pixel parity. Different times and paths do not measure comparative performance, clock accuracy or FPS. Audio remains open. Evidence was hash-verified before normal nonforced box shutdown. Race records, backend identities, tests and limits. Full historical27 and remaining per-app coverage remain open.

November 8: Quake II browser world and blaster on both backends

Published a16f5b820, run 20261108-quake2-bounded-world, verifies ordinary Game → Easy, textured world/weapon/HUD, forward movement, keyboard turning and blaster discharge/projectiles on both current browser routes. The coordinator inspected both projectile frames and software forward progress.

App Software GL WebGL
Original Quake II demo Owning OpenGLHostBridge software context/shared memory, guest/render CPU software flags 1; world, movement, turning and discharge reviewed. Owning live WebGLBackend v1 on Intel/Mesa ANGLE, guest/render CPU software flags 0; world, movement, turning and discharge reviewed.

Executing modules match f9650607…. Nineteen exact ACKs and fourteen independent key releases complete without gameplay UNKNOWN or replay. The accepted size metadata repair remains unchanged; bounded request-path tests prevent the prior oversized-wait failure. Both guard encounters cause incoming player damage; outgoing hits, kills, door/pickup and onward level progress remain unverified. Different timing and views do not prove parity, FPS or audio quality. Evidence was hash-verified before normal nonforced box shutdown. Backend identities, captures, tests and limits. Full historical27 and remaining per-app coverage remain open.

November 8: Heretic II browser docks and door on both backends

Published 7000de90d, run 20261108-heretic2-current-browser-backends, verifies current original campaign progress on both browser routes:

App Software GL WebGL
Original Heretic II demo Owning GL endpoint 1 software context/shared memory, CPU software flag 1; textured docks, narrative trigger and opened door, raised threshold reviewed. Owning GL endpoint 1 WebGLBackend v1, CPU software flag 0, live Intel/Mesa ANGLE context; textured docks, narrative trigger and opening door reviewed.

The coordinator inspected the WebGL door-before/after pair and software doorway capture. Both actual CPU/render modules match f9650607…; the separate legacy WebGL endpoint on the software route is not its GL owner. Forty-six exact ACKs and forty independently released holds accompany movement and attack poses. Player damage and spell light do not establish an outgoing hit or kill. Full interior traversal, objective completion, matched pixels, FPS and audio quality remain unverified. Evidence was hash-verified before normal nonforced box stop. Captures, actual backend identities and limits. Full historical27 and remaining per-app coverage remain open.

November 8: Hexen II browser world and traversal on both backends

Published 3e75cd4b7, run 20261107-hexen2-current-browser-backends, verifies current browser courtyard movement, ramp ascent and skull-door traversal:

App Software GL WebGL
Original Hexen II demo Owning OpenGLHostBridge software context/shared memory, CPU software flag 1; textured chamber and close spider reviewed. Owning live WebGLBackend, CPU software flag 0, non-lost ANGLE SwiftShader context; textured chamber reviewed.

The coordinator inspected both post-door captures. Live CPU/render modules match f9650607…, and the published audio repair is present. Software presents establish active rendering, not FPS; its separate legacy WebGL endpoint is not the GL owner. Thirty exact ACKs and independently released holds accompany the progress. No attack was submitted: player damage and a visible spider do not prove an enemy hit or kill. Viewpoints/timing differ, so pixel parity and performance remain unverified. The earlier CLI notice-only and MIDI-blocked observations below retain their original scope rather than describing this run. Backend identity, reviewed captures and lifecycle. Full historical27 and remaining per-app backend/gameplay cells remain open.

November 7: Current GLTron browser gameplay on both backends

Published 175f0258b, run 20261107-gltron-current-browser-round, verifies ordinary native steering in serial cooperative browser rounds on both routes:

App Software GL WebGL
Original GLTron 0.70 Main OpenGLHostBridge, software context and gl_sw_enabled=1; live turn and bent yellow trail reviewed. Owning WebGLBackend.gl matches the observed context and gl_sw_enabled=0; live turn and bent yellow trail reviewed.

The coordinator inspected both steer-live-after.png captures and their paired state records: direction 3 → 0, trail offset 0 → 1, player alive before/after with speed 8.5 and winner -1. Actual module is f9650607…, with the published audio fix present. The toolbar label is not backend evidence; stale overlay text and different viewports are disclosed. This proves scoped gameplay on both routes, not matching pixels, FPS, physical GPU performance or audio quality. The synthetic-clock CLI menu limitation remains a separate observation. Backend ownership, captures, original inputs and limits. Other per-app cells and full historical27 coverage remain incomplete.

November 7: Forsaken threaded recovery and GLTron audio primitive

Published 6c682a2ab follows the short startup sweep with a longer original Forsaken observation. Threaded software Direct3D is black at ten seconds, reaches the original menu at twenty seconds, and proceeds to readable Volcano. Ordinary thrust moves from the starting room into the metal corridor; firing reduces PULS from 2000 to 1946. The coordinator inspected the matching room and corridor captures. This does not establish a persistent threaded renderer failure, a hit, WebGL parity or complete crash freedom. The exact clear/caller preceding transient black intervals remains uncaptured. Evidence: 20261107-forsaken-threaded-black. Source, controls, screenshots and limits.

Published ac1bcecee fixes headless waveOut idle-time banking: a drained queue previously counted its idle gap as playback of newly submitted buffers, causing premature completion. Duration/FIFO/header-reuse/rate/callback regressions and the locked build pass. The original cooperative GLTron still submits no first scene and reaches no menu. The remaining investigation follows batch-clock publication, completion eligibility, consumer wake and runnable producer budget; the primitive repair is not a menu fix. Evidence: 20261107-gltron-audio-service. Regression, original result and missing observations. Full historical27 membership and per-app software/WebGL gameplay coverage remain incomplete. These observations do not change that requirement.

November 7: Hexen II post-notice progress and control repair

Published 10bf01e42 fixes stdin control replies being suppressed by diagnostic trace windows. Actual production Notepad reproduces the missing reply before the fix; success, error, snapshot and ordinary key replies pass afterward. Focused control regressions and the locked build passed.

The cooperative original passed its notice and reached native GL plus MIDI open-callback execution. Missing replies stopped that run as UNKNOWN; the threaded arm was not launched. Its final logo texture is not a composited menu or world capture. Evidence was verified before normal shutdown of bx_25tc5y5f. Observed progress, regression and limits.

Published eee2adf57 follows through with repaired transport: the cooperative original reaches a reviewed composited courtyard with fists and HUD, but black scene regions and incomplete HUD glyphs remain visible. The threaded arm creates native GL and then fails at MIDI stream open. The host incorrectly rejects its local main VM when an auxiliary Worker backend exists. The fix admits only the verified main instance; callback lifecycle, foreign-owner rejection, actual Worker regressions and the locked build pass. The original was not rerun after the fix. Backend selection does not prove auxiliary-thread execution, and empty PCM does not prove playback. Evidence was verified before normal archived stop of bx_xtyxkm7m. Playback, parity and historical27 completion remain open. Composited world, MIDI repair and limits.

Published fece02eca then reaches the threaded composited courtyard and records visible forward movement toward a stone wall. Black models and HUD regions persist. Bounded draw state identifies enabled texture names with no image levels and positive primary colors. The ordinary shader incorrectly sampled incomplete textures; the repair preserves API enable state while skipping incomplete texture application, with failing-before/passing-after WebGL1/2 pixel regressions, focused suites and locked build passing. These original captures predate the shader fix. Published cd98b939d reviews a fresh original with the repair: previously black statue/tree geometry becomes grey and HUD blocks become a white strip; missing artwork and broad black world regions persist. The first observed failed image call, number3550, sets WAT GL_OUT_OF_MEMORY before host upload. CPU image storage fails before definition metadata is written; DIB backing exhaustion was subsequently reproduced and repaired in published ef2709a76: GL CPU images now use reclaimable heap backing, with production capacity/lifetime/copy regressions and the locked build passing. Visible repair and storage failure.

Published 4e9e3fe17 verifies the allocator repair in a fresh registered original: textured statue/tree, illustrated HUD, courtyard floor/walls and ramps replace the grey/white/missing artwork. All 5,563 image definitions reach host uploads; call 3,550 has error zero and sampled DIB occupancy remains 621 pages. Ordinary movement produces a flask inventory icon and changed views, followed by damage, zero health and a natural return. No confirmed enemy hit, objective or visual parity follows. Evidence was hash-verified before normal archived shutdown. The prearmed watchdog was absent at final inspection; its unexplained lifetime is a prerequisite for the next original, rather than claimed continuous coverage. Restored artwork, gameplay and limits.

MIDI open/callback execution was observed, but no playback or auxiliary guest execution was established. Evidence was retrieved and hash-verified before normal archived stop of bx_hqm2pn99. Threaded movement, texture repair and remaining boundary.

November 7: Heretic II after texture storage repair

Published 1815b027c verifies textured Silverspring docks, roofs, sky, Corvus, stone and HUD on the registered threaded/headless-GL route. Ordinary movement reaches waterside views; health falls from100 to26 near a visible rat, followed by the native Load Game book. No attack input, enemy defeat, doorway traversal or objective was proved. The bounded observer records3,002 image definitions and matching uploads with error zero; later image coverage is unknown.

The revised detached supervisor survives launcher-session termination in focused mock tests and records original watchdog heartbeats through normal stop. This addresses the next-run lifetime prerequisite without explaining the earlier HexenII watchdog disappearance. Evidence was hashed before nonforced archived shutdown. The graphics slot next runs the existing current-module crash sweep; this one GL app does not establish historical27 membership or D3D parity. World progress, lifecycle checks and limits.

November 7: Repaired-module startup crash sweep

Published 3b89643d3 records twelve serial20-second windows across QuakeII, QuakeIII, GLTron, HexenII, HereticII and Forsaken, cooperative and threaded. Every window reaches its harness deadline without a classified escaped trap, unimplementedAPI, guest exit, setup failure or external timeout. This is startup coverage, not gameplay or complete crash freedom. HexenII remains at its notice. Cooperative GLTron and threaded Forsaken end black; their opposite modes show menus. Native contexts, backend selection and unequal scheduler progress do not by themselves identify a renderer defect. No speculative production fix landed.

All twelve actual module reads match the tested allocator build; runtime/native pins and20 sweep regressions pass. Evidence was hash-verified before normal archived shutdown. Later Forsaken progression and the partial GLTron audio repair are recorded above; exact historical27 membership and full backend coverage remain open. Six-app results and limitations.

November 7: Tunnel, Twist and Boids browser coverage

Published fa63787b0 closes the four remaining Worker configuration harness gaps after exercising boolean readiness and the exact capture path against production guest/render Workers locally. Twelve native compositor captures cover software Tunnel Worker and WebGL Tunnel, Twist and Boids Worker cases. Tunnel and Twist show scene motion. Boids shows a moving terrain boundary and one blue triangle; a recognizable flock remains unverified. Earlier cooperative WebGL Boids observations also lacked a flock through 120 seconds. Advancing draws and zero errors do not establish complete flock rendering.

The WebGL launcher forces ANGLE SwiftShader; the reported Intel renderer string is not physical GPU evidence. Unique evidence and both modules were SHA-verified before normal nonforced shutdown of bx_88fguqk8. No emulator defect, FPS, audio, parity or historical 27-app completion is claimed. Worker configuration evidence. Published 794273255 completes that discriminator: all 13 birds and 130 triangles survive into completed GPU geometry with exact XYZ/RHW preservation. Two birds intersect the Worker view while 11 lie outside; all 13 lie outside the separately timed cooperative view. Six native compositor captures agree with those counts. These observations establish neither a full visible flock nor correctness of the original model/matrix projection. No rendering defect was demonstrated in that run. Evidence was hash-verified before normal nonforced archive of bx_z65jbe9z. Geometry and visible-flock evidence. Published bae5dfb76 completes the original-model transform comparison: an independent scalar oracle exactly reproduces all 4,680 submitted position components across 39 bird draws. A separate high-precision calculation places all 13 birds outside the frustum in each of three observation windows, explaining the terrain-only phase. This excludes a model-to-submission transform mismatch in those draws; upstream x86 simulation, camera construction and native Windows parity remain unverified. Evidence was SHA/size verified before normal nonforced archive of bx_2svdb5d8. Original-model transform evidence. Published f04a2048d verifies current original SimGolf bitmap rendering and native construction: a new tee and green, Hole 1 open for play as a scenic 127-yard par 3, and a native championship course export. The 86,831-byte CSE was recovered and SHA-verified; a resumable career save or played round remains unverified. The actual owning WebGL 1 bitmap is 800×600, hwnd 0, composed through GDI; browser provider reports ANGLE Mesa Intel UHD 620. Both executing modules are e41511fe…. Twenty-eight actual Chromium release controls, full production Worker qualification and the locked build passed. Evidence was verified before normal archived shutdown of bx_bevw9s9b. This adds named scene and native interaction evidence, without claiming historical27 completion or parity. SimGolf course and export evidence.

November 7: Architecture, Jazz and Rock & Roll scene motion

Published 17a9d604d adds 12 named cases for original registered Architecture, Jazz and Rock & Roll savers: software and WebGL, each with cooperative and guest Worker execution. All 36 captures were inspected by the worker; the coordinator inspected Architecture's threaded WebGL motion pair and Rock & Roll scene. Objects change pose and extent against fixed backgrounds, with continued owning execution and no recorded escaped trap during the observation windows.

All six WebGL cases report zero errors and fallbacks; their actual provider is ANGLE Vulkan SwiftShader. Software draw completion counts remain unavailable. The locked build, three focused checks and postflight pins passed. Evidence and both modules were retrieved and hash-verified before normal nonforced shutdown. No implementation defect was established. FPS, audio, Windows-reference correctness, backend parity and historical27 membership remain open. Named cases, actual providers and evidence limits.

November 7: three registered savers across both backends and schedulers

Published c04cc7024 records motion in the original Falling Leaves, Geometry and Sci-Fi savers: software and WebGL, each with cooperative and guest Worker execution. The worker inspected all 36 preferred captures across 12 named cases, showing objects moving relative to fixed backgrounds. The coordinator also inspected Geometry's threaded WebGL motion pair. Fourteen launches had no recorded escaped trap, unimplemented-API trap or page exception.

Two initial trials affected by an observer's viewport change remain excluded from the preferred motion evidence; fixed-viewport followups cover those cases. The locked build and three focused regressions passed. All 2,218 captured successful response bodies matched pinned bytes; 62 additional bodies were unavailable to DevTools and are not individually certified. Evidence and both modules were retrieved and verified before normal box shutdown. This establishes named scene motion, with timing, visual-reference parity, audio and full historical27 membership still open. Named cases, captures and limits.

November 7: registered Direct3D samples across both schedulers

Published 3a006e140 records 12 successful registered startup cases for dx_globe, dx_viewer and dx_flip3dtl: software and WebGL, each with cooperative and threaded execution. The worker inspected every final startup capture and ordinary controls. Viewer loads the original mslogo.x through File/Open; Globe switches to wireframe through its Render menu; Flip3DTL rotates. The coordinator also inspected the threaded WebGL Viewer model and Globe wireframe captures and sent them to the owner.

The locked build and three focused sample regressions passed. The texture-wrap recording fixture failure found during that run was subsequently reproduced and repaired in fe9458c93. The fixture explicitly observes per-primitive storage and every submitted UV view, preserving its existing software pixel and wrap oracles. The focused test and locked build pass; production rendering is unchanged. A separate actual WebGL native16 regression passed 16 configurations and 96 dependent pixel pairs across both shipped modules and direct/Worker execution, with 56 negative controls rejected. That pixel evidence is separate from the recording fixture's UV checks. Fixture repair and native16 evidence.

No emulator repair was justified. Globe texture bands still need an aligned Windows reference; full historical27 membership, FPS, audio and backend parity remain open. Evidence was retrieved and verified before normal nonforced box shutdown. Cases, controls, hashes and test limitation.

November 7: Quake III browser WebGL movement, firing and torch pair

Published 524cc8138 verifies the authentic PK3 transfer and 913 runtime pins. The owning OpenGLHostBridge uses WebGLBackend v1 on Intel/Mesa ANGLE, with no software batches or fallback reported. The coordinator inspected the settled Q3DM1 courtyard, forward-progress and torch-approach captures: ordinary balanced W input moves toward the lit arch while the weapon, HUD and world remain readable.

The later published fe995f4db completes ordinary firing and the requested same-camera torch pair on the owning WebGL 1 backend, Intel/Mesa UHD 620. The coordinator inspected both torch captures: ammunition changes 92→82, the flame remains visible, and the arch, trim and crosshair stay aligned. Four fixed architecture rectangles totaling 59,150 pixels are byte-identical. The run verifies all 914 runtime pins and retrieves its evidence before normal box shutdown. No emulator defect was demonstrated. Full historical27 coverage, FPS, audio and backend parity remain open. Firing, paired captures, identities and cleanup. Earlier movement and transfer evidence.

November 7: Quake III browser software torch and WebGL menu

Published e0194cce7 repairs stale launch metadata: the controls file contains 567 bytes, while its recorded size was 488. The production-loader regression, focused checks and locked build passed. Findings 2661845f9 record original browser software GL gameplay, respawn, movement and firing. Paired skull-torch captures show ammunition 96→94 with health 21 and readable flames in both; the coordinator personally inspected both captures. Actual GL endpoint 1 uses software, despite the UI WebGL label and a separate legacy WebGL endpoint.

The serial WebGL run verifies the owning OpenGLHostBridge WebGLBackend and Intel/Mesa ANGLE provider at level selection only. A CDP Tab timeout followed by an incorrectly queued Enter prevented a gameplay claim; recovery released both keys and emptied the queue. WebGL world rendering and torch parity remain unproved. A fresh worker tests finally-release and command-success gating before original WebGL follow-through. All 909 pins passed and the sole box stopped normally after evidence retrieval. No FPS, audio or full27 claim follows. Captures, renderer identities and input boundary.

November 7: native GL gameplay with the Worker scheduler

Published 999436b27 records original Quake II movement, turning and visible blaster fire, and GLTron cycle turns through native llvmpipe GL. Quake II created no auxiliary guest threads; GLTron created a guest Worker that executed two slices and RPCs. Balanced controls were released and both originals exited normally. Retained round-result text does not establish GLTron survival or victory. No new emulator defect was demonstrated. Browser parity, FPS and audio remain open; the browser Quake III follow-through is recorded above. Gameplay captures, thread observations and limits.

November 7: Forsaken browser WebGL combat

Published 8304ab98e verifies ordinary Volcano movement, look and firing on Chrome 151, with the owning WebGL2 worker on Intel UHD 620. Inspected captures show ammunition 2000→1918, “ENEMY KILLED YOU,” then a readable respawn with four lives and restored shield/hull. No emulator mismatch was demonstrated. The white death interval is excluded from readable-world claims; intermediate damage, enemy kills, level completion and backend parity remain unverified. The box stopped normally after evidence retrieval. The subsequent native GL Worker-scheduler gameplay observations are recorded above. Captures, identities and limits.

November 7: Descent 3 installation prerequisite

Published c2316ac83 records current browser startup stopping before GL at missing installer-derived DirectX registry metadata, followed by normal exit. The existing registry regression passes; no rendering defect or current flight progress was demonstrated. Published 35952117c completes bounded installation recovery: 25 installed media hashes match, but no completed authentic DirectX snapshot or original setup media was found. The recovered genuine DX7a installer checkpoint still requires six WININIT operations; importing its version alone would misrepresent installation. Supported restart completion remains necessary before another flight check. The graphics slot now tests original Forsaken browser WebGL combat. Startup evidence and recovery prerequisite. Installation recovery evidence.

November 7: repaired original Quake III torch follow-through

Published eef6cb6cb closely matched the retained skull-torch scene on repaired native GL without the black quads before firing. A bot kill interrupted firing at that camera. Separate inner-room firing consumed four machinegun rounds; ordered decal draws and compositor samples preserve darkened RGB with opaque presentation. No new emulator defect was demonstrated. Exact historical cause, browser parity and synchronized firing at the retained camera remain open. Original captures, owning pixels and runtime limits.

November 7: opaque GL window presentation repaired

Published 36a212cac fixes two demonstrated presentation defects: destination alpha leaked into opaque window composition, and native canvas readback read the restored draw framebuffer instead of the presented default framebuffer. Public GL/WGL regressions pass on standard and compatibility modules and reject the old behavior; focused rendering tests and the locked full build pass.

The preceding original Quake III run captured Q3DM1 wall-fire decal state on native GL. It did not match the retained skull-torch viewpoint and preceded the repair. The exact torch shapes, post-repair original rendering and browser parity remain unverified. A fresh worker continues that rendering investigation. Owning draw evidence, regression oracles and run limits.

November 7: original native GL gameplay after the shader repair

Published 7e26c8485 records original runs containing native GL shader repair aaaf6e370, with standard module e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. These are Node routes on Mesa llvmpipe, separate from browser WebGL evidence.

Title Observed route and ordinary gameplay Remaining limits
Quake II demo Cooperative native GL: Game/Easy reaches the world; paired W and Right change the view. Threads reached menus only.
Quake III demo Cooperative native GL reaches Q3DM7 and moves forward; threads reaches Q3DM1, respawns and moves forward. Cooperative Left press lacked release when the guard ended. Torch-adjacent black shapes remain under investigation.
GLTron Cooperative native GL: cycle/trail, paired A turn, crash/winner and return to menu. Threads reached menus only; visual parity remains unproven.
Forsaken demo Registered software D3D in both modes: Volcano world, visible green fire/ammunition decrease and changed view after Up. This supplies no GL or browser evidence for Forsaken.

The eight serial startup checks reported no escaped trap or unimplemented API. Their short captures alone did not prove gameplay; six controlled followups supplied the observations above. The sweep selected a Quake III logo texture, so followups used compositor captures. The coordinator reviewed selected world and input captures. No new repair, FPS, audio, complete crash freedom or full historical27 coverage follows from these runs. Evidence was retrieved before normal box shutdown. Exact identities, captures and route limits.

November 7: OrganicArt rectangle reproduced on both renderers

Published 8b1dae070 recovers the run-local browser adapter and records eight confirmed software/WebGL scene layers, all with the same lower-right white rectangle. The coordinator inspected a scene layer from each renderer. Native traces report missing CALogo.gif and installation registry values; software submits white geometry at the rectangle bounds. This corroborates the existing missing authentic logo-input explanation; no emulator rendering repair was justified. Matching original installation metadata/assets remain needed.

The actual loaded module is e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. The two routes were observed at different natural scenes, so these captures do not establish frame parity or FPS. All 117 retrieved artifacts passed hash checks and the fresh box stopped normally. Run, route corrections and evidence limits.

November 7: current GTA2, Tomb Raider and Diablo II gameplay evidence

These runs advance individual title/backend cells. Their linked notes record actual source and module identities, reviewed captures, inputs, owning renderer counters, and cleanup. They do not establish current-main coverage for the entire historical corpus or resolve the exact historical 27-ID membership. Older dated tables below retain their historical scope.

Title Current WebGL evidence Current software evidence Remaining limits
GTA2 Wild Demo Walking, native Enter vehicle entry and distinct SHARK driving translation; owning WebGL endpoint reports zero errors/software fallback. Coordinator reviewed entry/driving captures. Published 27fdd86e9. Run and review. Separate software run verifies settled walking, native Enter BUG entry and driving translation against road/gantry landmarks. Owning legacy/software endpoint 1 has no GPU adapter and reports advancing direct software completions, error 0. Published 9d8ebf6e9; coordinator also inspected the retained driving pair. Software run. Combat, audio, FPS and backend parity unproven; old fullscreen failure did not recur. Software module bytes are identified; fresh compilation provenance is not inferred from staging HEAD.
Tomb Raider II demo Venice: settled forward movement, small turn, pistol draw and a visible muzzle-flash/lighting response to Control. Owning Worker WebGL reports zero errors or software fallback. Latest published 58d054088. Run and review. Production software selection demonstrates settled forward movement and persistent pistol draw. Owning software completions advance with no GPU adapter, WebGL triangles, errors or fallbacks. Coordinator reviewed captures. Published c22e58ae8. Software run. Jump, target hits/damage, audio, software firing and FPS comparison unproven.
Tomb Raider III demo Explicit New Game reaches Jungle. Forward movement settles at a changed position/camera; a separate Space action leaves both pistols raised. Owning Worker WebGL draws advance, with zero reported errors or software fallback. Published d5c39cdf2. Run and review. Settled Jungle translation and persistent pistol draw, personally reviewed by coordinator. Owning software completions advance with no GPU adapter, GL triangles, errors or fallback. Published c0b3d06ad. Software run. Initial long input returned to title; its cause is unproven. Jump, combat, audio, FPS and backend parity remain unproven.
Diablo II demo Single Player reaches town; ordinary ground-click movement changes terrain position and persists after idle. Actual owning WebGL counters advance. Published 68240bf8b. Town movement. Supported software selection reaches town and demonstrates persistent ground-click movement, with advancing owning software draw counters and no GPU adapter. Published 7437aa212. Software run. Later traversal remained in town: field/enemy combat and FPS parity are unproven. Later attempt.

The earlier WebGL Tomb Raider runs used actual Worker module 6cd858302895de91df5af17e3a3742281e31f369ddb6528d7b96252b3423caac. The later TR2 WebGL firing run used actual guest and render Worker module 2f9fd9f5fb1841afe68f47f61d9870ea617a73f2c45253cbf662c2479b723e55, with its source commit and retained patches recorded in the linked note. The TR2 software run used actual Worker module c1fbcf285cd0e7f0468d8c22af9ddf4b7edc7b7f092af77e41d51f3da2b7c186. The TR3 software run used actual Worker module 762e723515df6ce8c197fd0d23e0b84869657b0b49e673ad4ac2c48f78b8991f. Missing historical provenance files are disclosed in their notes rather than reconstructed. Each run used one original on a fresh separate boat and retrieved evidence before normal stop. None demonstrated a new emulator defect or changed production code.

November 4: original MDK2 browser/WebGL gameplay controls observed

The ordinary localhost candidate selector now retains the existing MDK2 registration. The original executable ran through Launch → New Game → Medium in an inspected browser with a real owning Worker WebGL1 context (ANGLE, Intel Mesa), canonical module cffca869b541e6eff12ae3a6b74c4476977e0a7e23cd43d8ec5ba8faecd4bcee. Reviewed captures include the original startup/cutscenes, tutorial, Up/Down movement and Space entering and exiting the Sniper Bullet view. Exact source, input, backend and run limitations. The software CLI result below retains its separate scope; shooting/jumping, other title/backend gaps and exact historical 27-ID membership remain open.

November 3: MDK2 prerequisites and bounded software gameplay verified

The current unsupportedMdk2 list in test/test-wgl-proc-availability.js is empty. Its separate fontWglApis list checks the three implemented layer/font exports. This is a static inventory observation, not a fresh runtime test or proof that an original game will encounter no further unsupported API. The historical checkpoints below retain their original scope.

Subsequent published change Accepted evidence and limits
8028df6ba complete wglCopyContext transaction All twenty server mask bits, atomic failures, direct and delayed Worker ABI checks, WebGL1/2 pixels
d2ef199a9 shared texture publication Original cyan-pixel failure repaired in WebGL1/2 through direct and Worker routes
7baae9487 raw Worker drawable fixture and response cloning Production drawable contract and software response clone repair
36f205f11 three edge-flag exports Outgoing boundary flags, intact polygon edges, native/Worker/WebGL1/2 pixels
b3e5c1f90 boundary polygon two-sided lighting Collinear-leading-vertex orientation repair and cyclic-start controls

These accepted results close the earlier context-copy and shared-texture prerequisites. They do not establish original MDK2 gameplay, complete visual correctness or backend parity for the historical 27 applications.

The 23-export evaluator family is published in 4ec8732ba: real curves, surfaces, mesh topology, automatic normals, map/query state, lists and context state pass native, queued Worker and WebGL1/2 tests plus eighteen regressions. A zero-evaluation mutation is rejected. Evidence and limits. Coordinator verified all 135 local artifact hashes and actual STOPPED state with a completed verified snapshot. The compatibility binary was omitted from local recovery; its hash/log and exact-source rebuild recipe remain retained. Selection/feedback is published in 91da9cb835: all eight APIs use real primitive clipping, depth/name hit records, feedback tokens, overflow/error and render restoration semantics. Locked build, native, actual queued Worker, WebGL1/2 and twenty-one regressions pass; a hit-suppression mutation is rejected. Evidence and limits. Coordinator verified181 local artifact hashes and86 recovered payload hashes, including BOTH canonical and compatibility modules, plus actual STOPPED state and completed verified snapshot. Original MDK2 gameplay remains unverified. Color-index is published in c96b1b241: thirteen APIs implement genuine typed state, arrays, queries, lists and context semantics. Locked build, native, actual queued Worker, WebGL1/2 and twenty-four regressions pass; a lost-index mutation fails as expected. Coordinator verified167 local and102 recovered payload hashes, both modules, actual STOPPED state and verified snapshot, and inspected restored WebGL pixels. Evidence. The final three observed prerequisites, wglCreateLayerContext, wglUseFontBitmapsA and wglUseFontOutlinesA, are published in 20fed4877 and 856be2942. Retained semantic and original-run evidence covers native, actual queued Worker, WebGL1/2, rejecting controls, locked builds and 34 scoped regressions. Coordinator review checked all 356 recovered files, both modules and the completed snapshot followed by normal STOP. The original run reached an inspected 640×480 Demo attract scene; this does not establish player control. The subsequent original New Game → Medium route is verified in 081a33477: ordinary Up/Down movement and reversible Space sniper mode, with reviewed captures and 314 sequential control requests (maximum 64 batches). Run evidence and limitations. Coordinator checked all 204 final local hashes, including both modules and 106 recovered payload files. This is explicit software CLI gameplay; the November 4 browser result above adds bounded WebGL1 controls. Shooting/jumping and per-title software/WebGL evidence for the full historical corpus remain open.

Recovered original corpus table; exact 27-ID membership still unresolved

The count predates this table. Commit 2d722ab3766c288d13796ed9dad3f1ceeca8f1eb (September 22) introduced tools/gfx-app-census.js and reported 27 OpenGL/Direct3D candidates among 225 registry apps, including seven D3DRM screen savers. Its commit message explicitly retains icewind_dale_demo and tworld as possible launcher noise. The original scanner used DLL/name literals in registered executables and DLLs; it did not establish runtime API use. Reproduce this provenance with git show --format=fuller 2d722ab37 and git show 2d722ab37:tools/gfx-app-census.js. The commit does not enumerate all 27 IDs. Recovering that exact output still requires the contemporaneous registry and binary inputs or a retained census; running today's expanded scanner against today's files cannot establish the historical membership. No app is removed on the basis of the noise warning. Source-only evidence: scratch/runs/20261105-gld3d-historical-census-provenance/.

On November 6, the exact September 22 registry and scanner were evaluated directly from Git objects against the currently retained binary paths. The registry still has 225 apps. This produced 89 candidates including DirectDraw, and 28 candidates with a family other than DirectDraw; the original commit reported 87 and 27 respectively. The reconstruction records SHA-256 and size for each available input among 333 unique paths, with four missing inputs. It does not establish that today's files equal the September 22 binaries. The script and complete results are retained in scratch/runs/20261106-historical-census-candidate/.

A hash crosscheck against notes at that same September 22 revision anchors 41 of the 329 retained inputs to historically documented SHA-256 values. All 41 retained files were rehashed and their sizes checked. These anchors include non-graphics inputs and do not prove census-time file availability. The historical Icewind Dale note documents the same 6,283,264-byte executable retained today, SHA-256 b94816d10029cb99c0315f175330c917be2fff298394e853e2764973d8d13af4. Changed Icewind executable bytes therefore do not explain the reconstruction's eight OpenGL candidates versus the reported seven. Availability at scan time, other input identities and the exact 27-ID membership remain unresolved. The reproducible matches and historical note identity are retained as historical-hash-anchors.json and icewind-identity-followup.json in that run. A missing hash citation is not evidence that an input changed.

The 28 reconstructed candidates are generally, arcanum_demo, ut2003_demo, ut2003_demo_server, ut2004_demo, icewind_dale_demo, baldurs_gate_chapters_1_2_demo, morrowind, simgolf_demo, black_white_2_demo, quake2_demo, halflife_uplink, warcraft3_demo, pawn, tworld, winamp, mcm, ptct, dx_globe, dx_viewer, dx_flip3dtl, and the seven screen savers listed below. These are a reproducible candidate list for follow-up, not a replacement historical 27-ID manifest or runtime coverage evidence. The missing paths are the historical Atomic Bomberman executable and the Quake II, Diablo II and Half-Life installer paths; other registered files can still produce hits for an app with a missing path.

The original version of this file is retained at commit d5b3b56cb (October 6, task GLD3D-CORPUS-27-20261006). It describes the goal as every OpenGL and Direct3D registry app working on both software and WebGL. Its "Apps that really use OpenGL or Direct3D" table contains 26 grouped rows and 37 named entries after splitting slash/comma groups. These include seven screen savers with abbreviated names, SDK examples, a server variant, and the heavy titles; the table is not an unambiguous 27-ID manifest.

Reproduce the source with git show d5b3b56cb:docs/gl-d3d-corpus-status.md. Preserve its entries as historical candidates when reconciling coverage; do not select 27 by dropping grouped examples or heavy titles. All 37 names now resolve to current registry IDs as described below; the exact intended 27-ID set and current evidence for both backends remain open. The original software sweep explicitly capped runs at 120 seconds and did not run its WebGL arm, so its menu/loading rows and older WebGL citations do not establish current gameplay or parity. This recovers a concrete provenance source without claiming that the historical membership question is solved.

The seven abbreviated screen-saver entries do map unambiguously to current registrations in lib/apps.js:

Historical entry Current registry ID Executable
scr_architec scr_architec ARCHITEC.SCR
fallingl scr_fallingl FALLINGL.SCR
geometry scr_geometry GEOMETRY.SCR
jazz scr_jazz JAZZ.SCR
oasaver scr_oasaver OASAVER.SCR
rockroll scr_rockroll ROCKROLL.SCR
scifi scr_scifi SCIFI.SCR

The remaining 30 historical names are already exact current registry IDs:

blood2_demo tomb_raider_2_demo tomb_raider_3_demo gta2_demo mw3
diablo2_demo darkstone_demo arcanum_demo dx_boids dx_flip3dtl
dx_tunnel dx_twist mcm dx_globe dx_viewer halflife_uplink
simgolf_demo quake2_demo warcraft3_demo ptct ut2003_demo
ut2003_demo_server ut2004_demo alien_shooter crimsonland pawn
pirates_2004 black_white_2_demo morrowind winamp

On November 5, splitting the original table's first column at slashes and commas reproduced 26 groups and 37 unique names. Looking up every name in require('./lib/apps.js').APPS, with only the six abbreviated saver aliases above expanded, resolved all 37 to unique registry IDs and executable paths. The complete path mapping and source hashes are retained in scratch/runs/20261105-gld3d-historical-registry-map/result.json; the registry SHA-256 was 146c8059f639a1571e6c9e925a5ab7676ffd547a56327863ea72b6eb3aa22455. This checks registry declarations, not installed media or execution. In particular, ut2003_demo and ut2003_demo_server share an executable path but remain distinct registrations and historical candidates.

This static mapping retains all 37 historical candidates. The savers' paired October 30 browser observations are recorded below; mapping adds no runtime coverage. The intended exact 27-ID membership remains unresolved.

Earlier November 3 checkpoint: 51 unsupported names

The published lookup test at c1f211fdc records 51 unsupported mandatory MDK2 names. The earlier 60-name checkpoint below is historical. These subsequent commits account for the nine-name reduction:

Published change Newly implemented exports Retained evidence
345270459 line and polygon stipple 3 test/test-gl-stipple.js, test/test-gl-stipple-web.js
e318fe1c5 interleaved arrays 1 All GL 1.1 layouts and direct/queued/WebGL validation
1e48691f4 shared object namespaces 1 Shared context objects and lifetime
3785da2fa texture residency 1 Oni prerequisite evidence
024c732b0 double color and texture priority 2 Oni prerequisite evidence
c1f211fdc all sixteen logic operations 1 Stored-width native, WebGL and Worker validation

Published bd1a15a25 then preserves exact integer color-source words before float32 conversion through current state, arrays, display lists and provoking vertices. Computed colors retire that provenance. Run 20261103-gl-exact-integer-source-recover passed resolved native direct and actual queued zero-fallback controls, real WebGL1/2 and browser Worker pixels, information-loss rejecting controls, focused regressions and the locked build. The two context/hardening oracle failures were reproduced with pre-change GL bodies. The coordinator previously verified the retained evidence hashes and 4,129 source pins before accepting this repair; the temporary boat stopped. This repairs the information-loss prerequisite identified in the logic-operation investigation. Published c1f211fdc completes glLogicOp with all sixteen operations on native RGB555/RGB565/RGBA8 and actual WebGL1/2, including browser Workers. The coordinator verified 450 retained artifact hashes and 4,108 current source pins, 32 passing production invocations including the locked build, 520 WebGL records and 15 intended rejecting controls. The native shared queue records 336 submissions with zero fallback. The temporary boat is verified stopped. Run 20261103-gl-full-logic-operations and coordinator review 20261103-coordinator-gl-logic-review retain the source/module identities and checks. Context copying was the next bounded prerequisite at this checkpoint; its subsequent implementation is recorded above.

These results do not qualify an original MDK2 replay, its first window or its gameplay, establish backend parity for the historical 27 applications, or reconstruct that historical membership. The next acceptance remains actual runtime and pixel evidence at the scope of each requirement.

November 2: MDK2 prerequisites and accepted unit movement

Published 181f4295fe repairs MDK2's six retained software drawable cases. Direct and queued routes now cover selected front/back reads and writes, independent destination blending, NONE with depth/stencil updates, drawable rebinding and lifecycle. All 25 focused runs, four affected reruns and the final locked build pass in 20261102-mdk2-drawable-color-buffers. The follow-up df747405e verifies those cases again and records the then-remaining WebGL boundary: FRONT read/draw and BOTH draw return error 1282; cross-drawable rebinding fails. That probe uses the actual consumer dispatcher with a recorder, without GPU pixel evidence. Published 71fb8fb279 subsequently repairs that boundary with per-drawable FRONT/BACK attachments and lifecycle handling. Its run 20261102-mdk2-webgl-drawable-attachments passes 167 pixel assertions each in WebGL1 and WebGL2, 13 endpoint checks and 21 actual browser Worker checks on ANGLE Intel UHD Graphics 620 / Mesa 23.2.1.

Published 47d4a08225 also repairs rejected-setter state restoration across presentation: blending, scissor and viewport now survive the next same-context draw. Run 20261102-mdk2-accumulation-implementation passes 321 checks each in WebGL1 and WebGL2, 13 endpoint checks, 26 actual browser Worker checks, 32 focused runs and the locked build. That browser reports ANGLE SwiftShader; it does not establish hardware GPU qualification.

Published b50eb7b0bf adds seven rectangle aliases with direct/queued native pixel and ABI checks in 20261102-gl-rectangle-family. Published 69fa4b52c then implements glLoadMatrixd, glTranslated and glScaled, including double argument decoding, selected-stack postmultiplication, display-list snapshots and Begin error atomicity. Run 20261102-gl-double-matrix-implementation passes 60 checks each through actual resolved direct and queued routes, meaningful negative controls, eight focused regressions and the locked build on an exact 7,186-input snapshot. Evidence was retrieved and hash-checked before its temporary boat was stopped. See matrix evidence. Published ab45b7d43 subsequently implements signed per-drawable accumulation, including glAccum and glClearAccum. Run 20261102-mdk2-accum-implementation-remote passes 261 native observations each through resolved direct and queued routes, 584/588 actual WebGL1/WebGL2 checks, 525 browser Worker checks, meaningful old-source and semantic negatives, 27 focused regressions, the union gate and the locked build. All 7,202 snapshot inputs and 21 tested committed blobs match; evidence was retrieved before its temporary boat was stopped. Published a6a1248b9 then implements native glGetDoublev with binary64 output, authoritative state conversion and synchronous queued visibility. Run 20261102-gl-getdoublev-implementation-resume passes 157 ABI checks per direct/queued route, 30 regression variants, 14 rejecting controls and the locked build. Its evidence archive hash was independently verified and its temporary boat stopped. That result left 175 mandatory names unsupported. Published 4631ff857 subsequently repairs the already-resolved glRotated; its double argument and matrix semantics do not change the missing-export count. See rotation evidence.

At this November 2 checkpoint, fourteen subsequently accepted families reduced the unsupported set to 60 mandatory names in the lookup test:

Accepted implementation Newly implemented exports Evidence
9bafbc319 typed normals 8 Normal conversions, lighting and direct/queued validation
9f56285e2 typed texture coordinates 25 Numeric adapters, dimensional defaults and display lists
7170d9afa typed vertices 14 Numeric adapters, homogeneous coordinates and direct/queued validation
80ebba348 raster positions 23 Clipping, raster lighting/texgen, bitmap pixels and WebGL1/2 validation
8480aed7f integer fog/light/material 6 Pname conversions, lighting, lists and direct/queued validation
bb2cef8fe indexed light/material queries 4 Stored-state conversion, ABI, lists and direct/queued validation
0cd72e8bb texgen setters and queries 6 Double transforms, numeric conversions and direct/queued validation
342d0433b texture environment and parameter queries 5 Context, unit and bound-object state with rendered pixels
cda0ebd87 user clip planes 2 Six eye-space planes, primitive clipping and WebGL pixels
451708e35 one-dimensional textures 3 Storage, sampling, integer vector parameters and rendered pixels
a8b85a569 texture copy/readback and pixel transfer 7 Framebuffer and texture storage, transfer, packing, lists and WebGL pixels
57cf33689 boolean and pointer queries 2 Byte-sized boolean conversion, saved array pointers and resolved ABI checks
3a49ef5e2 pixel maps 6 Per-context tables, numeric conversions, mapped transfers and literal pixels
93793279e raster pixels 4 Draw/copy pixels, zoom, pixel-store conversion and WebGL depth equality

Each family passed its focused semantic and regression checks, rejecting controls, and locked build; the coordinator verified retained evidence hashes and committed blobs. Texgen passed 672 literal checks and 560 matching observations per direct/queued route, two saved-source and 15 semantic rejecting controls. Texture state passed 437 literal checks and 367 matching observations per route, two saved-source and 12 semantic rejecting controls, and WebGL1/2 pixel checks. User clipping passed 154 literal checks and 98 matching observations per route, 40 WebGL1/2 pixel checks, two saved-source and nine semantic rejecting controls. Their final suites passed 62, 63 and 62 commands respectively, including focused regressions, generated checks and the locked build. The coordinator verified 71/79/77 remote evidence hashes and 11/22/23 tested committed blobs respectively; evidence was retrieved before each temporary boat stopped. Genuine 1D textures passed 308 checks and 91 matching full-frame observations per direct/queued route, 36 WebGL observations and 68 validation outcomes including 15 rejecting controls and the locked build. The coordinator verified 252 evidence hashes and 24 tested committed blobs; the temporary boat stopped after retrieval. Texture copy/readback and pixel transfer passed 29 positive regression/build entries, 367 checks and 360 literal oracles per direct/queued route, 35 WebGL checks per backend, and nine intended negative controls. The coordinator verified 45 evidence hashes, 24 tested committed blobs, accepted builds and 13 handoff hashes; the temporary boat stopped after retrieval. Boolean/pointer queries subsequently passed 272 literal checks per direct/queued route, saved-source and 16 compiled semantic rejecting controls, focused regressions and two locked builds. The coordinator verified retained evidence hashes and helper absence before accepting 57cf33689; its temporary boat stopped after retrieval. Pixel maps subsequently passed 935 checks per direct/queued route, 13 semantic and two saved-source rejecting controls, 20 focused commands and three locked builds. The coordinator verified 162 final and 96 retrieved evidence hashes plus 7,230 accepted source pins; its temporary boat stopped and guard was absent. These eight genuine exports reduce 72 to 64 mandatory misses. Raster pixels 93793279e subsequently passed 92 literal checks per direct/queued route, WebGL1/2 including depth-copy EQUAL, ten compiled semantic controls, saved-source rejection, 17 regressions and the locked build (32 final steps). The coordinator matched all 264 retrieved evidence hashes and 7,233 tested source inputs; helpers were absent and the temporary boat stopped after retrieval. These four genuine exports reduce 64 to 60 mandatory misses. Fresh real-worker integration is queued after the two game lanes; no original MDK2 run is claimed. Integer lighting passed 301 checks per direct/queued route, two saved-source and fourteen semantic rejecting controls, 36 regressions, two generated checks and the locked build (57 commands). The coordinator matched all 116 retained evidence hashes and 13 committed blobs. The four indexed queries passed 461 checks per direct/queued route with 373 matching observations, two saved-source and 13 semantic rejecting controls, 38 regressions, two generated checks and the locked build (58 commands). All 77 retained evidence hashes and eight committed blobs were independently matched. See query evidence. See the rectangle evidence. No original MDK2 replay, first window or gameplay has been qualified. See MDK2 notes.

Dark Reign 2's later dd82ceb03, run 20261101-dark-reign2-unit-movement, qualifies selected-unit movement after consumed ordinary Shift-M and a ground click. A matched neutral interval and stationary world landmarks distinguish unit translation from camera movement. Exact arrival of all six units, combat, mission completion and backend parity remain open. This reuses accepted evidence; no new original run or current-main pixel comparison was performed for this update. See the movement evidence.

November 1: accepted native progress and remaining backend gaps

These results reuse already reviewed original-game evidence. They do not establish software/WebGL parity or reconstruct the historical 27-app membership.

Title Accepted native evidence Remaining gap
Redline 0da33e84b, run 20261101-redline-world-loading-headroom: WORLD loading completes, firearm/HUD appears, and consumed ordinary Up moves the player toward the shutter/wall in paired before/held/released PNGs. One original exits 0 after 358.407 seconds; 70 original/configuration pins and 129 actual loaded-file hashes match. Current WebGL comparison, combat outcome, FPS and audio. Loading/cinematic changes alone were excluded from movement proof.
Dark Reign 2 d9aca1952, run 20261101-dark-reign2-native-mission-input: Demo Mission 1 terrain, soldiers, objective, command bar and minimap are visible; ordinary mouse selection is verified. One original exits 0 after 327.256 seconds; 29 original/manifest pins and 142 runtime-read hashes match. Later dd82ceb03 verifies selected-unit movement after an ordinary order (see November 2 reconciliation above). Mission completion, GPU mip support, backend parity, FPS and audio remain open. Arrow dispatch and changed framing alone were excluded from movement evidence.

The later native control runs add these bounded checkpoints:

Title Accepted native evidence Remaining gap
Outcast 039f7f5a2, run 20261101-outcast-original-weapon-control: ordinary weapon draw, aim, fire with ammo80→79, release and holster in reviewed PNGs, with consumed WM/DirectInput. One original exits0 after170.150 seconds. No enemy hit, projectile trajectory, impact, damage, kill or combat outcome. The initial NPC interaction was excluded from firing proof.
Dethkarz ca136b967, run 20261101-dethkarz-native-fire-control: reviewed before/Shift-held/released captures show a laser flash/bolt and weapon-bar response. Consumed WM down/up and actual GetAsyncKeyState returns0/8001/8000/0 establish the ordinary Shift path. Two originals exit0 after192.197 and80.305 seconds. Exact DirectInput keyboard consumption is unproved: observed DirectInput reads request16-byte mouse state. No impact, damage, combat outcome, charged-shot mechanics, FPS or audio claim.
Incoming 16f6dd886, run 20261101-incoming-native-phase-progress: reviewed hostile-fire captures show score0→250→500→6500, followed by PHASE COMPLETE and a helicopter FOLLOW WAYPOINT prompt. Consumed WM and acquired DirectInput keyboard events support the firing controls. One original exits0 after551.986 seconds. Later 4088515ae, run 20261101-incoming-helicopter-margin, proves ordinary helicopter flight in phase2/10: reviewed control-held/released captures show radar-dish flyover, banking and hill approach, with consumed WM/DirectInput controls. One original exits0 after472.113s. Waypoint arrival, objective change and helicopter phase completion remain open; the earlier transition prompt alone did not prove movement. No full playthrough, WebGL comparison, FPS or audio claim.

These later runs also reuse module dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4, retain unchanged original/source identities and cleanup evidence, and add no current WebGL comparison. See Outcast notes and Dethkarz notes, and Incoming notes.

MDK2 remains before its first window. The historical f0eda338f, run 20261101-mdk2-prewindow-cause, traces the first rejected lookup to glAccum: 212 requested GL/WGL names were rejected at that point, including 208 unimplemented procedures; the November 2 update above records the current 72-name remainder after the accepted implementations listed above. The loader enters error2 and then waits for an error string from its suspended file worker. One78.087-second original exits0 with a reviewed blank teal image and zero ordinary inputs. This is a concrete export-implementation blocker, not jump or gameplay evidence. The existing lookup fixture fails a stale-handle assertion; scratch-adapted checks pass the actual rejection ABI. Neither proves gameplay. See MDK2 notes.

The Redline and Dark Reign 2 runs loaded module dd43e4a6624d2449d7133b731d687c32530f7fcf0041a69ba71a09f78c09aaf4 and retained before/after identities and process-cleanup evidence. See Redline notes and Dark Reign 2 notes for exact commands, controls, observer limits and screenshot interpretation.

The subsequent unchanged GL transform and lighting fixtures passed in 0a33e2db4, run 20261101-gl-descriptor-transform-lighting (3.246 and 3.356 seconds, both exit 0). This covers production descriptor assertions; it adds no game-pixel, GPU or full-corpus acceptance. See the fixture note.

The unchanged stencil/depth fixture also passes in 3fc209ff3, run 20261101-gl-stencil-depth (7.990 seconds, exit 0). Assertions cover masked and scissored clears, alpha/stencil/depth ordering, stencil comparisons and operations, allocation failure and resize, context/attribute/display-list state, and immutable queued draw snapshots. The expected allocation refusal and caught unsupported readback trap belong to the passing assertions. All 12,982 production function bodies match the accepted module; three fixture functions/exports and an omitted custom section mean the whole module is not byte-identical. This establishes bounded native software and manual queue-replay coverage, without proving real worker-thread execution, GPU parity, original-game pixels, general crash freedom or the full 27-app goal. See the stencil/depth note.

October 31: known-source browser gameplay and remaining rendering blockers

Shogo 7c510cfe0, run 20261031T2006Z-shogo-pinned-browser, shows first-person street walking, a Bullgut pickup, turning and Ctrl firing (ammo40 to38). The coordinator inspected captures12/13/14, including weapon flash. All69 local target pins matched before/after, along with source, compiler, module and served bytes. Cleanup verified all23 owned runtime processes absent, ports closed, profile gone and browser lock free. New Game/difficulty screens were not separately captured; targeted hits, kills, mission progress, audio and FPS remain unproven. This owner-pinned local build has no independent upstream package comparison. See Shogo notes.

Blood II 64c9d009b, run 20261031T1949Z-blood2-pinned-browser, now shows native elevator exit, walking, turning and pistol fire (ammo50 to48). The coordinator inspected captures09/11, including muzzle flash and casing. All98 owner-pinned local BUILD109 files plus the manifest matched before/after; source/compiler/module and served-byte checks passed. All24 runtime processes exited and both ports/profile/browser lock were released. Health100 to29 to0 and the death camera were observed; the cause and any enemy hit, kill, objective, jump, audio or FPS remain unproven. See Blood II notes.

SiN b16c76d3c, run 20261031T1953Z-sin-pinned-browser, reaches native menus and the scripted HQ/helipad/city route into first-person Air Assault on FCB. Captures13/16 show turning; capture17 ends at RESTART LEVEL. Forward translation, settled on-foot play, firing and jumping remain unproven, and no restart cause or emulator repair is established. All23 local pins matched; runtime cleanup completed at20:01:05. Evidence applies to tested 0d9913225; later publication merged unrelated font/runtime changes without executing them. SiN 2d14f5852 adds reviewed Air Assault aim, barrel animation and wall/roof impact particles; no enemy damage, objective or exact restart cause is proved. FreeSpace 9ed9d4b54 completes genuine DX7a setup pending four WININIT DLL replacements; flight remains untested. The reboot follow-up retains that checkpoint pending authentic Win98 ordering and failure/retention semantics; its offline prototype is uncommitted and has not completed installation. Nox subsequently gained native Bunker movement and reviewed jumping/landing in 20261031T2127Z-nox-jump-action. See FreeSpace notes, Nox notes and SiN notes.

Heretic II 6e91d9608, run 20261031T1941Z-heretic2-original-browser, reaches tutorial and Easy docks gameplay with native walking, turning, spell attacks and jumping/landing. The coordinator inspected captures07/13/32/35. All1825 upstream package members matched; source/compiler/module and fixture pins matched before/after. All24 runtime processes exited, both ports closed and the browser lock was released. No hit, kill, objective, audio or FPS is established; tutorial-caption carryover and health100 to94 remain unexplained. See the gameplay notes.

Darkstone accepted combat reconciliation

The accepted October30 software result be70e8f86, run 20261030T2206Z-darkstone-software-combat, already establishes native manual skeleton combat: EXP0 to79 and surviving LIFE17/35 after a held-left attack. Its screenshot review and publication are recorded in the processed completion ledger; no repeat runtime or new acceptance is needed. WebGL skeleton combat was accepted separately in dc988e7cb, run 20261030T2158Z-darkstone-combat. Both use retained module 4649ee5f at committed main 10952264c; this is not current-main parity, FPS, audio, save/reload or child-renderer qualification. The outer harness exit143 remains unexplained despite local cleanup and acknowledged quit. See the original combat notes.

Blood II's check dbd1a8adf, run 20261031T1932Z-blood2-native-browser, stopped before runtime. The owner subsequently designated the local DEMO V1.0A (BUILD 109), pinned by its 98 file hashes, as the target build. All 98 files and the manifest were rechecked in 20261031T1950Z-blood2-owner-pin. The 7.17 GB compilation is no longer a provenance target or prerequisite; do not download it. The browser lane is unblocked. See the exact executable and manifest pins. Historical table entries below do not establish a new current-main run.

UT2003 ordinary visible-menu SOFTWARE route is now verified by 62ee90c86, run 20261031T1628Z-ut2003-native-menu-play: native Instant Action, Antalus and PLAY reach a joined match without console input. The coordinator inspected joined HUD18, firing24 (ammo150 to148, health100) and scoreboard27 (remaining 14:42; previous26 was14:46). All25 saved focused checks pass, with the same frozen9b4e39a2/295 source/seven compiler/881 original/107 served identities. Native press/release drained the DirectInput ring; the retained menu image covered loading. Additional Enter/click during loading prevent attributing joining to one press. No generic emulator defect was demonstrated. All runtime exits0 and FULLRELEASE16:36:32; actual SOFTWARE executor has no queue error. This supersedes the software menu gap below. Matching original ordinary-menu WebGL gameplay is now accepted in 09cedce52, run 20261031T1653Z-ut2003-webgl-menu-execution: inspected joined HUD17, firing23 (ammo150 to149, health100), and scoreboard26 remaining14:48 (previous25 was14:51). Native movement/look/fire completed; owning WebGL2 receiver completed19,248 with pending0/failedfalse, and24 focused checks pass. The same frozen module/source/input closure was retained. All runtime exits0, FULLRELEASE16:58:19. This closes the UT2003 ordinary-menu backend gap; full27, FPS, audio, player kills and match completion remain open.

UT2004 documentation 2342e9288 records original DM-Rankin software CLI play in 20261031T0427Z-ut2004-match: native movement, look and firing, health100 and ammo100 to96 then94. The coordinator inspected the readable firing frame. The runtime completed 8,425 batches in636.949 seconds and exited zero; 998 original inputs and1,283 retained pins verified unchanged. The default guestWorker browser follow-up 09eb0ecbe also verifies original DM-Rankin movement, look and firing on the actual software endpoint: health100, ammo100 to99 then98. The coordinator inspected the readable world/HUD frame. Run 20261031T0451Z-ut2004-browser verifies 998 original inputs, 292 source files and 107 served responses. Repair fb698d57e then moves the windowed-device HWND into shared atomic process state so presentation instances retain the correct client origin24,44. Run 20261031T0514Z-ut2004-origin verifies the actual software browser route on new module 9b4e39a2, with its exact295 source files and seven compiler inputs retained. The coordinator inspected the complete top clock and bottom HUD during native movement/look/fire, health100 and ammo100 to98. Visible DeathMatch and PLAY clicks work at their displayed coordinates. The full build and focused real-instance/pixel/Reset/DirectDraw tests pass; the new regression rejects the saved old source. Enemy hits/kills, match completion remain unverified. Documentation f81b9a35f closes the WebGL native match arm in 20261031T0629Z-ut2004-webgl: visible menus reach DM-Rankin, W/S movement, mouse look and firing100 to98. The coordinator inspected initial and changed-view full world/HUD frames. Actual owning WebGL2 completed68,885 GL draws and20 LockRect calls;26 trusted inputs, 998 originals,295 source files, seven compiler files and107 served pins pass. The fresh locked build passed. Render640x480 inside browser1024x768; ANGLE SwiftShader, no hardware-GPU claim. All19 owned PIDs, watcher and ports are cleaned. FPS/audio/kills/match completion remain open. See the UT2004 native route.

UT2003 diagnosis b2647538c in 20261031T0531Z-ut2003-readback attests actual WebGL endpoint1 on the same exact 9b4e39a2 build. DRAW29 rejects a multi-level cube texture with LOD bias-0.5; READBACK48 inherits that queue error, and the first LockRect completes with INVALIDCALL after token clearance and ESP+20. The coordinator inspected the original crash dialog and reran the causal check successfully;881 original inputs,295 source,7 compiler and107 served pins verify unchanged. Repair 5fa12b1be now implements generic cube LOD sampling with face-local mip atlas/filtering. All8,880 WebGL1/2 pixel checks and the full locked build pass; saved old source rejects. Run 20261031T0542Z-ut2003-cube-lod retains fresh 9b4e39a2, exact295 source/seven compiler inputs and881 originals, 107 served pins and88 trusted native events. DRAW29/READBACK48/LockRect complete successfully. The coordinator inspected native DM-Antalus world/HUD, health100 and ammo146 after movement/look/fire through the game's console open DM-Antalus route. Actual WebGL2 records21,049 completed draws with no failures. Runtime cleanup recovered an interrupted finalizer and verifies all owned processes/ports/profile closed. Menu PLAY, FPS, audio, kills, match completion and a matching software-arm qualification remain unverified. Software follow-up a14e5e309, run 20261031T1533Z-ut2003-software-antalus, stopped before guest input: its custom frozen-response handler omitted COOP/COEP, leaving isolation and the default guest worker inactive. The inspected NVIDIA startup and 85 completed software operations do not qualify gameplay or establish an emulator defect. All 295 source, seven compiler and 881 original inputs verify; all runtime resources closed. The next attempt must verify HTTP isolation headers and actual isolation/default guest worker before native input. Follow-up 98e807e69, run 20261031T1543Z-ut2003-software-isolated, fixes those scratch-response headers and verifies isolation/default guest worker before native input and throughout the route. The coordinator inspected software Antalus at “Fire to Start”; the ordinary console route remains explicit. Actual software execution completed12,389 operations with no queue error. The disk guard stopped before join; queued join22/23 never executed. No joined HUD, movement, look, firing, health/ammo/clock progression or software gameplay qualification. All frozen inputs and22 focused checks pass; all runtime exits0 and whole-process cleanup verifies. The next bounded join attempt requires more disk headroom. cd0272077, run 20261031T1603Z-ut2003-software-join, then correctly declined launch: actual predecessor cleanup/process gate passed, but free space was57,539,072 bytes below the2.6GB reserve. Its14 preparation checks pass; there is no new runtime or gameplay evidence. The next attempt follows disk recovery and uses a2.8GB initial reserve. Follow-up 950e4940e, run 20261031T1612Z-ut2003-software-join-reserve, completes the software gameplay arm on the same frozen identity. The coordinator inspected joined HUD15, forward17, firing21 (ammo150 to148, health100), and scoreboard24 elapsed00:33 (earlier00:24). Native console open DM-Antalus and FIRE join remain explicit; visible menu PLAY is unresolved. Actual software Device gpuFALSE completes301,598,102 samples; final queue25,202 completed of 25,220 submitted, no queue error. All17 captures,63 balanced trusted events, 107 served pins and24 focused checks pass. Disk minimum2,701,918,208 bytes; runtime exits0 and full process cleanup verified16:17:30. FPS/audio/hits/kills and match completion remain unqualified. A fresh worker investigates visible menu PLAY after the current Diablo browser releases, without repeating accepted console gameplay.

Deus Ex documentation 940371775 verifies the default guestWorker browser Training route on actual SoftDrv: native approach/retreat, first door opened with right mouse, then crossed into the red-floor room. The coordinator inspected that gameplay frame. Run 20261031T0447Z-deusex-browser verifies 131 original/support inputs, 292 source files and 105 served responses. Follow-up 2546fb407, run 20261031T0606Z-deusex-key-door, verifies NanoKey pickup, slot0/left-click unlocking, right-click opening and traversal of the next locked door. The coordinator inspected pickup/unlock/corridor frames;74 trusted inputs and the same131 original/292 source/105 served pins pass. Actual guest canvas640x480 differs from browser viewport1024x768; 1024x768 guest rendering is unqualified. Runtime cleanup verified. Training completion and FPS remain unverified. Follow-up 3d04b4841, run 20261031T0617Z-deusex-next-exercise, verifies crowbar pickup/equip and breaking an upper wooden supply crate with native input. The coordinator inspected pickup, intact/broken crate and Jaime lock-picking follow-up frames; 102 trusted events/five interaction deltas and131 original/292 source/105 served pins pass. Runtime cleanup verified. Lock-picking was not attempted in that run. Later accepted SoftDrv/default-worker runs progressed through the training exercises to the crate room. 208c41480 retains the preferred native count7 checkpoint with inspected fresh-page Load/gameplay validation in 20261031T1220Z-deusex-crates. Grounded-jump follow-up b47dfcf2b, run 20261031T1403Z-deusex-grounded, establishes positive vertical velocity before forward input, but insufficient stationary range and tall-wood collision. 2fde12461, run 20261031T1450Z-deusex-reposition, creates a genuine native two-small-crate stack: the exported upper actor names the lower actor as Base. The coordinator inspected restored gameplay, the stack and the failed climb; 450 balanced trusted events, 56 captures, frozen module/source/original pins and cleanup checks pass. Stack access remains unresolved; no stable96 landing or training exit was reached. Follow-up cc4aa4124, run 20261031T1521Z-deusex-stack-recovery, validates that exact diagnostic count8/time01:22:17 through native Load, original actor/save crosscheck and inspected gameplay. Count7 remains preferred; the different 1301Z diagnostic count8 is not promoted. A native air-crouch attempt reached positive vertical velocity 300 but retained height 43 during sampled flight, reaching crouched height 16 only after returning to the old wood support. It establishes no new stack footing, intermediate support, 96 landing or exit. The interrupted original worker's launcher exit 143 remains unexplained; recovery closed the inherited runtime with actual Chrome/controller/wrapper/monitor exits of 0. These runs retain module 8459597c, 292 source files, 131 original/support inputs, 105 canonical served responses and19 DLL mounts. They establish further native SoftDrv gameplay, without qualifying a matching WebGL renderer, FPS/audio or training completion. See the Deus Ex browser notes.

Pirates documentation a7725adaf verifies native sailing near St. Kitts and Nevis in 20261031T0458Z-pirates-knownsource; the coordinator inspected the ship/HUD and persistent white coast. Four matched mesh sequences retain attached properties into effective renderer state; ordinary map slots are empty while a separate array is populated. Follow-up 83ae79866 in 20261031T0520Z-pirates-maps proves the original separate-map writer and encoded reader contract. Four matching writer/default-setup mesh sequences include two actual large-draw correlations; the coordinator inspected the native coast image. Selection follow-up de5d26f5f in 20261031T0535Z-pirates-selection proves original terrain name selection and encoded NSB stage references. The original factory returns null for66 terrain requests; nine same-mesh correlations explain subsequent default-shader draws. The coordinator inspected sailing.png and reran the causal evidence check successfully. At that diagnostic stage, rejection inside original factory lookup 5bd720 remained unresolved; cache/library lookup, requirements and creation must be distinguished. Those diagnostic runs reproduced no crash and landed no repair; 1,374 original inputs, 292 source files and467 served responses verified.

Subsequent production fixes c6fa4878a (bootstrap retirement) and 44d8b27fd (D3D9 MaxStreams cap16) are on main. Later original default-worker WebGL runs reach textured coast, town/port, sailing and combat. The stream-cap fix is distinct from the earlier shader-factory diagnosis; these outcomes do not prove every factory-null event fixed. Accepted 70589baac, run 20261031T1423Z-pirates-native-turn, reaches the LovelyMary selector, actual hull contact and a fading native boarding-deck duel with50 Englishmen and40 pirates. The coordinator inspected selector/contact/deck images; 108 balanced trusted events,57 captures and cleanup checks pass. No duel was played to an outcome, no capture was established and no new save was accepted. Final device error is unknown; one readonly VFS helper failed, so no all-helper-success or VFS-validation claim applies. Frozen module 078d81cb retains296 source files, 40 support/compiler inputs,1,374 originals,106 source URLs and3 auxiliary pins. Follow-up 18aa9d6a5, run 20261031T1503Z-pirates-boarding-duel, stalls at the native title on both page attempts. The coordinator inspected that failure; 34 balanced DOM-trusted events do not prove guest consumption. No native Load, new combat or fencing occurred in that run. Earlier boarding evidence remains valid. Cheap follow-up ce79828e8, run 20261031T1515Z-pirates-title-load-cause, compares 13 retained recipe files and finds identical bytes. It establishes no causal harness drift, guest input consumption or supported repair and launches no new runtime. Title-to-Load recovery, duel outcome, software parity, FPS/audio and full corpus qualification remain open. See the Pirates boundary notes.

Diablo II follow-up a58fb9844, run 20261031T1523Z-diablo2-den-checkpoint, retains four whole unedited native DENMAP save files and verifies fresh-page character selection, native Load, inspected town gameplay and ordinary movement with actual isolation/default guest worker/WebGL. The coordinator inspected the restored gameplay frame. The earlier page lacked isolation/default guest worker because its custom responder omitted headers; its Blood Moor Zombie combat is outside the requested worker configuration and does not qualify that target. Den entry remains unverified. Frozen module 7f9da4e9 has unknown compiled-source identity; 144 frozen inputs and 25 original runtime files verify unchanged. The disk guard stopped cooperatively above 2 GB and all owned runtime resources closed. This is native checkpoint evidence, not full both-backend corpus coverage. Follow-up 23affb46a, run 20261031T1543Z-diablo2-isolated-den, verifies isolation/default guest worker from launch through all59 captures. The coordinator inspected native restored town08 and Blood Moor combat82; 124 balanced trusted events and frozen/original/save identities pass. The disk guard stopped before Den entry or a newer checkpoint; no new save is claimed. All runtime exits0 and whole cleanup verifies at16:03:08. See the Diablo II native restore notes.

Forsaken documentation 50bb30050 records a known-source WebGL frame investigation in 20261031T0427Z-forsaken-frame. The coordinator inspected thrust-short-gpu.png: briefing text, loading bar and incomplete world exist inside the GPU color target, alongside projectiles. Earlier firing frames show a readable room, but sustained readable world coverage remains unverified. Follow-up 20e6b97a8, run 20261031T0552Z-forsaken-boundary, retains two complete clear-to-clear intervals, one Flip each. All16 owning Execute/transport/GPU submissions match. Six world draws per frame contain 670 triangles already collapsed at Execute entry and change zero pixels; nondegenerate object draws do change pixels. Exact565 uploads from alternating backings explain room versus briefing fragments; depth-only clear preserves color. The coordinator inspected both frame colors and reran the offline check:873 original/292 source/107 runtime pins pass. Follow-up 24b5bd55a/7867342ee, run 20261031T0617Z-forsaken-first-mutation, reconstructs 876 initialization deltas and 29 original color-only stores: engine-transformed output becomes the next source through Unlock1417, diverging at PROCESSVERTICES1420. The startup/attract trace does not independently identify Volcano. A fresh native Volcano retest of separate-output candidate 41f5c4c7 preserves 162 source vertices over 12 Executes; two complete frames redraw 642 world triangles. The coordinator inspected a complete room/HUD and changed pitch with firing 2000→1998, and reran mutation/source/redraw checks. Owned processes and ports are closed. Repair df174b3cf is now on remote main. The final allocation- failure guard passes injected OOM/retry, 32 color-only cycles, six focused tests and the full locked build; both old-source and defective-guard mutations reject. Final module d2c1df73 pins 295 runtime sources and seven compiler files. Native candidate screenshots remain distinct. Fresh final-build run 20261031T0713Z-forsaken-repaired-gameplay, commit 501f52cf1, now verifies native Volcano room-to-corridor movement, look and repeated firing. The coordinator inspected four room/corridor/instruction captures; 873 originals, 295 sources, seven compiler inputs, 609 responses and 95 trusted events pass. All 59 observed owned PIDs are absent, ports closed and profile gone. Deaths/respawns and white pause frames are disclosed; outer wrapper143 is unexplained despite browser exit0 and verified cleanup. Enemy kills, hidden-door opening, uninterrupted play, hardware GPU, FPS and completion remain unverified. See the Forsaken investigation.

The earlier gameplay runs pin module 8459597c to the retained ReVolt locked build's exact 292-file source snapshot, including then-concurrent font edits; observed HEAD is not substituted for compiled source. The UT2004 repair uses the separately retained 9b4e39a2 source/compiler snapshot described above. Own runtimes were cleaned. These results extend the per-app evidence without establishing software / WebGL parity, audio, real-time performance or completion of the full27.

October 31: Re-Volt attachment repair reaches native WebGL driving

Repair 59a615fc5 fixes GetAttachedSurface matching for allocations created before DirectDraw 7: unused DDSCAPS2 extension words no longer reject an otherwise valid back buffer. DX7 allocation validation remains active. Run 20261031T0403Z-revolt-backbuffer reproduces the original failure on locked build 44aa9976 and reaches the menu, starting grid and native Up/Right/Down driving on repaired build 8459597c. The coordinator inspected the race frame: the car is off-road and overturned. Movement is verified; checkpoint, lap and race completion are not.

Both builds retain exact source snapshots, including concurrent dirty font work; neither is described as a clean commit build. The default guest-worker route uses explicit WebGL at stable 1024×768, with all 444 original input files unchanged before and after. Locked builds and the focused native COM regression pass; the new regression also rejects the saved pre-fix source. Two older synthetic fixtures fail identically before and after, outside the repaired lookup. Own browsers, controllers, ports and profiles were cleaned; the after controller's exit 143 remains unexplained. No software gameplay or full-corpus parity is established. See Re-Volt investigation and the run's retained recovery recipe for the next controlled-lap attempt.

October 31: Tunnel, Globe and Wormhole startup crash sweep

Run 20261031T0240Z-dx-more-startups launched registered dx_tunnel, dx_globe and dx_wormhole through the cooperative CLI software route, without injected input. All three exited zero with no CRASH, RuntimeError, unreachable or unimplemented-API signature in their logs. Tunnel completed 2,905 batches in 15.064 seconds, Globe 16,834 in 15.130 seconds, and Wormhole reached 100,000 batches in 9.700 seconds. Personally inspected PNGs show textured tunnel walls, the registered textured Globe scene, and Wormhole's yellow-green wire grid. Globe's appearance is not a fresh scene-correctness finding; preserve the original-scene distinction in its investigation note.

The module hash is 8333a35b1efa694462033e87f62ce4a3f5497941d1423540b39a590c9d31134a; compiled-source identity is unknown. Module, region-map, registry and runner hashes stayed unchanged across execution, as did all 38 registered input files per app. Exact commands, process exits, hashes and observations are retained in executions.json, inputs.json, after.json, assets.json and summary.json. No source changed or emulator defect was established. These bounded software startup observations do not establish WebGL parity, gameplay, audio, performance or completion of the historical 27-app set. No matching candidate manifest entry was identified, so this remains investigation evidence without a dashboard result.json. All three child processes ended; no browser, server or Boat was created.

October 31: bounded Twist and Boids cooperative startup crash sweep

Run 20261031T0237Z-dx-launch-sweep launched the registered dx_twist and dx_boids executables with test/run.js --no-build --no-threads, without injected input. Twist completed 6,641 batches in 15 seconds; Boids reached the 100,000-batch limit in 5.892 seconds. Both runners exited zero, with no CRASH, RuntimeError, unreachable or unimplemented-API signature in the preserved logs. Personally inspected final PNGs show the textured Twist object and colored Boids flock triangles with cyan boundaries.

The loaded local module hash is 8333a35b1efa694462033e87f62ce4a3f5497941d1423540b39a590c9d31134a. Module, region-map mirror, runner and registry bytes stayed unchanged across both executions; their hashes and exact commands are retained in inputs.json, executions.json and summary.json. Compiled-source identity is unknown; the launch-time HEAD is not proof of that identity. These are cooperative CLI software startup observations, not new WebGL, threaded, scene-parity, gameplay, audio or performance acceptance. No matching candidate manifest ID was found, so the bundle remains investigation evidence without a dashboard result.json. Both child processes ended; no browser, server or Boat was created. No source changed or emulator defect was established.

October 31: bounded OrganicArt original-input recovery found missing source

Run 20261031T0058Z-organicart-recovery searched the accessible retained local sources without repeating the native logo-toggle or substitute-image controls. No authentic CALogo.gif or source-verified Plus! 98 installation registry defaults were recovered. Consequently no new scene replay or production change is claimed.

The search inspected 538 ZIP indexes and 104 TAR/TGZ member lists, plus one ISO listing. Its only OrganicArt executable match was a byte-identical copy of the existing OASAVER.SCR in /tmp/caesar3-packet.tgz; it supplies neither the missing logo nor installation settings. All 47 CAB listing attempts failed on unrelated retained game installers' InstallShield format; their contents remain unverified. A UTF-8/UTF-16 settings search over 380 retained INF/REG/INI/THE/SCN files found no logo/default-setting evidence. The original executable's 57-resource inventory has no named CALogo resource and the PE has no GIF87a/GIF89a signature; encoded/compressed assets are not ruled out.

The concrete missing input is the authorized original Plus! 98 installation media/metadata or a verified installed logo/settings export for this package. The historical source lead in directanimation-design.md:146 is /Volumes/PLUS98/PLUS98.CAB and memory reference_plus98_iso.md; neither is present here. This lead does not establish that the CAB contains the logo. Native fallback values do not establish installation defaults, so an arbitrary logo-disable setting remains unjustified.

The run preserves the search recipe, archive results and limitations, 192-file screensaver SHA-256 manifest, original archive/member hashes and verification.json: all 85 prior registered inputs and 192 retained files remain unchanged. Original OASAVER.SCR SHA-256 is 470e28bcbc3cf92c514cd8e8bbf9eae399c59f79ed2adbf7bf2b91d25cfb477c. No browser, server, profile or Boat/test box was created; no runtime cleanup is outstanding (box ID/state/expiry not applicable). This was a local source recovery attempt within the 25-minute bound, not a fresh rendering test.

October 31: OrganicArt rectangle isolated to native logo input

Run 20261031T0046Z-organicart-logo supersedes the startup-only native registry comparison in 20261031T0034Z-organicart-rectangle/logo-ab. Both new WebGL arms reached original scenes by 45 seconds and ran for 210 seconds. With native DisableCALogo DWORD 0, personally inspected 45/150-second scenes contain the lower-right white rectangle; with DWORD 1, inspected 45/150/210-second scenes do not. Read-only guest observations confirm the native disable flag is respectively 0/1. The enabled 210-second scene also lacks the rectangle, consistent with the app's hide-logo path; this is not a permanent obstruction.

The white area is present in both the 640×480 backing canvas and DirectDraw frame layer. In the saved enabled layer, all 3,060 pixels of [550,434,640,468) are white; the corresponding disabled sample has zero white pixels. A separate native trace records DisableCALogo -> 0, no CALogo override, and SearchPath("CALogo.gif", path=".") -> NOT FOUND. No such file is among the 84 registered scene assets or available screensaver files. A fourth launch keeps the logo enabled and sets the app's native CALogo REG_SZ to the unchanged, already-registered BOLD6.GIF: lookup and file reads succeed, and personally inspected scene/layer captures show an orange/brown texture in the former white rectangle. This deliberate image substitution diagnoses the missing default logo input; it does not restore the authentic Computer Artworks logo.

No production code or default registry setting was changed, and no guest memory was patched. A renderer repair is not established. The remaining repair dependency is authentic Plus! 98 logo assets and/or installation registry defaults, followed by an original-input replay. The native disable key is a demonstrated workaround, not an accepted production repair. Randomized scenes differ between arms; this is not same-frame parity, and software was not rerun in this follow-up.

All four launches use original OASAVER.SCR /s, the same 85 executable/scene input hashes, retained module SHA-256 d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5 (source fix de2e338c4, full source unknown), Chrome 151/ANGLE SwiftShader, and guest threads disabled. Exact capture hashes, native traces, byte checks, reproduction instructions and cleanup are retained in the run's investigation.json, verification.json and recovery.txt. Concurrent SDK Viewer changes to lib/apps.js and generated file-size metadata were handled by recovering byte-identical served versions from launch commit 49340601e; all 920 served-response hashes verify against current or retained exact bytes. The first controller was later reaped with status 143 after its cleanup record; the signal's sender is unknown. Both later controllers exit 0. All three browsers are absent, six HTTP/CDP ports closed and profiles removed. Browser work ran approximately 00:45:50–00:56:32 UTC, within 25 minutes. No Boat/test box was used (ID/state/expiry not applicable).

October 31: OrganicArt rectangle reproduced on both routes

Run 20261031T0034Z-organicart-rectangle reproduces the lower-right white rectangle over original OrganicArt scenes on software and WebGL. Personally inspected WebGL frames 2/3 and software restored.png show different scenes, so this is not a same-frame parity check. WebGL records one visible 640×480 Screen Saver window and host GPU enabled. Software point inspection places the rectangle inside the displayed canvas, rather than a separate HTML overlay. Its owning drawing operation and cause remain unresolved; no emulator repair is claimed.

All three launches used retained module SHA-256 d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5 (source fix de2e338c4, full source unknown), Chrome 151 with ANGLE SwiftShader, registered /s startup and guest threads disabled. Each launch verified 85 executable/scene input pins and 230 served-response hashes. The short software arm captured startup only. A secondary CDP connection temporarily resized the longer software arm; its inspected capture followed viewport restoration. Layer extraction failed when that target closed, so no backing-layer evidence or cause is inferred. All browsers and servers closed and profiles were removed; no Boat was used. investigation.json records these limits and exact captures.

October 30: PTCT paired original browser rendering

Run 20261030T2351Z-ptct-paired-browser captures the original registered Please the Cookie Thing intro on software and WebGL. Two personally inspected frames per route show gray/white geometric beams changing shape and position on black. The software route reports gl_sw_enabled() == 1 without a GPU presentation layer; WebGL has an owning GPU layer whose sequence advances from 132 to 169. Normal startup uses the registered dialog command, without a guest patch. Guest threads were disabled.

Both routes loaded canonical module SHA-256 d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5 (source fix de2e338c4). Each registration's executable and all 144 served response hashes were verified. Chrome 151 was launched with ANGLE SwiftShader flags; the GPU vendor string was not queried. Both browsers and static servers closed, private profiles were removed, and no Boat was used. Exact timestamps, input pins, capture hashes and cleanup checks are in investigation.json and the two route verification.json files. This establishes visible animation, not full intro correctness, same-frame parity, FPS, physical GPU performance or original 27-app membership.

October 30: six screensavers, consolidated paired browser captures

Run 20261030T2340Z-screensaver-browser-batch completes inspected scene captures for the six siblings of Architecture on both browser routes:

App Personally inspected result
Falling Leaves Leaves and shadows change over the textured leaf background on both routes.
Geometry Gold polyhedra change arrangement over the blue radial background on both routes.
Jazz White musical-note clusters change on black on both routes.
OrganicArt Software changes from a patterned tunnel to rings; WebGL rings change orientation. A white lower-right rectangle remains unexplained.
Rock & Roll Guitars and picks change orientation on both routes.
Sci-Fi Birdlike objects render over the textured alien landscape on both routes; motion was inspected only on software.

These are original registered executables and scene assets, launched normally with /s, guest threads off, Chrome 151 and ANGLE SwiftShader. Every launch served retained canonical module SHA-256 d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5 (source fix de2e338c4). Per-frame owning-host GPU state matches the selected route. The six registrations contain 114 executable/scene input entries, all hash-pinned. Of 2,270 served responses across the batch and retries, 2,263 were reverified against current bytes or retained matching HEAD bytes. Seven intermediate lib/host-audio.js responses cannot be reverified after concurrent worker edits; their exact recorded hashes remain in the route verification records. This is an explicit provenance limitation, not a full immutable source replay.

Initial WebGL Jazz captures caught startup; a longer fresh run produced the inspected notes. The software browser target closed after OrganicArt, with no established cause; a fresh bounded run captured only Rock & Roll and Sci-Fi. All four browsers are absent, static ports closed and private profiles removed; no Boat was used. investigation.json records exact inspected images, verification counts and limitations. This establishes visible scenes, with motion where stated, not same-frame parity, FPS, physical GPU performance, complete scene correctness, complete saver cycles or the full 27-app goal.

October 30: Architecture browser scene on both routes

Runs 20261030T2338Z-architecture-webgl and 20261030T2338Z-architecture-software launch registered original ARCHITEC.SCR /s, with its seven registered scene assets, through the normal browser launcher. Personally inspected WebGL frames 2/3 and software frames 1/2 show the textured marble interior and changing central green objects. WebGL frame 1 records startup's white area; it is superseded by the later rendered captures, not counted as a scene failure.

Both Chrome 151/ANGLE SwiftShader runs load retained canonical module d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5 (source fix de2e338c4), with guest threads off. Backend records verify the owning host's GPU enabled only for WebGL. All eight registered input hashes and every served-response hash are verified separately in each run. Exact identities, inspected frames and cleanup appear in investigation.json. Both browsers are absent, their six ports closed and private profiles removed; no Boat was used. This establishes a visible animated scene on both routes, not same-frame pixel parity, FPS, physical GPU behavior, a full screensaver cycle or completion of the historical 27-app set.

October 30: longer Boids captures establish flock rendering

Run 20261030T2348Z-boids-scene-diagnosis extends the short paired captures below. Two personally inspected frames per route now show colored flock triangles and cyan boundaries changing on both software and WebGL. The short captures did not establish a missing-geometry defect; no emulator change was made. Both launches served the retained canonical module d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5 from source fix de2e338c4. All 38 registered inputs and 179 served response hashes per route verified against local bytes. Host d3dimGpu was present only on WebGL. Chrome 151 used ANGLE SwiftShader; page Worker threading was requested off. The per-app wine.threads default remained true and does not prove Worker execution; no thread-mode comparison was performed.

The browsers and static servers are closed and profiles removed; no Boat was used. This proves observed flock animation, not same-frame parity, physical GPU behavior, FPS, full scene correctness, or recovery of the historical 27-member list. Shared SDK Viewer aliases retain their documented provenance.

October 30: paired Boids motion, incomplete scene acceptance

Runs 20261030T2335Z-dx-boids-webgl and 20261030T2335Z-dx-boids-software retain the original boids.exe and each contain two personally inspected browser frames. Cyan polygon outlines change position on both routes. The images do not establish a correctly rendered flock; this is narrower than the earlier colored-triangle observation. Both Chrome 151/ANGLE SwiftShader runs loaded canonical module d83e413f… (source de2e338c4) with guest threads disabled. Owning-host GPU state was verified enabled for WebGL and disabled for software. Each run verifies all 38 local SDK-directory input hashes and all served response hashes; the input manifest includes the renamed ProgressiveMesh aliases described below. investigation.json records exact module identity, backend and frame paths. Both browsers were stopped, all six ports closed and private profiles removed. No Boat was used. Full scene correctness, pixel parity, FPS and historical corpus membership remain unqualified.

October 30: Globe motion verified on both repaired-build routes

Runs 20261030T2330Z-dx-globe-webgl and 20261030T2330Z-dx-globe-software each contain two personally inspected captures of the original SDK Globe executable. The shaded, textured sphere changes orientation on both routes. This closes the earlier unverified WebGL motion finding after GPU Execute extent repair de2e338c4; it does not resolve the historical texture-band correctness question or establish pixel parity.

Both Chrome 151/ANGLE SwiftShader runs loaded retained canonical module d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5. WebGL had the owning host's d3dimGpu and GPU selection enabled; software had neither. Guest threads were disabled in both. Each run verified 38 local SDK-directory input hashes and all 182 served-response hashes. Its investigation.json records frames, build/backend evidence and cleanup: browser absent, three ports closed, private profile removed, no Boat. The different orientations and sphere sizes are not a same-frame comparison; FPS, full scene correctness and completion of the historical 27-app corpus remain unqualified.

October 30: original Tunnel WebGL motion verified after extents fix

Run 20261030T2325Z-dx-tunnel-gpu-extents retested original dx_tunnel with the canonical module retained from the passing build for GPU Execute extent repair de2e338c4. Two personally inspected browser captures show changed tunnel orientation and texture position. A bounded trace of 284 copy/present events contains 94 distinct scene-region hashes on both primary surface and render target, excluding the text overlay.

Chrome 151/ANGLE SwiftShader loaded module d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5. All served-response hashes match their local inputs. The run's result.json records module identity, captures and verified cleanup: browser exited, three ports closed, private profile removed, no Boat. This closes the earlier unverified WebGL motion finding for Tunnel; pixel parity, full scene correctness, performance and completion of the 27-app corpus remain unqualified.

October 30: Twist WebGL animation repaired

The original dx_twist browser run 20261030T2320Z-dx-twist-gpu-extents now visibly deforms and rotates its textured sphere. Three captured frames were personally inspected. The bounded copy trace contains 101 distinct scene-region hashes on both the primary surface and render target; changing copy rectangles cover the sphere. Before the fix, run 20261030T2316Z-dx-twist-copy-trace copied only the FPS text rectangle after the initial frame, leaving the primary scene unchanged despite target updates.

The GPU executor omitted ExecuteBuffer status extents that D3DRM consumes to choose its windowed copy rectangle. Accepted draws now report conservative, clipped coverage while Execute still owns that status, independently of later batch submission or readback. Wrong-target, inactive and empty reports are ignored. Triangles, lines, points, clipping and deferred submission have focused coverage, including the actual compiled WASM export.

The locked canonical build and five focused D3DIM tests pass. The original Chrome 151/ANGLE SwiftShader run loaded retained module d83e413fbc38ca5aa5447a7be62d10010783fd5490c98ec07c076553897950e5; served-response integrity and browser/port/profile cleanup are recorded in its result.json. No Boat was used. This establishes visible animation for this sample; pixel parity, complete scene correctness, performance and the historical 27-app corpus goal remain unqualified. Earlier stationary-frame findings below describe the previous build.

October 30: Twist motion boundary narrowed; no repair yet

Two original WebGL diagnostic runs, 20261030T2305Z-dx-twist-motion-probe and 20261030T2307Z-dx-twist-frame-boundary, used the same verified canonical module 7e8d3194… from source 8cfa97b0d. Runtime-only wrappers sampled 120 distinct submitted vertex hashes and 120 distinct full readback/DIB hashes. That establishes progressing geometry submission and readback, not visible animation: whole-frame hashes can change with the guest FPS overlay.

Four later samples excluded the overlay and measured the same scene-region hash, 3605449039, while draw count advanced from 14,092 to 14,125. Three closely spaced screenshots were personally inspected and show the same sphere orientation. This narrows the unresolved animation gap without identifying its cause or establishing scene correctness. No production source was changed. Each run retains its diagnostic logs and investigation.json; both browsers exited, all six ports closed, private profiles were removed, and no Boat was used.

October 30: paired original Twist renders; motion differs

Runs 20261030T2304Z-dx-twist-browser and 20261030T2304Z-dx-twist-software contain inspected original dx_twist initial/final frames. Both render a dark textured object with a yellow/black radiation symbol. Software shape and orientation change; WebGL frames appear stationary. WebGL animation is unverified, and the cause remains undiagnosed. This is a concrete follow-up gap rather than a scene-correctness pass.

Both loaded module 7e8d3194… from the passing canonical build of 8cfa97b0d, with unchanged source hashes recorded in the copied build manifest. Chrome 151/ANGLE SwiftShader observed guest threads enabled; WebGL used D3DIMGpu, while software selected DX2–7 WAT without it. Each run verified 39 input hashes and 38 served SDK responses. Its investigation.json references captures, route/build evidence and cleanup. Both own browsers exited, six ports closed and private profiles were removed; no Boat was used. FPS overlays are not benchmarks. Pixel parity, complete scene correctness and the historical 27-app goal remain unqualified.

October 30: Tunnel renders from a verified canonical build

Runs 20261030T2317Z-dx-tunnel-browser and 20261030T2317Z-dx-tunnel-software contain inspected initial/final captures of original dx_tunnel: gray textured tunnel walls with yellow/black edges. Software frames show changed orientation; WebGL animation is not established. Chrome 151/ANGLE SwiftShader used D3DIMGpu on WebGL and the DX2–7 WAT software rasterizer without D3DIMGpu on software. Guest threads were enabled.

The locked canonical build passed on source 8cfa97b0d; all 1,148 tracked build-source file hashes stayed unchanged across compilation, with no tracked source patch. It produced module 7e8d3194…, matching each loaded module. This newly verified build reproduces the earlier retained module hash; it does not retroactively establish the earlier runs' complete host/source identity. Each run verified 39 inputs and 38 served SDK responses. Build logs, source manifest, captures, route observations and cleanup are in its investigation.json. Both browsers exited, six ports closed, and private profiles were removed. No Boat was used. The guest FPS overlay is not a benchmark; these captures do not establish pixel parity, full scene correctness, or completion of the 27-app corpus.

October 30: paired original Flip3DTL browser animation

Runs 20261030T2311Z-dx-flip3dtl-browser and 20261030T2311Z-dx-flip3dtl-software each contain inspected initial/final captures of original dx_flip3dtl: a Windows 95 textured cube changes orientation on both routes. The window identifies itself as Flip3D DISPLAY HAL. The guest FPS overlay is not an independently measured benchmark.

Chrome 151/ANGLE SwiftShader used WebGL with D3DIMGpu in the first run; the second selected the DX2–7 WAT software rasterizer and had no D3DIMGpu. Guest threads were observed enabled in both. Both loaded retained module 7e8d3194…, compiled source revision unknown, and verified all 39 input hashes and 38 served SDK response hashes. Each run's investigation.json records evidence and cleanup: browser gone, three ports closed, private profile removed, no Boat. These observations establish animated textured output on both routes, not pixel parity, full scene correctness, performance, or completion of the historical 27-app corpus.

October 30: Viewer original-mesh startup failure

20261030T2308Z-dx-viewer-browser/investigation.json records a fresh original viewer.exe browser attempt with WebGL selected. The normal registry replaces camera.x, mslogo.x, and sphere2.x with tetra/cube test fixtures. For this run only, the launch manifest instead used the original SDK files from dx_globe, retaining the original Viewer executable. No fixture mesh was fetched; 39 input hashes and 38 served SDK responses passed verification.

The inspected startup dialog says Failed to load camera.x — D3DRMERR_NOTFOUND. A trusted click on OK left no Viewer windows or running app. This does not establish a rendered original scene; prior fixture-based Open/menu checks do not close this gap. Original camera.x uses binary xof 0303bin ProgressiveMesh/InlineData representation; the failure's cause remains unknown. Chrome 151/ANGLE SwiftShader loaded retained module 7e8d3194…, compiled source revision unknown. The browser, three listening ports, and private profile were cleaned up; no Boat was used.

October 30: paired original Globe browser renders

Runs 20261030T2302Z-dx-globe-browser and 20261030T2305Z-dx-globe-software add inspected initial/final captures of original dx_globe. Both show the titled Direct3DRM sample with a dark shaded sphere and yellow/black textured band. Software captures show a change in orientation; WebGL animation is not established. Their different apparent sizes and orientations are not a pixel-parity comparison or correctness verdict.

Chrome 151 used ANGLE SwiftShader with guest threads observed enabled. The WebGL UI/console and D3DIMGpu identify that route; the software UI/console identify the DX2–7 WAT rasterizer, with no D3DIMGpu. The queried host backend property was unavailable, and OpenGL gl_sw_enabled() does not identify the D3DIM route. Both loaded retained module 7e8d3194… (full hash in each loaded-module.json), compiled source revision unknown. Each run verified 39 input hashes and 38 fetched SDK responses. Browser processes and servers are gone and temporary profiles removed. Evidence and limits are in each run's investigation.json; no FPS, full scene correctness, current-main build, or complete historical 27-app claim follows from these captures.

October 30 Boids browser animation

The original dx_boids sample rendered moving colored flock triangles and cyan boundaries in two inspected browser captures. Run scratch/runs/20261030T2255Z-dx-boids-browser/ pins the loaded module 7e8d3194fd189a8e0089b41ee5972b08b67fd5650d6dfbe2a2fbb9d3c323a327 (compiled source revision unknown), 39 inputs, served asset hashes, WebGL host selection and a D3DIMGpu instance. Chrome used ANGLE SwiftShader and guest threads. Readback-shadow performance warnings remain; this is animated render evidence, not complete correctness, pixel parity, FPS, or recovery of the original 27-member list. Browser and servers were closed.

The paired 20261030T2258Z-dx-boids-software/ run now adds two inspected software captures with moving colored triangles and cyan boundaries. It uses the identical retained module and SDK inputs; all 39 input and 38 served SDK response hashes pass. Host backend software, no D3DIMGpu, and the UI's DX2–7 WAT rasterizer message identify the route. Both routes produce animated output; different animation times preclude pixel parity claims. No FPS or complete-scene correctness measurement. Its browser, servers and profile were cleaned up.

October 30 Pawn browser move

Run 20261030T2250Z-pawn-browser-move adds inspected original Pawn browser gameplay: native e2–e4 and engine e7–e5 reply in the composited window, WebGL selected in Chrome/ANGLE SwiftShader, child GPU frame sequence 2→25. Six original inputs and fetched Pawn payloads were hash-checked; loaded module 7e8d3194…, compiled source unknown. Threads were actually enabled. Exact backend implementation/export state was not captured; no hardware-GPU or complete chess-game claim. This does not establish all 27 historical members.

October 30 Tile World census lead replay

Run 20261030T2242Z-tworld-input replayed the registered original Tile World on SDL's GDI route. Root inspected Level 1 before/after ordinary keyboard movement: remaining chips 6 → 3 and time 200 → 138. Exit 0 after 650 100,000-block batches. Exact module 7e8d3194 and original input hashes are retained; compiled source revision is unknown. This validates bounded gameplay for an explicitly named historical census row, while providing no GL/D3D renderer qualification. Full-level completion and browser replay remain open.

October 30 Pawn move and extended Jazz scene

Run 20261030T2238Z-pawn-software-move exercised the original Pawn executable on software D3D9: e2–e4 input followed by the engine's e7–e5 reply. Root inspected the composited desktop, including both moved pawns and yellow engine move outlines. Exit 0, 20,000 batches; all six original input files hashed. Module 7e8d3194 is pinned; its compiled source revision is not independently established. This adds bounded CLI software gameplay, not browser parity.

The later 20261030T2236Z-jazz-scene run resolves the sparse Jazz observation below: after 180,000 batches the inspected primary surface shows five large lit musical notes and many smaller notes radiating against black. Exit 0; seven original inputs verified, retained module 1dbddeaf / source d4fb06b46. Neither observation reconstructs the missing historical 27-ID membership list.

October 30 four further original screensaver observations

Run 20261030T2226Z-d3drm-four-scenes retains five bounded software D3DIM executions of Architecture, Jazz, OrganicArt and RockRoll. All exited 0; 117 original input identities were hash-checked. Root inspected Architecture's textured marble interior, RockRoll's intermediate guitars/picks scene, and OrganicArt's extended-run green/red looping forms against orange. The initial OrganicArt run ended with an offscreen texture selected; the extended run's final capture selected the primary surface. Jazz showed small white fragments against black, so its full scene remains unqualified. Loading frames and offscreen textures are not evidence of a failed scene renderer. The later extended Jazz observation above supersedes this run's sparse-scene limitation.

The exact retained module/source is 1dbddeaf / d4fb06b46, as below, preceding the later GLTron repairs. Commands, durations, surface diagnostics, captures, input hashes and inspection notes are in the run. These are bounded scene observations, not visual correctness, browser parity, current-main validation, performance measurements or completion of the historical 27-app goal.

October 30 Falling Leaves and SciFi software observations

Run 20261030T2222Z-d3drm-scenes retains original scr_fallingl and scr_scifi software D3DIM executions, each reaching 120,000 batches and exiting 0. Root inspected the final images: gold/green leaf forms over a leaf backdrop, and flying gold forms above terrain under a red sky. Commands, logs and asset hashes are retained. These scene observations use the same retained 1dbddeaf module/source d4fb06b46 as Geometry below, before the subsequent GLTron repair. They establish rendered scenes in bounded runs, not visual correctness, browser parity, current-main validation, FPS or completion of all 27 apps.

October 30 Geometry software scene observation

Run 20261030T2217Z-geometry-render retains two original scr_geometry /s CLI executions on software D3DIM with the real guest d3drm.dll, one per cooperative/threaded scheduler. Both exited 0 after 20 seconds without a reported guest crash. Root inspected the cooperative intermediate/final and threaded final images: gold 3D geometry renders against a blue background. Logs, images and five original asset hashes are retained in the run directory.

The tested module is 1dbddeafbeff685596095d18a4fd9240820b00b0fa293d21ac03393f9f652fb1, retained from the original GLTron run on source d4fb06b46; this does not validate subsequent main edits. Browser WebGL, backend parity, FPS and other scenes remain unverified. Loader warnings for absent shell32/ole32 DLL files remain in the logs. This observation advances one known member's software evidence; it does not resolve historical membership or qualify all 27 apps.

October 20 historical membership and receipt matrix

This is the evidence-table deliverable for GLD3D-CORPUS-27-20261006, on documentation base 4d57bb66dae7da0a488b9546ccb0b4f4f7ba2434. It extends this canonical document; the dated broader tables below remain historical records. The original 27-ID list is not recovered, and the goal is not complete. No runtime, census execution, asset download or new build was performed.

Membership provenance and its limits

The retained memory /home/user/.claude/projects/-home-user-wine-assembly/memory/project_gl_d3d_corpus_goal.md (modified September 26, originating session bd610256-7b4f-4ed1-afa0-9c3cf3412fbb) records the September 22 count: 27 apps, with overlapping family totals GL7 / D3DRM10 / D3DIM9 / D3D8 7 / D3D9 2. Original census commit 2d722ab3766c288d13796ed9dad3f1ceeca8f1eb independently records 27 among 225 registry entries. Neither contains the complete output. The memory's SHA-256 is edf52a63ba936a24d25b1cd3e21cd4a939e4fb6c92d45fa7a66d93e65fc4d8e4. The original registry blob is 65f56177c3bbafe8f4ce0818e5369e44f235777c; the original census-tool blob is 45425e4ab7eed19a9aef67bc76365f4baa0885b5. These pin the recovery inputs that survive, not the missing binary inventory.

Membership labels in the table below are deliberately distinct:

Thus nine IDs are explicitly tied to the historical count, with 18 membership slots unresolved. E/M rows below are useful evidence mappings, not assignments to those slots. There are 30 distinct IDs below because the evolving memory names more than the original denominator; this is not a reconstructed 30-app or 27-app manifest. generally, mentioned in the memory only to exclude its DirectDraw route, is not added.

The bounded local-history search checked 145 retained project JSONL/gzip files for the originating UUID and the original 27/87-of-225 census text, then specifically checked the four retained September histories for census references. It found memory copies/references, not the original stdout. The originating session file is absent from that project directory. The exact search summaries, memory hash and path checks are retained under scratch/runs/20261020-gld3d-evidence-matrix/. This is a local search result, not a claim that the original session no longer exists elsewhere.

Missing provenance is specifically the September 22 census's enumerated stdout, or that originating session's tool result. A reconstruction instead needs the original registry/tool revision and the contemporaneous executable/DLL inventory with identities and path resolution, including which inputs were missing. Current gfx-app-census.js also scans mounted companion PEs and IID bytes; the original tool did not cover that same set. gl-name-census.js names static reach, not calls, and API-loader/driver binaries can name all GL entry points. Neither tool can turn current assets, API conformance tests or family counts into historical membership/gameplay.

Reading the receipt columns

H means a retained committed report says the scene/input was reviewed; the original run bundle is absent locally in this audit, so there is no new image/backend revalidation. L means the stated local receipt was read; only Quake II's October 1 before/after PNGs were newly viewed here. U means the required evidence is unknown or not identified. A report's source is the tested source, not this document's base or a later fix commit.

All run IDs below resolve relative to scratch/runs/. Exact IDs are printed in full; where only a date or glob survives, that is an explicit missing-ID dependency. The audit's validation.json records existence and hashes of available references without changing evidence. No row is a current-main both-backend pass. Software means the emulator backend, not a game's internal software renderer. A Node --headless-gl result remains a Node bridge result, not proof of browser WebGL. FPS is unqualified unless a specific metric and limitation are stated. Of the 25 exact distinct run directories cited here, only the October 1 Quake II software directory is locally present; the other 24 are absent. All 11 linked reports exist in the pinned Git base. The shared checkout's Unreal-family and Black & White 2 notes are older than that base; their later findings in this matrix were read from the base's Git objects.

Receipt groups used in the table:

Per-app evidence and missing dependencies

App / membership Emulator software receipt and tested identity WebGL receipt and tested identity Screenshot / input / FPS status and precise remaining dependency
scr_architec / S H S6: 20261006T0250Z-scr_architec-w4-software, interior renders. H B6: 20261006T1840Z-gld3d-webgl, renders. Reviewed-render reports only; interactive gameplay N/A; FPS U. Recover both bundles and per-capture backend/module identities. Short white loading captures are not failures.
scr_fallingl / S H S6: 20261006T0250Z-scr_fallingl-w4-software, leaves/shadows. H B6: 20261006T1840Z-gld3d-webgl, leaves/shadows. Historical opaque-shadow defect; the October 10 lit-alpha correction in the screensaver note supersedes “black leaves are intended.” Its paired corrected render receipts are not here. Input N/A, FPS U.
scr_geometry / S H S6: 20261006T0250Z-scr_geometry-w4-software, renders. H B6: 20261006T1840Z-gld3d-webgl, renders. Recover scene captures/identities; historical distant-fill/wireframe issues cannot be judged from API tests. Input N/A, FPS U.
scr_jazz / S H P10: 20261010T1850Z-gld3d-current-main-probes, renders. Earlier software ID only 20261006T02*-scr_jazz-gld3d-sw. H B6: 20261006T1840Z-gld3d-webgl; later P10 is Node GL, not another browser pass. Recover exact early ID or P10 bundle and paired frames; input N/A, FPS U.
scr_oasaver / S H S6: 20261006T0250Z-scr_oasaver-w4-software, renders. H B6: 20261006T1840Z-gld3d-webgl, renders with stray-box caveat. Scene/timer changes require mid-scene series. 20261011T0640Z-scr-default-clock-w4, source dcbdf0521, explains tick dependence; absent here. Correctness still not revalidated; input N/A, FPS U.
scr_rockroll / S H S6: 20261006T0250Z-scr_rockroll-w4-software, renders. H B6: 20261006T1840Z-gld3d-webgl, renders. Recover paired scene/identity evidence; input N/A, FPS U. Exit capture can be a scene transition.
scr_scifi / S H S6: 20261006T0250Z-scr_scifi-w4-software, renders. H B6: 20261006T1840Z-gld3d-webgl, renders. Recover paired captures; tick dependence and later lit-alpha correction as above. Input N/A, FPS U; source fix alone is not a new render pass.
icewind_dale_demo / N U for a GL/D3D route. Later sweep reports DirectDraw as registered. U for GL/D3D; no backend pass inferred from static opengl32 text. Original membership explicit, actual 3D use unproved. Exact traced run ID/source missing; preserve launcher-noise classification rather than inventing a rendering failure.
tworld / N U for a GL/D3D route; later sweep classifies tile game as non-3D. U for GL/D3D. Same membership-versus-use distinction. Recover runtime trace/source if resolving API scope; no 3D gameplay/FPS acceptance.
quake2_demo / E H Q8: 20261008T000901Z-quake2-software-ordinary. Also L 20261001T230900Z-quake2-universal-software: browser shared-Worker software, source label d1ea48a9 plus documented snapshot corrections, module e3924487eea80ca213d1e4f7865b7823d1f41ea3d5fdb484a382ee2b69f912b8. H Q8: 20261007T213620Z-quake2-ordinary-traversal, browser WebGL ordinary traversal. October 1 result/benchmark and two PNGs locally inspected: textured room/HUD and changed viewpoint, endpoint gl/software; keyboard movement is the original report's claim. Q8's stronger ordinary forward/reverse/idle bundles absent. October 1 0.850 completed publications/s is one regressed sample, not general gameplay FPS or paired acceptance. Recover historical membership and Q8 bundles.
warcraft3_demo / E H W8: 20261008T0400Z-warcraft3-software-runtime. H W8: 20261008T0156Z-warcraft3-campaign-world-runtime. Reports review selection/opposing movement on both OpenGL arms; FPS/audio/combat U. Recover bundles and original membership. This says nothing about D3D8 gameplay; preserve laptop Warcraft III ownership/pause.
halflife_uplink / E H dated table: September 22 corridor/HUD on software GL; exact run ID/source U. H 20261006T2100Z-hl-uplink-lazy-mci, browser GL corridor/HUD after fix 5f4bac8c; exact tested source/module U. 4-gameplay-webgl-lazy.png absent with bundle; source fix is not exact run identity. Ordinary input/FPS not separately established by these summaries. Recover both identities and original membership.
simgolf_demo / M H dated table: software gameplay; exact October 6 ID/source U. H 20261006T0503Z-simgolf_demo-webgl-route-w5, browser GL course/build bar; tested source/module U. Historical screenshot review, not fresh controls/FPS qualification. Recover software ID, browser bundle/identities and original membership. Cursor edge-scroll can explain a black view.
ptct / M L 20261030T2351Z-ptct-paired-browser/software: changing geometric beams, WAT rasterizer enabled. L same run root: changing beams on owning GPU layer; same canonical d83e413f…, source de2e338c4. Two inspected frames per route; one executable input and 144 served hashes each verified, cleanup verified. Full intro correctness, parity, FPS and historical membership remain U.
mw3 / M L 20261110-mw3-proof-repair-weapon-gameplay-1030: original browser software D3D, module 79f0842e…, source 1c59017d5; reviewed movement, turn and visible laser beam/red indicator/raised heat. L 20261110-mw3-webgl-menu-consumption-1045: original owning Worker WebGL, module 79f0842e…, source 44d457af2; reviewed cockpit movement 449→441 m and turn to compass10. Visible firing consequence remains unverified on this current run. Movement, turn and firing behavior, b3b0be723: reviewed terrain/heading changes and range 449→441→428→410 m; normal GetDeviceData consumed balanced input. Tested preparation base 1c42d2302; compiled module provenance recorded separately. Named target selection, 53e1139dd: Firefly/Owens target panels; increasing enemy range does not prove approach. Both boxes normally archived. November 10 software evidence: normal owning GetDeviceData caller 0x4f76c8 consumed balanced Space DIK57 press/release sequences9/10; coordinator reviewed full recovered/case-software/weapon.png in run 20261110-mw3-proof-repair-weapon-gameplay-1030. November 10 WebGL1045 observed 48 paired owning GetDeviceState completions at caller 0x4fa56c, nine returning the held mouse button, followed by reviewed Pilot Entry and ordinary cockpit control. Coordinator reviewed full recovered/case-webgl/movement-settled.png and turn.png; run report. Space press/release returned together 757 ms after release, and both prearmed captures were late; the live guard stopped subsequent gameplay commands. Current WebGL Chrome/Node and manager/watchdog actual parent waits plus separate EOF retained; box independently stopped normally. The earlier software run retains its original local-manager wait gap. Software beam verified; impact, damage/kill, sustained playability, FPS, current paired-backend parity and historical membership U.
dx_boids / M L 20261030T2348Z-boids-scene-diagnosis/software: moving colored flock triangles and cyan boundaries, canonical d83e413f…. L same run, root frames: moving colored triangles/boundaries, same module/source de2e338c4. Two inspected frames per route; 38 input and 179 served hashes each verified; backend and cleanup verified. Full scene correctness, FPS, parity and membership U.
dx_tunnel / M L 20261107-graphics-worker-readiness/case-software-dx_tunnel: original Worker software, changing textured tunnel/radiation sign. L same root, case-webgl-dx_tunnel: changing textured tunnel, owning Worker WebGL2/SwiftShader, module e41511fe…; zero observed GL errors/fallbacks. Personally reviewed retained 15/45-second PNGs on both routes; runtime closure and cleanup pins JS snapshot 38b9f7d319 plus patches and inputs. Compiled source commit, same-frame parity, full scene correctness, measured FPS and historical membership U. Review hashes: 20261107-tunnel-twist-retained-review.
dx_twist / M L 20261110-twist-current-paired-scene-motion-0900/case-software-dx_twist: original changing textured body; actual guest GPU enable false, no owning render endpoint. L same root, case-webgl-dx_twist: changing body/visible faces; actual owning Worker WebGL2/SwiftShader, zero reported errors/fallbacks/software draws. All six full 15/30/45-second PNGs personally reviewed. Current paired scenes and recovery: locked build/source e2063597c7, live guest/render module 79f0842e…, served-source and stable-owner bindings. Paired software/source motion gap closed. Full scene correctness, measured FPS, reference/backend parity and historical membership U.
dx_flip3dtl / M H P10: 20261010T1850Z-gld3d-current-main-probes, textured cube. H 20261006T1935Z-gld3d-webgl-recheck, browser render; later P10 Node GL reports 33k draws/zero fallback. Recover paired images and browser source/module; sample interaction/FPS U. Original membership remains unproved.
dx_globe / M L 20261030T2330Z-dx-globe-software: original textured sphere visibly rotates; canonical module d83e413f…, source de2e338c4. L 20261030T2330Z-dx-globe-webgl: original textured sphere visibly rotates on owning D3DIMGpu. Also L 20261107-globe-original-band-oracle: natural original Worker WebGL draw/readback, module e41511fe…, source 6d465979a plus recorded patch. October paired frames establish motion. November original band oracle, afd8bde95, matches all12 selected pixels around six black/nonblack transitions exactly using independent scalar and60digit Decimal calculations; root reran Decimal check from raw inputs (20261107-globe-corpus-review). Bands at these samples are expected texture content. The remaining101 of5347 robust interior pixels differ by one channel value; full scene/seam correctness, same-frame backend parity, FPS and historical membership remain unqualified.
dx_viewer / M Historical fixture rendering only; original plain-Mesh assets unavailable. L 20261030T2308Z-dx-viewer-browser: rejects renamed PM input with D3DRMERR_NOTFOUND; no original scene. 20261030T2334Z-viewer-input-diagnosis proves camera/mslogo/sphere2 aliases equal PM files. Prior original-mesh provenance claim withdrawn. Restore actual plain-Mesh inputs before original-scene retest; fixture/menu checks remain narrower. FPS and membership U.
mcm / M H P10: 20261010T1850Z-gld3d-current-main-probes, routed Stunt Quarry race, 200 s. H P10 Node GL routed race, 225k draws/zero fallback reported; older September 20 browser report lacks exact ID here. Recover route/input/frame receipts; don't promote Node GL to browser. FPS and original membership U.
arcanum_demo / M H 20261008T1431Z-arcanum-gameplay-qualified, browser cooperative software D3DIM, crash-site westward movement. H same run, explicit browser WebGL, westward movement. Review pins module d8d4096f957ed51cecab589b5a1ec7bf402336f13959ce0710faf2f7d3c68960; full tested closure is missing provenance/pins.json (branch base 7e4ce05ce alone is insufficient). WebGL 434/20.127145 s = 21.562919 visible presentations/s, idle scene, not scanout. Software FPS rejected because observer missed OffscreenCanvas. Recover bundle/membership.
ut2003_demo / M Reviewed original ordinary-menu Software gameplay: 62ee90c86, run 20261031T1628Z-ut2003-native-menu-play; DM-Antalus PLAY/join, movement/look/fire, health100/ammo150→148, advancing scoreboard clock. Actual Software Device gpu=false and completed executor samples retained. Reviewed original ordinary-menu WebGL gameplay: 09cedce52, run 20261031T1653Z-ut2003-webgl-menu-execution; PLAY/fire-to-join, movement/look/fire, health100/ammo150→149, advancing scoreboard clock. Actual owning WebGL2 executor; ANGLE SwiftShader, no hardware-GPU claim. Both accepted runs pin module 9b4e39a2…, 881 original/295 source/7 compiler/107 actual served files; original default guestWorker and -d3d -window -nosound retained. Reviewed captures, balanced native input and durable cleanup recorded in Unreal notes. Earlier cube-LOD repair 5fa12b1be remains causal evidence; ordinary-menu/backend gaps are closed. Exact historical membership, hits/kills, FPS/audio, same-frame parity and match completion remain unqualified.
ut2003_demo_server / M H October 6 DM-Antalus sweep claim, but exact reviewed software result.json was already missing in the October 10 table. H B6: 20261006T1840Z-gld3d-webgl, browser DM-Antalus/HUD. Recover software ID/result/source and browser bundle. COMCTL seed applies to both registry entries, but the client failure above is not proof of a server run failure. Ordinary control/FPS/membership U.
ut2004_demo / M Reviewed October31 default guestWorker browser software DM-Rankin: native movement/look/fire, health100/ammo98, full clock/HUD and visible-coordinate menu clicks after fb698d57e. Run 20261031T0514Z-ut2004-origin, module 9b4e39a2, exact295 source+7 compiler inputs; focused/full build pass, old-source regression rejects. Earlier CLI 2342e9288 and browser 09eb0ecbe also reach gameplay. Reviewed original ordinary-menu default guestWorker WebGL2 DM-Rankin gameplay, run 20261031T0629Z-ut2004-webgl, documented 939ff7db0: PLAY/fire-to-join, native forward/look/fire/retreat, clock20:00→19:51, ammo100→98, health100→96. Owning endpoint1 completes68,545 backend draws with zero observed executor failures; ANGLE SwiftShader, no hardware-GPU claim. Both browser backend arms have reviewed gameplay. WebGL pins module 9b4e39a2…, matching295 source/7 compiler files,998 original inputs and107 served snapshots; exact launch snapshot bf9c495b6. Same WASM hash does not permit mixing historical host files. Captures/input and cleanup retained in Unreal notes. Enemy hits/kills, match completion, audio, FPS, same-frame parity and historical membership remain unqualified.
morrowind / M No reviewed software movement receipt identified; software fix unit tests do not qualify the world. H 20261005-morrowind-prison-movement, browser Worker/WebGL forward/reverse/idle; source 94d18605, module 2e2fd8d1ca62cd87f0bc09312bc4836108df610d00cb7771bb5ae22755eceedc. Metadata review reports four reviewed gameplay images; bundle absent. Preserve no-heavy-Morrowind restriction. Recover evidence/membership; no software gameplay/FPS claim.
black_white_2_demo / M Reviewed retained October23 CLI D3D9 software tutorial: 20261023T1030Z-black_white_2_demo-claudea7d5-goldscroll, source aa09d05dcd6acf80d8f506c1bef1a00a92869119, recorded module 43375d0fca46b3e9684c18d6ee76cffe743ed55eee03a0188fca171174ae2dc8. Gold Scroll objective completes after the sixth click, followed by scripted field/storehouse camera pan; 10000 s reported zero failed commands. No reviewed WebGL gameplay receipt identified in this reconciliation. November7 retained review 20261107-bw2-retained-software-review personally inspects f0/f1/g1/g5 and hashes all listed artifacts/screenshots; source-head, module-hash record and successful build log agree with result. Executing module bytes were not recovered here. B&W2 notes retain consumer/input traces. This is historical CLI software evidence, without free play, isolated drag attribution, browser/current-main proof or paired-backend parity. Historical membership remains unknown; preserve pauses.
winamp / M No CLI software MilkDrop visualization receipt. H browser 20261007-winamp-milkdrop-complete-closure: source 0cbc1c6ffd, module 7c6864f8e224a0d42c6743b3c088ce79349101182295b79b1e4fa8a4c3911cc0; ordinary Start ends in illegal-operation (2), not a backend rendering pass. Review reports 11 images and complete-loader rerun. Capture actual exception/API; do not repeat obsolete “needs Winamp 5” inference. Visualization/FPS unqualified; recover bundle/membership.
pawn / M Recorded October 30 software D3D9 e2–e4 drag and engine e7–e5 reply, run 20261030T2238Z-pawn-software-move, documented in 4c2fdf3df. Composited desktop was inspected; exit 0 after 20,000 batches and six original inputs hashed. Recorded October 30 ordinary browser WebGL-selected drag/reply, run 20261030T2250Z-pawn-browser-move, documented in 26465f29b; Chrome/ANGLE SwiftShader, child frame sequence 2→25, threads enabled. Pawn notes and the October 30 sections above supersede the earlier board-only observations. Both pin module 7e8d3194…; compiled source revision remains unknown, and browser backend implementation/export state was not captured. November 6 reconciliation found both original run directories absent from scratch/runs/, so captures and identity receipts cannot currently be independently rechecked. Recover bundles and historical membership; no fresh execution, hardware-GPU, full-game or FPS claim.

What this delivers, and what remains unqualified

The nine explicitly linked members now have individual software/WebGL evidence rows, and all 21 additional named historical leads have separate provenance labels and dependencies. This closes the documentation omission only to the proved subset, not the historical denominator or runtime goal. Local historical imports named history-scr_* remain unreviewed/unknown-source evidence and cannot replace the missing dated run bundles.

The strongest remaining dependencies are (1) original enumeration provenance, (2) recovery of the dated run bundles and source/backend/PNG/input receipts, and (3) actual paired gameplay or rendering correctness where only menus, API tests, Node bridge routes or defective scenes were reported. Missing local bundles are not evidence of deletion everywhere and do not invalidate the committed historical reviews. Conversely, a merged caps/mip/Surface/COM fix cannot turn those reviews into current-main render qualification.

Existing broader-app work, including Darkstone, Hype and Pirates, remains in the dated sections of this canonical table; none is silently assigned a missing September membership slot. Preserve laptop Warcraft III/Heroes II ownership, all game/heavy-run pauses, CMR3's automated-review hold, and other denied-action restrictions. This document authorizes no rerun, replacement game, Claude activity, deployment or main integration.

October 17 retained Darkstone movement review

Later October 15 captures supersede the camera-only result below. Both software and WebGL show ordinary ground-click movement with a following camera. The software run exited successfully; the WebGL run retained two movements but ended at its input-queue deadline (exit 1).

At a closely matching gate/tree/feet pose, WebGL has a substantially larger black foreground triangle; software still has a smaller wedge. Neither is established as correct. This is movement evidence with an unresolved rendering defect, not full backend qualification. Audio, combat, FPS and touch remain unmeasured. No new runtime was performed on October 17.

See the reviewed comparison for exact source/module identities, PNG hashes, draw-completion counters and cleanup receipts. Runs: 20261015-darkstone-software-movement and 20261015-darkstone-locomotion. Identical-input foreground draw comparison remains the next diagnostic; the old D3D9 capture format alone cannot preserve legacy D3DIM live resources and framebuffer/depth history.

October 15 paired Darkstone result (17:40 UTC)

Fresh source fadaff08a, module 9b57583c…, reached Town through ordinary character creation on both software and WebGL. Root reviewed before/idle/held ArrowRight screenshots in both arms: camera control is verified. Main renderer endpoint 1 reported advancing completed native software draws in the software arm and advancing WebGL submissions in the WebGL arm. Reported WebGL errors and all six fallback sites were zero; declinedClears=1 remains a caveat. Both arms exited normally and independent postflight verified idle/preserved inputs. Full archive SHA-256 is 25252b50372c11589f30439d0a13ab39b09576b5844c1855721c26a429c2aa8f. See the paired receipt. This supersedes backend-unknown for this new run only. Locomotion, combat, audio, FPS, child renderer ownership and original-27 membership remain open.

October 15 current integration status (14:38 UTC)

This section supersedes the earlier same-day startup and unmerged-palette statuses below; those sections remain dated evidence.

The earlier Invisible War transport review below is historical: the corrected chain was subsequently merged as 3fdb29373, an ancestor of current main, after all eight combined native arms completed with their expected results. The exact negative control failed with RECEIVER_RETIREMENT_DID_NOT_SETTLE and the restored receiver suite passed. See the combined native result and merge receipt. Original Invisible War gameplay, pointer marshalling and both renderer routes remain unproved; native conformance does not establish those results.

Before that integration, callback transport remained on task branches. Selective port 8793e9f47 and endpoint checkpoint 17e713cd2 preserve main's existing lock, generation pins and INITIALIZING/LIVE/FAILED lifecycle; the older e2d10ebb0 branch must not replace those semantics. Exact candidate canonical module a948a826… passed eight native loader/transport suites. Combined local/Worker and Worker/Worker native endpoint races, nested initialization, publication acknowledgements and lookup retirement passed at 4050bacb3.

Independent review then found scheduler teardown could discard a retiring Worker and reuse its TID/stack before termination. Fix b5b69a714 passed actual Worker ABI checks and canonical native validation dca0e6201: lost retirement replies and generation-pin cleanup failures retain ownership; successful termination releases the real native stack exactly once. The old code fails the intended ownership assertion; restored code passes. Root verified archive SHA-256 22140338c794795536c124019079412943316e4a879b24014b4d51ea763f04fe under scratch/runs/20261015-iw-retirement-native/ and its logs/postflight.

The earlier review blocked integration because PID-wide broker admission deadlocked independent send chains and blocked a second receiver thread behind the first. Root reproduced both failures with real Node Workers, without guest WAT, in scratch/runs/20261015-iw-broker-root-review/. Worker /root/iw_independent_send_fix implemented per-thread, exact-waiting-frame admission and regressions before the combined validation linked above. Do not reopen this historical merge blocker as though that correction were absent.

The original 27-app membership evidence and per-game software/WebGL gameplay coverage remain incomplete. None of these focused tests closes that goal.

October 15 source-only next-validation reconciliation

Against 73f27af1b, Age of Wonders II's viewport repairs 9a438f1b9 / 201de22d9 and Start-click repair b6e18c196 are ancestors. Its dated row's terrain/UI defect is superseded by those merged fixes and the October 10 re-note. This audit does not independently re-review the absent October 10 capture bundle or certify a backend pair. Darkstone's complete installed manifest is locally recoverable, but its original installer and retained gameplay bundle are absent on this host. Neither absence invalidates the historical observations below. The two exact next validation routes, fixture hashes and missing evidence are in the bounded handoff.

October 15 reviewed follow-up (09:36 UTC)

The dated table below preserves earlier results. Signed SHORT2/SHORT4 GPU inputs and software declaration-component defaults are now on main ec4844f16: canonical build, seven CPU suites and actual WebGL1/2 pixel comparisons passed. Coordinator independently checked twelve GPU artifacts and 266 tested-source hashes. Original .kkrieger run 20261015-kkrieger-short-original eliminates the earlier declaration errors but still reaches EIP zero near 310 seconds, with 2 GiB memory and no OOM. Its 42 artifacts were independently hash-checked and final screenshot viewed. No gameplay is qualified. The next lead is a missing Buffer8 interface; the game's actual QueryInterface HRESULT remains uncaptured. See docs/re-notes/kkrieger-beta.md for the bounded diagnosis and uncertainty.

The palette candidate remains unmerged. Explicit state144 TRUE zero-RGBA handling passed canonical build and CPU checks; new Worker coverage passed both synchronization modes on d582f0da5, including native-index duplicate RGB, held state snapshots, cache transitions and depth. Worker observations read an XRGB DIB whose unused byte is normalized to 255, not framebuffer alpha. Source-alpha behavior is checked through blend/test/depth effects; direct GPU tests separately read framebuffer RGBA. Browser software-path fixtures exposed incorrect endpoint routing and an assertion on XRGB's unused byte; their failures are retained, corrected tests are being rerun, and no full browser acceptance is claimed yet. A cache-identity mutation must still fail the strengthened oracle. Default/FALSE filtered-key semantics, alpha-palette/key precedence and legacy stencil remain separate open work.

Original 27-app membership evidence

The original census commit is 2d722ab3766c288d13796ed9dad3f1ceeca8f1eb (22 September). Its message records 27 of 225 apps and seven D3DRM screensavers, but the commit contains only the census tool and a CLAUDE.md edit, not the complete per-app output. The tool depends on the executable/DLL files present when run. A current-asset rerun therefore cannot authenticate the historical membership. The later expanded 69-app census is a different set, not a recovered original list.

The original stdout or session transcript associated with memory session bd610256-7b4f-4ed1-afa0-9c3cf3412fbb is still missing from the inspected local archives. The 57 references in scratch/recovery-gl27-20261014/REPORT.txt refer to later dated run IDs, not 57 original September runs. Recover the original output, or explicitly label a historically matched reconstruction; do not infer a completed 27-app denominator from the current status table.

October 14–15 targeted follow-up

Authorized Boat exec access now supports separate-box builds and browser tests. The older blanket remote-access blocker is resolved; missing historical runs and incomplete per-app software/WebGL qualification remain separate gaps.

Scope Verified change and evidence Remaining work
.kkrieger D3D9 startup c79b943bb implements signed Q8W8V8U8 storage/upload/sampling; full build and resource, software, WebGL1/2 and pipeline checks pass. 2e1d68489 fixes recycled sparse heap clearing that overwrote live mesh edges; baseline regression fails and fixed regression passes. Original demo advances beyond the persistent mesh loop. Runs 20261014-kkrieger-q8w8v8u8, 20261014-kkrieger-mesh-writer, 20261014-kkrieger-heap-fix. Both fixes are on main. Census 20261014-kkrieger-memory-census proves backing shortage: 2,162,688 requested versus438,272 free; tables have capacity. Report on main 6cb760da2. Registered2GiB route merged9e8afa2e5 uses174,718,976 extension bytes with OOM0 through300s. Follow-up 20261015-kkrieger-registered-600s then exits around306s after243 generic declaration errors; correlation is not proven causality. Diagnostics/exit guard mergedcb0bf751a; exact original declaration capture is next. Retained heap space is not classified as a leak. No gameplay or both-backend acceptance.
Serious Sam cooperative startup 43f04c7c4 prevents the reproduced low-EIP crash by making synchronous-fault file consumers resident on the cooperative backend. Four browser arms distinguish original failure, eager control, fixed policy and Worker behavior; runs 20261014-sam-browser-*. Fix is on main. Karnak gameplay/profile 20261014-sam-low-fps measures3.18–4.05 actual GLfps on forced SwiftShader; pending GPU completion dominates this box. Report/probe 9b53a3795 on main. Logging-off control and Safari hardware cause remain unmeasured. HUD double-count fix 471dee7c2 passes focused mixed GL/GDI regression, not a performance improvement.
Legacy indexed palette candidates Preserved c2192e014 passes a separate-box full build and four real-VM CPU/page-watch/resource/software tests. Evidence 20261014-palette-remote-validation. Still unmerged. Source reconciliation f8e4676ce preserves candidate and main format63 changes. Exact reconciled candidate164dc33ad passed canonical build/four CPU suites; diagnostic-only fixture correction2b28b3e27 passed17 actual framebuffer checks each on WebGL1/2, zero fallback. Root reviewed original runtime hashes and pixel results in 20261015-palette-validation; forced SwiftShader, reported Intel string is spoofed. Worker reports clean explicit-sync lifetime run125426 with7pixel observations/3blocked-before-drain barriers, pending root receipt review.13of14 follow-up configs pass; Boids route test assumes software default despite current WebGL default and needs explicit route selection. Default-lazy Worker variant, stencil rejection, baseline GPU failure and filtered-key/alpha-key contracts remain gaps.

These are scoped results, not a new full-corpus sweep. Preserve the historical rows below as dated evidence; no original 27-app completion is claimed.

The user's goal: every OpenGL and Direct3D app in the registry works on both the software and the WebGL backend. This is the one shared status table for it (task GLD3D-CORPUS-27-20261006). Codex coordinates the remaining work. The Claude fleet is stopped. Codex coordinates current assignments on the board. D3DRM rows marked w4 refer to historical work by claude:65967384 (docs/re-notes/plus98-dx-screensavers.md), not a live worker.

Reconciled on 2026-10-07 against main 30f1e268. This preserves the dated measurements below; it is not a new full-corpus run. Tomb Raider III and Half-Life Uplink have later retained browser evidence that supersedes their earlier failures. Menu, loading, excluded and unmeasured rows still do not prove the user's full software-and-WebGL gameplay goal.

Selective October 9 reconciliation adds retained Morrowind movement and Pirates sailing evidence, current Age of Wonders II movement, and the Black & White 2 startup repair. This is not a new full-corpus or both-backend run. Missing Pirates logs are listed below; the log provenance gaps are not passes.

October 15 retained evidence update

Darkstone now has a fresh, recovered browser run: scratch/runs/20261015-darkstone-town-2, source controller c7e23d5f0, loaded module 04ede400070b3acbdcc2ad5f3c033bcf051e3d2da9a4d28a449324d23ab0ba17. Ordinary character creation reaches Town. The road and gate remain fixed during five idle seconds, then rotate under held ArrowRight. This proves camera control; it does not prove champion locomotion, combat, FPS or audio. The recovered archive SHA-256 is 60900452d07daac7bfa60ba9e08cf5217ef66bcb1e1bca93388c034477053540.

The software backend was requested, but the owning snapshot only exposed replay progress, which also includes non-draw commands. Backend qualification therefore remains unknown for this run. The October 10 paired-backend claim below remains a historical report; its scratch/runs/20261010T2116Z-darkstone bundle is absent on this host, so it has not been independently revalidated here. Do not replace that historical claim with a fresh both-backend pass.

Main 99932f66b adds endpoint-owned D3DIM completed-call counters, separately for requested software and GPU fallback, published on existing natural fences. Focused dispatcher/adapter tests cover endpoint isolation, flips and GL exclusion, and throwing native calls. These counters were not in the Town run and are not pixel-write counters. A fresh paired software/WebGL validation is queued; no original 27-app membership or full-corpus completion follows from this update.

October 19 reconciliation, without a new corpus sweep

Surface4/7 ABI support is on main 73433fabe. The combined RTC/Surface build passed both ABI candidates, eight DirectDraw controls in both dispatch modes, and RTC native Worker/host/RPC checks. This qualifies the tested API behavior, not original-game rendering or both-backend coverage. Legacy DirectDraw Surface4/7 cube-face storage remains unsupported: CreateSurface rejects extended DDSCAPS2/3/4 with E_NOTIMPL. This is distinct from the existing Direct3D8/9 cube-resource and shader paths; those already have implementations and must not be re-created as part of the legacy follow-up. Their presence does not establish original-game or both-backend qualification.

The original September 22 report names 27 measured apps without listing all IDs. The later table below has 26 rows expanding to 37 registry IDs; it must not be relabeled as the original 27-app manifest. Recover the contemporaneous census output to establish membership. The broader goal remains every registered GL/D3D app on both backends.

October 10 follow-up, without a new corpus sweep

The historical measurements below remain dated observations, not current-main passes. In particular, the newer fixes do not complete this corpus goal.

October 13 evidence reconciliation

This updates retained evidence against main fac1dc0ba; it is not a new corpus sweep.

App / scope Evidence now available Remaining gap
Diablo II GPU POINTLIST fix 782bc73d4 passed 40 pixel cases and ordinary browser town/control in paired 20261013T2055Z-diablo2-pointlist-control / -candidate, canonical Wasm ba6f25ea2d467b3fba00110d83575d48ac1aeb8794a474052190998067723259. Control 38,701 fallbacks/27,556 readbacks; candidate 0/1,001 with zero backend errors. Those optimization arms both use WebGL. Separate 20261013T2115Z-diablo2-software-town on2cf59bcff/ba6 now verifies ordinary town/ground movement with live software endpoint, no GPU adapter and command progress. Both backend routes therefore have reviewed browser gameplay evidence on recorded versions. Different RNG/time prevents a performance comparison; logical/display FPS, audio/device remain unqualified.
Neverwinter Nights (new-game extension) Real client attributes, stencil and color masks are merged. 20261013T2045Z-nwn-color-prelude on 99e780866 / canonical ba6f25ea… reaches Prelude world; ordinary ground clicks change position/camera and open Pavel conversation, then Continue advances dialogue. Root reviewed captures. Local-only corpus registration 46a456ab merged. Evidence is CLI software with synthetic batch clock, not browser/WebGL, FPS, audio, combat or physical-device qualification. Registered browser run 20261013T2100Z-nwn-registered-browser now also proves Prelude ground interaction and Pavel Continue; its exact GL backend remains unattested.
Arcanum Retained 20261008T1431Z-arcanum-gameplay-qualified explicitly pins both backends, ordinary movement and module identity. Software FPS remains unqualified; later missing-knob notes do not erase this run.
UT2004 Merged fc0d7bf10; 20261010T2240Z-ut2004-sw-menu-input-w5 retains software DM-Rankin and Up movement. Historical installer route, not a fresh registered-app or full backend-pair certification.
UT2003 20261010T2140Z-ut2003-server-sw-w4 and both named artifacts are present; TLS/OOM fixes are merged. Unattended listen server does not prove player movement; module hash absent.
Daikatana Retained software 20261006T1830Z-daikatana-demo-gameplay and explicit CLI WebGL 20261010T2140Z-daikatana-webgl-w4 record world, ordinary movement and weapon pickup. Fresh browser 20261031T0256Z-daikatana-webgl records the live GL endpoint's backend: webgl, inspected original world/HUD and native ArrowUp movement, with 458 unchanged input entries and 106 verified host snapshots. DLL repair remains merged. Historical module hashes are incomplete. Fresh browser module 8333a35b… has unknown compiled source identity; Chrome ANGLE SwiftShader is not hardware-GPU evidence. Combat, level completion and full renderer parity remain unqualified. Do not duplicate generic CLI WebGL qualification.
FreeSpace Historical flight predates removal of fabricated DirectX metadata. Genuine DX7a setup9ed9d4b54 now produces real files and registry but requires restart: four DLL replacements remain queued. Fresh worker01a193a3-efe3 implements/tests generic WININIT completion from retained private VFS before flight; heavy work waits current Nox release. No registry fabrication or current flight claim.

The falling-leaf repair, registered saver clock and headless resize fix are merged. Historical 14/14 crash revisits are retained launch evidence, not a fresh full-game certification. MilkDrop remains held after automatic review; Morrowind remains paused and Heroes II belongs to the laptop.

How the set was measured

node tools/gfx-app-census.js --family=gl,d3drm,d3dim,d3d8,d3d9 --list (fixed in 3a7b0ecb to read every PE an app mounts and the IDirect3D IID bytes) names 69 apps that can reach a 3D API. Static reach is not use, so every one was run: three CLI workers, software arm only, each app traced for the calls that create a 3D device (Direct3DCreate8/9, IDirect3D*::CreateDevice, a DirectDraw QueryInterface for IDirect3D*, wglCreateContext) and photographed. Every screenshot cited was looked at.

Apps that really use OpenGL or Direct3D

app API at runtime software WebGL (date = when verified) blocker / note
blood2_demo D3DIM (Device3) gameplay gameplay (2026-10-06, browser; in-level, HUD 100/50; scratch/runs/20261006T1840Z-gld3d-webgl) —
tomb_raider_2_demo D3DIM (Device2) gameplay gameplay (2026-10-06, browser; Venice alley) —
tomb_raider_3_demo D3DIM (Device2) gameplay gameplay (2026-10-06, later browser route; Jungle and ordinary movement, scratch/runs/20261006T2030Z-tr3-web-title-input) Earlier title result superseded by route timing correction 48ffb5e5; no input repair needed.
gta2_demo D3DIM (Device3) gameplay gameplay (2026-10-06, browser; city map, HUD) —
mw3 D3DIM (Device3) gameplay (cockpit) menu (2026-10-06, browser, no input; an Escape in the page's timing QUITS the demo, exit 3); gameplay (2026-09-20) route: Escape at batches 10-41 skips the Zipper intro
diablo2_demo D3DIM (Device3) menu (hero select) menu (2026-10-06, browser; Single Player / Exit) gameplay needs > 120 s
darkstone_demo D3DIM (Device2) menu in dated sweep menu (2026-10-06, browser) Later retained Town before/after images show champion, HUD and substantial camera rotation; root reviewed them on October 8. Original backend, module and command log were not recovered, so neither backend is newly qualified. See scratch/runs/20261007-darkstone-retained-gameplay-audit/result.json; do not count this as a never-played new title.
arcanum_demo D3DIM (D3D7) gameplay + player movement (2026-10-08); FPS unqualified gameplay + player movement (2026-10-08); idle crash-site 21.56 presentations/s current source/module/backend pinned; 434 visible submissions / 20.127 s, not scanout; retained earlier gameplay, not a new never-qualified game; handoff
dx_boids / dx_flip3dtl / dx_tunnel / dx_twist D3DIM renders renders, all four (2026-10-06, browser; scratch/runs/20261006T1935Z-gld3d-webgl-recheck) —
mcm D3DRM over Device2 gameplay (race) gameplay (2026-09-20); not re-run (long route) w4
dx_globe / dx_viewer D3DRM renders renders (2026-10-06, browser) globe texture seam (w4)
scr_architec, fallingl, geometry, jazz, oasaver, rockroll, scifi D3DRM renders (w4 rerun at --tick-ms-per-batch=2) 7 savers render (2026-10-06, browser; fallingl dark leaves and oasaver green field as on software) fallingl black leaves, oasaver stray box: w4
halflife_uplink OpenGL gameplay (corridor + HUD) gameplay (2026-10-06, lazy-file browser route after 5f4bac8c; scratch/runs/20261006T2100Z-hl-uplink-lazy-mci/4-gameplay-webgl-lazy.png) MCI lazy park now waits for IO; subclassed dialog buttons receive queued mouse input. Earlier black result superseded.
simgolf_demo OpenGL gameplay gameplay (2026-10-06, browser; course + build bar) —
quake2_demo OpenGL (ref_gl) gameplay + ordinary forward/reverse/idle traversal (2026-10-08, temporary browser, guest Worker/WAT software OpenGL, accepted reference 096889e1 / module fb1be916; reviewed run) gameplay + ordinary forward/reverse/idle traversal (2026-10-07, temporary browser, guest Worker/WebGL, module fb1be916; reviewed run) Original ref_gl Game → Easy; support-beam approach/reverse and stable idle landmarks reviewed on both backends. Software world visible by 49.617 s after selection, WebGL by 54 s. Software endpoint selection verified in host and owning render Worker, without GL WebGL fallback. Fresh software execution uses the unchanged accepted reference host/module closure, not a new current-main build. FPS/audio/network unqualified; software identity and cleanup, WebGL identity and cleanup. Historical 2026-09-23 gameplay retained.
warcraft3_demo OpenGL Prologue gameplay, Thrall selection and two opposing ordinary move orders (2026-10-08, temporary browser; guest OpenGL software; reference f62ab3c9 / module 4dc5ac2c; scratch/runs/20261008T0400Z-warcraft3-software-runtime) Prologue gameplay, Thrall selection and two opposing ordinary move orders (2026-10-08, temporary browser; reference f62ab3c9 / module 4dc5ac2c; scratch/runs/20261008T0156Z-warcraft3-campaign-world-runtime) Coordinator reviewed WebGL displacement against fixed hut, stone circle and rocks; matching live host/owning WebGL and all675 contained hashes verified. Coordinator also reviewed software selection and opposing displacement against the same landmarks, verified matching host/owning software endpoints and all666 contained hashes; integrated on main 0d9515322. No engine change. FPS/audio/combat/campaign completion and a current-main build remain unverified. Identity, route and cleanup. Software identity, route and cleanup. Historical SetPixelFormat repair 579ee802 retained.
ptct OpenGL renders, correctness unverified beams draw (2026-10-06, browser) 0.35 presents/s on software
ut2003_demo D3D8 menu menu (2026-10-06, browser, no route); gameplay (2026-09-25) each frame ~1 s on software
ut2003_demo_server D3D8 gameplay (listen server renders DM-Antalus) gameplay (2026-10-06, browser; DM-Antalus, HUD) —
ut2004_demo D3D8 splash at 120 s menu (2026-10-06, browser, no route); gameplay (2026-09-25) slow
alien_shooter D3D8 Mission 01 gameplay, movement and firing (2026-10-06; scratch/runs/20261006T0612Z-alien_shooter-w4-gameplay2) Mission 01 gameplay, ordinary movement and aiming (2026-10-08; scratch/runs/20261008T0114Z-alien-shooter-webgl, accepted reference f62ab3c9 / module 4dc5ac2c) Root reviewed player movement and opposing aim poses; live host and owning WebGL proved,109 artifact hashes verified. No engine patch. FPS/audio/combat completion/current-main build unqualified. Handoff.
crimsonland D3D8 Tutorial gameplay and movement (2026-10-06; scratch/runs/20261006T0500Z-crimsonland-w4-survival) Tutorial gameplay, movement and aiming (2026-10-08; owning WebGL, reference f62ab3c9 / module 4dc5ac2c; scratch/runs/20261008T0105Z-crimsonland-relative-input) Ordinary trusted relative mouse motion establishes real menu hover before clicking; arrow keys move player/terrain and relative input moves aim. Harness correction only. FPS/audio/combat/Tutorial completion and current-main build remain unqualified.
pawn D3D9 gameplay (board) board (2026-10-06, browser); gameplay (2026-09-23) —
pirates_2004 D3D9 menu after ba161dfb + 12408feb + d7f5a429 retained sailing captures (October 2; record reports Worker/WebGL, source commit unknown, six logs missing) Ship displacement/date progression visible; white terrain remains. Module hash recorded as 7c5f97f5...; source commit unknown. Run. No fresh backend qualification; recover provenance or revalidate. Local1371 manifest paths present (1,294,935,424bytes); lack of room for a duplicate local archive is not itself a remote-transfer blocker.
black_white_2_demo D3D9 island flyover (October9, normal intro/menu with Worker boot fix efcef0022 and live relative-input opt-in); player control pending not freshly measured Shared PE tail mapping fixes startup. Current flyover advances but has recurring rejected draws with declaration/FVF zero, despite zero main-Worker declaration-rejection counters; origin unproven. Stale right-edge pixels and garbled menu text remain. Live run; no ordinary gameplay/FPS claim. Historical diagnostic intro-skip gameplay remains separate.
morrowind D3D8 no reviewed software gameplay in this reconciliation prison-ship gameplay, forward/reverse movement (October5; Worker/WebGL) Reviewed run, source 94d18605, module 2e2fd8d1; four review-image hashes and retained module rechecked October9. Partial typing loss and inset viewport remain. Not current-main, full character creation, audio or FPS qualification.
age_of_wonders2_demo D3DIM gameplay and army movement (October9, ff0c1f07b) not measured by this run Reviewed run. FVF RESERVED1 stride repair prevents Miles callback overwrite; movement20→13 verified. Terrain/UI overlap subsequently repaired by merged 9a438f1b9 / 201de22d9; October 10 re-note reports clean panels on both renderers. Raw later bundle absent on this host; current backend/control pair, FPS/audio remain unqualified.
winamp D3D8 (MilkDrop) not run (CLI) ordinary playback/Start opens MilkDrop then illegal-operation error (2026-10-07, browser) original1.04e passes music gate but no visualization; fault/API cause unknown, native DLL probe refusals retained (run findings)

Glide is not in this goal's scope, but the sweep saw it: nfs3_glide_demo gameplay; diablo2_glide_demo title at 120 s (software Glide 1-2 batches/s and its frame fills only 512x384 of 640x480); hitman_glide_demo crashes on the CLI software arm (NULL object call in EngineData.dll 0x0ff6da1f during level load; its re-note has the browser arm getting further).

In the census, but not 3D at runtime

The IID evidence over-reports, exactly as the census warns: a DirectDraw game that links dxguid.lib carries every DirectX GUID. Confirmed 2D (or GDI) by trace: jazz2_demo, moorhuhn, moorhuhn_2, gallinelle (probe IDirect3D2/7, never create a device), pocket_tanks, heroes3_demo, captain_claw_demo, aoe1, aoe2, nfs3_demo (its own softtria.dll renderer), ut348_demo (SoftDrv), generally, generally_track_editor, baldurs_gate_chapters_1_2_demo, icewind_dale_demo, scummvm_fotaq, tworld, dungeons_of_dredmor(_release) (SDL; no GL calls), arena_gog/daggerfall_gog/ultima4_gog (DOSBox output=surface; opengl is a config option, not pursued), scr_win98, spider, and the four DirectAnimation theme savers (scr_corbis/fashion/horror/wotravel). aoe1 and pirates_2004 also match through a setup.exe/dxdiagn.dll that only lists DirectX files.

Open blockers in the GL/D3D set, ranked

  1. DirectX version detection -- both fixed. Crimsonland: 1fdd9a64, a versioned D3D8.DLL stub (DirectX 8.1). Pirates: the DxDiag query was only its error-message chooser; the real failure was MaxTextureBlendStages = 0 in our caps (ba161dfb), then a failed CoCreateInstance (Miles A3D) re-running its thunk (12408feb), then the CLI ignoring bigMemory (d7f5a429).

  2. Crimsonland Tutorial control qualified on both backends: retained October 6 software movement and October 8 ordinary WebGL movement/aiming are reviewed. WebGL required actual relative mouse input for Pointer Lock; absolute page coordinates had not established the menu target. This is a harness correction, not an engine patch. Survival, combat, FPS, audio and Tutorial completion remain unverified.

  3. Throughput, not correctness: ut2003/ut2004/arcanum reach gameplay only past the 120 s cap on the dated software sweep. Quake II's fresh ordinary software run on the accepted reference build reaches the world by 49.617 s after Easy, superseding its earlier cap result.

  4. Remaining WebGL coverage: the October 6 spot checks and later TR3/Uplink follow-ups are retained above. Pirates has retained sailing pictures with incomplete provenance; Morrowind has reviewed WebGL movement, and BW2 currently reaches its ordinary island flyover. Their software/WebGL pairs remain incomplete, and several other rows only prove menus or rendering. Do not reopen completed spot checks or mark the overall gameplay goal complete.

  5. D3DIM PBO warning: emulator lifecycle audit and standalone Chrome151 reproduction are complete. The standalone4x4 case reproduces the warning with correct pixels; the emulator trace has216 writes/215 reads and no overwrite violation (one capture-end pending read). Do not queue a speculative renderer repair or repeat the old causal trace. Browser-side warning remains; performance impact unmeasured. Receipt.

Outside the 3D set but found here: the DirectDrawFactory IID typo (fixed fa4be36a; the theme savers then run the existing DirectAnimation shim, frames unverified until their JPGs can be decoded: skia-canvas's native binary is missing on this box at the software sweep; the later browser photo-grid check is recorded above); aoe2's Unicode IDirectPlay4 QueryInterface (handed to the AoE lane on the board); PathAppendA for dungeons_of_dredmor (fixed by w5 5ca54afc).

D3DIM asynchronous PBO warning: source audit, 7 October 2026

No proven unsafe PBO reuse yet. Do not implement the old TODO's proposed ring or fence wait from the warning alone. No engine edits, tests, browser, build or performance measurements were performed. Drakan's acceptance fixture remains unchanged.

Exact evidence

source-receipt.json hashes the source copies and original MW3 console. The console contains 145 instances of the shadow-copy-discard warning. The relevant lib/d3dim-gpu.js at origin/main e710d101 is byte-identical to feature commit fcfa4989. The shared checkout is older: this audit deliberately uses Git object contents, not its working-tree file.

In that source, _flip lines 457–498 collects an existing t.inflight at line 468 before writing the same PBO. _completeInflight lines 501–520 binds t.pbo, calls getBufferSubData, unbinds and deletes the fence. The normal path therefore orders readback before reuse. No explicit clientWaitSync exists; synchronous collection can still block.

The dead-target branch at lines 362–365 discards an unread pending result, deletes its sync, destroys the device and removes the target. It does not itself reuse that PBO. Resizing at lines 207–212 fences before destroying/replacing the target. These branches need live identity evidence before attributing the warning to discarded reads. _completeInflight clears its JS pending record before GL collection; an exception would leave no retry record. A WebGL validation failure can also return normally without copying. Neither is established in the saved run.

What Chromium actually warns about

Primary source is pinned to Chromium revision d03948c49f64c93042f36929fc9a89d1e688c6e6, not claimed to match the installed browser revision:

Existing coverage and minimal next proof

test/test-d3dim-gpu-async-flip.js covers one queued flip, unrelated-range deferral, original-DIB collection, global collection, and synchronous fallback. It has no repeated-flip, discard/recreation or multiple-PBO identity case. Its mock getBufferSubData reads a variable last assigned by bufferData, rather than the currently bound buffer: extend that mock before trusting identity coverage.

First proposed source-only tests, once authorized: actual executor with a binding-aware GL mock; two differently colored consecutive frames; assert collect-old before write-new and correct old/new DIB bytes. Two targets/contexts must never collect each other's buffer. Dead target discards once and recreation gets a fresh resource identity; size growth collects before deletion. Negative control removing the collect-before-reuse call must fail on overwritten old-frame bytes, not a missing helper. These prove application ordering, not Chrome shadow behavior.

Then, only under a separate serialized browser grant: one short MW3 ordinary menu diagnostic, capped 10 seconds/256 detailed events with total counters and explicit dropped-event flag. Identify the actual executor and context, then assign stable WeakMap IDs to contexts, PBOs, syncs and targets. Wrap existing bindBuffer, bufferData, readPixels, fenceSync, getBufferSubData, deleteSync, deleteBuffer and executor target/collect lifecycle seams. Record args, bound pack-buffer identity, byte ranges, target/backing identity, entry/return/throw and warning timestamps. Forward exactly once with original receiver/arguments/results/errors. No extra readback, wait, flush, binding, getError or pixel mutation. Restore only own wrappers, report foreign replacement. Preserve observer overhead and unknown on cap/context mismatch.

Discriminating outcomes: a same-resource second write with no completed collection supports a missed/discarded-read lifecycle; a complete ordered collect between writes refutes that simple explanation and requires checking actual Chromium revision, validation and context before a change. A returned JS call alone cannot certify successful GL copying. Do not infer GPU corruption or a performance gain from warning counts. Any performance A/B belongs on separate boats with matched useful work.

Winamp fixture identity, October 7

Static PE resource inspection identifies test/binaries/winamp.exe as 2.9.1.0. The registry mounts plugins/candidates/vis_milk.dll (430,592 bytes), not the separate vis_milk2.dll (425,472 bytes, Winamp 5.6.6 strings). No Winamp 5 executable was found in the test/binaries filename inventory. The registered plugin itself mentions a feature requiring Winamp 2.90 or later; that does not prove its full host requirement. The older “needs Winamp 5” table entry is therefore an unverified prerequisite, not an established missing-file blocker. Next inspect the exact registered plugin initialization/version gate before acquiring a different host. No runtime result is implied. Exact hashes and resources: scratch/sweep-reconciliation-20261007/winamp-fixture-identity.json.

October 7 PBO executor regression follow-up

The binding-aware test/test-d3dim-gpu-async-flip.js now passes on unchanged executor SHA-256 e3370b31c0a63aff9420dc89f044f121c87bb2140a7683f4917f199a17680b8f. It proves same-PBO two-frame collection ordering and distinct DIB contents, correct currently bound buffer despite another allocation, reversed collection of independent contexts, rejection of a foreign-context PBO, and dead-target discard/removal without a subsequent flip reusing that target. The GPU and surface metadata are fixture objects; this is the actual JavaScript executor, not actual Chrome/GPU execution. Resize/recreation is not newly covered.

Under the explicit pure-JS lease, candidate PID 2407020 exited 0 and private in-memory --negative-control PID 2407027 exited 1 at the intended first-frame byte assertion: actual [17,17,17,17], expected [48,32,16,255]. The control removes only collection before reuse, so setup or missing exports cannot explain that expected failure. Both processes completed within 46.221 ms total at 08:58:53.624Z; this duration is a cleanup receipt, not a benchmark.

Receipt/logs: scratch/d3dim-pbo-audit-20261007/test-validation.json, candidate.log, negative-control.log. Test SHA-256: f9cfd55d29972759b7595aedea386bb4cb106d390d5fa4aeb05919b6a6dba13e. No engine change or claim that Chrome's warning is fixed. The bounded live context/PBO lifecycle observer above remains the next causal diagnostic.

Registered MilkDrop initialization gate, October 7

The exact registered vis_milk.dll export at 0x100299c0 returns its header directly. Its initializer at 0x10029a90 contains no blanket Winamp 5 rejection at the historical music check: it queries the host using SendMessageA(WM_USER, 0, 0), accepts a result at least 0x4000, and otherwise accepts SendMessageA(WM_USER, 0, 0x68) == 1. Only the latter failing path shows “This plugin can't run without music” and returns failure. This identifies an older-host playback condition, not proof of successful rendering. Raw exact-PE disassembly is retained at scratch/sweep-reconciliation-20261007/milkdrop-init-disassembly.txt.

Current test/test-winamp-visualizers.js documents and exercises MilkDrop header enumeration on the existing host; its historical remaining-gap comment names cross-thread IsPlaying delivery, not a Winamp 5 dependency. Main 2dab96f8 later changed Worker-to-main SendMessage delivery and is in the pinned 3b8189f5 runtime ancestry, but its WinBoard regression does not prove the MilkDrop route. Next run the actual registered plugin with ordinary playback/Start and observe the returned playback query and D3D device creation under each relevant thread mode. Keep software/WebGL rendering unqualified until that evidence exists. No host download or new runtime was performed for this source audit.

Retained software evidence reconciliation, 8 October 2026

Root reviewed Crimsonland s0.png and s-moved.png from the October 6 Tutorial run: terrain moves beneath the player and the HUD remains visible. Root reviewed Alien Shooter m17500.png and m-fire.png: the soldier moves from the road to the fence and the later image shows firing. The run records identify source commits 95e04dd8 and d699e4df, respectively; both omit WASM hashes. These are retained historical software results, not fresh tests of current main. Neither establishes WebGL gameplay, FPS or audio.

October 9 retained-evidence limits

Pirates sailing images were inspected again: the ship leaves Port Royale and the calendar advances from January11 to February3. The result record reports WebGL but the following referenced files are absent from scratch/runs/20261001-pirates-worker-sailing-after/:

The screenshots alone cannot restore the missing source/backend execution record. These are retained pictures, not newly verified current-main gameplay. Morrowind has stronger retained evidence: review.json records W/S1200ms and idle1500ms, four screenshot hashes match, and the retained wasm matches 2e2fd8d1ca62cd87f0bc09312bc4836108df610d00cb7771bb5ae22755eceedc. DiabloII lazy-default evidence ends at ActI loading; do not promote it to gameplay merely because its regression result says passed.

Current-main status, 10 October 2026 (GL-D3D-CORPUS-STATUS-20261010)

One row per app that draws through OpenGL or Direct3D at runtime, combining the best dated evidence above, the per-app re-notes and scratch/runs, and a fresh two-arm probe on current main. The sections above are unchanged.

Census. node tools/gfx-app-census.js --family=gl,d3drm,d3dim,d3d8,d3d9 --list at c50336dcc: 90 of 289 registry apps reach a 3D API statically (ddraw 82, d3drm 14, d3dim 68, d3d8 13, d3d9 4, gl 25). The 21 apps registered since the October 6 sweep are classified below. Every census app appears exactly once, either in the 3D table or in the not-3D list.

Fresh probes. Boat bx_u968n4h4 (4 vCPU, llvmpipe, Xorg :0, ops services off), main cac339517, wasm 6a798f18. Each run is a bounded test/run.js --app=ID --no-close --png; software arm is the default (D3DIM/D3DRM), --d3d9-renderer=software (D3D8/9) or --gl-renderer=software; GPU arm is --headless-gl (+ --d3dim-gpu for D3DIM/D3DRM). Runs, final frames, log tails and the probe script are in scratch/runs/20261010T1850Z-gld3d-current-main-probes/. A probe without an input route only proves boot-to-menu and a clean exit; gameplay columns cite the routed runs.

Legend: gameplay = ordinary control proven in a reviewed run; renders = the 3D scene draws (no gameplay concept); menu / loading = how far the best run got; P = reached in the 10-10 probe on current main.

Apps that draw through OpenGL or Direct3D

app API software WebGL / GPU first open blocker (row)
alien_shooter D3D8 gameplay 10-06 20261006T0612Z-alien_shooter-w4-gameplay2 gameplay 10-08 20261008T0114Z-alien-shooter-webgl —
crimsonland D3D8 gameplay 10-10 20261010T0515Z-crimsonland-cli-frames gameplay 10-08 20261008T0105Z-crimsonland-relative-input —
diehard_nakatomi_demo D3D8 gameplay 10-10 20261010T0230Z-diehard_nakatomi_demo-gameplay gameplay 10-10 20261010T0315Z-diehard_nakatomi_demo-web —
ut2003_demo_server D3D8 gameplay (DM-Antalus, sweep row) gameplay 10-06 20261006T1840Z-gld3d-webgl software row has no reviewed result.json
ut2003_demo D3D8 gameplay 10-10: DM-Antalus, joined, HUD 100/150, Up walks to the boulder, Left turns (20261010T2050Z-ut-family-both-arms) gameplay 10-10 (--headless-gl): DM-Antalus, Up walks into a weapon pickup, Left turns (same run) headless-GL frame drawn 4,24 px off the client origin (HEADLESS-GL-CLIENT-OFFSET-20261010)
ut2004_demo D3D8 menus 10-10 (Instant Action game-type list) but clicks land late and ~40 px high, so gameplay was not reached before the boat TTL; DM-Rankin spawn 09-29 (re-note) gameplay 10-10 (--headless-gl): DM-Rankin spawn with bots fighting, Up moves down the corridor, Left turns (same run) software menu input mapping (UT2004-SW-MENU-INPUT-20261010); GPU menu list boxes draw empty
morrowind D3D8 world renders, no reviewed software movement gameplay 10-05 20261005-morrowind-prison-movement software run needs a heavy route (user: no Morrowind runs)
winamp (MilkDrop) D3D8 not run MilkDrop opens then illegal-operation fault 10-07 MilkDrop fault (TODOS line "Diagnose original Winamp MilkDrop exception")
pawn D3D9 gameplay 09-23; board renders P gameplay 10-03 page; board renders P —
pirates_2004 D3D9 menu retained sailing pictures, provenance gap PIRATES-ROUTE-FOLLOWUP, PIRATES-TERRAIN-REVIEW
black_white_2_demo D3D9 island renders 3300 s, 0 failed commands 10-10 20261010T1638Z-bw2-raster-budget-boat-w6; control not shown not measured since 09-15 (wrong scene then) GAMEPLAY-black_white_2_demo
blood2_demo D3DIM Dev3 gameplay gameplay 10-06 —
tomb_raider_2_demo / tomb_raider_3_demo D3DIM Dev2 gameplay gameplay 10-06 (20261006T2030Z-tr3-web-title-input) —
gta2_demo D3DIM Dev3 gameplay gameplay 10-06 —
mw3 D3DIM Dev3 gameplay (cockpit) cockpit gameplay + forward movement 10-10, --headless-gl --d3dim-gpu, 49,818 GPU draws, 0 fallbacks (20261010T2110Z-mw3-webgl-cockpit-w5) —
diablo2_demo D3DIM Dev3 menu / Act I loading menu 10-06 gameplay needs a longer route
darkstone_demo D3DIM Dev2 Town gameplay 10-10, software D3DIM: camera turns under Right input (20261010T2116Z-darkstone, w5) Town gameplay 10-10, --headless-gl --d3dim-gpu: camera turns under Right input, 62,238 GPU draws, 0 fallbacks (same run) —
arcanum_demo D3D7 gameplay 10-10 20261010T0810-arcanum-control-frames gameplay 10-08 (backend knob not recorded) —
age_of_wonders2_demo D3DIM gameplay 10-09 20261009T0306Z-age-of-wonders2-gameplay gameplay + army movement 10-10, page WebGL on a GPU-less boat Chrome (20261010T1145-aow2-viewport-clip-fix, after 9a438f1b9) —
mcm D3DRM / Dev2 race P: Stunt Quarry, riding, 200 s routed run race P: same route, 225k GPU draws, 0 fallbacks — (first current-main WebGL race since 09-20)
dx_flip3dtl D3DIM renders P (textured cube) renders P (33k GPU draws, 0 fallbacks) —
dx_globe D3DRM renders P renders P (233k GPU draws) —
dx_viewer D3DRM renders P renders P —
dx_boids / dx_tunnel / dx_twist D3DIM renders 09-23 renders 10-06 20261006T1935Z-gld3d-webgl-recheck —
scr_jazz D3DRM renders P (first dated software run) renders P —
scr_architec, geometry, oasaver, rockroll, scifi D3DRM renders 10-06 …-w4-software renders 10-06 …-w4-webgl oasaver stray box (minor)
scr_fallingl D3DRM renders, leaves draw as black silhouettes same picture SCR-FALLINGL-BLACK-LEAVES-20261010
zuma_deluxe D3D7 (9 lit-off 2D quads) + DDraw gameplay 10-10 20261010T0710Z-zuma_deluxe-control-frames-w6; title P gameplay 10-04 20261004-zuma-adventure-gameplay; title P (9 GPU draws) —
avp_alien_demo / avp_marine_demo D3DIM (execute buffers) gameplay 10-06 (20261006T072800Z-avp-alien-demo-forward-walk, 20261006T073600Z-avp-marine-demo-gameplay); menu P gameplay 10-10, both demos in-game on WebGL (20261010T1930Z-avp_alien-webgl-w4, 20261010T1930Z-avp_marine-webgl-w4) —
carmageddon2_demo D3DIM Dev2 gameplay 10-06 20261006T075600Z-carmageddon2-demo-gameplay race with control 10-10, --headless-gl --d3dim-gpu, accelerating among opponents (20261010T2100Z-carma2-webgl-w4) —
carmageddon_tdr2000_demo D3D7 race 10-10: opponents racing, HUD (20261010T2035Z-tdr2000-loader-stall-w6). The earlier "stall" was a long CRT-heavy load: use --batch-size=500000 race with control 10-10, --headless-gl --d3dim-gpu, lap 1/4 with kills after Up (20261010T2100Z-tdr2000-webgl-w4) software steering not separately proven
cmr2_demo D3D7 gameplay 10-06 20261006T1650Z-cmr2_demo-gameplay-w6; menu P in-race with control 10-10, --headless-gl --d3dim-gpu, Sweden stage at 11 mph after Up (20261010T2100Z-cmr2-webgl-w4); the menu glyph blocks were fixed by 8918a13de —
colin_mcrae_rally_demo D3DIM gameplay 10-06 20261006T061500Z-colin-mcrae-rally-demo-gameplay; options P in-race 10-10 20261010T1930Z-colin-webgl-w4 —
drakan_demo D3D6 gameplay 10-06 20261006T1600Z-drakan_demo-gameplay; menu P in-level with control 10-10, --headless-gl --d3dim-gpu, on the dragon, turn and forward (20261010T2100Z-drakan-webgl-w4) —
quake2_demo OpenGL gameplay 10-08 20261008T000901Z-quake2-software-ordinary gameplay 10-07 20261007T213620Z-quake2-ordinary-traversal —
warcraft3_demo OpenGL gameplay 10-10 20261010T0330Z-warcraft3-demo-control-frames gameplay 10-08 20261008T0156Z-warcraft3-campaign-world-runtime D3D8 path (registry without -opengl) reaches only the menu
halflife_uplink OpenGL gameplay 09-22 gameplay 10-06 20261006T2100Z-hl-uplink-lazy-mci —
simgolf_demo OpenGL gameplay 10-06 gameplay 10-06 20261006T0503Z-simgolf_demo-webgl-route-w5 —
anachronox_demo OpenGL gameplay 10-06 20261006T1940Z-anachronox_demo-gameplay-w6 gameplay 10-06 20261006T2030Z-anachronox_demo-web-w6 —
daikatana_demo OpenGL gameplay 10-06 20261006T1830Z-daikatana-demo-gameplay gameplay 10-06 20261006T1910Z-daikatana-demo-web (renderer not recorded) —
descent3_demo OpenGL gameplay 10-06 20261006T1320Z-descent3_demo-gameplay-w6; pilot screen P in-flight gameplay 10-10: cockpit, yaw and laser fire, energy 100 -> 99, 664k GL draws (20261010T2030Z-descent3-webgl-flight-w4). It needed the lib/gpu-backend.js fix below; before it, the run crashed at the pilot screen —
baldurs_gate2_demo DirectDraw as registered: 0 GL calls at startup and menu on both arms, no 3D Acceleration key in baldur.ini (20261010T1930Z-bg2-renderer-w4) movement 10-09 (native CLI, no result.json) gameplay 10-10 20261010T0146Z-bg2-browser-lock-fix (DirectDraw) its OpenGL path (3D Acceleration=1) is unexercised
ptct OpenGL renders P (untextured beams) renders P (72 draws, 24 presents in 50 s) correctness unverified (0 texture uploads)
deus_ex_demo SoftDrv by default; D3DDrv / OpenGlDrv optional D3DDrv on software D3DIM 10-06; OpenGlDrv intro 10-06 D3DDrv menu + Training 10-06 20261006T1115Z-deusex-renderer-bench GAMEPLAY-deus-ex-demo

Census apps that do not draw through OpenGL or Direct3D

The October 6 list above still holds: jazz2_demo, moorhuhn, moorhuhn_2, gallinelle, pocket_tanks, heroes3_demo, captain_claw_demo, aoe1, aoe2, nfs3_demo, ut348_demo, generally, generally_track_editor, baldurs_gate_chapters_1_2_demo, icewind_dale_demo, scummvm_fotaq, tworld, dungeons_of_dredmor, dungeons_of_dredmor_release, arena_gog, daggerfall_gog, ultima4_gog, scr_win98, spider, scr_corbis, scr_fashion, scr_horror, scr_wotravel. Glide (out of scope): nfs3_glide_demo, diablo2_glide_demo, hitman_glide_demo.

New since that sweep, each with a reviewed gameplay run on its own renderer: asghan_demo, braveheart_demo, crusaders_mm_demo and populous_tb_demo draw with their own software renderer through DirectDraw (the registered exe; their D3D/Glide exes are not registered). croc2_demo never calls CreateDevice, even with a HAL seeded. disciples2_demo is DirectDraw only. driver_demo's config.exe offers only Glide.

Found and fixed on the way: desktop GLSL explicit LOD

--headless-gl crashed Descent 3 at its first GL_CLAMP draw: extension 'GL_EXT_shader_texture_lod' unsupported in fragment shader. lib/gl-compat.js's border-sampling program is ESSL. The Node bridge exposes desktop GLSL, where that extension is GL_ARB_shader_texture_lod and the function is texture2DLod. lib/glide-backend.js already made that rewrite for its own shader. portToDesktopGLSL in lib/gpu-backend.js now makes it for every shader. test/test-headless-gl.js compiles such a program. On the boat it fails with the exact Descent 3 error on unchanged main, and passes with the fix. Browsers are unaffected: their WebGL1 knows the EXT name.

test-headless-gl.js was already red on main before this change, for two reasons. Its run.js regex predated HEADLESS_GL || GLIDE_RENDERER === 'software' (now fixed here). Later, its resize assertion right edge stale: drawing buffer was not resized fails under llvmpipe both with and without this change: HEADLESS-GL-RESIZE-STALE-20261010. Fixed in 636e0c8b7: once a context has drawn, Mesa keeps its drawable at the old size until the next swap, so resizeContext now swaps once after resizing.

What is still open, in order

  1. Pairs still short of gameplay (GL-D3D-PAIRS-20261010, split on the board 2026-10-10 20:40Z). Closed on 10-10: avp_alien/marine, colin_mcrae and descent3 on WebGL (w4); aow2 on WebGL (w5); mcm on both arms; tdr2000 on software (w6); the cmr2 WebGL glyph blocks (8918a13de); driver_demo's Glide path on WebGL (--headless-gl, llvmpipe on a boat Xorg): the "Lose the tail!" chase with its HUD, GAME OVER at 01:04.48 without input as on software, and with ' + Z the car burns out, turns and reaches the garage exit past 01:04 (d10ba697, scratch/runs/20261010T2200Z-driver_demo-glide-webgl-d10ba697). Closed later on 10-10: mw3 on WebGL (w5); cmr2, tdr2000, drakan and carmageddon2 on WebGL (w4); ut2003 on both arms and ut2004 on WebGL (1863d2b5); darkstone_demo on both arms (w5).
    • Remaining after later evidence: the historical diablo2_demo pair and ut2003_demo_server software result remain to reconcile. Arcanum's explicit software/WebGL crash-site movement was reviewed in 20261008T1431Z-arcanum-gameplay-qualified; software scene-qualified FPS remains open, and the missing full provenance bundle noted above is not repaired by that gameplay acceptance. Daikatana's explicit CLI WebGL 20261010T2140Z-daikatana-webgl-w4 and browser 20261031T0256Z-daikatana-webgl supersede its missing-WebGL entry; the historical module/source identity and combat/parity limits above remain. These are retained observations, not fresh current-main qualification. The historical UT2004 software menu-input blocker is superseded by October31 fb698d57e and the reviewed software DM-Rankin browser gameplay evidence above. The separate WebGL native-menu/movement/look/fire result f81b9a35f, run 20261031T0629Z-ut2004-webgl, is also accepted above; FPS, audio, enemy hits/kills and match completion remain open on both arms.
    • Not split: bg2 is DirectDraw as registered.
  2. Heavy D3D8/D3D9 titles: Pirates (PIRATES-* rows), BW2 (GAMEPLAY-black_white_2_demo), Morrowind software (heavy route needs user sign-off).
  3. Remaining visual validation: Falling Leaves' opaque-shadow defect was corrected by the October 10 material-alpha repair documented in the screensaver notes. The October 30 software scene observation is retained above. The November 3 paired software/browser WebGL evidence validates original alpha127 and source-alpha blending on both routes (a4f232a32; run20261103-falling-leaves-paired-shadow-alpha). Native reference parity and visible edge/texture artifacts remain unqualified; this is not complete visual correctness or historical27 completion. PTCT texture correctness also remains unverified.
  4. Fault: Winamp MilkDrop (existing TODOS row).

Current-main refresh, 11 October 2026 (GFX-CORPUS-MATRIX-20261011)

Census. node tools/gfx-app-census.js --family=gl,d3drm,d3dim,d3d8,d3d9 --list on 4cbde5326: 92 of 294 registry apps reach a 3D API statically (ddraw 82, d3drm 14, d3dim 68, d3d8 13, d3d9 4, gl 24). Three of today's registrations are absent only because their executables are fetched, not kept, on the ops box: re_volt_demo, forsaken_demo, midtown_madness_trial (all DirectX 5/6 D3DIM, run below).

Probe. Boat bx_az67mx8t (4 vCPU, llvmpipe, Xorg :0, ops services off), main 4cbde5326, the 10-10 probe.js: one unrouted test/run.js --app=ID --no-close per arm, 35 s wall clock, final frame as PNG. Software arm is the default (D3DIM/D3DRM), --d3d9-renderer=software (D3D8/9) or --gl-renderer=software; GPU arm is --headless-gl (+ --d3dim-gpu). Up to three probes shared the 4 vCPUs, so batch counts are lower than an idle box. Apps that stop on a start-up dialog were re-probed with Enter at batches 30/90/200/400 and --stuck-after raised. Frames, logs, summaries and the scripts: scratch/runs/20261011T0600Z-gfx-corpus-matrix-w4/. This is a regression sweep at a fixed short budget; the gameplay evidence in the 10-10 table still stands where the probe reached the same point.

app software (35 s) GPU (35 s) note
dx_flip3dtl, dx_globe, dx_viewer, dx_boids, dx_tunnel, dx_twist renders renders samples animate per presented frame, so final frames differ by arm
scr_jazz, scr_rockroll, scr_fallingl, scr_geometry renders renders
scr_architec white at 35 s; renders at 60k batches (series) same a long first load, not a fault
scr_scifi, scr_oasaver uniform frame through 2.9M batches (120 s) on the default CLI clock same render with the 10-06 flags --tick-ms-per-batch=2 --dx-surfaces (scifi 237,900-byte frame): D3DRM-SCR-DEFAULT-CLOCK-UNIFORM-20261011. Resolved: a clock artifact (scene timers expire before the first Flip, so the device is rebuilt every frame); the seven savers now register cliTickMsPerBatch: 2, see re-notes/plus98-dx-screensavers.md
tomb_raider_2_demo, tomb_raider_3_demo in-level after Enter on the setup dialog same; identical frames at batches 60k/120k
gta2_demo, darkstone_demo, cmr2_demo, colin_mcrae_rally_demo, avp_alien_demo, avp_marine_demo, blood2_demo, mw3, diablo2_demo menu menu
drakan_demo, mcm, midtown_madness_trial menu after Enter on the start-up message box same midtown: vehicle menu; mcm: profile name entry
carmageddon2_demo intro, then exits itself (code 0) at ~112k batches same unattended only; the 10-10 routed race is unaffected
zuma_deluxe black at 35-40 s; title at 100 s — slow load, not a fault
re_volt_demo Acclaim splash (8.4k batches in 35 s) same new; menu not reached within budget
forsaken_demo dies: guest heap exhausted, Unable to allocate thread TLS vector same FORSAKEN-HEAP-EXHAUSTION-20261011; w5's routed command (--batch-size=200000, 700 batches) still runs clean
pawn board board
crimsonland, alien_shooter, ptct launcher / setup dialog same need their routes
quake2_demo menu menu, 1.78M GL calls
descent3_demo pilot screen same, 352k GL calls
anachronox_demo intro intro, 36k GL calls
halflife_uplink menu menu, 0 GL calls within 35 s slow start (7.4k batches)
warcraft3_demo menu menu, 504k GL calls
daikatana_demo menu menu, 1.94M GL calls
serious_sam_demo first-run message box same needs the registered route (re-note)
quake3_demo not run not run fixture: QUAKE3-FIXTURE-INSTALL-STEP-20261011
winamp main window main window MilkDrop not exercised

Not re-probed (heavy, or fresh 10-10 evidence): arcanum, aow2, tdr2000, diehard, deus_ex, ut2003/ut2004, morrowind, pirates (Codex), bw2.

Regressions since 10-10: none found. Every app that reached a point on 10-10 reached it again on both arms, and the two arms agree wherever they can be compared at the same batch. New failures: Forsaken's heap exhaustion on an unattended run, quake3's unreproducible fixture, and the scifi/oasaver default-clock frame.

Historical census recovery anchor — 18 October 2026

The original tool addition is commit 2d722ab3766c288d13796ed9dad3f1ceeca8f1eb (22 September). Its commit message independently records 27 OpenGL/Direct3D apps among 225 registry apps, seven scr_* D3DRM screensavers, and the launcher-noise cautions for icewind_dale_demo and tworld. The retained project_gl_d3d_corpus_goal.md memory identifies the originating session as bd610256-7b4f-4ed1-afa0-9c3cf3412fbb and records family counts GL7, D3DRM10, D3DIM9, D3D8 7, D3D9 2 with overlap.

Neither source enumerates all 27 IDs. A matching transcript/census output was not found in the checked local Claude project files. This narrows recovery to the original tool and registry revision plus the contemporaneous binary inventory, or the originating session's saved output. Re-running today's tool against today's registry is not equivalent; scanning old registry paths against changed/missing binaries is also insufficient without identity checks. No historical census was rerun here and no membership or backend qualification is inferred from the counts.

Current registry availability audit, October 19

Machine-readable inventory pins main 4b0c7d8b3 and the matching scanner hash. All 347 registry entries were considered against files available on this coordinator. Static names/imports/IIDs identify 201 graphics candidates, 135 with a family beyond DirectDraw; these are potential routes, not runtime coverage or gameplay passes. IID-only matches can be false positives.

Thirteen entries lack one or more directly declared executable/DLL paths, and eleven also lack their local file manifests. The inventory records exact missing paths per app. A missing manifest prevents discovery of its companion modules, so an empty family set is not proof of no 3D use. Recovery is needed for those entries; it does not block validation of other available originals. This is the current registry snapshot, not a reconstruction of the missing September 22 historical 27-ID set. Software/WebGL runtime evidence and reviewed gameplay remain separate requirements.

Software GL presentation bounds assessment — 1 November 2026

Source-only run 20261101-gl-presentation-source-assessment examined the short-pitch/large-output concern in lib/gl-render-worker.js presentation. Its local check uses bits + pitch * height, while row copies read width * 4 bytes. An arbitrarily changed descriptor could violate that relationship; this observation alone does not show a producer-reachable bug.

The ordinary window producer validates drawable dimensions in 1..4096 in lib/gl-software-producer.js, unbinds a prior bitmap, then sets the validated default dimensions. The WAT default-size setter enforces the same range. The window target constructs its front with d3d9_create_surface(w,h,32,1); that constructor allocates the aligned row pitch and backing together. For these dimensions, pitch is exactly 4*w, at most 16384, and backing is at most 64 MiB. Allocation failure returns no surface. Bitmap binding clears the window front, and presentation skips a zero front.

The remaining snapshot path was checked: gl_sw_worker_draw restores snapshot dimensions into raster-worker instance globals, but does not restore gl_sw_front. The production raster workers instantiate their own module; lib/guest-rpc.js supplies the front exported by the producing guest instance. The front-global write sites in the raster WAT either clear it or derive it from the constructor above. Thus the inspected worker snapshot path does not establish the proposed producer-front counterexample.

No endpoint test, original game, browser or build ran for this assessment; no source fix is claimed. The queued repair hypothesis requires a reachable producer counterexample before dispatch. Arbitrary descriptor corruption, other surface users, runtime lifetime safety, backend parity and the historical 27-app qualification remain unproven. source-pins.json records the eight inspected files and revision; this is a scoped source assessment only.

Native GL original gameplay refresh — 7 November 2026

Original-title evidence and recovery recipe records run 20261107-graphics-fixed-gl-originals, published as 7e26c8485. After native GL shader repair aaaf6e370, eight original startup sweeps (four titles, coop and threads) ended without escaped traps or unimplemented APIs. Six ordinary-input followups supplied the following inspected progress.

Original Executed route Inspected progress and limits
Quake II Node native GL, coop First-person world, weapon/HUD and view changes after paired W/Right inputs. Threads reached the menu only.
Quake III Node native GL, coop and threads Q3DM7/Q3DM1 worlds with bots, weapon/HUD; threads death, respawn and paired W movement. Coop W was balanced, but the later Left release was cut off by the harness deadline. Torch-adjacent black shapes remain unresolved.
GLTron Node native GL, coop Moving cycle/trail and paired A direction change, followed by a crash/winner screen. Threads reached the menu only.
Forsaken Software D3D, coop and threads Volcano world, visible firing/ammo change and paired Up view change. This does not establish a GL route.

The tested source was ce0330c444ea43742259624db5e9cd0240b2f156 plus the recorded untracked lib/wininit-boot.js patch. All 14 original runs loaded standard WASM SHA-256 e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467. The native environment was Node 24.22.1, @node-3d/glfw 7.4.1, @node-3d/webgl 6.0.3 and Mesa llvmpipe LLVM 20.1.2. Native GL and WGL regressions passed before originals; loaded-module audits matched the pinned inputs. The run retained 66 unique inspected PNGs and normally stopped its fresh box after evidence recovery.

The short startup sweep selected a Quake III logo texture rather than its display; the world claims above use compositor followups. One Q2 teardown/Q3 launch overlapped by 0.107 seconds in supervisor bookkeeping; later launches required terminal records. No new production fix was justified by this run. These results establish the named Node routes and observed controls. Browser WebGL, full visual parity, FPS, audio, game completion and the missing historical 27-ID census remain unproven. Current runtime evidence does not reconstruct that census.

2026-11-07 10:00 UTC: Wormhole/Bellhop corrected report, commit5a5f148ff, establishes zero qualified cases: a missing transitive helper import falsely satisfied broad rejection assertions. Eight observations/24 captures are supplemental only; Wormhole uses DirectDraw and Bellhop remains at an empty GDI connection dialog, with no active D3D/WebGL draws proved. The coordinator independently rehashed all185 retrieved artifacts and inspected both scenes; bx_24t85jnv was normally archived after retrieval. Fresh boids-current-frustum-1000 requires positive complete helper transport before allocation and natural completed-geometry evidence before any rendering diagnosis. No parity/FPS/audio/full27 claim.

2026-11-07 10:13 UTC: Boids current-frustum report, commit79cf7b2e1, records an original cooperative WebGL launch followed by a passive observer image-base lookup failure. No completed geometry or reviewed scene discriminator was captured; flock visibility and parity remain unknown.25 unique artifacts including modules were recovered and verified before normal shutdown of bx_c5kztwfa. Fresh graphics-remaining-worker-configs-1013 covers the four remaining Tunnel/Twist/Boids Worker configurations, requiring production metadata and scheduler qualification before any passive observer is reused. No full27/FPS/audio claim.

2026-11-07 10:23 UTC: Worker configuration report, commit46c5d37a7, records a harness failure before the first software Tunnel Worker capture: CDP tried to serialize the guestWorker controller returned by readiness. All four missing configurations remain open; no emulator defect established.25 unique artifacts including both modules were verified before normal nonforced shutdown of bx_8ptdehqn. Fresh graphics-worker-readiness-1023 must exercise a scalar readiness result and the exact capture branch with a live production Worker locally before another allocation.

2026-11-07 15:45 UTC: Expanded startup report, commit ecf85c4d9, records six serial 20-second cooperative/threaded runs. None escaped with a trap, exit, setup error or external timeout. Hexen II remained at its opening notice and did not reach GL or MIDI startup. Heretic II reached native GL book menus in both modes; Forsaken reached software Direct3D menus, with three auxiliary Workers executing in threaded mode. The worker inspected six PNGs, verified 54 retrieved artifact references and normally stopped bx_kku6pzkp. Native GL/WGL preflight and 20 sweep regressions passed; no production repair was justified. Fresh graphics-post-notice-1544 targets ordinary Hexen II notice dismissal and the uncovered startup path. These observations do not establish gameplay, visual parity, FPS, audio quality or the historical 27-app census.

Original game progress — 18 November 2026

Age of Wonders II durable lifetime report records run 20261118-aow2-durable-native-lifetime-terrain on the retained 87ef39648 module graph, rather than current main. Reviewed before/after frames show five selected units moving and movement points decreasing from 20/20 to 14/20. Mountains, trees and a tower render, while much ground stays black. The sampled Device7 strip and 1,355 completed draws exclude the chosen black ground point; the last captured upload there is black. This does not identify the producer responsible or establish a rendering repair. The durable launcher survived about 805 seconds. Ordinary Quit remains unproved after a native send-callback drive-limit failure; owned cancellation completed cleanup.

Blood II armed native watch report records run 20261118-blood2-armed-native-watch-music-cause. Reviewed frames show BUILD109 menus, level loading and a first-person world with weapon/HUD. The native watch disables chaining/uops and changes dispatch timing; this is diagnostic-route evidence, not a music repair or default-runtime stability claim. Four of five discovered instances were armed, with no observed watched music-object invalidation. World arrival followed the gameplay cutoff, so movement, firing and enemy consequences remain unproved. Quit input expired before native consumption; owned cancellation completed cleanup. Neither run establishes backend parity, audio quality or ordinary Quit.

NFS III Flip publication and original scene — 18 November 2026

Native window-present association report records run 20261118-nfs3-native-window-present-association. Repair 1fe5e1baf reserves PRESENT capacity before Flip swaps primary/back-buffer DIBs, preserving the prior publication during a synchronous wait. Native old-fail/new-pass, retirement/reuse/foreign-target negatives, exact-module normal Worker rendering controls and locked builds passed. Window/DIB/creator and sequence checks remain intact. Launcher correction 49af4ab72 separately retains native/browser/watchdog waits after submitter TERM or KILL.

The original uses retained c9 sources, 135 media files and the recovered b8c8fa6c module with the producer correction. Reviewed whole captures show the cockpit, road, rain and opponents at 0 MPH, with positive window and render target associations. The driving frame's native scene was 1329 ms old, beyond the unchanged 250 ms fence; no bind, DOWN or UP was sent. Driving, steering, Pause/Continue and ordinary Quit remain unproved. A disk-floor watchdog ended the detached run with retained cleanup; the earlier attached run has an explicit outer-wait evidence limitation. Neither result establishes the cause of the remaining scene lag or intermittent white composition.

Drakan original main menu — 18 November 2026

Post-warning startup report records run 20261118-drakan-post-warning-startup-readiness. Preparation repairs 2b8ba1b67 and 5eeed9ab8 separate completed warning input from passive startup readiness and wait for verified native launch receipts. Authored native/browser controls and locked builds passed. In the original, ordinary ShowWindow gave the main window focus about 11 seconds after the completed warning Enter and independent UP. No forced focus or visibility was needed. Reviewed whole frames show publisher/intro screens, the demo title and the authentic New Game/Load/Save/Options/Credits/Quit menu.

A finite menu Enter and its independent release timed out UNKNOWN. Local keydown/keyup delivery does not establish native consumption or UP acknowledgment; level entry, movement, attack and ordinary Quit remain unproved. Cleanup retains actual waits and EOFs. Seven selected PNGs were recovered byte-for-byte from persisted image-tool responses after unexplained deletion, matching prior hashes and sizes. Two full stdout logs remain missing; only exact earlier excerpts survive. Recovery introduced no runtime request or new screenshot.

NFS III primary-source arbitration repair — 18 November 2026

Completed publication investigation records published repair cf7044e37. The CPU fallback content heuristic selected an unrelated opaque-white offscreen surface over the submitted GPU primary. Native HWND, DIB, creator, CPU stamp and sequence validation now selects the actual primary before that heuristic. Native and normal-browser old-fail/new-pass whole-pixel controls, lifetime/foreign-target/CPU-write negatives and locked builds passed. Guest-slice, GDI and publication fences remain intact.

Fresh registered originals used the retained c9 source graph plus recorded patches, module b91a6cdc…, 135 unchanged media pins and eight live-owner module readbacks. Whole reviewed frames show the Corvette cockpit, road and rain; the compositor selects retained GPU pixels. The native table distinguishes the black primary at slot14 from the unrelated white offscreen surface at slot18.

There was no physical input. A local expired-review admission bug was corrected in c2256a77a; the later fresh run closed when its 4215ms screenshot exceeded the preselected 3000ms capture deadline, before guarded input. Driving, steering, Pause/Continue and ordinary Quit remain unverified. All13 scopes retain26 actual native/watch waits with both EOFs and no recorded leaks. These rendering results do not establish backend parity or the complete historical27-app goal.

Evidence: scratch/runs/20261118-nfs3-deferred-publication-visible-frame/result.json.

Drakan heading, flame attack and ordinary Quit — 18 November 2026

Prepared New Game report, published through 94a08881d, records a fresh native original on the retained a51912dd… module with genuine COMCTL32/COMMCTRL and unchanged media/settings. Ordinary reviewed New Game and Normal clicks reach Wartok Canyons. The coordinator reviewed whole mounted-level frames: a left pointer action changes mountain/tree landmarks, and a subsequent primary click produces orange flame at the dragon's mouth. Native consumption and independent releases are recorded. Heading control and attack effect are verified; forward travel and enemy damage remain unverified.

Ordinary Escape, highlighted Quit and the genuine in-progress YES confirmation lead to guest/native exit0 without cancellation. Final input consumption and independent UP succeed; native/watch actual waits retain both EOFs and no leaks. Exit occurs at664.838 seconds, before the720-second gameplay cutoff. Earlier UNKNOWN scopes remain closed. The preparation repair uses measured native return latency and an authored old2500ms-fail/new4000ms-pass control, with its policy fixed before the successful original. Focused native/browser controls and locked builds passed. This establishes the named native gameplay route; authentic-browser gameplay, backend parity and complete27-app coverage are not claimed.

Evidence: scratch/runs/20261118-drakan-prepared-new-game-mouse-route/result.json.

ROCKROLL current production browser WebGL motion — 18 November 2026

Current-graph report, published as 0886eb336, records a fresh registered /s original on published 5732320bb and module 9862d124…. The coordinator reviewed both whole frames: a large edge-on guitar ring changes to a smaller tilted ring of face-on guitars and picks. The 403,241 changed pixels include substantial natural geometry motion; no input was sent. Actual WebGL1 Intel/Mesa execution records completed draws 3,939→8,371 and PRESENTs 101→289, with three matching actual module receipts and eleven unchanged media paths. This is separate from earlier historical graphs.

Native/browser/watchdog waits retain exit0, both EOFs and no surviving processes or external TERM. Browser close is harness teardown, not ordinary guest Quit. Focused controls passed; a heavy authored control's earlier UNKNOWN remains closed. No production fix was needed. The report accurately retains its shared build refusal on concurrent UT callback WIP, which was excluded from the tested immutable published graph. A post-compile comment-normalization hash correction is recorded separately from canonical served-source pins. Backend parity, audio, FPS and the complete historical27-app result remain unproved.

Evidence: scratch/runs/20261118-rockroll-current-webgl-natural-motion/result.json.

GLTron browser turns, round outcome and ordinary Quit — 18 November 2026

Browser gameplay report, published as d08044e4a, records the original registered game on pinned published 44221dc3875 / module 9862d124…, through ordinary guest/render Workers and browser WebGL. All 92 registered media hashes match. Later shared main changes were excluded from this immutable run. No production code was changed.

The coordinator reviewed whole initial-arena, first-corner, second-corner and outcome frames. In one round, A turns the intact yellow cycle's trail upward; S adds a rightward segment and second corner. Trusted browser key edges, actual native input-dequeue logs, independently armed releases and all-up renderer keys accompany these consequences. Resume without further steering ends with PLAYER 3 WINS and player 1 crashed. Ordinary Quit opens credits; Escape then produces a retained owning-process exit0 before host browser close. The transitional title image named paused-baseline is explicitly excluded from arena proof.

Focused controller/capture checks passed. The report retains its historical shared-build refusal; the callback gate later passed. Successful browser, display and release children have actual waits and both pipe EOFs, and the native supervisor records child exit0/no leaks. Outer-launcher waits were not retained; direct log descriptors do not provide pipe EOF evidence. An earlier controller exit1 remains unexplained, with supervisor cleanup recorded separately. Neither those gaps nor closed diagnostics are promoted into successful gameplay evidence. This proves the named browser controls, outcome and ordinary Quit, without a complete 27-app, backend-parity, FPS, audio or game-completion claim.

Evidence: scratch/runs/20261118-gltron-current-browser-turns-round/result.json.

Alice browser player HUD and W movement — 18 November 2026

Completed player-control report, published as 03dcc660f, records normal browser WebGL guest/render Workers on immutable 57af0a44c / module 1a8a78d4…, with sixteen unchanged media pins. Natural startup, New Game/Easy, loading and the intro reach third-person Alice with the player HUD. The coordinator reviewed neutral and post-W whole frames: the timber doorway, lamps and passage move relative to the character. A later corrected neutral/W pair supplies additional movement evidence. Autonomous intro frames are not used as control proof.

The reused controller mismatched A/S DOWN and UP codes during middle navigation. That segment is excluded. Matching ordinary releases cleared the held keys; fresh W and Space inputs then had passive physical-UP checks. Jump remained visually inconclusive. The proposed mouse attack was NOT_SENT because its release child missed the parent's fixed one-second prearm acceptance. The controller closed the browser before ordinary Quit, so neither attack nor game Quit is verified. No Alice runtime defect or production repair was established.

Focused helper/evidence checks and the locked build passed. The second original's browser, display and supervisor actual waits retain both EOFs and no leaks; the first original has explicit outer/browser wait gaps, and separate EOF values for a matching-release tool were unavailable. Cleanup is distinct from guest Quit. This establishes the named browser player movement, without jump, attack, backend-parity, complete 27-app or game-completion claims.

Evidence: scratch/runs/20261118-alice-current-browser-player-hud/result.json.

NFS III browser throttle, Pause and resume — 18 November 2026

Browser driving report records helper repair 95df0faa9 and report c0b550a66. The registered original uses 135 unchanged media pins, module a6e2bb71… and pinned published source on the actual Glide/WebGL browser Worker route. This is not a D3D-backend claim. The helper moves full identity validation before the short independent-release clock; focused old-fail/new-pass controls and the locked build passed.

The coordinator reviewed whole fresh-original frames: throttle changes 2→15 MPH with RPM increase and road/house approach. Ordinary Escape opens the original Single Race - Paused menu. A later Escape resumes the cockpit and advancing race clock; the earlier Continue click alone does not establish the resume. Left input visibly turns the steering wheel, but coasting prevents an isolated road-turn claim. Earlier idle creep is not used as throttle evidence.

Ordinary Quit reaches System confirmation, but activating System does not produce an observed native exit within the preselected 15-second window. That scope closes without replay; game Quit remains unverified. Native supervisors report actual waits/no leaks and browser/display/release waits retain both EOFs, but the second original lacks its enclosing-launch receipt (outer tool143, EOF unknown). Complete cleanup attestation is explicitly not claimed. These results establish the named browser throttle/Pause/resume controls, without ordinary Quit, causal road steering, backend parity, complete 27-app coverage, FPS or audio claims.

Evidence: scratch/runs/20261118-nfs3-current-browser-driving-pause/result.json.

Blood II normal browser movement and firing — 18 November 2026

Browser world/control report, published as c5b4c4c54, records unchanged BUILD109 media on pinned 7741a4cfd and module 3e2030df…, normal 100000-step slices and five guest Workers. Music, chaining and uops remain enabled. This is separate from the earlier diagnostic native-watch world evidence. Helper media-alias and DirectDraw-menu admission repairs passed focused old-fail/new-pass controls and the locked build; no engine or music repair is claimed.

The coordinator reviewed whole initial-world/forward frames: the player moves out of the elevator into the corridor. After a restart, the reviewed firing pair changes ammo 50→49 with health 100. Turning is also recorded. Player death is observed separately, but neither its killing actor nor enemy damage/kill is proved. Actual key-code matching, independent releases and physical-neutral checks accompany the controls. The initialized WebGL bridge's sampled counters do not establish GPU authorship of the world pixels; visible rendering and backend identity remain separate observations.

Quit reaches the promotional card; after dismissal, worker 4 traps at 0xc06edb. The retained stack matches a multimedia timer callback. AM18 attribution is a bounded inference, with current occupied module generation/bytes and timer kill/selection/detach ordering still missing. No speculative lifetime fix or ordinary Quit claim is made. Browser/display/release waits have both EOFs and supervisors report no leaks, while two enclosing-launcher waits and board-watcher EOF remain explicit gaps. No complete cleanup, backend parity, full 27-app, FPS or audio claim is inferred.

Evidence: scratch/runs/20261118-blood2-current-browser-world-combat/result.json.

Forsaken normal browser movement and firing — 18 November 2026

Browser world/control report, published as 6efb02c76, records the unchanged registered demo's 873 media pins on source 5af3bae93 and module 3616571e…. The normal production browser reaches the Volcano world. The coordinator reviewed whole neutral, firing and movement captures: ammo decreases 2000→1994, and the ship advances toward the doorway. A later paired firing interval reaches 1980. Actual RenderWorker WebGL1 draw/readback and Blt presentation records accompany the world, with zero observed producer fallback/decline events. This supplies browser WebGL evidence separate from the older native software observations.

Player damage and subsequent white imagery do not establish enemy damage or defeat. Pause and return to the main menu still work. Ordinary EXIT GAME and promotion dismissal terminate the guest with code 13 and the original D3DAppShowBackBuffer/bRenderingIsOK error, so clean exit 0 is not claimed. No causal rendering or teardown repair was established in this run.

Original enclosing-tool exit 143 leaves browser/display/driver waits and EOFs missing. Separate bounded cleanup verifies the recorded processes absent but does not supply those missing waits. No complete original cleanup, backend parity, full 27-app coverage, FPS or audio claim follows from this evidence.

Evidence: scratch/runs/20261118-forsaken-current-browser-world-combat/result.json.

Hexen II normal browser encounter and ordinary Quit — 18 November 2026

Independent release and encounter report, published as 20765d3df, uses helper 34c794d54, unchanged registered demo media, normal production Workers and module 1504acf2…. Independently recovered guest/render bytecode hashes match the served module; actual rendering is WebGL1 on the reported ANGLE/Mesa device.

The coordinator reviewed whole courtyard, ramp, chamber and combat frames. The living player moves from the courtyard up the ramp into a chamber with two visible spiders. Health changes 81→50 while the player remains upright; later death is recorded separately. Incoming damage is proved. Enemy damage, defeat and guest attack consumption remain unproved. Mouse1 has an independent release ACK and physical state 0, but the DOM ledger lacks mouseup; the first Escape DOWN exits browser pointer lock. These limits are not hidden by transport ACKs.

Ordinary menu Quit and a balanced Y confirmation produce process exit code 0 before host cleanup. Runtime browser/display/supervisor/watchdog/enclosing waits have drained EOFs and no containment leaks; the separate board watcher retains an EOF-receipt gap. The helper repairs a process/document release reader; the run-local controller also adopts atomic command publication. Neither establishes a game engine or compositor defect. The isolated old-fail/new-pass control does not identify the predecessor's missing exception. The final authored startup visibility annotation was corrected using raw PNGs; visible notice transition evidence comes from a separate authored scope. Some run-local helper versions lack contemporaneous prelaunch hashes; the authenticated runtime graph is separate from that limitation. This establishes the named browser gameplay and Quit result, without software-backend parity, full 27-app coverage, FPS or audio claims.

Evidence: scratch/runs/20261118-hexen2-dialog-release-browser-gameplay/result.json.

Age of Wonders II normal browser army movement — 18 November 2026

Worker identity and map report records published source 3c6d7d115 and module 1504acf2…, unchanged registered media and real required Borland/MSS companions. The coordinator reviewed whole party-selection and two arrival frames: ordinary destination/confirmation inputs move the party southeast and farther south, with points 20/20→14/20→10/20. Pre/post identity checks and independent release waits accompany both moves. All 216 Worker identity reads complete; neither the prior identity timeout nor the callback-drive failure reproduces. The bounded trace change is diagnostic, without a causal callback repair claim.

Ground remains black while objects and HUD render. Native surface 65 already contains zero bytes at ground point (600,300) before transfer to the backbuffer. The sampled Device7 strip excludes that point; a geometrically covered neighbor maps to zero bytes in texture 68. Actual WebGL executor upload/readback at the ground point is black, with no retained draw covering it. This localizes the next investigation before transfer/upload, without establishing an intended nonblack source, missing geometry, texture defect or WebGL raster defect. The executor is observed directly; its vendor/renderer string was not collected.

The controller closes at its fixed command-wait deadline after the terrain probe. A later inspection is rejected locally before any endpoint request; ordinary Quit is unattempted. Runtime child/enclosing waits and EOFs are retained, with earlier control and board-watcher outer-tool gaps recorded separately. These named gameplay results do not establish terrain correctness, backend parity, full 27-app coverage, FPS or audio.

Evidence: scratch/runs/20261118-aow2-worker-identity-boundary-map-progress/result.json.

Hexen II actual software browser encounter — 18 November 2026

Software encounter report, published as a5bff2fa9, uses clean Git 0b55b3c2f, authenticated guest/render module a6002d5d… and 18 unchanged media/control references. The actual selector is gl-gps=software; natural completed receipts have positive nativeDraws, nativeBytes and presents from the WAT rasterizer. The software result is supported by execution evidence as well as the selector.

The coordinator reviewed whole courtyard, ramp-advance and close-spider frames. The player starts upright at health 85, advances into the chamber and receives incoming damage, reaching health 5 with two spiders visible. Mouse1 has a trusted window-capture DOWN/UP pair and physical state 0. Its DOWN arrives 11.906 seconds after the health-5 frame; the following view is death/loading. A living attack, enemy hit and kill remain unproved. The window-capture ledger explains the current document-ledger omission without recovering the earlier WebGL run's missing event record. CTRL remains the registered crouch binding.

Ordinary menu Quit/Y produces process exit 0 before cleanup. This adds a named software gameplay/exit result alongside the separate WebGL run. Backend parity, enemy damage, full 27-app coverage, FPS and audio remain unqualified. No new runtime repair is claimed; the report retains failed-control, command timing and cleanup limitations with their actual receipts.

Evidence: scratch/runs/20261118-hexen2-software-living-enemy-consequence/result.json.