Registered DX samples: November 7, 2026

Run scratch/runs/20261107-graphics-dx-samples/ adds named coverage for dx_globe, dx_viewer, and dx_flip3dtl. All three started and rendered in cooperative and threaded execution with software and WebGL D3DIM: 12 bounded startup cases, 20 seconds per title/mode/backend, followed by ordinary-control observations. No escaped guest trap, unimplemented-API failure, or new reproducible original-scene rendering failure was found in the accepted runs. No emulator implementation was changed. The registered Globe's historical band observation remains open. Historical27 membership remains unresolved.

Source, closure, and box

One fresh box, bx_a2jkhpsf, was created at 2026-11-07T07:58:02.490Z with a persisted UUID Idempotency-Key (boat-create-key.json). Its original expiry was 2026-11-07T08:33:02.506Z. Only one browser/runtime ran at a time. No held-risk lane was reused, stopped box resumed, or public service deployed.

Git reconstructed source commit fe995f4db6d9d1b794f94e0959fd0717bed3cfda, then applied explicit binary-capable tracked and new-file patches. Git remains the source archive; no source/media archive was made. source.json, dirty-source.patch, and new-source.patch record the shared checkout's tested state. Patch SHA-256 values are respectively 476714a55e6515fb874d6d1cdac572bd7dede67aeb5892223b770eed9682b29d and fb82c82addff18a7842dea81e108ae591e1d3990991ca15f1ba4f14edb482476. This is the tested source identity, independently of this report's later commit.

runtime-pins.json records 1,560 paths with size/SHA-256: registered EXEs and scene files, runtime source, fonts, available shared DLL directory, icons/TLB, package metadata, and both built modules. All registered lists matched 20261107-corpus-next-inputs/registered-pins.json. The ignored binaries -> test/binaries link was reconstructed explicitly. An expanded 166-file icon/TLB inventory was checked against the same Git commit already reconstructed before launch. Postflight rehashed all 1,560 files with zero mismatches, including unchanged original inputs. Media is referenced by path and hash, without repeated copies in evidence.

Shared DLL closure matters: Globe/Viewer actually load pinned d3drm.dll and oleaut32.dll; their OLE32 imports use the existing builtin route. Imports and the complete available DLL directory are pinned, rather than assuming the coordinator's scene lists included them. Loader probes for executable-directory d3dxof.dll return 404 before shared-DLL resolution; raw logs preserve this.

The locked local build passed (build.log). Retained modules:

Module SHA-256 Execution evidence
canonical e41511fe861363fbe1639d80f20cbcdd7240f46937c93a801eaa256ca094c467 Observed compiled/instantiated module and CPU/render-worker module messages
compatibility 04acbb89f66ad9814836b5e2c34de659b518f4e1994f50c4ee8069c991f02f72 Pinned and retained; not exercised by this Chrome

Chrome was 151.0.7922.108, viewport 1100×820, cross-origin isolated, served only on box loopback. The browser's supplied unsafe-SwiftShader permission did not determine D3D ownership. Cooperative WebGL records actual D3DIM GPU targets; threaded WebGL records render-endpoint statistics and context identity. Both report ANGLE / Intel Mesa UHD Graphics 620 (KBL GT2), OpenGL 4.6 Mesa 23.2.1, WebGL 2.0. Accepted WebGL startup statistics report zero GPU errors/fallbacks. Software observations have no D3DIM GPU target; threaded guest configuration records d3dimGpu:false. Flip3DTL additionally records software render-endpoint completion counters and no GL renderer. Globe/Viewer software endpoint counters are empty, so no draw-count or performance conclusion is drawn from them.

served-and-module-audit.json checks 4,149 retrievable HTTP 200 bodies against the closure with zero mismatches. Another 439 response bodies were unavailable to Puppeteer's body reader; this is explicitly not complete wire-body hashing. Direct pre/post file hashes and actual module instrumentation provide the closure/loaded-module evidence. Optional bundle/auth/build-info/favicon/app-icon 404s and DLL search probes are retained, not silently counted as successes.

Named observations

Registered app Software coop / threads WebGL coop / threads Ordinary route and inspected images
dx_globe Both 20-second starts pass Both 20-second starts pass Render menu → Wireframe changes solid textured globe to visible triangle edges in both modes/backends; later images show animation
dx_viewer Both 20-second starts pass Both 20-second starts pass File → Open Mesh; select exact listed mslogo.x; Open; Renderer → Wireframe. Loaded lettering and wireframe geometry are visible in all four cases
dx_flip3dtl Both 20-second starts pass Both 20-second starts pass Original Windows-textured cube, HAL-labelled window/HUD, changing orientation. Software additionally receives ordinary F7 down/up and visibly shrinks the cube

Accepted startup artifacts are recovered/startup4-webgl-* and recovered/startup-software-*, with raw console/guest logs, state snapshots, 10/20-second PNGs, response receipts, and trial results. Control artifacts are controls2-webgl-* (five trials), controls-webgl-dx_globe-coop (the sixth visually successful route), and controls-software-* (six). Images were personally inspected, including all twelve final startup images, Globe menu / wireframe views, Viewer loaded / wireframe views, and Flip3DTL rotation / F7 before-after views. Native browser input, trusted event observations, paired press/release acknowledgements, and final released mouse state are preserved where recorded in actions.jsonl and snapshots. The earliest Globe control trial predates the action-receipt extension; its native input observations and captures remain available.

These are the registered scene files. Globe retains registered sphere3.x (56 vertices / 108 faces, SHA-256 bb8bd85a4f13ef27777690295264f7f2475b3b4bb4e936e3b3394ee08c05360f). The separately named archive Globe's 296-vertex / 588-face mesh and viewport fix are outside acceptance. Viewer opens the original registered mslogo.x (SHA-256 bbe967bbe686d5b3bc25979b70abf2d8ce6aea4b821f4ced9ca59d7902a6727d) through its ordinary file dialog, not a replacement fixture/configuration. The focused Viewer selection regression's synthetic fixture is supporting test evidence only. Flip3DTL's historical Frame04211 is not an exact-frame requirement; no displayed FPS figure is treated as measured performance.

No software/WebGL frame parity is claimed. Captures at equal wall time have different animation progress. No change was inferred from those differences, and aligned-callback parity was not established. The historical Globe band remains visible; the earlier trace analysis records original-texture bands, valid RGB565 texels, 108 primitives and matching per-triangle wrap UVs. This run neither changes that diagnosis nor claims the band fixed.

Tests and bounded investigation

The four supplied route tests were inspected before choosing inputs. Three focused tests executed on the reconstructed box passed:

The additional unchanged test/test-d3dim-texture-wrap.js failed in its GPU recording fake, after its software assertions passed. The fake has a 16-bit target whose device implements only draw; current _submitImmediate also calls _storeNative16, which requires device.gpu.gl. The failure is an undefined synthetic GL device, not an escaped API/trap in a registered scene. Its raw stack and test log are retained. No shared source/test was edited.

A run-local diagnostic, texture-recording-control.cjs, bypasses only _storeNative16 while asserting that the target has no device.gpu, then runs the original test unchanged. It passes the existing software, descriptor-field 33, WRAPU/WRAPV/WRAP0 and GPU per-triangle coordinate assertions. This isolates the recording-fixture incompatibility; it does not test native16 storage or turn the unmodified test into a pass. The documented wrap contract under test is short-seam interpolation, independently per triangle and enabled U/V axis; no observed contract violation justifies an implementation repair here.

An early Globe control assertion also incorrectly expected checked bit 4 from menu_child_flags; it returned 0 although the captured renderer visibly switched to wireframe. The harness assertion was removed only for subsequent run-specific observations; raw failure and successful images remain. This is not reported as an emulator fix. An implementation change would need a reproducible original failure and old-failing regression first.

Initial picker/provisioning trials are preserved separately: hidden app-picker options prevented initial launch; missing ignored binaries link caused the first three startup3-webgl launches to return 404. Only complete clean startup4-webgl and software batches count as startup coverage. Read-only retrieval encountered HTTP 500 and one hash mismatch while an active log grew; completed evidence was subsequently recovered from the running box with fresh read-only requests. No unknown guest input was replayed and no stop/resume was used to recover it.

Recovery, cleanup, and next work

final-02-inventory.json / final-02-verified.json record 404 recovered evidence files, 13,792,616 bytes, SHA-256 verified. This includes unique captures, raw logs, failed test output, terminal phase records, and postflight.json. Earlier partial retrieval attempts are retained and are not the authoritative final inventory. Source patches and both modules are retained separately locally.

Postflight found all 42 recorded browser/server PIDs absent, no active run processes, all eight server ports closed, all 34 attempted trial browsers closed, and no trial profile directories. Only after verified recovery, the explicitly delegated normal nonforced stop completed at 2026-11-07T08:17:00.728Z. Final box state is archived, stop operation stop_12c3924caa965bc0 completed, forced=false (boat-final-state.json). No runtime ownership remains to hand back. Local free disk remained above 2 GB.

Next useful work is to repair the narrow recording-test fixture separately, and, if investigating Globe further, capture the registered scene's draw stream at aligned animation callbacks against a documented/native reference. Do not replace its mesh to obtain acceptance. No candidate-corpus manifest entry was found for these three registry IDs; this is investigation evidence with no unrelated candidate attachment, full27, audio, FPS, or native-parity claim.