Forsaken native additive-command origin, 2026-11-12

The original application explicitly submits additive render-state commands before the following scene Execute. The retained command bytes, instruction parsing, natural state stores and packet state agree. No runtime mismatch was proved; no blend, clear, texture or other runtime repair was made.

This continues the paused-worker/raw-state investigation at eec039872. That run's first retained white witness remains GPU30810, 10339/0, texture handle 208 with ZWRITE=1. Its snapshot/replay path is closed by its own evidence. This fresh run does not substitute its different first white witness for that historical one, or confer independent parent acceptance. Parent full artifact review remains pending.

Evidence is in scratch/runs/20261112-forsaken-native-additive-origin-1715/. The source base is eec0398723f94568c7d524e9a94e791ad407651f; exact binary source/helper patches, pins and compressed-byte recovery hashes are retained. Git is the source archive. No source-tree or media archive was created.

The final 1645 attach, private pinned translator/boundary, input identity, ACK/cycle and telemetry helpers are byte-identical. The added run-local observer reads Execute object metadata at the public API entry and command headers/state records at existing natural Execute trace imports. It records handle, buffer, offset, size/count, bounds, raw state bytes, parser triples and return value. It adds no guest writes, calls, render fences or state substitutions. Its order shares the existing natural post-store/enqueue boundary serial. Causal command selection is bounded by owned enqueue, not later publication.

Controlled byte/triple/count/bounds/high-address checks pass. Actual Chrome CPU and GPU production bootstrap controls pass, and production Encoder bytes and messages match an unobserved Chrome control. The first CPU bootstrap control's fixed 250 ms ready sample was premature; its Chrome closed normally. A fresh control with a bounded readiness wait passes. Both results are retained. Prior negative attach and identity controls are reused with their recorded hashes. The actual local Transport.provision whole-input seam, stale-pin rejection, safe Python imports and transitive JS imports pass before allocation. All 1,557 remote provision paths are checked, followed by remote final import/pin checks before Chrome. Runtime, normal module and media pins match 1645.

Fresh normal original menus, NORMAL difficulty and VOLCANO briefing lead to gameplay with five lives and SHLD128/HULL128. Native a600/a700/a300 movement is retained. The corridor capture already shows four lives and respawn before the Control1800 primary pulse; preserve that difference from the predecessor. post-combat-passive shows the bright rotating death scene, and post-death-passive at 17:23:54.165Z shows the white viewport with yellow fragments. There are eleven full PNGs with nineteen bound identity fields, eight independently armed/released native pulses, sixteen trusted input edges and eleven command ACKs. There is no replay.

The actual owner is WebGL2 / Intel UHD 620 through ANGLE/Mesa 23.2.1, endpoint 1. Both owning worker bindings identify normal module 04c534192ee3fe39e388536c2ce710c12c47a5cd2ca4d95c19f9d00b1e8f0a17. CPU target is BDD784E98E3F72FDA3099766B5780935; GPU target is C84956A423D2752920C085168E330472. These claims come from actual owner, epoch, module and frame bindings, not the page's dropdown or counters alone.

The fresh run's first retained selected-pixel white after the combat-ready fence is GPU19828, 6504/1, in diagnostic-tick-94.json.gz, drained at 17:22:20.299Z. It changes [214,211,231,255] to white using texture handle 244, ZWRITE=0 and ONE/ONE. This precedes the corridor capture and the primary pulse. Its public scene Execute15988 uses handle 136455784, buffer 19178716, buffer size 32767, instruction offset 3200 / length 452, flags 1, and returns zero. Before it, Execute15986 uses state-only handle 136458464, buffer 19799476, size/length 48 and offset 0. Its header 08080500 explicitly requests five 8-byte records: (14,0), (27,1), (19,2), (20,2), (21,2). Parser triples, bounds, stores, live/owned/publication/replay/workspace/description agree.

The later scene witness GPU19857, 6518/0 changes [255,243,255,255] to white with texture handle 210, ZWRITE=1 and ONE/ONE. Its source scene Execute16009 also enqueues texture handle 208 with precisely the predecessor's eight retained raw state values, before the 210 draw. The 208 packet is 6517/0; it is not a new selected-pixel white-transition witness. Its application source is handle 136458432 (0x082230c0), buffer 27853204 (0x01a90194), size 6784, instruction offset 5184 / length 1192, viewport 136454224, flags 1.

Immediately before that scene:

The runtime's src/09aa-handlers-d3dim.wat Execute walker reads those same header triples and forwards opcode 8 to d3dim_exec_state_walk. The latter reads consecutive (state,value) pairs at 8-byte stride and calls d3dim_set_render_state in src/09ab-handlers-d3dim-core.wat. For these observed commands bSize is exactly 8; enums and values agree through lowering. All 33,723 retained changed-state applications match an explicit source record. There are zero retained command/parser/application mismatches and zero observer errors. This excludes that parser/state-store path for the witnessed inputs; it does not prove the application intended the final rendered image.

Coverage limits are material. There are 158 drains, 510,689 boundary rows, 385,754 CPU rows and 8,607 GPU events. A long command-8 passive capture causes 3,803 boundary drops (serial gap 403314 to 407118), fifteen GPU event drops, and one missing Execute24023 return in that gap. No return is inferred for it. There are 203 unjoined replay keys, including later tail keys. Eleven ground drains reach the existing 32-record limit, including tick94. Its boundary and GPU event streams have no recorded drops, but its capped ground records still preclude an exact/global first-white claim. Native texture samples remain observation-time level-zero candidates, not an exact sampled-texel or original asset/writer oracle. Existing predecessor startup drops and its own trace/tail limits remain unchanged. The selected-pixel observation does not timestamp the first instant the entire viewport turns white.

The next specific unobserved producer boundary is the original guest path at 0x469275 (observed return 0x46928e) selecting the state-only buffer 0x00ff59ec before the scene path at 0x4417f2 (return 0x44182f). Why it selects that additive pass, and the preceding texture/vertex/asset production, remain unproved. White output is not evidence that those native requests are wrong. No mandatory fix follows from these results; further work needs a specific producer or texture/state mismatch rather than a rendering workaround.

Chrome and browser runner have actual zero parent waits and both EOFs. The outer API process 8238 is independently observed exited/known with exitCode=0 and untruncated streams. All 302 unique remote outputs were downloaded twice independently and reread/hash-verified locally before delegated normal stop. Box bx_vkm7au3e initially expired at 19:15:39.178Z; its final state is archived / expiry null. Stop stop_2ff25b68c4c915cb completed at 17:24:59.215Z, forced=false, error=null. Local lifecycle client, actual owning parent and independent deadline watchdog all have zero waits and EOF evidence. No missing runtime wait is filled from another process's result. Do not resume or delete the box. Sampled actual local disk stayed above 1.5 GB; no RAM gate was used.

Only this note is committed. Runtime code and original binaries are unchanged; there is no changed-original before/after or locked-build claim, and no parity, FPS, audio, enemy-defeat, level-completion, full-27 or broad crash-free claim.