Deus Ex upstream light provenance: module buffer identified; writer proof missing

No production emulator defect was demonstrated, and no production code changed. The precise blocker is missing ordered writer/initializer and allocation-lifetime evidence for the zero light, selected palette entry and texture mask. This run narrows the upstream locations using original PE bytes and retained natural samples; it does not establish the upstream cause.

Predecessor: owning pixel inputs, commit 5c55e600d. Evidence: scratch/runs/20261104-deusex-upstream-light-provenance/. Local starting HEAD was 38238efd5b786b6a00b594ec5a6f775a3ce90787.

Address provenance and candidate writers

The predecessor's actual SoftDrv base was 289816576. Original bytes at RVA 0x36457 load EBP with module RVA 0x752f0, resolving to 290296560. The retained boundary at 0x36472 has exactly that EBP. The sampled pixel's light pointer 290296568 is module RVA 0x752f8, after the original add ebp,8 at 0x364d8. This light address is a module buffer, rather than an independently identified heap allocation. Its initializer remains unknown.

The original setup path loads MM3 from descriptor ECX+0x18 and MM4 from ECX+0x20, then stores them to shift RVA 0x7d3e8 and mask RVA 0x7d3e0 at instruction RVAs 0x36487 and 0x3648e. The latter resolves to the historical mask address 290329568. At retained boundary 0x3649e, MM3 and MM4 are zero; this is consistent with the observed mask but is not an ordered before/after store witness. The incoming descriptor address and its initialized bytes were not captured.

The pixel's palette base is 412029080 and texel is 1, giving selected entry 412029088. Original setup RVA 0x36464 loads EDX from pointer-global RVA 0x7d308. Static direct stores to that pointer occur at 0x26596 and 0x265ba; the second follows an indirect call with a pushed 0x800 size. These are candidate selection/allocation sites, not proof of which site produced this palette or initialized entry 1.

Static light-buffer writes also exist, for example 0x3d5ee onward using [ESI+EDI*8] after loading ESI with module RVA 0x752f0, and 0x3d693 / 0x3d69e using an address derived from that buffer. Their natural execution and overlap with the sampled qwords are unproved. No historical address was used as a watch target in a new runtime.

audit.py independently parses PE section mappings, verifies the exact original instruction encodings for the EBP setup and shift/mask stores, checks the captured module-relative addresses and owning slot 0, and retains bounded disassembly excerpts and input SHA-256 values. The retained per-distinct-PC samples do not carry sequence numbers or per-sample Draw identity; they cannot establish adjacency, a last writer, or a complete chain through these candidate sites.

Verification and observation limit

On retained runtime bx_mpnr7rhj, the unchanged predecessor producer-next-boundaries-fixed.cjs passed its actual original PE/Engine/SoftDrv WASM fixture tests again: optional absent/true/false controls, zero/positive pixel results, owning main Worker MMX slot 0, observer memory unchanged, cold successor and warm predecoded 0x36577 landing. The original-instruction replay again matched all 16 captured pixels and all eight MMX registers against the independent integer reference, plus all seven discriminating controls. Only output destinations were redirected to this run's namespace; predecessor files were not edited. These are offline fixtures, not new natural observations.

The local static audit passed after correcting a text-excerpt parser that initially tried to parse the Disassembly heading as hexadecimal. That failure and correction are recorded; they do not indicate a guest defect. Four new remote artifacts were recovered with full SHA/size verification.

No Chrome was launched. At dispatch/read time the fixed 10:17Z Chrome-close deadline left under four minutes. A new observer capable of correlating fresh palette allocation, descriptor inputs, module-buffer writers and the owning Draw was not implemented and preflighted within that window. The required complete actual PE/Worker/dependency HTTP/cache/CJS/ESM preflight therefore was not claimed for a new driver. No browser/reload, guest call/write, forced Draw, repaint, input, PNG export or speculative production patch occurred. The natural-observation allowance was unused; this run does not extend the predecessor's coverage.

Next evidence needed is an ordered, bounded read-only observation of the setup descriptor and pointer-global selection, correlated with the fresh owning Draw, followed back to actual overlapping light/palette writers and allocation lifetime. Original instructions plus those captured inputs must then be replayed against an independent reference before selecting a repair. The current evidence cannot distinguish expected guest state from an upstream emulator defect.

Custody

API inspection at 10:14:13Z found no Chrome process. Final box info is retained; the box is left RUNNING (API idle, healthy/hydrated), expiry 2026-11-04T10:27:15.142Z. Coordinator owns shutdown and receives custody back. All new remote artifacts are recovered. No lifecycle, allocation, source-box, local heavy browser, subworker, successor or deployment action was taken. This is a note-only commit; no new full-build claim applies. Source recovery uses Git SHA and small owned-file patches, with no source/media/module archive.