Original Deus Ex: owning pixel inputs explain sampled zeros; upstream state remains unresolved

The original optional producer does execute, and sixteen actual owning-slot pixel calculations correctly produce zero from the captured inputs. An independent integer reference and exact-address replay of the unchanged original instructions match every sampled stored pixel and all eight MMX registers. No production arithmetic defect is demonstrated in this block. Why the upstream guest selects these zero light/material inputs remains unresolved: expected guest state and an upstream emulator defect are still distinct possibilities. No renderer patch, guest write/call, forced Draw, repaint or game input was used.

This follows the zero-surface cause investigation, commit 628db82fa0b1413faf9b7879d38431b2e6439173. All new evidence, helpers, raw results and tests are in scratch/runs/20261104-deusex-owning-pixel-inputs/. The predecessor's rejected thread-1 MMX samples remain rejected; the new evidence reads the initialized MMX slot 0 of the actual owning main Worker, whose guest thread ID is 1.

Natural predicate and original producer

The sole Chrome PID 14443 ran 09:55:15.676Z–09:55:57.194Z on November 4. Original native readiness passed at 09:55:28.329Z. The ordinary scheduler completed the bounded observation after 1,386,712 blocks / 27,602.940 ms, below 10M blocks / 120 seconds. Chrome closed normally, without supervisor intervention or a first-failure record. There was no second browser or reload.

The read-only helper preserves the published collector's return values and Draw/ESP acceptance. The original Draw has nine RAW identities, 192 Render ESP save/restore pairs, and a verified saved-caller return candidate. One MainLoop loop-check and one Tick-return boundary were observed during this interval; this does not identify a level or prove gameplay. The terminal has no diagnostic, module or trapping-run error. Lifetime and viewport-to-host binding remain unknown.

For this invocation, engine object 272560288, viewport 422613888, scene 425000968, Draw entry ESP 396359800, saved caller 288637550, Worker slot 0, guest thread 1 remain correlated in each retained pixel record. The original positive dimensions are 640×480. The missing optional branch is now observed directly:

These are actual target/entry/return-stack observations, beyond dimension checks or unrelated SoftDrv activity. producer.json retains seven branch checkpoints, two callee entries and their code bytes, with zero optional-record drops.

Actual pixel inputs and differential result

The extension observed 785,333 SoftDrv boundaries / 253 distinct PCs, retaining one register/code/MMX sample per distinct PC with zero sample drops or errors. There were 2,993 entries to original SoftDrv RVA 0x36509; the first 16 were retained with their following store boundaries, and 2,977 were intentionally omitted by the fixed cap. All sixteen pairs are in the same owning Draw frame. This is bounded sampling, not every-pixel coverage or a complete instruction trace.

The first sample is at original SoftDrv base 289816576, RVA 0x36509:

Input / result Actual value
ESI / EDI 389 / 1346711552
MM3, MM4, MM5 all zero
MM6 ff64c217ffc4b517
MM7 1fabe140e301d047
Texture coordinate addend, RVA 0x7d3a8 00020000fffc0000
Texture shift / mask, RVAs 0x7d3e8 / 0x7d3e0 32 / 0
Calculated texture index / actual texel 0 / 1
Texel guest / WASM address 271384456 / 135073672
Palette entry 1 guest / WASM address 412029088 / 140767392
Selected palette entry eight zero bytes
Light bytes at EBP 290296568 24 zero bytes
Store guest / WASM address 1346712330 / 474297098
Following PC / ESI / stored pixel RVA 0x36577 / 390 / 0x0000

The full 2,048-byte palette contains 1,528 nonzero bytes; it is the selected entry that is zero. RGB packing masks/multipliers are initialized: 00f800f800f800f8, 0000fc000000fc00, 2000000420000004; the dither qword at RVA 0x546f0 is 0404020402020102. The zero texture-index mask is separate from these packing masks. All sixteen samples have zero MM3/MM4/MM5, zero light bytes, texel 1 and zero selected palette entry; their nonzero MM6/MM7 values exclude the old wrong-slot all-zero MMX observation mistake.

oracle.cjs independently evaluates the original integer lane operations, signed multiply-high, wrapping adds, saturation, packing, masks and shifts. replay.cjs loads the original EXE and SoftDrv on the exact runtime WASM, loads SoftDrv at the captured base, verifies the unchanged relocated code bytes, and reconstructs the sampled inputs at the same guest pointers. It restores all registers and owning-slot MMX state, then executes the original block. For all sixteen samples, stored pixel, ECX and all eight MMX registers match both natural execution and the reference. Exact guest pointers do not recreate the original allocator history; memory placement in the isolated fixture differs.

Seven discriminating fixture controls also match the reference:

Changed input(s) Pixel Texture index
Positive palette entry only 0x0000 0
Positive MM4 light only 0x0000 0
Positive palette entry and MM4 light 0xffff 0
Both positives, texture mask 255 0xffff 47
Same, shift 31 0xffff 95
Same, shift 64 0xffff 0
Negative MM4 light, positive palette 0x0000 0

Controls are offline fixture inputs, never substituted into the natural game. These tests support correct computation for the sampled block and controls; they do not certify all MMX instructions, upstream memory generation, other pixel routes or all unsampled writes.

Source/upload correlation and observer correction

The fresh Lock pointer is 1346711552, WASM source 474296320. All nine selected Engine producer snapshots see the same 614,400-byte zero source. All sixteen stores lie within it. The final actual upload presentation 0x200001 uses source 474296320, HWND 0x10003, attached DX-owner canvas identity 10, version 8, write sequence 38. Its canonical bytes remain zero. Ten actual page import observations have zero errors and no limit exhaustion. The predecessor's source 475783168 / canvas 12 is historical, not substituted for this fresh allocation. Fresh host and 640×480 surface PNGs were personally viewed: both application areas are black, with no gameplay claim. Their exact hashes are in analysis.json and the recovery manifest.

The natural helper's unexpectedNext counter is 16. Its whitelist admitted cold block successors 0x36581 / 0x366bb, while the warmed natural decoder already had a target at 0x36577 and stopped there after the store. This is an observer coverage-check defect, not a guest failure. The raw helper, counter and records are preserved unchanged. The original bytes include the store and increment before that landing; replay predecodes it and reproduces all sixteen natural states. Separate producer-next-boundaries-fixed.cjs admits the valid landing and passes actual-WASM cold/warm boundary tests with zero flags. It was not installed in another browser. The corrected helper's full memory comparison also confirms that observation itself does not write guest memory.

Preflight, provenance, recovery and remaining gap

Before Chrome, 542 closure/helper/module/media pins passed. Tests covered the predecessor's corrected main-slot MMX helper, actual shipped owning Worker, original Engine optional absent/true/false branch fixtures, original SoftDrv zero/positive pixel fixtures, original PE/Engine/Core/WinDrv/Render Draw/ESP suite, and complete next-driver integration. Original Engine gate bytes execute in the fixture; its predicate/producer stubs are controlled fixture callees. Full-driver DOM/readiness modeling remains distinct from actual Worker/WASM execution. Actual Puppeteer 25.7.0 CJS/ESM imports, cache graph 98 page / 13 Worker scripts, 144 pinned HTTP resources, isolation/no-store and readiness positive/negative controls passed. Raw terminal/producer/publication JSON was persisted before the brand-correct asynchronous PNG export.

An initial offline test wrongly expected the next block boundary before an unbroken conditional branch; its failure is retained in prebrowser-attempt1/. The fixture expectation was corrected before the final passing preflight. The later natural warm-cache whitelist flag was repaired only offline as above. There was no unexpected runtime failure, retry, new Chrome or forced execution.

The closed natural namespace recovered 106 artifacts / 7,040,067 bytes, full SHA/size checked, at 09:56:36.212564Z. All 542 pins remained unchanged and Chrome/driver were absent. Separate offline replay/correction artifacts were also recovered and SHA checked. analyze.cjs validates hashes, exact frame/owner, original branch return PCs, pixel/store ranges, counters, controls and upload correlation. The runtime WASM SHA-256 is 129a71a01a5fbc8b07d6fa242a5a3ebc6a29f5aed9fb4b6a47d9318a764c1f52.

source-provenance.json records local session HEAD, relevant uncommitted binary diff, and exact reconstruction of 277 tracked served source inputs: 275 match 628db82fa0b1413faf9b7879d38431b2e6439173; the two D3dim files match recorded Git blobs at 5a834874fd889eceb98dc72a5c6b01946972a01d and e947be8ad308f3eae93bd5b2914406ca704fbced, with their small binary adjustment patch retained. These files were not edited. Tested WASM identity is pinned separately; no current-main rebuild is claimed. New owned small helpers are retained as patches, and media is referenced by existing path/SHA. No source archive, module copy or media copy was made. Only this note is committed; no production code changed, so no new full-build claim is made.

Precise remaining gap: the producer of the zero light stream at 290296568 and following qwords, the selection/initialization of zero palette entry 1 at 412029088, and the zero texture mask at 290329568. Their upstream provenance must be observed before choosing a production fix. The present sampled pixel arithmetic agrees with expected x86 integer semantics for these inputs; that does not establish that the guest was supposed to choose these inputs or that every pixel has the same cause.

Custody: bx_mpnr7rhj is left RUNNING (API idle, healthy/hydrated), expiry 2026-11-04T10:27:15.142Z, coordinator owns shutdown. Chrome is closed and all unique evidence is recovered before the 10:17Z close / 10:22Z recovery deadlines. No lifecycle action, source-box access, local heavy runtime, subworker, successor, Heroes II or deployment was used. The one authorized natural observation is consumed. Ownership returns to the coordinator; this single assignment stops.