Original Deus Ex zero surface: producer executes; natural arithmetic inputs remain missing

The black surface is already zero after the original producer runs. A fresh natural Draw passed the 640×480 scene gate and executed SoftDrv's original pixel loop before Unlock uploaded the same zero storage. The evidence still does not distinguish expected guest startup/material/light state from an emulator arithmetic or memory defect. No production renderer patch is justified. The smallest repaired observable is a run-owned, read-only MMX reader using the actual owning main instance's MMX slot, with actual-WASM positive/negative tests. Its repair was tested offline; no second browser ran.

This follows c0666fa0919b8f99d1745b280e3e34a0cd5432b5 owning natural evidence, the natural surface correlation, and original ESP-aware Draw notes. The older blank-world note describes a different historical launch/capture and is not evidence of the cause in this original natural launch. Accepted predecessor artifacts are referenced by full SHA in historical-refs.json, without rerunning historical packaging. All new evidence is under scratch/runs/20261104-deusex-zero-surface-cause/.

What the original instructions and fresh execution establish

Original Engine RVA 0x8ee89 tests scene fields +0xa8 and +0xac and branches past the virtual producer call at 0x8eeb6 when either signed dimension is not positive. Actual original Engine instructions, loaded by the pinned WASM loader, pass controlled 640×480 inputs and bypass for zero width, zero height and negative width. The natural scene pointer was 425000968; both observed gate boundaries record 640×480. This excludes that dimension bypass for this invocation. It does not by itself identify the level, texture, light or expected visible scene.

The sole Chrome PID 88503 ran 09:35:12.313Z–09:35:38.381Z on November 4. Native GIsRunning=1 and visible HWND 0x10003 passed at 09:35:20.536Z. The read-only extension was installed on the owning Worker before the ordinary collector started. It wraps the existing observer boundary, preserving its return value and all Draw/ESP acceptance rules; it does not execute guest calls, write guest memory, alter scheduling, force Draw, repaint, input or upload.

The extended 10M-block/120s ceiling completed naturally after 898,943 blocks / 16,140.805 ms. Frozen terminal: saved caller 0x1134426e, thread 1, yield 0, halt 1, no diagnostic/module/trapping-run error. All nine original Draw checkpoints, 192 Render ESP save/restore pairs and nine RAW identity records are retained. Identity errors are zero. Lifetime and full binding remain unknown. MainLoop loop-check/Tick-return counts are both zero in this interval; a complete Draw is not promoted to a new MainLoop witness.

The producer extension recorded 870,480 in-frame boundaries, nine selected Engine checkpoints and 745,335 SoftDrv boundaries / 252 distinct PCs. Every distinct SoftDrv PC has a first register/code sample: zero sample drops, zero recorded observer errors. These are basic-block observations, not a complete instruction trace or an every-pixel store ledger.

At Lock return, after successive original producer calls, at the scene-branch join RVA 0x8eeb9, immediately before Unlock at 0x8ef6f, and at Unlock return, the same canonical 614,400-byte source has zero nonzero bytes. Thus a missed host upload cannot explain the absence of nonblack bytes at these producer boundaries. An intermediate nonzero write followed by clearing is not excluded by these snapshots.

Original SoftDrv first-PC samples include RVA 0x36509 with ESI 433, followed by 0x36577 with ESI 434 and EAX 0. The unchanged original straight-line block between these PCs loads a texel, multiplies its palette/light vector, packs RGB, executes movd eax,mm0, stores AX to [edi+esi*2], then increments ESI. The recorded EDI is 1348198400, the natural Lock pixel pointer; that first pixel address is source+866. This is evidence of a zero result in the original pixel-producing route, beyond merely witnessing Render's ESP reuse. First-PC samples are not a complete per-invocation edge history; they do not establish that all other pixel writes computed zero or why this one did.

The page retained 13 actual import observations, zero errors and no limit exhaustion. Final presentation 0x200001 uploaded storage 475783168, matching Lock/Unlock; target HWND 0x10003, canvas identity 12, attached and DX-owner canvas matches, version 6, write sequence 75. Canonical bytes remain all zero. The predecessor's storage 475574272 is retained as historical data, not substituted for this fresh allocation. Surface rotation and distinct frame-layer canvas identities remain recorded; they are not assigned an unproved lifetime/binding meaning.

Discriminating original pixel test and the observation defect

A separate offline fixture executes the unmodified loaded original SoftDrv RVA 0x36509 pixel block, through its real MMX instructions and scalar store, on the exact runtime WASM. With controlled texture/step/mask/light inputs:

Fixture Stored original 16-bit pixel
Zero palette/light vector 0x0000
Positive palette/light vector 0xffff

These are fixture inputs, not reconstructed natural game inputs or a manufactured frame. They show this original instruction block can produce and store nonblack pixels through the same interpreter. They do not certify every MMX operation, exclude an input-dependent emulator defect, or prove natural black is expected. The exact outputs and module/media identities are retained in the offline logs and original-soft-producer-test.json.

The new natural helper also attempted MMX reads at MMX_FILE + guestThreadId*128. All those natural MMX samples are rejected. The guest-visible current thread ID is 1, but the owning main instance initializes its execution slot as 0. Source analysis of the default MMX base and init_thread, plus actual-WASM local and shipped owning-Worker fixtures, demonstrates the distinction: movd mm0,eax with EAX=0x13579bdf writes slot 0; the guessed thread-1 slot remains zero. A successful MMX-to-EAX readback independently confirms the value. Consequently the captured all-zero MMX arrays are an observer defect, not proof that the guest's actual MMX register file was zero.

The browser namespace stays immutable. Separate producer-fixed.cjs requires owning Worker slot 0 and reads its initialized main MMX slot using the pinned generated region map. Its actual-WASM test observes the positive MM0 sentinel, checks dimension positives/negatives and original SoftDrv bytes, counts repeated PCs, rejects duplicate installation, skips absent frames, forwards the original observer result, and compares guest memory before/after observation unchanged. test-mmx-worker.cjs separately verifies the actual shipped owning Worker and reproduces the old reader's false-zero negative. These are offline fixes/tests, not retroactive replacements of the natural arrays. No production source changed.

The dimensions alone do not establish execution of Engine virtual call +0x70: the optional viewport +0x34 object and its virtual +0x60 predicate can still bypass it. The natural helper did not record that predicate or this callee entry. SoftDrv execution elsewhere within this Draw does not fill that specific gap.

The exact remaining causal gaps are that optional predicate/callee boundary, and the owning main-slot MMX inputs, texture texel/palette/light values and shift/mask globals at the first original pixel calculation, together with its output/store boundary. Without those values, expected zero shading and an input-dependent interpreter/memory defect cannot be separated. The corrected read-only MMX observer is executable and tested; texture/palette/global input capture is still missing. No claim is made that that complete next observation is already implemented or browser-qualified.

Preflight, captures, evidence and custody

Before the one Chrome, all 525 exact closure/helper/module/media pins passed; the source, native readiness recipe and original module/media were unchanged. Actual original EXE/Engine/Core/WinDrv/Render WASM fixtures, owning Worker, Draw/ESP suite and complete next driver passed. The driver fixture models only DOM/readiness and the new observer install/read messages; its WorkerLink/WASM execution is real. The new helper's own tests use real WASM module mappings. Actual Puppeteer 25.7.0 CJS/ESM imports, generated map, cache graph (98 page / 13 Worker scripts), 144 served SHA/isolation/no-store resources and selector / readiness positive/negative controls passed before launch.

Two offline preflight attempts failed before Chrome: first the modeled driver fixture lacked the new install message; then a retained first-failure artifact made the repaired fixture fail its clean-output assertion. Both failed attempts and their logs were preserved; the fixture and output isolation were corrected before final preflight. A local launch-receipt script then raised a missing json import after spawning the one driver; launch identity was recovered read-only. The launcher was not rerun. The actual browser has no first-failure artifact and closed normally. Raw terminal/result persistence precedes PNG export.

I personally viewed both fresh PNGs: the host application area and the 640×480 surface are black. They are not gameplay success:

The closed browser namespace recovered 138 artifacts / 7,411,702 bytes, full SHA/size checked, snapshot 09:36:04.990658Z; all 525 original pins remained unchanged and Chrome/driver were absent. Separate offline artifacts have a full SHA manifest. analyze.cjs checks all recovered hashes, natural raw identities, checkpoint/producer counts, source/upload equality, rejected MMX evidence and original pixel positive/zero controls. Git provenance records local main fb9b73fedc815ae88fcbabd0d28e5c57d7490d13, the inherited exact served manifest, relevant binary diff and owned new small-file patches. Runtime WASM SHA-256 is 129a71a01a5fbc8b07d6fa242a5a3ebc6a29f5aed9fb4b6a47d9318a764c1f52. No source archive, module copy or media copy was made. Only this note is committed; there is no production-code commit or new full-build claim.

Custody: bx_3k2ukgfj remains RUNNING (API idle, healthy/hydrated), expiry 2026-11-04T09:55:33.875Z; coordinator owns shutdown, hardclose 09:45Z and recovery 09:50Z. No browser remains. No new box, lifecycle operation, source-box, local heavy browser, subworker, successor, Heroes II or deployment was used. The authorized observation is consumed. The next action is the precise original branch/pixel-input observation above under separate authorization. This task stops after publishing this note and returning custody.