Conan demo: WORLD writer provenance, 2026-11-05

Task: NEW-GAME-CONAN-DEMO-20261010. Evidence is in scratch/runs/20261105-conan-world-writer/; Git is the source archive.

This investigation follows bfba3542a and the first-draw findings. The predecessor captured device 136450064/generation 1, queue 62083, with a zero WORLD upper 3x3, nonzero translation, NORMAL input, lighting and specular enabled. WORLD slot 10 agreed between the producer and executor. Those facts did not identify a writer.

Scope and diagnostic

The ordered legacy-alpha16, AoW2 callback and Aranna releases were observed at 21:29:42Z, 21:50:39Z and 22:10:42Z. Independent cleanup verified 611 absent PIDs, 215 refused ports, 74 absent profiles, four free locks and no heavy runtime. The gate was 22:11:54Z; the hard deadline is 22:56:54Z. Before that gate only run-owned helpers were prepared; no shared code, tests, builds or guests ran.

The diagnostic is disabled by default. An existing capability query enables value-only packets on the existing trace import. The actual owning CPU copies 16 source and destination words, per-instance sequence, device, owning program or block, guest source address, EIP, thunk, entry ESP and return address. There is no added guest export, live-export call, guest write, matrix substitution or change to the strict renderer failure.

Operations: 1/2 live SetTransform before/after; 3/4 recorded SetTransform; 5/6 Apply; 7/8 Capture; 9/10 MultiplyTransform; 11 reset program replacement; 12 initial program attachment. CreateDevice retires its 32-byte stdcall frame before attachment, so that route reconstructs the entry ESP when recording caller context. Other routes retain their original API frames.

Host storage is bounded: 32 incomplete packets, 16 device histories of 32 rows, 64 block origins and 16 last live writers. Draw lowering copies WORLD words and history independently of the render payload. The actual submission receipt attaches generation and sequence; each device keeps 32 such snapshots plus its first zero-upper-3x3 draw. Rows and snapshots are deeply frozen. The last live writer survives unrelated later recorded writes; Apply retains a copied recorded/captured source origin when that origin remains in the bounded map. Sequences are per owning CPU instance; cross-thread total mutation order is not claimed. Missing origins and packet errors must remain explicit.

Semantics and controls

SetTransform sets the selected transformation state. State blocks capture or record state; Apply changes only captured states. MultiplyTransform must multiply the supplied matrix by the selected state. The existing emulator handler succeeds without multiplying. That separate source gap is not proof that Conan uses it or that it caused this failure; this diagnostic preserves the existing behavior and distinguishes its before/after state.

Published diagnostic: ffe79b45f on main. Focused stateblock tests passed, including copied words, all mutation routes, bounded history and last-live-writer retention across later recorded writes. The required locked full build passed after the caller-context correction. The existing real owning-WebGL-Worker strict-normal diagnostic passed all 24 controls.

The generated ordinary x86 PE calls imported D3D9 creation and actual COM vtables, including Set/Get/MultiplyTransform, Begin/EndStateBlock, Apply/Capture, Reset and lit DrawPrimitiveUP. The production owning guest Worker and actual shared renderer Worker passed disabled and enabled cases. All 12 operation codes, original guest return addresses, copied input words after caller-memory overwrites, Apply source origins and exact device/generation/sequence correlation were checked. A preceding healthy draw snapshot still contains identity WORLD after the later zero-WORLD write; the failing snapshot contains zero WORLD. Both snapshots are deeply immutable. Strict singular failure remains identical.

Earlier generated-control attempts exposed two helper mistakes (returning code instead of the completed PE, and assuming a synthetic boolean lighting flag) and the CreateDevice trace-context error described above. Failed logs and correction notes are retained; none used original media. The final strengthened control passed with exit 0. No broken source was committed.

The genuine 37-file Conan installation and registry were rehashed in place. Manifest SHA-256: adf7c78e9ab56519edd53479b2151f04e18496668dec450ba2af12cc0b90cd38; registry SHA-256: 64969ff6f94d58b54431e18f937af3374bda5755e233ad62517840e1611d8de3. Media remains at scratch/game-media/conan-installed-20261105/.

Single original session

The sole changed-diagnostic original was started at 22:23:19Z after publication and all controls. It closed normally at 22:43:30Z. This was the only original session. The actual owning-Worker module was 2,174,894 bytes, SHA-256 d55b0003873cf0fe7c340c61d98e0bd21dc57190252822eba9acb7c02297111a, matching the public controls. Its unique bytes are retained in the run. The original used diagnostic commit ffe79b45f; the later passive guard below was not exercised by a second original.

Personally reviewed captures show logos, menu, Warrior selection, Kordava loading and the town/rooftop/street/temple cinematic. Escape did not skip it. No player movement or second player control response was reached. FPS and audio remain UNKNOWN.

first-failing-provenance.json and asserted causal-boundary.json identify device 136450064, generation 1, endpoint 1, queue 66517. The completed prefix is 66516. This is also the first lowered zero-upper-3x3 WORLD draw. Subsequent lowering through 66519 does not overwrite the retained failed snapshot.

The last live writer is SetTransform, operation 2, serial 112385; its before packet is operation 1, serial 112384. Guest instruction 0x41cd03 calls vtable slot 44; return address 0x41cd09, decoded-block EIP 0x41ccee, thunk 0x7502aa0, entry ESP 0x74ff0dc, guest thread 1, owning Worker slot 0. The source matrix is guest stack address 0x74ff2d0; destination program 0x6d71d4. The preceding WORLD has identity upper 3x3 and translation (48.4705047607, 256.5792236328, 4.7315182686), determinant 1.

The caller already supplies twelve +0 words followed by [1109070214, 1131938626, 1099106956, 1065353216]: translation (38.7671127319, 248.0127258301, 16.3801498413), homogeneous 1. Copied caller bits equal the before packet's source, after packet's destination, lowered WORLD and executor WORLD, exactly. No packet errors or incomplete packets were observed. Lighting and specular are enabled, normals are present; the failing draw is 12 triangles, stride 32. Strict singular failure remains intact.

No divergent emulator transformation is proved. This SetTransform copy is faithful to its input. The exact unresolved boundary is the upstream guest producer of stack matrix 0x74ff2d0. The bounded disassembly in guest-failing-world-writer.txt shows 0x41ccee loading that pointer from [esp+0xe4], then passing WORLD (0x100) at 0x41ccfd. It does not explain why the matrix was already zero before entry. The last writer is neither Apply nor reset. The retained 32-row tail contains SetTransform before/after only; this does not prove that Conan never used MultiplyTransform earlier.

Passive guard and handoff

Review found that the diagnostic's MultiplyTransform source lookup used g2w even though the existing success-noop handler never read that source. A focused negative control with unmapped pointer 0xf0000000 reproduces a changed quiet- miss sentinel even with diagnostics disabled (passive-negative-before.log). The guard replaces that lookup with a read-only mapped-page probe across all 64 bytes; it neither faults pages in nor requests lazy DIB materialization. Unavailable source packets are reported as two explicit collector errors when enabled. They are never fabricated from sentinel contents. This is a diagnostic passiveness repair, not a claimed cause of Conan's singular matrix.

Published passive guard: b487c7514 on main. The regression checks disabled and enabled cases, preserves the sentinel and existing success result, and checks explicit unavailable-source reporting. The public x86/owning-Worker controls additionally call MultiplyTransform with an unmapped pointer, preserve WORLD and continue through all mutation routes and both healthy/failing draws. Their actual module is 2,175,008 bytes, SHA-256 3d59cca0579d2ccfdec1561d7403c371af771943f6c8385ce814eacb008c6821. Final focused tests, public controls and the required locked full build are recorded separately from the original controls.

Recovery uses Git commits and passive-guard.binarydiff, not a source archive. diagnostic-source.json, source pins, exact module hashes, in-place media hash verification, command logs, screenshots and original terminal evidence remain in the run. owned-runtime-cleanup.json records independently checked absent PIDs/profiles, refused ports and free locks after all work. No test box was created, stopped, deleted or resumed; box ID/state/expiry are not applicable. The next investigation should trace the guest matrix producer upstream of 0x41ccee, preserving this exact SetTransform-to-executor boundary.