Conan: first singular-normal draw retained, 2026-11-05

Task NEW-GAME-CONAN-DEMO-20261010. Evidence: scratch/runs/20261105-conan-first-draw/. Predecessor: player-control investigation, commit 5d27fd1e0.

The diagnostic repair is published as b3ebcdbb92b2870883c4f531171145ab8cab240a. The sole changed original session retained the first failing draw before retirement. Its WORLD matrix has twelve positive-zero float words followed by a translation and 1. Normals are supplied and lighting is active. The original producer of that matrix is still unknown; no causal rendering repair or player-control acceptance is established. The strict singular-transform exception remains unchanged.

Passive executor evidence

lib/d3d9-render-worker.js catches the exact singular-normal exception at the actual WebGL device.draw executor, copies bounded state into a deeply frozen diagnostic, and rethrows the original error. The physical worker supplies its owning endpoint ID. lib/d3d-command-stream.js transports the diagnostic with the original failure and makes a private, deeply frozen, nonwritable received copy before queue retirement.

The retained data comprises immutable command identity; world/view/projection values and float bits; vertex declaration and normal usage; lighting, material-source, fog, normalization and texture-generation flags; light count/types; and the executor stack. Bounds are 16 attributes, 8 stages, 8 light types and 8192 stack characters. No vertex, texture, successful-command history or full stream is retained. The actual original failure diagnostic occupies 3679 bytes as compact JSON.

test/test-d3d9-normal-diagnostic-web.js runs the production manager/receiver and actual WebGL executor in a physical owning Worker. Generated controls additionally wrap the actual adapter in that Worker to inspect the pretransport error and mutate the private failed command afterward. Twenty-four assertions prove exact identity, freshness against preceding state and caller mutation, pretransport and received immutability, bounded data, payload retirement, sticky strict failure, isolation of a healthy device, and a newly created generation-2 namespace. This is not a claim that a reset recovers the failed generation. The baseline reached the real exception and failed for absent evidence; the changed controls pass.

Ordered gate and sole original

AoW2callback1945 FINAL 20:34:39, AoE3wow64setup1950 FINAL 20:43:09, then Aranna-passive-resource2005 FINAL 20:54:13 were observed in order. Independent gate checks at 20:55:14.671993Z found 496 absent PIDs, 190 refused ports, 61 absent profiles, four free locks, no heavy runtime and 7,323,598,848 free bytes. The execution deadline was 21:40:14.671993Z. Before that gate only source reading and small run-owned inputs were prepared.

The changed original ran once, 20:58:34.310Z–21:10:58.797Z, using the predecessor app-config/browser recipe, defaultViewport:null, actual guestWorkers and owning WebGL Worker. Chrome and driver both exited 0 through ordinary quit. Commands were issued once after captures were inspected; pending commands were never replayed. No second original, extra guest, box, subworker or public deployment occurred.

The existing 37-file install was verified in place and reused unchanged at scratch/game-media/conan-installed-20261105/:

source-before.json records the exact source commit and runtime file hashes; runtime-relevant.binarydiff and runtime-untracked-source.json record launch-tree state. executor-identity.json retains actual owning-worker script hashes and the unique observed wasm module, SHA-256 b6c36db2aa538aa8c6e411d1cbe421d7898f791d65003b5ed54b8564495289d2. The loading guest Worker had already closed when final identity was read; main guest and owning renderer identity were retained. Identity observation set no breakpoint and made no guest writes.

Exact first failing draw

first-failing-draw.json, after-temple-state.json, first-draw-analysis.json and console.log retain the original evidence.

Field Observed value
First error 2026-11-05T21:09:53.208Z
Endpoint / device / generation 1 / 136450064 / 1
Failed sequence / successful prefix 62083 / 62082
Submitted at snapshot / in flight 62086 / 0
Error / opcode EXECUTION singular normal transform / DRAW 5
Geometry TRIANGLELIST 4, 12 triangles, stride 32
Declaration POSITION FLOAT3 offset 0; TEXCOORD0 FLOAT2 offset 12; NORMAL FLOAT3 offset 20
Guest shaders / native fixed no vertex shader, no pixel shader, nativeFixed false
Lighting lighting 1, specular 1, two point lights
Other normal flags normalizeNormals 0, localViewer 1, colorVertex 1
Material sources diffuse/ambient/specular/emissive all 0
Texture generation none; stage indices 0–5, transform flags all 0
Exact determinant world 0; view -1; projection approximately 0.10000998824715648

WORLD, in captured flat float order:

0 0 0 0
0 0 0 0
0 0 0 0
38.767112731933594 248.01272583007812 16.380149841308594 1

All first twelve uint32 words are exactly 0x00000000; the last four words are 1109070214, 1131938626, 1099106956, 1065353216. Full view/projection bits are in the retained diagnostic. Exact rational determinant checks used the captured float values, not an epsilon heuristic.

The owning stack is normalMatrix (lib/d3d9-fixed.js:18) from the lighting branch of compile (line 145), through TLCache.compile, Device.draw, adapter execute, receiver drain and physical-worker control. This rules out the proposed unused normal / texgen-only explanation for this draw. Ordinary repeated sticky error logs continue until quit; they are not additional captured failing draws.

Contract, producer review and next discriminator

Microsoft documents normals as transformed by the inverse transpose of the complete world-view matrix, with normalization afterward if enabled: camera-space transformations. Its lighting equations use normals for diffuse and specular terms. D3DXMatrixInverse can fail inversion. These contracts provide no basis here to substitute a matrix, suppress the exception or discard the draw. They do not establish native rasterizer behavior for this singular, lit input.

Source review found d3d9_transform_index maps VIEW 2 to slot 0, PROJECTION 3 to slot 1 and WORLD 256 to slot 10. d3d9_transform copies exactly 64 bytes through the guest mapping into the canonical slot or recorded state block. The D3D8 setter routes through that implementation. The draw bridge copies view/projection/world from the matching slots using .slice() in lib/d3d9-host.js. The existing generated state-block test passes default identity, independent slots, immutable recorded copies, last-write semantics and Capture/Apply. This supports the mapping, but it does not reconstruct the original setter or prove the original guest computed the zero matrix correctly. No evidence says the original called MultiplyTransform.

The next discriminator is a bounded last-writer trace on the owning guest producer for WORLD 256: retain SetTransform versus state-block Apply/MultiplyTransform, device and state version, caller return address/EIP and the raw 64-byte guest source; compare the canonical slot-10 bytes and draw-bridge copy to the executor bits at the first singular DRAW. This distinguishes guest-supplied zero scale from corruption before or during state production. That requires a separately authorized future original session; this task did not run a second session.

Visual review and validation

Reviewed captures show logos, New Game, Warrior difficulty, Kordava loading, town and story cinematics, rooftop/street/eastern-edge/temple views, and then a black scene retaining the temple subtitle when the queue failed. One ordinary Escape changed the cinematic view. It supplies no movement/control evidence. Both final black captures were reviewed. No changed player position followed by a second control response was established. FPS and audio are unknown.

Focused checks passed: the 24 owning-WebGL-Worker diagnostic controls; test-d3d9-fixed-plan-cache.js, test-d3d9-fixed-cascade.js, test-d3d9-worker-survives-draw-failure.js, test-d3d9-shared-render-worker.js, test-d3d-neutral-command-stream.js, and test-d3d9-stateblocks.js. flock /tmp/wine-build.lock bash tools/build.sh passed before the diagnostic commit and original launch. No rendering semantics changed afterward. The early cleanup probe during that build correctly reported the build lock occupied; a subsequent independent cleanup passed after the build and again after the original session.

Final independent cleanup, source comparison, media-pin verification, artifact hashes and recovery instructions live in the run directory. Owned process/port/ profile cleanup is checked independently of ordinary browser shutdown. No assigned boxes exist (ID/state/expiry: not applicable). Git is the source archive: evidence contains exact commits, relevant binarydiffs and owned new-file patches, unique logs/captures/module, and no source archive or copied game media.

Conan has no entry in test/candidate-corpus/manifest.json; per the capture workflow, evidence remains attached to this investigation rather than an unrelated dashboard candidate. node ops/check-run-evidence.js exits 0 with zero containment errors; its historical missing/invalid records are retained separately from this run.