Conan: nonsingular producer result becomes zero before WORLD submission

NEW-GAME-CONAN-DEMO-20261010 remains short of verified player-controlled gameplay. One changed original session narrowed the first unresolved producer interval to guest 0x42a42e → 0x42a5a6, on the same CPU thread, stack frame, node and matrix. The matrix is nonsingular after the node method returns and has twelve zero words at the second boundary. No causal emulator mismatch was isolated, so this change publishes findings only. It does not diagnose a game bug or repair the renderer by accepting singular state.

Evidence is scratch/runs/20261105-conan-upstream-world-producer/ (ignored). verified-boundary.json records the passing offline assertions; first-producer-boundary.json contains the selected raw words, registers, stack, x87 state and decoded values. The preceding boundary and its passive diagnostic commits are documented in the WORLD-writer note.

Exact observed boundary

All following rows are copied from the actual owning guest Worker: endpoint 1, slot 0, guest thread 1. The matrix address is 0x74ff2d0. Every selected row has zero unavailable reads. At both interval endpoints ESP is 0x74ff1ec and the renderer caller's node is 0x88496f0.

Diagnostic serial Guest boundary Observed matrix
32195 0x42a41c, before node-method call Destination scratch contents; not a producer result
32196 0x4ab9f0, composer entry Destination still scratch; input node is 0x884ae70
32197 0x42a42e, node-method return Rotation/translation, determinant 0.9999999826119712
32198 0x42a5a6 First twelve words are +0; determinant 0
32199 0x42a5bf Same zero words
32200 0x42a65f Same zero words
32201 0x41c990, consumer entry Same zero words; return address 0x42a71c
32202 0x41ccee, WORLD-writer block Same zero words

The nonsingular result at serial 32197 is approximately:

 0.79863548  0.60181504  0  0
-0.60181504  0.79863548 -0  0
-0           0           1  0
38.76711273 248.01272583 16.38014984 1

The translation and homogeneous last word remain bit-for-bit unchanged when its first twelve words become zero. The final words are exactly:

0 0 0 0 0 0 0 0 0 0 0 0 1109070214 1131938626 1099106956 1065353216

The forwarded composer input at serial 32196 has scale (1,1,1) and finite quaternion (0,0,-0.31730467081069946,0.948323667049408). Its return address is 0x492d32, after the call in node method 0x492cf0. This method can forward to another node; the renderer node and composer node differ here. The measured nonsingular return rules out blaming this composer invocation's output for the subsequent zero matrix. A later first-zero-at-composer-entry packet observes reused destination scratch, not a returned composer result.

The consumer's actual return 0x42a71c confirms the indirect call at 0x42a716. Its fifth argument is the matrix pointer: [entry ESP+0x14], later [consumer block ESP+0xe4]. Its tenth argument identifies the renderer node: [entry ESP+0x28], later [consumer block ESP+0xf8]. These offsets and copied registers/stack establish the join; two uses of a stack address alone would not.

Strict failure and the remaining gap

The original first fails with D3D9 fixed function: singular normal transform at queue 72140, completed prefix 72139, device 136450064, generation 1, endpoint 1. Its first-zero WORLD snapshot is also queue 72140. The last live writer is SetTransform (operation 2), block 0x41ccee, return 0x41cd09, source guest 0x74ff2d0. Its source words, stored WORLD words, lowered snapshot and executor words all match the words above. The previous run failed at queue 66517; this run reproduces the same matrix contents and strict error with a different preceding command count.

The measured interval contains multiple calls. In particular, 0x42a5a6 is the return from 0x42a5a1 → 0x41cfd0, but there is no observation at that callee's entry. Calls earlier in the interval include 0x401000, 0x43a110, indirect node/renderer methods and conditionally 0x458c10/0x412e80. Their actual branch sequence and the first store that changes the matrix are not captured. It would therefore overstate this result to attribute the mutation to 0x41cfd0 alone.

Read-only disassembly identifies candidate matrix-producing paths inside 0x41cfd0: 0x431bc0, table-derived matrix pointers through 0x4130a0/0x431ff0, products through 0x414f60, and REP MOVSD copies back into the supplied matrix. Those paths and their operand matrices are not dynamically distinguished. The explicit inverse-helper call 0x42a5ba → 0x415450 is after serial 32198's zero observation, so that call site cannot be the first zero-producing step. MultiplyTransform remains the old success-noop; this observation does not establish it as the cause.

The next bounded observation should locate the first write to this 64-byte matrix in 0x42a42e–0x42a5a6, preserving the store's PC, operands and active call context. If the write is a product or copy, capture its input matrices and their producer before testing documented x86 semantics. No second original session or live instrumentation change was attempted here.

Passivity controls and limits

Before the original, a run-owned compiler transform added copied-value packets at selected delivered block boundaries. It never calls guest setters, evaluates lazy flags, changes the debug mode, plants breakpoints, materializes DIB memory, or calls the faulting g2w/gl32 readers. Reads use the direct affine window or already-PRESENT sparse PTEs and return an explicit unavailable word on rejected reads. Raw x87 physical words/top/control/status/tag are copied without popping or evaluating the stack. Guest disassembly is evidence only.

An independently generated ordinary x86 PE ran through the production owning CPU Worker and shared WebGL Worker with both baseline and diagnostic modules. Its ordinary FLD/FADDP/FSTP result and a CMP/JNE branch survive observation; unmapped object observation does not cause a guest fault. Existing public COM controls additionally preserve copied setter inputs, state-block origins, Reset identity, all twelve writer routes, queue correlation, the old Multiply no-op observation, and the original strict lit-singular draw error. Final producer-controls.json is PASS. The first un-stamped control pass is preserved in controls-unstamped/; the final stamped controls passed before launch.

This proves the tested functional passivity cases, not universal timing invariance or complete CPU correctness. The original delivered 40,185 complete snapshots with zero packet errors or incomplete packets. The 1,000,000-snapshot cap was not reached. The host keeps 128 recent rows, 64 zero events with 16 preceding rows, and four writer rows per node for at most 1,024 nodes. First zero at each watched PC is saved immediately. Interior fused instructions and discarded history remain unobserved. The 1,745 captured completed initial scale-read nodes had no all-zero scale; that is not an exhaustive inventory of node constructors or later writes.

For semantics, the Microsoft SetTransform contract identifies the supplied matrix as the transform input. The Intel architecture manuals are the reference for x87 arithmetic, stack operations and stores. No observed instruction here was shown to violate those contracts. A mathematical quaternion cross-check in the offline helper is explicitly not a bit-exact instruction oracle.

Recovery, ordinary input and cleanup

Compilation was pinned to Git 25cd5f11331d880ea27277b8b2f0427fdf1bcee2; launch was pinned to 71a0ca4706f6a0dc0724eb6014569985f0c6bcfb. The intervening change is coordinator documentation. identity-verification.json verifies 301 relevant source files unchanged, 18 served JavaScript script identities, and the diagnostic module in three actual page/guest/render contexts. Relevant tracked binary diffs are empty; the pre-existing untracked lib/wininit-boot.js is recorded as a new-file patch without editing it. diagnostic.source.patch records the run-owned instrumentation hunks, and helper identities are recorded separately. No source archive was made.

Module Bytes SHA-256
Baseline, byte-identical to shipped production 2175008 3d59cca0579d2ccfdec1561d7403c371af771943f6c8385ce814eacb008c6821
Diagnostic actually executed 2176480 1c6a57797a3e29d5f6d6837d66ea72a8b160cc9620c494abe62eaa66ddfe5525

The existing installation at scratch/game-media/conan-installed-20261105/ was hashed in place: all 37 files match its manifest, manifest SHA-256 adf7c78e9ab56519edd53479b2151f04e18496668dec450ba2af12cc0b90cd38, registry SHA-256 64969ff6f94d58b54431e18f937af3374bda5755e233ad62517840e1611d8de3. No game media was copied.

The ordered predecessor releases and independent cleanup passed at 23:33:40.622874 UTC, establishing a fixed deadline of 00:18:40.622874 UTC. The single original launched at 23:36:20 and closed normally at 23:48:59, exit 0. Captures were inspected before each ordinary key action: Escape to the main menu, Enter to Difficulty, an initially unaccepted short Enter, then an accepted longer Enter to Kordava loading. kordava-loading.png is still the Difficulty screen despite its filename; warrior-confirmed.png shows actual loading. Later captures show the town/street/temple cinematic. Escape did not establish a skip to player control. No player movement plus second response was verified. FPS and audio remain unknown.

No shared production code changed, so there is no causal fix or code regression commit and no code-change full-build requirement to claim. The two run modules were compiled under /tmp/wine-build.lock; generated production controls and offline boundary/identity assertions passed. Final owned-process, port, profile, lock and disk checks are recorded in owned-runtime-cleanup.json. No boxes, public deployment or Heroes II were used.