Conan palette producer: observation repair and remaining runtime gap

Run: scratch/runs/20261106-conan-palette-producer/.

This investigation does not identify a causal emulator defect or establish player-controlled gameplay. It repairs the run-owned observation path, but the single original-game launch was disqualified and stopped before any palette producer rows arrived. No production CPU, RPC, renderer, or guest bytes changed.

Why the original launch is disqualified

The first locked build and byte-identical baseline modules used a6d8048f5887c2e9b74b94021e1df98e5c54df88. A concurrent pull advanced source to 277b92b72175d9a60cd835baadd034f9309195d2 before diagnostic compilation. The diagnostic module therefore gained the source-declared d3dim_gpu_cooperative_enable export. Initially the worker incorrectly treated this as diagnostic compiler retention and removed only that export entry. That adjustment did not establish source equivalence. The coordinator's 02:09 source-drift warning on the messageboard was missed before launch.

The one original launched at 02:24:24 UTC and was closed at 02:25:06 immediately after the error was recognized. There were no gameplay inputs, palette-update observations, or retries. Its startup observation has 16,729 real decoder-map arrivals, zero producer rows, and zero observer errors. Its actual page, owning CPU Worker, and renderer Worker module identities were recovered and verified; this establishes which diagnostic bytes ran, not equivalence to the earlier baseline. original-disqualification.json records the exclusion.

The personally inspected startup.png shows a black viewport during startup and loading logs. It provides no evidence of movement, combat, or any other visible player response. The existing candidate remains unqualified for ordinary gameplay.

The invalid comparison's modules, contracts, source patches, and control results are preserved under source-drift-controls/. Earlier rejected attribution attempts are preserved under failed-control-1/ through failed-control-4/.

Run-owned passive observer

The previous kind-31 diagnostic packets did not cross the production RPC forwarder. The replacement uses supported kind 30 with distinct tag 0x43504d50 and complete, ordered decoder records. It does not broaden lib/guest-rpc.js. Production owningWorker receives the actual decoder emissions; the generated controls do not inject maps or set CPU registers through live exports.

Decoder records initially describe staging addresses. Passive metadata follows the cache's existing memory-copy publication to their execution addresses. The host retains bounded whole records, records eviction ordinals and absent keys, and resets only the corresponding thread slot on arena rewind.

Store attribution also needs the original instruction record for fused x87 execution. A rejected block-executor hypothesis produced no origin-alias activity and was removed. The observed palette store instead uses the short x87 pipeline: the final FSTP record survives even when the first handler record is rewritten. The observer copies that final record before the store. It also has scalar and island attribution and a tree-store metadata hook. The tree family has no separate generated attribution case in this run.

Passive guest reads accept only the direct image window or already committed sparse pages, with explicit bounds and cross-page handling. They never invoke g2w, a pager, or guest fault handling. An unmapped object control must remain unavailable without causing a guest exception. Actual reserved-page guest reads are tested separately through the real SEH/pager path.

The intended original capture follows dynamic owner pointers and index 72 at 0x431e30, copies cache flags, parent indices, four table pairs, decomposition words, channel records, and palette stores, and retains the uncached update when a later cached call occurs. It does not assume the previous allocator's owner/base addresses remain valid. Guest disassembly is read-only evidence.

Validation and source recovery

After stopping the original, the worker rebuilt current main under flock /tmp/wine-build.lock bash tools/build.sh. The replacement module recipe requires identical canonical closure hashes for baseline and diagnostic and exact imports/exports without export normalization. Baselines must equal both shipped production modules byte for byte.

The generated public CPU/COM controls cover scalar stores, scalar FSTP, REP, fused x87 islands, and the lazy palette updater, including a matrix crossing committed sparse pages. They exercise all twelve public WORLD mutation routes, canonical producer/executor state, the normal strict singular-transform error, and a real guest child thread. Equivalence compares eight GPRs, EFLAGS, FNSTENV, FNSAVE, guest thread identity, and numerical results. A bounded 128-record map is churned with more than 512 ordinary guest blocks to test real eviction.

The independent native x87 oracle has 24 passing multiply/FCOMP controls across three precision-control words, including signed zero and unordered comparison. The generated SEH controls compare baseline/diagnostic, standard/compat, eager/ lazy, and committed/reserved reads. Reserved reads must invoke the actual guest handler exactly once, commit/fill the page through public Win32 calls, and resume with 0x12345678; committed reads must not invoke the handler.

The final same-closure tests passed: 20 producer cases and 16 SEH/pager cases, ending at 02:27:43 and 02:28:03 UTC respectively. Both modules use canonical closure SHA-256 dba4d3b013cc1f620d76cf7ae984c28d428c88c4e8b239894f6dddcd72b215b4. The earlier cross-source PASS results are not substitutes for these tests.

Module SHA-256
Standard production/baseline 881a4ffc816bb2a752febff01551ade99dfc5680dc8a3a6a4222f0bb8d2c1951
Compat production/baseline 721dea351a4ac638cc275b9fa30101be050b0836eb6d44de0230e9edb38e2079
Standard diagnostic 920e3c0d8266602e244235f223c56ec58bd97d5abe08067dd9e8d8423bc2255b
Compat diagnostic a8cd4394934a90f51e63c5aba54c5d4253c4dbfba747367fad49b3b3d7e9bb0b

Each final diagnostic palette case receives 1,150 decoder records and 1,611 relocations while evicting 1,756 records. The first actual fused palette FSTP is attributed to 0x430e1d; the second sparse-page store is 0x430e24. Independent generated inputs produce binary32 2 and 4 respectively. The uncached flag is 0 on the first call and 1 on the second. These are generated fixture facts, not measurements of original Conan inputs.

To reproduce the controls, run the current locked build, then prepare-modules.cjs baseline standard, baseline compat, diagnostic standard, and diagnostic compat, followed sequentially by producer-controls.cjs and pager-controls.cjs. The run's original immutable deadline guards must be renewed only under a new authorized runtime window; the existing one-shot original guard must not be removed for a retry.

Git is the source archive. source-pin.py records exact Git SHA, relevant binary diffs, owned helper new-file patches, and hashes. The run preserves unique tested modules and logs without source archives or media copies. Original media remain at scratch/game-media/conan-installed-20261105/; verify-inputs.py checked all 37 existing inputs and registry pins in place. The EXE SHA-256 is c343ef6a51a88c5b01a3c3e1511035359d25336fbe010123bd2e02962b77df56.

Remaining producer question

The prior matrix-write investigation proved the 64-byte REP copy and preceding first product components. Those are not new mismatch evidence. This run captured no original 0x431e30 producer inputs, so it cannot determine whether a cache decision, channel selection, x87 operation, or earlier input first differs from expected behavior.

The next bounded original investigation needs the validated same-source observer and copied actual inputs at that lazy updater. For each candidate discrepancy, compare the exact operation against documented x86/Win32 semantics and an independent result. The run-owned analyze-palette.py can compare captured scale multiplies and channel FCOMP decisions with the native oracle, but it was not run on invented or generated data as evidence about original Conan. Zero scale or skipped channels alone would not prove an emulator defect.

No causal repair is justified by the evidence in this run. No further original launch is held by this worker.

Runtime release

Ordered releases were point-a1 discriminator 0025, then Aranna recovery 0100. Independent cleanup opened the gate at 02:03:02.647705 UTC, with fixed deadline 02:48:02.647705. All runtime work ended before that deadline. Final independent cleanup confirmed 321 owned PIDs absent, 18 ports refusing connections, 18 profiles absent, all four runtime/build locks free, no heavy process, and 5,932,699,648 free bytes on /home. owned-runtime-cleanup.json is the receipt. No boxes were assigned or used; box ID, state, and expiry are not applicable.