AoW II: creation captured, earliest producer detail lost

2026-11-06, worker aow2-earliest-dib-writer0545. Run: scratch/runs/20261106-aow2-earliest-dib-writer/ (ignored evidence).

Black terrain remains. No emulator divergence was demonstrated, so no emulator, guest, input, or media implementation was changed. This run establishes the terrain surface's creation descriptor, successful return, backing identity and observer-local generation. It does not establish the intended first drawing contribution: a host collection failure lost the detailed CPU events at terrain startup. That is the first remaining evidence failure, rather than evidence of missing geometry.

Gates and tested identities

The exact ordered FINAL releases were Aranna0500, Conan0510, then KKnD0535. Independent checks after all three found 221 absent PIDs, 20 refused ports, 14 absent profiles, four free locks, no heavy process, and sufficient disk. release-gate.json fixes the active window at 06:38:19–07:23:19 UTC. Preparation before this gate stayed inside this run. No boxes or subworkers were assigned or used.

Tested source SHA: 6c48fd87416794f12059e8fc9290ae784c760c92 on main. The 301-file relevant source manifest was identical before controls, before the original and after it; relevant binary diffs and separate new-file patches record concurrent uncommitted source, without a source archive. The untracked lib/wininit-boot.js patch is recorded as shared-tree context, not this worker's change. See *-source.json, *-relevant.binarydiff, and *.new.patch.

Both shipped modules remained byte-identical through the locked full build:

Module Bytes SHA-256
wine-assembly.wasm 2178481 41d3ed6a2b52b46d25550dd54f582c574a7acc48f2024c3e01f878bdafbe470f
wine-assembly.compat.wasm 2181127 2c94cba0b6138e51d2d69ed19f938f2cb6abd7b83cd3cc0363ce232119cbaf7a

The original HTTP response was the standard module above, status 200, with matching retained response bytes. Controls used both modules. Actual source responses and local diagnostic transformations are hashed in http-served.jsonl, controls-http-served.jsonl, served-transforms.jsonl, responses.json and producer-controls.json. These record served identities; they are not an independent CDP extraction of instantiated WASM bytecode. identity-guard.json verifies source, module and six runtime helper hashes.

The 1,064 existing AoW fixture paths, sizes and hashes matched before and after; no media was copied. The executable is test/binaries/win98-games-a-d/Age of Wonders2 demo-SW/extracted/AoW2.exe, SHA-256 a10590e5dbd013d154b00ea53e66670f4e74d38ab33adb2523f0a662af8f89f7, preferred image base 0x400000. Full path/hash lists are fixtures-before.json and fixtures-after.json.

Passive observation and controls

Only run-local JS was added. surface-observer.js consumes existing API entry, API exit and dx_trace hooks, preserving original hook arguments, receiver and return. It reads current RegionMap fields and the existing pure JS guest-pointer translator, including sparse mappings. It adds no guest calls, setters, uploads, fences or readbacks. Creation scans are bounded to four MB; retained API details, births and points have explicit quotas. Generation is a local counter of observed allocation events, not a newly invented guest field. API completion records unchanged dispatch EIP, no yield and the known stdcall stack adjustment.

Current source symbols were located by name, including $handle_IDirectDraw_CreateSurface in src/09a8-handlers-directx.wat:2637, $handle_IDirectDrawSurface_BltFast at 5960, GetDC at 6644 and Lock at 6841. The existing raw prepacking trace is at 14463. Native fields come from the generated RegionMap rather than an old note's movable addresses. The observer can retain the first raw Device3/7 primitive source bytes at dispatch entry, before FVF packing, without inferring stride, coverage or an intended texel.

flock /tmp/wine-build.lock bash tools/build.sh passed. focused.sh passed the run-local observer controls plus existing draw-trace, GPU batching, GPU points, fallback-counter and worker-parity tests.

The final public COM controls passed all four standard/compat × disabled/armed cases in actual owning CPU/GPU Workers: 27 draws each, identical GPU pixels and counters, 159 CPU and 57 GPU events when armed, no observer errors/drops. Every recorded hook's before/after architectural state was unchanged; architectural entry/exit records also matched across armed/disabled cases. This included nonzero GPRs, unresolved lazy flags, live x87, the complete FPU/XMM files, and eight MMX getter values. Genuine public VirtualAlloc supplied a sparse stack, cross-page DDSD and sparse output. Public creation checked normal return, backing, generation and zero initialization; other controls checked a known raw RGB565 write through Lock/Unlock, production GDI GetDC/SetPixel/ReleaseDC, Blt, upload, covering draw and Clear. See producer-controls.json and helper hashes there.

Two earlier generated fixture failures are retained: the fixture initially lacked production gdi_surface_create imports; next it changed an API selector without updating that thunk's API-name string. Both fixture issues were corrected before the passing controls. No emulator implementation was changed to make these pass. attempt1-*, attempt2-* and their command logs preserve the failures.

What the sole original establishes

One original ran 06:43:21–06:55:07 UTC, after the controls and another independent resource cleanup. Chrome 151.0.7922.108, local port 45305, debugger 46441, cross-origin isolated, main CPU Worker slot 0 with child Workers, legacy WebGL render endpoint 1. The observed renderer was Intel ANGLE Mesa UHD620; requested SwiftShader command-line flags do not justify calling this a SwiftShader run.

boundary-findings.json is the corrected concise analysis. The earlier boundary-analysis.json remains available but its naturalReadbacks filter incorrectly searches for natural-readback instead of actual target-readback; use the corrected findings for that field.

The persistent creation return at CPU observer order 451, invocation 964080:

Field Observed value
Requested DDSD 124 bytes, flags 0x1007, width 1024, height 525, pitch 0, lpSurface 0
Requested caps / pixel format 0x6040; RGB565, flags 0x40, 16 bpp, masks f800/07e0/001f
Return HRESULT 0, completed, ESP 122682432 → 122682452 exactly +20, EIP 4397449 unchanged, yield 0, return PC 4397461
Output / object pointer 165260888, returned COM object 136454664
Surface slot 65, local generation 1, type 2, flags 0, native pitch 2048, format 1
Backing native RT 131807264, native DIB 484888576
Initial storage all 1,075,200 logical bytes zero; selected point (600,300) bytes [0,0]

The requested descriptor is unchanged on return. This non-alias allocation's zero storage follows the explicit $zero_memory call at current src/09a8-handlers-directx.wat:2854, before COM object creation. That is ordinary allocation initialization, not proof of a failed intended drawing operation.

Cumulative successful completed results report 13 Device7 DrawPrimitive calls targeting slot 65, followed by one SetRenderTarget call whose entry target was 65. Their detailed entries were lost as described below. No conclusion about their source data, packing, coverage, selected-point contribution or correctness can be drawn from HRESULT alone. There is no retained target-65 Lock/GetDC/GDI writer detail establishing an earlier producer; the missing batch prevents a blanket absence claim.

On GPU endpoint 1, order 2, command sequence 1 ordinal 0, the first target upload names exactly RT/DIB above, rows [0,525], native sample [0,0], changing RGBA [0,0,0,0] to [0,0,0,255]. The natural existing readback at endpoint order 15, sequence 8, route legacy-fence, writes the same DIB over [0,116,1024,525] and still samples [0,0]. Existing CPU DX observations then name source 65 and destination 18 at BltFast coordinates (0,0), with source sample zero. CPU observer, existing DX observer and GPU endpoint use separate sequence spaces; these receipts do not supply an invented total cross-worker event order.

Exact remaining failure and recovery

There are 42 retained collection segments, nine persistent birth records, 613 retained CPU events, 23,239 existing DX events and 156 GPU events. Observer errors and reported new CPU/GPU drops are zero. This is not complete collection: CPU order jumps from 311 to 808, losing 496 events, orders 312–807. Persistent births and cumulative call/result summaries survive that gap. Eight legacy DX segments separately report 7,903 dropped events, including startup segments.

faults.json records a detached guest Worker evaluation failure. The collector drains CPU Workers sequentially, then waits for all later Workers and GPU before writing a segment. If a later Worker is detached, already-drained earlier CPU rows are discarded. This run therefore did not preserve the first terrain producer's raw API/prepacking records, even though its persistent creation and summary counters remain. Missing detail must not become a missing-geometry claim.

Recovery requires a fresh assigned run, not another original here:

  1. Persist each individual Worker drain immediately, with its slot and target; handle detached Workers independently and retain the error. Collect GPU independently so its failure cannot discard CPU receipts.
  2. Add a focused host collector control that deliberately detaches a later Worker after an earlier drain and proves those earlier rows survive.
  3. Re-run both module public COM armed/disabled pixel and architecture controls, current identity/media checks, ordered release gates and locked full build.
  4. Re-resolve current symbols and capture the exact creation-to-first-upload lifetime, first 13 Device7 draws before packing, source textures/rectangles and an actual writable-pointer/raw-store interval if needed. Determine which intended source should affect the chosen point before proposing a repair.

No documented graphics/Win32/x86 semantic difference was established here. No late triangle-only probe or second original was attempted.

Reviewed play and cleanup qualifications

reviewed-captures.json names personally inspected images and hashes. Ordinary balanced 180 ms mouse presses progressed Scenario → Single → Start, dismissed the dialog, selected the party, plotted and confirmed movement. party-base.png shows six selected portraits and 20/20. The coordinator independently read confirmed-arrival.png as 12/20. Later final-world.png and the valid native party-counters-final.png show five portraits, a green-5 party southeast of the city and 10/20. This preserves evidence of commanded movement without claiming all six units moved or explaining the changed selection. Black ground persists.

After confirmed arrival, a screenshot helper accidentally applied Puppeteer's default 800×600 viewport. Its first two panel clips were invalid. No game input was sent during that interval. The final helper restored the recorded 1280×813 viewport before the final valid panel and world captures. Failed helper versions, geometry receipts and invalid captures are retained, not presented as evidence of a game defect. No game restart occurred. No FPS or audible-output claim.

terminal.json records ordinary Chrome closure with exit 0 and removed profile. The surrounding tool supervisor ended 143 without writing its final exit receipt; the Node driver's exit code is therefore unknown. supervisor-audit.json preserves that distinction and observes all recorded original PIDs absent. Final independent PID/port/profile/lock/disk checks are in final-cleanup.json. No box exists to hand back (ID/state/expiry all not applicable). The final artifact-hashes.json verifies retained evidence; result.json is written last.