AoWII earlier CPU source: bounded original observation
Run: scratch/runs/20261117-aow2-ground-cpu-writer-visible-repair/.
No production repair: valid application black versus an API/raster/coverage
defect remains UNKNOWN. Visible terrain acceptance remains unmet.
The predecessor 5b00e9af644843f75b41474593300bb6b43eb1fc established actual
process1000 webgl1 D3DIM endpoint1/gen1 and upload403538 target9: native
[600,300] was already RGBA(0,0,0,255). Its later1239 draw bounds excluded that
point through natural readback405841. This run begins upstream of that boundary.
Offline offline.cjs/offline-map.json checks the retained source bands and
actual upload call sites in src/09abg-d3dim-gps.wat, SHA256
c7e648501924975352bde60afee7d503d415fe366a2d6699a8b83fe72fa316ab, equal
to pinned base0a823095. Rows0..512 and512..525 are opaque; respectively
480845/524288 and12085/13312 pixels are black. The render-target identity
excludes texture/palette branches but does not distinguish trimmed prepare,
CPU target rectangle upload, or parked-primary merging. Stream records alone
do not bind the earlier DIB caller.
One fresh local original reused module
87ef39648a61ed6f54ca4ab8de1926de02c651eb2d99655015dc7e5de5407009, the
predecessor sealed media/module/source inputs, and its strict adapter. Exact
media hashes and requested paths are in original/requested-media-pins.jsonl;
actual worker modules are in original/actual-workers.json. No media copies,
new backend inference, forced guest application calls, or forced pumps.
Only main CPU worker slot0 received an import wrapper, before instantiation,
around existing dx_trace/gpu_stream_kick. Observer SHA256
0637bb6df5704f6105beb2af34ce63902234a69c20bb23284d409a401fe181c0;
original/observer-transform.json binds original/transformed worker hashes.
It preserves original arguments, receiver, return and real exceptions. Copies
are bounded snapshots; observer errors stop observation, not execution.
Maximum384 events,4096bytes/read,128vertices/retained raw call; no memory stores.
Model controls and a small real Wasm import bridge pass ownership, copying,
memory invariance, disposal and error controls. Initial qualification maximum
2.90us/hook; strengthened whole-memory control maximum6.47us/hook (final rerun1.82us/hook). These are
observer controls, not a native graphics API counterexample.
original/terrain-lifetime-cpu.json binds the fresh natural chain:
- Allocation slot65/gen1 requests1024x525 RGB565, pitch2048, masksf800/07e0/001f. The allocation hook precedes initialization; it does not establish creation completion or the last writer.
- At ordinary Device7 DrawPrimitive return PC0x4459fe, target3 maps surface65 DIB485216256. Native[600,300] bytes are00 00. The application supplies primitive5, FVF0x1c4, vertex pointer0xbc3c600, count4. EIP0x4459dd is a block location; the actual call is at0x4459fb.
- Natural BltFast at return PC0x431bfb copies source65 rectangle [0,0,1024,525] to backbuffer18 at[0,0], flags0x10. Destination DIB480124928 sample is42 08 before the copy, source sample00 00. Subsequent natural kick binds target9 to backbuffer18 with sample00 00. The11 event reports1075200 copied source bytes. Both surfaces are RGB565; zero converts to opaque black.
Observation ends after that transfer and two natural kicks:16 retained events, 21602 copied bytes,12367 hook calls,217.07ms measured observer time. It does not identify the predecessor's exact late upload403538 occurrence.
Concrete blocker: the existing kind15 raw-input hook is in Device3, whereas
this original calls Device7. Thus the observer retained the actual caller and
source pointer but no raw vertex bytes, texture inputs or effective coverage.
No individual earlier natural writer, missing coverage, wrong format, or
intended color is proved. The field named surface.flags in observer JSON
reads object+8 (misc0), not actual surface flags at+28; no conclusion uses it.
An attempted offline Wasm function-name mapper found that its older source
manifest does not match this module; names remain UNKNOWN. No second original
or unchanged UNKNOWN retry was performed.
Whole original/world-ready.png visibly retains black terrain. A party click
does not prove army selection or movement. Prior audio/thread/two-move20→14→8
and nativeQuit0 evidence remains referenced from
20261117-aow2-required-dll-audio-gameplay-progress/result.json; it is not
relabelled fresh affected-original verification. Historical missing keeper
receipts remain UNKNOWN.
Fresh ordinary Escape/No returned to the menu, then Quit/Next/Exit reached
both outro screens and hid the game window (native-ready.png). The2500ms
exit check still saw running=true and no Program-exited marker, failed, and
triggered local shutdown. Fresh ExitProcess(0) is therefore UNVERIFIED.
The independent finite watchdog was prearmed with actual /home free floor
1500000000B and no RAM gate. custody/supervisor-wait.json records controller
failure, driver1, Chrome SIGTERM, watchdog0, actual child waits/EOF, adopted
waits, no remaining scope/children. parent-wait.json records supervisor1
with actual wait/EOF. Outer helper session returned143 and failed to write
keeper-wait.json: that receipt remains missing. Unique custody files were
retrieved with hash readbacks before own /dev/shm profile removal. No boxes:
IDnull, stateNOT_CREATED, expirynull, stop receipt not applicable.
No production code changed, so no repair/native pixel counterexample or canonical code build is claimed. Next useful evidence would bind the actual Device7 raw inputs and effective state/coverage before its first terrain draw, with corrected prequalification. That requires a separately authorized run.