Forsaken: retained explosion atlas and target order, 2026-11-12

For GPU24138, source7608/1 only, the original SMLEXPL.PPM reproduces all 16 distinct retained native RGB565 texels, and smlexpl.off reproduces the four retained triangles' atlas rectangles. The immediately preceding opaque scene draw supplies the observed destination color. This closes the sampled asset-content and immediate target-order boundaries; the complete DIB/upload, live loader-to-destination copy, and exact GPU sampling remain unresolved. No runtime defect or production patch is proved.

Parent: guest state selection, 9bae6e1815a41455f66aecca813d5270e56b2140. Run: scratch/runs/20261112-forsaken-texture-target-provenance-1805/. This is read-only analysis of that parent's retained original evidence and installed media, with no new original execution or emitted observer change. The later requested-selector scene GPU24337, source7680/0, texture210 is excluded. The selected20011 call does not explain this earlier witness.

Asset bytes and original writer boundary

The witness binds texture handle244, actual entry131812992, DIB476602368, 256×128, pitch512, 16bpp format1. Its decoded cache generation is d3dim18; the submitted image has that same key. Among all 54 installed PPMs, 12 have these dimensions. Only test/binaries/candidates/forsaken-demo/installed/data/textures/SMLEXPL.PPM matches all16 coordinates under top-down RGB565 conversion. No vertically flipped candidate matches all16. Asset SHA256: 3decc7c3b982d147253021851673355d9a14a677b90c9e07081eb358c149515a.

Examples from the actual retained DIB/cache samples, with native bytes in little-endian order:

Coordinate Native bytes Decoded RGBA
92,26 ff ff 255,255,255,255
92,27 fa ff 255,255,214,255
23,14 5d ff 255,235,239,255
34,7 65 bb 189,109,41,255
35,7 03 8a 140,65,24,255
28,25 c3 20 33,24,24,255
29,25 00 00 0,0,0,255

analysis.json retains all16 original RGB triples, native bytes, decoded values and DIB-relative offsets. All20 sample entries, including repeated coordinates, agree with bit-replicated RGB565 decoding. This is a content match at retained coordinates, not a hash of all65,536 DIB bytes.

The original 564-byte data/offsets/smlexpl.off begins with one texture name, SMLEXPL.PPM, then count16. Its rectangle descriptors match the submitted UV bounds exactly: triangle13 uses rectangle4 (79/256,0)..(115/256,38/128); triangles17,18,21 use rectangle2 (21/256,0)..(46/256,26/128). Thus both sampled content and independent atlas coordinates identify the explosion asset. This does not imply that the selected sample or the full viewport should be white.

Original EXE SHA256 remains 366ec5d0943daca839836296f9fb52724e94dc155630c6b0c30d7181106b5e24. Read-only original disassembly identifies the P6 surface loader at 0x4cccb0, its Lock return0x4ccf19, and Unlock return0x4cd6e4. The16-bit path at0x4cd288 reads RGB bytes through table0x7f5480, divides by mask-derived quantization divisors, shifts the channels and stores a WORD at0x4cd3b0. It also avoids quantizing a nonblack texel to black. Gamma-table identity and the live filename are not recorded here.

The retained CPU tick30 has this real loader's Lock/Unlock for temporary surface243, this136456088, backing guest23003564: rows159047–159052, both returns0. Allocation row159053 creates destination surface244; row159054 frees243. These are allocation/free kind21/22 events, not Texture::Load receipts. The inherited CPU observer excludes the Load API and kind23 copy trace. Consequently the specific243→244 copy and its filename/bytes are not proved by this adjacency. The sample/atlas content match is stronger evidence than assuming an allocation-order join. All addresses above are historical identities discovered in retained records, not instructions to read those addresses in another owner.

Upload and sampling limits

Nine relevant current source files match final1735's SHA256 pins, including d3dim-gpu.js, d3dim-draw-trace.js, the D3D9 backend/fixed/shader files, gpu-backend.js and the three D3DIM/DirectDraw WAT fragments. The recorded source commit plus empty relevant binary patch reproduces those files.

In that pinned path, _textureLevel copies the decoded RGBA into the d3dim18 cache; the first use on a target supplies those pixels, and the backend's keyed texture path uploads/reuses that key. The retained sample records read the cache's decoded pixels and the observation-time native DIB. The submitted texture uses repeat U/V, linear min/mag and mip0, with no key, palette or palette alpha. All16 native/cache decode pairs match. This excludes a decode discrepancy at these coordinates; it does not directly read back the GL texture or prove its complete upload.

The observer's loop stops as soon as four covering triangles are found. The witness contains60 submitted triangles (180 owned vertices), but retains only triangles13,17,18,21. Their constant vertex RGB is64. Applying an ideal bilinear level-zero model at the retained UVs gives these source RGB contributions:

Triangle Ideal modulated RGB
13 64.000,64.000,63.448
17 64.000,63.838,63.332
18 45.672,27.247,9.599
21 28.798,18.866,12.241

Adding just these four to the observed destination predicts approximately 243.469,255,255, rather than observed255,255,255. This is not a proved rendering mismatch: later covering triangles may be absent because of the four-hit cap, visibility/depth is not recorded, and this arithmetic is not the actual GPU fragment/sampling oracle. Neither four hits nor the nominal60 triangles establishes how many fragments passed depth.

Ordered target contents at 320,240

The target entry131805408 alternates DIBs. The relevant current backing is 471527424, not the other buffer472166400. Retained GPU order gives:

GPU order Actual operation/source Color or bytes
24117 Earlier readback to471527424 native c5 83
24129 Natural upload of471527424, command7605 same c5 83 → 132,121,41
24130 Accepted clear, flags2, command7605 depth clear; color stays132,121,41
24134 Opaque scene source7607/0, texture entry131811904 132,121,41 → 0,12,8
24137 Opaque scene source7607/1, texture entry131812480 0,12,8 → 41,166,189
24138 Additive source7608/1, explosion texture244 41,166,189 → 255,255,255
24139 Completed ordinary readback to471527424 native ff ff

These actual before/after joins all agree. The immediate destination is the preceding opaque scene's output, not an already-white target or the other buffer's contents. This resolves the immediate prior-color ordering for this pixel; it does not prove every prior CPU write or frame-wide order. The clear wrapper records target metadata before _clear reconciles the target's backing, so its target field still names472166400 while its after pixel belongs to471527424. The nested upload24129 and subsequent actual draw targets identify the transition. That diagnostic field timing must not be interpreted as a cross-buffer color clear.

Verification and remaining witness

python3 scratch/runs/20261112-forsaken-texture-target-provenance-1805/analyze.py passes the asset scan, native/decode pairs, atlas rectangles, target history, actual loader rows, nine frozen-source pin checks and EXE identity. It hashes 70 referenced inputs without copying source or media. Small original disassembly slices and the exact source/owned-file patches are retained in the run. No build or production regression was needed for this docs-only finding; inherited observer controls were not rerun or promoted to fresh ones.

The next bounded observation should capture all covering triangles of one natural submission, its existing uploaded texture generation, and the unchanged depth/target context; then derive an actual sampling/fragment oracle. Separately the remaining loader edge is live filename/gamma/source surface → destination244 with a completed kind23 copy receipt. Any observer change still needs emitted passivity, return/error identity, actual CPU/GPU bootstrap controls and final freeze/import gates before allocation; reuse final1735 attach/private translator, passive Execute/selection, 19-field identity, ACK/readiness and prearmed independent release helpers. No blend override or guest write/call is justified by this result.

Actual /home free fell from1,518,219,264 to1,497,341,952 bytes during read-only work; no new native action or box creation followed or preceded that crossing in this task. No filesystem reserve change was applied. New box N/A, stateNOT_CREATED, expiryN/A, stop receiptN/A. The predecessor's archived box remains historical evidence; it was not resumed, stopped again or deleted. This task makes no new gameplay, input-causality, global-first-white, full-scene parity, FPS or audio claim.