Legacy texture-operation capability correction — 2026-10-23

Demonstrated defect on exact input 5406777ecbcef9cf9b03b40d592d722bb0086de0: d3dim_fill_device_desc7 advertised 0x003FF at offset 180, claiming ten texture operations. Both legacy routes implement SELECTARG1, SELECTARG2 and MODULATE, while the other advertised operations silently modulate. Correct the mask to 0x0000E. This changes one production constant and its explanation; no new combiner operations or other capability fields are implemented.

The Microsoft-owned legacy d3dcaps.h defines these capability bits in operation order. Its D3DPRIMCAPS and D3DDEVICEDESC7 declarations establish the 236-byte structure, texture-operation DWORD at 180, and stage-limit WORDs at 184/186. Microsoft's operation documentation defines the corresponding behavior; the legacy extended-capability documentation confirms that each capability asserts operation support. Checked 2026-10-23. The older sized D3DDEVICEDESC writer has no texture-operation store and is unchanged. This is not the D3DPRIMCAPS texture-blend mask.

DISABLE was checked separately. Setting COLOROP=1 does not unbind the texture: the actual setter stores it, d3dim_bound_texture_entry still resolves it, both software outer triangle dispatches reach the textured raster sink with operation 1 unchanged, and the actual GPU descriptor/lowering retains an active texture and emits MODULATE. The software inner combiner also modulates. Thus the DISABLE bit is removed on direct outer-path evidence. An alpha-only DISABLE with color enabled is undefined in the documented API; no support claim is inferred from that case.

The distinguishing operands are texture 0xA0C08040 and diffuse 0x8060A0E0. All seven removed operations produce 0x50485038 (red 72), with the GPU lowering emitting operation 4 for both color and alpha:

The retained bits are SELECTARG1=0x2, SELECTARG2=0x4, MODULATE=0x8. Nine independent RGB/alpha combinations execute the actual production combiner and actual legacy GPU lowering, checking both operations and TEXTURE/DIFFUSE arguments. Reversed operands, white identity and zero inputs also execute the software combiner. Both software outer dispatches are exercised. The GPU draw uses the actual WAT descriptor and JS lowering; allocation, texture upload and submission are explicit fixture doubles. Software raster submission is an explicit recording sink, not a substituted combiner.

Run from this checkout:

node --max-old-space-size=256 test/test-legacy-texture-operation-caps.js --negative-controls
node --max-old-space-size=256 test/test-d3dim-gpu-batching.js
node --max-old-space-size=256 test/test-d3dim-gpu-points.js

All three pass. The exact baseline runs the same distinguishing behavior and fails specifically at the executed capability writer assertion. Three capability mutations (restore DISABLE, restore MODULATE2X, drop MODULATE), a software SELECTARG1 operand mutation and a GPU SELECTARG lowering mutation are rejected. The complete 236-byte baseline/candidate descriptors differ only at bytes 180/181, with masks 1023/14; all 232 bytes outside that DWORD are identical.

Evidence is adjacent in legacy-texture-operation-caps-20261023/: baseline-results.json, candidate-results.json, descriptor-comparison.json, both *-excerpt-manifest.json files and seals.json. The manifests bind every compiler path to exact Git blob, full-file SHA-256 and size, and each production form to byte offsets, line ranges, byte count and hash. The extraction recipe keeps forms verbatim and wraps them with explicit fixture imports/globals and exports. WATX consumes top-level forms. The final assembled source is 32,939 bytes, SHA-256 f05616f67b30f52efec42aeeeb17f66278af5a59790d9e3045cada42d3c4d48a; each compiled variant is 4,380 bytes with one 64 KiB memory page. No full-source compilation or 512 MiB native guest occurs. Compiler identity and byte-exact excerpt checks pass before acceptance compilation.

The first focused manifest measured 16 files/1,330,408 bytes. Adding the actual fill_primcaps dependency required a revised measured selection of 17 files/ 2,388,036 bytes before that additional checkout. The complete closure includes the GPU's transitive local imports and all four authentic compiler stages. The shared clone was --no-checkout, with explicit non-cone file patterns installed before checkout. The donor and installed tools remain standing, read-only inputs. Installed Node is v24.18.1 at /home/user/.nvm/versions/node/v24.18.1/bin/node, 123,656,816 bytes, SHA-256 f3432a45b03b2da0d270095fdd8813dc34cbea73f5fc8b18c7a384b7cf9b333a.

Resource evidence includes first-admission.json, complete-admission.json, expanded-admission.json, resource-before-publication.json, prelaunch-baseline.json, direct-capture-allowance.json, and the executable guard/recovery recipes. The pre-publication projection is 36,334,951 of 67,108,864 NEW bytes, including an 8 MiB private peak, at least 16 MiB future external growth and 8 MiB reserve; projected host free is 4,479,762,246 bytes. Standing SQLite is excluded, while every sampled positive SQLite/WAL/SHM growth is charged cumulatively without shrink credit. Task outputs/session events, Git/index, copied evidence, guard, and an 8 KiB conservative baseline/board allowance are charged. The final live receipt remains in /tmp/legacy-texture-operation-caps-oct23/final-resource.json after publication.

All substantive commands retain streaming capture, per-file 24 MiB, process ownership and cleanup checks. Node uses a 256 MiB heap, 60-second CPU/wall and fresh MemAvailable >=805,306,368; Python uses a 60-second wall. The actual 32-byte capture rejection control terminated its child, left the sentinel absent and verified owned-group cleanup. Direct instruction/source/web reads and receipt reprints have a conservative cumulative 327,680-byte charge, in addition to actual guarded child output; the combined limit is 524,288 bytes. No substantive resource-denial retry occurred. The explicit 2 MB selection estimate failed before the extra checkout; its replacement was an exact measured closure within unchanged resource limits. Routine fixture corrections included missing import signatures, WATX API/result shape, a baseline excerpt path and mutation-receipt preservation. execution.jsonl records all actual guarded argv/exits and bounded output; unguarded-bootstrap-attempts.json records the initial missing Python alias and inspection-helper name collision. The evidence snapshot has 64 guarded commands, 12 nonzero exits (including intentional rejection controls/baselines and corrected fixture failures), all with verified cleanup. Later validation/publication commands are retained in the live execution log rather than recursively copying that log into itself.

Qualification is limited to source and small fixtures: actual capability writer, combiner, selected outer dispatch, descriptor and legacy GPU lowering. This is not full COM, native guest, real WebGL pixel, gameplay, or full-build acceptance. No runtime access, package installation/download, public deployment, shared integration, main merge, worker launch or donor/root tracked-file/index/ref mutation was performed. Original native qualification checkout, fixture09dc, memory watcher, Shadow Man approval hold and runtime403 holds remain preserved. No box is assigned (ID/state/expiry N/A); coordinator retains lifecycle ownership.

Publication is only task/legacy-texture-operation-caps-oct23, using strict trusted SSH and no force. Final commit/actual remote SHA are recorded in /tmp/legacy-texture-operation-caps-oct23/publication.json and the root append-only messageboard. No integration is requested by this worker; coordinator decides independent acceptance and integration.