Legacy framebuffer paired blend factors and stored alpha

Task: D3D-LEGACY-TEXTURE-CAPS-AUDIT-20261013. Run: scratch/runs/20261105-legacy-framebuffer-blend/.

Source-only preparation

Preparation stayed source-only behind Aranna loader1840, AoW2 browser1850, then AoE3 native-system-info setup1900. Their exact FINAL runtime AND file RELEASE ALLBUILDSFINISHED records arrived at 19:27:34Z, 19:35:35Z, and 19:46:30Z. Before any test/build/browser or shared renderer edit, the gate independently verified 408 absent PIDs, 157 refused ports, 44 absent profiles, four free locks, no heavy process, and sufficient disk. Execution began at 19:48Z with a hard deadline of 20:32:50Z. No subworkers or boxes were used.

The preparation source SHA and relevant uncommitted diff are recorded in source-sha-preparation.txt and relevant-preparation.diff. Git is the source archive; no source-tree snapshot or media copy was made.

Primary contract

primary-contract.json records five selectively read Microsoft DX6.1 SDK topics from the retained CHM at scratch/runs/20261105-legacy-advertised-decal-blend/directx61.chm. Its SHA-256 is ade41d4cac0d000629ef5035827f4c6439dd360bc619b6832b33c53f27379c32. The original Microsoft-authored help was preserved by the preceding task in a third-party version-pinned SDK mirror; original distribution authenticity has not been established. Only five specific topics were read, without extracting or duplicating the SDK tree.

devdoc/live/directx/imref_15ic.htm defines D3DBLEND: BOTHSRCALPHA (12) is equivalent to separate SRCALPHA/INVSRCALPHA factors. BOTHINVSRCALPHA (13) uses inverse source alpha for the source and source alpha for the destination, overrides the destination blend selection, and is supported only as SRCBLEND. imref_5fdw.htm identifies SRCBLEND=19, DESTBLEND=20, ALPHABLENDENABLE=27, defaults ONE/ZERO, and places texture blending before framebuffer alpha blending. imover_8204.htm describes framebuffer source and destination blend pairs and alpha-test rejection. imover_68bo.htm states the weighted-sum equation. ddref_5n3r.htm defines DDPF_ALPHAPIXELS as alpha information interleaved with color and dwRGBAlphaBitMask as its alpha-channel mask. Thus a public 0x41 flag and 0xff000000 mask establish eight stored alpha bits; a 0x40 flag and zero mask do not. Enum identities alone are not the oracle.

Setter-to-draw route and causal baseline

Public Device3 SetRenderState routes to d3dim_set_render_state, which stores the raw factor unchanged; no 12/13 normalization occurs there. The software triangle path loads slots 332/336 into source/destination blend factors and forwards them to d3dim_blend_rgb when gl_sw_rgba_span is clear. That RGB helper and its channel helper lack 12/13 cases. lib/d3dim-gpu.js independently normalizes source factors 12/13 to the documented pairs. The separate native renderer's d3d_software_blend_pack is not this path.

The first rendered baseline records source d3ce00646d5872c7e0da88cb9450981d455c0364; subsequent baseline records may name its documentation-only descendant. The relevant renderer diff was empty until the first repair edit at 19:49:43Z. Every execution records its actual source and shipped-module hashes. The initial browser command found no Puppeteer-downloaded executable and launched no browser; all actual runs explicitly used the already installed /usr/bin/google-chrome.

All acquisition, state, Lock/Unlock, and DrawPrimitive/DrawIndexedPrimitive operations use real public x86 COM calls. The software, direct WebGL, and production GuestWorkerLink/RenderWorker routes consume both shipped modules; no excerpt module or substituted draw implementation supplies the proof.

With source RGB (51,153,204), destination RGB (180,45,90), and source alpha 85, the SDK equations require factor 12 to produce (137,81,128) and factor 13 to produce (94,117,166), regardless of the conflicting destination selection. Both modules' software routes instead produced (51,153,204). Direct WebGL and the owning queued Worker already passed these RGB cases. See baseline-rgb-actual/ and baseline-13/.

Explicit RGB targets publicly return flags 0x40 and alpha mask zero. Unused stored X bytes 17/231 are varied independently and are never interpreted as destination alpha. Factors 8/11 becoming zero on such a target are not used as bug evidence.

Separate A8R8G8B8 targets publicly return flags 0x41 and mask 0xff000000. On both modules and all routes, ONE/ZERO should preserve source alpha 85; software instead stored zero and both GPU routes stored 255. With source alpha 255 and destination alpha 85, source DESTALPHA (7) should yield (17,51,68), alpha 85; it instead yielded full source RGB. INVDESTALPHA (8) and SRCALPHASAT (11) should yield (34,102,136), with alpha 170 and 255 respectively; both instead yielded black. Each factor has its own unchanged module baseline in baseline-alpha/, baseline-destination/, baseline-factor8/, and baseline-factor11/.

Repair and regression

Software triangle draws now lower source factors 12/13 to the documented pairs at consumption time. Raw GetRenderState values and immutable queued snapshots remain unchanged. The opaque RGB helper and its fast paths remain in place. A8R8G8B8 framebuffer blending now uses actual destination alpha and writes output alpha through the existing RGBA arithmetic, retaining the legacy ZERO/SRCCOLOR RGB truncation path. GPU upload and readback preserve the alpha byte only for A8R8G8B8; opaque target handling is retained. No capability mask changed.

test/test-legacy-framebuffer-blend-web.js is the permanent regression. regression-checked/result.json records PASS for both shipped modules on software, direct WebGL, and the actual owning queued Worker:

The independent byte equations cover source alpha 0/85/170/255, conflicting destination selections, explicit ZERO/ONE and source-alpha pairs, disabled blending, GREATER alpha tests, indexed/non-indexed draws, and two stored destination-alpha values. Selected products divide exactly by 255, avoiding unspecified rounding. RGB and alpha expectations are computed separately from production blend helpers, descriptors, and shader code.

Focused indexed-texture/light-map/blend/format/depth, Device3/7 state aliases, lit-material alpha, GPU batching, depthless Z, asynchronous Flip, and the existing legacy DECAL/COPY/DECALALPHA/ADD public browser regression all pass. flock /tmp/wine-build.lock bash tools/build.sh passes in build-final.log. The tested module hashes, relevant binary diff, owned new file patches, captures, cleanup checks, and artifact hashes are retained in the run. Git supplies source recovery; no source archive or media copy was created. Representative baseline/repaired captures were personally reviewed; alpha correctness comes from actual channel readback, not the opaque preview.

This closes the demonstrated paired-factor and A8R8G8B8 framebuffer failures. The broader capability audit remains open. Other alpha target formats, unsupported destination uses of source-only factors, native Windows, physical GPU hardware, and game/backend parity are not qualified here. No game, held investigation, laptop Heroes II lane, or public deployment was run. Box ID/state/expiry are not applicable.