DX7 computed diffuse alpha: exact unresolved authority boundary

2026-11-04. Run: scratch/runs/20261104-legacy-computed-alpha-contract/. This bounded offline investigation ends at the permitted authority boundary. The retained primary DX7 material-source rules are concrete, but they do not explicitly connect the selected source's alpha to the computed diffuse output alpha. No supported defect was demonstrated. No producer was changed and test/test-legacy-computed-alpha-contract.js was deliberately not created: there is no independently supported expected alpha to assert.

This is an unresolved contract result, not full RGBA coverage. The existing computed RGB and point/spot RGB results stand.

Retained authority and its limits

Read the two preceding notes, their tests, and the retained original Microsoft DX7 SDK pages served by the archival mirror. No network retrieval was made. primary-pins.json references 67 entries from the two preceding primary manifests and the earlier retained alpha-state investigation, with existing paths, exact URLs, byte lengths and SHA-256 hashes. Original response bytes remain at those existing paths; none were copied into this run. authority-excerpts.json records the relevant lines and source identities.

The central sources, under scratch/runs/20261104-legacy-supported-computed-lighting/primary/, are:

Primary DX7 source SHA-256 Established rule / remaining limit
D3DMATERIALCOLORSOURCE 23023ace4691e70611ef2b0b500b9cd94f1c2ed2d4eec838e19ec331377dbe58 MATERIAL=0, COLOR1=1, COLOR2=2 select current material, vertex diffuse, vertex specular. No computed-output alpha equation.
Render states 7bcb2f8393d9615e54150ac6c0fe2b79a5a849614a36de1eba3be51e08c97ff0 DIFFUSEMATERIALSOURCE applies only with COLORVERTEX=true. Default is COLOR1. LIGHTING enables/disables lighting. No resulting alpha provenance stated.
Per-vertex Color States 33134f6a7bb7e587d705e070c394be241c2ada05bab76830ed20164bd8044594 Vertex colors can supply material components for lighting. Does not define their alpha's path to the output.
Diffuse Reflection Model 597355b616c005d0c35e9aa39e77bdfab3d697bcf27dc0dd947f6bd899b8c7e9 Ambient/emissive/diffuse equation and selectors support RGB. No separate alpha equation.
Material Properties 4674407b624e9e40c4a90e5c013f6eb48bd97bd325da3b1a29d636d925d0680a Material properties carry RGBA blending information; emissive text also mentions transparency. Does not state which alpha survives lighting.
Color Values for Lights and Materials 134808ba0945a3823c2745f5acc31df258173dececbcf441db4961e2cb68c245 Lights do not use alpha. That fact does not determine the output alpha.

D3DMATERIAL7, D3DCOLORVALUE, D3DCOLOR, Processing Vertices, ProcessVertices, Alpha-Blending States and Alpha Texture Blending were also checked. Field definitions and later pixel blending do not supply the missing lighting-stage equation. The retained October 28 DX7 alpha title-search result lists texture, blend and test topics; it is not a complete SDK full-text search and does not prove that the needed document does not exist. No complete DX7 SDK documentation archive was located in the searched repository scratch paths. No archives were unpacked. DX6 rules, D3D9 rules and current producer behavior were not substituted for DX7 authority.

The precise missing proposition is the source and computation of the alpha byte written to a DIFFUSE destination by D3D7 ProcessVertices with TRANSFORM|LIGHT and LIGHTING=true. In particular:

Explicit state What remains unsupported
COLORVERTEX=false, each selector 0/1/2 Whether output alpha is material diffuse alpha, independently of selectors, and whether other material alpha contributes.
COLORVERTEX=true, DIFFUSEMATERIALSOURCE=MATERIAL Whether output alpha copies material diffuse alpha or follows another equation.
COLORVERTEX=true, DIFFUSEMATERIALSOURCE=COLOR1 Whether input diffuse alpha supplies output alpha.
COLORVERTEX=true, DIFFUSEMATERIALSOURCE=COLOR2 Whether input specular alpha supplies output diffuse alpha. This is distinct from output specular alpha's fog role.
LIGHTING=false or LIGHT omitted The corresponding bypass/copy behavior needed for state-changing alpha controls.

An explicit DX7 statement resolving these rows is needed before selecting expected RGBA bytes. If non-endpoint material alpha is used, float-to-byte conversion needs its own support or inputs whose permitted conversions agree. The exact missing document title/URL is unknown; the links above are inspected sources, not claims that an unretained page contains the answer.

Why the existing tests cannot close this boundary

The existing public tests execute actual x86 COM ProcessVertices and Lock in both shipped variants, with nonzero source/destination indices, immutable source allocations, destination sentinels, HRESULT/stdcall and stack guards. Their expected records explicitly exclude output alpha from acceptance. The independent audit.py here checks the input and output evidence directly: all 42 positive output allocations (126 vertices) use material diffuse alpha 1 and alpha 255 for every present COLOR1/COLOR2. All 126 observed output alpha bytes are 255. These inputs cannot distinguish material-only, selected-source or constant-opaque alpha hypotheses. The observed 255 values are not oracles.

test/test-d3d7-fvf-lighting.js does distinguish material and COLOR1 alpha in its helper-based checks. It inserts test exports and directly calls the lighting helper; its alpha expectations have no primary DX7 citation. It remains useful as a neighboring implementation regression, but neither its comments nor its successful run supplies missing primary authority. The current producer passes the selected diffuse color's alpha to its packer (d3dim_d7_light_vertices in src/09ab-handlers-d3dim-core.wat). That is an implementation observation only. It provides no reason to repair or certify the behavior. No peer guest-worker, cache or driver file was edited.

Verification and recovery

The following existing tests pass in this run:

The first dimensional invocation omitted its required LEGACY_PV_EVIDENCE variable and failed before execution. The corrected invocation passes; both logs are retained. These are neighboring regression results, not new alpha acceptance cases or native Windows/GPU parity.

Input main is d5f03571b1a6b35fb233b27f3159c2f05c72ff1d; entry HEAD was 7a9fac338882b23a6b9fde7b12cffda1fa0fa259. Validation HEAD is recorded in identities.json (peer documentation advanced main). Tested module identities, unchanged across the build:

Existing module path SHA-256
build/wine-assembly.wasm 4417305d70322467441aef51e888b0fe66451d9c60da0c734eeb23f5a22c0a6b
build/wine-assembly.compat.wasm 07fb4e75d05177271097dcd67d84e2629272b50073074bbdea8825ce10b4ed34

The run retains exact Git identities, relevant binary diff, this owned new document's patch, input path/hash pins, unique test logs/readbacks and a hash seal. Git is the source archive. No source snapshots, archives, module/media copies or repeated setup-only acceptance test were created. Existing calc.exe is referenced by path/hash in identities.json.

The task stops at this exact unresolved authority boundary. The next required input is primary DX7 computed-output-alpha authority, not another run of the same nondiscriminating inputs. Offline operation was preserved: no Chrome, game, boat API/allocation/lifecycle, subworker, deployment or approvals. Assigned box ID/state/expiry: NA. Future runtime and shutdown remain owned by the coordinator.