Public D3D7 ProcessVertices point/spot RGB lighting

2026-11-04. Evidence: scratch/runs/20261104-legacy-point-spot-public-lighting/. The public route repaired by 61802c92d8661f59a8cf0c90a5254a94b0d42e56 also computes the supported point/spot RGB cases below. No production failure was demonstrated and no production source was changed. The new regression is test/test-legacy-point-spot-public-lighting.js. It reuses the previous public x86 COM fixture mechanics, with independently authored literal RGB expectations.

The existing DX7 computed-output alpha authority gap remains open. All 72 positive output alpha bytes are retained but excluded from acceptance. This suite establishes neither a full RGBA contract nor a general rounding rule.

Retained primary rules

Read the previous note and retained DX7 diffuse-reflection, material-source enumeration and per-vertex-color pages. Read-only HTTPS retrieval followed their links to these original Microsoft DX7 SDK pages and equation images. primary/manifest.json records exact URLs, lengths and SHA-256 values; unaltered response bytes are retained. The equation and spotlight sign images were visually inspected. Network retrieval is independent of the offline run.

Primary DX7 page SHA-256
Light Attenuation Over Distance c89ebff8b754c2be162da9328273a615b6e92509ed3e7be1b1875e989435136a
Attenuation equation daf3b5290e64f096b3300ee3755d9b5c3b300d7fd92bffd8555d1c8752fb322c
Spotlight Falloff Model 9d3825745d1e24fb343f0f35d5e8fd2dce4eeac99c9edca358e9a849a438e5fe
Falloff equation 72b3ea7830b87d952bc2b981e2dcb9e534b12b3e0ae4f99d540a762a77d13ef1
D3DLIGHT7 1001fb12ec17af11e2f0321092e64487b0e95237b931279146eb4490886584a9
D3DLIGHTTYPE f696a686a9b3faecebc60ee909e4417bf5535aee53f194e4b741628f5f26c359

For point/spot lights, distance is d = |light.position - world.vertex|. When d > range, the light contributes nothing; equality is within range. Otherwise attenuation is 1 / (a0 + a1*d + a2*d*d). Directional lights ignore position, range and these coefficients. These cases use positive denominators and attenuation at most one, avoiding degenerate/clamping rules.

With L the spotlight's outward unit axis and D the unit vector from vertex to light, the SDK illustration explicitly gives rho = -L dot D. This sign resolves the text's loose description of the vectors. Let ct=cos(theta/2) and cp=cos(phi/2): spotlight intensity is zero at/below cp, one at/above ct, and ((rho-cp)/(ct-cp))^falloff between them. Both angles are total cone angles, supplied in radians. Distance attenuation multiplies this term.

RGB then follows the previously retained DX7 equation: Ia*Va + Ve + sum A*(max(0,N dot D)*Vd*Ld + Va*La). All normals here are unit vectors and the world/view transforms are identity. Global ambient and emissive are explicitly zero. Independent per-light diffuse and ambient channels make their attenuation observable separately. MATERIAL=0, COLOR1=1 and COLOR2=2 select material, vertex diffuse and vertex specular, respectively; COLORVERTEX and all selectors are explicitly set/read.

Cases and independent arithmetic

The twelve positive states run on each shipped module:

State Discriminating geometry/equation Expected RGB, vertex 0 / 1 / 2
Point constant distances 1,3,9; denominator 3 (67,67,0) / (67,67,0) / (67,67,0)
Point linear distances 1,3,9; denominator d (201,201,0) / (67,67,0) / (22,22,0)
Point constant + linear denominator 2+d (67,67,0) / (40,40,0) / (18,18,0)
Point quadratic distances 1,2,4; denominator 1+2d² (67,67,0) / (22,22,0) / (6,6,0)
Point range distances 1,3,9; range=3, denominator d (201,201,0) / (67,67,0) / black
Directional distant position, range=.25, coefficients 3,7,11 ignored (201,201,0) at all three
Point COLOR1 diffuse/COLOR2 ambient red/green/blue vs green/blue/red source colors (201,201,0) / black / black
Point COLOR2 diffuse/COLOR1 ambient swapped selectors, same distance attenuation black / (0,67,0) / (22,0,0)
Spot inner/penumbra/outside axis, 3-4-5 ray, outside ray (204,186,0) / (102,74,0) / black
Spot falloff squared penumbra intensity 1/4 instead of 1/2 (204,186,0) / (51,37,0) / black
Spot distance × cone denominator 3, independent green intensity .71 (68,60,0) / (34,24,0) / black
Spot range range=4 excludes all three distances black at all three

Point diffuse red and ambient green intensities are .79: their unattenuated byte-scale value is 201.45. Spot ambient red is .801 and diffuse green .73; the distance-times-cone case uses green .71. Inner/outer cosine thresholds are .9/.7. At (3,0,4), distance=5, rho=.8, diffuse cosine=.8 and cone intensity=.5 (or .25 for falloff=2). The outside vertex (4,0,2) has rho=1/sqrt(5), safely below .7. Cone angles are packed to f32 before SetLight.

audit.py independently reconstructs source bytes and the 104-byte D3DLIGHT7 layout, decodes source/destination FVF, and evaluates the RGB equation using Python Fraction arithmetic. It proves that truncation and nearest-integer conversion agree for every tested nonzero channel, with at least .04 byte margin from either interval edge. Float packing/trigonometric error is far smaller than this margin; the audit separately bounds packed-angle cosine error below 1e-7. No producer helper, producer output or intermediate producer color supplies the literal expectations. This does not claim every possible conversion method is equivalent or establish non-endpoint rounding semantics.

Projection uses w=2z, viewport (3,5,64,32), giving (35+16x/z, 21-8y/z, .5, 1/(2z)). The audit checks its exact f32 bytes. During fixture development an outside ray with nonbinary x/z exposed an arithmetic-order geometry discrepancy, and a combined spot green input exposed a truncation/nearest ambiguity. Inputs were revised before acceptance. fixture-development-quantization.json retains the latter input and mismatch; it is an oracle-development artifact, not a demonstrated producer failure.

Actual public calls and controls

Execution starts with actual x86 DirectDrawCreateEx/QueryInterface, creates a D3D7 device and vertex buffers, sets/reads matrices, viewport, material, light, enable and render states, then calls public ProcessVertices and Lock. No draw, GPU path or host-side replacement computes the actual output. The fixture supplies the exact shipped Math.cos/Math.pow host imports used by the Wasm stage; all lighting, attenuation and packing remain in Wasm.

Every input uses LIGHT|TRANSFORM (0x401), flags zero, source index 1, destination index 2 and count 3. The entire five-record source allocation remains unchanged, including 0xa5 guards. Six-record destinations keep 0xcd prefix/suffix records. Geometry is asserted byte-for-byte and only the three RGB bytes in each written color are accepted. Descriptor and Set-input immutability, guarded output pointers, capacities, HRESULTs, stdcall cleanup, caller RET, bounded completion and stack canaries are checked. Zero-count, source-end, destination-end and null-source calls check HRESULTs and complete destination immutability through fresh public Lock/Unlock.

Totals for positives: 24 allocations, 72 vertices, 216 RGB bytes, 2,808 checked destination bytes, 96 guard ProcessVertices calls and 1,124 stack-checked public calls. The independent audit verifies these counts.

Seventeen controls per module change one decisive state: point/directional type, position, range, constant/linear/quadratic coefficient, spot direction, theta, phi, falloff, spot/point type, diffuse/ambient material source, diffuse/ambient light intensity, enable, and spot range. All 34 controls successfully set/read state and complete ProcessVertices, retain the original RGB table, and fail it. RGB failure is deferred until after source, geometry, destination guards and Unlock/Release checks, so controls exercise those too.

Verification, identities and remaining boundary

Focused evidence run, ordinary no-evidence invocation, independent audit, previous supported-computed-lighting regression, D3D7 FVF lighting, ProcessVertices dimensional regression and ProcessVertices fog regression all pass. flock /tmp/wine-build.lock bash tools/build.sh passes. The test uses the ordinary default test tier and needs no evidence environment variable.

Starting recorded Git identity: de6a33f257d93e8814d666bdec62361161569ac4; validation head: d17c8d23c59e7efbcf7fd16226065758ab962d9a (peer docs only). Original and current modules are identical across the serialized build:

Existing path Original/current SHA-256
build/wine-assembly.wasm 4417305d70322467441aef51e888b0fe66451d9c60da0c734eeb23f5a22c0a6b
build/wine-assembly.compat.wasm 07fb4e75d05177271097dcd67d84e2629272b50073074bbdea8825ce10b4ed34
test/binaries/calc.exe cb71fcc9726f756421519a215c39cf89efd930c9ae0fa5bbb59b64050faef52b

identities.json, relevant binary diff, owned new-file patches, dependency hashes, primary manifests, test logs and SHA256SUMS retain recovery evidence. Git remains the source archive. No source-tree snapshots or module/media copies were made. No peer guest-worker, cache-observer or tools file was edited.

Remaining boundary: explicit primary DX7 computed-output alpha and its material-source interaction are still needed before a full RGBA oracle. Specular lighting, degenerate light configurations, arbitrary conversion, nonrigid normals, native Windows/GPU parity, 27-app coverage and gameplay remain unproved. No Chrome, game, boat allocation, subworker, deployment or runtime lifecycle operation was performed. Assigned box ID/state/expiry: NA. The coordinator retains ownership of future runtime work and shutdown.