Legacy unbiased raster repair: first unsupported dependency

No production repair was implemented or accepted. The first unresolved dependency is a solid-triangle coverage mechanism that generates the required fragment invocations and carries primitive-local edge information. The current solid backend submits original indexed vertices directly to GL. Correcting attributes inside its fragment shader cannot restore an invocation that geometry omitted. This is a concrete boundary in the existing implementation, not a claim that a portable replacement cannot be written.

Predecessors: integer-center diagnostic at 930f2cc11d95f498dc40daff39d8f224e5ccd73b, and nonsolid public filtering failure at e8647f27fa54b1e37afbb5faeaef8729099a7c89. The independent public oracle, tolerance 2, queued path and five retained observations are unchanged. MIN LINEAR still expects 113.752/42.28/103.404992, with retained queued Worker WebGL1 readback 121/50/107. No fixed LOD or fallback bypass was added.

Derived separation and the missing implementation

For a consistently oriented triangle, write each unmodified edge as E_i(x,y) = A_i*x + B_i*y + C_i. Coverage accepts positive values and accepts zero only for a top/left edge. If an agreed integer coordinate representation makes every edge evaluation integral, this comparison can equivalently use E_i >= 0 for inclusive edges and E_i - 1 >= 0 for exclusive edges. That one-unit adjustment belongs only in the coverage predicate. It must never be used in interpolation weights, and it is not a displacement of the vertices. The representation must follow the actual clipping/snapping contract; choosing a convenient grid in a test does not establish that contract.

Unbiased screen barycentrics independently define the attribute planes:

Q = sum(lambda_i * RHW_i)
N_u = sum(lambda_i * RHW_i * u_i)
u = N_u / Q
du/dx = (N_ux * Q - N_u * Qx) / Q^2
z_device = sum(lambda_i * z_device_i)

V and smooth perspective color require their own numerators and the same Q. Affine color uses its own affine plane. Device depth is affine, rather than another numerator divided by Q. Any reconstruction must reach the fragment combiner and the mip footprint calculation before sampling, with the gradients at the corrected position. Exact numerator/Q transport is analytically valid; first-order correction of divided UV and reuse of displaced gradients are not.

The strict-interior witness (0,0),(8,0),(0,8+1/32) covers integer center (4,4) with all barycentric weights positive. The existing displacement evaluates coverage at (4+1/16,4+1/16), which lies outside. Even exact rational attribute reconstruction cannot run at the omitted fragment. For the shared positive diagonal (0,0)->(8,8), the same displacement is tangential. It cannot choose which complementary hardware edge policy owns a tie.

The ES2 specification guarantees a single owner for matching shared edges but does not select D3D's top-left owner; see section 3.5.1 of the OpenGL ES 2.0 specification. The test uses an opposite complementary tie policy as a portability counterexample, not as a newly measured GPU policy. WebGL1 derivatives and fragment depth writes also require the optional OES_standard_derivatives and EXT_frag_depth extensions. Their presence alone would not solve missing coverage.

The complete solid submission spy reaches Device.draw's final gpu.draw for WebGL1/2, POSITIONT/transformed input and list/strip/fan topology: twelve cases, two submissions each. It observes original vertex bytes and indices, only the original three input attributes, no conservative solid expansion, and no fragment edge/depth reconstruction. Each case reuses one actual backend program-cache entry. Ordinary fixed draws request neither optional extension. The earlier assembly capture separately retains the actual mip wrapper. These are backend execution/source-boundary checks, not GLSL compilation or rendered pixels.

Reusing the existing outline expansion is not an implemented alternative: Device.draw gates it to non-solid modes and WebGL2, and its coverage is for points/lines. A solid replacement needs its own coverage and primitive-data transport, handling clipping, generated edges, snapping, culling, depth/stencil and unbiased attribute/LOD planes. Merely removing the shift delegates ties to GL; retaining it and changing UV leaves coverage, color and depth defects. No such partial candidate was installed. Clipping/snapping behavior remains unchanged because production remains byte-for-byte unchanged.

Tests and retained rejecting gates

The extended diagnostic preserves the 69 selected interpolation samples and 14 original edge cases. On inspection those 69 samples comprise 66 strict interiors and 3 boundary samples; the predecessor's blanket description as interiors was too broad. Reversing triangle winding with corresponding vertex attributes adds 69 rational UV/color and affine-depth transport checks. The horizontal/vertical fixtures were corrected to share both seam endpoints.

A second independently written coverage implementation uses BigInt edge coefficients on an authored 1/32 fixture lattice. It agrees with the original edge predicate on 490 paired grid cases covering both diagonals, horizontal, vertical and oblique seams, both windings. Rectangle interiors have no holes and adjacent triangles have no double ownership. Fourteen diagonal samples show ownership-sensitive blend and depth differences under the legal opposite tie policy. These are analytic blend/depth controls, not rendered tests.

test/test-d3d9-fixed-top-left-bias.js still characterizes the retained axis displacement, but now explicitly rejects promoting it to a general exact top-left guarantee. Its strict-interior and tangential-diagonal controls pass. No production source comment was treated as evidence of portability.

node test/test-legacy-integer-center-interpolation.js --evidence=scratch/runs/20261104-legacy-public-mip-nonsolid-filter/browser/result.json
node test/test-legacy-integer-center-interpolation.js --require-unbiased-raster
node test/test-legacy-integer-center-interpolation.js --require-integer-center
node test/test-d3d9-fixed-top-left-bias.js
node test/test-d3d9-fixed-plan-cache.js
node tools/test-tiers.js --check
flock /tmp/wine-build.lock bash tools/build.sh

Diagnostics, axis controls, plan cache, tiers and the locked full build PASS. --require-unbiased-raster deliberately exits 1 with MISSING_SOLID_COVERAGE_TRANSPORT; --require-integer-center deliberately exits 1 at perspective U 0.3735632183908046 versus 4/11. Both failures are retained. The bounded current-source recognizer must be extended for a future repair; passing these diagnostics alone would never qualify that repair.

Evidence and runtime handback

Run: scratch/runs/20261104-legacy-unbiased-raster-repair/. Recovery uses its exact Git SHA, relevant.patch, owned new-note patch and file hashes. Git is the source archive. Relevant production and public-fixture diffs are empty. No source tree/archive or media was copied into evidence.

The locked build reproduces the retained original modules exactly:

There is no changed candidate module identity. The predecessor browser modules remain in place; no duplicate module recovery was needed. No browser was allocated or launched, so no staged HTTP/bootstrap validation, controlled old/candidate rendered comparison or remaining public cell was executed. The first unsupported dependency was established offline, before staging. There is no rendered repair, full27 or physical-GPU claim.

Assigned runtime bx_e5cc38rs was independently checked APIidle, healthok, hydratedtrue and actual exec exit 0, then ACKed on the shared messageboard. It is handed back RUNNING, expiry 2026-11-04T05:40:37.336Z; shutdown belongs to the coordinator. No persistent helper was started. No lifecycle action, other-box access, local Chrome/game runtime, subworker, successor, deployment or Heroes II work occurred. Next required implementation is the solid coverage/primitive transport contract, followed by rendered oblique shared-edge, perspective color/depth and canonical queued Worker WebGL1 regression before the remaining matrix.