Legacy integer-center interpolation: offline causal boundary

No production repair was established. lib/d3d9-fixed.js and lib/d3d9-backend.js remain unchanged. The new offline diagnostic independently reproduces the first public expected color, exposes the emitted mapping's interpolation error, and retains a rejecting integer-center gate. It does not qualify rendered pixels, queued execution, or native D3D conformance.

Read predecessor note at e8647f27fa54b1e37afbb5faeaef8729099a7c89: docs/re-notes/legacy-public-mip-nonsolid-filter-20261104.md. Its original scratch/runs/20261104-legacy-public-mip-nonsolid-filter/browser/result.json is retained in place and checked by this diagnostic. All five actual/expected observations agree with the independently authored sampler. The fifth still fails at tolerance 2. Neither the public fixture nor another worker's helper was edited.

What the source establishes

The legacy lib/d3dim-gpu.js triangle submission builds POSITIONT attributes and fixed-function state, then submits them to the shared backend. This route uses Fixed.TLCache.compile/Fixed.compile, rather than draw.nativeFixed. In lib/d3d9-fixed.js, screen vertices receive (0.5-1/16) in both D3D screen axes before the GL y flip. Clip W is the reciprocal of input RHW. UV and diffuse outputs retain their vertex values. The transformed fixed vertex path has the same screen displacement. lib/d3d9-backend.js compiles this fixed vertex source unchanged; its mip fragment wrapper consumes ff_uv. The new test captures actual backend program assembly at createProgram, stopping before uploads or GL compilation. This is explicitly a source test, not a substitute for the queued public path.

The backend's separate draw.nativeFixed branch adds (0.875,-0.875)*W/extent after its shared clip IR, which encodes the same half-pixel-minus-1/16 screen shift. Guest programmable VS adaptation instead adds (1,-1)*W/extent. Neither branch was changed or runtime-tested here.

Subtracting the required half-pixel center conversion leaves a geometry translation of (-e,-e), with e=1/16. In an unsnapped analytic rasterizer, unchanged vertex attributes therefore evaluate at original screen position p+(e,e). This follows from translation invariance of barycentric coordinates; it is not a measured statement about GPU snapping or the actual queue.

Independent interpolation and coverage

For triangle vertices (x_i,y_i), unmodified edge equations give screen barycentrics lambda_i(p). Coverage uses their signs and exact top-left ties; interpolation uses their unbiased values. With input reciprocal W q_i:

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

The oracle takes the integer pixel position directly. Its barycentric method is cross-checked against an independent affine-plane solve. No geometry bias is inserted into the expected values. Affine color is checked with constant Q, avoiding an assumption about legacy color perspective policy. Smooth perspective color is separately checked as the emitted GLSL transport model. Device depth is affine in screen space, not another attribute divided by Q.

Hand-authored oblique triangle (1,1),(9,1),(5,9), at integer center (4,3): lambda=(1/2,1/4,1/4). For RHW (1,1/2,1/4) and UV (0,0),(2,0),(0,4), both U and V are 4/11. Color (0.2,0.6,1) with constant Q is 0.5; smooth perspective color with the stated RHW is 19/55. Depth (0.1,0.7,0.9) is 0.45.

Interior attribute Integer-center expected Emitted-shift analytic result
Perspective U 0.363636363636 0.373563218391
Affine color 0.5 0.5078125
Device depth 0.45 0.45859375

Both sample locations are inside the triangle. These errors are independent of shared-edge ownership. There are 69 interior sample positions on two oblique triangles, cross-checking both UV planes, perspective color, and depth transport. Fourteen shared-edge cases cover horizontal, vertical, both diagonal slopes, oblique seams, and reversed windings. Every analytic pair has exactly one owner.

The existing finite shift is not itself a general exact top-left rule. For triangle (0,0),(8,0),(0,8+1/32), (4,4) is strictly inside, but (4+1/16,4+1/16) is outside. For edge (0,0)->(8,8), the shift is tangential and leaves its edge equation exactly zero. These are analytic counterexamples, not claims of new observed GPU coverage failures. They rule out treating the old axis-aligned source test as sufficient coverage acceptance for a repair.

Retained public failure

The new sampler uses independently authored grid channels and explicit four tensor-product weights. It imports neither the public oracle nor production sampling code. At the public integer center (2,3), the MIN LINEAR fixture has U=0.387, V=0.568, LOD 1, and an 8x8 level. Its texel-space bilinear fractions are 0.596 and 0.044.

MIN LINEAR model RGB
Independent integer-center oracle 113.752,42.28,103.404992
Existing emitted translation, analytic only 120.752,49.78,106.974992
Original queued Worker WebGL1 readback 121,50,107

All five retained observations match the displaced analytic prediction within 0.6 channel units. This narrows the hypothesis to a concrete emitted-source interpolation defect and a matching explanation of the readback. It does not uniquely prove that defect caused the runtime failure: there is no controlled rendered repair experiment, and rasterizer quantization was not measured. The original public failure, tolerance, queued path and oracle are preserved.

Missing repair contract and validation

A defensible repair must decouple exact coverage ownership from attribute evaluation. Removing the shift would change the known cockpit tile behavior; correcting only texture UV would leave color/depth errors and potentially perspective LOD errors. An arbitrary oblique edge cannot be represented by one global translation that affects only ties.

An analytic correction, if retaining shifted geometry, must transport the affine numerator AND reciprocal-W plane before division:

u(p) = (N_u(p+e) - e.x*N_ux - e.y*N_uy)
     / (Q(p+e)   - e.x*Qx   - e.y*Qy)

Device depth needs its own affine-plane transport. Subtracting first-order derivatives of already divided UV is not exact; the focused test rejects that shortcut on perspective data. These calculations do not establish a portable shader implementation, clipping behavior, rasterizer snapping, WebGL1 extension availability, corrected LOD derivatives, depth ordering, or unchanged actual top-left ownership. Exact per-edge coverage and unbiased interpolation, with those constraints, remain the missing shared rasterization contract. No speculative production correction was made.

Commands executed:

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-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

The first command validates analytic controls and preserved evidence; it is not a passing integer-center acceptance test. The second deliberately exits 1 at the current lowering's perspective-U mismatch. The ordinary diagnostic explicitly characterizes the existing lowering; its bounded source recognizer and displacement controls need updating with a repair. Top-left policy, plan cache, tier membership and the serialized full build all pass. The passing diagnostic controls and build do not close the failing acceptance gate.

Run: scratch/runs/20261104-legacy-integer-center-interpolation/. Exact source SHA, relevant binary diff, owned new-file patches, first failing gate, results and hashes are retained there; Git is the source archive. No source archives, media copies, browser/game/large CLI runtime, box access, lifecycle actions, subworkers, deployment or successor task. No box is assigned; ID/state/expiry are not applicable. Coordinator retains all runtime ownership. Next required work is an explicit coverage/interpolation implementation contract and its rendered oblique shared-edge, perspective, depth and public queued regressions; the remaining public spatial matrix is still unexecuted.

Retired 2026-11-17

test/test-legacy-integer-center-interpolation.js was deleted. Everything it measured is gone: lib/d3d9-fixed.js (GPU stream v2 slice 9.2, 4b489cb1b), the legacy D3DIM GPU route (51ca79f5d) and lib/d3d9-backend.js (a6ee42aed). Its (0.5-1/16) two-axis displacement, the backend createProgram spy and the retained readbacks all described that lowering. Nothing in 09aj IR maps screen positions anymore: D3D9/D3DIM vertices reach the GPU stream as POSITIONT pixels, and lib/gpu-stream-webgl.js maps integer centres (+0.5 x, +Y_CENTRE y lift for top-left ties). That mapping and its tie rule are covered by test/test-gpu-stream-webgl-subpixel.js.