Experimental solid transport reaches actual WebGL pixels

The released predecessor implementation now passes actual WebGL1/2 compiler, rendered coverage/interpolation and the unchanged public spatial mip oracle. This task deliberately reclaimed and preserved the uncommitted backend and focused test described in legacy-solid-coverage-transport-20261104.md. It publishes those implementation changes, including the predecessor's offline WebGL1 array-index correction, after rendering them. No additional runtime compiler failure arose in this task. solidCoverageTransport: true remains an explicit experimental Device constructor option; the default is unchanged. This is not a normal-path or full27 repair.

Concrete implementation and dependency fix

The backend submits one conservative invocation rectangle per original triangle, including indexed list/strip/fan parity. Immutable original inputs travel in uniforms; GPU vertex replay and homogeneous clipping produce conservative bounds and primitive-local coverage. Fragment edge ownership uses integer D3D sample locations, while original XYW planes independently reconstruct color, UV and device depth. Ordinary 2D mip derivatives use numerator/Q transport. Blend and depth tests remain GL operations after the final coverage discard.

The original missing strict interior cannot be recovered by fragment-only interpolation correction. This mechanism supplies those missing invocations without shifting input geometry. The first predecessor WebGL1 rejection was solid_scratch[count]: fragment arrays require constant or loop-symbol indices. solid_append now scans loop-indexed slots with a destination predicate. Both append sites use it. The original first-error log/shaders remain intact in scratch/runs/20261104-legacy-solid-coverage-transport/; the earlier standalone glslang success is not used as WebGL acceptance.

What actually ran

Evidence: scratch/runs/20261104-legacy-solid-shader-rendering/. The run driver, exact source seal, Git SHA, initial backend diff and new-test patch reconstruct the tested revision. The browser driver is run-specific; recovery uses its retained file and the predecessor's existing pinned runtime. Only two changed small source files and run-specific driver/seal were uploaded. The intact 37-file dependency closure, Puppeteer 24.31.0 actual import, static server import, both module hashes, served script bytes and HTTP isolation passed before Chrome. No source archive or media copy was made.

One browser ran 2026-11-04 05:13:56.950Z–05:14:10.059Z, Chrome 151.0.7922.108. There was no retry, reload, second browser or post-browser production edit. The backend SHA-256 tested and committed is d31cc9432b6c111946aeb5d2c43a8e72107f755cddc2e30cbedaa721d4bedd16. The later focused-test additions validate retained evidence; the rendered fixture function is unchanged from the tested revision.

CDP reports SwANGLE/SwiftShader; public direct-route observations retain an Intel/Mesa renderer string. This discrepancy is unresolved. These are actual WebGL compiler and pixel tests, not a physical-GPU portability claim. There is no application CPU rasterizer or fallback added by this change.

In that browser, the option-disabled control reproduced omitted strict interior at (4,4) for (0,0),(8,0),(0,8+1/32): RGBA 0/0/0/255. Its real queued Worker public MIN LINEAR draw reproduced 121/50/107 against the unchanged 113.752/42.28/103.404992 oracle, tolerance 2.

The candidate passed both actual WebGL versions:

There are 30 direct cases and 1,142 grid comparisons per version. Seam blending is ownership-sensitive; affine-depth occlusion is tested separately. Full seam ownership under all depth/stencil modes remains unverified. Captures were recovered and inspected, but direct screenshots occur after device destruction and show a cleared canvas; raw readback assertions are the coverage evidence.

After direct cases, both retained modules passed the actual queued Worker and direct public routes under WebGL1 and WebGL2: eight cells, 120 observations, three queued sampler transitions in each Worker cell, zero reported errors or fallbacks. MIN LINEAR now reads 114/42/103 in every cell against 113.752/42.28/103.404992, tolerance 2. The unmodified public fixture also checks magnification/minification separation, point/linear transitions, fractional/trilinear mips and perspective UV/mip gradients. No fixed LOD, UV-only adjustment, global geometry shift or weakened oracle is used.

Module files are referenced at existing build/ paths, with full SHA-256:

Offline checks and scope

The actual Device.draw transport regression independently rejects the original main backend (two submissions instead of four), then passes twelve candidate transport combinations and 48 submissions, preserving source bytes and inputs. The new --render-evidence=PATH arm rejects missing Worker cells, stale backend hashes, old witness pixels and reported fallbacks. It validates the two direct versions and eight public cells separately from submission mocks.

Focused transport/evidence, original-backend and evidence negative controls, retained interpolation diagnostic, top-left controls, plan cache, backend validation-only, fixed shader and position-type tests, test-tier membership, and flock /tmp/wine-build.lock bash tools/build.sh pass. The ordinary --require-unbiased-raster and --require-integer-center gates still exit 1; their failures are retained and not reclassified as passing.

Recheck retained rendering without launching a browser:

node test/test-legacy-solid-coverage-transport.js --render-evidence=scratch/runs/20261104-legacy-solid-shader-rendering/browser-result.json

Before any promotion, the following remain unresolved: full-domain clipping and snapping/exact-sign ownership; W=0/eye-plane degeneracy; nonlinear, projected and cube gradients; programmable mutable-global reset during vertex replay; user clip distances and hardware uniform limits; culling/two-sided stencil, alpha/discard and guest depth-output interactions; transformed/programmed rendered coverage beyond these fixtures; hardware portability; and performance. The current per-primitive draw and repeated fragment vertex replay are expensive and unqualified for normal gameplay. Narrow successful tests do not resolve these limits. No game or full27 run occurred.

The assigned box bx_e5cc38rs was independently verified RUNNING by actual exec, API idle/healthy/hydrated, expiry 2026-11-04T05:40:37.336Z. Browser/server closed normally; /proc inspection found no remaining Chrome/Chromium/Node task process. Unique results and ten captures were recovered before 05:30Z. The coordinator retains runtime/shutdown ownership; this worker performed no allocation, resume, stop/delete, subworker launch or public deployment.