GL evaluator family, 2026-11-03
This change implements the 23 remaining evaluator exports with evaluated geometry and pixels. It does not qualify original MDK2 gameplay; 24 names in the observed mandatory lookup inventory remain unsupported.
Input: a2ae4283da7db2687a6e3d853a0103f6678ed451 on shared main.
Evidence: scratch/runs/20261103-gl-evaluator-family/.
The primary reference is the OpenGL 1.1 specification, 4 March 1997, section 5.1, section 5.4, section 6.1 and table 6.17. The implementation provides the required minimum maximum order of eight, stores the original double coefficients and domains, and evaluates tensor Bernstein polynomials in double precision. Analytic partial derivatives, including the homogeneous quotient rule, produce normalized automatic surface normals. Float entry points convert once on input; the native ABI consumes each double's two x86 stack words.
All eighteen map targets have private per-context state. Grid endpoints are selected exactly at integer zero and the specified grid division count. Evaluation uses the highest enabled vertex and texture dimension, preserves current attributes, and sends generated attributes through ordinary vertex packing, lighting and texturing. Index maps retain their coefficients and enable state; the existing drawable routes are RGBA.
Lists own copied canonical command values and control points. Evaluation reads the executing context's current maps, grids and enables during replay. Shared list lifetime uses the existing shared-object namespace; maps themselves remain private. EVAL_BIT snapshots include definitions, grids and enables; ENABLE_BIT copies only enables. Context copy preserves independent map storage.
The focused native fixture resolves actual opengl32 and wglGetProcAddress thunks, checks stack cleanup/canaries, and has direct and queued Worker modes. It checks polynomial values, padded/transposed strides, homogeneous coordinates, derivatives, normal precedence, all scalar/vector forms, queries, invalid calls, attribute/context isolation and lists. Captured production draw records feed real WebGL1/2 tests for point/line/fill meshes, automatic-normal lighting, evaluated color material and texture coordinates.
Static checks completed: fragment balance, include manifest, duplicate-function ratchet, handler stack inspection and test tier membership; both new JavaScript fixtures pass syntax checking. These checks do not establish runtime correctness. The primary PDF was readable through the web tool; a direct curl fetch returned HTTP 403. No PDF download or new runtime result is claimed.
Recovery: the first worker ended with a model-capacity error after its 19:39:33 static milestone. The recovery worker preserved that implementation and the same evidence run. It added early Begin/End rejection before evaluator attribute-frame changes, and extended the prepared fixtures to all eighteen map targets and a full 8-by-8 homogeneous patch. These remain prepared tests until a native execution is recorded below. Runtime ownership remains with Alice until an explicit release; no predecessor or foreign boat is used.
The recovery worker observed Alice's explicit runtime release at 19:59:14Z and
created fresh boat bx_bfypzcje, TTL 3600, original expiry
2026-11-03T21:00:10.115Z. This task delegates its normal shutdown to this
worker after complete local recovery and SHA verification. All heavy work runs
there. The baseline is 5,839 exact Git blobs at the input revision plus nine
explicit owned source/test changes; unrelated shared-checkout WIP is excluded.
The locked full build passed. Both native software and the actual queued Worker
pass the focused suite; both log the same instrumented module SHA-256
ea6f09ae6635e14601cfdab4eaaf5224233deb3097e260ac97b30cb0302d82cb
(2,134,544 bytes). Worker assertions require queued submissions and zero
fallbacks. Direct and Worker production captures each pass real WebGL1 and
WebGL2 mesh, automatic-normal diffuse lighting, evaluated color-material and
four-color texture pixel assertions, with GL error zero. Inspected native PNGs
show the filled patch, its grey diffuse-light result, and the four texture
quadrants. No original executable or game encounter is part of this fixture.
Retained first failures: the build's address-literal gate recognized the numeric
EVAL_BIT|ENABLE_BIT mask as an address; composing the two named flag values
fixes it without changing the gate. The native fixture initially used an
unaligned Float64Array on a valid four-byte-aligned guest output; DataView now
reads it. One browser launch received API HTTP 500 without running. The next
browser run exposed a 4 KiB fixture memory too small for the authentic
glGenTextures output at guest 8389908; 16 MiB keeps that output mapped.
All polynomial, state, error and pixel assertions are retained.
All eighteen serial regression invocations pass: resolver availability, display
lists, attribute stacks, context-copy native/Worker/WebGL, edge flags
native/Worker/WebGL, shared objects native/WebGL, software and WebGL drawables,
lighting, interleaved arrays, and server state native/Worker. Existing tests and
their assertions are unchanged except the resolver inventory now partitions the
same fifty observed names into supported edge/evaluator APIs and the twenty-four
still unsupported names. A mutation replacing the polynomial result with zero
fails immediately at cubic x: 0 != 8; its distinct module hash and rejection log
are retained. No availability is claimed solely from lowering the missing count.
Reproduce in a fresh finite-TTL test environment with npm ci --omit=optional,
then flock /tmp/wine-build.lock bash tools/build.sh,
node test/test-gl-evaluators.js,
node test/test-gl-evaluators.js --worker, and
node test/test-gl-evaluators-web.js. Set GL_EVAL_CAPTURE on a native/Worker
run and GL_EVAL_CASES on the browser run to check either production capture;
GL_EVAL_PIXELS saves native PNGs and GL_EVAL_MODULE_OUT saves the exact
instrumented module. The evidence run retains complete commands, source seals,
initial failures, final logs, captures and module bytes.
Final verification checks all 5,842 staged source/test inputs against the pinned
baseline plus owned delta. The canonical module is SHA-256
03c7ffae336b2e250ea3b1869707321192fe4bdfb2d166d2d1b578ece634388a
(2,134,098 bytes); the compatibility module is
b47f499f30faea08b1fee1755ce0306c2615576e4c25584e97a6dd8597fc9437.
The final locked build passes with the repaired fixtures in place. Runtime is
Node 24.22.1 and Chrome 151.0.7922.108 using ANGLE/SwiftShader. The accepted
scope is this API family and the named regressions, not application gameplay
or a historical corpus-wide backend qualification.