Public DD7/D3D7 MIRROR independent U/V: explicit WebGL1

PASS on both retained modules, through direct production WebGL and real GuestWorkerLink → RenderWorker queued rendering. Every observed context reports WebGL 1.0 (OpenGL ES 2.0 Chromium). Physical hardware and full27 remain unknown.

Module Direct exact RGB observations Queued exact RGB observations Held batches Rendered negatives rejected
Canonical 480 1,440 480 20
Compatibility 480 1,440 480 20

The unchanged public x86 DD7/QI D3D7/CreateSurface/CreateDevice7 vtables, independent continuous-coordinate oracle, signed periods/seams, four texels, all nine independent WRAP/MIRROR/CLAMP U/V pairs, point → linear → point, public state readbacks and exact caps checks come from the predecessor fixtures. No tolerance or assertion was weakened. Direct pixels additionally match gl.readPixels; every queued observation checks endpoint identity, actual context VERSION and zero GPU errors, fallbacks and native/fallback rasterizer draws. Every held production batch changes live state before draining and checks the captured old state followed by the new state.

Each module/path rejects WRAP-for-MIRROR, stale U, stale V and both coupled-axis policies under point and linear filtering. All 40 controls reach successful public DrawPrimitive7 and an actual RGB mismatch at PIXEL ORACLE, with WebGL1 identity already checked. Bootstrap failures cannot qualify these controls.

First failure and repair

The cheap explicit-version control first failed before browser execution: legacy device forwards explicit WebGL version: undefined !== 1. D3DIMGpu discarded the requested version when constructing D3D9Backend.Device; the legacy worker adapter also discarded endpoint options. The repair forwards webglVersion through both constructors into the existing D3D9Backend webglVersion → WebGLBackend apiVersion route. Omitted versions retain the existing automatic selection. Worker fence stats now expose actual gl.getParameter(gl.VERSION) alongside renderer and endpoint identity. No sampler defect was observed after this option-forwarding repair.

The fixture accepts --webgl=1, --webgl=2 or --webgl=auto (default). Explicit requests assert actual VERSION throughout direct/queued observations. Cheap controls cover constructor forwarding, independently reported worker VERSION, invalid arguments and unavailable explicit-version context creation: WebGL1 never requests WebGL2, and WebGL2 never requests WebGL1.

Runtime, source and recovery

Baseline sampler fixture: c6884b0987e358aaa3aadeb367d13a445f217b02. Actual staged source: f3ba696fb842a3295d8e1d99f4cc15c15ad6d1ec plus scratch/runs/20261104-legacy-mirror-webgl1/relevant.patch. The 35 minimal predecessor inputs passed hash verification before staging; only the fixture and three repaired production scripts were uploaded. source-seal.json pins all resulting inputs. No unchanged module rebuild was used to reproduce: the mandatory subsequent locked build independently reproduced both hashes.

Chrome151.0.7922.108 / Node24.22.1 on assigned bx_4hx68dxv. CDP identifies ANGLE Vulkan SwiftShader Device (Subzero), driver5.0.0; page/worker Intel/Mesa renderer strings do not establish physical hardware. One browser attempt ran 03:51:49.767Z–03:53:07.581Z (77.814 seconds). Helpers closed at03:53:07.582Z. Stable final recovery verified all eight unique files (result, log, four captures, both modules) by full SHA-256, and subsequent checks confirm browser15954/driver absent and 35/35 source pins unchanged. Boat's multiline command receipt caused a local JSON parser error after execution; driver.log preserves it. Runtime result is PASS; no browser retry.

The box remains RUNNING/APIidle, healthy and hydrated, expiry 2026-11-04T04:38:03.251Z; shutdown ownership returns to the coordinator. Recovery and cleanup completed before04:28Z. No lifecycle mutation, local Chrome, subworker, public deployment, source archive or media copy was used.

Tests and limits

Run evidence: scratch/runs/20261104-legacy-mirror-webgl1/. verification.json checks 3,840 exact observations, 960 held batches, 40 rendered negatives, actual WebGL1 versions and retained module hashes. The first failing control, repaired planners, relevant binary diff, new note patch, full build, focused logs and stable recovery manifest are retained there.

PASS: explicit-version controls/planners, predecessor planner, D3DIM batching, depthless targets, points, generic GPU backend, shared render-worker protocol, tier membership and flock /tmp/wine-build.lock bash tools/build.sh. The existing fallback-counter test fails fallbackDrawTexture: 0 !== 1 with both baseline and repaired D3DIMGpu; both logs are retained, with no assertion change. This is an existing test limitation, separate from the passing pixel contract. WebGL2 runtime remains the predecessor's result; its explicit-option controls pass here. Other sizes/formats, mip/perspective behavior, physical GPUs, original games/full27, resource lifetime and the rejected queued-texture diagnostic are unexecuted or excluded.

Reproduce on an authorized box with the sealed inputs and existing dependencies:

LEGACY_MIRROR_GPU_SOURCE=<full-Git-SHA> \
LEGACY_MIRROR_GPU_MODULE_DIR=<retained-module-directory> \
LEGACY_MIRROR_GPU_EVIDENCE=scratch/runs/<fresh-id>/browser \
  timeout --kill-after=15s 840s node test/test-legacy-mirror-gpu-worker-contract.js --webgl=1

--plan --webgl=1 runs cheap controls only and cannot qualify browser coverage.