Public DD7/D3D7 mip selection: bounded direct WebGL subset
Canonical-module direct WebGL1 and WebGL2 PASS: 17 observations / 13 public draws each, with four rendered negatives rejected in total. Overall browser execution stopped at the first Worker setup failure; queued rendering and compatibility-module rendering are unexecuted. No production sampler defect was observed. CDP identifies SwiftShader; physical hardware is unknown.
test/test-legacy-public-mip-webgl.js adapts the existing public x86 COM
fixture using checked, unique source seams. DirectDrawCreateEx obtains DD7,
QI obtains D3D7, CreateSurface creates a 16×16 five-level X8R8G8B8 mip chain,
and public GetAttachedSurface7 acquires each child. Public Lock/Unlock fills
each authored level with a distinct solid color and checks dimensions/format.
CreateDevice7, SetTexture7, SetTextureStageState7/readback, BeginScene7,
DrawPrimitive7 and EndScene7 all use acquired vtables and real x86 CALLs with
stdcall/guard checks. There are no fixture-only WAT exports or module patches.
GetCaps7 checks exact triangle 0x030303ff and line 0x030003c3 filter masks.
The subset covers mip NONE at strong minification, POINT→LINEAR→POINT and NONE restoration, POINT footprints 0.25 through 32768 (including level clamping), LINEAR and trilinear half-level blending, and five perspective samples selecting four distinct levels. The independent oracle differentiates perspective U/Q, V/Q across a 32-pixel triangle, derives the largest texture-space footprint, and computes nearest-level selection or inter-level blending. It calls no production sampling helpers. Expected colors come from those calculations and authored texture inputs, never software-rendered output. Production derivatives and mip selection remain active; no fixed LOD is supplied.
Every observation checks public render-target Lock pixels and independent
gl.readPixels, actual context VERSION and zero GL errors. Both paths report
13 GPU draws, five texture-level uploads, zero errors and zero fallbacks.
The original mip test's tolerance of two channel units is preserved; maximum
observed error is 0.5. Deliberately setting mip NONE throughout rejects a
level-zero-only implementation; leaving POINT stale when LINEAR is requested
rejects a stale mip filter. Each negative reaches successful public DrawPrimitive7,
actual WebGL readback and a discriminating PIXEL ORACLE failure.
First setup failure and bounded repair
Both retained modules first passed browser-free public bootstrap, including five acquired mip levels, public fills/dimension checks and caps. Cheap controls also check constructor/worker WebGL-version forwarding, unavailable explicit versions, public-call fixture seams, analytic blend/restoration values and negative discrimination at the unchanged tolerance.
The single browser attempt passed canonical direct WebGL1/2 and their negatives,
then failed before any Worker COM calls:
SharedArrayBuffer transfer requires self.crossOriginIsolated.
The new fixture's custom HTTP routes returned before static-server's ordinary
isolation-header path. A browser-free HTTP control reproduces the missing
header (null versus same-origin). The final candidate sets COOP/COEP in
the custom handler and runs HTTP controls against the page, module and guest
worker routes before bootstrap/browser execution. Those controls pass.
The repaired HTTP setup has not been rerun in Chrome: the assigned single
browser attempt was already consumed. Worker pixel/held-batch code is present
for a future authorized run, but none of it qualifies coverage here. No
production code was changed. The runtime's FAIL status and first error remain
intact in evidence. A separate local command-wrapper parser mistook Boat's
streaming attached receipt for command completion; this did not rerun the
browser. The terminal runtime result and recovered artifacts establish outcome.
Provenance and recovery
Run: scratch/runs/20261104-legacy-public-mip-webgl/.
Staged Git baseline: 5b975ccce19bdeacc24fcf455b30f32d536152c9 (includes
7d1a1289ea698cff150a5227a33f2dfae62fad9c explicit-version forwarding).
tested-new-file.patch reconstructs the exact initially tested fixture;
source-seal.json pins all 36 staged inputs. Final candidate changes are
recorded separately in relevant.patch and owned new-file patches. The assigned
minimal staged runtime was reused; only the new fixture was uploaded.
- Canonical: 2,143,712 bytes,
1a021b0f9ef7eeba12637c66e8929fdd75383e16c431ee318f812af8dbd78e58. - Compatibility: 2,146,325 bytes,
74b34125e175895dc3ca843ccaa555f934611e3bd676e1e95f25aae9c2e7be66. - Existing
test/binaries/calc.exebootstrap:cb71fcc9726f756421519a215c39cf89efd930c9ae0fa5bbb59b64050faef52b.
Chrome151.0.7922.108 / Node24.22.1 on bx_4hx68dxv. CDP reports ANGLE Vulkan
SwiftShader Device (Subzero), driver5.0.0. The browser ran
04:02:03.187Z–04:02:07.823Z (4.636 seconds); browser23405 and its driver are
absent, server/helpers closed. Stable recovery verifies eight unique files
(two module binaries, two captures, two results and two logs) by size/full SHA;
36/36 source pins remained unchanged. No source archives or media evidence
copies were made. The box remains RUNNING/APIidle, healthy/hydrated, expiring
2026-11-04T04:38:03.251Z; lifecycle/shutdown ownership returns to coordinator.
PASS: bootstrap on both modules; bounded canonical direct WebGL1/2 pixels and
negatives; final planner/oracle/options/HTTP controls; predecessor WebGL1 planner;
test-tier regression/membership; flock /tmp/wine-build.lock bash tools/build.sh.
The mandatory build reproduced both retained module hashes. The new fixture
is registered as E2E. No fallback-counter detour was performed.
Next: one authorized browser run of the repaired fixture, starting with queued Worker isolation and then compatibility-module rendering. Further uncovered cases: non-solid mip texels and MIN/MAG discrimination, vertical-only footprints, partial/malformed chains, per-level formats/padded pitch, changed lower-level uploads, legacy combined filters, other dimensions/formats, physical GPU and original games/full27. Resource lifetime/rejected queued textures are excluded.
node test/test-legacy-public-mip-webgl.js --plan
LEGACY_PUBLIC_MIP_SOURCE=<full-Git-SHA> \
LEGACY_PUBLIC_MIP_MODULE_DIR=<retained-module-directory> \
LEGACY_PUBLIC_MIP_EVIDENCE=scratch/runs/<fresh-id>/browser \
timeout --kill-after=15s 840s node test/test-legacy-public-mip-webgl.js
--bootstrap performs public bootstrap on both pinned modules without Chrome
or draws. --plan performs no Wasm instantiation or browser rendering.