GL 1.1 interleaved client arrays

glInterleavedArrays now resolves through the real WGL and opengl32.dll procedure paths (API id 4314, GL opcode 177). Native client state implements all 14 GL 1.1 formats. Draws continue through the existing native encoder, software rasterizer, shared render worker and WebGL consumers.

The layout and enable rules were checked against GL 1.1 section 2.8, table 2.5, printed pages 25–26. The independent literal fixture is test/helpers/gl-interleaved-layouts.json; production code contains its own format cases and does not read that fixture.

Format Packed stride (bytes)
V2F 8
V3F 12
C4UB_V2F 12
C4UB_V3F 16
C3F_V3F 24
N3F_V3F 24
C4F_N3F_V3F 40
T2F_V3F 20
T4F_V4F 32
T2F_C4UB_V3F 24
T2F_C3F_V3F 32
T2F_N3F_V3F 32
T2F_C4F_N3F_V3F 48
T4F_C4F_N3F_V4F 60

Zero stride selects the packed stride. Positive strides, including overlapping records, are retained. Setup stores guest addresses without reading array memory. Invalid formats and negative strides report GL_INVALID_ENUM and GL_INVALID_VALUE before changing descriptors. Setup during an executing or compiling Begin/End reports GL_INVALID_OPERATION. Vertex arrays are enabled; index and edge arrays are disabled; optional arrays are enabled only when present. Absent descriptors remain untouched. Only the client-active texture array changes. Client stacks and context switching preserve this state; server attribute stacks do not restore it. Client array enables have native boolean, integer, float, double and glIsEnabled queries.

Array setup executes immediately in both display-list modes and is not list content. Array draws and elements inside lists now capture owned scalar attribute/position values. Replay uses the then-current transforms and disabled current attributes, without reading freed/replaced array or index storage or restoring client descriptors. Queued draws own their packed records before source memory changes; deleting a context fences its pending draw.

The frame-state regression fixture previously tried to update an undefined texture level. It now defines a genuine 16×16 level before the valid update and separately checks undefined-level rejection. Production texture validation is unchanged.

Verification

Evidence: scratch/runs/20261102-gl-interleaved-streamed-resume/ (ignored). Ordered Oni96725 then Braveheart25443 terminal/result/final/exact release and scoped helper absence were independently consumed before shared edits or boat use. Their 50 and 272 final artifact hashes matched. Prior untested drafts were reviewed; the old incomplete test assembler was repaired before installation.

The unchanged source was compiled and invoked through the real module lookup and Win32 dispatcher. It exited 1 at real export glInterleavedArrays (null lookup), with test module SHA-256 bf6c35fdd9061fc73ce052bbc86139f0454204f5613083bf733a8c9b2ffb645b. This is an executed old failure, not an inferred export comparison.

Final direct and actual queued software runs each check 84 literal draws (14 formats × packed/padded strides × DrawArrays/DrawElements/list), owned 19-float vertex payloads and inside/outside pixels. State checks cover both client texture units, all enable transitions, preserved absent descriptors, invalid setup, unmapped-pointer setup, overlapping strides, both list modes, client/server stacks, context deletion/recreation and pending-queue fencing. The queue requires positive queued-draw counts and zero fallbacks.

Both routes also feed their native-produced records through the production browser decoder, bridge, frontend and shaders on actual WebGL 1 and 2. Each route passes 504 color/normal-lighting/projective-texture pixel cases. Pixel oracles include byte color normalization, literal lighting values 191/255, and nearest-sampled texture colors selected by real projective q. Installed Chrome uses ANGLE SwiftShader; this does not establish hardware-GPU coverage.

Repeatable commands:

node test/test-gl-interleaved.js
node test/test-gl-interleaved.js --worker
node test/test-gl-interleaved-web.js
node test/test-gl-interleaved-web.js --worker
node test/test-gl-interleaved.js --negative-controls
node test/test-gl-interleaved-web.js --negative-controls
flock /tmp/wine-build.lock bash tools/build.sh

The four semantic controls deliberately corrupt packed stride, owned list color capture, shader normal consumption and projective q. Each exits 1 at its specified descriptor/value/pixel assertion. The positive tests retain real behavior checks; lookup success alone cannot pass. The availability regression was changed only after direct/queued/WebGL coverage passed: 56 observed names remain rejected. No original MDK2 or held game retry was performed.

Twelve focused regressions passed: frame state, command stream, texcoord4, server attributes, display lists, stipple, procedure availability (including its three broken-lookup controls), client arrays, client attributes, sparse pages, boolean/pointer queries and shared render-worker protocol. Queued stipple is checked separately. The first focused attempt stopped on missing bundled fonts; required existing font bytes were subsequently streamed and SHA-verified, then the suite passed. No fonts or browsers were downloaded.

The locked full build passed on the finite-TTL boat. Because that streamed tree has no Git metadata, its API-table gate could only do the index check; the real shared checkout's node tools/check-api-table.js also passed append-only versus HEAD. Exact final source/compiler/test/font verification matched all 3,995 pinned inputs, with no mismatches. The source context is main fd69d481859504a3ea234557b08f70171ad09fff plus this change; its underlying DirectDraw repair is ab3596819df988f9eea258604f66a0104eea7d36.

Executed/built artifact Bytes SHA-256
Instrumented native test module 2073436 aba030f7c3ca2ee583a0f6e03dfce55d8b33f69d319c224dc6a6a4e6fe57490e
Production full-build module 2073181 46c90ae60c2d47d0451e5b972c88ce8d0444500e7e734fd0683ca3ca18a7667a
Compatibility full-build module 2075754 f1e1c74334fb3dc439e89fa62fa2740b0b2776fff0ac50eff1acf0139d5eada9

Test-only accessors invoke the actual lookup, dispatcher and production GL functions; no implementation is replaced. The instrumented module adds those accessors to the same full source closure. Compressed identities pin every compiler member and src/api_table.json. The shared local build remains the older 307168f9e0475fd3c10f87cbc5d98f05006c6d4494d2f564381a9a0ebe289fe6 module; it was not substituted for the actual tested module. Recovery recipes and verified built artifacts are retained in the run.

Transport and limits

Fresh own boat bx_pefugusd had original expiry 2026-11-02T21:19:20.407Z, recorded before use. Source was streamed directly without a local source archive or checkout copy. The first owned stream was aborted and its process group removed when headroom fell below the 2,000,000,000 byte floor. Its supervisor minimum was 1,974,009,856 bytes; the continuous observer recorded 1,969,983,488. Thus this run cannot claim whole-task floor compliance. No source extraction or tests ran from that partial transfer. Its size/SHA and cleanup evidence were retrieved and verified before its owned remote partial was removed. No foreign artifact was removed by this worker.

After the coordinator independently verified and retired its released redundant transport caches, the retry used a 64 MiB reserve and passed at minimum 2,087,272,448 bytes free. The 81,509,424-byte compressed stream SHA was 060bb886d5d013761f18564961352babd6fe9bfa7292805fae7144be2d72559f; all 8,841 exact members were size/SHA-verified after extraction. Later transfers, retrievals and extraction retain the floor guard and their own recorded minima.

All 66 remote evidence snapshot members were retrieved and SHA-verified, including both production modules, the instrumented module, actual test logs, final source identity, and cleanup metadata. The modules were also checked against their separately recorded build identities. Remote test/browser helpers and temporary browser profiles were absent before shutdown. Only then was the explicitly delegated stop requested for bx_pefugusd; it reached STOPPED at 2026-11-02T19:43:13.659Z. No old or foreign boat was resumed or stopped. final-retrieval-verification.json, stop-request.json, box-stopped.json and the final local helper audit preserve that ordering. The original expiry above is retained independently of the stopped boat's changed TTL metadata.

The main commit/pull/push result is preserved in publication.json; its explicit file list excludes foreign/laptop/HeroesII work. This is GL conformance coverage, not MDK2 gameplay acceptance or a public deployment. Remaining unsupported families still gate any original MDK2 retry.