feat(agc): vertex clip and cull distances

[MichonGoddijn231849]

What

Draws whose vertex stage exports clip or cull distances (PA_CL_VS_OUT_CNTL clip/cull enables and VS_OUT_CCDIST) were refused with clip distances, layer, viewport or auxiliary vertex exports. The recompiler already turns the misc and clip/cull position exports into ClipDistance/CullDistance outputs from ShaderVertexStageInfo::paClVsOutCntl, but the driver never passed the register, so the field was always 0.

• Driver: decodeVertexInfo passes PA_CL_VS_OUT_CNTL's clip/cull enables and VS_OUT_MISC/CCDIST bits. They are part of the draw lookup key and the vertex-info comparison, the recompile cache key, and the serialized request (version 18; main is at 17).
• Preparation (with @johnmaia's b599cfc): registration (from the registered state's PA_CL_VS_OUT_CNTL) and draw-time preparation set the bits the same way the draw does, so a draw whose register matches finds the artifact prepared ahead instead of preparing one at use. A program whose code has no pos1..pos3 export gets no bits at all (PositionExportLayout, cached per snapshot), so its artifact does not depend on PA_CL_VS_OUT_CNTL. That also covers johnmaia's 0x01240000 case, where only the misc-vector bit was set.
• State: up to eight clip and cull distances are accepted; more are refused by the decode and the precheck. State::paClVsOutCntl keeps the value for the validator's messages and its cache key.
• Validator: vertex ClipDistance outputs of f32[1..8] (they share one array with main's all-zero clip position distance from 9db923e, which takes the index after the guest's) and vertex CullDistance outputs of f32[1..8], at most eight together. A clip or cull capability on a device without shaderClipDistance/shaderCullDistance is refused with the PA_CL_VS_OUT_CNTL value in the message.
• Device: shaderCullDistance is enabled when available, next to main's shaderClipDistance, and CullDistance joins the device capabilities.

In GTA V the reflections in glass and some glass textures were missing: their draws clip against a plane. tibzejoker's Story mode stop (register 0x207 = 0x00400001, one clip distance) is this path.

Tested

Windows 11 Pro, AMD Ryzen 5 7600, NVIDIA GeForce RTX 5080 (driver 32.0.16.1074), 32 GB DDR5-5600, MinGW-w64 GCC 15.2.0.

• New execution tests:
  • agc_driver_vertex_clip_cull_distances: device draws for each enable pattern (no bits, clip planes 0, 1 and 0x5, cull plane 0 with every vertex negative and with one not, sparse cull plane 0x4), checking which pixels are written.
  • agc_driver_vertex_clip_cull_draw and ..._no_key: the position export layout of a pos0-only program (no bits), a pos1 program (only the layout bits), cached and at a code offset; then the same draws through PM4 submission, with the draw key on and off (the cache then keys on the decoded vertex stage, which must include PA_CL_VS_OUT_CNTL). Both shut the driver down at exit (thanks @Scramblrrrr for the teardown crash).
• Unit tests: four clip and four cull distances pass the validator, nine are refused; a device without the clip or without the cull feature refuses that capability only; four distances pass the decode and the precheck, nine are refused by both; the field survives serialization, changes the recompile cache key, and a version 17 request reads without it; position components decode to the right clip and cull indices.
• Main's agc_driver_zero_position_vertex and agc_position_invariance still pass.
• GTA V (PPSA04264), Prologue, on the earlier head: the reflections and glass textures render.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes. Written with Claude (Anthropic) under my direction; I reviewed and tested it.

🤖 Generated with Claude Code

[oneandonlydean]

Heads-up on the request serialization: #303 (mine) also moves it from version 5 to 6, for nonConstantImageOffsets. Git merges the two cleanly because the version lines are identical, but then a version 6 request written by a build with only one of them is misread by a build with both. Whichever of the two lands second should move to version 7. #426 currently has both fields under version 6.

[MichonGoddijn231849]

Thanks, good catch. #303 landed first, so this one moved to version 7: it writes 7, reads paClVsOutCntl only from version 7, and still reads #303's field from version 6. Rebased onto main with a round-trip test for the field.

[MichonGoddijn231849]

Rebased onto current main and the rebased #375 (now 7aa98bf). One more fix on top: #752 took recompile request version 7 for the conservative Z export, which is the version this PR used for PA_CL_VS_OUT_CNTL. A version 7 request written by main would have been misread, so this PR now writes version 8 and reads paClVsOutCntl only from version 8. The round-trip test still covers the field. Builds; the driver tests pass apart from agc_driver_scalar_sop1, which fails without this PR too.

Met on the current head: drawing tests (VertexClipCullDistances, VertexClipCullDraw), capabilities gated on device features, request versions 16-19 replayed in agc_shader_memory.