shader: a domain program reading a control point at an index computed from data is refused as "not affine", and a computed hull read leaves its patch when the relative patch is also data; fix branch ready #8051

Two related gaps in tessellation lowering's computed control-point reads (#4727, and the select/phi extension 9f429e65a):

1. Domain reads are refused. Onimusha: Way of the Sword (PPSA27836) picks one of its patch's three control points from data in the first scene: v_cmp results feed v_cndmask/v_add_co_ci to make an index of 0, 1 or 2, and v_lshl_add_u32 scales it by the 32-byte control-point stride before a buffer_load from the off-chip ring. Main refuses it:

   shader stage 3: tessellation lowering: the tessellation ring or LDS address is not affine in the control point and relative patch at pc 0x354
   

   Computed indices are accepted only for hull LDS reads. Also, the index is computed from data the domain program read from its own patch, and independentOfRelative follows that read's address into the relative patch index, so even a hull read would be refused for an index derived that way.

2. Hull reads point into another patch when the relative patch index is also data. A computed read keeps the program's own address, which contains relative * inputs * localStride. When the program also uses its relative patch index as data, every relative base is replaced by the grouped patch instead of 0, so that term stays and the read lands in another patch's slots. Nothing refuses it.

Fix branch: https://github.com/TheSovietWaffle/AnyPS5/tree/fix/recompiler-domain-computed-control-point (one commit on main 8c9161eac; Optimization/src/TessellationLowering.cpp, SpirvBackend/src/SpirvModuleEmitter.cpp, tests in tests/TessellationLowering.cpp and tests/execution/Tessellation.cpp, the hull input control points entry in docs/dev/debt/shader.md).

• Off-chip EvaluationInput reads get the same dynamic term as hull LDS reads; it must be a whole number of control points, and the backend clamps the domain gl_in index to the last output control point as it already does for hull reads.
• A value read from the patch's own ring or LDS no longer counts as depending on the relative patch index: the lowered read takes its patch implicitly.
• When the relative patch index is used as data, computed reads (hull and domain) subtract their relative patch term before the bases are replaced. Programs that do not use it as data keep the program's own address, as 9f429e65a's tests require.

Checks:

• New GPU case DomainComputedPoint in agc_driver_tessellation: a copy of DomainCode that reads control points 1 and 2 at cndmask(0, 1, FLT_MAX < x0) * 16 + base, which is always the static address at run time. It draws the same patches as DomainCode and passes on an RTX 3080 (617.42) and an AMD Radeon RDNA 2 iGPU (1002:164e, 32.0.21045.5002); on main it fails with the same "not affine" refusal.
• New unit case in agc_tessellation_lowering: a hull read at a computed index in a program that also references its relative patch now gets address - patch * 96; without the subtraction it fails. 9f429e65a's cases pass unchanged.
• The retained Onimusha domain request prepares (22,268 SPIR-V words); with it the title ran past the first scene for over 4 hours (with local workarounds for AGC: pixel shaders that export a stencil reference value abort on devices without VK_EXT_shader_stencil_export (NVIDIA; Demon's Souls, 1080p renderconfig) #4276 and AGC: compute buffer read of a native 4-sample D32 depth target needs sample-preserving publication (Onimusha: Way of the Sword, PPSA27836) #6570, so this is not a claim that the title works).
• With both fix branches merged on main 8c9161eac, all 806 agc tests on an RTX 3080 (617.42) and an AMD Radeon RDNA 2 iGPU (1002:164e, 32.0.21045.5002): 805 pass on each. The one failure on each (agc_driver_pass_input_hazard on the RTX 3080, agc_trimmed_import on the Radeon) fails the same way on unchanged main.
• check_conventions.py passes.

Not verified: the console's result for an index past the patch (the debt entry records the clamp).

AI-assisted: yes (Claude Code).