fix(shader): ignore unused descriptor fields for scalar buffer loads

[eljackjoestar-cmd]

What

A scalar-only s_buffer_load_* resource can carry non-buffer fields in its descriptor. The recompiler previously rejected word 3's TYPE; this PR normalizes word 3 only for scalar-only resources. Review also found that Kena reached ShaderResources::addGuestBuffer and stopped on reserved bit 30 of word 1. The same scalar-only normalization now clears that bit before building the driver-visible descriptor, preserving base address, stride and record count.

Vector, mixed scalar/vector, formatted, writable, atomic and image-alias resources retain their validation. No driver-wide acceptance of reserved bits is added.

Semantics source: AMD RDNA2 ISA §7.2.1, printed page 52 (PDF page 60), describes the scalar buffer fields as base address, stride and record count, with other fields ignored: https://gpuopen.com/download/rdna2-shader-instruction-set-architecture.pdf . Native console/ISA comparison of the ignored reserved bit is not claimed.

Lost Soul Aside (PPSA26358), captured shader 0x12c2298400, previously stopped on unsupported descriptor TYPE. A Kena run reported in review exposed the later reserved-bit rejection. This addresses those validation rules; no new Kena game run or completed gameplay is claimed.

Tested

Rebased onto current main f87799aa3, Windows 11 Pro, GCC 15.2, Release, SPIRV-Tools OFF. Full Windows build passed.

The new native regression failed before the reserved-bit normalization with AGC graphics: buffer descriptor has reserved bits set. With the fix, every lane receives the expected scalar value across both reserved-bit states and four word-3 patterns (including image-shaped TYPE and all bits set). Vector-only and mixed scalar/vector resources still reject non-buffer TYPE during recompilation and reserved word-1 bits during driver dispatch.

• NVIDIA RTX 5070, driver 617.14: four focused tests passed without skips in 1.93 s: agc_driver_buffer_descriptor_runtime, agc_shader_memory, agc_driver_buffer_load_store, agc_driver_scalar_wide_loads.
• AMD Radeon integrated gfx1036, driver 32.0.21036.18: the three Vulkan buffer/scalar tests passed without skips in 1.25 s.
• Conventions and whitespace checks passed. Linux/full-suite/CI and complete game playability are not claimed. These checks use this branch alone on the stated main, not the combined title integration.

Checklist

• Based on current main; no equivalent open PR found
• One topic per PR
• No comments in code added
• Unimplemented paths throw; no silent stubs
• No investigation files, requests, logs or screenshots added to the repository
• No third-party code added
• Generic scalar-load behaviour; vector/mixed validation retained
• Depends on: none
• Unblocks: PPSA26358 Lost Soul Aside (PPSA26358): bring-up and remaining blockers #4401 (one descriptor-validation blocker; later blockers remain)
• GPUs tested: NVIDIA RTX 5070 / 617.14 and AMD Radeon integrated gfx1036 / 32.0.21036.18, Windows 11 Pro 10.0.26300
• AI-assisted: yes, OpenAI Codex

[alfalfa-star]

Tested on Kena: Bridge of Spirits (PPSA01802), Linux/RADV (RX 9070 XT), on main 0214069 with this PR merged.

Kena has the same pattern: a pixel shader binds a T# in buffer slot 1 and reads it only with scalar loads. The read extent is 108 to 112 bytes, first used at pc 0x94. One example of the words is 20f56700 caa00000 03ffc3ff 909c0fac.

• Without this PR, those draws stop at buffer descriptor uses an unsupported type (about 7.1k skips in a 5 minute run).
• With this PR the recompiler accepts them, and the type refusal is gone. The same draws now stop in the driver instead: ShaderResources::addGuestBuffer checks word 1 bit 30 before anything else and refuses with buffer descriptor has reserved bits set (1,353 skips in the same window). The T#'s word 1 (0xcaa00000) has that bit set, and this PR only normalizes word 3.
• Read as a V#, the base is 0x20f56700 (T# word 0 plus the low 16 bits of word 1). That address isn't registered guest memory here (RegisteredReadableEnd returns 0). Stride × NUM_RECORDS from the T# words also comes out at several GB, over maxStorageBufferRange.

So for this title the scalar-only case also needs the driver path to stop validating fields that s_buffer_load doesn't use, and to decide what an unmapped base reads. As a local test only, I bound it as an empty buffer, and Kena then reaches its combat tutorial with the arena lit. I'm not claiming that is the right upstream behaviour.

[oneandonlydean]

The conventions check fails because the body lost the template sections (What, Tested, Checklist); please put them back.
alfalfa-star's Kena run shows the same scalar-only T# now stops in the driver instead: ShaderResources::addGuestBuffer refuses word 1 with buffer descriptor has reserved bits set. If s_buffer_load ignores those fields, the driver check needs the same rule for scalar-only buffers, otherwise the titles still stop. Can you cover that here?

[eljackjoestar-cmd]

Restored the What, Tested and Checklist sections and rebased onto f87799aa3. The scalar-only snapshot normalization now also clears reserved bit 30 of word 1 before the driver-visible binding is built, while preserving the base, stride and record count. The general driver check remains strict for vector/mixed and other non-scalar-only uses.

The new native case failed before the change with AGC graphics: buffer descriptor has reserved bits set. With afd9b92b0, all scalar readbacks pass across both word-1 states and four word-3 patterns; vector-only and mixed uses still reject reserved bits during dispatch and non-buffer TYPE during recompilation. Full Windows Release build, four focused NVIDIA checks and three AMD Vulkan checks pass without skips; conventions pass. I have not rerun Kena itself, so this is an end-to-end driver regression result, not a new Kena gameplay claim.

[oneandonlydean]

Rebased onto main: kept both debt entries in docs/dev/debt/shader.md; build and shader tests pass.

[p3n1cillin]

Heads-up @eljackjoestar-cmd: a first-parent bisect points at this merge for a regression in Surround Sound Studio (PPSA99051). It exits about 3 s in with [ps5-agc] submission retirement failed on Windows/NVIDIA, and is fine on eecf90953. Details and the bisect table are in #5876.