[Celegans12]

What

A variant whose SPIR-V emission throws is never remembered: findOrCompileVariant keeps only successful variants and the first attempt uses up the prepared program, so every later request reruns PrepareResourceProgram and the emitter to hit the same error. Plan failures already have a memo; this adds one for emission.

• compileVariant reports an exception from SpirvEmitter::Emit (or SPIRV-Tools validation), and the source entry keeps it per code address, binding layout and resource specialization. A repeat rethrows it before the disk-cache lookup.
• That's exact: a successful variant already serves every snapshot with that layout and specialization, the source key holds the target, and the address keeps a relocated copy's message right.
• Descriptor population failures (they depend on the snapshot) and out-of-memory errors are not kept. Nothing goes to the disk cache, and APS5_NO_FAILURE_MEMO=1 turns it off as for plan failures.
• PPSA26344 v1.008 with fix(shader): throw for constant lane reads outside the host subgroup #997: five wave64 pixel programs reading lane 63 (0x80002f0b00 among them) fail to emit on 32-lane hosts and make up most skipped title-load draws, 12-30 ms each.
• feat(agc): implement image and sampler tables indexed at runtime offsets #757-feat(agc): implement deferred faults for reads through null SRT pointers #759 and feat(agc): implement image_load through image tables #920 (ours) edit the bdaAbiVersion line after the emission and merge cleanly; perf(agc): baseline shader recompilation on registration #639 (draft) rewrites both functions and already needs a rebase.

Tested

• Windows 11, MinGW-w64 GCC 15.2.0, based on main ee391a5: configure -DBUILD_TESTING=ON, full build, --target libs, ctest:
• 368/368 tests pass (ctest --output-on-failure)
• Same build on lavapipe (Mesa 26.2.4 llvmpipe, 8-wide subgroups, VK_DRIVER_FILES): 367/367 pass; agc_driver_flat_store is left out there because it takes several minutes on Windows llvmpipe
• New verifyEmissionFailureMemo in agc_shader_disk_cache (CPU only): a pixel shader's clock read without VK_KHR_shader_clock emits once per layout, stride and address (on four threads too), twice with APS5_NO_FAILURE_MEMO=1. Fails on main.
• PPSA26344 v1.008 with fix(shader): throw for constant lane reads outside the host subgroup #997, before: skipped draws took up to 5.2 s per 10 s.
• Not tested: Linux (CI), console behaviour.

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 (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes

[adriadam10]

Checked on RADV (build + full ctest on current main, 395/395; agc_shader_disk_cache fails without the change). The memo key matches what the variant cache already reuses (layout + specialization) plus the address, population and OOM failures stay out, and nothing reaches the disk cache. Merging.