fix(agc): preserve registered array slices on CPU write-back - #6367

What

Preserve GPU results during CPU write-back when the selected texture-array slice is registered but unused slices in the descriptor extend outside registered memory.

For example, only slice 0 of a four-slice array may be registered and have pending results. StorageTexture::writeBackLayers previously checked registration of the entire declared image and discarded slice 0's valid results because the unused slices were unregistered. It also refreshed unsaved guest bytes outside the ranges being stored.

Validate each retained write-back range instead, and intersect unsaved snapshot reads with those ranges. Keep the existing overhang handling and drop behavior when an actual destination has been unregistered. No public API or shader-semantic changes; no title-specific checks.

The regression registers only the first or last slice while keeping the complete backing allocation mapped. It checks the drawn pixels and untouched unused slices. A separate retirement case unregisters the active allocation before flushing and verifies that no stale results are written.

SuperTuxKart PPSA69421 supplied real two-slice descriptors whose active first slice fits its registered allocation while the entire descriptor does not. This PR does not claim a title unblock, a black-rendering correction, or a performance improvement.

Related work checked before publication: #6202 scopes refresh and adds independent coverage in the same test file; #6197 changes import selection, including within writeBackLayers; #6280 changes mip handling within that function; #6246 adds an allocation lease for idle stores. None implements this retained-range registration check or snapshot-read intersection. These are independent mechanisms; retain their changes and this regression when reconciling textual conflicts. No dependency on those PRs or the shader experiment tracked by #1413.

Tested

Based on upstream main 73e924a; PR commit 540c3a1.

Windows (10.0.26200), Intel Core i7-13700, 32 GiB RAM, NVIDIA RTX 4070 12 GiB, driver 591.86. GCC 15.2.0, Ninja, Release, BUILD_TESTING=ON, ANYPS5_ENABLE_SPIRV_TOOLS=ON; checked-out submodule revisions match the repository gitlinks. Rebuilt the complete libs target and focused test binaries; preserved a complete 126-library runtime snapshot with SHA-256 hashes locally.

• Before: with the new regression and upstream's original production Texture.cpp, agc_driver_array_layer_target fails in 1.56 s:

  CPU write-back discarded the registered slice because an unused array slice was unregistered
  

• After: 14/14 focused tests pass in 17.25 s total, each under 30 s, with ANYPS5_GPU=RTX 4070 and ANYPS5_REQUIRE_VULKAN=1:
  • agc_storage_kept_unit_settle, agc_storage_upload_lifetime, agc_storage_upload_failure, agc_storage_split_writeback, agc_host_import_retirement, agc_storage_cpu_snapshot;
  • agc_driver_graphics, agc_driver_array_layer_target, agc_driver_depth_fast_clear, agc_driver_depth_fast_clear_submit, agc_driver_depth_storage, agc_driver_depth_storage_reuse, agc_driver_storage_one_slice_alias, agc_driver_storage_write_back_overhang.
• python tools/check_conventions.py --base upstream/main and git diff --check upstream/main...HEAD pass; tracked working tree clean.

Remaining validation: fresh standalone SuperTuxKart and vkQuake baseline/candidate comparisons, a completed race, controls/audio checks on this candidate, and a 15-minute stability run. Earlier successful interactive rendering/controls/audio used a separate experimental runtime and are not acceptance evidence for this PR. Linux, lavapipe and other GPU vendors have not been tested locally. No speed claim or performance acceptance result.

Checklist

• Based on current main; no other open PR implements the same fix (related function/file overlap listed above)
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt (no comments added)
• Unimplemented paths retain strict failure behavior; no silent stubs added
• No notes, investigation .md files, images, logs or game files added to the repository
• New third-party code is a submodule built from source (none added)
• Generic behaviour, not specific to one title
• Depends on: none
• Unblocks: none claimed; regression workload SuperTuxKart PPSA69421
• GPUs tested: NVIDIA RTX 4070, driver 591.86, Windows 10.0.26200
• AI-assisted: yes (OpenAI Codex)

Extended validation remains pending; this PR makes no title-unblock or performance claim.