perf(agc): reuse readonly fallback copies - #5469

What

Retain immutable raw copies of tracked read-only fallback buffer regions through the recorder's existing storage-snapshot cache. Unchanged readers can bind the same buffer; writes are flushed and freshly collected before lookup so a CPU store within the same collection epoch cannot serve stale bytes. The existing cache supplies its entry/byte budgets and preserves buffers held by older consumers.

Reuse is limited to uploads holding the GPU mutex, using the active recorder on the same device, with a nonzero write-watch generation and a fully committed, live, non-addressable region that no descriptor writes. Writers, sparse regions, untracked memory and unlocked preparation keep separate copies. APS5_NO_READONLY_FALLBACK_REUSE=1 disables reuse for comparison; APS5_PROFILE_DRAW reports avoided logical copy footprints.

Related open #1364 includes immutable buffer reuse alongside upload batching and shader-data streaming. This is a draft alternative for the reuse portion using current main's existing cache, with no batching or streaming redesign. Ordering with #1364 has not been agreed and needs maintainer/author coordination before this becomes ready for review. The checklist explicitly records that overlap rather than asserting its absence. No title FPS or frame-time improvement is claimed by this PR.

Tested

Fresh Release build against main a0f5916a293067ca840e835e58a8a55a0b764b4b, Windows 11 Pro / MinGW-w64, SPIRV-Tools disabled, native NVIDIA Vulkan with Khronos validation and shader cache disabled.

The new native fixture passes (0.556 s): unchanged-handle reuse, a CPU overwrite within one explicitly established collect epoch, queued GPU consumers retaining old bytes, writer bypass, a real GPU fill published before the next reader, generation-zero bypass, unlocked preparation, cache eviction with a live consumer, private recommit, shared-section CPU writes/remap at the same virtual address, and sparse descriptor clipping/bypass. Mapping changes use the existing allocation invalidation callback without registering the fallback memory for import. Disabling reuse in the identical binary fails at the expected unchanged-handle assertion (0.460 s).

CTest: 5/5 passed, covering the new fixture, unchanged CPU stamps, written-buffer ordering, copied-direct ordering and the draw snapshot cap (3.74 s total, each test below 1 s). The dedicated draw descriptor writes and snapshot override check also passed (0.450 s). No cases skipped or Vulkan validation errors reported. Linux/lavapipe and other hardware vendors have not been tested.

Checklist

• Based on current main; open-PR overlap reviewed and disclosed: perf(agc): reuse immutable buffer copies and batch uploads #1364 (draft alternative; ordering remains pending as described above)
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw; no new stubs
• No notes, investigation .md files or images added to the repository
• New third-party code is a submodule built from source (no new dependencies)
• Generic behaviour, not specific to one title
• Depends on: none (perf(agc): reuse immutable buffer copies and batch uploads #1364 is an alternative, not a build prerequisite)
• Unblocks: none (generic fallback-copy retention; no title bring-up claim)
• GPUs tested: NVIDIA GeForce RTX 4070 Laptop GPU, Windows driver 32.0.16.1692, Windows 11 Pro 10.0.26300
• AI-assisted: yes