fix(agc): read GPU timing queries without waiting on unwritten ones

[hypermoist]

What

Recorder::readGpuTiming reads a batch's timestamp queries after the batch's fence was waited on, so every stamp the batch wrote is already available. It still passed VK_QUERY_RESULT_WAIT_BIT. Begin and end stamps are separate calls (BeginGpuTiming / EndGpuTiming), so a timed range whose end never ran, for example a draw that failed between the two and was skipped, left its end query unwritten, and vkGetQueryPoolResults then waited forever on the reaping thread. This only happens with GPU timing on (APS5_PROFILE_GPU, draw profiling or a performance trace).

Without WAIT_BIT the call returns VK_NOT_READY for such a batch and its timings are dropped, as they already are for any other non-success result.

Tested

Linux (CachyOS, GCC Release, NVIDIA GeForce RTX 3060): builds; the AGC group of ctest passes except agc_driver_store_over_untracked, which fails on unmodified main too. tools/check_conventions.py passes; tools/pr_overlap.py --branch reports no line overlap. The stall was found profiling PPSA08427 with APS5_PROFILE_GPU=1 on my local branch; the game was not rerun on this exact rebase.

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 under Silent stubs in their library's technical debt file
• 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
• Unblocks: none
• GPUs tested: NVIDIA GeForce RTX 3060 (proprietary driver)
• AI-assisted: yes

🤖 Generated with Claude Code

https://claude.ai/code/session_01Q8tinbP1MUpUaDsbNWnGnu

[github-actions]

Some comparisons could not be completed because a pull request conflicts with the base branch. Review these shared files manually before concluding that there is no overlap.

Pull request	Shared files not compared
PR 5431	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 5609	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6185	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6186	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6188	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6539	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6615	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6946	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 6947	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp
PR 7642	core/libs/prx/libSceAgcDriver/Graphics/src/Recorder.cpp

[Neithennn]

Windows NVIDIA run, since the description lists Linux only.

Windows 11 Pro 26200, NVIDIA GeForce RTX 4080 (driver 617.42), AMD Ryzen 7 5800X, WinLibs GCC 15.2.0 r7, Ninja, Release, bundled dependencies, BUILD_TESTING=ON, ANYPS5_NO_SHADER_CACHE=1. Head 44d31ef9d merged onto main ff23657f3; full build succeeds.

• agc_performance_trace, agc_driver_draw_program_trace, agc_driver_shader_clock, agc_subgroup_clock, trace_digest, performance_compare: pass.
• ctest -R "agc_|shader|recompiler|spirv|video_out" -j 8: 622/623, same as unchanged main on this machine; the one failure is agc_driver_store_over_untracked (Windows: StoreOverUntracked fails at "the first result was not stored" #6347/AGC: StoreOverUntracked's NeighbourCopied case intermittently misses its pending result on Windows / RTX 4080 #6383), also on main.

These tests do not seem to reach Recorder::readGpuTiming. I tried APS5_PROFILE_GPU=1 on agc_driver_graphics: it stops at missing Vulkan function: vkCreateQueryPool both with this PR and on unchanged main, so the changed path is not exercised here. No title run.

AI-assisted: yes (runs done with Claude Code).

[JJenkx]

I reproduced this on AMD Radeon RX 6650 XT / RADV Mesa 26.2.4-arch1.1, Arch Linux, current main b7cedd900. A real-queue test deliberately leaves one range end unwritten, completes another range in the same batch, then submits and syncs. With main, the all-complete control passes but the incomplete case times out at 15 seconds. With your exact 44d31ef9d change, both pass, and complete batches before/after the incomplete batch retain valid GPU timings. The test executable is byte-identical across the negative control and fixed runs.

Added the test-only follow-up as draft #6514, explicitly depending on this PR. Validation source with your fix; standalone test commit is based directly on main if useful to cherry-pick. Full libs build and four relevant tests pass on both RADV and lavapipe with the fix. Lavapipe also passes the unfixed case, so software-driver-only testing does not expose this hang. No title or FPS claim.