[Dripl1x]

What

On Windows a host import of private memory under 1 MiB is compared until its deferred settle, and the deferred return to the write watch is skipped while a live storage image overlaps the range (the rewatch stamps it as a CPU store, which the image would take as a write over its results). SuperTuxKart (PPSA69421) views its small textures (96 KiB to 352 KiB) from storage images, so they stayed compared for good and every draw sampling one took the full revalidation walk with a byte compare.

• Imports larger than one 64 KiB block now settle at the import, as the 1 MiB ones did, and stay in the write watch. 64 KiB imports, the size a title maps around every submission, keep the deferred settle.
• The cost moves to the import: each new private import from 128 KiB up to 1 MiB waits for its settle read (the [buffers] line's "settled inside their window" counts them). In SuperTuxKart they are made once; a title that makes such imports every frame would pay a wait per import.
• APS5_SETTLE_FLOOR_KIB sets the floor (1024 restores the old one). The retire test needs a 128 KiB import on the deferred path and sets it to 1024; with the default floor, an import replaced while its deferred read runs is now only reached by 64 KiB imports and has no test of its own.

SuperTuxKart main menu with the tutorial prompt (about 59 draws a frame), on a stack based on main a0900b9d99 (#7464, #9402, two local direct-memory changes and three other local changes), APS5_PROFILE_DRAW over 10 s, the old floor against this one on the same build:

	revalidations	time	full walks	draw recorded in
APS5_SETTLE_FLOOR_KIB=1024	33583	1339 ms	19435 (19355 for a failed collect)	104 us
this PR	18287	315 ms	204 (audits only)	82 us

Dumped menu frames 0 to 6 of both matched within a mean difference of 0.001. The menu runs at the title's 60 FPS cap either way, so this shows as spare time per frame, not FPS: on current heads (main 9d4fb4fb70 + #7464 + #9402 + this and three other local changes, the menu without the tutorial prompt, about 33 draws a frame) 59.94 FPS with and without it, the longest idle gap of queue worker 0 and the committer per frame 6.7–7.0 ms (median) with it against 5.9 and 6.2 ms without (a third run without it lost the window focus and is left out).

Overlap: #9941 also changes importRunAt (it marks partial-block imports settleOnDemand and settles them on demand before an initial multisampled DCC import, for #9724). With this PR a private import of more than 64 KiB, partial or not, settles at the import, which may already cover #9724's 784 KiB import; that was not tested on those titles. Whichever lands second needs a small rebase there.

Tested

This head: main 9d4fb4fb70 + this commit; merges cleanly into 5e6ae6e3d0. Windows 11 Pro 26300, Ryzen 7 5800H, RTX 3070 Laptop GPU (NVIDIA 621.01), WinLibs GCC 15.2.0 r7, Release: every ctest entry of agc_driver_recorder_tests (86, the agc_import_settle_* entries included), agc_driver_host_import_windows_tests, agc_driver_host_import_at_map_tests, agc_driver_host_import_teardown_tests, agc_host_import_retire_tests and agc_driver_guest_memory_lifetime_tests (6) passes.

Linux (Ubuntu 26.04 in WSL, llvmpipe), on the measurement stack: agc_driver_recorder passes (the change is Windows-only code).

Titles: SuperTuxKart menu as above. vkQuake (PPSA99010), on the earlier stack: 27 imports settled at the import over a run, 82–85 FPS with this series and without it. Not run: the full CTest suite, other titles, other GPUs.

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
• Unblocks: none
• AI-assisted: yes (Claude Code)

[github-actions]

This pull request is labeled perf, so it needs the summary line of a capture on main and one with the change, from the same title, scene and machine, as the Performance report form asks. Please paste both aps5-summary.jsonl lines, attach the two traces and add the output of tools/trace_digest.py.

[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 1368	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 1395	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 1896	docs/dev/ENVIRONMENT.md
PR 4472	docs/dev/ENVIRONMENT.md
PR 5021	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 5925	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6186	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 6187	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6188	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6192	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6197	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6209	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 6481	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6918	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 6924	docs/dev/ENVIRONMENT.md
PR 7085	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake
PR 7449	docs/dev/ENVIRONMENT.md
PR 7457	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp docs/dev/ENVIRONMENT.md
PR 7514	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp docs/dev/ENVIRONMENT.md
PR 7604	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp docs/dev/ENVIRONMENT.md
PR 7642	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 7643	docs/dev/ENVIRONMENT.md
PR 8004	docs/dev/ENVIRONMENT.md
PR 8484	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp docs/dev/ENVIRONMENT.md
PR 9012	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp docs/dev/ENVIRONMENT.md
PR 9289	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 9345	docs/dev/ENVIRONMENT.md
PR 9356	docs/dev/ENVIRONMENT.md
PR 9387	docs/dev/ENVIRONMENT.md
PR 9852	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp docs/dev/ENVIRONMENT.md
PR 9884	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake
PR 10193	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 10678	docs/dev/ENVIRONMENT.md
PR 10757	docs/dev/ENVIRONMENT.md
PR 10894	docs/dev/ENVIRONMENT.md
PR 10920	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp
PR 11435	docs/dev/ENVIRONMENT.md
PR 11436	docs/dev/ENVIRONMENT.md
PR 12157	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 12324	docs/dev/ENVIRONMENT.md
PR 12544	docs/dev/ENVIRONMENT.md
PR 12746	docs/dev/ENVIRONMENT.md
PR 12848	docs/dev/ENVIRONMENT.md
PR 12849	docs/dev/ENVIRONMENT.md
PR 12970	docs/dev/ENVIRONMENT.md
PR 13148	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 13251	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake
PR 13486	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp docs/dev/ENVIRONMENT.md
PR 13707	docs/dev/ENVIRONMENT.md
PR 13793	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 13809	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 13863	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 13920	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 13928	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 13929	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 14051	docs/dev/ENVIRONMENT.md
PR 14344	docs/dev/ENVIRONMENT.md
PR 15059	docs/dev/ENVIRONMENT.md
PR 15136	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 15169	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 15296	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 15302	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp
PR 15529	core/libs/prx/libSceAgcDriver/tests/Recorder.cpp core/libs/prx/libSceAgcDriver/tests/cmake/agc_driver_recorder_tests.cmake docs/dev/ENVIRONMENT.md
PR 15548	docs/dev/ENVIRONMENT.md
PR 15570	core/libs/prx/libSceAgcDriver/Graphics/src/GuestBufferMemory.cpp core/libs/prx/libSceAgcDriver/tests/Recorder.cpp docs/dev/ENVIRONMENT.md

[MarcZX10]

Reviewed exact head 90b9574ad52da6ecf26c1cec51bd1fea2587d592 against refreshed upstream main c8d34aa8f4d8058d57169cc1e423ba6db00faaf0.

The threshold change is narrowly scoped to Windows private imports and preserves the 64 KiB deferred-settle path; the new 128 KiB retirement case restores the old 1 MiB floor for that test. The recorded Windows suite and SuperTuxKart measurements are useful, but they were run on a stack based on 9d4fb4fb70 with three additional local changes. The current branch is based on that same 9d4fb4fb70 commit, while current main has 20 commits not present here, including subsequent AGC import/write-watch changes. A merge-tree preview is clean, but that does not verify the changed settle behavior with the current integrated import paths.

Please update this branch to current main and rerun the focused Windows tests on the final composition: the recorder suite, host-import Windows/map/teardown suites, retire and guest-memory lifetime tests. Report which tests actually ran and whether the existing GPU/title comparison still holds on that base. Keep the environment floor override exercised so the retire test still covers deferred settling.

Local review: current-main merge-tree is clean; current-main conventions and diff-check pass. Hosted eligibility, intake, labels, report, and conventions pass; hosted product/relinker jobs are skipped. I cannot reproduce Windows/GPU tests locally. This request is for fresh integration evidence, not a claim that the implementation is incorrect.

[tibzejoker]

One thing from Grand Theft Auto V (PPSA04264) that may matter for the larger floor, since it involves the same path on Windows.

• What happened with perf(agc): settle Windows imports under 1 MiB without waiting #6678 on GTA V. The deferred settle returned small private imports to the write watch. Some of that memory was later mapped a second time as a shared view (the zlib output buffers of its streamer, 16 KiB MEM_MAPPED views). The view was then armed read-only, so the game's stores into it faulted. Each resumed fault can clobber the System V red zone below rsp (Windows: resuming a write fault overwrites the guest red zone when AVX state is live #361). inflate_fast keeps state there, and the streamer thread crashed after a single resumed write (#9136, with a per-rip fault log showing the fault and then the crash on the same thread).
• What covers it now. fix(agc): keep a rewatched import's memory compared once it is shared #9159, merged yesterday, keeps rewatched memory compared once it becomes shared. feat(relinker): frame red-zone leaves without a frame pointer by moving their rsp operands #5640's relinker framing covers the red zone itself, but only for a title relinked with --frame-red-zone, which is off by default.
• Why it may matter here. This PR returns more private memory to the write watch at the import (every import from 128 KiB to 1 MiB instead of only the 1 MiB ones). On main with fix(agc): keep a rewatched import's memory compared once it is shared #9159 that should be covered. Any case fix(agc): keep a rewatched import's memory compared once it is shared #9159 misses would then show up as more resumed write faults in guest code, and as crashes in titles that keep data below rsp.

If it helps, I can run GTA V's open world on your head and on its main, without --frame-red-zone, with the per-rip resumed-fault log on, and post the counts by memory type and whether the streamer crash comes back.

AI-assisted: yes.