[notquiteog]

main's Game-free source regressions check is red: 15 source checks fail at de86601, from PRs that were merged without a rebase onto each other. This makes them pass again, and fixes the three real bugs that were behind four of them.

What was wrong

Real bugs (the checks were right):

• RX_VK_CAPTURE_KEEP_OLD=0 and RX_VK_CAP_BLOCKS=0 turned the switch on (vk_memory.cpp, perf(vk): capture buffers in size classes, kept up to 256 MB and made ahead on a worker #2437). Both are on/off switches read as set-means-on. They read through env_on now, as line 1054 of the same file already did. (test_env_on)
• Mouse camera: the view snapped back after first-person aim or Z-targeting (host/mouse_native.c, fix: native camera positioning bugs #2617). The mouse's kept angles were reset when the game takes the stick, but not when the camera leaves states 3 and 4, so on return the old angles were written over where the game had left the camera. (test_mouse_native: yaw 4 where -1 is expected)
• DLSS could be widened into Ultra Performance below a third (dlss_checks.h, fix(vk): resolve DLSS permanent failure latching and improve dynamic resolution handling #2574). (w + 2) * 3 >= ow fixed 853x480 at 2560x1440 (853 × 3 = 2559) but also let a size one or two pixels under the third through, which mode_for then widens a range for. It is w * 3 + 2 >= ow now: a third rounded down, which is NGX's own Ultra Performance size, and no less. (test_dlss_checks: "below a third")

Stale translations: three English texts changed in vk_menu.cpp after the 13-language table was made (#1930's ocarina sentence in the gamepad hint, #2362's RADV note in the Mesh shaders hint, #2435's keyboard face in the Button prompts hint). Their 39 entries in menu_i18n.h are keyed and translated again. They are machine-made like the rest of the file and not reviewed by speakers. (test_menu_i18n)

Fixtures that cut source by its text and missed what the source gained:

Check	What it lacked
test_wayland_libdecor_hint, test_display_startup_failure	window_title() (#2116)
test_texel_buffer_ownership	dead_import_bytes (#1906)
test_audio_biquad, test_audio_aux	env_on in reverb_frame
test_mods_size_table	a stub for hos_xinput_pads (#2435)
test_readback_part	its marker matched #2458's forward declaration of wait_recorded
test_hdr_encode, test_framegen_preset, test_rumble_menu	kEngine's retry and temporal_reset_failure (#2574)
test_linux_docker_packages	its parser took the workflow's --no-update for a PPA

How it was tested

• python -I -B tests/run_all.py on this branch (Linux, the emsdk's clang, glslang 16.6.0): 468 passed, 0 failed, 22 skipped (Windows-only, or needing dotnet / PIL / a Vulkan device). On de86601 without it: 15 failed.
• The Vulkan back end builds from the branch (Linux, clang, with the NGX SDK).
• In the game, RTX 5060 Ti 16 GB, driver 615.71.09, Linux (COSMIC, Wayland), Kokiri Forest, 2048x1152 output: DLSS at Ultra Performance (683x384), Performance, Balanced, Quality, 75%, 83%, 92% and Auto, 54 to 60 fps, each feature made once (feature 2 made ... 683x384 (683x384 to 683x384) -> 2048x1152, mode 3), no device lost.

Not tested

• The mouse camera fix and the two capture switches are proven by their source checks, not by a run with a real mouse or with the switches set.
• Windows (the fixtures are OS-neutral; CI will say).
• The three re-translated menu texts were not read by a speaker of any of the 13 languages.

🤖 Generated with Claude Code

[Joey-Magnus]

Current main already has every part of this, through other PRs. On main at 51da046, tests/run_all.py has only test_lift_cache failing here (a local issue), and every check this PR fixes passes. The PR now conflicts in nine test files.

• The capture switches read through env_on: #2651 (1101a8e), gpu/vk/vk_memory.cpp:937/960/995.
• The mouse camera's kept angles after leaving states 3 and 4: #2653 (0ec31ff).
• The stale translations: #2654 (6946562).
• The source fixtures (window_title, dead_import_bytes, kEngine's retry, the readback marker, the mods stubs): #2648 (9776c36). The docker PPA parser: #2652 (f82caa6).

The one piece main did differently is scene_size_fits in gpu/vk/dlss_checks.h. Main kept (w + 2) * 3 >= ow, and #2648 made tests/test_dlss_checks.cpp:29-33 require that one and two pixels under a third fit. So w * 3 + 2 would now fail main's own check. If you think sizes two pixels under should be refused (because mode_for widens them), that's worth its own small PR arguing against those test cases. Otherwise I'd suggest closing this one as superseded.

[notquiteog]

Closing: main has all of this now through #2648, #2651, #2652, #2653, #2654 and #2661, as the review says. Thanks for checking it against main. The one thing that differs is scene_size_fits: main keeps the two-pixel tolerance with tests that say so, which is a deliberate choice and fine by me.