Merge plan after v0.3.11: corrections after review

[itaskaev2]

@valentderah, a reply to your review of the merge plan for the release after v0.3.11. You were right on every point; the plan below is corrected and, as of 2026-11-07, brought up to date with what went into dev and what changed on our PRs.

Every point of your review holds after checking the code:

• Grow the texture heap a mod's HeapDef asks for, so HD texture packs load (#673) #713 reserved 10 GiB for everyone.
• Hold the game's dynamic resolution at 100 % and let the scene grow past 4096 wide #573 keeps {"DYNAMIC_RES", "0", ...} as the default.
• Add a perf log summary tool for comparing settings #487 still takes the percentiles from per-window means. In summarise, one FRAME_RE match is one window, and percentile() runs over those. 3cae2df fixed the bound count only.
• Fix shrine floor flicker with reusable linear-depth snapshots #690 and Sample square 2D arrays as cubes through a cube-compatible twin #647: we did not know about your earlier rejection of Fix shrine floor flicker with reusable linear-depth snapshots #690, and in both PRs we did not reproduce the bug.
• Section 1 was our own PRs, rated by us. You decide what goes into the release.

Merged into dev

• Count play time across runs and show it on the Game tab (#720) #729 (play time) through task/playtime, e0f8121. GitHub shows the PR as closed because it went in through your branch.
• Grow the texture heap a mod's HeapDef asks for, so HD texture packs load (#673) #713 (HD texture packs), caf53eb, with your follow-up df9ba65.
• Store the data writes of mods' ExeFS patch files into main before the game runs #835 (IPS data writes of ExeFS patch files), 6cb41a5.
• Grow a content mod's HeapDef when an HD texture pack above it grows the texture heap #858 (HeapDef of a content mod below an HD texture pack), da86c39.

Thanks for all four.

Our open PRs

PR	Base	State
#790 (The Second Quest)	dev	dev merged in (0dfaee5), mergeable. The Second Quest 1.4.3 runs with MSFix (7 hooks) on six saves, and with Link moving on five (runs, with movement). The plugin part waits for your comments on the step 2 design, #845; no code before them.
#732 (achievements)	dev	Your eight points are done (reply): rebased on dev 2327db9, achievements = 0 by default (no thread, no files while off), a live value unlocks only on 3 reads in a row with the world loaded, mem_guest_readable, --probe-stats, one SARC and dictionary reader in romfs.c, menu_test_focus documented, the tab padding kept (the Russian label is shorter). Mergeable. Waiting for your re-review and the 1.2.1 run.
#497 (setup and menu languages)	main	Conflicts with main resolved (60adfbd), mergeable. The 11 setup strings main added (DLSS download, NSZ/XCZ files, no display) are translated into the seven new languages; the 44 menu strings main added fall back to English there for now, as missing entries do in ui_text.c. Still a draft. I can move it onto dev if you want it there.
#367 (working-loop skill for coding agents)	main	Mergeable, unchanged.

Found on 2026-11-06: HD pack above a content mod

With the HDPacks.com 4K pack and The Second Quest both on, in the default order (4kTextures sorts first), the game crashes at start: a write to address 0 in main+0x1d4200. The Second Quest's HeapDef grows RSDB (9728 -> 16000 KiB) and UI (45688 -> 52000 KiB) for its own files. df9ba65 keeps those sizes when the content mod is above the pack; with the pack above it, the pack's table is still swapped for the game's own, so the mod's sizes are lost.

Fix: #858, merged (da86c39). It reproduced on plain dev with the pack and only the mod's romfs folder (no plugin needed). With the fix the pack's growth goes onto the mod's table and the save loads (0 validation errors). The pack alone and the mod alone behave as on dev. --selftest, --probe-cfg and --probe-mods pass.

Others' PRs, not ready by your rules

• Hold the game's dynamic resolution at 100 % and let the scene grow past 4096 wide #573: dynamic resolution is off by default for everyone; our 4080 test says nothing about the target hardware. It also conflicts with dev now.
• Add a perf log summary tool for comparing settings #487: the percentiles come from window means.
• Fix shrine floor flicker with reusable linear-depth snapshots #690 and Sample square 2D arrays as cubes through a cube-compatible twin #647: the bug was not reproduced (repro-before-fix).
• Add PlayStation button prompts #741, Add a standalone controller input monitor on Windows #669, Support the 1.1.1 update (0x30000) #657, Measuring tools for Linux and the Steam Deck: profile_deck.sh, scripted spin and walk runs, run_game.py on Linux #748: features or tooling; no claim from us beyond "works on our machine". Add a standalone controller input monitor on Windows #669 conflicts with dev now.

Closed: #595 (Breath of the Wild) on 2026-11-06. Merged since: #743, a83b3b8.

The dev-next-release branch: only a reference for git diff, as you said. We withdraw the fast-forward suggestion.

#713 test results (merged)

Built on dev e0f8121. --selftest, --probe-cfg and --probe-mods pass. Both runs used the Kakariko night save:

• Without mods: 3759 frames, 0 validation errors. The block is the same 6 GiB as on dev.
• With the HDPacks.com 4K pack: its HeapDef asks for TextureHeap 3925880 KiB, so the heap and pool grow by the difference and the block becomes 6 + 4 = 10 GiB. 4279 frames, 0 validation errors.

[itaskaev2]

The Second Quest status on the split mod PRs, tested today with The Second Quest 1.4.3 (romfs, the ExeFS IPS and the subsdk9 plugin, MSFix) on game 1.4.3. Each PR was rebuilt the way the installer does it (exefs_patches.py, lift with the patches, fold, build_game.py), then started from a Great Sky Island save.

PR	The mod	Run
#790	On. The Second Quest: on, romfs exefs subsdk, ok; MSFix of The Second Quest 1.4.3 runs, 7 hooks on; 1 ExeFS patch files on, 4 bytes of main's data stored; 3301 files served (2427 replaced, 874 added).	The save loads. 6480 frames, 0 validation errors, 0 shader failures, no fault. The first minute after the load has pipeline-build hitches (cold pipeline cache for the mod's content: worst frame 4.6 s, mean 42-50 fps); the next minute runs at 58.6 fps mean, 1% low 44.9 fps.
#835	Off as a whole, as intended. has a code module (subsdk or main), not supported; the whole mod stays off, the same as on dev.	The game runs without the mod: the save loads, 6666 frames, 0 validation errors, 59.4 fps mean, 1% low 47.1 fps.

So with the full mod, only #790 runs it; #835 neither breaks the game nor changes it when the mod brings its plugin.

Not checked: the mod's own quests and the Master Sword behaviour MSFix changes. These runs show only that the mod loads and the game plays on the save.

[itaskaev2]

@valentderah, the checks in docs/building.md that need no game files, run on dev (da86c39), and the PRs that make them green.

Red today, one small PR each:

Check	Why it fails	PR
tests/settings_doc_check.py	PROBE_MOVIE_BENCH (3cd2410) has no row in docs/settings.md	#859
tests/worktree_test.py	the test's scratch .gitignore lists only LINK_FILES[0], so CLAUDE.local.md shows in git status	#860
tools/check_platform.py	11 Win32 uses in vk_dlaa.c; the NGX loader now goes through runtime/platform/ (DLAA tested on an RTX 4080 SUPER)	#861
totk.exe --probe-services	the Speakers row (963901c) lacks pt-BR and zh-Hans	#862
totk.exe --probe-services	the probe expects English when the Windows UI language has a region (ru-RU); the menu correctly shows ru. Passes on en-US, so you would not see it	#863
tests/excl_mt.py	the ABA rows: the exclusive monitor is a compare-and-swap	#868

Older PRs, carried onto dev with their authors' commits unchanged (each original has a comment pointing to the carry):

• Give the exclusive monitor a store generation, so a store that puts the value back fails STXR (#674 by zzs181, on dev) #868 = Give the exclusive monitor a store generation, so a store that puts the value back fails STXR #674 by zzs181 (store generation). Our commit on top fixes a link error it caused in tests/diff. diff/func (889 functions), diff/lift, diff/atomic and diff/simd: 0 failing. A 12-run A/B in game shows no speed difference within the run-to-run spread: 112.2 fps on dev, 112.4 fps with it.
• Add tests/run_all.py: one entry point for the tests (#675 by zzs181, on dev) #864 = Add tests/run_all.py: one entry point for the tests, each with its interpreter and needs #675 by zzs181 (tests/run_all.py). Our commit adds the tests that came after it and tests/diff. On dev: 10 pass, 3 fail, the three above.
• Compile the runtime on both hosts in CI, and build the AppImage in a pinned container (#656 by jkoehler11, on dev) #865 = Compile the runtime on both hosts in CI, and build the release AppImage in a pinned container #656 by jkoehler11 (compile gate on both hosts, pinned AppImage container). Our commit fixes the gate under Windows' own Python. Both of its workflows already passed on Compile the runtime on both hosts in CI, and build the AppImage in a pinned container (#656 by jkoehler11, on dev) #865 here: Windows in 3m15s, Linux in 8m58s.

With all of them in, tests/run_all.py is green without game files, and #865's workflows compile the runtime on every PR. Once that holds, two settings only you can change:

1. make those two workflows required checks for dev;
2. switch Compile the runtime on both hosts in CI, and build the AppImage in a pinned container (#656 by jkoehler11, on dev) #865's test step to tests/run_all.py and drop the continue-on-error on excl_mt.

Each PR stands alone. The only interaction: after #861 and #865 are both in, check_platform.py prints a note that the vk_dlaa.c allowlist entry from #865 can go. It is a note, not a failure.

[itaskaev2]

Linux results, run locally after the comment above: Docker 28.0.4, the image from #865's tools/linux/Dockerfile (Ubuntu 24.04), the source in a Docker volume.

What	Result
#865's Linux workflow, step by step	the pinned toolchain (LLVM 23.1.4, sniper sysroot, runtime deps) builds in 327 s; check_platform.py ok; compile_gate.py --undefined: 116 files, 0 failed, 4 skipped (game headers); flags_check, diff/simd (159976 cases), diff/atomic (12672 cases): 0 failing; excl_mt: the ABA rows fail, as on dev (continue-on-error)
#861 merged with #865, through the same gate	ok runtime/vk_dlaa.c, ok runtime/platform/linux.c, 116 files, 0 failed, every plat_* symbol defined; 0 Win32 uses left in vk_dlaa.c
#868 on Linux	excl_mt PASSED: ABA rows at status 1 at all three widths, contention exact; diff/atomic 0 failing

So the Linux side of #861 is now compiled with the release toolchain, not only in pieces, and #868 fixes excl_mt on both hosts. Details: #865 (comment 6693679063), #861 (6693679568), #868 (6693680097).

A bind mount of a Windows checkout fails at the sysroot: NTFS is case-insensitive, and the sniper SDK has both usr/include/FL and Fl. A Docker volume or the runner's ext4 has no such problem.

[itaskaev2]

@valentderah, an update on this list, and the PRs added since.

Merged this morning: #859, #860, #861, #862 and #863, thank you. #864, #865 and #868 are still open as described above.

New: a fix for the Intel crash at start, #911 (fixes #681, likely #774).

• Cause. The first pipeline cache autosave calls vkGetPipelineCacheData while the stage library prebuild still builds into the same cache. Intel's driver faults there: igvk64.dll+0x15db72, reading 0x170.
• When it shows. Only with a cold driver shader cache. With a warm one, the prebuild ends before the first save.
• The fix. A reader-writer lock keeps builds apart from the cache copy only. The file write stays outside it.
• Tests on an Iris Xe laptop, with the driver cache deleted before every run:
  • without the fix: 0.3.12 crashed 6 of 6, dev 2 of 2;
  • with it: 0 of 6 and 0 of 3.
  • On an RTX 4080 SUPER: same frame rate, and the cache is still saved.

Older PRs carried onto dev with their authors' commits:

Carry	Original	What we added
#913	#638 by STS78721 (Reflex, today's rebase ce44de3, unchanged)	1. With Reflex on, NVIDIA stopped at the first recorded submission: rec_submit refuses a pNext. The recorder now carries the present ID. 2. vkLatencySleepNV ran without the present lock: 20 validation threading errors, now 0.
#914 (draft)	#228 by Lucas-blip-png (FSR 3.1, 553 commits behind)	The merge into dev's new vk_dlaa.c. The DLSS modes now shrink the scene for FSR without nvngx_dlss.dll. vk_fsr3.c no longer includes windows.h. Settings docs and translations. The square blocks of the old base are gone. Open: FSR 3.1 loses the device on an RTX 4080 SUPER, on #228's own head too.
#915	#624 by ipavlin98 (SMAA, against main)	conflicts only
#916	#573 by ProSep (dynamic resolution held at 100 %)	the cfg.c list conflict only
#917	#676 by ProSep (sync ahead)	the cfg.c list conflict only

Each PR has its numbers from the laptop, and from the 4080 where it matters. CI is green on all six. As before, each one stands alone. If an author rebases the original, the carry can be closed.

[itaskaev2]

Closing: everything merged from this plan is in v0.3.13 (#729, #713, #835, #858, #859-#863, #911). The PRs still open, with their tests and a suggested order, are in the new plan: #965.