[bmjask]

This draft preserves the combined playable branch codex/ocarina-playability at 26748913090b705421f0719b35a0d17426c22485. It remains an integration checkpoint; changes are being extracted into independently reviewable PRs against current main. The installed Mac app is not replaced by these branch operations.

Independent review PRs

• NVN: opt-in bounded compute dispatch ordering #1887 — bounded compute dispatch ordering, opt-in. Focused Metal-validation tests: 6,986 checks in each encoder mode, plus documentation checks.

• NVN: preserve indirect conversions across disjoint mesh writes #1888 — preserve indirect cache entries across proven-disjoint mesh writes. Adapted to main's native mesh conversion rather than restoring the older implementation. Focused Metal-validation tests: 22 checks, plus documentation checks.

• Experimental opt-in fractional timing for Ocarina high-FPS play #1890 — opt-in build-gated fractional timing, with contract 2 moved away from upstream dynamic resolution.

• Extract conservative raw-load grouping and high-product shader lowering #1891 — conservative shader lowering: legacy raw-load grouping and exact unsigned high products. The broader pointer-site/mesh optimization remains separate.

All four split PRs (#1887, #1888, #1890 and #1891) are ready for review. Full regression on the extracted current-main versions remains pending. Their descriptions distinguish historical combined-branch A/B evidence from tests run on the extracted commits. No current-main performance gain is claimed from historical measurements alone.

The save editor is excluded from this PR-splitting effort. It remains in the historical combined branch; no separate save-editor PR is being submitted.

Remaining extraction

Experimental high-FPS timing is now extracted in #1890 with contract 2 at +0x44 and fresh focused tests; gameplay validation of that adaptation remains pending. #1891 extracts the conservative shader passes, not the integration branch's broader pointer-site cache ABI, mesh grouping or every GPU experiment. Those remaining changes still need overlap/dependency review and independent validation.

Guest-clock timestamp projection is already upstream (267ef2d3, PR #1676), as are several archive, rendering and shader fixes. Those should not be duplicated in new PRs.

Historical evidence

The following report describes an older checkpoint, not the current branch tip or the newly extracted PRs. It is retained for provenance. The older chain ID, validation counts, exclusions and packaging requirements below apply only to that checkpoint.

Historical checkpoint report (64a54c9 / ff5c23c)

Ocarina of Time (0400486024DEC000) needs host-service, rendering and shader-translation changes to run on the Mac. This draft checkpoints the playability branch, including its game profile, performance caches, experimental save editor, focused tests and recorded validation limits.

The current checkpoint is 64a54c97 (runtime checkpoint ff5c23c8 plus fractional-clock tests). It includes the measured mesh/compute grouped-load improvements, restored mesh-pair coverage, guest-clock projection fixes, and experimental build-gated fractional timing used by the separate HighFPS package. Build shaders and host together for chain ad4eab41ef63; rewrite the supported game module for the fractional timing contract. High-rate mode is opt-in through DM_FPS_TARGET=120; the normal target remains unchanged.

Measured behavior

Two short 60 FPS forest comparisons showed 5.8% and 5.5% lower median GPU time from the ordinary-compute grouped-load extension, following the earlier mesh improvement. A separate combined timing/compute run averaged 93.7 FPS under a 120 FPS request. An instrumented short forest run measured simulation speed within 0.06% of the 60 FPS reference. These are route-specific results, not stable 120 FPS, whole-game playability, or broad audio/movie validation. The game note records conditions and limitations.

Validation

• Matching Mac host rebuild passed. Local checks passed: 264 NVN checks, 84 frame-step checks, six ARM64 stub execution tests, 17 shader-chain tests, 11 documentation tests, six grouped-load tests and 34 mesh-wrapper tests. Later local checks also passed five dynamic-helper and 20 reflection tests.
• The current static regression comparison against ad2b936b has 42,850/42,850 successful shaders in both revisions, no added or newly failing shaders, no increased loss annotations, and no interface failures across 34,474 checks. Expected changes: 2,649 MSL sources, 2,234 reflection records and 1,406 variant records.
• Full scripts/regress.py check completed with CHANGED, exit 0, for runtime checkpoint ff5c23c8 against ad2b936b. Both revisions reached the route frame gate, with zero new REDs, zero pipeline-prebuild failures and no new shader/interface failures. Both measured 59.5 FPS at the 60 FPS request and 18.0 ms frame p99. Candidate preparation took 14.29 s versus 10.87 s; this stayed within the harness tolerance. The subsequent 64a54c97 change adds tests only. This validates the short forest route, not the whole game or 120 FPS.
• Ocarina is the only locally available game. The general NVN changes can reach other games, which remain unchecked. iOS full-game/device validation is also pending.
• The separate experimental HighFPS package passed signature verification and a copied-save startup test. Existing installed-app and personal-save hashes were preserved.

Work not included

Constant-metadata address spaces, fused task/mesh initialization, upstream extended multiplication (#1741), archive-period rollover, and task/mesh render-pass ordering remain isolated experiments. They are not part of this pushed checkpoint. The installed HighFPS app is not automatically replaced by branch pushes.

Front-ends: recognize title 0400486024DEC000; rebuild shaders and host together for the declared chain, rewrite supported modules when enabling the fractional timing contract, and retain the normal per-title save directory. Treat the 120 FPS request as experimental, not a guaranteed frame rate.

Full local regression report

Ocarina was selected because it is the only locally available game and directly exercises the changed NVN/mesh/compute paths. CHANGED is expected from the grouped-load and restored shader-coverage changes. Candidate screenshots at 70 and 78 seconds were inspected: the forest, Link and HUD render intact; animation differences remain, so this is not pixel-identical rendering proof. The baseline was repeated after a startup covered-window event; the accepted retry had no covered-window event.

regress.py check (full): CHANGED

Baseline ad2b936ba83ca03ab924b845326bf2012b458b37 ad2b936, candidate ff5c23c8e8983708fe09efc48e7cd9f3b8d396f6 ff5c23c; Apple M4 Max, macOS 27.2, Xcode 27.2; 4083.3 s

game	status	shaders (base → cand)	newly failing	newly building	changed MSL	interface (new / known)	boot	fps p50	frame p99 ms	peak MB	prepare s	new RED
OcarinaOfTime 0400486024DEC000	CHANGED	42850/42850 → 42850/42850	0	0	2649	0 / 0	ok	59.5 → 59.5	18.0 → 18.0	9975 → 9917	10.87 → 14.29	0

OcarinaOfTime 0400486024DEC000: CHANGED

• CHANGED: 2649 shaders changed (MSL 2649, reflection 2234, variants 1406)
  • changed shaders: 001ef25305e1bd7e, 001f62d77a5447e6, 0031f9442b11b682, 00359b6009aed142, 003f7645f98dae69, 003fa45149288240, 0041cb1d10d87b8c, 005abfb28692a48c, 00ac145f91deb2c7, 00b0635ae2b9613b, 00bb2a1e99a98106, 00fa422e4f968603, 012ad2c75b59f3d9, 01780967ff13361f, 0193bf98ac2101db, 0195afef02530406, 01b4c2e8ffdf5893, 01cee45f33ef14ad, 01d03d59b3855728, 01eaf34a114e432f, 0207d3440d0e8dda, 020fdeffc388ba6d, 023f5bbaaeeff52f, 026942f6f9659042, 028fdf520010c667, 02a33fe918156109, 02f639fd402d35ba, 0308e88978d44ed3, 0316d96ea530cf26, 0342ab5e045200ee, 039de553a54efefc, 03a0adf10e22ddc9, 03a2a106a130e659, 03b4b2bd24119d23, 03e1d6d1364b6f66, 03eac1c8abe5bd7b, 04033d299ca5cfe1, 040bde5e6e671e68, 0414ac23d09a0d71, 04344d0b3b6c2846 …
  • screenshots (s: dHash bits / RGB diff / baseline noise): 70: 4/2.09/0, 78: 4/1.67/0
  • baseline: exit None, first frame 12.8 s, max frame 4598, reached True, prepare 10.87 s, prebuild 573.5 s (30515 built, 0 failed), fps 59.5, p50/p99 16.64/18.0 ms, peak 9975 MB
  • candidate: exit None, first frame 17.3 s, max frame 4329, reached True, prepare 14.29 s, prebuild 688.3 s (30515 built, 0 failed), fps 59.5, p50/p99 16.64/18.0 ms, peak 9917 MB
  • time: chain 183.2 s, smoke 3899.7 s

[Anatharias]

there seems to be a problem with shader tables :

[RED] the split shader tables differ from game profile 0400486024DEC000's: shader_map.tsv: 4,710 row(s) only in the profile, 4,707 only in the split profile: ps_lib.bin 0 pixel 269e3cb1372760f6 /Lib/Terrain/tera_resource.Nin_Ounce_NVN2.release.sarc.zs/tera_common.sharcb#BAHS@0x10038 split: ps_lib.bin 0 pixel 4ad910dcbc2de272 /Lib/Terrain/tera_resource.Nin_Ounce_NVN2.release.sarc.zs/tera_common.sharcb#BAHS@0x101d0 pso_pairs.tsv: 0 row(s) only in the profile, 175 only in the split split: 06f14fa278dddec5 00cbadd92d5092f3 Either these game files are not the versions the profile registers, or shader_split.py now reads them differently and the profile's tables were not split again (the split's tables: /Users/xxxx/Games/SWITCH2/0400486024DEC000_build/shader/pkg.partial/tables)