Skip to content

Release v0.126.0 - #692

Merged
leonardoaraujosantos merged 4 commits into
mainfrom
release/v0.126.0
Oct 6, 2026
Merged

leonardoaraujosantos merged 4 commits into
mainfrom
release/v0.126.0

Conversation

@leonardoaraujosantos

Copy link
Copy Markdown
Contributor

claycore v0.126.0

The release branch for v0.126.0: twenty-two PRs since v0.120.1 (#655, #667–#669, #673–#689, #691), ABI 0.121.0 through 0.126.0, scene and .clayspace format minor 19 → 20 (#673). The notes are docs/release-notes/v0.126.0.md; the history entry is in docs/RELEASE.md. Engine paths are identical to main at abdd1d49; the branch adds only the notes, the history entry and the device gate stamp.

The device gate

Passed at 7049e19c on the reference iPad15,5, iOS 27.0.1 (24A446): 75 cases, seven cold sessions with 1800 s cooldowns after a thermal probe, thermal state nominal at both ends of every case, treeDirty: false, canary steady at x1.18. Scored against the baseline #690 re-recorded from v0.120.1's engine on 27.0.1, so REGRESSION compares v0.120.1 to v0.126.0 on the same OS: median ratio 0.996 over 75 cases, none above both the 1.4 tolerance and the 0.125 ms floor, largest rise 1.04x.

The first run, at 7406eb1d before #691, failed: mask_extrude x4.97 (26992 ms p95 at 1000 stamps against an 8154 ms budget) and mask_extract x6.26, both through #667's extrude. With #691 they read 1.01x and 1.04x (mask_extrude 52.4 / 377.9 / 3680 ms at 10 / 100 / 1000 stamps against 51.6 / 374.5 / 5297). The stroke cases read 0.81–0.89x, named in the notes as a candidate for #688 and not a claim.

The run needed the iPad's Wi-Fi off and the device under an air conditioner; with Wi-Fi on, post-update on-device inference heat-killed four sessions 36–80 s in before any run completed. The gate skill (merged in #690) records that and the preflights that catch it.

Release checklist

python3 tools/release_check.py at fe97a210: every row passes — version 0.126.0 in cmake, ABI and wheel; configure, build, tests; parity (cpu, 48 cases, 1,411,877 assertions); layering; dialect; licenses; task-symbols; bindings (773 capabilities); kernels; abi; openspec; device (75 cases on iPad15,5 at 7049e19); the four hardware gates carried as waivers at 4401b354 (the only kernel-relevant change since 59e42ccf is include/clay/eval/bake_volume.h, #680, no kernel arithmetic); benchmarks; wheel quickstart. "all gates passed — safe to tag".

After merge

Tag v0.126.0 on main (git tag -a v0.126.0 -m "claycore v0.126.0"); the release workflow drafts the GitHub release with the wheels, kernels and xcframework attached and generated notes. Replace the draft's body with docs/release-notes/v0.126.0.md before publishing.

leonardoaraujosantos and others added 4 commits October 6, 2026 01:49
Covers the twenty PRs since v0.120.1 (#655, #667-#669, #673-#688), ABI
minors 0.121.0 through 0.126.0, and the scene format minor 19 -> 20.
The device gate block is a placeholder until the iPad run is recorded.

Carries the four manual hardware gates forward as waivers at 4401b35:
the only kernel-relevant change since 59e42cc is #680's per-brick cull
band in include/clay/eval/bake_volume.h, which changes what a brick
compiles, not kernel arithmetic, and the parity corpus is unchanged.
The multires stroke-session consumer now takes a clay_multires_delta,
so its known limit is gone; the at-a-glance table, the session section,
Upgrading and the RELEASE.md entry say so.
The first device gate failed on mask_extrude and mask_extract at 5-7x,
caused by #667's per-sample gradient; #691 restored the cost. The notes
record the failed run, the cause, the fix and its small value change.
The gate passed at 7049e19 against the iOS 27.0.1 baseline #690 recorded
from v0.120.1's engine, so REGRESSION compares v0.120.1 to v0.126.0 on the
same OS: iPad15,5, 75 cases, seven cold sessions with 1800 s cooldowns,
nominal throughout, canary steady at x1.18, median ratio 0.996, none above
both the 1.4 tolerance and the 0.125 ms floor, largest rise 1.04x.

The first run at 7406eb1, before #691, failed on mask_extrude (x4.97,
26992 ms at 1000 stamps against an 8154 ms budget) and mask_extract (x6.26);
with #691 they read 1.01x and 1.04x. The stroke cases read 0.81-0.89x,
named as a candidate for #688 and not a claim. The run needed the iPad's
Wi-Fi off and an air conditioner; the notes say why.
@leonardoaraujosantos
leonardoaraujosantos merged commit 1946ac2 into main Oct 6, 2026
16 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant