Release v0.120.1

[leonardoaraujosantos]

Cuts v0.120.1, a patch on top of v0.120.0 (published 2026-09-18, the latest published release). It covers the eleven PRs merged since that tag: #635–#638, #645–#648, #651–#653.

Why a patch

No C ABI symbol was added or removed, and no struct was re-laid out: 708 clay_*( symbols on both sides, and every - line in the clay.h diff is the CLAY_ABI_PATCH define or a comment. That is what moves this repo's minor. pyclay did gain one keyword argument (layer_height on the multires and sculpt-layer Layer brushes), which the Python API's SemVer would call a minor; the notes say so and follow the 0.24.2 precedent. The version moved in all three files in 7f6cf38f.

What lands

• docs/release-notes/v0.120.1.md. It opens with a one-row-per-PR summary of everything since v0.120.0, then a section per change, What moves under you (formats, then entry points), the ABI diff, the gates, Known limits (including everything carried from v0.120.0), and Upgrading by risk.
• docs/RELEASE.md: the 0.120.1 history entry.
• tests/device/last-gate.json: this release's device gate.
• A correction to v0.120.0's notes. Their gates section was written before the gates ran and was never updated: it said all gates were stale, when as tagged the hardware gates were waived and the device gate passed on iOS 27.0. It is corrected with a dated note above the original text. The published v0.120.0 release body is updated to match; its tag and artifacts are unchanged.
• make-voxel-sculpt-layer-creation-a-step (Make a voxel sculpt layer's creation, and each edit inside it, an undo step #651) archived: all 11 tasks were ticked, but it was left unarchived. The roadmap header is recounted to 21 capabilities, 240 archived, 41 open.

The headline changes, for a caller who changes nothing

• Undoing one Move segment reports the grab, not its node (#639) #648: undoing one Move segment refills the grab's reach, not its whole node: 1,000 / 1,440 / 4,000 → 12 bricks at 1 / 10 / 40 grabs. This is the undo cost ClaySpaceDesktop reported in clay_document_undo_bound reports a grab's whole node, so undoing a Move refills the node and its cost grows with the chain #639.
• An unconfined lattice reports an unbounded tape.bounds (#640) #645: an infinite repeat grid reports unbounded tape.bounds. Meshing, rasterizing, baking, pyclay's Volume constructors and clay_sdf_smooth_begin now refuse one without a region, instead of silently dropping material (5,824 of 5,824 samples were outside the box).
• Narrow tape bounds per operator, gated through the C ABI; finish fold-the-layers-with-an-operator #637: bounds are narrowed per operator. A mesh sized by resolution= gets finer at the same resolution; nothing is lost at a fixed voxel size.
• Bound an edit by the smooth combines after it (#650) #653: an edit's bound widens behind a smooth sibling (25 → 0 stale bricks).
• Keep a grab the deformer cull cannot judge in the brick's tape (#649) #652: a full brick build keeps a grab the deformer cull cannot judge.
• Finish unify-the-undo-history: voxel sculpt-layer operations are undo steps #647 and Make a voxel sculpt layer's creation, and each edit inside it, an undo step #651: voxel sculpt-layer operations, layer creation and edits inside a layer are now undo steps.
• Rebind a multires sculptor on any renumbering, and gate the Layer ceiling on the host's path (#627) #636 and Average the layered smooths over the whole rim ring, and give SculptLayerStroke.stamp a layer_height #646: multires rebind fixes (one reachable through the C ABI), whole-ring rim smoothing, and pyclay's Layer brushes moving again.
• The crash journal is the one format change: v0.120.0 cannot replay a v0.120.1 journal that recorded a sculpt-layer operation.

Verification

release_check.py --skip-slow: every row PASS, exit 0.

• bindings checked against a real built pyclay (766 capabilities).
• dialect covers the cpu, cuda and metal profiles.
• parity compared cpu only; cuda, opencl and vulkan were not built.

Device gate: PASSED, and this time against a real baseline.

• Run: iPad15,5 on iOS 27.0 (24A437), claycore 7f6cf38f, 75 cases, seven cold sessions, thermal state nominal at both ends, no RunningBoard kill lines.
• It is the first same-OS comparison since the iOS 27.0 re-baseline, so REGRESSION could fire. v0.120.0 had no such check.
• Median ratio 0.998 over 75 cases; no case above both the 1.4 tolerance and the 0.125 ms floor.
• The canary drifted x1.65 while thermalState read nominal, so sdf_flatten and sdf_relax were measured warm. Their budgets held.

The four GPU hardware gates are carried as waivers at 59e42ccf. Nothing they depend on changed since then: git diff over the full kernel-relevant set is empty. No GPU backend ran against this tree.

Closes #634: stroke_build reads 0.324 ms against the baseline's 0.488 (0.66x), within 1.09x of its iOS 26.5.2 figure. So the 1.64x at v0.120.0 neither persisted nor came from the OS. Whether that reading was high or something in this range recovered it, a one-shot gallery case cannot say; the notes state the ambiguity rather than resolve it.

After merge

Tag v0.120.1 on the merge commit. The release workflow builds the wheels, the xcframework with its SwiftPM checksum and the kernels artifact, and opens a draft. The draft's body will be replaced with docs/release-notes/v0.120.1.md, the artifacts checked, and the release published.