Describe changes

Updates @tenphi/tasty from 3.3.1 to 3.5.0 (3.4.0 + 3.5.0). No changes were needed in the kit — the two releases are additive, and the new holdKeyframes() / ownKeyframes() reach consumers automatically through the existing export * from '@tenphi/tasty' in src/index.ts.

Two things matter to apps built on the kit:

Injected CSS is garbage-collected again. Once Tasty's render path became hook-free it stopped disposing the classes it injected, so reference counts never fell back to zero and gc() / cleanup() / tastyDebug.cleanup() all read a non-zero count as "still in use". Nothing was ever evicted, and a long-lived page accumulated injected rules for its entire lifetime — the failure mode is a session that gets slower the longer it stays open, worst in a UI that mounts many short-lived styled subtrees (dialogs, popovers, table cells). What a style is worth keeping is now decided by the DOM. touch() and StyleUsage are deprecated but still exported, and tastyDebug no longer reports Unused: 0 classes for a page holding hundreds of stale rules.

Local @keyframes are content-addressed. A @keyframes block written inside a styles object is now emitted as fade-<hash of its steps>, resolved identically by the client, the SSR collector and the RSC pass. Two components authoring animation: fade 1s over different steps get two rules instead of one definition silently winning — which previously also let server and client disagree about which class to use. Three components here use local @keyframes (LoadingAnimation, Spin/Cube, SpinsContainer); all render unchanged. keyframes()-registered global animations are unaffected and keep their authored names.

3.4.0 is pure performance: LRU cache reads are 1.79x faster in their steady-state shape, and a repeat SSR render of one component is 1.8x faster end to end now that tasty() factories memoize their chunk cache keys.

Verification

Merged with main (25 commits) — the only conflict was pnpm-lock.yaml, resolved by regenerating from main's lockfile rather than hand-merging, then re-pinning tasty to 3.5.0. The diff against main is 4 files: package.json, pnpm-lock.yaml, the changeset, and the size note. Main's @tenphi/eslint-plugin-tasty 1.0.5 is preserved and re-pointed at tasty 3.5.0.

All re-run after the merge:

• pnpm test — 102 files, 2202 passed / 1 skipped
• pnpm test:browser — 12 files, 141 passed
• pnpm lint — clean (only the pre-existing consistent-token-usage warnings), Prettier clean
• tsc --noEmit — byte-identical to main: the same 18 pre-existing errors, no new ones

Worth noting for reviewers: two intermediate full-suite runs failed 17 and 2 tests respectively, in different files each time (Form, then ItemTable), while the machine was also running builds. Four runs total, two fully clean, and the failure sets never repeated — timing flake under load, not a regression. main in a clean worktree passes the same suite, and the merged branch actually runs faster than main (45.6s vs 58.8s back-to-back on a quiet machine), consistent with 3.4.0's LRU work.

Main now sets batchInjection: true and opens TastyBatchProvider windows in Root, Portal and Overlay (#1344), so this bump also picks up 3.5.0's fix for tastyDebug reads taken inside a batch window reporting empty.

Size, both sides freshly rebuilt on the same machine, against the new main:

Both stay inside budget, so no limit is raised. Matching deltas on both entries are the signature of a change inside Tasty's always-included core.

All is now down to 1.40 kB, the thinnest it has been — but most of that was spent before this bump: main climbed 508.95 → 512.39 kB on its own (+3.44 kB of Board chrome, tokens/resolve, useScheme, TextInput autoComplete and the batching wiring), leaving 2.61 kB before tasty was touched. The next change here will likely need to raise it.

One calibration note, since .size-limit.cjs warns that macOS reads ~1.6 kB lower than the Linux runner: on this machine it did not. CI reported main at 500.38 KB / 118.96 KB binary, which is 512.39 kB / 121.81 kB decimal, and the local rebuild of the same commit read 512.39 kB / 121.81 kB — an exact match on both entries. CI has since measured this branch too, at 501.56 KB / 120.13 KB binary — 513.60 kB / 123.02 kB, matching the local readings the budgets were checked against. Assuming the 1.6 kB gap would have raised this budget for no reason. I recorded that in the file without retiring the warning.

Checklist

• Pipeline is passed — all checks green, including Measure size of the library
• Tests are passed successfully
• Changeset(s) is(are) added
• You have passed the threshold of the library size
• Commit message follows commit guidelines

Closes: N/A

Other information

Local @keyframes naming is the one behaviour change worth a second look at review time: any CSS or test that refers to a local keyframes name literally would no longer resolve. Nothing in this repo does, and Chromatic covers the three animated components visually.

Tasty 3.6.0 was published mid-review (2026-08-28 17:29), after this branch was measured. This PR deliberately stays on 3.5.0, which is what was asked for and what every number above was measured against. A plain pnpm install will resolve the ^3.5.0 range up to 3.6.0, so the lockfile pin is what holds it — worth a follow-up rather than an in-place retarget of this branch.

🤖 Generated with Claude Code