4 min readrelease
Six updates, one pixel diff: shipping a dependency round you can trust
How Unown updates packages without regressions: staged gates, an A/B pixel diff of the landing page, and the word budget the changelog almost broke.
- #release
- #dependencies
- #verification
v0.0.54 is a maintenance release: six package updates, nothing new on the screen. That is exactly why it deserves a post — "safely update everything, prove nothing regressed" is the kind of task that sounds boring and quietly isn't, because the two honest ways to fail it are not updating anything (drift) and updating everything at once (an unattributable breakage). This round did neither. Here is the shape of it.
The policy did the first half of the work
Unown's update rules were already written down, so the round started from
constraints instead of vibes: no semver-major bumps (humans opt into
majors), exact pins move as deliberate exact bumps, the playwright
trio moves in lockstep, and the vendored pi-repo/ workspace is an
upstream-sync decision — frozen, not a routine bump surface.
A sweep of all 28 root packages against the registry found the truth Dependabot-style updates hide: only six updates existed within policy. next 16.4.2 → 16.4.3 and eslint-config-next → 16.4.3 (LTS backports), marked 18.1.1 → 18.1.2 (CommonMark fixes), then three exact-pin bumps — eslint 10.13.0 → 10.14.0, lucide-react 1.61.0 → 1.66.0, framer-motion 14.6.0 → 14.9.0. Everything else is at latest. The only major available anywhere is TypeScript 7, the native port — parked, by policy, for a human with a spare afternoon.
Every stage got its own gate
The updates landed in four stages, and after each one: typecheck, lint,
the 2063-test unit suite, a production build, the 15-test e2e lane, and a
bun audit --prod pass. A regression that appears after stage four is a
regression you can name. The engine suite (5326 tests) and the bridge suite
(243) never re-ran against the updates — not out of laziness, but because
the round touches no pi-repo/ manifest or lockfile: their dependency
graph cannot move. Knowing what a change cannot affect is half of
verification.
Mark got the sharpest gate: it renders every blog post, so the blog data was regenerated under it — byte-identical, all 37 posts and the feed.
The part you can't unit test: motion
framer-motion 14.6 → 14.9 is the one update that changes what the page
does, not just what it links against — 14.7 moved background-color
animations to the main thread and touched AnimatePresence positioning, and
the landing page's reveals ride exactly those APIs. No unit test sees that.
So the round took the only measurement that can:
- The landing page's own screenshot lane shot all 22 sections — both colour
schemes, plus the
prefers-reduced-motionlane — against the production build. Twice on the new tree (a noise floor), once on the old tree (rebuilt from the pre-round lockfile). - A per-pixel diff compared the passes.
The result: the reduced-motion lane is pixel-identical between the old and new dependency trees — three shots, zero differing pixels. 21 of 25 full shots identical; the four differing frames differ only in the live star-count and animated-terminal regions that a same-build double-pass had already shown to be dynamic content. Settled rendering did not move.
One honesty note worth keeping: a VLM design reviewer — our own Agnes lane — had been briefed to hunt for motion regressions and confidently reported several, including "half-finished reveals" in the reduced-motion frames, where nothing animates at all. Pixels beat opinions. Brief your reviewers neutrally, then measure.
The catch that never left the machine
The round's one real regression almost shipped — and it wasn't a dependency. Folding the changelog for release regenerated the landing rail, and the new release stop's chrome pushed the page to 3,930 words against a 3,920-word scroll budget. The e2e budget gate caught it locally, before the release commit left the machine. The fix followed the landing contract's precedent — and taught it something new: the sanctioned first move (compressing the round's own listed changelog title) turned out to be a page-level no-op, because the rail only prints titles for Added and Behavior entries, and this round's entries were all Changed and Fixed. The stop costs its chrome and nothing else — so the twelfth ceiling amendment (3,920 → 3,950) carries exactly that, plus one stop of headroom, documented in all three places the budget lives. The breach never reached CI, let alone the deploy.
Receipts
The full record — what moved, what stayed and why, every gate, the A/B numbers — lives in the research record and the session log; the release stop is in the changelog. The next round is already queued — the settings ledger's spend meter and memory-health card, each with a waiting benchmark to flip. But this one ends the way a maintenance release should: nothing on the screen changed, and we can prove it to the pixel.
Spotted something wrong or worth adding? The record is the repo — open an issue.