fix(Countdown): stop the ring rendering NaN and negative dash lengths, and make twenty-two snapshot cases assert something

[IgorShevchik]

Linked issue

Resolves #454

Follow-ups filed: #477 (Calendar prop leak), #479 (Table has no aria-sort).

Type of change

• Documentation (updates to the documentation or readme)
• Bug fix (a non-breaking change that fixes an issue)
• Enhancement (improving an existing functionality)
• New feature (a non-breaking change that adds functionality)
• Chore (updates to the build process or auxiliary tools and libraries)
• Revert (undoing a merged change — retitle this PR revert(Scope): ...)
• Breaking change (fix or feature that would cause existing functionality to change)

Description

#454 describes a spec case whose fixture never satisfies its precondition: the snapshot records the fallback, the case reads as coverage, and nothing fails when it stops being true. It lists fourteen.

Two things changed the shape of the work before any of them were touched.

Measuring first. The issue suggests a guard that groups snapshot entries by identical body. Building it first, as a measurement, found 916 of 4361 entries in 301 groups — a fifth of the corpus — against the fourteen named. So the fourteen are a hand-picked sample, and "fix them all" was never the task.

Checking upstream first, at the maintainer's instruction: our components we analyse ourselves; upstream's we check against upstream. That turned out to matter more than anything else here.

Twenty-two cases with one symptom and seven different causes

cause	count
inherited — upstream's spec has the identical vacuous case	13	fixed here, diverging from upstream's tests
unreachable in happy-dom	1	removed, with the reason recorded
a component defect	1	Countdown
a recorded divergence	1	Empty
a slot that does not exist	1	DescriptionList
a dead array	1	Table actions
a second-order fixture defect	1	Table sorting
reachable, but not from a declarative matrix	1	ChatMessages viewport
genuinely just a wrong fixture	2

Thirteen of twenty-two would have been misdiagnosed as test bugs and "fixed" by adjusting fixtures. Including, most importantly, the one where the component is broken.

The component defect

Countdown's guard checked seconds < 0 but not === 0 — and 0 is the default. So <B24Countdown use-circle />, the minimal usage from the docs, computed 0 / 0 and rendered stroke-dasharray="NaN 283", an invalid SVG value that was the only dasharray in either snapshot file: no snapshot had ever rendered a valid ring. seconds is typed number | string, so any non-numeric string reached the same place.

Review then found two more things in the same eight lines, both of which the first attempt got wrong:

• The total < 0 guard is load-bearing, not decorative as an earlier revision of this branch claimed. That claim rested on trying -1, which reaches 1 unaided — as does every whole number, since totalSeconds floors and the fraction is Math.floor(total) / total. Fractions do not: without the guard seconds="-0.5" renders an empty ring and seconds="-0.2" renders stroke-dasharray="-4245 283". The vacuous test had been deleted and a comment left inviting the next reader to delete the guard.
• The formula overshoots at the end of a normal countdown. It subtracts a tick's worth of arc so the ring keeps step with the digits, so at totalSeconds === 0 it yields -1 / total and the last frame rendered stroke-dasharray="-28 283". Clamped to [0, 1].

Every one of these has a case that fails when its guard is removed, verified by removing it.

Where upstream was the answer, and where it wasn't

Empty's ['with avatar', …] looked like a dropped prop — upstream's Empty does carry avatar?: AvatarProps and renders it through UAvatar. It is not: the ledger entry for 86cd25c5 says "b24ui's Empty diverges … no avatar". Removed, citing the record. Without checking, deleting the case would have erased the only evidence of a missing feature had it been one.

DescriptionList is ours, so there was nobody to ask. First attempt passed items on the theory the slot renders per item; the snapshot did not move. DescriptionListSlots declares ten slots and no default at all — Vue was dropping it silently.

Thirteen others — Alert's close, CommandPalette's empty, Progress's status, Tabs's custom, ContextMenu's, NavigationMenu's, Stepper's valueKey, the five descriptionKey cases, Table's expanded — have byte-identical cases in upstream's own specs. Not porting slips; fixtures that now improve on upstream's.

The Table cluster, where reaching the branch was not enough

The actions menu never opened, so every label in the column's items array appeared zero times — the whole array was dead while looking like coverage. open: true, portal: false takes it to 40.

Sorting was unreachable in all 35 entries. sorting is a defineModel, so two cases hand the state in. And that alone would not have been enough: the fixture gave asc and unsorted the same icon, so even a sorted case could only ever have distinguished desc. Three icons now, verified by comparing the three snapshot bodies pairwise rather than by assuming.

Virtualization plus row pinning is removed: under happy-dom the virtualiser renders a single row, so there is nothing to pin.

One case that was wrongly declared unreachable

ChatMessages' viewport slot was removed on the grounds that happy-dom cannot open it. Review disagreed and was right. showAutoScroll comes from a scroll handler comparing scrollHeight against scrollTop + clientHeight; happy-dom reports all three as 0, but stubbing scrollHeight and dispatching the event runs the component's own handler. It is not a renderEach case — that matrix cannot dispatch anything — so it is a test of its own, asserting hidden before and shown after.

The guard

test/utils/indistinguishable-snapshots.spec.ts, over scripts/indistinguishable-snapshots.mjs shared with the regenerator (pnpm snapshots:baseline), with a 293-group baseline.

It deliberately does not demand every entry be unique — that would be 876 red entries and the rule repealed within the week. Most collisions are not defects: a renderEach matrix is cheap, and a color that only tints a closed popup is genuinely indistinguishable. The cheapness is the problem; nothing told you which entries were distinct.

So it pins what exists and fails on a new collision. Three properties:

1. the corpus parsed before any conclusion is drawn from it — otherwise "no new collisions" and "nothing compared" read the same;
2. no new collision;
3. no stale baseline entry — a case that stops colliding fails until its line is removed.

The third is what makes the baseline a record rather than a licence: it can only shrink. 916 entries in 301 groups at the merge base, 876 in 293 now.

Review found three holes in the first version, all fixed here:

• readdirSync is not recursive, so test/components/content/__snapshots__ was outside the guard entirely — with two unrecorded collisions in it. The walk now covers everything under test/.
• the self-check counted distinct bodies, so a parser change that lost every duplicate would have passed it. It counts entries.
• none of the guard's own properties were tested. Deleting the stale half, or collapsing the key to the filename, left the suite green. There is now a describe block over synthetic groups; both mutations fail it.

PORTING.md

§7 gains the rule this work produced, because the next person will hit the same fork:

Compare against upstream at our cursor, not at HEAD. We match — fix the fixture. We differ and it is recorded — remove the case, citing the record. We differ and nothing records it — that is the finding; fix the port.

With the two edges it does not cover: a component absent from upstream is not automatically ours (§1 makes renames mandatory), and the procedure finds what a port dropped, never what it added.

Also found

• Calendar: four button-config props fall through to the picker root as attributes #477 — four Omit props fall through Calendar's omittedProps and render as nextmonth="[object Object]". viewControl is in the list; its four siblings are not.
• a11y(Table): a sortable column header never exposes its sort state #479 — aria-sort appears zero times in Table.vue. Sorting changes the icon for a sighted user and tells a screen reader nothing. Upstream does not emit it either. It surfaced because I used aria-sort as a metric for "did the sorted branch render", got zero, and found the branch had rendered and the metric was wrong.

Checklist

• I have linked an issue or discussion.
• I have updated the documentation accordingly.

Gate run locally: lint 0 · typecheck 0 · test 314 files, 7168 passed, 6 skipped.