test: pin label vocabularies with laravel-telemetry 2.4 metric inspection

[sylvesterdamgaard]

The overlay's counters attribute across bounded vocabularies — publish states, Stache key groups, static-cache outcomes, Glide presets, auth events. The tests asserted membership: "a series with action=published exists". Every branch produces a valid value, so those assertions pass on one branch and stay silent about the other four. Several tests also hand-rolled a loop over collect(), because the fake offered nothing better.

telemetry 2.4's recordedMetrics() does, so the tests now pin the full observed set.

What each one can catch that it could not before

FrontendRequestTest — two foreach ($fake->collect() …) loops with toContain/not->toContain become assertLabelValues('http.route', [...])->assertSeriesCount(1). Previously a third route value, or the /{segments?} catch-all surviving beside the logical route, passed.

PublishStateTest — new test; the five per-state tests are untouched. Saves all four states in one run and pins action to exactly [published, draft, scheduled, expired]. A per-state test cannot see a status resolved once and reused across saves — all four collapsing onto one action still passes every one of them. It also covers expired, which entrySaveAction() supports but nothing exercised.

ListenersTest — the Glide preset test now fires three different ad-hoc param sets and pins preset to [thumbnail, custom] with a series count. That is the cardinality leak the listener's own comment warns about, and nothing tested it. Auth events: three membership assertions → one exact set, so a fourth value (an unmapped event falling back to a class name) fails.

StaticCacheTest — miss/write and invalidate/flush are exact sets with a count, so double-counting under a wrong outcome is visible.

MetricsProviderTest — collect(...)->first(...) + samples[0] read only the first sample; a duplicate pages series or a stray collection passed silently while doubling the dashboard total.

CacheKeyGroupsTest (new) — drives ~70 realistic Stache keys through the real cache.operations counter and asserts the exact six-bucket vocabulary plus assertCardinalityBelow(25) / assertLabelCardinalityBelow('key_group', 10), and that total() equals 3 × the key count so nothing is dropped. Boundedness is the property the classifier exists for; the unit test could not express it, and nothing covered the fact that Hooks::register actually installs the classifier.

tests/Unit/CacheKeysTest.php is deliberately untouched — already exact per key, and rewriting it into a cardinality test would trade precise coverage for nothing.

Verified, not assumed

Three deliberate breakages, each reverted. The one worth quoting: dropping 'expired' from RecordContentChange::entrySaveAction() leaves all five pre-existing publish-state tests green and fails only the new one:

Metric [statamic.content.changes] observed label [action] values [draft, published, saved, scheduled],
expected exactly [draft, expired, published, scheduled] (never observed: expired; unexpected: saved).

Removing classifyCacheKeysUsing() from Hooks::register is invisible to the unit test and fails CacheKeyGroupsTest. Making the Glide preset fall back to raw param keys fails with unexpected: fit, w, w,h.

Tests only — git diff main -- src/ is empty. pint, PHPStan level 9 and 79 tests all green.

Noticed, not touched

src/StaticCaching/StaticCacheTelemetry.php:91 says "count each outcome once per request", but the guard is a single last-value slot: $request->attributes->get(RECORDED_KEY) === $result. It suppresses immediate repeats only, so a miss → write → miss sequence in one request — which the comment itself says happens, via middleware plus nocache — counts miss twice. Either the comment overstates the guarantee or the slot should be a per-outcome set. No test here encodes the current behaviour as correct.