Fix hourly time labels dropping an hour on DST transition days

[jaideeppyne]

Changes

On a day that a site's timezone moves its clocks forward, the hourly graph silently loses the last hour of the day.

time_labels/1 built the time:hour labels by adding 0..n hours to the naive start of the range, where n came from DateTime.diff(last, first, :hour). That diff is absolute, so a 23 hour day yields 23 labels ending at 22:00. ClickHouse buckets the rows with toStartOfHour(toTimeZone(timestamp, tz)), which does produce a 23:00 bucket, so that row has no label to land on. The same range also emits a label for 02:00, an hour that never happened. Fall-back days get the opposite: 25 labels for 24 buckets, the last one belonging to the next day.

New York, 2024-03-10, before:

["2024-03-10 00:00:00", "2024-03-10 01:00:00", "2024-03-10 02:00:00", ..., "2024-03-10 22:00:00"]

after:

["2024-03-10 00:00:00", "2024-03-10 01:00:00", "2024-03-10 03:00:00", ..., "2024-03-10 23:00:00"]

I found this with a property check: every hour bucket ClickHouse can produce for a range must have exactly one label, and no label may exist for an hour that doesn't. I used Tzdata as the oracle for what the local hours of a day really are. Sweeping all 596 zones over every day from 2015 to 2030 gives 3,483,024 (zone, day) pairs; 6,284 of them were wrong before, 0 after. Both directions of the bug show up in every zone that observes DST, and Pacific/Chatham and Antarctica/Casey were also wrong because their transitions aren't on the hour.

time:minute already handles this correctly, and the day/week/month labels are pure Date arithmetic, so this only touches time:hour.

Scope I did not cover: a sub-day window that straddles a backwards transition in a zone whose transitions aren't on the hour can still miss one bucket. In practice that is Pacific/Chatham on its April switch, reachable via "Last 24 Hours" or a custom API range. Fixing it properly needs the bucket-to-range overlap test rather than a naive hour walk, which is a lot more code. Happy to add it if you want it covered.

The three new tests fail on master and pass here. I also ran test/plausible/stats/, main_graph_test.exs and test/plausible_web/controllers/api/external_stats_controller/ before and after: identical results, no new failures.

I used an AI assistant to help write the property-check harness and this patch. I ran everything and checked the results myself.

Tests

• Automated tests have been added

Changelog

• Entry has been added to changelog

Documentation

• This change does not need a documentation update

Dark mode

• This PR does not change the UI

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}