v0.2.3 — completing the review fixes

The v0.2.2 fixes were themselves re-reviewed, and several turned out to be
partial. No sensor values change; no reconfiguration needed.

The miss that mattered

Transactions had their own pagination loop, a few hundred lines away from
the shared one fixed in v0.2.2 — and it kept both of the same bugs. A server
answering "here are 0 records, there are 100" produced 0 points earned this
month. There is now exactly one pagination implementation; the duplication
was the entire cause.

Also fixed

• Advancing the offset does not help against a server that ignores it. Two
  identical pages turned one 10-point transaction into 20. Records are now
  matched by id, and a page containing only repeats is an error — matched by
  id rather than by content precisely so that Trappers' own duplicate tag
  reads, which are separate records sharing a date, still work normally.
• A fast poll could still schedule a redundant one. If a refresh finished
  just before the slot it was woken for, the next poll was scheduled a minute
  later instead of three hours later.
• The monthly reset timestamp could name the wrong month. A poll starting
  at 23:59:59 on the 31st computes that month's figures and can finish a second
  into the next one — and would then file them under the new month.
• The euro rate accepted a catalogue split exactly 80/20, treated an unreadable
  expiry date as "never expires", and read "Cadeaukaart € 25,00 2026 editie"
  as 25.002026. The rate is also re-checked after midnight now, since article
  availability is expressed in whole days.

The leak guard, where two holes were in the code written to close a hole

• The new commit-message check skipped lines starting with #, assuming git
  strips comments. git commit -m does not, so a commented-out secret went
  straight through the very hook added to catch it.
• Files with non-ASCII names were committed unscanned, because git renders
  such paths escaped and the escaped form matches no file.
• The hook's own test suite counted any failed commit as a successful block,
  so a broken test environment would have reported passes without ever reaching
  the hook — the same fail-open the guard exists to prevent, one level up.

171 tests, 34 hook cases.