v3.1.1

Nine tickets from the pending queue, cleared in three batches. Two behaviour
changes; everything else is documentation the measurements had outgrown, and
the tests that now hold those claims to the code.

Fixed

• _sample_filter no longer builds one OR arm per track. SQLite had to
  evaluate all of them against every row of the row_number() subquery, so
  the batch AU-050 introduced cost O(batch squared): 64 tracks took 140 ms
  against the real export where the per-track implementation it replaced took
  65 ms, and 365 tracks took 3.8 s. One IN over a composite key restores a
  flat cost per row -- 64 tracks in 41 ms, 0.62x of the pre-AU-050 figure at
  every width measured -- while keeping the two statements AU-050 bought. The
  batch reads fewer bound parameters than before, not more (CUI-0019).
• parse_datetime converts an ISO timestamp carrying an offset into the
  requested zone, and recognises a negative one. The branch tested for "+",
  so -05:00 reached neither it nor the UTC branch and died in the naive
  strptime below, taking the whole file with it. GPX and TCX write UTC Z
  and were converted; KML writes the local offset and was not, so the three
  exports of one run disagreed by an hour anywhere outside Hong Kong -- and
  _TIMESTAMP_FORMAT carries no offset, so the evidence was gone by the time
  the row was written. Verified byte-identical over 1,087 activity records and
  328,752 track points (CUI-0006).
• The static sampler rounds half to even, as Python does, instead of half up.
  22.6% of the (length, points) pairs the sampler can be asked for over the
  123 track lengths the export holds land on a tie, where the two deployment
  modes drew different indices. None is reachable while TRACK_POINTS is 600,
  which is both above every stored track and even -- an odd limit is what
  makes a tie possible at all (CUI-0021).

Documented

• track(points: 1) returns the last row alone rather than the first, in the
  docstring and in the SDL description, matching downsample deliberately
  (CUI-0004).
• MAX_QUERY_DEPTH stays 5 and says outright that it cannot fire against this
  schema; an alias flood is wide rather than deep, and MAX_QUERY_TOKENS is
  what refuses one. Two tests hold the limiter to its edge, one of which goes
  red if anyone deepens the type graph (CUI-0003).
• Every wall clock quoted by MAX_TRACK_FIELDS_PER_REQUEST names the scale it
  was read at, and the Vercel headroom is read off the export rather than
  inferred from a 600-row fixture (CUI-0020, CUI-0017).
• A missing string column reads as empty text rather than None, recorded
  where the choice is made and pinned by tests (CUI-0014).
• The CI prose no longer repeats pages.yml's branch list or job steps, which
  had gone stale twice in one session, and points at the workflow instead
  (CUI-0015).

Known

• Nothing bounds the list fan-out: 166 aliased activities fields with no
  track at all take about 3.1 s on the real export, eleven times the worst
  legal track shape and within 4.9x of the 15 s function limit (CUI-0027).