Standings history for current_rating-ranked tournaments

[amrtgaber]

Summary

Implements #305: /standings/history now serves current_rating-ranked tournaments — the 501 guard is gone, completing the rank_by feature surface (#290).

The sweep gains a second value-resolution mode alongside the peak ratchet, reusing all shared machinery (daily anchors + reorder-stamped buckets, roster visibility gating, team series, response shape):

• Last-value semantics: an entrant's as-of-bucket current rating is the latest recorded value at-or-before the bucket — the in-window match log (points at completion time) merged with the snapshot observations (recorded on every current change since Per-tournament scoring config: rank_by (peak_rating | current_rating) #290), with the log winning ties, mirroring _frozen_current_by_profile's precedence. Non-monotone by nature: declines chart as declines.
• Exact-first baseline, no reconstruction fold: the rating held entering the window (the first in-window game's old_rating — exact, games are the only mover) → else the latest pre-start observation → else earliest in-window observation carried back → else a flat line at live current_rating (the recorder's self-seeding argument). Current-mode tournaments postdate the Per-tournament scoring config: rank_by (peak_rating | current_rating) #290 recorder, so observation coverage is complete from first poll — none of the peak mode's log-reconstruction (#357) is needed.
• No clamp: the live clamp exists for rebased-away peak noise; current has no such invariant. The peak path is byte-identical (its pre-event-peak query is now skipped in current mode — one query saved).
• Ranking per bucket: as-of-bucket current desc, live max_rating desc (static tiebreak), name — the live table's current-mode order, so the final bucket provably matches /standings.

Contract (additive): StandingHistoryPoint.current_rating and TeamStandingHistoryPoint.combined_current_elo — exactly one of the two rating fields per point is populated, matching rank_by (the same split as the summary's rating cards); combined_peak_elo becomes nullable but is always populated in peak mode, so existing consumers see identical bytes. Consumers regenerate orval clients.

Tests

594 passed. New: a decline charts as a decline with the reorder captured (the case peak mode can't express); final bucket equals the live /standings current-mode order; baselines hold before the first data point (entry old_rating exact for a player with games, carried-back observation for a snapshot-only player); team series sums current with null peak fields. The five existing peak-mode whole-point assertions updated for the additive null keys — peak values and positions unchanged.

Closes #305