Per-tournament scoring config: rank_by (peak_rating | current_rating)

[amrtgaber]

Summary

Implements #290: the ranking metric becomes per-tournament configuration instead of platform behavior. rank_by = peak_rating (server_default — every existing row and the launch event's behavior, unchanged) or current_rating.

Model + contract (all additive; consumers regenerate orval clients):

• Tournament.rank_by — non-native Enum stored as the StrEnum value, mirroring MatchState. Expand-only migration 7f07220ac5cb (NOT NULL + server_default → safe under revision rollover; validated on a throwaway Postgres 16 together with its parent).
• rank_by on TournamentRead/Create/Update (explicit null → 422, like leaderboard_id).
• TournamentSummary.highest_current_rating — the rating card for current-mode tournaments; exactly one of the two rating cards is ever populated, highest_peak_rating stays the peak-mode card, pre-existing consumers see null in the new field.

Read paths threaded (app/routers/tournaments.py):

• /standings: SQL ORDER BY ranks on the configured metric (other rating + name as tie-breaks); post-window re-rank and the surfaced value of the ranked metric freeze, the other rating stays live.
• /teams/standings: member order, combined_rating_sum/combined_rating_average, and team order all follow the metric; field names deliberately unchanged (same documented-semantics approach as Team standings: rank teams by combined PEAK rating, not current; add per-member max_rating #158).
• /summary: the rating card ranks by the configured metric via the generalized _rank_metric_by_profile (keeps the Mirror FE stat cards on /summary; resolve displayName into name #244 correction in peak mode).
• /standings/history: peak mode only — 501 for current-mode tournaments. The sweep's carried-in baselines, ratchets, and live clamp are peak-specific machinery; an honest refusal beats a chart whose order contradicts the live table. Tracked as a follow-up issue; /progression covers rating-over-time meanwhile.

Freeze semantics for current mode (the issue's open design point, resolved): new _frozen_current_by_profile behind a _frozen_metric_by_profile dispatcher. Current is a last-value metric, not a ratchet, so the fold differs from peak: (1) the new_rating of the last game started by the bound — authoritative, includes the buzzer-beater that settles after the bound, durable in our matches table for any in-window game; (2) fallback, the last recorded snapshot at-or-before the bound; (3) live current. No clamp — clamping exists for rebased-away peak noise only.

Recorder widened (insert_rating_snapshots): diffs the (max_rating, current_rating) pair instead of max alone, so current-mode freezes have observations to fall back on. Volume stays bounded by play (≤ one row per rated game per tracked pair); self-seeding and rebase-down behavior preserved.

CLAUDE.md's three convention bullets (sort order, team aggregates, freeze) updated to the config-aware wording.

Non-goals (per #290)

No points/W-L/bracket modes — this is the config seam plus the two modes with known semantics; the platform stays a ladder-standings API.

Tests

581 passed. New coverage: rank_by round-trip/default/invalid on create + PATCH (incl. explicit-null 422); current-mode standings order with peak surfaced untouched; frozen-current via match log (beats a stale snapshot, post-window play ignored), snapshot fallback, live fallback; team sums/order on current; summary card routing both modes; history 501; recorder current-only-change append.

Merge order: after #303 (stacked — shares the migration lineage and tournament-write test context; #299's one-line CLAUDE.md fix also lands first and is a trivial rebase here).

Closes #290