Two-way player bare-name residuals in browse-page audit / compare-standings subsystem

[alhart2015]

Follow-up from #190 (player_key sweep). The #190 branch swept the enumerated roster/lineup/swap layers off bare-name keying, but three residual bare-name spots remain in the browse-page audit + compare-standings subsystem, which #190 did not enumerate. Each collides a two-way player's hitter and pitcher rows (same name). All latent today (no two-way player rostered) and surfaced repeatedly by loop-review on the #190 branch.

1. Compare endpoint — web/season_data.py:1123 compute_comparison_standings resolves the dropped player via find_player_by_name(roster_player_name, ...) (first bare-name match). The route (season_routes.py ~1719) reads roster_player as a bare HTTP arg with no type, so the compare view can only ever price the first (hitter) row of a two-way player; the pitcher-row comparison is unreachable. Fix needs the compare route + frontend to thread roster_player_type (the sibling other_player side already passes other_type; the roster-audit template already emits entry.player::entry.player_type).

2. Audit SGP/gap display — lineup/roster_audit.py player_sgp / fa_sgp are keyed on bare name, so a two-way player's two rows collide and the browse-page drop-candidate SGP/gap shows the wrong row's value.

3. audit_index — web/season_routes.py:1442 looks up audit_index[(drop_player.name, fa_name)], and audit_index (built from the ROSTER_AUDIT cache) is keyed on bare (player_name, fa_name) tuples. Two-way rows collide in the cache. Fixing requires AuditEntry.player to carry the key/type and a cache-shape change.

These form one coherent cluster: the audit subsystem's identity model (AuditEntry.player, audit_index, the serialized ROSTER_AUDIT cache, and the compare route's HTTP arg) is name-based end-to-end. Fixing properly means re-keying that model on player_key (name::player_type), which is a larger cross-cutting change than #190's scope.

Found by loop-review on fix/190-player-key-sweep, 2026-07-12.