fix: address PR #104 review feedback

- Correct idempotency guard docstring in update_season_stats() to
  accurately describe the last_game FK check limitation: only detects
  replay of the most-recently-processed game; out-of-order re-delivery
  (game G after G+1) bypasses the guard. References issue #105 for the
  planned ProcessedGame ledger fix.
- Fix migration card_type comment: 'batting' or 'pitching' → 'batter',
  'sp', or 'rp' to match actual seeded values.
- Remove local rarity fixture in test_season_stats_update.py that
  shadowed the conftest.py fixture; remove unused rarity parameter from
  player_batter and player_pitcher fixtures.
- Update test_double_count_prevention docstring to note the known
  out-of-order re-delivery limitation.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/services/season_stats.py

@@ -2,9 +2,15 @@
 season_stats.py — Incremental PlayerSeasonStats update logic.
 
 Called once per completed StratGame to accumulate batting and pitching
-statistics into the player_season_stats table. The update is idempotent:
-if this game_id has already been processed (detected via last_game FK),
-the function returns early without double-counting.
+statistics into the player_season_stats table.
+
+Idempotency limitation: re-delivery of a game is detected by checking
+whether any PlayerSeasonStats row still carries that game_id as last_game.
+This guard only works if no later game has been processed for the same
+players — if game G+1 is processed first, a re-delivery of game G will
+bypass the guard and double-count stats. A persistent processed-game
+ledger is needed for full idempotency across out-of-order re-delivery
+(see issue #105).
 
 Peewee upsert strategy:
 - SQLite:    on_conflict_replace() — simplest path, deletes + re-inserts
@@ -364,8 +370,15 @@ def update_season_stats(game_id: int) -> dict:
     Accumulate per-game batting and pitching stats into PlayerSeasonStats.
 
     This function is safe to call exactly once per game. If called again
-    for the same game_id (detected by checking last_game FK), it returns
-    immediately without modifying any data.
+    for the same game_id while it is still the most-recently-processed
+    game for at least one affected player (detected by checking last_game
+    FK), it returns early without modifying any data.
+
+    Limitation: the guard only detects re-delivery if no later game has
+    been processed for the same players. Out-of-order re-delivery (e.g.
+    game G re-delivered after game G+1 was already processed) will not be
+    caught and will silently double-count stats. See issue #105 for the
+    planned ProcessedGame ledger fix.
 
     Algorithm:
         1. Fetch StratGame to get the season.
@@ -396,8 +409,10 @@ def update_season_stats(game_id: int) -> dict:
     season = game.season
 
     with db.atomic():
-        # Step 2 — Double-count prevention: check if any row already
-        # carries this game_id as last_game
+        # Step 2 — Double-count prevention: check if any row still
+        # carries this game_id as last_game. Note: only detects replay
+        # of the most-recently-processed game; out-of-order re-delivery
+        # bypasses this guard (see issue #105).
         already_processed = (
             PlayerSeasonStats.select()
             .where(PlayerSeasonStats.last_game == game_id)

migrations/2026-03-17_add_evolution_tables.sql

@@ -90,7 +90,7 @@ CREATE INDEX IF NOT EXISTS player_season_stats_player_season_idx
 CREATE TABLE IF NOT EXISTS evolution_track (
     id SERIAL PRIMARY KEY,
     name VARCHAR(255) UNIQUE NOT NULL,
-    card_type VARCHAR(50) NOT NULL,       -- 'batting' or 'pitching'
+    card_type VARCHAR(50) NOT NULL,       -- 'batter', 'sp', or 'rp'
     formula VARCHAR(255) NOT NULL,        -- e.g. 'hr', 'k_per_9', 'ops'
     t1_threshold INTEGER NOT NULL,
     t2_threshold INTEGER NOT NULL,

tests/test_season_stats_update.py

@@ -156,11 +156,6 @@ def make_play(game, play_num, batter, batter_team, pitcher, pitcher_team, **stat
 # ---------------------------------------------------------------------------
 
 
-@pytest.fixture
-def rarity():
-    return Rarity.create(value=1, name="Common", color="#ffffff")
-
-
 @pytest.fixture
 def team_a():
     return _make_team("TMA", gmid=1001)
@@ -172,13 +167,13 @@ def team_b():
 
 
 @pytest.fixture
-def player_batter(rarity):
+def player_batter():
     """A batter-type player for team A."""
     return _make_player("Batter One", pos="CF")
 
 
 @pytest.fixture
-def player_pitcher(rarity):
+def player_pitcher():
     """A pitcher-type player for team B."""
     return _make_player("Pitcher One", pos="SP")
 
@@ -399,15 +394,18 @@ def test_decision_integration(team_a, team_b, player_batter, player_pitcher, gam
 def test_double_count_prevention(team_a, team_b, player_batter, player_pitcher, game):
     """Calling update_season_stats() twice for the same game must not double the stats.
 
-    What: Process a game once (pa=3), then call the function again. The
-    second call should detect the already-processed state via the
-    PlayerSeasonStats.last_game FK check and return early with 'skipped'=True.
+    What: Process a game once (pa=3), then immediately call the function
+    again with the same game_id. The second call detects via the
+    PlayerSeasonStats.last_game FK check that this game is still the
+    most-recently-processed game and returns early with 'skipped'=True.
     The resulting pa should still be 3, not 6.
 
     Why: The bot infrastructure may deliver game-complete events more than
-    once (network retries, message replays). Without idempotency, stats
-    would accumulate incorrectly and could not be corrected without a full
-    reset.
+    once (network retries, message replays). The guard prevents
+    double-counting when the replayed game is still the last game
+    processed for those players. Note: this test only covers same-game
+    immediate replay — out-of-order re-delivery (game G after G+1) is a
+    known limitation tracked in issue #105.
     """
     for i in range(3):
         make_play(