Tune knock-early thresholds and fix failing test suite

Tighten should_knock_early() so AI no longer knocks with projected
scores of 12-14. New range: max_acceptable 5-9 (was 8-18), with
scaled knock_chance by score quality and an exception when all
opponents show 25+ visible points.

Fix 5 pre-existing test failures:
- test_event_replay: use game.current_player() instead of hardcoding
  "p1", since dealer logic makes p2 go first
- game.py: include current_player_idx in round_started event so state
  replay knows the correct starting player
- test_house_rules: rename test_rule_config → run_rule_config so
  pytest doesn't collect it as a test fixture

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

server/models/game_state.py

@@ -237,7 +237,7 @@ class RebuiltGameState:
         self.initial_flips_done = set()
         self.drawn_card = None
         self.drawn_from_discard = False
-        self.current_player_idx = 0
+        self.current_player_idx = event.data.get("current_player_idx", 0)
         self.discard_pile = []
 
         # Deal cards to players (all face-down)