fix(sports): harden the B2/B3 capabilities against bad config and subclasses

Review pass on the phase B2-B4 changes. Every fix here is the same shape
as the crashes this PR already fixed in the hockey and baseball
extractors: a config or feed value that is present-but-wrong reaching
arithmetic or a comparison on a path with no guard.

celebrations:
- celebration_duration is coerced and floored at init. It is compared
  numerically in display() *outside* any try block, so a string from a
  hand-edited config propagated a TypeError straight out; zero or
  negative armed a celebration that could never render.
- A render failure now disarms instead of staying armed. It previously
  retried the same broken render on every frame for the rest of the
  window -- a traceback per frame, and no scorebug either.
- prune_score_baselines() for the live set. Only _check_for_win removed
  entries, so a game that left the live list any other way leaked its
  baseline and the dict grew all season.
- display() reuses has_active_celebration() rather than repeating its
  window comparison, and log lines carry a [Celebrations] prefix.

rotation:
- MAX_WEIGHT ceiling. A cycle is sum(weights) long and each step scans
  every game, so an unbounded weight from a misread config spins the
  display thread -- on a Pi that stalls rendering outright.
- register_rotation_strategy rejects a non-subclass factory at
  registration instead of failing frames later inside schedule().
- schedule() previews through type(self), so a subclass overriding
  next_game is previewed with its own ordering -- which is what the
  method promises.

sports_scroll:
- scroll_speed / scroll_delay coerced. dict.get(key, default) only helps
  when the key is absent; present-but-null reached the multiplication
  inside __init__ and the display failed to construct at all.
- update_scroll_position and get_visible_portion moved inside the try.
  They ran outside it, so a raise there reached the plugin's frame loop
  despite the comment promising none can.
- prepare_and_display guards the subclass call, so one sport's bad
  payload cannot take down the shared orchestration for the others.
- _current_game_type spells "nothing active" as "" in both classes; the
  manager said None while the display said "".

Not taken: the report that baseball's favorite-team debug path still
reads event-level status. Verified against current code -- there are no
remaining game_event["status"] reads in that file; it was fixed in
2486bdb and the finding is stale.

Gates: 747 core unit, 66 plugin safety.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FgbA8SMutQQpXkMG8LMmC4

docs/SPORTS_UNIFICATION.md

@@ -166,9 +166,9 @@ self.rotation = get_rotation_strategy("swrr", weight_for=self._live_weight)
 `weight_for` is supplied by the host, so the *favorites* policy stays with the
 plugin and `rotation.py` never learns what a favorite is. An unknown strategy
 name degrades to `simple` rather than raising: the name comes from user config,
-and a typo should cost the boost, not the scoreboard. A plugin needing an
-ordering core does not ship calls `register_rotation_strategy` instead of core
-growing a branch.
+and a typo should cost the boost, not the scoreboard. When a plugin needs an
+ordering that core does not ship, it calls `register_rotation_strategy` to add
+its own — rather than core growing a branch for it.
 
 `test_sports_capabilities.py` checks each strategy against a **verbatim
 transcription** of the plugin code it replaces, over every live-game shape up to