rivals-api 5.0.0

Overall, hero, and class win-rate functions now default to provider summaries
without requesting match history. This changes the default counting semantics;
summary participation counts can differ from match-based attribution.

Migration

• Pass method="precise" to player.stats.win_rate(), hero_win_rates(),
  class_win_rates(), heroes(), or classes() for match-based calculations.
  Precise overall rates verify completed match outcomes; hero/class rates use
  longest-played-hero attribution with explicit unresolved coverage.
• Applicable MCP tools accept method="normal" (default) or "precise".
  Season and mode selectors and existing response envelopes are retained.
• Legacy hero/class exact aliases precise; estimate aliases normal summaries.
  Explicit legacy overall matches.fetch_win_rate(method="exact") retains
  history-summary behavior; use precise for completed-outcome verification.

Changes

• Share scoped summary calculations across the three win-rate functions. Keep
  provider counts intact, never average percentages, and expose selection,
  disagreements, private snapshots, missing coverage, and count bases.
• Persist collected history and match details in SQLite across client restarts.
  Refresh each provider's newest pages until a cached match from a completed,
  chronological traversal is found, then reuse its retained history tail.
  Interrupted or unordered scans do not establish stopping boundaries.
• Configure storage with cache_dir, RIVALS_API_CACHE_DIR, or
  persist_cache=False. Calculation caches remain short lived; incomplete
  details are retried. Incremental reuse assumes no older-history backfill
  behind a cached boundary. Neither method guarantees full game-history coverage.
• Add regression coverage for summary-only requests, fractional counts,
  missing modes, provider disagreements, restart reuse, pagination boundaries,
  interrupted scans, and storage failures.

Validation: 202 tests passed; Ruff passed; wheel and source archive built and inspected.