v0.1.0

Added

• ContextManager — add/extend/clear messages, fit() to a token budget, total_tokens(), and last_fit_report()
• FitReport — kept, dropped (each a Dropped(message, reason)), tokens_used, tokens_budget, fits, num_dropped
• Six pluggable strategies in contexttrim.strategies:
  • RecencyDrop — drop oldest first
  • MiddleDrop — drop from the middle ("lost in the middle"), preserving head and tail
  • RoleWeighted — drop lowest-scored roles first; pin system
  • ImportanceWeighted — role weight × recency decay ÷ length penalty
  • ToolResultMerge — merge/dedup adjacent tool results, then truncate the largest if still over budget
  • SemanticCluster — drop least topically-relevant messages via pure-Python TF-IDF cosine similarity
• Strategy base class for writing custom strategies
• Character-heuristic token counter by default (default_token_counter), with an injectable token_counter
• Zero external dependencies — pure Python stdlib (3.9+)
• 24 unit tests covering the manager, all six strategies, token counting, and no-mutation guarantees