Extract grounding/caching cluster from query.py into dedicated mixin

[aaronsb]

Context

PR #276 code review (finding #6) noted query.py grew to 1126 lines. The grounding and caching code forms a natural cluster:

• _get_polarity_axis() — tier-1 polarity cache
• _get_graph_generation() — generation reading with SAVEPOINT
• calculate_grounding_strength_semantic() — per-concept grounding
• calculate_grounding_strength_batch() — batch grounding with chunking
• Module-level cache globals (_polarity_cache, _grounding_cache, locks)
• _parse_embedding(), _parse_confidence() helpers

This cluster has no dependency on the rest of QueryMixin (search, learned knowledge) and could be a GroundingMixin in api/app/lib/age_client/grounding.py.

Also Includes

• Warm cache optimization (review finding Add Librarian Tool Set to MCP Server #7): Skip pool connection when all concepts are cache hits. Currently acquires a connection just to read graph generation even on 100% warm.
• Docstrings: Document the two-tier cache invalidation model on key methods:
  • calculate_grounding_strength_semantic() — polarity axis caching, vocab generation check
  • GraphFacade._execute_sql() — pinned connection, generation-based ensure_fresh
  • graph_facade.py module docstring — full pipeline: topology (sub-ms) → cached hydration

References

• PR feat: generation-aware hydration caches (ADR-201 Phase 5f) #276 (ADR-201 Phase 5f)
• Phase 5e in .claude/todo-adr-201-graph-accel.md