perf(retrieval): optimize map-nav evidence packing

[suguanYang]

Summary

Production Logfire recorded retrieval mapnav phase=evidence_pack seconds=22.309 chunks=20 for a namespace with 644 active documents. The slow path was local composition work, not an LLM call or a database query.

This change:

• Uses the lightweight path_titles lookup for evidence headers instead of get_structure, which can materialize and count a complete subtree.
• Replaces repeated full-selection rendering in trim/refill with an exact incremental length tracker.
• Finds header-only owners by walking parent chains instead of comparing every owner pair.
• Preserves the existing evidence budget, ordering, indentation, truncation, and refill behavior.

Validation

• uv run pytest apps/api/tests/contract/test_retrieval*.py (41 passed)
• uv run ruff check packages/shared-python/shared/services/retrieval/nav/nav_compose.py apps/api/tests/contract/test_retrieval_lazy_tree_contract.py
• uv run pyright packages/shared-python/shared/services/retrieval/nav/nav_compose.py apps/api/tests/contract/test_retrieval_lazy_tree_contract.py
• Random differential check: incremental lengths matched full rendering through delete/re-add sequences.
• Synthetic 3,000-chunk benchmark: about 55s before the change, about 0.34s after the change; final evidence remained 4,967 chars with 43 chunks kept.

No database migration or backfill is required.