Profile and optimize Phase 2 species classification performance

[NewGraphEnvironment]

Problem

Phase 2 species classification is the pipeline bottleneck. ADMS (11k segments, 5 species) takes ~70 min sequential, ~14 min per species. The expensive operation is frs_classify() which runs fwa_upstream() network traversal for every segment against every break point.

With workers = 4 this parallelizes but is still ~18 min.

Profiling needed

• frs_classify() SQL: the fwa_upstream() call is O(segments × breaks). With 11k segments × 32k breaks that's a lot of ltree comparisons
• Index strategy on breaks table (now has label + source columns)
• Whether batching or partitioning the classify query would help
• Phase 1 is fast (<80s even with 533k crossings) — no optimization needed there

Potential approaches

1. Add indexes on breaks table blue_line_key column
2. Batch frs_classify() by blue_line_key groups
3. Pre-filter breaks to only BLKs present in the working table
4. Use spatial index instead of ltree for segment-break matching
5. Docker PG tuning (already high work_mem/parallel workers, but could profile)

Context

Phase 2 conn fix in v0.7.0 means sequential mode reuses conn correctly. Parallel mode (workers > 1) works but requires PG_*_SHARE env vars pointing at the target DB.

Relates to #70