Skip summary judge LLM when no similar summaries meet the 0.4 threshold, and add Go/Python ingest benchmarks.

Summary judge now searches existing summaries first and only calls the LLM when a match scores at or above 0.4; otherwise items are added deterministically. Includes benchmark CLIs for comparing Go and Python pipeline timing with real LLM calls.

Co-authored-by: Cursor <cursoragent@cursor.com>

Author: ved015