hotfix: fix faulty intent parsing, enhance search, simplify ICP scoring

• Replaces deterministic ICP scoring with an LLM judge (icp_judge) that returns an overall score, lead stage, localized summary, and 3–5 self-chosen criteria with per-criterion scores and notes (icp_evaluation, score_reasons).
• Improves discovery accuracy by sanitizing parsed intent (no inferred geography/funding unless mentioned in the query), centralizing query expansion in discovery_query_builder, and applying exclude terms to search queries during backfill/find-more rounds.
• Filters hollow enrichment profiles via enrichment_quality and adds automatic discovery backfill (up to 2 rounds) when substantive lead count falls below the requested limit.
• Splits ranked output into visible leads and reserved_leads for find-more reranking; streams completed payload as { leads, reserved_leads }.
• Hardens the enrichment pipeline with per-company error handling, substance checks before scoring, and restored/expanded lite search tool configuration.