feat: remove legacy deterministic path — Claude subagent only

The evaluate command no longer pretends to analyze documents with regex.
It ingests the PDF, loads criteria, spawns the agent panel, and waits
for Claude to decompose the document via MCP.

Removed:
- enrich_document() and extract_first_phrase()
- Regex extraction loop (extract_all, to_claims_and_evidence)
- Deterministic validation (validate_deep/core)
- Keyword alignment (align_via_ontology fallback)
- Semantic embeddings step
- Mechanical debate (identical scores converging with themselves)

Kept:
- PDF ingest, criteria loading, agent spawning
- SPARQL rule mining, gate verdict, web verify, badge
- Report generation, visualization, memory
- Orchestration task generation (always runs now)

The pipeline is honest: no score without Claude's decomposition.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: fabio-rovai