Build deterministic analysis core

[Spbd1]

Motivation

• Provide a deterministic analysis core that extracts claims, locates exact evidence spans, and retrieves taxonomy candidates before any LLM integration.
• Reduce false positives by applying healthy-reasoning suppressors and conservative lexical matching rules so neutral prose does not produce many candidates.
• Ensure components are fast and scale to large taxonomy packs so retrieval works with 1000+ rows for practical QA and testing.

Description

• Implemented sentence-level claim extraction in engine/argument_risk_engine/extraction/claim_extractor.py with typed Claim objects, preserved character offsets, marker-based claim-type detection, precedence rules, and short-fragment filtering.
• Added exact evidence handling in engine/argument_risk_engine/explanation/evidence.py via EvidenceSpan and find_evidence_spans that return exact substring matches and never fabricate evidence, plus a legacy-compatible evidence_span wrapper.
• Built an in-memory indexed taxonomy in engine/argument_risk_engine/retrieval/inverted_index.py with tokenization, stopword/generic-term suppression, field-specific indexing (name, synonyms, signals, trigger_patterns, definitions), activation/deprecation checks, candidate-only flags, and healthy-suppressor detection.
• Implemented conservative lexical retrieval in engine/argument_risk_engine/retrieval/lexical_retriever.py producing RetrievedTaxonomyEntry objects with retrieval_score, matched terms/fields, retrieval_reason, false_positive_risk, healthy-pattern suppression (penalties), diagnostics, and index caching; and added candidate_filter and retrieval_diagnostics helpers for final filtering and observability.
• Added/updated unit tests in tests/test_claim_extractor.py and tests/test_retriever.py to cover sentence splitting, offsets and claim types, exact evidence spans, large neutral retrieval behavior (1000 rows), deprecated exclusions, candidate-only rows, and healthy-reasoning suppressor behavior.

Testing

• Ran lint/format checks and auto-fixes with python -m ruff check --fix and final python -m ruff check, and issues were fixed by the automated pass (no remaining ruff errors reported).
• Ran targeted tests with python -m pytest tests/test_claim_extractor.py tests/test_retriever.py -q which passed, and then ran the full test suite with python -m pytest -q which completed successfully.
• All automated tests passed: 25 passed, 3 warnings (pytest) and the lint checks completed successfully.

---

Codex Task