Leg 28 llm responses

[Ankita-Advitot]

• Add three new files that make up the LLM semantic-validation step: Extraction/validate_llm/llm_field_checks.py (prompt + parsing), Extraction/validate_llm/ollama_client.py (Ollama HTTP client), Extraction/validate_llm/test_prompt.py (standalone tester).
• llm_field_checks.py sends one Ollama call per case, covering only fields with real semantic risk (SEMANTIC_FIELDS whitelist: name, role, address, office_address, designation, district, state, purpose, nature_of_disposal) across persons, judges, lawyers, organizations, and hearings. Case numbers, dates, IDs, and other rule-checkable fields are excluded — those are already handled by the existing rule engine.
• The model returns invalid-only JSON ({"checked": N, "invalid": [{"n", "reason"}, ...]}) instead of a verdict per field, with a checked-count guard (one retry, then skip on persistent mismatch) so an incomplete response can't be silently trusted.
• A small closed reason-code vocabulary (wrong_content, missing_location, missing_designation, narrative_text) — an unrecognized/hallucinated code is treated as valid rather than kept under a relabeled reason.
• ollama_client.py provides ask_json() for this JSON-object response shape, plus _call_ollama()/ask_valid()/ask_lines() as supporting/alternate helpers.
• Extraction/main.py: wire the new check into the pipeline — build the per-entity field lists (persons, judges, lawyers from missing_advocates, organizations from new_parties, hearings, case-level district/state) and add primary/secondary drop logic: name is primary for persons, judges, lawyers, and organizations — if missing or flagged wrong, the whole entity is dropped before the Neo4j insert phase runs. Every other checked field is secondary — a flag there only nulls that field.
• Extraction/validate_llm/engine.py / init.py: expose validate_case_llm()/check_case() as the entry point main.py calls.
• shared/config.py: add the Ollama URL/model env vars this step needs.
• requirements.txt: fix two pins that don't exist on PyPI — onnxruntime==1.27.0 → 1.23.2, pandas==3.0.3 → 2.3.3.