Add OpenAI two-pass extraction pipeline with QC, compaction, and review integration

[karilint]

Motivation

• Introduce a production-ready two-pass OpenAI extraction backend to detect raw evidence (PASS1) and produce ETS-structured trait records (PASS2) while reducing token use and noisy analytical content.
• Improve robustness of the existing pipeline by adding QC/normalization, deterministic deduplication, timing instrumentation and safer handling of schema drift in importer code.
• Surface PASS1/PASS2 artifacts and QC to the UI and persist necessary fields on extraction runs and assertions for review and downstream persistence.

Description

• Add OpenAI adapter recode_extraction.adapters.openai_client.OpenAITwoPassClient and prompt builders in openai_two_pass_prompts.py, plus lightweight service exports in openai_two_pass.py.
• Implement orchestration support in recode_extraction.services.orchestrator.RecodePipelineRunner with a new openai_two_pass path that runs per-page PASS1 calls, compacts evidence (pass1_compaction.py), runs PASS2, runs QC/normalization (qc.py), and persists candidate ExtractedAssertionModel rows; add timing logs and safer page-number handling.
• Add trait vocabulary service trait_vocabulary.py to build abbreviation dictionaries and trait lists from the DB, cached for performance.
• Extend data models and DB via migrations to store pass1_evidence_package, pass2_structured_package, qc_summary, qc_errors and extend unmapped_reason length.
• Add UI and view updates to show PASS1/PASS2 availability, QC summary and allow selecting the OpenAI backend when enabled; add settings flags and many RECODE_OPENAI_* settings in settings.py.
• Add review flow improvements to services.review to accept prefilled ets_payload when persisting approved assertions and to validate presence of review columns at runtime.
• Tighten imports module to avoid import cycles and fix validation regex escaping in base_validation.py.
• Harden BaseImporter.get_or_create_source_location to handle None values and database schema drift when coord_text lacks a default.
• Add requirements for openai and pydantic and a collection of unit tests and fixtures covering adapter, compaction, QC/normalization, orchestrator flows, review persistence, and importer edge cases.

Testing

• Ran unit tests for the new OpenAI pipeline and helpers including app/tests/recode_extraction/test_openai_two_pass_pipeline.py, test_openai_client_adapter.py, test_openai_qc_normalization.py, test_openai_schema.py, test_pass1_compaction.py, test_openai_two_pass_pipeline.py, and updated review/importer tests, and they all passed.
• Ran existing orchestrator and review tests (app/tests/recode_extraction/test_orchestrator.py, test_qc_review.py) which succeeded and show timing logs and new persistence behaviors.
• Adapter-level OpenAI calls are mocked via the mock_openai fixture in app/tests/recode_extraction/conftest.py to avoid live API calls during tests.

---

Codex Task