Retire the legacy prose-JSON repair stack

[Ndevu12]

Summary

Native structured outputs (#38) made the prose-JSON repair machinery redundant: schemas are attached to agents via pydantic-ai output_type, validated natively, and retried with validation feedback by the framework itself. This change completes that migration by removing the parallel legacy path entirely — −5,307 / +150 lines — leaving one LLM call path (run_structured / try_run_structured) with the existing heuristic fallbacks as the safety net.

What was removed

• src/utils/ repair stack (8 modules, ~4,400 lines): enhanced_response_handler, fallback_processing, content_quality, model_adaptation, enhanced_validation, quality_monitor, retry_manager, response_models, plus the repair-narrative message formatters that only they used.
• src/retrieval/helpers_modules/ (~700 lines): JSON extraction/validation/recovery helpers whose only remaining consumers were the stack itself and query expansion's legacy branch. (src/retrieval/helpers.py is unrelated and stays.)
• Legacy branches in synthesis (per-paper extraction and collective pass), gap analysis, and query expansion, along with the handler/session_id plumbing threaded through extract_papers, synthesize_collective, run_synthesis, analyze_gaps, SynthesisStage, and GapAnalysisStage.
• Configuration: the llm.structured_outputs flag (settings, config/default.yaml, .env.example, docs) and the now-unused extraction_max_retries / collective_max_retries synthesis knobs. Research loop and verification now gate on synthesis.llm_enabled alone.
• Tests: the deleted stack's suite (test_graceful_response_handling.py); synthesis and structured-output tests now mock src.models.structured.try_run_structured directly, preserving the original behavioral assertions (identity pinning, circuit breaker, max_llm_papers cap, heuristic fallbacks, call-count checks).

What did not change

• Heuristic fallbacks at every LLM call site — a failed structured call still degrades gracefully, per paper and per stage.
• The circuit breaker, max_llm_papers cap, concurrency control, and identity pinning (paper ID/title always come from retrieval metadata).
• Ranking, retrieval, and evaluation behavior: golden-set metrics are byte-identical to main (verified by running the harness on both).

Verification

• pytest: 411 passed, 1 skipped (count drops from 442 because the deleted stack's 31-test suite is gone)
• ruff check src tests setups: clean
• scripts/check_docs_policy.py: OK (63 nav pages, 6 hubs)
• Golden-set harness: R@5 0.93, R@10 1.00, nDCG@10 0.97, MRR 1.00, citation validity 92% — identical on main and this branch