PROD-AI-REPORT-08 (LLM Receives Empty Data) & PROD-AI-REPORT-09 (Analysis Bypasses Quality Gate)

Release digest

PROD-AI-REPORT-08 (LLM Receives Empty Data) — Root cause: ai_report_generator.py created a plain psycopg2 cursor (tuples) but accessed rows as dicts, silently falling into except with empty stats/reviews. Three related bugs fixed:
- Added cursor_factory=psycopg2.extras.RealDictCursor for dict-based row access (backend/ai_report_generator.py).
- Removed broken dict(zip(columns, stats_row)) conversion that produced {"col_name": "col_name"} once cursor returned dicts.
- Added NULL/NaN-safe str(x or "").strip() for review text fields that previously crashed on None.strip().
PROD-AI-REPORT-09 (Analysis Bypasses Quality Gate) — POST /api/analyze now routes through ReportOrchestrationService.resolve_sections() with force_regenerate=True instead of saving raw LLM output directly to DB (backend/routers/analytics.py).
PROD-AI-REPORT-10 (No Force-Regenerate) — Added force_regenerate parameter to resolve_sections() so "Run AI Analysis" always generates fresh content, skipping cached hallucinations (backend/services/report_orchestration_service.py).

Added tests/test_ai_report_data_integrity.py (9 tests):
- Stats serialization correctness (numeric values, not column names)
- NULL/NaN-safe review text assembly (None, NaN, all-NULL)
- Quality gate hallucination blocking (Trust Score N/A, no employee reviews)
- Orchestration force-regenerate contract
Full targeted regression: 9 passed.
Added docs/reports/2026-03-07_PROD_REPAIR_SEARCH_SENTIMENT_AI_REPORT.md (updated with PROD-AI-REPORT-08/09/10 closure).