synthesize-reportlisted

# Role REPORT SYNTHESIS. Runs ONCE at end of pipeline. Sole output: `synthesis/REPORT.md`. # Invocation /ultra-analyzer:synthesize-report <run-path> # Inputs 1. `<run-path>/state/manifest.json` 2. `<run-path>/findings/*.md` — filter to those with PASS verdict in `validation/findings/*.json` 3. `<run-path>/topics/done/*.md` — topic specs for included findings 4. `<run-path>/topics/failed/*.md` — for Appendix B 5. `<run-path>/config.yaml` + `<run-path>/seeds.md` — for coverage assessment # Mode branch: SINGLE vs CHUNKED Before loading anything: ```bash du -sk <run-path>/findings/ ``` - If total findings size < 400 KB → SINGLE mode: load all findings, produce one REPORT.md in one pass. - If >= 400 KB → CHUNKED mode: - Pass 1: produce `synthesis/REPORT.draft-chunks/{p1,p2,p3}.md` (one chunk per priority tier, each reading only its tier's findings). - Pass 2: merge pass that reads only the three chunks + manifest + reconcile notes, produces final REPORT.md. # Triangulation (redundancy pairs) For each pair in manifest.redundancy_pairs: 1. Load both findings files. 2. Compare: - Direction: same sign? (+/+, -/-, or +/-) - Magnitude: within 2x? - Confidence: both >=0.6? 3. Write `<run-path>/synthesis/reconcile/<pair-id>.md` with outcome: - **REPLICATED**: same direction, within 2x magnitude → merge as single finding with doubled evidence weight. - **DIRECTION-CONFIRMED-MAGNITUDE-UNCERTAIN**: same direction, 2-3x magnitude divergence → report direction only,