validate-findingslisted

# Validate Findings Opt-in coordinator-only LLM validation pass over a consolidated v4 report: adds AI assessment, verdict, and confidence per finding. Under the v4 contract producers emit `likelihood`/`impact`/`relevance` themselves, so the typical run leaves the floats untouched; re-estimate them only when the consolidator left them absent (partial producer output that still satisfied the schema). NOT part of the automatic review pipeline — invoke after `consolidate_reports.py assemble` when a triage-quality validation pass is wanted. **Argument**: `$ARGUMENTS` — path to the consolidated `report.json`. Edited in place. ## Inputs - A consolidated v4 report on disk (output of `consolidate_reports.py assemble`). - The producer commit (when `metadata.commit` is present) for best-effort source lookup via `git show`. ## Per-finding loop **Never pre-build an id-keyed assessment lookup before `consolidate_reports.py assemble` runs.** `assemble` calls `assign_ids()`, which sorts each section's findings by `overall_severity` desc (then integer `severity` desc) before assigning sequential `CMT-`/`SEC-`/`CODE-` IDs — which finding lands in an ID slot depends on the severity sort, not fetch order. Run the loop in-place on the already-assembled report, reading each finding's own current fields — never values pre-computed by assumed id. If bulk pre-computation is unavoidable, key the lookup by a field `assemble` never mutates (`comment_id`, `thread_id`, `location`, or a content hash