apex-ray-improvelisted

# Apex Ray Improve ## Purpose Run a post-merge learning pass. The goal is not to review the PR again; it is to decide whether Apex Ray should learn from what happened through repo memory, rules, eval labels, telemetry interpretation, coverage tuning, or config changes. ## Process - Identify the PR number, repository root, base branch, merge commit, and whether the PR is merged. If the PR is not merged, label the output as a review-feedback learning pass instead of a post-merge pass. - Collect PR signals with GitHub CLI when available: `gh pr view <number> --json number,title,state,mergedAt,mergeCommit,baseRefName,headRefName,author,comments,reviews,files,url` and review-thread comments from `gh api repos/<owner>/<repo>/pulls/<number>/comments --paginate`. - Separate Greptile comments, human comments, CI/bot comments, and author follow-up commits. Treat comments as evidence, not ground truth. - Inspect Apex Ray artifacts when present: `.apex-ray/reports/`, `.apex-ray/evals/cases/pr-<number>/`, `.apex-ray/evals/runs/*/pr-<number>/`, `.apex-ray/eval/labels/`, local review telemetry, and PR eval telemetry. - If a comparable eval case is missing and the user asked for a fresh analysis, capture or replay narrowly with `apex-ray eval capture-prs --pr <number>` and `apex-ray eval run-prs` rather than running a broad historical benchmark. - Compare external findings with Apex Ray findings. Call out missed issues, duplicate findings, false positives, findings outside scope, and tru