fix(pr-review): credit human answers when dispositioning open questions

[Troublor]

Problem (Fixes B + C)

The pr-review action's open questions could never be closed by answering them. Concrete case — grafana-dashboard #15:

1. Reviewer asked 2 open questions (round on 20684249).
2. A maintainer answered both in a PR comment — one with live Prometheus output (count by (server_role, instance) …) proving the concern didn't apply.
3. The next round (on fix commit eaae1731) had that answer in its conversation.timeline, resolved the 2 findings correctly, but still returned disposition: open for both questions, reasoning only from the code delta:

   "reason":"Panel 13 query and legend unchanged in the delta; still not live-verified…"

The questions asked for a live-metric check the CI reviewer structurally cannot run (no Grafana access in the job), so a code-delta pass can never auto-close them — only crediting the human answer can. Result: questions sat "awaiting an answer" forever, and the review never reflected the maintainer's reply.

Root cause

Not the pipeline — the apply/render code already flips a question to ✅ Answered when the model returns answered, and the timeline already includes the human comment. It's the analysis prompt: the contract said "verify factual claims against the code," which the model applied to questions too and so discounted an out-of-CI human answer.

Fix

• action.yml (main + retry prompts) — weigh findings and questions differently:
  • A defect visible in the code still needs code verification before dismissal (a comment can't erase a reproducible finding).
  • A question is information the reviewer lacked, so a credible maintainer answer resolves it — even when it reports a live/runtime check this job can't reproduce — unless it contradicts the code. Disposition questions from the whole discussion, not just the delta.
• review_pipeline.py (C) — make the living status comment's guidance consistent with the review body: reply to a finding in its thread; answer an open question in a reply on this PR, replacing the contradictory "answer … in the review threads."
• test_review_pipeline.py — lock in the new prompt guidance. Full suite green (78 tests).

Scope note

This is the correctness half (B + C). The trigger-side change (A — let answering a question without a new push kick off a re-review, gated so the heavy reviewer doesn't run on every comment, reusing the same manifest for full history) is larger and lands as a follow-up PR.

🤖 Generated with Claude Code