task: deterministic pre-flight for the PR review agent

[serina-mcfall]

Parent PRD

#109

Objective

A script that, given a PR number, emits a structured record of everything about that PR which can be decided without a model — so no reviewer spends a token on a question with one correct answer.

Definition of done

• Given a PR number, the script emits machine-readable output naming: the merge-base diff, the PR body and title, whether a closing keyword is present, the labels, and every required check with its conclusion.
• It resolves the nearest AGENTS.md or CLAUDE.md to each changed path and lists them, so a later reviewer reads the rules that actually govern each file.
• A missing or unreadable input reports SKIP with a reason and never PASS — absence of evidence is never reported as evidence.
• It exits non-zero when a required input cannot be read, so a caller cannot mistake a broken run for a clean PR.
• A control suite covers each check, including one case per SKIP reason, and each control fails if its check is deleted.
• The script uses no model and makes no inference call.

Impacted components

• launchpad/ — new script path to be decided in the PR
• .github/workflows/ — invocation, only once adr: where the PR review agent runs — GitHub Actions or through Buzz #110 decides where it runs
• .github/PULL_REQUEST_TEMPLATE.md and launchpad/AGENT_PR_TEMPLATE.md — read to determine body compliance

Out of scope

Any judgement about whether the code is good. No findings, no severities, no model call. This produces facts for later stages and nothing else.

Posting anything to the PR — publishing belongs to the task that owns the single review comment.