Benchmark the PR experience: script the marginal-client release checklist #208

Context

The review-quality benchmark (#179, ADR 0008) answers "does marginal find the right bugs?". It doesn't answer "does marginal behave right on a PR?": a fixed finding gets resolved, a disputed one gets closed with credit, a draft or lockfile-only PR gets no model call, a PR can't switch off its own review. Today those are checked by hand, by opening PRs from the marginal-client release checklist and reading the result (ADR 0007). Nothing is scored, and a regression shows up only if someone happens to run the right scenario. #140's last run left scenarios 2–5 unchecked.

Proposal

Turn the marginal-client checklist into a scripted, scored suite that runs on the same triggers as review quality: release PRs and workflow_dispatch.

Each scenario is a script of steps on marginal-client plus checks on what GitHub shows afterwards:

# tests/integration/pr_experience/scenarios.yml
- id: fix-after-finding
  steps:
    - open_pr: {branch: bug-round-down}          # scenario 1's bug
    - expect: {inline_findings: ">=1"}
    - push: {patch: fix-round-down.patch}
    - expect: {thread_resolved: true, reply_contains: "✅ Resolved"}
- id: lockfile-only
  steps:
    - open_pr: {branch: lock-bump}
    - expect: {model_calls: 0, summary_contains: "1 file(s) not reviewed"}

The runner opens each PR, waits for the review, checks the expectations through the GitHub API, then closes the PR and deletes its branches, so marginal-client is left as it started.

Scope

• Start with the checklist scenarios that need no human: 1 (real bug), 3 (clean), 4 (fix after a finding), 6 (lockfile-only), 7 (draft), 8 (opt-out label), 9 (secret in a URL), 10 (PR silences its own review).
• Scenario 5 (disputed finding) needs a reply from a second account; follow-up.
• Soft gate like review quality: warns in the run summary, never fails the build.
• Results go to the run summary and score.json, next to the review-quality results.

Cost

Scenarios 6, 7 (draft part) and 8 make no model call. The others are one review each, plus one more per later push (scenario 4), so about 8 reviews per run. Run it only on release PRs and on demand, like review quality.

Acceptance criteria

• One command (make pr-experience) runs the scenarios above on marginal-client and writes a pass/fail table.
• The workflow runs on a release PR and on demand, and warns without failing.
• A run leaves no open PRs or scenario branches on marginal-client.
• The marginal-client README checklist points to the suite for the scripted scenarios.