spike: sample merged PRs in agentv and classify eval-case yield

[christso]

Part of #1155.

Objective

Validate the premise behind the PR/issue-mining direction by sampling real merged PRs in this repo and checking how many convert to useful eval cases. If yield is too low, the dependent work (sub-issue on skill extension) should be rescoped or dropped.

Scope

• Take the 20 most recent merged PRs in main.
• For each, classify as one of:
  • useful: a plausible eval case — the PR title/body is a task spec, the diff represents a behavioral change an agent could reproduce.
  • not useful: typo fix, version bump, dep update, pure refactor with no behavior change, doc-only, or too small to be meaningful.
• Record classification with a one-line reason per PR.
• Report yield percentage.
• Recommendation: proceed, rescope, or drop.

Acceptance signals

• A short markdown note (in this issue, as a comment, or linked from the research repo) with a table: PR number, one-line summary, classification, reason.
• Yield percentage computed.
• Yes/no/rescope recommendation with rationale.
• No code changes.

Non-goals

• Not building any tooling to mine PRs programmatically — this is manual classification.
• Not evaluating quality of generated cases (we aren't generating any here).
• Not extending beyond 20 PRs unless initial signal is borderline.

Rule of thumb

• ≥50% useful: proceed with the skill extension as proposed.
• 30-50%: proceed but narrow the scope (e.g., filter by label, commit message pattern, PR size).
• <30%: rescope or drop — the premise doesn't hold for this codebase.

Blocks

Sub-issue for agentv-eval-writer extension (see #1155).