Bound Autopilot's verify rounds and state the program size before fan-out

[michael-denyer]

What changed

• Only findings the diff causes go back to the owner. Step 4 of playbooks/autopilot-full.md sent every proven finding to the owner. A reviewer can always find one more defect in the files a PR touches, so a PR reached a clean verdict only when a reviewer found nothing. The root now files a defect that trunk already has as an issue, or lists it as a declared gap, and does not hold the merge for it.
• Rounds have a ceiling. The first round runs the full swarm. A later round re-runs the gates and the probe behind each blocking finding, with one review lane scoped to the fix-forward diff. After two fix-forwards the root reports the PR blocked and the operator chooses what happens next.
• Review lanes report what the diff breaks. The brief asks for no mutation or fuzz campaign unless the operator asked for one.
• The root states the program's size before the first owner starts. It names the owners, the first-round lanes and the two-round ceiling, and waits for the operator's go above three owners.
• The rule is pinned. tests/skill-rules.test.mjs pins the three sentences, and the tools/forks.json entry describes the wider fork.

Why

One seven-PR program ran under the old step 4. Each PR went through up to two audits and two fix rounds. The session transcripts record 67 subagents and about 815M cache-read tokens. After the first round, most findings were defects that main already had.

Verification

• bun test tests/skill-rules.test.mjs gives 20 pass, 3 fail against main's playbook and 23 pass, 0 fail with the edit.
• bun tools/generate.mjs --check passes.
• The dry run of tools/sync.mjs for pstack at the pinned SHA lists the file as a policy fork and reports no failure.
• bun test tests/ gives 953 pass, 35 skip, 24 fail on macOS. All 24 are in tests/worktree-audit.test.mjs, which fails the same 24 on main on this machine. The other test files together give 701 pass, 22 skip, 0 fail.

Open risks

• The rule is prose. Nothing stops a root from running a third round.
• I did not check playbooks/orchestrate.md, autopilot-stack.md or shipping.md for the same loop.
• No Autopilot program has run under the new wording.

Proposed CHANGES entry

Autopilot-full sends an owner only the findings its diff causes, runs the full swarm in the first round only, stops a PR at two fix-forwards, and states the program's size before the first owner starts.