A crew reviewer's changes-requested queues the builder's follow-up within one tick, with the review as its brief #1213

A crew reviewer's changes_requested now queues the builder's feedback run in autonomous and assisted, and queues nothing in off. The new tests pass, but the preview screenshot and bun x tsc --noEmit are not done.

The step that dropped it. It was the active-run lookup in the pull.reviewed handler in src/crew/dispatch.ts. That lookup was activeRunFor(db, "agent", issue.id), which ignores whose run it is. The reviewer's own review run is still running on the pull when it posts its verdict. So the verdict was treated as a word said into that run and was never queued for the builder.

A reviewer's retry waiting in the queue did the same thing inside queueAgentRun (its "one queued follow-up is enough" check). The fix is that a feedback run only counts runs of the agent being woken.

I did not rule out emailVerified (agents are always verified), agentFor, the run cap or the breaker, but I did not test them. They stay in force.

What changed

• src/core/runs.ts: activeRunFor and hasQueuedRun take an optional actorId.
• src/crew/dispatch.ts: both the steer path and queueAgentRun scope to the builder for feedback. A reviewer's verdict in off returns before queuing. A person's comment in off behaves as before.
• test/crew-loop.test.ts, new block "a crew reviewer's changes-requested wakes the builder": in autonomous and assisted, a review posted while the reviewer's own run is running, with a retry queued behind it, queues one feedback run for the builder. In off, nothing is queued.
• Against main's src/, all three new cases fail. Against this branch, all three pass.

Already there, unchanged. The builder's brief already quotes the review body and its line comments (taskPrompt in src/crew/brief.ts). now over MCP already lists the run with why: feedback and the pull it is on. The pull page already shows queued runs.

Not done

• No preview screenshot: I can't take one from here. Please capture it from this pull's preview.
• bun x tsc --noEmit couldn't run, because the bun type definitions weren't found in the sandbox.
• The sandbox refuses unix sockets, so every test that calls serve() fails to start, and I couldn't run them here. That includes test/crew-loop.test.ts's first block, crew.test.ts and review-route.test.ts. My new block runs without serve().
• The case where a live steered session hears the verdict can't be reproduced in a test, since the steering map is private to runner.ts. I tested the queued-retry form of the same bug instead.
• main also queues builder runs in off on a reviewer's verdict, which the issue asks to stop. Only a reviewer's verdict is now gated; I left a person's comment in off alone.
• The three-round bound is left to the next issue.

To try. Run bun test test/crew-loop.test.ts -t "wakes the builder" on a machine that allows sockets. Or post a changes_requested as a crew reviewer on a crew pull and look at now or the pull page.

Closes #1175