follow-ups: critic still unavailable on the deployment; sessions don't use --summary/--default; mention-less connectors

From docs/reports/e2e-slack-test-2026-09-20.md (B9 remainder, room finding 5, N3). Three smaller items, filed together or split.

1. bug/investigation: the configured critic never completes (B9 remainder)

The issue-393 fix removed the misdiagnosed 120 s wrapper cap — confirmed live: the session ran the round in the background per the new guidance and critic policy printed the 900 s timeout. But the-loop critic run <critic> still exited 2 with "timed out" and no envelope, twice, in well under 900 s. The session recorded the round as unavailable in evidence/critic-review.md (honest, and the human gate accepted the gap) — but the deployment has had zero completed critic rounds across both e2e runs. Needs its own investigation on the harness side (cursor-agent invocation, env, or the wrapper's error propagation: the improved whose-limit-was-hit message did not surface — the output was a bare "timed out").

2. enhancement: the spawned session should pass --summary and --default

The issue-393 machinery for summary-led gate messages (the-loop ask --summary) and the one-tap "Defaults are fine" button (--default) shipped, but the live session used neither, so gate digests still paste the document (cut mid-sentence, "Answer with…") and the button never appeared. The spawn prompt / skill should instruct the session to attach a 2–3 sentence summary to every approval hook and a default to every ask.

3. docs: connectors that cannot render a mention

The Slack MCP connector types @the-loop as literal text, so listen: mentions (the default) hears nothing from it; raw <@BOT_USER_ID> … markup works. One line in the channels doc: "an integration that cannot render a mention entity should send <@BOT_USER_ID> … verbatim, or declare the room --listen all". Related polish: session.autoclosed logged merged: false for an issue whose delivering PR merged three minutes earlier, and the hosted service's process log goes to /dev/null on this deployment, making every logger.debug diagnostic unreachable in operation.

Fixed (branch fix/n1-selection-grammar-refusal-reason)

1. Critic timeout surfacing — the round's error now leads with the observed duration vs the-loop's limit, and when a round dies well short of the limit it says plainly that the cap that fired is the CALLER's tool timeout, not the-loop's (run it as a background process). The critic executable itself failing on the deployment (cursor-agent producing no envelope) is a deployment-side problem no code change here fixes; the session records an unavailable round honestly, which is the right behaviour.
2. --summary / --default adoption — skills/the-loop/reference/collaboration.md now tells the session to give every the-loop ask a --summary and, where there is a safe one, a --default, so the room leads with the agent's account and offers the one-tap button.
3. Connector doc — docs/capabilities/channels.md documents that a mention-less integration must send raw <@BOT_ID> markup or declare the room --listen all.