docs(agent): let a scheduled run merge behind an independent review

The runs produced five green PRs and no way to land them without a person. What makes that safe to automate is not CI — the author writes the tests — but a reviewer in a fresh context: last night's caught two guards that left the whole suite green when deleted, an over-narrow sweep, and an overlay covering the artwork it was editing.

So the gate is seven conditions, and a Kind/Feature is never one of them. It can exist at all because a push to main publishes nothing here: release.yml is workflow_dispatch and the four publishers key on v* tags, so a wrong merge costs a revert rather than a package nobody can unpublish. The section says so, and says it is void if that changes.

.pi/prompts/next-issue.md

@@ -133,10 +133,71 @@ status and `GET /api/v1/repos/yonlu/yellowjacket/actions/jobs/<id>/logs` for
 the log — and fix it. Two consecutive failed CI runs on the same cause: stop,
 comment what you know on the PR, and leave it for a human.
 
-**Do not merge.** Comment on the issue linking the PR, leave
-`Status/In Progress` on, and end the run.
+**Do not merge, and do not review your own work.** Comment on the issue
+linking the PR, leave `Status/In Progress` on, and end the run. Whether
+this merges is decided by a reviewer that did not write it — see below.
+A CI run is not a review: it proves the tests you wrote pass, which is
+exactly the thing an author is worst placed to judge.
 
 ## Finally
 
 Report in three lines: which issue you took, what state it is in
 (PR open / CI green / stopped and why), and any issues you filed.
+
+---
+
+# The merge gate
+
+This half is **not** run by the author. It is run against a PR by
+someone who has not seen the branch before, and it decides whether the
+work lands on its own or waits for a human.
+
+A push to `main` publishes nothing here — `release.yml` is
+`workflow_dispatch` only and all four publishers key on `v*` tags — so
+the cost of a wrong merge is a bad commit on `main` and the time to
+revert it. That is the whole reason this gate can exist. If that ever
+changes, this section is void.
+
+**Merge only when every one of these is true.** Any single no means
+leave it open.
+
+1. An independent review of the diff returns **MERGE** or **MERGE WITH
+   NITS**. `CHANGES NEEDED`, or a review that could not reach a verdict,
+   means a human looks.
+2. `CI / check (pull_request)` **and** `CI / e2e (pull_request)` are
+   `success` on the PR's current head. Ignore the `(push)` contexts —
+   they are `skipped` by design and Gitea folds `skipped` into a
+   combined state of `pending` that never clears.
+3. The PR is mergeable with no conflicts, rebased onto current `main`.
+4. The diff touches **none** of: `.gitea/workflows/`, `.releaserc.yml`,
+   `packaging/`, `build/`, `scripts/gitea-release.sh`,
+   `backend/database/sql/schemas/`, `backend/database/staleshape*.go`,
+   `go.mod`, `go.sum`. These either publish to somewhere a mistake
+   cannot be taken back from, or can destroy a database that a user
+   cannot rebuild.
+5. The issue is `Kind/Bug`, `Kind/Testing`, `Kind/Documentation` or
+   `Kind/Enhancement`. **A `Kind/Feature` is a design decision and is
+   never auto-merged**, however green it is.
+6. The diff is under ~600 changed lines across under ~15 files. Past
+   that, "a reviewer read it" stops being a claim anyone should take on
+   trust.
+7. The PR does not claim to have verified something no tier here can
+   see. A change whose evidence would have to come from a physical
+   device is reported, not merged.
+
+**When it merges**, use `{"Do":"merge"}` on
+`POST /api/v1/repos/yonlu/yellowjacket/pulls/<n>/merge`. Then **check
+the issue actually closed** — a squash or an edited merge message drops
+the `Closes` footer — and close it by hand with a comment naming the
+commit if it did not. `unclaim.yml` strips `Status/In Progress` on
+close; if the label is still there a minute later, strip it yourself.
+
+**When it does not merge**, say so on the PR in one paragraph: which
+condition failed and what would satisfy it. Leave the PR open, leave
+the label on, and file the review's substantive findings as issues so
+they are searchable rather than buried in a PR comment.
+
+**A nit is not a blocker, and it is not free either.** A `MERGE WITH
+NITS` merges, and each nit worth keeping becomes an issue. Do not fix
+nits on the branch: that is a second author pass with no second review,
+which is the thing this gate exists to prevent.