docs(agents): AGENTS.md's Maintenance summary contradicts the contract on promotion, external PRs and trusted authors

> 🤖 Generated by the Agentic Engineer

## Evidence

`AGENTS.md`'s `## Maintenance` section summarises the shared cross-repo conventions, and three of its claims now contradict the monorepo `AGENTS.md` they defer to. This matters because `AGENTS.md` is the file GitHub Copilot code review reads directly, and the one every agentic tool loads for this repo — a stale line here misleads every future agent and reviewer silently.

Read at `main`, `AGENTS.md` lines 359–364:

| `AGENTS.md` says | Superseded by |
|---|---|
| *"the maintainer's promotion to `ready` is the go-signal"* | Maintainer direction 2026-07-16 retired the human promotion gate. An agent self-promotes when it genuinely knows the change is ready: programmatically tested, green review at the current head, and tried/evaluated as a user. |
| *"**never merge external PRs**"* | Maintainer direction 2026-08-08: *"you are responsible to drive all prs to merge … Contribution PRs is also your responsibility just be careful!"* External PRs are now driven and merged; what stays forbidden is **running** their branch locally (review statically, let CI be the execution surface). |
| trusted authors are *"`devantler`, `ksail-bot`, `dependabot[bot]`, `github-actions[bot]`, and `renovate[bot]`"* | The portfolio trusted set also carries **`app/cursor`** (`cursor[bot]` on REST surfaces), added when Cursor Automation became a trusted PR author, plus the narrow path-scoped grant for the agent-skills updater App. |

The first two are the ones that change behaviour: read literally, an agent working in this repo would park every finished draft waiting for a human, and would refuse to merge an outside contribution it is now responsible for.

## Affected audience and impact

Any agent or reviewer operating from this file — the scheduled engineer lanes, Copilot code review, and any contributor pointing a tool at the repo. The failure is silent: the guidance reads authoritative and self-consistent, and nothing in CI notices that it disagrees with the contract it claims to summarise.

There is a second-order cost too. The section opens by saying the shared conventions are *"defined centrally in the devantler-tech monorepo `AGENTS.md`"* and that readers should *"follow that document rather than relying on the summary below"* — which is right, but the summary is what a tool with only this repo in context actually sees.

## Expected behaviour

The summary agrees with the contract it defers to, or it stops restating the parts that keep changing.

## Proposed direction

Two options, and the second is probably better:

1. **Correct the three claims in place.** Smallest diff, but it re-arms the same drift: this summary has now fallen behind three separate maintainer directions, so it will fall behind the next one too.
2. **Shrink the summary to the parts that are genuinely KSail-specific and link out for the rest.** The section already says the monorepo file is authoritative; leaning on that removes the copy that keeps going stale. Keep the KSail specifics that have no central home — the validate commands, the generated-artifact list, the experimental-gate mechanism, the task menu — and replace the restated cross-repo paragraph with a pointer.

Whichever is chosen, the same pass should check the other agent-instruction surfaces in this repo for the same three claims (`.github/instructions/**`, and the `CLAUDE.md`/`GEMINI.md` shims), since the portfolio convention is that a change to one updates every file that repeated the fact.

## Acceptance criteria

- [ ] No statement in `AGENTS.md` contradicts the monorepo contract on promotion, external-PR merging, or the trusted-author set.
- [ ] The execution guardrail stays explicit and unweakened: an external contributor's branch is never checked out, built, or run locally, whatever the merge policy says.
- [ ] Every other agent-instruction file in the repo repeating those facts is updated in the same PR.
- [ ] The result is either accurate or explicitly deferential — no restated cross-repo rule left without a source.

## Rough size

S. A prose change; the work is deciding between the two options and sweeping the sibling files.

Read the monorepo AGENTS.md contract first, then compare AGENTS.md lines 359–364 and inspect .github/instructions/** plus the CLAUDE.md and GEMINI.md shims for the same claims. Decide whether to correct or remove duplicated cross-repo guidance; done means all repeated statements are accurate or deferential and the external-branch execution guardrail remains explicit.

Written by the indexing model from the issue text.