chore: enforce one-command agent handoffs

[rorybot]

What changed

• make ordered human command sequences explicitly count as missing automation
• require one verified, idempotent launcher to own cwd, preflight, orchestration, logging, exit status, and retry behavior
• state that host, container, device, credential, privilege, and physical boundaries do not waive the launcher requirement
• synchronize the rule across AGENTS.md, CLAUDE.md, and the reusable project template
• add a CI guard, invoked by the existing lane guard, to prevent instruction drift
• move issue Chore: review, reconcile, and template agent instructions #108's local mirror to In progress

Why

During #158, an agent hit host-exec exit 127 and delegated a multi-command host preflight checklist to Rory. The existing wording said to prefer one launcher, but did not make packaging that launcher a mandatory prerequisite for a human handoff or explicitly preserve the rule across execution boundaries.

Impact

Agents must now automate complex handoffs before asking Rory to run anything. If one physical, privileged, or interactive action is irreducible, the launcher carries the workflow to that boundary and emits one precise next action.

Validation

• bash scripts/ci/check-agent-instructions.sh
• HEAD_REF=chore/agent-instructions-108 BASE_SHA=origin/master HEAD_SHA=HEAD bash scripts/ci/lane-guard.sh
• git diff --check

Refs #108