Who Guards the Update? Transactional Publication Semantics for Continually Learning Embodied Agents

A better controller can still be installed unreliably. A predecessor invocation may remain live after a candidate becomes current, and a stale request may target a superseded installation. We define a conditional installation contract for one executor slot: check the expected installation, admit stateless candidates, revalidate authority, exclude active invocations, and preserve accounting on rejection. We compare quiescence-free atomic publication (QF-AP) with guarded replacement (GR). Across 200 primary schedules per protocol, QF-AP produced retired predecessor commands in 72 schedules and GR in none. In a separate 50-schedule above-period stress stratum per protocol, QF-AP produced old/new invocation overlap in every schedule; GR prevented overlap but skipped service in every schedule. A corrective replay exposes the counter-hazard: QF-AP publishes immediately but permits one retired command; GR prevents that command but delays publication by 1.137 ms. Here “transactional” refers only to software-authority publication; it implies neither crash atomicity, persistent recovery, nor reversible physical effects. The experiments use hosted guards and simulated control, without foundation-model adaptation, privileged controller execution, or downstream command fencing.