2026-08-14 — Gates die with their eval, exactly like tasks; the probe that said otherwise could not fail

A field report from a tenant embedding Remontoire in a host application: a namespace-level (def gate (q/gate n)), re-created by an eval that drove the host's sweeping reload path (clj-reload), came back dead — the gate was born a child of the eval and cancelled with it. Verified here against quiescent 0.9.0, and extended to the narrower question the report left open: plain (require ns :reload) behaves identically. A probe namespace loaded via require inside an eval had its top-level task cancelled and its top-level gate's control-task cancelled at eval settle — within 205 ms by a compelled watcher. Parenthood is dynamic extent at creation; def lifts nothing out; every load path that re-evaluates def forms inside an eval mints eval-children. Boot and plain nREPL loads are safe for the same reason the eval chain is rooted (§9): no ambient task, so top-level resources are roots.

Gates have no task-space escape, mechanically confirmed: (q/compel (q/gate n)) throws IllegalArgumentException — compel accepts only Tasks — and the smuggle @(q/compel (q/task (q/gate n))) fails too, since the gate parents to the inner task and dies the moment it settles. The tenant conversation's mechanism-story ("compel insulates with a wrapper task") is falsified in the detail — there is no wrapper, there is a type error — while its conclusion stands stronger: tasks have two reload-safe shapes (compel at creation, capture the settled value), gates have exactly one — never re-create (defonce), or mint off-tree.

Mid-trial this suite briefly "falsified" the report — several def'd gates read alive minutes after settle — and the falsification was itself falsified: the probe was (try (q/gate-task g …) :alive (catch CancellationException …)), and enqueue on a cancelled gate does not throw. It returns an already-cancelled task ("Gate cancelled.") whose exception surfaces only at deref, get-now or grounding — with a stack pointing at Gate/enqueue, the construction site, not the use site. Direct isCancelled on the control-task told the truth throughout: every gate dies promptly. A liveness probe that cannot fail reads as survival.

That same property is the manual-worthy sharp edge, beyond the def-boundary gotcha itself: a dead gate is silent at the call border. gate-task succeeds, the task it returns is cancelled, and cancellation is a control signal — then/catch/handle never run, only finally sees it; grounding it cancels the surrounding expression instead of failing it. After a reload that re-created a def'd gate, every request through it evaporates as cancellation, invisible to error-shaped handling. Manual entry pending.