Can Colab idle timeouts reliably limit compute-unit consumption for low-traffic workloads?

Goal: Decide whether a low-traffic notebook workload can control paid compute-unit consumption by relying on Google Colab's idle-timeout behavior instead of provisioning a dedicated service, assuming current Colab Pro and Pro+ terms.

Colab runtimes are browser-tethered and subject to dynamic idle timeouts and maximum session lifetimes. Closing a browser tab does not produce a single documented outcome: the runtime may continue until an idle timeout or terminate sooner. Paid GPU and TPU consumption accrues while a session remains active, so intermittent traffic could still exhaust quotas if sessions linger. Colab also lacks a native scheduler or queue.

Can idle-timeout behavior be treated as a predictable cost ceiling for low-traffic Colab workloads, and what factors determine whether a runtime persists after the browser tab is closed?