Everything's safe when we have a reference implementation that people test against in a useful timeframe, and Linux is not currently that implementation. We churn far too quickly.

ACPI is formally specified, but there's no meaningful way you can write a specification that documents every single possible implicit assumption in an implementation. You could potentially write a Linux-based test suite, but we'd still need someone to run that on every single piece of ACPI-based ARM hardware (including the unreleased ones) towards the end of every kernel cycle.