Astra update reports fixes for early stops and self-checking problems

A September 12 quality update for Astra says several issues have been fixed: skills written for previous models were triggering too often or preventing the model from checking its work, and an opt-in context-management experiment could cause early stops or replies to older messages. The update says the experiment was disabled and badly configured engines that caused measured quality degradation for some traffic were removed. It says these fixes and smaller improvements should bring more consistent follow-through, better tracking of the latest message and better checks on work.