v0.9.4

The AI now sees an automation's recent runtime errors from the HA log on its first review attempt — combined with self-correction, a 'no fix' outcome means the model had all available evidence. The panel shows recent Helmsman and automation log entries in a collapsible table. LLM proposals are sanitized of null entries before validation.