- annotate queued and interrupt inputs with source/output channel context
- drain all pending interrupts instead of only one
- avoid double-processing by routing interrupts exclusively:
active LLM -> intercept channel, idle -> input queue
- keep interrupt source/channel in payload and system context