Fix panic when sub-agent tool completes after primary agent's turn ends

[tcdent]

The ToolEvent::Completed handler was unconditionally draining injectable
notifications and calling finish_turn()/begin_turn() on the transcript,
even for sub-agent tool completions. When the primary agent had already
finished its turn, this created an orphaned active turn that nobody would
finish, causing a panic ("Cannot begin turn: previous turn not finished")
on the next user submit.

Guard the notification injection behind an is_primary check so only the
primary agent's tool completions manipulate the transcript turn lifecycle.

https://claude.ai/code/session_01RimYSQyprjibcxqX8XBmrD