An install while a chat runs (the onboarding plugins card, Settings, a terminal
install through the serve backend) made the plugin's MCP tools callable at once,
but its notice waited on the one-shot turn-note channel for the next USER
message. Onboarding is one long turn, so in a live run the notice arrived six
minutes late, after the model had already failed to use the plugin.
Notices are now queued per plugin key (agent/plugin_live_notes.py; a reinstall
or update replaces the plugin's older notice) and popped onto the next tool
result before it is appended and flushed: a new row, so no cached byte moves.
The notice goes after the content, outside the untrusted-data wrapper web, MCP
and browser results get. A turn-budget spill of that result keeps the notice on
the inline row. When no tool runs again this turn, the next user message still
carries it, and a delivered notice is never sent twice.
(cherry picked from commit 41cbb879a98a5bbcf57018a5c03d6207b57b819c)
