The MCP plugin sends the rollbacks it saw with its next task response
and then forgets them. When that response came after the server had
given up on the task, the server dropped it and the agent never heard
about the rollback. A reload of the tab or of the plugin lost them too.
The server now sends the rollbacks of such a late response back to the
plugin, which reports them again with its next response. The plugin
keeps the rollbacks not reported yet in its local storage, one entry
per file, so a reload does not lose them. Both MCP plugin manifests ask
for the allow:localstorage permission for that.
The backend rejects a save after the call that made the changes has
returned, and the workspace then undoes them. A plugin, or the MCP
agent, went on working with shapes that were no longer there and never
learned why.
The rollback event now names the files that lost commits and the pages
and page shapes their changes touched, and plugins get it as the new
`contentrollback` event (`ContentRollback` type) through a subscription
to the store stream, so no rollback is dropped by debouncing.
The MCP plugin keeps the rollbacks it is told about and sends them with
its next task response, failed or not. The server collects them per tool
call and puts a warning first in that tool's response, naming what was
affected; the instructions tell the agent to read the pages again and
redo what is missing.
