Test connection in the AI Insights settings lists models at the Ollama
address it started with. Editing the address while it ran did not
retire it, so the late reply reported "Connection works" and filled
the model list for an address that was never tested; a finished result
also kept claiming success after the address was edited. A provider
change already retired the request.

An address edit now retires a running test and clears its result.

Test: tests/test_ai_insights_settings.py copies the page's
retire/run/report/accept functions and the endpoint handler into a
QtTest with stubbed processes; both endpoint scenarios fail on main.

Documentation: ai-insights.md (Test connection, verification list) and
development.md. No CHANGELOG entry: AI Insights is unreleased and its
Added entry still describes the behavior.


Claude-Session: https://claude.ai/code/session_01TvjnVJPgkwU6fxn8uL6N8c

Co-authored-by: Claude <noreply@anthropic.com>