- Fix `ParamsPartial` batching to match backend payload fields (`event_type` and `params`), so partial tool params are dispatched correctly.
- Ignore late `ParamsPartial` events after terminal tool states, preventing completed tool cards from reverting to `receiving` and staying expanded.
- Replace the `<thinking_end>` marker with an explicit end flag for thinking chunk events across backend transport and frontend handling.
- Add stream statistics collection and summary logging for AI stream handlers to improve streaming diagnostics.
- Sync frontend log level with `app.logging.level` at runtime so logging changes take effect immediately in the UI.