- update the watcher design and automation guide to treat --categorize as sequential ATVM sub-runs rather than one parent run with internal phases
- document that categorized runs should send one Mattermost status per completed grouped sub-run instead of one parent-only final post
- add a --categorize option to the watcher start helper so categorized mode is explicit in watcher startup
- update the watcher implementation to track categorized sub-runs separately, write per-subrun state, and post each completed grouped run once
- update the watcher cancel helper so it writes a final CANCELLED state into state.json before stopping the service
- record cancellation timestamps and a cancellation note in the watcher state file for clearer post-run inspection
- update the watcher service docs so the documented cancel behavior matches the state-file handling
- add the per-run ATVM watcher service package under atvm/watcher-service, including the Python watcher, systemd template unit, helper scripts, and deployment docs
- document the watcher-service install and operating model, including one-run-per-instance behavior, Mattermost posting rules, and the best-practice /opt/atvm-watcher-service install path
- clarify ATVM run approval semantics so `approve` means run without watcher and `approve with watcher` means run and start the watcher
- update the ATVM automation guide and AGENTS rules so watcher usage and approval behavior are explicit and consistent
