Patrol thread can hang forever and freeze latest_image

The camera API can serve frozen images for hours while everything looks healthy.

Seen on 2026-09-16 at chateau-eau-douadic (Pi 192.168.255.170, two Reolink cameras):

• latest_image returned the exact same JPEG for all 8 camera/pose pairs over 12 minutes
• camera 192.168.1.11 served a night IR frame at 12:05, while a live capture on the same camera at the same second returned a sharp daylight image
• the engine kept scoring those dead frames and logging No wildfire, so the site was blind with no error anywhere

Cause

The patrol thread was alive but blocked in a PTZ call. In camera/adapters/reolink.py the 9 requests.post calls have no timeout, so move_camera, called by patrol_loop, can wait forever. The stop flag was set meanwhile and the frozen thread never saw it.

From there the API cannot recover:

• start_patrol only checks thread.is_alive(), so it answers already_running and never starts a new thread
• patrol_status also checks the flag, so it answers patrol_running: false
• last_images is never updated again, and latest_image carries no timestamp, so nobody can tell the frame is old
• the stuck detector skips when the patrol is not running, which is exactly this case

What to change

1. Add a default timeout to every request in the Reolink adapter, with one helper instead of 9 call sites.
2. Share one "is patrol running" check, and let start_patrol replace a thread that is alive but has its flag set.
3. Store the capture time with each image in last_images, expose it, and let the engine refuse a frame that is too old.
4. Do not disable the stuck detector when the patrol is not running.
5. Engine side: core.py stops the patrol only when it is already stopped (inverted condition), and is_day is overwritten at each pose, so the last camera of the loop decides day or night for the whole site.

Workaround

Restart the camera API container. The engine starts a clean patrol on the next cycle.

Start with camera/adapters/reolink.py and trace move_camera through patrol_loop, then inspect the patrol state and image handling named in the issue. Read core.py for the inverted stop condition and is_day overwrite. Done means stalled requests recover, patrol status and replacement behave consistently, image age is exposed and enforced, the detector still checks stopped patrols, and day/night state is correct across poses.