the open in immich button in the photo of the day plugin did nothing, and the cause was boring and classic: the url pointed at a private tailscale ip instead of the public domain. the backend default got switched to the proxy domain found in the nginx proxy config, and because the state file still held a cached pick with the old raw ip in it, the fix got a serve time rewrite too. any cached url gets mapped to the domain and saved back, so tomorrow's pick is correct even before the backend reloads. the proxy route verified over https and the button works.
then the scare: vram usage looked oddly low, which usually means the machine learning service silently fell back to cpu. it had not. the gigabyte resident on the gpu was the ml worker itself with its preloaded models, and they are the small variants on purpose, the base clip model, the face detection models, and the mobile ocr model. added up, they land right around a gig. that is the correct steady state, not a fault.
the earlier scare of a nearly full card turned out to be container ram shown in a docker stats column, not gpu memory. the weights live in vram now, so host ram shows almost nothing. vram only climbs to two or three gigabytes while the ml actually processes, smart search, face indexing, ocr, and then settles back. no fallback warnings, no restarts, gpu idle at zero percent utilization. the button was the real bug, the vram number was the healthy one.