v0.2.1

Updating is now one step. After uv tool upgrade inference-vitals, run inference-vitals start: it
sees that an older version is still running and restarts it. inference-vitals --version and the first
line of status show the version; so does the page's footer.

start tells the truth about what started.

• If another program already holds a port (often Prometheus or macmon run by brew services, or a
  second copy of inference-vitals), start names the port and the usual fix and starts nothing there,
  instead of saying "started" while showing someone else's numbers.
• A service that stops at once is reported with the last line of its log.
• start exits with an error when anything did not start, so scripts and AI agents can tell.
• Cutting its output short (inference-vitals start | head) no longer stops it halfway.
• On Linux without an NVIDIA GPU, start works (0.2.0 crashed there) and status says why the GPU
  exporter is not used.

The page

• Loads cleanly the first time: its server dropped most of a first visit's ~40 requests, leaving tiles
  empty until the next refresh.
• Charts with a limit end at the real limit: "Mac memory and swap in use" tops out at your Mac's own
  memory, not a round 100 GB. Other charts round up less.
• The GPU memory chart shows a fine-tune's peak against your Mac's GPU limit (read from Metal, as MLX
  reads it), even with no model run reporting one. If a fine-tune swaps, you can see why.
• The battery chip follows plugging in and out within seconds (it could lag a minute and a half).

Fine-tuning

• A fine-tune that was killed no longer holds the page for a week: an hour after it stops reporting, the
  page goes back to the model run. A finished one stays for a day, so last night's run is still there in
  the morning.
• A slow trainer shows up straight away (mlx-vlm on an 8B showed nothing for its first 10 steps), and
  its time per step and time left are the trainer's own (they read 9.6 s and 28 minutes for an 85 s
  step and a 4-hour run).
• During a fine-tune, Favourites shows its loss, its speed, its memory against the GPU limit, and swap.
  For a model run it now starts with four charts (the machine's memory joins), so its grid is full.

Under the hood: idle power is worked out once a minute by Prometheus (the costliest query, before);
fewer requests per refresh; the code is split into queries.py, grafana.py and monitor.py. Tested
on macOS and in a Debian container, and with Grafana 13.

v0.2.0

Full Changelog: https://github.com/urbanmorph/inference-vitals/commits/v0.2.0