Berth v0.3.1

Fixed

• berth update and berth update --check always ask GitHub; before, a "no newer release" the
  harbor had kept for up to a day could answer them right after a release.
• A pre-release or pseudo-version build (0.3.0-rc, 0.2.1-0.2026...) now sees its release as
  newer; a pre-release is still never offered as an update.

v0.3 parks engines in system RAM instead of unloading them. A switch back to a parked engine skips the reload from disk. My workstation has 62 GB, and Strata alone pins 40 GB, so I can't measure parking at the sizes you run.

I'm looking for:

• 64, 96 or 128 GB of RAM with a 24 GB card. You'd park a vLLM or ComfyUI engine, switch away, switch back, and send me the times.
• AMD on Linux. That's the next release, through a vendor-neutral probe.
• Windows with an NVIDIA card. The Windows node leaves experimental in the same release, and one of you already runs it there. contrib/windows/check.ps1 compares Berth's per-process VRAM with Task Manager and writes a report.
• Two or more GPUs, any mix. Multi-GPU leases come in the release after that, and I want real topologies before I write them.

What a run takes:

1. Install with berth service install, or download a binary from the v0.3.1 release.
2. Run the failure scenarios in contrib/chaos (needs PyTorch; one scenario waits out a 10-minute drain limit).
3. Run the swap bench in contrib/bench/swap between two of your models.
4. File the results with the Drain report form on GitHub. A run where it falls over helps me as much as a clean one.

Each tester gets a row in the README's Tested hardware table, with your handle.