v0.81.30

Added

• GCN/Vega (wave64) safe mode, opt-in on the official engine: type
  GGML_METAL_WAVE64_SAFEMODE=1 in Extra arguments (any KEY=VALUE there is now passed
  as an engine env var). Validated coherent on an RX 580; off by default, no-op on
  Apple/RDNA. Extra arguments now route uppercase KEY=VALUE tokens to the environment.
• Per-session server logs with date-and-time filenames (survive a crash) that
  auto-delete after 3 days, plus a self-contained log header (version, engine, model,
  GPUs, args/env). Start/stop the server and pick a model from the Logs tab.
• System info on the dashboard: Mac model and macOS version.
• Reset options to defaults button (keeps models and the models folder).

Fixed

• eGPU full speed: forces VRAM-resident buffers on external GPUs (was streaming over
  Thunderbolt at ~0.8 t/s), automatically when an eGPU is selected, with a manual toggle.
• CPU-threads selector capped to the machine's actual thread count.
• Port no longer shows a thousands separator (8080, not 8.080).

Changed

• UI polish: dashboard cards aligned to equal height, server quick-settings (port,
  discoverability) without layout jumps, cleaner model estimate rows.

Full Changelog: v0.81.29...v0.81.30