[shuk777]

What & why

Fixes found while accepting the 0.2.0 Windows arm64 release on an RTX Spark (N1X) machine.

Desktop app: keep enough commit memory for generation (b1fac34)

On a GPU that shares system memory, Windows charges VRAM allocations to the commit limit (RAM + page file). edge0-strata's --expert-cache auto fills the free VRAM, so the official Edge0 145B Q3K left ~0.5 GB of commit free after loading on a 128 GB machine (66.5 GB visible to Windows, 102.5 GB commit limit). The first request commits ~6.5 GB more, and the engine crashed with an out-of-memory error. It worked only after retries, when a dying engine still held VRAM at start-up and the new cache came out ~9 GB smaller.

Without changing the engine or the machine, the app now:

• waits after the ready line until the committed memory has been steady for 3 s (at most 30 s) before reporting the model loaded;
• when less than 10 GiB is left, reloads the engine with --strata-arg --vram-reserve-mib N sized to the shortfall (at most twice, up to 24 GiB), and remembers N per model in memory-reserve.json so later loads start with it;
• stops reloading a model when a reload frees no commit memory (a GPU with its own memory);
• shows an English warning when less than 8 GiB is still free, and adds the likely cause to the error when the engine dies mid-generation.

Only models run by edge0-strata get the reserve; macOS has no commit limit and skips the check. A smaller cache drops the least used experts: an engine that started with ~9 GB less cache decoded at 34.4 tok/s against 34.1–34.7 tok/s with a full one.

Desktop app (cfe48bd, 11d9abd): use the edge0-infer beside the app when no executable is set; read the engine status once a model is loaded so the sidebar names the device.

Docs (2b735d6, 57d6840): link the Edge0 145B and 120B Hugging Face repositories; drop the stale start.sh / model-bench.json note.

Area

• Engine, CLI or server (edge0-infer/)
• Model asset tools (edge0-infer/tools/assets/)
• Chat app examples (edge0-infer/examples/)
• Model docs (docs/)
• Other

How was it tested?

• macOS (Apple Silicon): cargo test in edge0-infer/examples/desktop/src-tauri (28 passed, including the reserve sizing and per-model memory tests), npm test (214 passed), tsc --noEmit.
• Windows 11 arm64, RTX Spark N1X: release build of the app (tauri build) succeeds. The crash and its cause were reproduced and measured there (commit limit, per-process GPU memory, Resource-Exhaustion-Detector events). Loading the official Edge0 145B Q3K with the reload is being tested on that machine now; results will be added here.

Checklist

• ctest --preset cpu -L "L0|L1" passes locally in edge0-infer/ (engine changes)
• Ran the higher test levels this change affects (see edge0-infer/docs/testing.md)
• Added/updated tests for new behavior
• Updated relevant docs / READMEs (incl. translations if user-facing)