v0.5.1 — live-proof defect batch (fp8/overrides, admission guard, GUI lifecycle)

[0.5.1] — 2026-07-06

Same-day follow-up to 0.5.0: every defect found while live-proving the 0.5.0 fleet
(GUI sweep, browser-driven training, vision serving), fixed and review-gated.

Fixed

• Per-load overrides survive restarts (#51) — solo loads persisted kv_cache_dtype
  / max_model_len / aliases only to the stacked-instance manifest; the primary model
  boots from config.json and silently lost them (a VLM loaded with kv_cache_dtype=auto
  came back on fp8 → corrupted output). All override fields now persist and reset
  correctly, with no cross-model inheritance on the launch config.
• fp8 KV cache off by default for multimodal models (#51) — fp8 KV corrupts VLM
  generation on GB10 (text models unaffected); vision models now default to auto,
  explicit operator fp8 still honored via a provenance flag.
• Stacked-load admission guard (#53) — stacked loads require explicit
  gpu_memory_utilization (400) and reject when the node's projected total exceeds
  0.9 (409) — the missing check let a default-sized second model hard-crash a node
  (unified memory). Gate runs before any existing instance is touched.
• Training accepts on-disk models (#53) — wizard cards submitted disk slugs
  (org--name) that HF rejects; wizard now sends canonical repo ids and the container
  builder maps any on-disk reference to its mount across all four disk layouts.
• Training/merge no longer require live PyPI (#53) — peft wheel vendored per node
  with a tolerant, import-guarded install chain (a review-caught event-loop freeze in
  the wheel fetch was fixed with executor offloading).
• Node-targeted launches (#53) — the launch form loaded on the head regardless of
  node selection; now routes through /api/cluster/load with the chosen node_id,
  plus a GMU input (validated on distributed launches too).
• Stacked instances visible (#53) — /api/nodes surfaces per-node instances;
  dashboard renders STACKED sub-cards (ports 8001+) with their own unload controls.
• GUI lifecycle polish (#54) — Server-view LLM "Load" button actually loads
  (was a "coming soon" toast); cancel-download failures surface a toast and recover
  (true event delegation on the queue); training wizard shows an honest empty state
  instead of hardcoded not-on-disk models; dead renderModels() view purged.

Added

• AutoData v2.2 live results (#52) — first live val-set run on the fleet:
  weak-solver lift +0.458 (0.500 → 0.958 on 24 held-out GT probes),
  exact-McNemar p=0.0005, 44% yield, early stop at target in round 1.

---