fix(models,engine): persist solo per-load overrides across restart; kv fp8 off for multimodal

[webdevtodayjason]

Summary

Fixes the two fp8/override defects found live on the 0.5.0 fleet, plus a small AutoData report polish.

• Solo-load override persistence: the primary model boots from config.json, which never received per-load overrides (feat(launch): served-model-name aliases + per-load config overrides #41 persisted them only to the stacked-instance manifest) — a VLM loaded with kv_cache_dtype=auto came back after restart on fp8 and produced corrupted output. Every solo load now persists the full override set to NodeConfig, with absent fields reset to defaults (no cross-load inheritance).
• Adversarial review caught + fixed a would-be regression: the first cut leaked the previous primary's overrides (quantization, trust_remote_code, served_model_name…) onto the next model's launch config via the shared mutable config — reproduced against the branch, fixed via _resolved_overrides() applied to the actual launch config, with regression tests inspecting the real backend object. Also fixed the live-vs-persisted divergence on same-model reloads.
• Multimodal fp8-KV auto-skip: fp8 KV corrupts VLM generation on GB10 (verified: garbage on fp8, perfect on auto; text models unaffected). The NVIDIA backend now defaults multimodal models (vision_config / VL architectures in the local config.json) to kv_cache_dtype=auto. Explicit operator fp8 is still honored via a new kv_cache_dtype_explicit provenance flag. Not-yet-downloaded repo-ids keep the fp8 default (documented ceiling).
• AutoData valset meta reports now surface best_lift/best_score/best_significant.

639 passed, 1 xfailed; ruff clean. Review-gated (1 blocker + 2 majors found, all fixed with regression tests).

🤖 Generated with Claude Code

https://claude.ai/code/session_017NziXzqT1L9kj2T1byA3Ak