fix(training,models,web): slug mapping, offline peft, stacked-load admission guard, node-targeted launch, stacked visibility

[webdevtodayjason]

Summary

The five blocker/major defects from today's live GUI proofs on the 0.5.0 fleet, all reproduced with evidence before fixing:

• Slug mapping (blocker): training wizard's downloaded-model cards submitted on-disk slugs (qwen--...) that HF rejects (HFValidationError) — instant job failure. Wizard now submits canonical repo ids; the container-command builder defensively rewrites any on-disk model reference to its /ainode-models/<slug> mount across all four disk layouts the registry tracks.
• Offline peft (major): training/merge containers died without live PyPI (pip install && …). peft wheel now vendored into AINODE_HOME/wheels (fetched once via executor — a review blocker found the first cut froze the whole aiohttp event loop for up to 120s on cold cache), import-guarded tolerant install chain, &&→;.
• Stacked-load admission guard (major): the missing check that let a default-gmu stacked load hard-crash spark-2 (unified memory). Stacked loads now require explicit gpu_memory_utilization (400) and reject when the node's projected total exceeds 0.9 (409) — gate runs before any existing instance is touched (review catch), primaries unaffected.
• Node-targeted launch (major): the GUI launch form always loaded on the head; now routes through /api/cluster/load with the selected node_id, plus a gmu input (validated + forwarded on distributed launches too — review catch).
• Stacked visibility (major): /api/nodes now surfaces per-node instances; dashboard renders STACKED sub-cards for ports 8001+.

662 passed, 1 xfailed; ruff clean; node --check on app.js. Review-gated: 4 blocker/major findings across two lenses, all fixed with regression tests.

🤖 Generated with Claude Code

https://claude.ai/code/session_017NziXzqT1L9kj2T1byA3Ak