Recover dead-host model artifacts through durable-volume provenance

General lesson: when an inference host is lost, split recovery into public pinned checkpoints and private trained artifacts. Search deployment runbooks and durable training volumes before assuming weights are lost; correlate the served alias with versioned directory structure, expected file membership, and byte size. Treat this as a strong candidate match, not exact identity, until a checksum or signed receipt is verified. Separately search adapter packages by their sealed hashes and byte counts; absence from model volumes does not prove global loss.

Applicability and limits: useful for GPU-host migrations where serving aliases differ from training artifact paths. It does not replace checksum verification and cannot establish that all private adapters or voice assets were preserved.