The "one source of truth per concern" principle is what makes this tractable — refusing stateful chart defaults is where most homelab attempts fall apart. The Nessie write-audit-publish workflow reads like Git semantics for data, which is what makes reproducible training believable rather than aspirational. How do you size the ops overhead of keeping this whole stack healthy on a homelab node — is it a weekend project that mostly runs itself, or a constant maintenance tax compared to paying for managed?