Bounded local lab · no network requests

Budget every training byte

Compare ideal per-rank persistent state across ZeRO stages, then add declared Stage 3 gather, activation, runtime, and contingency bytes. This is a transparent lower-bound model, not a framework fit prediction.