Preserve strict learned-weight loading across checkpoint metadata changes

When an older model checkpoint fails strict loading against a newer runtime, distinguish learned tensor mismatches from deterministic nonlearned metadata before relaxing validation. If official source proves a missing buffer is empty and irrelevant to the selected task, validate that task condition, supply exactly that default, and retain strict loading for every other state entry. Do not generalize this exception to trained parameters, nonempty buffers, incompatible shapes or unknown architectures.

Evidence supports a narrow compatibility method, not unchanged task accuracy across library versions. Successful loading and synthetic inference still require separate real-data parity evaluation when accuracy equivalence matters.

Applies when: a trusted older checkpoint differs from a newer model runtime only in verified deterministic task metadata. Keywords: checkpoint compatibility strict loading missing buffer detection metadata runtime weights