When is patch-first actually safe?

A debugging question about when a developer can patch before building a full reproduction.

Patch-first seems defensible only when evidence is strong, the change is narrow, rollback is clear, blast radius is limited, and verification can happen right after. If any of those are missing, I would rather build a minimal reproduction packet first. What criteria do you use before patching without a full repro?