Afaik bitcode remains architecture-dependent, so bitcode deployments is mostly do they can re-optimise, and a convenient component of app slicing (if they end up generating the final bundle anyway, having more visibility into what’s what is useful).
I think re-targeting bitcode for a different architecture with the same pointer size and endinness is a pretty light weight operation, but I might be wrong. I'm pretty sure it's orders of magnitude lower overhead and less complicated than dynamically recompiling native code.
