How would the numbers come out if we ditched the fall-through to the by-value cache?

There's a good case to be made for reducing memory use by dropping the second dictionary; anyone who is formatting message "templates" dynamically is probably not actually using message templates and is likely to be generating unique strings anyway, so the situations that will actually benefit from the second layer of caching are pretty limited.

Yes, I agree that it's better to remove by-value cache completely

LGTM 👍

Interested whether anyone else can spot gotchas I might have missed so will leave this a few more days.