my point is that even the "type-independent compressions" (LZW,
deflate, etc) are at best ineffective (and at worse pessimal)
when applied to different kinds of media types for which they
were designed. you therefore don't want to gratuitously apply
them to random content-types.
Actually this isn't true -- for one thing, there's still an advantage to be had
by compressing base64 at the transport level. And for another, the
type-independent conversions sometimes do manage to squeeze a little more juice
out out that the type-specific conversions didn't manage. And when that's not
possible, they are smart enough to tell this is the case and opt out; the
result is a little CPU lost on the sending side (very cheap), effectively no
bloat, and no significant CPU loss on the receiver.
The quality of these things really has gone up over the years, so with the
proper use of modern options pessimal outcomes are eliminated.