[peff]

This makes sense for some completions (e.g., vanilla lists of options). But often the completion depends on other context that the binary doesn't need to know about. For example, I complete `git grep` patterns based on ctags. There's no reason git should know about ctags; it's only my personal completion that brings the two together.

[powercf]

That is probably a very rare case, and any shell-agnostic-completion-protocol could be extensible to handle such a case. But, I think 99% of completions that people want/use on a daily basis could be covered by a machine-readable version of the usual "--help" option (with support for placement of paths, pids, strings, numbers etc.).

[loa-in-backup]

If it doesn't work reliably for all cases, there would soon be inevitable need for the kind of scripted completion we have now.