I was curious myself so I investigated in Godbolt https://zig.godbolt.org/z/84qYfvejq (see export fn xy at top), and indeed, the code size is a lot less. The compiler is able to figure out splitAny and splitScalar with constant values though and will generate the same code.

How about splitAll instead of splitFull? I think "all" complements the "any" well.
And then instead of "scalar" we could say "one". It sounds very simple... but that's kind of the point:
splitOne, splitAny, splitAll. Now the names are also the same length. Not sure though.

Alternatively, or in addition, what if we have a sensible default which would be simply called split? Maybe we can make splitFull the default so it becomes split as before... I do think that would be okay because that behavior is what most other split functions from other languages use too, such as JavaScript or Python. And from a logical perspective it makes the most sense too.

And then for any advanced uses one can use one of the two new functions you added here.

We already do this for other functions too. There is std.mem.lastIndexOf, and std.mem.lastIndexOfAny and std.mem.lastIndexOfScalar.

I agree with keeping the name of the new splitFull as just split. The name makes enough sense and imo its best to avoid any type of breaking change if its not necessary, even something as small as a name change.

The reason for deprecating split and tokenize in favor of only the explicitly suffixed versions is that the default behavior is not consistent between them: the default for split is what-is-now splitFull, but the default for tokenize is what-is-now tokenizeAny. Not sure how to square that circle without changing the default behavior of tokenize (doesn't sound good) or forcing at least tokenize to be explicit (i.e. deprecrating the un-suffixed version). I just went with deprecating both since at some point it will cause people to need to explicitly choose which to update their code to use (similar to the ensureCapacity -> ensureTotalCapacity/ensureUnusedCapacity split that happened previously).

Other than that,

• I think All as a suffix is a bit too ambiguous since it sounds like it could apply to the buffer param or the delimiter param (EDIT: Full suffers from this too)
• Scalar matches indexOfScalar and other std.mem functions, and same with Any (indexOfAny)

EDIT: Also, just for clarity, in this PR std.mem.split is an alias for splitFull and std.mem.tokenize is an alias for tokenizeAny with a deprecation doc comment on each. Added this bit to the OP since it's easy to miss it in the diff.