I agree, I think syntax could be defined as "the whole problem" if you see it holistically. Approaches that avoid scaling an intermediary grammar from lower-level tokens like the Lisp or Forth construction of "the parse directly maps to a data structure, and the data structure has the semantics" are robust. The reason why they aesthetically offend comes down to familiarity and tooling: infix expressions "look like math" - they please someone with prior training - but often act to hide important machine-level details. And the grammar helps to spread around the architecture of the language so that a syntax error compiles as a different semantic. Versus an approach like APL with a richer token set, or a ColorForth that packs in more semantic value by assigning tags to each word and making that part of the presentation.

I've moved towards designing languages now that operate over CSV source. That adds an extra dimension while still enabling convenient editing - just turn off all the parsing behavior in the spreadsheet and you can edit it as "plain text". Although, column alignment isn't always desirable in this case.