Huh, that's a really interesting approach. I just wrote my first Pratt parser a month ago, and one of the most annoying things was debugging infinite loops in various places (I had both tokenizer bugs where no characters were consumed and parser bugs where a token was emitted but not advanced). It's doubly annoying in Zig, because the default test runner won't print out stdout at all, and won't print stderr unless the program terminates by itself (Ctrl + C doesn't print). I resorted to building the test and running it manually, or jumping into a debugger to figure out recursion issues. It's working now, but if (really when) I run into issues in the future I'll definitely add some helper functions to check emitting invariants.
How about another way, which is memoization: at each position in the source code we never attempt to parse the same production more than once. This solves infinite looping as discussed by the author because the “loop” will be downgraded by the memoization to execute once. Of course I wouldn't literally use a while loop in code to represent the production. I would use a higher-level abstraction to indicate one-or-more or zero-or-more in the production; indeed I would represent productions as data not code.
This also has another benefit of work sharing. A production like `A B | C B` will ensure that in case parsing A or C consumes the same number of characters, the work to parse B will be shared, despite not literally factoring the production into `(A | C) B`.
That's a slick way, would you essentially have a second counter that you'd set to the current cursor whenever you use `.currentToken()` or something like that?
This also has another benefit of work sharing. A production like `A B | C B` will ensure that in case parsing A or C consumes the same number of characters, the work to parse B will be shared, despite not literally factoring the production into `(A | C) B`.