Diagnostics for metadata parsing is pretty bad!

[ErichDonGubler]

Right now, diagnostics are pretty bad. TODO: show examples, they're pretty easy to make, like with parens, empty intermittents, etc.

Current leads to fix this:

• I don't know how to force chumsky to say, "Actually, you should report issues you have with the inner property or subtest parsers, and not just bail with an error trying to parse another top-level test as a last resort." a la nom::prelude::cut.
• We might be able to get around this with multiple passes to parsing. I'm not sure what the performance ramifications of this are, but it might make parsing easier to reason about, and provide this as a benefit. We'd have to do some thinking about how to keep diagnostic spans aligned between parsing passes, too.

[ErichDonGubler]

A relevant chumsky upstream discussion: zesterer/chumsky#522

GitHub

Fail parser immediately (skip other choices) · zesterer/chumsky · Discussion #522

Let's say that I have this parser: it either parses a function definition or an expression. let fn_ = keyword("fn").padded() .ignore_then(ident_) .then(args) .then(fn_body); choice(( fn_, expr, )) ...