1. You have users.
2. You can demonstrate a marked improvement in performance and/or error reporting with the hand-written parser.
Otherwise, deriving a parser from a grammar should absolutely be the default. Writing parsers by hand is error-prone, causes security vulnerabilities[0], and is very tedious work.
[0]: http://langsec.org/
When you outgrow it, you'll know - and by then, hopefully you'll have funding/users/contributors/etc to support the dev effort of rolling your own parser.
The problem we're now facing with Happy is that it doesn't really have support for error recovery. We're trying to implement LSP so recovery is important in order to have proper syntax highlighting and error reporting in an IDE.
Wondering how well ANTLR covers this, I was quite surprised to find out that Happy has so little support for it.
A parser for a Forth is a one-liner or a short loop in most languages.