Fix a bug with named NEVER_MATCH expressions

[markw65]

Found while working on #452.

The condition was backwards, so the error got misreported:

Before the fix:

% echo 'start "start" = []' | node bin/peggy.js -t "x"
Error running test
Error: Expected , or undefined but "x" found.
 --> command line:1:1
  |
1 | x
  | ^

After the fix:

% echo 'start "start" = []' | node bin/peggy.js -t "x"
Error running test
Error: Expected start but "x" found.
 --> command line:1:1
  |
1 | x
  | ^

There was already a test, but it was testing for the incorrect result...

[Mingun]

Unfortunately, the new error also not so good, because it is misleading. The grammar

start 'start' = []

cannot have any expectations because it just cannot match any input. Probably, it would be better to emit compiler error here instead. This grammar does not expect "start" rule at this position, it expect literally nothing. Even EOF is not expected there.

[Mingun]

Good catch! The bug probably was introduced during rebase while adapt initial PR to the new code after another merged PR.

Actually, the whole buildSequence could be replaced by [op.FAIL, nameIndex] if node always fail and there are no actions or semantic predicates (because they can have side effects), but anyway, it is better to do in a separate pass.

[markw65]

if node always fail and there are no actions or semantic predicates (because they can have side effects)

There are already comments about adding side-effect analysis so we can do that kind of thing...

[markw65]

This grammar does not expect "start" rule at this position, it expect literally nothing

But that was true whether or not it's a named rule.

Probably, it would be better to emit compiler error here instead

I guess thats ok if it's a start rule that always fails - but my examples only used start rules to exemplify the problem. You could use it as a catch all where a promising start goes off the rails:

echo 'start = [ab] fail / [cde]; fail "bzzt try again" = []' | node bin/peggy.js -t "ax"

Ok, its still not a great example, but it could be useful to have a named, NEVER_MATCH expression.