Emit {0,n} for at_most so the regex compiles on the outlines-core backend

[lollinng]

Problem

Term.at_most(n) / at_most(n, term) (the QuantifyMaximum DSL term) compiles to a regex of the form (...){,n}:

return f"({to_regex(term.term)}){{,{term.max_count}}}"   # -> "(a){,5}"

Python's re accepts the lower-bound-omitted {,n} form, but the default generation backend — outlines-core (Rust regex / regex-automata) — does not (it only supports {n}, {n,}, {n,m}). So any structured generation using an at_most quantifier raises when the Index/Guide is built:

>>> from outlines_core import Index, Vocabulary
>>> v = Vocabulary.from_pretrained("gpt2")
>>> Index("(a){,3}", v)
ValueError: Failed to build DFA error building NFA ...
>>> Index("(a){0,3}", v)   # OK

The whole at_most quantifier is effectively unusable with the default backend.

Fix

Emit the equivalent (...){0,n}. Semantically identical (0 to n repetitions), but valid for both re and outlines-core. A unit test asserted the old (a){,5} output; updated to (a){0,5}.

Verification

to_regex(at_most 5)            -> "(a){0,5}"
outlines_core.Index("(a){0,5}") -> builds OK   (previously "(a){,5}" failed)
pytest tests/types/test_dsl.py -k "to_regex or at_most" -> 2 passed

[RobinPicard]

Good catch, thanks!

[RobinPicard]

A test in test_to_regex.py is failing actually. Can you fix it please?

[lollinng]

Fixed in 802114b. The stale assertion was in tests/types/test_to_regex.py::test_to_regex_simple (line 76) — it still expected the old (a){,2} form. Updated it to (a){0,2} to match the new output. (The earlier PR only updated the equivalent assertion in test_dsl.py.)

- assert to_regex(a) == "(a){,2}"
+ assert to_regex(a) == "(a){0,2}"

The other CI failures (ollama multimodal 400s) are unrelated to this change.