fix: reject negative counts in exactly/at_least/at_most/between

[chuenchen309]

What

QuantifyExact, QuantifyMinimum, and QuantifyMaximum (backing exactly(), at_least(), at_most()) accepted negative counts with no validation. QuantifyBetween (backing between()) validated min_count <= max_count but was also missing a non-negative check.

Why this matters

A negative count silently produces a regex quantifier that Python's re engine doesn't recognize as a quantifier at all (e.g. {-1}), so re falls back to treating it as literal text:

from outlines.types.dsl import exactly, to_regex
import re

term = exactly(-1, "a")
pattern = to_regex(term)          # '(a){-1}'
re.compile(pattern).fullmatch("a")        # False
re.compile(pattern).fullmatch("a{-1}")    # True  <- silently changed meaning

Instead of failing at construction time with a clear error, the term silently becomes something that matches a completely different (and almost certainly unintended) string.

Fix

Add a __post_init__ guard to all four Quantify* dataclasses, mirroring the existing min_count <= max_count check already on QuantifyBetween.

Testing

• Added regression tests to tests/types/test_dsl.py::test_dsl_init covering all four classes; confirmed they fail without the fix (DID NOT RAISE ValueError) and pass with it.
• Ran pytest tests/types/ (232 passed) and pre-commit run --files src/outlines/types/dsl.py tests/types/test_dsl.py (mypy + ruff clean).
• Searched the codebase for other direct construction sites of these four classes — none found; all usage goes through the exactly/at_least/at_most/between factory functions and Term instance methods, which pass through unchanged.

Disclosure: I used an AI coding assistant (Claude) to help identify this bug and draft the fix. I reviewed the diff, wrote/ran the regression tests, and ran the full local test suite + pre-commit before opening this PR.

[github-actions]

📚 Documentation preview: https://dottxt-ai.github.io/outlines/pr-preview/pr-1922/

Preview updates automatically with each commit.