fix(citations): escape substring quotes before fuzzy span matching

[RohitMudili]

Describe your changes

CitationMixin._get_span interpolated the LLM-generated quote directly into a
regex fuzzy-match pattern:

minor = quote
s = regex.search(f"({minor}){{e<={errs_}}}", major)

Because the quote is never escaped, any regex metacharacter in the model's
substring_quotes — unbalanced parentheses/brackets, *, +, ?, \, etc.,
all common in real prose — is interpreted as part of the pattern. This either
raises regex.error (crashing validation) or silently matches the wrong span,
instead of the documented behavior of finding the span or dropping the quote.

Reproduction:

context = "The margin is 50% (approx) this quarter."
Answer.model_validate(
    {"substring_quotes": ["50% (approx"]},   # unbalanced paren from the LLM
    context={"context": context},
)
# regex.error: missing ) at position 19

Fix

Escape the quote with regex.escape() before building the pattern so it is
matched literally. Fuzzy (edit-distance) matching is unaffected — a quote with a
one-character typo still resolves within the default edit budget.

Changes

• instructor/v2/dsl/citation.py: regex.escape(quote) before the fuzzy search.
• tests/v2/test_citation.py: new tests — metacharacter quotes resolve instead of crashing, non-matching quotes are dropped, fuzzy matching still works, and no-context input is untouched. (These fail on main and pass with the fix.)
• CHANGELOG.md: entry under [Unreleased].

Issue ticket number and link

Closes #2431 — #2431

Checklist before requesting a review

• I have performed a self-review of my code
• If it is a core feature, I have added thorough tests.
• If it is a core feature, I have added documentation.

[RohitMudili]

Hey @jxnl @ivanleomk — this fixes the regex.error crash in CitationMixin reported in #2431. When the model returns a substring_quotes value containing regex metacharacters (unbalanced parentheses/brackets, *, \, etc. — common in ordinary prose), validation blows up with regex.error instead of resolving the span or dropping the quote. Would appreciate a review when you get a chance!

TL;DR: _get_span() interpolated the raw LLM quote straight into a fuzzy pattern f"({minor}){{e<={errs_}}}" with no escaping. Wrapped it in regex.escape() so the quote is matched literally; fuzzy (edit-distance) matching is unaffected. Includes regression tests that fail on main and pass with the fix.

[ErenAta16]

Applied this locally and ran the new test file against current main: all 8 pass, including test_fuzzy_matching_still_works_after_escaping, which was the thing I wanted to check, regex.escape() on the quote before wrapping it in the {e<=N} fuzzy syntax doesn't interfere with the edit-distance matching itself, since escaping only affects how literal characters are interpreted, not the fuzzy-match operator appended outside the escaped substring.

Only blocker is the DIRTY merge state, looks like just a CHANGELOG.md conflict from newer entries landing since this was opened, the actual fix and tests applied cleanly.