fix(citations): honor the errs tolerance cap in CitationMixin fuzzy span matching

[lntutor]

What's broken

CitationMixin._get_span escalates the allowed regex edit distance in a loop whose guard is evaluated before the increment:

errs_ = 0
s = regex.search(f"({minor}){{e<={errs_}}}", major)
while s is None and errs_ <= errs:   # errs defaults to 5
    errs_ += 1
    s = regex.search(f"({minor}){{e<={errs_}}}", major)

When errs_ == errs (5) the body runs once more, bumping errs_ to 6 and searching with {e<=6}. So the effective fuzzy tolerance is errs + 1 = 6 — a hallucinated substring_quote that is 6 edits from any real source span is accepted and silently rewritten to that unrelated context substring, instead of being dropped as "not found":

class Answer(CitationMixin): pass
Answer.model_validate({"substring_quotes": ["0123ZZZZZZ"]}, context={"context": "0123456789"}).substring_quotes
# before: ['0123456789']   (6 edits away -- should be [])

This weakens the citation verification the mixin exists to provide.

Fix

Change the guard to errs_ < errs so the final search runs at exactly {e<=errs}. Quotes within errs edits still match.

Tests

Adds tests/dsl/test_citation.py (within-tolerance matches; beyond-tolerance dropped). The existing citation coverage in tests/coverage/test_dsl_small_coverage.py — which relies on a legitimate 1-edit match — stays green (16). Distinct from #2430 (regex-escape crash), which leaves the loop bound untouched.

[jxnl]

Consolidated and shipped in #2495. Closing this focused patch as superseded; thank you for the contribution.