fix(tokenizers): relax whitespace at full-cite template boundaries

[jaysoma]

Fixes

Fixes: #338

Summary

Glued citations (846F.2d746) currently extract as nothing: the full-cite
template requires exactly one literal space at the volume-reporter and
reporter-page boundaries. This extends #305's whitespace relaxation to those
two boundaries:

• full_cite template becomes r"$volume\s*$reporter,?\s*$page"
• _relax_ws now trims its trailing \s* so the reporter group cannot
  capture boundary whitespace (reporter='U.S. ') once the mandatory
  template space is gone.

Measured on the 7,740 glued-citation opinions from the bulk scan (details in
#338): 7,737 recovered citations across 4,141 documents, no losses of
previously-extracted citations, existing test suite unchanged. New regression
tests cover glued full/short cites, composition with #305, group cleanliness,
and a non-citation negative.

AI Disclosure

• No AI tools were used to create the content of this PR.
• Parts of this PR were created with the help of an AI tool, and I have carefully reviewed all of its content and take full responsibility for it.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[CLAassistant]

All committers have signed the CLA.

[jaysoma]

I ran this patch against real-world data to validate it beyond the unit tests:
a bulk scan of the CourtListener opinion dump (10.8M rows) found 7,740 opinions
containing glued citations, and I replayed both pristine main and this branch
over the stored context windows of every one of them (11,748 unique windows
across 7,489 documents after deduplication).

Recovery: 7,416 windows in 4,141 documents gained citations that main
extracts nothing from — 6,568 FullCaseCitation, 1,166 ShortCaseCitation,
3 SupraCitation. Known specimens (846F.2d746, 443U.S.307,
42U.S.C.§1983) all extract with correct volume/reporter/page groups.

Regressions: the diff initially showed 14 windows where the branch lost a
citation main had found. On inspection:

• 9 are a bare SupraCitation("supra,") replaced by the short citation it
  pointed at (Attaway, supra, 757F.2d at 1241-42 — main only saw the
  supra; the branch recovers 757 F.2d at 1241 and the overlap filter
  absorbs the supra, preserving the referent as antecedent_guess). Net
  information gain.
• 3 are garbage UnknownCitation blobs (e.g.
  544(2007);28U.S.C.§1915.Dismissal…) replaced by real citations.
• 1 was a true regression, now fixed by the second commit: in
  347U.S. 483,349 U.S. 294 the first citation's consuming trailing
  boundary swallowed the comma, starving the second citation's leading
  boundary. Notably this swallow exists on main for unglued text —
  get_citations("347 U.S. 483,349 U.S. 294") returns only the first
  citation — and only the pure-Python tokenizers are affected;
  HyperscanTokenizer reports overlapping matches and never had the blind
  spot. TokenExtractor.get_matches now resumes scanning at the end of the
  token content rather than the end of the boundary-consuming match, which
  aligns the tokenizers and fixes both the glued and unglued cases.

Over-capture, documented honestly: when OCR glues a preceding number
onto the volume, the digits are inherently ambiguous: …U.S. 137, 8 S. Ct. 38… glued to 1378S. Ct. 38 extracts volume 1378. About 285 of the ~7.7k
newly matched instances (≈4%) are in this class (recognizable here by
impossible S. Ct. volumes). Two things worth noting: (1) main already
extracts volume 1378 from the properly-spaced string — eyecite has no
volume-plausibility gating anywhere, so this is pre-existing behavior that
glued input merely produces more of; (2) a fabricated-looking volume still
fails any downstream lookup, whereas main extracts nothing from these
windows at all. A volume-plausibility check against reporters-db ranges
would be a natural follow-up if there's interest.

Also observed: the overlap filter now sees pairings it has no rule for
(Supra followed by adjacent ShortCase, parallel ShortCase pairs) and
logs "Unknown overlap case" warnings; outcomes are reasonable but
order-dependent (the supra is absorbed in one processing order, kept
alongside in the other). Happy to file that separately.

Replay harness and full per-window diffs are available if useful.