Take a printed non-English references heading before weaker list guesses

[thesanogoeffect]

Before: a paper with a German "Literatur" list could lose the whole list. The locator's printed-heading check only knew English headings, so an earlier, weaker step won. On EXCITE dev this happened twice. On one paper the front-role model took a one-row body section ("Auswertung") as the header; on another the layout fallback took the last unclassified section, a two-row talk title. In both cases the segmenters then ran on body sentences.

After: a printed references heading in any language the heading grammar lists wins over those guesses. An English heading still goes first when both are printed. The German "Quellen" compound headings are recognized too.

How:

• The printed-heading override in RefLocator.collect_reference_rows tries English headings first, then any heading that matches the multilingual grammar as a whole.
• A front-role header whose section has no dated row is skipped.
• The header-text fallback runs before the "last unclassified section" layout fallback.
• ref_patterns adds "Zitierte und verwendete Quellen", "Verwendete Literatur", "Literatur und Quellen" and "Quellen und Literatur". Bare "Quellen" is left out because it also opens German table notes.

Two front-role tests used "Список литературы", which the printed-heading step now catches first. They now use a heading outside the grammar, so they still test the model path.

Evidence: I ran EXCITE dev (174 papers) on declan with the first commit (cb22d3d) against the 685765d baseline. The setup was identical and the scorer was held constant.

	Baseline	cb22d3d
ref_matching_f1	0.773	0.784
ref_matching_f1, end-list papers	0.816	0.828
ref_matching_f1, German papers	0.729	0.744
Gold references cut out cleanly	3,782	3,834

Footnote, mixed and English papers were unchanged. Only five papers changed which section they took as the list. 32707 regressed (21 → 9 clean) because its list sits under both "References" and "13 Literatur" and the German half was picked. The second commit (040a5de) restores English-first order for that case. I have not re-run dev for it: the order only differs where an English heading exists, so I expect 32707 back at the baseline and the other four unchanged.

Known leftover: on 40753 the "Literatur" section also takes in a publisher's series list in the back matter, so 15 of its 16 references now come out clean but about 34 false entries come along with them. Splitting a section at large page gaps would need care for books with one list per chapter, so it's not in this PR.

Tests: the new tests/extract/test_ref_locator_printed_heading.py fails on main and passes here. The extract, ocr and structure suites pass.

[codecov]

Codecov Report

❌ Patch coverage is 83.33333% with 3 lines in your changes missing coverage. Please review.
✅ Project coverage is 90.56%. Comparing base (eeaf588) to head (040a5de).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #213      +/-   ##
==========================================
- Coverage   90.56%   90.56%   -0.01%     
==========================================
  Files         309      309              
  Lines       57337    57343       +6     
  Branches    10383    10383              
==========================================
+ Hits        51929    51933       +4     
  Misses       3449     3449              
- Partials     1959     1961       +2     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.