Recognise word-anchored insertions; close quotes at the next item
The documented next step was the word-anchored structural insertion
("Vor den Wörtern „Aus dem Bereich Verkehr:“ wird folgender Absatz 5
eingefügt", Berlin Artikel 1 Nr. 2 b) bb)). StrukturEinfuegung now
carries an optional WortAnker, the recogniser has a pattern for the form
— placed before STRUKTUR_EINFUEGUNG, which bails out of recognition
entirely when the StellenParser cannot read "den Wörtern «0»" — and the
applier positions the new unit at the anchor line instead of a structural
boundary.
Two blockers turned up next to it, both measured on the corpus rather
than assumed:
The command was not cleanly recognisable at all. Its closing quotation
mark is missing in the official text, so the quote swallowed items cc)
and c); an anchor alone would have made the command look applicable with
a huge foreign quote as its payload. The docs called a boundary at
enumeration markers impossible, and that holds for the unguarded form: a
quoted amendment provision carries command language itself. Guarded by
four conditions together — the article's quotation marks demonstrably do
not balance, the next line's marker is at the same or a shallower level
than the one the quote opened on, that line carries command language, and
closing here leaves the rest of the article balanced — it fires exactly
twice in the whole sample corpus, both times where the missing mark
belongs: Berlin before cc), and GV. NRW. Artikel 2 before 12. Without the
guards it also fires inside quoted amendment provisions in the BayJG and
GModG documents; the balance gate rules those out.
The GV.-NRW. gazette encodes part of its umlauts decomposed (u + U+0308,
79 places, dozens of them "eingefügt:"/"angefügt:"). Invisible in the
text, but a different word to every command pattern, so those commands
could never match. TextBereiniger now normalises to NFC first — not
NFKC, which would flatten the official sentence numbers ¹²³ and take
SatzTeiler and Superskript their basis. Every other sample document,
every gii-XML stem and every hand-kept plain-text stem is already NFC, so
the pinned figures stay bit-identical.
Berlin Artikel 1 is thereby fully recognised (6 of 6 commands, none
unknown). One pinned statement changed for a stated reason: the NRW
acceptance test asserted the article-heading warning, which is now
preempted by the item boundary inside Artikel 2 — Nr. 12 survives
instead of being swallowed. 241 tests green.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Change-Id: Ic579bec6fecc2495a61b3dcf87c9ad71b42f34ae
13 files changed