OPENNLP-1850: Unicode normalization engine — CharClass, rungs, Dimension, confusables (1a/5)

[krickert]

Part 1a of the OPENNLP-1850 stack. Splits the former foundation PR (#1103) into a mechanical engine layer (this PR) and the offset/alignment layer (1b), as requested in review.

The dependency-free engine:

• CharClass/CodePointSet over the Unicode White_Space and Dash properties (cursor-based, no regex)
• the per-code-point substitution/collapse/strip rungs, the Dimension ladder, and the non-aligned TextNormalizer builder
• the bundled UTS Remove deprecated IndexHashTable class #39 confusables.txt with its LICENSE/NOTICE/rat-excludes bookkeeping

Confusables now loads lazily and recoverably (double-checked accessor; no classpath resource I/O in a static initializer), so a missing/unreadable resource is a catchable exception at call time rather than a class-poisoning ExceptionInInitializerError.

Merge bottom-up: 1a (this) → 1b alignment → tokenizer → DL → docs.

[krickert]

@rzo1 Thanks — both of your points on the foundation are addressed.

Split into 1a + 1b (done). I split the foundation along the history exactly where you suggested:

• OPENNLP-1850: Unicode normalization engine — CharClass, rungs, Dimension, confusables (1a/5) #1108 — engine (1a): CharClass/CodePointSet/UnicodeWhitespace/UnicodeDash, the per-code-point rungs, Dimension, the non-aligned TextNormalizer, and confusables.txtwith all its LICENSE/NOTICE/rat-excludes bookkeeping. Mostly mechanical substitution,and where the license review belongs.
• OPENNLP-1850: Offset/alignment layer — Alignment, AlignedText, buildAligned (1b/5) #1109 — offset/alignment layer (1b): Alignment, AlignedText, OffsetAwareNormalizer, buildAligned(), the *Aligned CharClass variants, and the dense span-mapping tests(binary-search mapping, expansion/deletion edge cases). The conceptually hard ~800 lines, isolated for a focused read.

#1104 (tokenizer) now bases on #1109. So the stack is now 1a → 1b → tokenizer → DL → docs, each well under your ~1.5k-real-code target, and the 10k-line confusables.txt datafile is contained in 1a. I closed #1103 pointing at the two replacements.

Static-initializer resource loading (done, and generalized). Agreed on the rule. All three bundled-data loaders that did classpath I/O in a static {} block now load lazily on first use through a double-checked accessor, so a resource the loader can't see surfaces as a catchable exception at call time rather than an ExceptionInInitializerError that poisons the class:

• Confusables (1a / OPENNLP-1850: Unicode normalization engine — CharClass, rungs, Dimension, confusables (1a/5) #1108)
• WordBreakProperty and ExtendedPictographic (tokenizer / OPENNLP-1850: UAX #29 word tokenizer and the layered Term model (2/4) #1104) — the latter two wrap their tables in a small immutable holder loaded via the same pattern.

The List.of(...) static blocks in UnicodeWhitespace/UnicodeDash are left as-is (no I/O, no classloader risk), as you noted.

Each layer builds and tests green on its own (mvn -pl … -am verify, plus checkstyle + forbiddenapis across the full reactor).