Pipeline order

Verified against BurmeseNLP._analyze():

Normalize
    → Syllable tokenize
    → Word tokenize (longest)
    → BMWE (token-sequence trie)
    → Contextual POS (MWE-aware)
    → Gazetteer NER (list match → Document.entities)
    → Phrase chunking (entity spans locked as NP)
    → Sentence segmentation (chunk/POS-aware)
    → Clause parse (→ Document.sentence_trees / clauses)
    → Document

Gazetteer matching runs after POS on post-BMWE words — POS tags are one of the signals GazetteerManager.find_all uses to reject short/ambiguous false-positive matches. Once a span is confirmed, every token in it is locked to PROPN in pos_tags itself (not just inside the chunker), so a name's trailing syllables don't keep whatever tag the context-blind first POS pass gave them (e.g. ထက် as CONJ, ဘို as VERB) once we know they're part of one name. The chunker then locks the same span to one NP so names are not split by POS patterns.

Sentence boundaries intentionally come after POS + chunks, so subject-marker သည် inside an NP is not confused with sentence-final သည် after a VP. Clause parsing runs last over the phrase covering within each sentence window.