INTERNALS · COMPILATION

Query compiler

Analysis candidates

Lexicon lookup combines the lemma, coarse POS, and user entries while preserving every compatible analysis. An explicit POS narrows the detailed tags but does not remove ambiguity inside the category.

An analysis carries stem alternation, particle or ending start state, derivation path, and source-capability requirements.

Program IR

A CandidateProgram contains anchor bytes, core assembly, a suffix automaton, boundary policy, and StructuralConstraint. Provenance is an ordered path of lexical, derivation, and ending rule IDs.

A resource-dependent constraint remains an explicit capability in the compiled plan. Matcher construction fails immediately when that resource is absent.

Query composition

A phrase keeps one plan per atom and joins them under --max-gap in the matcher. A disjunction merges alternative analyses and programs into one logical atom.

Disjunction origin analysis indices are remapped into the merged analysis array. Alternatives that produce the same span emit one result while preserving every origin.

Ordering and deduplication

Programs are sorted by deterministic keys built from anchors, execution conditions, and span projections. Analyses with identical execution conditions merge programs while unioning origins.

Plan limits bound program, anchor, and structural-state width. Exceeding a limit raises a compile error instead of dropping candidates.