Module autocorrect 

Runtime autocorrect primitives for Bangla text produced above Obadh.

This module implements the deterministic runtime shell for autocorrect: candidate generation, Bangla-aware edit scoring, and conservative ranking. Learned model weights can be added behind the same candidate/ranker boundary later without changing the transliteration hot path.

Structs

AutocorrectConfig

AutocorrectDecision

The result of AutocorrectEngine::decide over the in-memory Lexicon: the ranked candidates and a conservative replacement suggestion.

AutocorrectEngine

CandidateFeatures

CorrectionCandidate

CorrectionRequest

EditCost

FstCandidate

FstLexicon

FstLoanwordMatch

FstRepairedBaseline

FstSuggestOptions

FstSuggestResult

Lexicon

LexiconEntry

LexiconStats

LoanwordEntry

LoanwordLexicon

LoanwordMatch

LoanwordSearchOptions

LoanwordSuggestion

RomanRepair

RomanRepairOptions

RomanRepairedOutput

Enums

CorrectionSource

FstCandidateSource

FstSuggestError

LexiconArtifactError

LoanwordArtifactError

LoanwordSuggestionKind

RomanRepairKind

Constants

AUTOCORRECT_FEATURE_DIM

DEFAULT_FST_MAX_DISTANCE

DEFAULT_FST_PREFIX_CANDIDATES

DEFAULT_LOANWORD_FUZZY_CANDIDATES

DEFAULT_ROMAN_REPAIR_BEAM_SIZE

FST_MAX_LEVENSHTEIN_DISTANCE

LOANWORD_FUZZY_MAX_DISTANCE

Functions

build_loanword_bytes

default_loanword_fuzzy_distance

is_loanword_key

key_slip_repaired_outputs

Key-slip repaired baselines: single adjacent-key rewrites of input whose transliteration is a real lexicon word and differs from the untouched baseline. Mirrors super::roman_repaired_outputs — the caller supplies only plumbing closures (transliterate, is_lexicon_word); every numeric parameter is baked into this module.

roman_repair_beam

roman_repaired_outputs

weighted_edit_distance