/**
 * @copyright Sister Software
 * @license AGPL-3.0
 * @author Teffen Ellis, et al.
 *
 *   Punctuation-gap span bridging — the v4.4.0 corrective (and the long-deferred Saint-Albans
 *   span-merge, scoped to where it is provably safe).
 *
 *   The corpus alignment tokenizer drops standalone punctuation (corpus/src/tokenize.ts), so NO
 *   training row can label the periods inside "P.O. Box" — the model learns the tag perfectly
 *   (every letter piece at 0.93+ confidence) but emits it as fragments split at each dot, and span
 *   assembly surfaces only the first fragment ("p"). Measured on the v1.3.0 gate: dotted po_box
 *   leaders failed 98%, ALL truncations, while plain leaders passed — a structural expressivity
 *   limit of the label format, not a learning failure.
 *
 *   The fix is deterministic: AFTER decode, merge adjacent same-label spans whose gap consists only
 *   of punctuation/whitespace, contains at least one non-space character, and is short (≤ 3 chars).
 *   The non-space requirement is essential — space-only gaps ("Saint Paul" as two locality spans)
 *   are NOT bridged, because a space between two same-tag spans is often a real boundary (the
 *   Saint-Albans fragmentation wants this fix too, but it must come with its own evidence; this
 *   pass stays conservative by construction).
 *
 *   Runs beside the postcode/unit repair passes in the classifier, before tree-building.
 */
import type { DecoderToken } from "@mailwoman/core/decoder";
/**
 * Options for {@link bridgePunctuationGaps}.
 */
export interface BridgePunctuationOpts {
    /**
     * Structural spans (from the Stage 2.7 span proposer — ANNOTATION/QUOTED groups, delimiters inclusive) whose
     * boundaries no merge may straddle: M2's crossing constraint, the bridge's mirror image (the bridge merges across
     * WEAK punctuation; this blocks merging across STRUCTURAL punctuation). A merge is blocked when either span boundary
     * falls inside the gap being bridged — e.g. an apostrophe-quoted name whose closing quote sits in an
     * otherwise-bridgeable gap. Boundaries already inside a labeled token are the model's call, not the bridge's; only
     * gaps are policed.
     */
    blockedSpans?: ReadonlyArray<{
        start: number;
        end: number;
    }>;
}
/**
 * Merge same-label fragments separated only by punctuation gaps. Returns a new token array where the first fragment of
 * each bridged group is widened to the group's full char range (so span extraction reads the raw text straight through
 * the punctuation), and later fragments are dropped. Labels, ordering, and all non-bridged tokens are untouched.
 */
export declare function bridgePunctuationGaps(text: string, input: readonly DecoderToken[], opts?: BridgePunctuationOpts): DecoderToken[];
//# sourceMappingURL=span-bridge.d.ts.map