@mailwoman/neural 4.16.2 → 5.1.0

package/out/classifier.d.ts

@@ -7,13 +7,13 @@
  *   `@mailwoman/core` decoder. Single user-facing entrypoint: `parse(text)` returns an
  *   `AddressTree` ready for projection into JSON / tuple / XML.
  *
- *   Convenience wrappers `parseJson` / `parseTuples` / `parseXml` project the tree on the way out.
+ *   Convenience wrappers `parseJSON` / `parseTuples` / `parseXML` project the tree on the way out.
  */
 import { type SystemCode } from "@mailwoman/codex";
-import { decodeAsXml, type AddressTree, type Calibrator, type ComponentTag } from "@mailwoman/core/decoder";
+import { decodeAsXML, type AddressTree, type Calibrator, type ComponentTag, type SerializeJSONOpts, type SerializeTuplesOpts, type UnknownSpan } from "@mailwoman/core/decoder";
 import { type SpanProposerLexicon } from "@mailwoman/core/pipeline";
 import type { AnchorLookup } from "./anchor-inference.js";
-import { type FstMatcherLike } from "./fst-prior.js";
+import { type FSTMatcherLike } from "./fst-prior.js";
 import type { GazetteerLexicon } from "./gazetteer-inference.js";
 import type { InferResult } from "./onnx-runner.js";
 import { type QueryShapeLike } from "./query-shape-prior.js";
@@ -22,9 +22,9 @@ import { type StreetMorphologyPriorOpts } from "./street-morphology-prior.js";
 import { MailwomanTokenizer } from "./tokenizer.js";
 import type { ResolveWeightsOpts } from "./weights.js";
 /**
- * Structural type the classifier needs from a runner. Lets callers swap the Node-side `OnnxRunner`
- * for a browser-side runner (e.g. `@mailwoman/neural-web`'s `WebOnnxRunner`) without inheritance —
- * the classifier only ever calls `infer(ids)`.
+ * Structural type the classifier needs from a runner. Lets callers swap the Node-side `ONNXRunner` for a browser-side
+ * runner (e.g. `@mailwoman/neural-web`'s `WebONNXRunner`) without inheritance — the classifier only ever calls
+ * `infer(ids)`.
  */
 export interface NeuralRunner {
     infer(tokenIds: number[], anchor?: {
@@ -39,24 +39,24 @@ export interface NeuralAddressClassifierConfig {
     tokenizer: MailwomanTokenizer;
     runner: NeuralRunner;
     /**
-     * Label vocabulary in the order the model emits them. Defaults to Stage 2 (v0.3.0). Stage 2
-     * strictly extends Stage 1 at the same indices, so a v0.2.0 Stage 1 model loaded with this
-     * default still decodes correctly — its emissions only span the first 15 entries.
+     * Label vocabulary in the order the model emits them. Defaults to Stage 2 (v0.3.0). Stage 2 strictly extends Stage 1
+     * at the same indices, so a v0.2.0 Stage 1 model loaded with this default still decodes correctly — its emissions
+     * only span the first 15 entries.
      */
     labels?: readonly string[];
     /**
      * Decoding strategy:
      *
-     * - `"viterbi"` (default) — linear-chain CRF Viterbi with the BIO structural mask. Prevents
-     *   orphan-`I-*` sequences. If `transitions` is provided, uses learned scores on top.
-     * - `"argmax"` — per-token argmax. Faster but produces structurally invalid sequences. Use only for
-     *   debugging / comparison.
+     * - `"viterbi"` (default) — linear-chain CRF Viterbi with the BIO structural mask. Prevents orphan-`I-*` sequences. If
+     *   `transitions` is provided, uses learned scores on top.
+     * - `"argmax"` — per-token argmax. Faster but produces structurally invalid sequences. Use only for debugging /
+     *   comparison.
      */
     decode?: "viterbi" | "argmax";
     /**
-     * Optional learned CRF transition scores. Square matrix of size `labels.length × labels.length`.
-     * Added on top of the structural BIO mask. Future weights releases ship this; today's v3.0.0
-     * weights don't, so the structural mask alone is used.
+     * Optional learned CRF transition scores. Square matrix of size `labels.length × labels.length`. Added on top of the
+     * structural BIO mask. Future weights releases ship this; today's v3.0.0 weights don't, so the structural mask alone
+     * is used.
      */
     transitions?: number[][];
     /** Optional learned start-of-sequence transition scores per label. */
@@ -64,72 +64,66 @@ export interface NeuralAddressClassifierConfig {
     /** Optional learned end-of-sequence transition scores per label. */
     endTransitions?: number[];
     /**
-     * Optional postcode-anchor lookup (#239/#240). When set, `parse` builds per-piece anchor features
-     * from the text + this lookup and feeds them to the runner — for models trained with the anchor
-     * channel (exported with the `anchor_features`/`anchor_confidence` ONNX inputs). Omit for plain
-     * models. Load via `loadAnchorLookup` from `./anchor-inference.js`.
+     * Optional postcode-anchor lookup (#239/#240). When set, `parse` builds per-piece anchor features from the text +
+     * this lookup and feeds them to the runner — for models trained with the anchor channel (exported with the
+     * `anchor_features`/`anchor_confidence` ONNX inputs). Omit for plain models. Load via `loadAnchorLookup` from
+     * `./anchor-inference.js`.
      */
     postcodeAnchorLookup?: AnchorLookup;
     /**
-     * Optional gazetteer-anchor lexicon (#464, knowledge-ladder rung 3.2). When set, `parse` builds
-     * per-token candidate-tag-set clues (country/region/po_box/cedex/homograph) from the text + this
-     * lexicon and feeds them to the runner — for models trained with the gazetteer-anchor channel
-     * (exported with the `gazetteer_features`/`gazetteer_confidence` ONNX inputs). Omit for plain
-     * models. Load via `parseGazetteerLexicon` from `./gazetteer-inference.js`.
+     * Optional gazetteer-anchor lexicon (#464, knowledge-ladder rung 3.2). When set, `parse` builds per-token
+     * candidate-tag-set clues (country/region/po_box/cedex/homograph) from the text + this lexicon and feeds them to the
+     * runner — for models trained with the gazetteer-anchor channel (exported with the
+     * `gazetteer_features`/`gazetteer_confidence` ONNX inputs). Omit for plain models. Load via `parseGazetteerLexicon`
+     * from `./gazetteer-inference.js`.
      */
     gazetteerLexicon?: GazetteerLexicon;
     /**
-     * Channel choreography (#464, v0.9.13 postcode fix): when true, zero the gazetteer clue on pieces
-     * adjacent to a postcode-anchor hit (needs both `gazetteerLexicon` and `postcodeAnchorLookup`).
-     * Targets the region-clue→postcode CRF interference (~3pp US postcode).
+     * Channel choreography (#464, v0.9.13 postcode fix): when true, zero the gazetteer clue on pieces adjacent to a
+     * postcode-anchor hit (needs both `gazetteerLexicon` and `postcodeAnchorLookup`). Targets the region-clue→postcode
+     * CRF interference (~3pp US postcode).
      *
-     * PAIRING IS ESSENTIAL: set this IFF the model was TRAINED with the matching train-time
-     * choreography (`data.gazetteer_choreography`). The 2026-06-10 diagnostic showed the harm is
-     * WEIGHT-BAKED — applying this at inference on a model trained _without_ train-choreography does
-     * NOT recover postcode and adds train/inference skew. Only enable for a consolidation-era model
-     * trained with the train-time half.
+     * PAIRING IS ESSENTIAL: set this IFF the model was TRAINED with the matching train-time choreography
+     * (`data.gazetteer_choreography`). The 2026-06-10 diagnostic showed the harm is WEIGHT-BAKED — applying this at
+     * inference on a model trained _without_ train-choreography does NOT recover postcode and adds train/inference skew.
+     * Only enable for a consolidation-era model trained with the train-time half.
      */
     suppressGazetteerNearPostcode?: boolean;
     /**
-     * Default address-system conventions mode for every parse (see
-     * `ParseOpts.addressSystemConventions` for semantics — `"auto"` reads the model's locale head; a
-     * `SystemCode` pins it). Per-parse opts override this. Omit for the byte-stable pre-#511 default
-     * (no detection, no mask).
+     * Default address-system conventions mode for every parse (see `ParseOpts.addressSystemConventions` for semantics —
+     * `"auto"` reads the model's locale head; a `SystemCode` pins it). Per-parse opts override this. Omit for the
+     * byte-stable pre-#511 default (no detection, no mask).
      */
     addressSystemConventions?: "auto" | SystemCode;
     /**
-     * Punctuation-gap span bridging (the v4.4.0 corrective; see `span-bridge.ts`). The corpus label
-     * format cannot express punctuation inside a span, so dotted surfaces ("P.O. Box", "C.P.") decode
-     * as fragments. When true, adjacent same-tag spans separated only by short punctuation gaps are
-     * merged after decode. Per-parse opts override. Omit for the byte-stable pre-v4.4.0 behavior.
+     * Punctuation-gap span bridging (the v4.4.0 corrective; see `span-bridge.ts`). The corpus label format cannot express
+     * punctuation inside a span, so dotted surfaces ("P.O. Box", "C.P.") decode as fragments. When true, adjacent
+     * same-tag spans separated only by short punctuation gaps are merged after decode. Per-parse opts override. Omit for
+     * the byte-stable pre-v4.4.0 behavior.
      */
     bridgePunctuationGaps?: boolean;
     /**
-     * Stage 2.7 span proposer (M2+M3 from the punctuation survey, #518). When set, every parse runs
-     * `proposeSpans` (`@mailwoman/core/pipeline`) over the raw text and consumes the typed proposals
-     * two ways: (a) as additive emission priors — the phrase-prior path; the classifier conditions on
-     * the boundary hypotheses and can still disagree — and (b) ANNOTATION/QUOTED span boundaries feed
-     * the span bridge as merge-crossing constraints (no same-tag merge may straddle a structural
-     * delimiter). Build the lexicon with `buildCodexSpanLexicon` (`./span-proposer-lexicon.js`).
+     * Stage 2.7 span proposer (M2+M3 from the punctuation survey, #518). When set, every parse runs `proposeSpans`
+     * (`@mailwoman/core/pipeline`) over the raw text and consumes the typed proposals two ways: (a) as additive emission
+     * priors — the phrase-prior path; the classifier conditions on the boundary hypotheses and can still disagree — and
+     * (b) ANNOTATION/QUOTED span boundaries feed the span bridge as merge-crossing constraints (no same-tag merge may
+     * straddle a structural delimiter). Build the lexicon with `buildCodexSpanLexicon` (`./span-proposer-lexicon.js`).
      * Per-parse opts override.
      *
-     * DEFAULT ON (operator ruling 2026-06-12, after the #518 measurement closed both v0-win quadrants
-     * with no class down): omitting this builds the codex lexicon lazily with the frozen measured
-     * scales (biasScale 5.0 / annotationBiasScale 12.0). Pass `false` for the proposer-free baseline
-     * (the pre-2026-06-12 byte-stable default).
+     * DEFAULT ON (operator ruling 2026-06-12, after the #518 measurement closed both v0-win quadrants with no class
+     * down): omitting this builds the codex lexicon lazily with the frozen measured scales (biasScale 5.0 /
+     * annotationBiasScale 12.0). Pass `false` for the proposer-free baseline (the pre-2026-06-12 byte-stable default).
      */
     spanProposer?: SpanProposerConfig | false;
     /**
-     * Per-word BIO consistency repair (#727 + the admin-token fragmentation class). Default off →
-     * byte-identical. When true, every `▁`-delimited word's pieces are forced to ONE tag by a
-     * confidence-weighted vote over the post-prior emissions (see word-consistency.ts). Per-parse
-     * `ParseOptions.enforceWordConsistency` overrides this default.
+     * Per-word BIO consistency repair (#727 + the admin-token fragmentation class). Default off → byte-identical. When
+     * true, every `▁`-delimited word's pieces are forced to ONE tag by a confidence-weighted vote over the post-prior
+     * emissions (see word-consistency.ts). Per-parse `ParseOptions.enforceWordConsistency` overrides this default.
      */
     enforceWordConsistency?: boolean;
 }
 /**
- * Config for the Stage 2.7 span-proposer integration (see
- * `NeuralAddressClassifierConfig.spanProposer`).
+ * Config for the Stage 2.7 span-proposer integration (see `NeuralAddressClassifierConfig.spanProposer`).
  */
 export interface SpanProposerConfig extends SpanProposalPriorOpts {
     /** Codex-backed designator vocabulary (`buildCodexSpanLexicon`). */
@@ -145,51 +139,54 @@ export declare class NeuralAddressClassifier {
     private readonly endTransitions;
     constructor(cfg: NeuralAddressClassifierConfig);
     /**
-     * The default-ON Stage 2.7 config: codex lexicon (us/au/nz), frozen measured scales (the prior
-     * builder's own defaults). Built once per instance, only when a parse actually needs it.
+     * The default-ON Stage 2.7 config: codex lexicon (us/au/nz), frozen measured scales (the prior builder's own
+     * defaults). Built once per instance, only when a parse actually needs it.
      */
     private defaultProposer;
     /**
-     * One-call factory that resolves the weights package (or explicit paths), loads the tokenizer and
-     * ONNX runner, and returns a ready-to-use classifier.
+     * One-call factory that resolves the weights package (or explicit paths), loads the tokenizer and ONNX runner, and
+     * returns a ready-to-use classifier.
      *
-     * Resolution order: explicit paths in `opts` → `@mailwoman/neural-weights-<locale>` package →
-     * throws a single actionable error.
+     * Resolution order: explicit paths in `opts` → `@mailwoman/neural-weights-<locale>` package → throws a single
+     * actionable error.
      *
-     * **Node-only.** The dynamic imports keep `OnnxRunner` (onnxruntime-node) + `resolveWeights`
-     * (uses Node fs) out of the static dependency graph, so this file can be bundled for the browser
-     * by `@mailwoman/neural-web`. Calling this method in a browser will throw at runtime — use
-     * `loadNeuralClassifierFromUrls` from `@mailwoman/neural-web` instead.
+     * **Node-only.** The dynamic imports keep `ONNXRunner` (onnxruntime-node) + `resolveWeights` (uses Node fs) out of
+     * the static dependency graph, so this file can be bundled for the browser by `@mailwoman/neural-web`. Calling this
+     * method in a browser will throw at runtime — use `loadNeuralClassifierFromUrls` from `@mailwoman/neural-web`
+     * instead.
      */
     static loadFromWeights(opts?: ResolveWeightsOpts & {
         postcodeAnchorLookup?: AnchorLookup;
+        executionProviders?: string[];
     }): Promise<NeuralAddressClassifier>;
     /** Tokenize → infer → Viterbi (or argmax) → decoder tree. */
     parse(text: string, opts?: ParseOpts): Promise<AddressTree>;
     /**
-     * Like `parse`, but also returns the raw per-token logits and piece offsets needed for per-span
-     * logit aggregation (Option C joint-reconcile integration). Shares the ENTIRE decode path with
-     * `parse` (one `#decode`, #481) — including the repair passes, which previously ran only in
-     * `parse`: reconcile must consume the same tokens the argmax path serves users, and the repair
-     * opts were silently ignored here before. `logits` stay RAW (pre-prior, pre-repair) — they are
-     * the model's emissions, not the decode's opinions.
+     * Like `parse`, but also returns the raw per-token logits and piece offsets needed for per-span logit aggregation
+     * (Option C joint-reconcile integration). Shares the ENTIRE decode path with `parse` (one `#decode`, #481) —
+     * including the repair passes, which previously ran only in `parse`: reconcile must consume the same tokens the
+     * argmax path serves users, and the repair opts were silently ignored here before. `logits` stay RAW (pre-prior,
+     * pre-repair) — they are the model's emissions, not the decode's opinions.
      */
     parseWithLogits(text: string, opts?: ParseOpts): Promise<ParseWithLogitsResult>;
-    parseJson(text: string, opts?: ParseOpts): Promise<Partial<Record<ComponentTag, string>>>;
+    parseJSON(text: string, opts?: ParseOpts): Promise<Partial<Record<ComponentTag, string>>>;
+    parseJSON(text: string, opts: ParseOpts & SerializeJSONOpts): Promise<Partial<Record<ComponentTag, string>> & {
+        unknown?: UnknownSpan[];
+    }>;
     parseTuples(text: string, opts?: ParseOpts): Promise<Array<[ComponentTag, string]>>;
-    parseXml(text: string, opts?: ParseOpts & {
-        xml?: Parameters<typeof decodeAsXml>[1];
+    parseTuples(text: string, opts: ParseOpts & SerializeTuplesOpts): Promise<Array<[ComponentTag | "unknown", string]>>;
+    parseXML(text: string, opts?: ParseOpts & {
+        xml?: Parameters<typeof decodeAsXML>[1];
     }): Promise<string>;
     /**
-     * Guard against a silent label/emission shape overrun. When the model emits MORE logits per token
-     * than the configured label vocabulary (e.g. a Stage 3 bundle loaded with the default Stage 2
-     * labels), viterbi indexes past the transition matrix and dies with an opaque `Cannot read
-     * properties of undefined (reading '0')`. Fail fast here with a message that names the contract
-     * the caller violated.
+     * Guard against a silent label/emission shape overrun. When the model emits MORE logits per token than the configured
+     * label vocabulary (e.g. a Stage 3 bundle loaded with the default Stage 2 labels), viterbi indexes past the
+     * transition matrix and dies with an opaque `Cannot read properties of undefined (reading '0')`. Fail fast here with
+     * a message that names the contract the caller violated.
      *
-     * The opposite shape (model narrower than labels) is intentionally permitted — STAGE2_BIO_LABELS
-     * prefix-extends STAGE1_BIO_LABELS so a Stage 1 model loaded with Stage 2 labels decodes
-     * correctly via the first 15 logits. See labels.ts for the contract.
+     * The opposite shape (model narrower than labels) is intentionally permitted — STAGE2_BIO_LABELS prefix-extends
+     * STAGE1_BIO_LABELS so a Stage 1 model loaded with Stage 2 labels decodes correctly via the first 15 logits. See
+     * labels.ts for the contract.
      */
     private assertEmissionWidth;
 }
@@ -203,98 +200,92 @@ export interface ParseWithLogitsResult {
     }>;
 }
 /**
- * Per-call opts for `parse()`. Threading a precomputed `QueryShape` here turns on the soft-prior
- * bias path in the Viterbi decoder (Stage 2.4 boundary → Stage 3 encoder integration).
+ * Per-call opts for `parse()`. Threading a precomputed `QueryShape` here turns on the soft-prior bias path in the
+ * Viterbi decoder (Stage 2.4 boundary → Stage 3 encoder integration).
  */
 export interface ParseOpts {
     /**
-     * Precomputed `QueryShape` for this input (from `@mailwoman/query-shape`'s `computeQueryShape`).
-     * Known-format hits in the shape produce additive emission biases toward the matching BIO label.
-     * Typed structurally — no runtime dependency on `@mailwoman/query-shape`.
+     * Precomputed `QueryShape` for this input (from `@mailwoman/query-shape`'s `computeQueryShape`). Known-format hits in
+     * the shape produce additive emission biases toward the matching BIO label. Typed structurally — no runtime
+     * dependency on `@mailwoman/query-shape`.
      */
     queryShape?: QueryShapeLike;
     /**
-     * Maximum bias magnitude in log-odds units. Default 1.0 — adds up to ~e^1 ≈ 2.7× odds to the
-     * favored label. Confidence-scaled, so a 0.6-confidence format hit gets +0.6 max bias.
+     * Maximum bias magnitude in log-odds units. Default 1.0 — adds up to ~e^1 ≈ 2.7× odds to the favored label.
+     * Confidence-scaled, so a 0.6-confidence format hit gets +0.6 max bias.
      */
     queryShapeBiasScale?: number;
     /**
-     * Pre-built FST gazetteer matcher. When provided, gazetteer matches produce additive emission
-     * biases.
+     * Pre-built FST gazetteer matcher. When provided, gazetteer matches produce additive emission biases.
      */
-    fst?: FstMatcherLike;
+    fst?: FSTMatcherLike;
     /** Bias magnitude for FST gazetteer matches. Default 1.0. */
     fstBiasScale?: number;
     /**
-     * Pre-built street-morphology FST matcher. When provided, street-type affixes (Avenue, rue,
-     * Calle, Straße, …) produce additive emission biases toward `street_prefix`/`street_suffix` on
-     * the matched tokens AND toward `street` / away from `dependent_locality` on the adjacent name
-     * tokens. Closes the v0.6.1 dependent_locality vacuum; see
+     * Pre-built street-morphology FST matcher. When provided, street-type affixes (Avenue, rue, Calle, Straße, …) produce
+     * additive emission biases toward `street_prefix`/`street_suffix` on the matched tokens AND toward `street` / away
+     * from `dependent_locality` on the adjacent name tokens. Closes the v0.6.1 dependent_locality vacuum; see
      * `docs/articles/concepts/street-supplement-architecture.md` for the layered design.
      */
-    fstStreetMorphology?: FstMatcherLike;
+    fstStreetMorphology?: FSTMatcherLike;
     /** Override bias magnitudes for the morphology prior. */
     fstStreetMorphologyOpts?: StreetMorphologyPriorOpts;
     /**
-     * When true, run the deterministic postcode regex repair pass (v0.7 #35) on the decoded label
-     * sequence before tree-building. Detects postcode-shaped substrings (GB/CA/NL/US/FR/… patterns)
-     * and snaps/adds the postcode span to the matched shape, fixing the SentencePiece-fragmentation
-     * failures catalogued in the 2026-05-29 postcode diagnostic. Off by default — opt-in until the
-     * v0.7 gate confirms it. See `./postcode-repair.ts`.
+     * When true, run the deterministic postcode regex repair pass (v0.7 #35) on the decoded label sequence before
+     * tree-building. Detects postcode-shaped substrings (GB/CA/NL/US/FR/… patterns) and snaps/adds the postcode span to
+     * the matched shape, fixing the SentencePiece-fragmentation failures catalogued in the 2026-05-29 postcode
+     * diagnostic. Off by default — opt-in until the v0.7 gate confirms it. See `./postcode-repair.ts`.
      */
     postcodeRepair?: boolean;
     /**
-     * Per-word BIO consistency repair (#727 + the admin-token fragmentation class). Overrides the
-     * classifier's `enforceWordConsistency` config default for this parse. See word-consistency.ts.
+     * Per-word BIO consistency repair (#727 + the admin-token fragmentation class). Overrides the classifier's
+     * `enforceWordConsistency` config default for this parse. See word-consistency.ts.
      */
     enforceWordConsistency?: boolean;
     /**
-     * When true, run the deterministic secondary-unit regex repair pass on the decoded label sequence
-     * before tree-building. Detects designator-shaped substrings ("Apt 4B", "Ste 12", "Unit 9400",
-     * bare "#104", …) and snaps/adds the unit span, fixing the unit-drop weakness the three-arena
-     * capability eval surfaced (postal secondary-unit 0% neural). Off by default — opt-in until the
-     * v0.7.2 arena re-run quantifies its delta. See `./unit-repair.ts`.
+     * When true, run the deterministic secondary-unit regex repair pass on the decoded label sequence before
+     * tree-building. Detects designator-shaped substrings ("Apt 4B", "Ste 12", "Unit 9400", bare "#104", …) and
+     * snaps/adds the unit span, fixing the unit-drop weakness the three-arena capability eval surfaced (postal
+     * secondary-unit 0% neural). Off by default — opt-in until the v0.7.2 arena re-run quantifies its delta. See
+     * `./unit-repair.ts`.
      */
     unitRepair?: boolean;
     /**
-     * When true AND the input is detected ALL-CAPS (registry/compliance data like `214 JONES RD,
-     * ELKHART, TX 75839`), title-case the input before the model sees it. The model trains on
-     * mixed-case text, so all-caps is partly OOD — it drops/mis-bounds tokens (#690: `PALESTINE` →
-     * locality `ALESTINE`; all-caps locality 3/5 vs title-case 5/5). Detection-gated, so MIXED-case
-     * input is untouched (byte-stable). Off by default. On all-caps input the output values are
-     * title-cased (the SHOUTING is normalized away — better, and the resolver name-matches
-     * case-insensitively regardless).
+     * When true AND the input is detected ALL-CAPS (registry/compliance data like `214 JONES RD, ELKHART, TX 75839`),
+     * title-case the input before the model sees it. The model trains on mixed-case text, so all-caps is partly OOD — it
+     * drops/mis-bounds tokens (#690: `PALESTINE` → locality `ALESTINE`; all-caps locality 3/5 vs title-case 5/5).
+     * Detection-gated, so MIXED-case input is untouched (byte-stable). Off by default. On all-caps input the output
+     * values are title-cased (the SHOUTING is normalized away — better, and the resolver name-matches case-insensitively
+     * regardless).
      */
     normalizeCase?: boolean;
     /**
-     * Optional span-confidence calibrator (task #59). When provided, each decoded span's `conf=` is
-     * mapped through it (isotonic lookup table → calibrated probability of correctness). OPT-IN —
-     * omit for the byte-stable default softmax confidence. Build one via `createCalibrator`
-     * (`@mailwoman/core/decoder`) from `data/eval/calibration/isotonic-<locale>-<version>.json`.
+     * Optional span-confidence calibrator (task #59). When provided, each decoded span's `conf=` is mapped through it
+     * (isotonic lookup table → calibrated probability of correctness). OPT-IN — omit for the byte-stable default softmax
+     * confidence. Build one via `createCalibrator` (`@mailwoman/core/decoder`) from
+     * `data/eval/calibration/isotonic-<locale>-<version>.json`.
      */
     calibrate?: Calibrator;
     /** Per-parse override of the config-level `bridgePunctuationGaps` (see that doc). */
     bridgePunctuationGaps?: boolean;
     /**
-     * Per-parse switch for the config-level `spanProposer` (see that doc). `false` disables the
-     * configured proposer for this parse; `true`/omitted runs it when configured. Cannot enable the
-     * stage without a configured lexicon.
+     * Per-parse switch for the config-level `spanProposer` (see that doc). `false` disables the configured proposer for
+     * this parse; `true`/omitted runs it when configured. Cannot enable the stage without a configured lexicon.
      */
     spanProposer?: boolean;
     /**
      * Address-system conventions enforcement (#511 Tier A / #478's rules-as-constraints slice).
      *
-     * - `"auto"` — detect the system from the model's locale head (`locale_logits` output, v1.1.0+
-     *   exports; silently no-ops on models without it) and apply that system's codex conventions:
-     *   forbidden tags become a hard emission mask before Viterbi, and a conventions postcode shape
-     *   enables the snap-only postcode repair pass.
-     * - A `SystemCode` (`"fr"`, `"us"`, …) — apply that system's conventions unconditionally (callers
-     *   that already know the locale, e.g. the pipeline's BCP-47 region).
+     * - `"auto"` — detect the system from the model's locale head (`locale_logits` output, v1.1.0+ exports; silently no-ops
+     *   on models without it) and apply that system's codex conventions: forbidden tags become a hard emission mask
+     *   before Viterbi, and a conventions postcode shape enables the snap-only postcode repair pass.
+     * - A `SystemCode` (`"fr"`, `"us"`, …) — apply that system's conventions unconditionally (callers that already know the
+     *   locale, e.g. the pipeline's BCP-47 region).
      * - Omit — byte-stable default: no detection, no mask (pre-#511 behavior).
      *
-     * The detection threshold is deliberately high (0.8): the mask must never fire on a guess.
-     * Measured motivation: the 2026-06-10 v1.1.0 gate, where US suffix logic fired inside French
-     * parses (`street_suffix: "Rue"`) and digit-splits corrupted leading FR postcodes.
+     * The detection threshold is deliberately high (0.8): the mask must never fire on a guess. Measured motivation: the
+     * 2026-06-10 v1.1.0 gate, where US suffix logic fired inside French parses (`street_suffix: "Rue"`) and digit-splits
+     * corrupted leading FR postcodes.
      */
     addressSystemConventions?: "auto" | SystemCode;
 }

package/out/classifier.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"classifier.d.ts","sourceRoot":"","sources":["../classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAwB,KAAK,UAAU,EAAE,MAAM,kBAAkB,CAAA;AACxE,OAAO,EAIN,WAAW,EACX,KAAK,WAAW,EAChB,KAAK,UAAU,EACf,KAAK,YAAY,EAEjB,MAAM,yBAAyB,CAAA;AAChC,OAAO,EAAmC,KAAK,mBAAmB,EAAE,MAAM,0BAA0B,CAAA;AAGpG,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAA;AAEzD,OAAO,EAA0B,KAAK,cAAc,EAAE,MAAM,gBAAgB,CAAA;AAC5E,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,0BAA0B,CAAA;AAEhE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAA;AAEnD,OAAO,EAA0C,KAAK,cAAc,EAAE,MAAM,wBAAwB,CAAA;AAGpG,OAAO,EAA2B,KAAK,qBAAqB,EAAE,MAAM,0BAA0B,CAAA;AAE9F,OAAO,EAAuC,KAAK,yBAAyB,EAAE,MAAM,8BAA8B,CAAA;AAClH,OAAO,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAA;AAGnD,OAAO,KAAK,EAAE,kBAAkB,EAAmB,MAAM,cAAc,CAAA;AAGvE;;;;GAIG;AACH,MAAM,WAAW,YAAY;IAC5B,KAAK,CACJ,QAAQ,EAAE,MAAM,EAAE,EAClB,MAAM,CAAC,EAAE;QAAE,QAAQ,EAAE,aAAa,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC;QAAC,UAAU,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;KAAE,EAC9F,SAAS,CAAC,EAAE;QAAE,QAAQ,EAAE,aAAa,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC;QAAC,UAAU,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;KAAE,GAC/F,OAAO,CAAC,WAAW,CAAC,CAAA;CACvB;AAED,MAAM,WAAW,6BAA6B;IAC7C,SAAS,EAAE,kBAAkB,CAAA;IAC7B,MAAM,EAAE,YAAY,CAAA;IACpB;;;;OAIG;IACH,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;IAC1B;;;;;;;OAOG;IACH,MAAM,CAAC,EAAE,SAAS,GAAG,QAAQ,CAAA;IAC7B;;;;OAIG;IACH,WAAW,CAAC,EAAE,MAAM,EAAE,EAAE,CAAA;IACxB,sEAAsE;IACtE,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAA;IAC3B,oEAAoE;IACpE,cAAc,CAAC,EAAE,MAAM,EAAE,CAAA;IACzB;;;;;OAKG;IACH,oBAAoB,CAAC,EAAE,YAAY,CAAA;IACnC;;;;;;OAMG;IACH,gBAAgB,CAAC,EAAE,gBAAgB,CAAA;IACnC;;;;;;;;;;OAUG;IACH,6BAA6B,CAAC,EAAE,OAAO,CAAA;IACvC;;;;;OAKG;IACH,wBAAwB,CAAC,EAAE,MAAM,GAAG,UAAU,CAAA;IAC9C;;;;;OAKG;IACH,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B;;;;;;;;;;;;;OAaG;IACH,YAAY,CAAC,EAAE,kBAAkB,GAAG,KAAK,CAAA;IAEzC;;;;;OAKG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAA;CAChC;AAED;;;GAGG;AACH,MAAM,WAAW,kBAAmB,SAAQ,qBAAqB;IAChE,oEAAoE;IACpE,OAAO,EAAE,mBAAmB,CAAA;CAC5B;AAED,qBAAa,uBAAuB;;IASvB,OAAO,CAAC,QAAQ,CAAC,GAAG;IARhC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmB;IAC1C,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAsB;IACjD,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAY;IAGxC,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAU;IAC3C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAU;gBAEZ,GAAG,EAAE,6BAA6B;IAa/D;;;OAGG;IACH,OAAO,CAAC,eAAe;IAKvB;;;;;;;;;;;OAWG;WACU,eAAe,CAC3B,IAAI,GAAE,kBAAkB,GAAG;QAAE,oBAAoB,CAAC,EAAE,YAAY,CAAA;KAAO,GACrE,OAAO,CAAC,uBAAuB,CAAC;IAkGnC,6DAA6D;IACvD,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,WAAW,CAAC;IAWjE;;;;;;;OAOG;IACG,eAAe,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,qBAAqB,CAAC;IA2K/E,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAC;IAIzF,WAAW,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAC;IAInF,QAAQ,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG;QAAE,GAAG,CAAC,EAAE,UAAU,CAAC,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,CAAA;KAAE,GAAG,OAAO,CAAC,MAAM,CAAC;IAI7G;;;;;;;;;;OAUG;IACH,OAAO,CAAC,mBAAmB;CAW3B;AAED,wFAAwF;AACxF,MAAM,WAAW,qBAAqB;IACrC,IAAI,EAAE,WAAW,CAAA;IACjB,MAAM,EAAE,MAAM,EAAE,EAAE,CAAA;IAClB,MAAM,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CAC7C;AAED;;;GAGG;AACH,MAAM,WAAW,SAAS;IACzB;;;;OAIG;IACH,UAAU,CAAC,EAAE,cAAc,CAAA;IAC3B;;;OAGG;IACH,mBAAmB,CAAC,EAAE,MAAM,CAAA;IAC5B;;;OAGG;IACH,GAAG,CAAC,EAAE,cAAc,CAAA;IACpB,6DAA6D;IAC7D,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;;;;;;OAMG;IACH,mBAAmB,CAAC,EAAE,cAAc,CAAA;IACpC,yDAAyD;IACzD,uBAAuB,CAAC,EAAE,yBAAyB,CAAA;IACnD;;;;;;OAMG;IACH,cAAc,CAAC,EAAE,OAAO,CAAA;IAExB;;;OAGG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAA;IAEhC;;;;;;OAMG;IACH,UAAU,CAAC,EAAE,OAAO,CAAA;IACpB;;;;;;;;OAQG;IACH,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB;;;;;OAKG;IACH,SAAS,CAAC,EAAE,UAAU,CAAA;IACtB,qFAAqF;IACrF,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B;;;;OAIG;IACH,YAAY,CAAC,EAAE,OAAO,CAAA;IACtB;;;;;;;;;;;;;;OAcG;IACH,wBAAwB,CAAC,EAAE,MAAM,GAAG,UAAU,CAAA;CAC9C"}
+{"version":3,"file":"classifier.d.ts","sourceRoot":"","sources":["../classifier.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAEH,OAAO,EAAwB,KAAK,UAAU,EAAE,MAAM,kBAAkB,CAAA;AACxE,OAAO,EAIN,WAAW,EACX,KAAK,WAAW,EAChB,KAAK,UAAU,EACf,KAAK,YAAY,EAEjB,KAAK,iBAAiB,EACtB,KAAK,mBAAmB,EACxB,KAAK,WAAW,EAChB,MAAM,yBAAyB,CAAA;AAChC,OAAO,EAAmC,KAAK,mBAAmB,EAAE,MAAM,0BAA0B,CAAA;AAGpG,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAA;AAEzD,OAAO,EAA0B,KAAK,cAAc,EAAE,MAAM,gBAAgB,CAAA;AAC5E,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,0BAA0B,CAAA;AAEhE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,kBAAkB,CAAA;AAEnD,OAAO,EAA0C,KAAK,cAAc,EAAE,MAAM,wBAAwB,CAAA;AAGpG,OAAO,EAA2B,KAAK,qBAAqB,EAAE,MAAM,0BAA0B,CAAA;AAE9F,OAAO,EAAuC,KAAK,yBAAyB,EAAE,MAAM,8BAA8B,CAAA;AAClH,OAAO,EAAE,kBAAkB,EAAE,MAAM,gBAAgB,CAAA;AAGnD,OAAO,KAAK,EAAE,kBAAkB,EAAmB,MAAM,cAAc,CAAA;AAGvE;;;;GAIG;AACH,MAAM,WAAW,YAAY;IAC5B,KAAK,CACJ,QAAQ,EAAE,MAAM,EAAE,EAClB,MAAM,CAAC,EAAE;QAAE,QAAQ,EAAE,aAAa,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC;QAAC,UAAU,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;KAAE,EAC9F,SAAS,CAAC,EAAE;QAAE,QAAQ,EAAE,aAAa,CAAC,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC;QAAC,UAAU,EAAE,aAAa,CAAC,MAAM,CAAC,CAAA;KAAE,GAC/F,OAAO,CAAC,WAAW,CAAC,CAAA;CACvB;AAED,MAAM,WAAW,6BAA6B;IAC7C,SAAS,EAAE,kBAAkB,CAAA;IAC7B,MAAM,EAAE,YAAY,CAAA;IACpB;;;;OAIG;IACH,MAAM,CAAC,EAAE,SAAS,MAAM,EAAE,CAAA;IAC1B;;;;;;;OAOG;IACH,MAAM,CAAC,EAAE,SAAS,GAAG,QAAQ,CAAA;IAC7B;;;;OAIG;IACH,WAAW,CAAC,EAAE,MAAM,EAAE,EAAE,CAAA;IACxB,sEAAsE;IACtE,gBAAgB,CAAC,EAAE,MAAM,EAAE,CAAA;IAC3B,oEAAoE;IACpE,cAAc,CAAC,EAAE,MAAM,EAAE,CAAA;IACzB;;;;;OAKG;IACH,oBAAoB,CAAC,EAAE,YAAY,CAAA;IACnC;;;;;;OAMG;IACH,gBAAgB,CAAC,EAAE,gBAAgB,CAAA;IACnC;;;;;;;;;OASG;IACH,6BAA6B,CAAC,EAAE,OAAO,CAAA;IACvC;;;;OAIG;IACH,wBAAwB,CAAC,EAAE,MAAM,GAAG,UAAU,CAAA;IAC9C;;;;;OAKG;IACH,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B;;;;;;;;;;;OAWG;IACH,YAAY,CAAC,EAAE,kBAAkB,GAAG,KAAK,CAAA;IAEzC;;;;OAIG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAA;CAChC;AAED;;GAEG;AACH,MAAM,WAAW,kBAAmB,SAAQ,qBAAqB;IAChE,oEAAoE;IACpE,OAAO,EAAE,mBAAmB,CAAA;CAC5B;AAED,qBAAa,uBAAuB;;IASvB,OAAO,CAAC,QAAQ,CAAC,GAAG;IARhC,OAAO,CAAC,QAAQ,CAAC,MAAM,CAAmB;IAC1C,OAAO,CAAC,QAAQ,CAAC,UAAU,CAAsB;IACjD,OAAO,CAAC,QAAQ,CAAC,WAAW,CAAY;IAGxC,OAAO,CAAC,QAAQ,CAAC,gBAAgB,CAAU;IAC3C,OAAO,CAAC,QAAQ,CAAC,cAAc,CAAU;gBAEZ,GAAG,EAAE,6BAA6B;IAc/D;;;OAGG;IACH,OAAO,CAAC,eAAe;IAMvB;;;;;;;;;;;OAWG;WACU,eAAe,CAC3B,IAAI,GAAE,kBAAkB,GAAG;QAAE,oBAAoB,CAAC,EAAE,YAAY,CAAC;QAAC,kBAAkB,CAAC,EAAE,MAAM,EAAE,CAAA;KAAO,GACpG,OAAO,CAAC,uBAAuB,CAAC;IAsGnC,6DAA6D;IACvD,KAAK,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,WAAW,CAAC;IAYjE;;;;;;OAMG;IACG,eAAe,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,qBAAqB,CAAC;IAqL/E,SAAS,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAC;IACzF,SAAS,CACd,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,SAAS,GAAG,iBAAiB,GACjC,OAAO,CAAC,OAAO,CAAC,MAAM,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,GAAG;QAAE,OAAO,CAAC,EAAE,WAAW,EAAE,CAAA;KAAE,CAAC;IAQzE,WAAW,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG,OAAO,CAAC,KAAK,CAAC,CAAC,YAAY,EAAE,MAAM,CAAC,CAAC,CAAC;IACnF,WAAW,CAChB,IAAI,EAAE,MAAM,EACZ,IAAI,EAAE,SAAS,GAAG,mBAAmB,GACnC,OAAO,CAAC,KAAK,CAAC,CAAC,YAAY,GAAG,SAAS,EAAE,MAAM,CAAC,CAAC,CAAC;IAQ/C,QAAQ,CAAC,IAAI,EAAE,MAAM,EAAE,IAAI,CAAC,EAAE,SAAS,GAAG;QAAE,GAAG,CAAC,EAAE,UAAU,CAAC,OAAO,WAAW,CAAC,CAAC,CAAC,CAAC,CAAA;KAAE,GAAG,OAAO,CAAC,MAAM,CAAC;IAI7G;;;;;;;;;OASG;IACH,OAAO,CAAC,mBAAmB;CAY3B;AAED,wFAAwF;AACxF,MAAM,WAAW,qBAAqB;IACrC,IAAI,EAAE,WAAW,CAAA;IACjB,MAAM,EAAE,MAAM,EAAE,EAAE,CAAA;IAClB,MAAM,EAAE,KAAK,CAAC;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAA;KAAE,CAAC,CAAA;CAC7C;AAED;;;GAGG;AACH,MAAM,WAAW,SAAS;IACzB;;;;OAIG;IACH,UAAU,CAAC,EAAE,cAAc,CAAA;IAC3B;;;OAGG;IACH,mBAAmB,CAAC,EAAE,MAAM,CAAA;IAC5B;;OAEG;IACH,GAAG,CAAC,EAAE,cAAc,CAAA;IACpB,6DAA6D;IAC7D,YAAY,CAAC,EAAE,MAAM,CAAA;IACrB;;;;;OAKG;IACH,mBAAmB,CAAC,EAAE,cAAc,CAAA;IACpC,yDAAyD;IACzD,uBAAuB,CAAC,EAAE,yBAAyB,CAAA;IACnD;;;;;OAKG;IACH,cAAc,CAAC,EAAE,OAAO,CAAA;IAExB;;;OAGG;IACH,sBAAsB,CAAC,EAAE,OAAO,CAAA;IAEhC;;;;;;OAMG;IACH,UAAU,CAAC,EAAE,OAAO,CAAA;IACpB;;;;;;;OAOG;IACH,aAAa,CAAC,EAAE,OAAO,CAAA;IACvB;;;;;OAKG;IACH,SAAS,CAAC,EAAE,UAAU,CAAA;IACtB,qFAAqF;IACrF,qBAAqB,CAAC,EAAE,OAAO,CAAA;IAC/B;;;OAGG;IACH,YAAY,CAAC,EAAE,OAAO,CAAA;IACtB;;;;;;;;;;;;;OAaG;IACH,wBAAwB,CAAC,EAAE,MAAM,GAAG,UAAU,CAAA;CAC9C"}

package/out/classifier.js

@@ -7,14 +7,14 @@
  *   `@mailwoman/core` decoder. Single user-facing entrypoint: `parse(text)` returns an
  *   `AddressTree` ready for projection into JSON / tuple / XML.
  *
- *   Convenience wrappers `parseJson` / `parseTuples` / `parseXml` project the tree on the way out.
+ *   Convenience wrappers `parseJSON` / `parseTuples` / `parseXML` project the tree on the way out.
  */
 import { conventionsForSystem } from "@mailwoman/codex";
-import { buildAddressTree, decodeAsJson, decodeAsTuples, decodeAsXml, } from "@mailwoman/core/decoder";
+import { buildAddressTree, decodeAsJSON, decodeAsTuples, decodeAsXML, } from "@mailwoman/core/decoder";
 import { proposeSpans } from "@mailwoman/core/pipeline";
 import { detectAddressSystem } from "./address-system.js";
 import { normalizeInputCase } from "./case-normalize.js";
-import { buildFstEmissionPriors } from "./fst-prior.js";
+import { buildFSTEmissionPriors } from "./fst-prior.js";
 import { STAGE2_BIO_LABELS } from "./labels.js";
 import { repairPostcodeLabels } from "./postcode-repair.js";
 import { addEmissionMatrix, buildEmissionPriors } from "./query-shape-prior.js";
@@ -25,7 +25,7 @@ import { buildCodexSpanLexicon } from "./span-proposer-lexicon.js";
 import { buildStreetMorphologyEmissionPriors } from "./street-morphology-prior.js";
 import { MailwomanTokenizer } from "./tokenizer.js";
 import { repairUnitLabels } from "./unit-repair.js";
-import { buildBioEndMask, buildBioStartMask, buildBioTransitionMask, softmax, viterbi } from "./viterbi.js";
+import { buildBIOEndMask, buildBIOStartMask, buildBIOTransitionMask, softmax, viterbi } from "./viterbi.js";
 import { enforceWordConsistency } from "./word-consistency.js";
 export class NeuralAddressClassifier {
     cfg;
@@ -40,35 +40,35 @@ export class NeuralAddressClassifier {
         this.cfg = cfg;
         this.labels = cfg.labels ?? STAGE2_BIO_LABELS;
         this.decodeMode = cfg.decode ?? "viterbi";
-        const structural = buildBioTransitionMask(this.labels);
+        const structural = buildBIOTransitionMask(this.labels);
         if (cfg.transitions) {
             this.transitions = addMatrices(structural, cfg.transitions);
         }
         else {
             this.transitions = structural;
         }
-        this.startTransitions = cfg.startTransitions ?? buildBioStartMask(this.labels);
-        this.endTransitions = cfg.endTransitions ?? buildBioEndMask(this.labels);
+        this.startTransitions = cfg.startTransitions ?? buildBIOStartMask(this.labels);
+        this.endTransitions = cfg.endTransitions ?? buildBIOEndMask(this.labels);
     }
     /**
-     * The default-ON Stage 2.7 config: codex lexicon (us/au/nz), frozen measured scales (the prior
-     * builder's own defaults). Built once per instance, only when a parse actually needs it.
+     * The default-ON Stage 2.7 config: codex lexicon (us/au/nz), frozen measured scales (the prior builder's own
+     * defaults). Built once per instance, only when a parse actually needs it.
      */
     defaultProposer() {
         this.#defaultProposerCfg ??= { lexicon: buildCodexSpanLexicon() };
         return this.#defaultProposerCfg;
     }
     /**
-     * One-call factory that resolves the weights package (or explicit paths), loads the tokenizer and
-     * ONNX runner, and returns a ready-to-use classifier.
+     * One-call factory that resolves the weights package (or explicit paths), loads the tokenizer and ONNX runner, and
+     * returns a ready-to-use classifier.
      *
-     * Resolution order: explicit paths in `opts` → `@mailwoman/neural-weights-<locale>` package →
-     * throws a single actionable error.
+     * Resolution order: explicit paths in `opts` → `@mailwoman/neural-weights-<locale>` package → throws a single
+     * actionable error.
      *
-     * **Node-only.** The dynamic imports keep `OnnxRunner` (onnxruntime-node) + `resolveWeights`
-     * (uses Node fs) out of the static dependency graph, so this file can be bundled for the browser
-     * by `@mailwoman/neural-web`. Calling this method in a browser will throw at runtime — use
-     * `loadNeuralClassifierFromUrls` from `@mailwoman/neural-web` instead.
+     * **Node-only.** The dynamic imports keep `ONNXRunner` (onnxruntime-node) + `resolveWeights` (uses Node fs) out of
+     * the static dependency graph, so this file can be bundled for the browser by `@mailwoman/neural-web`. Calling this
+     * method in a browser will throw at runtime — use `loadNeuralClassifierFromUrls` from `@mailwoman/neural-web`
+     * instead.
      */
     static async loadFromWeights(opts = {}) {
         // /* webpackIgnore: true */ tells webpack to leave the dynamic import statement intact —
@@ -76,7 +76,7 @@ export class NeuralAddressClassifier {
         // + node:fs) and throws cleanly in a browser if called. Without the directive, webpack
         // pulls onnx-runner / weights into the browser chunk graph + then chokes on the Node-only
         // builtins they reference.
-        const [{ OnnxRunner }, { resolveWeights, readLabelsFromModelCard, readCrfTransitions, readRequiredChannels }, { parseAnchorLookup }, { parseGazetteerLexicon }, { PostcodeBinaryResolver }, fs,] = await Promise.all([
+        const [{ ONNXRunner }, { resolveWeights, readLabelsFromModelCard, readCrfTransitions, readRequiredChannels }, { parseAnchorLookup }, { parseGazetteerLexicon }, { PostcodeBinaryResolver }, fs,] = await Promise.all([
             import(/* webpackIgnore: true */ "./onnx-runner.js"),
             import(/* webpackIgnore: true */ "./weights.js"),
             import(/* webpackIgnore: true */ "./anchor-inference.js"),
@@ -89,7 +89,7 @@ export class NeuralAddressClassifier {
         const crf = readCrfTransitions(resolved.crfTransitionsPath);
         const [tokenizer, runner] = await Promise.all([
             MailwomanTokenizer.loadFromFile(resolved.tokenizerPath),
-            OnnxRunner.create(resolved.modelPath),
+            ONNXRunner.create(resolved.modelPath, { executionProviders: opts.executionProviders }),
         ]);
         // --- Soft-feed (#718 D1): feed the channels the SHIPPED model was trained against ----------
         // The anchor-trained en-us model goes OOD when scored anchor-OFF (the #566/#685 crater: country
@@ -164,12 +164,11 @@ export class NeuralAddressClassifier {
         return buildAddressTree(modelText, tokens, opts?.calibrate ? { calibrate: opts.calibrate } : undefined);
     }
     /**
-     * Like `parse`, but also returns the raw per-token logits and piece offsets needed for per-span
-     * logit aggregation (Option C joint-reconcile integration). Shares the ENTIRE decode path with
-     * `parse` (one `#decode`, #481) — including the repair passes, which previously ran only in
-     * `parse`: reconcile must consume the same tokens the argmax path serves users, and the repair
-     * opts were silently ignored here before. `logits` stay RAW (pre-prior, pre-repair) — they are
-     * the model's emissions, not the decode's opinions.
+     * Like `parse`, but also returns the raw per-token logits and piece offsets needed for per-span logit aggregation
+     * (Option C joint-reconcile integration). Shares the ENTIRE decode path with `parse` (one `#decode`, #481) —
+     * including the repair passes, which previously ran only in `parse`: reconcile must consume the same tokens the
+     * argmax path serves users, and the repair opts were silently ignored here before. `logits` stay RAW (pre-prior,
+     * pre-repair) — they are the model's emissions, not the decode's opinions.
      */
     async parseWithLogits(text, opts) {
         if (text.length === 0) {
@@ -183,9 +182,9 @@ export class NeuralAddressClassifier {
         };
     }
     /**
-     * THE decode path (#481): tokenize → anchor/gazetteer features → infer → priors → CRF/argmax →
-     * tokens → repairs. Both `parse` and `parseWithLogits` consume this — never fork it; the 2026-06
-     * audit found three drift surfaces in the previous duplicated copies.
+     * THE decode path (#481): tokenize → anchor/gazetteer features → infer → priors → CRF/argmax → tokens → repairs. Both
+     * `parse` and `parseWithLogits` consume this — never fork it; the 2026-06 audit found three drift surfaces in the
+     * previous duplicated copies.
      */
     async #decode(text, opts) {
         const { pieces, ids } = this.cfg.tokenizer.encode(text);
@@ -221,7 +220,7 @@ export class NeuralAddressClassifier {
             }))
             : logits;
         if (opts?.fst) {
-            emissions = addEmissionMatrix(emissions, buildFstEmissionPriors(opts.fst, pieces, this.labels, {
+            emissions = addEmissionMatrix(emissions, buildFSTEmissionPriors(opts.fst, pieces, this.labels, {
                 biasScale: opts.fstBiasScale ?? 1.0,
             }));
         }
@@ -306,25 +305,24 @@ export class NeuralAddressClassifier {
         }
         return { tokens, logits, pieces };
     }
-    async parseJson(text, opts) {
-        return decodeAsJson(await this.parse(text, opts));
+    async parseJSON(text, opts = {}) {
+        return decodeAsJSON(await this.parse(text, opts), opts);
     }
-    async parseTuples(text, opts) {
-        return decodeAsTuples(await this.parse(text, opts));
+    async parseTuples(text, opts = {}) {
+        return decodeAsTuples(await this.parse(text, opts), opts);
     }
-    async parseXml(text, opts) {
-        return decodeAsXml(await this.parse(text, opts), opts?.xml);
+    async parseXML(text, opts) {
+        return decodeAsXML(await this.parse(text, opts), opts?.xml);
     }
     /**
-     * Guard against a silent label/emission shape overrun. When the model emits MORE logits per token
-     * than the configured label vocabulary (e.g. a Stage 3 bundle loaded with the default Stage 2
-     * labels), viterbi indexes past the transition matrix and dies with an opaque `Cannot read
-     * properties of undefined (reading '0')`. Fail fast here with a message that names the contract
-     * the caller violated.
+     * Guard against a silent label/emission shape overrun. When the model emits MORE logits per token than the configured
+     * label vocabulary (e.g. a Stage 3 bundle loaded with the default Stage 2 labels), viterbi indexes past the
+     * transition matrix and dies with an opaque `Cannot read properties of undefined (reading '0')`. Fail fast here with
+     * a message that names the contract the caller violated.
      *
-     * The opposite shape (model narrower than labels) is intentionally permitted — STAGE2_BIO_LABELS
-     * prefix-extends STAGE1_BIO_LABELS so a Stage 1 model loaded with Stage 2 labels decodes
-     * correctly via the first 15 logits. See labels.ts for the contract.
+     * The opposite shape (model narrower than labels) is intentionally permitted — STAGE2_BIO_LABELS prefix-extends
+     * STAGE1_BIO_LABELS so a Stage 1 model loaded with Stage 2 labels decodes correctly via the first 15 logits. See
+     * labels.ts for the contract.
      */
     assertEmissionWidth(logits) {
         if (logits.length === 0)
@@ -338,12 +336,11 @@ export class NeuralAddressClassifier {
     }
 }
 /**
- * Loud-degrade warning for the `loadFromWeights` soft-feed (#718 D1) — the Node mirror of
- * neural-web's `warnOnUnfedTrainedChannels`. Fired ONCE per channel per process: a model-card that
- * declares a channel REQUIRED, paired with a package that didn't ship (or could not parse) its
- * data, runs that channel OFF. Structural fallback (the parse still works), loud console (a
- * silently anchor-OFF anchor-trained model is the #566/#685 OOD crater this fix exists to
- * surface).
+ * Loud-degrade warning for the `loadFromWeights` soft-feed (#718 D1) — the Node mirror of neural-web's
+ * `warnOnUnfedTrainedChannels`. Fired ONCE per channel per process: a model-card that declares a channel REQUIRED,
+ * paired with a package that didn't ship (or could not parse) its data, runs that channel OFF. Structural fallback (the
+ * parse still works), loud console (a silently anchor-OFF anchor-trained model is the #566/#685 OOD crater this fix
+ * exists to surface).
  */
 const warnedUnfedChannels = new Set();
 function warnUnfedChannel(channel, detail) {