@mailwoman/corpus 8.6.0 → 9.0.0

package/out/src/shard-recipes/sub-venue-sources.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"sub-venue-sources.d.ts","sourceRoot":"","sources":["../../../src/shard-recipes/sub-venue-sources.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;GAyBG;AAUH,OAAO,KAAK,EAAE,eAAe,EAAE,MAAM,yBAAyB,CAAA;AAC9D,OAAO,EAAyC,KAAK,oBAAoB,EAAE,MAAM,+BAA+B,CAAA;AAChH,OAAO,EAAuB,KAAK,iBAAiB,EAAE,MAAM,kCAAkC,CAAA;AAQ9F;;;;GAIG;AACH,wBAAgB,kBAAkB,IAAI,MAAM,CAE3C;AAED;;GAEG;AACH,wBAAgB,mBAAmB,CAAC,IAAI,GAAE,MAA6B,GAAG,oBAAoB,CAE7F;AAyBD;;;;GAIG;AACH,wBAAgB,WAAW,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAYjD;AA6FD;;GAEG;AACH,wBAAgB,eAAe,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAUrD;AAMD;;GAEG;AACH,MAAM,WAAW,eAAe;IAC/B,YAAY,EAAE,MAAM,CAAA;IACpB,MAAM,EAAE,MAAM,CAAA;IACd;;OAEG;IACH,OAAO,EAAE,MAAM,CAAA;IACf,kBAAkB,EAAE,OAAO,CAAA;IAC3B,gBAAgB,EAAE,OAAO,CAAA;CACzB;AAED;;;GAGG;AACH,wBAAgB,SAAS,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAEhD;AAED;;;;;;GAMG;AACH,wBAAgB,cAAc,CAAC,SAAS,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAazE;AA4BD;;;;;GAKG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,OAAO,CAMvD;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,gBAAgB,CAC/B,SAAS,EAAE,MAAM,EACjB,QAAQ,EAAE,eAAe,EACzB,SAAS,GAAE,SAAS,MAAM,EAAO,EACjC,OAAO,UAAQ,GACb,OAAO,CAgBT;AAED;;;;;;GAMG;AACH,wBAAgB,oBAAoB,CACnC,SAAS,EAAE,MAAM,EACjB,QAAQ,EAAE,eAAe,EACzB,SAAS,GAAE,SAAS,MAAM,EAAO,EACjC,OAAO,UAAQ,GACb,OAAO,CAIT;AAED;;;;;;;;;;;GAWG;AACH,wBAAgB,mBAAmB,CAClC,MAAM,EAAE,MAAM,EACd,OAAO,EAAE,oBAAoB,EAC7B,UAAU,GAAE,SAAS,iBAAiB,EAAwB,GAC5D,eAAe,EAAE,CA6CnB;AAED;;GAEG;AACH,wBAAgB,kBAAkB,CACjC,MAAM,EAAE,MAAM,EACd,UAAU,GAAE,SAAS,iBAAiB,EAAwB,GAC5D,MAAM,EAAE,CAEV;AAMD,UAAU,WAAW;IACpB,KAAK,EAAE,MAAM,CAAA;IACb,YAAY,EAAE,MAAM,CAAA;IACpB,QAAQ,EAAE,MAAM,EAAE,CAAA;CAClB;AAED;;GAEG;AACH,MAAM,WAAW,eAAe;IAC/B,YAAY,EAAE,GAAG,CAAC,MAAM,EAAE,WAAW,EAAE,CAAC,CAAA;IACxC,MAAM,EAAE,WAAW,EAAE,CAAA;CACrB;AAoBD;;GAEG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,oBAAoB,EAAE,MAAM,EAAE,MAAM,GAAG,eAAe,CA0BnG;AAED;;;;;;GAMG;AACH,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,eAAe,EAAE,YAAY,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,MAAM,GAAG,MAAM,GAAG,IAAI,CAkBlH;AAMD;;GAEG;AACH,MAAM,WAAW,QAAQ;IACxB,OAAO,EAAE,eAAe,EAAE,CAAA;IAC1B;;OAEG;IACH,MAAM,EAAE,MAAM,EAAE,CAAA;IAChB;;OAEG;IACH,QAAQ,EAAE,MAAM,EAAE,CAAA;IAClB;;OAEG;IACH,cAAc,EAAE,MAAM,EAAE,CAAA;IACxB;;;OAGG;IACH,WAAW,EAAE,MAAM,EAAE,CAAA;IACrB;;;;;OAKG;IACH,gBAAgB,EAAE,MAAM,EAAE,CAAA;CAC1B;AAED;;;GAGG;AACH,MAAM,MAAM,SAAS,GAAG,IAAI,CAAC,QAAQ,EAAE,QAAQ,GAAG,UAAU,GAAG,gBAAgB,GAAG,aAAa,GAAG,kBAAkB,CAAC,CAAA;AAErH;;;GAGG;AACH,eAAO,MAAM,gBAAgB,EAAE,SAM9B,CAAA;AAaD;;GAEG;AACH,MAAM,WAAW,SAAS;IACzB,QAAQ,EAAE,SAAS,eAAe,EAAE,CAAA;IACpC,eAAe,EAAE,SAAS,MAAM,EAAE,CAAA;IAClC,iBAAiB,EAAE,SAAS,MAAM,EAAE,CAAA;IACpC,SAAS,EAAE,SAAS,MAAM,EAAE,CAAA;IAC5B,OAAO,EAAE,OAAO,CAAA;CAChB;AAUD;;GAEG;AACH,wBAAgB,gBAAgB,CAAC,IAAI,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,GAAG,SAAS,CAiD1E;AA8BD;;;;;GAKG;AACH,wBAAgB,YAAY,CAAC,MAAM,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,KAAK,EAAE,SAAS,GAAG,SAAS,CA2DzF;AAED;;GAEG;AACH,wBAAgB,cAAc,CAAC,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,SAAS,GAAG,SAAS,CAQpE;AAuBD;;;;GAIG;AACH,wBAAsB,iBAAiB,CAAC,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,MAAM,EAAE,eAAe,EAAE,CAAC,CAAC,CA2CjH;AAED;;;;;;;GAOG;AACH,MAAM,WAAW,eAAe;IAC/B,UAAU,EAAE,eAAe,EAAE,CAAA;IAC7B,kBAAkB,EAAE,eAAe,EAAE,CAAA;IACrC,UAAU,EAAE,eAAe,EAAE,CAAA;CAC7B;AAQD,wBAAgB,oBAAoB,CACnC,OAAO,EAAE,SAAS,eAAe,EAAE,EACnC,iBAAiB,EAAE,SAAS,MAAM,EAAE,EACpC,SAAS,EAAE,SAAS,MAAM,EAAE,EAC5B,OAAO,EAAE,MAAM,GACb,eAAe,CAkCjB"}

package/out/src/shard-recipes/sub-venue-sources.js

@@ -0,0 +1,669 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   The READ half of the `sub-venue` shard recipe (#35 step 4): which surfaces a locale may emit,
+ *   what identifier follows them in that region, and which real names become the confound negatives.
+ *   `sub-venue.ts` owns the WRITE half (line rendering + the emit loop) and the recipe registration;
+ *   split because the two halves together run past the 750-line file cap, and this is the seam — one
+ *   side reads disk and the ledger, the other side never touches either.
+ *
+ *   Every rule in this file is a rule about EVIDENCE, and each one has a measurement behind it:
+ *
+ *   - {@link promotedSurfacesFor} — a promotion names a designator, a phrase AND a locale, because the
+ *       same token is a designator in one language and a disaster in another (`hall` is 0-of-3,273 in
+ *       Great Britain and 35-of-40 in France).
+ *   - {@link hasPromotedShape} — an `identifier-required` promotion is exercised only as
+ *       `<phrase> <identifier>`. The de-DE `halle` board is the founding case: its 168-hit confound
+ *       includes the CITY Halle (Saale), and only the identifier-bearing shape separates them.
+ *   - {@link buildIdentifierModel} — the identifier distribution is measured per REGION, because it
+ *       differs by country far more than the shared vocabulary suggests (GB gates 71% bare digit, ES
+ *       35% ranges).
+ *   - {@link isVenueSlotName} / {@link isSignIdentifier} — the filters that keep bus-stop codes, route
+ *       descriptions and street names out of the slots they would mislabel. Both were written FROM
+ *       smoke output, not predicted; the docstrings name the strings that produced them.
+ */
+import { readFileSync } from "node:fs";
+import { createRequire } from "node:module";
+import { dirname, resolve } from "node:path";
+import { DatabaseSync } from "node:sqlite";
+import { parseJSONStrict } from "@mailwoman/core/objects";
+import { dataRootPath } from "@mailwoman/core/utils";
+import { classifyIdentifier, readSubVenueJSONL } from "../tools/sub-venue-lexicon.js";
+import { SUBVENUE_PROMOTIONS } from "../tools/sub-venue-promotions.js";
+import { readTuples as readLocaleTuples } from "./locale.js";
+import { makeMulberry32, readTuples as readShardTuples } from "./scaffold.js";
+//#region Lexicon
+const require_ = createRequire(import.meta.url);
+/**
+ * The committed lexicon, resolved through the package manifest rather than by counting `..` segments — the same file
+ * sits at `corpus/data/` from both the source tree and the compiled `out/` tree, and `createRequire().resolve` finds it
+ * from either without a `__isCompiledTree` flag of its own.
+ */
+export function defaultLexiconPath() {
+    return resolve(dirname(require_.resolve("@mailwoman/corpus/package.json")), "data", "sub-venue-lexicon.json");
+}
+/**
+ * Read and parse the lexicon. Strict: a corrupt lexicon is a build failure, not a fallback.
+ */
+export function readSubVenueLexicon(path = defaultLexiconPath()) {
+    return parseJSONStrict(readFileSync(path, "utf8"));
+}
+//#endregion
+//#region Name filters
+/**
+ * Longest venue name kept for the venue slot. The extracts carry a tail of route descriptions and junction names
+ * ("Furnival Gate/Moorhead MH2", "SPEKE HALL ROAD/HILLFOOT AVE") that are not venue names at all; a length cap plus
+ * {@link isCleanName} removes the bulk of them without a hand list.
+ */
+const MAX_VENUE_NAME_LENGTH = 44;
+/**
+ * Shortest kept name. Below four characters a "name" is an airport code or a platform letter, not something that can
+ * stand in a venue slot.
+ */
+const MIN_NAME_LENGTH = 4;
+/**
+ * Longest attested sub-venue string, in whitespace tokens. `Terminal 1 Flugsteig B` is four and real; anything longer
+ * is a venue's own name that happens to contain a designator.
+ */
+const MAX_ATTESTED_TOKENS = 4;
+/**
+ * Reject a name that is a route description, a junction, or a code rather than a name: embedded `/`, `;`, `,`, `:`,
+ * parentheses, no letters, or a bare source code. Measured motivation, not taste — the GB extract's `platform` tier
+ * contributes 7,549 `other`-shaped names like `kntgwdgj` and `SPEKE HALL ROAD/HILLFOOT AVE`.
+ */
+export function isCleanName(name) {
+    if (name.length < MIN_NAME_LENGTH || name.length > MAX_VENUE_NAME_LENGTH)
+        return false;
+    // A colon in a name is a qualifier, not part of it — "Porte 4 : Ferrys", "Derby College: Ilkeston Campus".
+    if (/[/;,:()[\]<>|]/.test(name))
+        return false;
+    if (!/\p{L}/u.test(name))
+        return false;
+    // A name that is all lower-case with no space is a source code (`kntgwdgj`), not a name.
+    if (!name.includes(" ") && name === name.toLowerCase())
+        return false;
+    return true;
+}
+/**
+ * Head words that make a "name" something other than a venue name, checked on the FIRST token.
+ *
+ * Two populations, both found by reading the 2026-08-05 smoke output rather than predicted:
+ *
+ * - **Street types.** A bus stop is routinely named after the street it stands on, so the FR extract offers `Rue de la
+ *   Porte Bergault` as a `porte` confound. It IS a confound, but it is a STREET, and putting it in the venue slot would
+ *   train `Rue …` as a venue name — trading one mislabel for another.
+ * - **Stop qualifiers.** British stop names carry a position prefix (`OPPOSITE BRICKLEHAMPTON HALL`, `ADJ THE GREEN`)
+ *   that names a relationship rather than a place.
+ *
+ * Per-language and short on purpose: this is a head-token filter over four Latin languages, not a street-type
+ * gazetteer. `@mailwoman/corpus` cannot reach the shipped street-type lexicon (it lives behind the gazetteer build),
+ * and a longer list here would be a second, drifting copy of it.
+ */
+const NON_VENUE_HEAD_WORDS = new Set([
+    // en
+    "street",
+    "road",
+    "lane",
+    "avenue",
+    "drive",
+    "way",
+    "close",
+    "opposite",
+    "opp",
+    "adj",
+    "adjacent",
+    "outside",
+    "near",
+    "nr",
+    "stop",
+    // fr
+    "rue",
+    "boulevard",
+    "chemin",
+    "impasse",
+    "allée",
+    "allee",
+    "route",
+    "quai",
+    "place",
+    // es / ca
+    "calle",
+    "carrer",
+    "avenida",
+    "avinguda",
+    "carretera",
+    "plaza",
+    "paseo",
+    "camino",
+    // de
+    "straße",
+    "strasse",
+    "weg",
+    "platz",
+    "gasse",
+]);
+/**
+ * Street types that appear at the END of an anglophone street name, checked on the LAST token.
+ *
+ * The head-word filter cannot see these — English streets are `<name> <type>`, so `Strawberry Hall Lane` and `Guinea
+ * Hall Mews` reached the venue slot in the second smoke and would have trained a STREET as a venue name. The
+ * street-side confound already has its own class (`designator-street`), drawn from real (street, locality, postcode)
+ * pairings, so nothing is lost by refusing these here. German is suffix-compounded rather than suffix-worded and is
+ * handled by {@link GERMAN_STREET_TAIL} instead.
+ */
+const STREET_TAIL_WORDS = new Set([
+    "street",
+    "road",
+    "lane",
+    "avenue",
+    "drive",
+    "close",
+    "crescent",
+    "mews",
+    "terrace",
+    "walk",
+    "way",
+    "court",
+    "gardens",
+    "grove",
+    "rise",
+    "row",
+    "parade",
+    "esplanade",
+]);
+const GERMAN_STREET_TAIL = /(?:straße|strasse|weg|platz|gasse|allee|ring|damm)$/;
+/**
+ * Is this name usable as a VENUE-slot string (positive venue or negative confound venue)?
+ */
+export function isVenueSlotName(name) {
+    if (!isCleanName(name))
+        return false;
+    const words = name.toLowerCase().match(/[\p{L}]+/gu) ?? [];
+    if (!words.length)
+        return false;
+    if (NON_VENUE_HEAD_WORDS.has(words[0]))
+        return false;
+    const tail = words.at(-1);
+    return !STREET_TAIL_WORDS.has(tail) && !GERMAN_STREET_TAIL.test(tail);
+}
+/**
+ * Title-case a designator phrase for rendering (`flugsteig` → `Flugsteig`). Single-token by construction: every
+ * promoted phrase in the ledger is one word.
+ */
+export function titleCase(phrase) {
+    return phrase.charAt(0).toUpperCase() + phrase.slice(1);
+}
+/**
+ * Word-boundary containment, script-aware enough for the Latin legs this recipe runs.
+ *
+ * The boundary matters: without it `gate` matches Briggate and `wing` matches Wingate, which is how a harvest teaches
+ * itself that a Yorkshire street is a sub-venue. (The lexicon's own harvest carries the same rule, and it drops the
+ * boundary only for Han/Kana, which has no word boundaries and no leg here.)
+ */
+export function containsPhrase(lowerName, phrase) {
+    let from = 0;
+    for (;;) {
+        const at = lowerName.indexOf(phrase, from);
+        if (at === -1)
+            return false;
+        const before = at === 0 ? "" : lowerName[at - 1];
+        const after = lowerName[at + phrase.length] ?? "";
+        if (!/[\p{L}\p{N}]/u.test(before) && !/[\p{L}\p{N}]/u.test(after))
+            return true;
+        from = at + 1;
+    }
+}
+/**
+ * Identifier-shape classes a sub-venue string may sample.
+ *
+ * `other` is excluded: it is the junk bucket the lexicon's own examples advertise — `C15/C15A+C15B`, `Segelflug Start
+ * 06`, `152, 240`, `de.05374048.drabenderhoehezeithstrasse`. Everything else is a real identifier register.
+ */
+const USABLE_IDENTIFIER_SHAPES = new Set([
+    "digit",
+    "letter",
+    "letter-digit",
+    "digit-letter",
+    "range",
+]);
+/**
+ * One atom of a sign identifier: a short number (`5`, `205`), a single letter (`B`), a letter-then-number (`A12`,
+ * `B05`), or a number-then-letter (`2F`, `4S`).
+ *
+ * The **single** leading letter is the load-bearing part, and it is what the third smoke found. A letter-digit ref with
+ * a MULTI-letter prefix is not an identifier read off a sign, it is a network code: the lexicon's own examples of that
+ * shape are `BS04`, `BS07`, `PWP2`, `WSW3687`, `RQ8` — campus and platform codes — and the GB extract offers `Arundel
+ * Gate AG1` … `AG124`, fourteen bus stops on a Sheffield STREET called Arundel Gate whose stop codes begin with its
+ * initials. Admitting two-letter prefixes put all fourteen in the attested pool as `unit`.
+ */
+const SIGN_IDENTIFIER_ATOM = /^(?:[0-9]{1,3}|[A-Za-z]|[A-Za-z][0-9]{1,3}|[0-9]{1,3}[A-Za-z]{1,2})$/;
+/**
+ * Is `value` what a sub-venue identifier looks like on a sign — an atom, or a range of two?
+ *
+ * Applied on top of {@link USABLE_IDENTIFIER_SHAPES}, which classifies but does not bound: `WSW3687` classifies as
+ * `letter-digit` and is a station code.
+ */
+export function isSignIdentifier(value) {
+    const parts = value.split(/[/-]/);
+    if (parts.length > 2 || parts.some((p) => !p))
+        return false;
+    return parts.every((part) => SIGN_IDENTIFIER_ATOM.test(part));
+}
+/**
+ * Does a real extract name exercise `promoted` in one of the two shapes this shard teaches?
+ *
+ * Only `<phrase> <identifier>` and (English legs) `<modifier> <phrase>` qualify. That is stricter than "contains the
+ * phrase", and the 2026-08-05 smoke is why: the loose test put `Glasgow Clyde College - Langside Campus`, `Terminal de
+ * Ferry de Bilbao` and — worst — `Halle Wohnstadt Nord` into the attested pool, the last of which is a bare German
+ * `Halle` wearing a name where the ledger requires an identifier. A whole venue's name is not a sub-venue string, and
+ * an attested string that violates the promotion's own shape constraint is not attestation of it.
+ *
+ * The follower is checked with {@link classifyIdentifier} plus {@link isSignIdentifier} rather than "any following
+ * word": `Wohnstadt` is a word, `8` is an identifier, and the de-DE board turns on exactly that difference.
+ */
+export function hasPromotedShape(lowerName, promoted, modifiers = [], english = false) {
+    if (!containsPhrase(lowerName, promoted.phrase))
+        return false;
+    const escaped = promoted.phrase.replaceAll(/[.*+?^${}()|[\]\\]/g, String.raw `\$&`);
+    const withFollower = new RegExp(`(?:^|[^\\p{L}\\p{N}])${escaped}\\s+(\\S+)`, "u");
+    const follower = withFollower.exec(lowerName)?.[1];
+    if (follower && USABLE_IDENTIFIER_SHAPES.has(classifyIdentifier(follower)) && isSignIdentifier(follower))
+        return true;
+    if (promoted.identifierRequired)
+        return false;
+    if (!english || !promoted.modifierEligible)
+        return false;
+    return modifiers.some((modifier) => new RegExp(`(?:^|[^\\p{L}\\p{N}])${modifier}\\s+${escaped}(?![\\p{L}\\p{N}])`, "u").test(lowerName));
+}
+/**
+ * {@link hasPromotedShape} plus the length cap that makes a string usable AS a sub-venue span.
+ *
+ * The two are separate because the difference decides which POOL a name lands in. `navette n2 vers terminal 2g` carries
+ * the promoted shape and is five tokens: too long to be a sub-venue string, and disqualified from being a NEGATIVE
+ * precisely because it does contain the shape. It belongs to neither pool, and only splitting the test says so.
+ */
+export function matchesPromotedShape(lowerName, promoted, modifiers = [], english = false) {
+    if ((lowerName.match(/\S+/g) ?? []).length > MAX_ATTESTED_TOKENS)
+        return false;
+    return hasPromotedShape(lowerName, promoted, modifiers, english);
+}
+/**
+ * The surfaces a locale may emit as `unit`.
+ *
+ * Two inputs, and the difference between them is the advisory/binding split the ledger's docstring names:
+ *
+ * - The SHIPPED English vocabulary (`neural/venue-structure.ts`, re-declared in the lexicon as `shipped: true`) is a flat
+ *   English list with no locale gate. It is promoted-by-shipping for the English legs, because the span proposer fires
+ *   on it there today and the eval board's target cases are drawn from it.
+ * - {@link SUBVENUE_PROMOTIONS} adds the localized surfaces and SUBTRACTS the rejections. A rejection of a shipped
+ *   designator cannot un-ship it (nothing here stops the proposer firing on "Red Wing"), but it absolutely stops this
+ *   recipe generating a positive: en-US `wing` produces negatives instead.
+ */
+export function promotedSurfacesFor(locale, lexicon, promotions = SUBVENUE_PROMOTIONS) {
+    const language = locale.split("-")[0];
+    const rejected = new Set(promotions.filter((p) => p.decision === "reject" && p.locale === locale).map((p) => `${p.designatorID}|${p.phrase}`));
+    const out = new Map();
+    if (language === "en") {
+        for (const designator of lexicon.designators) {
+            if (!designator.shipped)
+                continue;
+            const key = `${designator.id}|${designator.id}`;
+            if (rejected.has(key))
+                continue;
+            out.set(key, {
+                designatorID: designator.id,
+                phrase: designator.id,
+                surface: titleCase(designator.id),
+                identifierRequired: false,
+                modifierEligible: designator.modifierEligible,
+            });
+        }
+    }
+    for (const promotion of promotions) {
+        if (promotion.decision !== "promote" || promotion.locale !== locale)
+            continue;
+        const designator = lexicon.designators.find((d) => d.id === promotion.designatorID);
+        const key = `${promotion.designatorID}|${promotion.phrase}`;
+        if (rejected.has(key))
+            continue;
+        out.set(key, {
+            designatorID: promotion.designatorID,
+            phrase: promotion.phrase,
+            surface: titleCase(promotion.phrase),
+            identifierRequired: promotion.shape === "identifier-required",
+            // A promotion marks a SURFACE usable; it does not widen the modifier grammar (the ledger's
+            // own words). Modifier eligibility stays the designator's, and only English legs read it.
+            modifierEligible: Boolean(designator?.modifierEligible) && promotion.shape !== "identifier-required",
+        });
+    }
+    return [...out.values()].toSorted((a, b) => a.phrase.localeCompare(b.phrase));
+}
+/**
+ * The phrases REJECTED in a locale — the negatives' vocabulary.
+ */
+export function rejectedPhrasesFor(locale, promotions = SUBVENUE_PROMOTIONS) {
+    return promotions.filter((p) => p.decision === "reject" && p.locale === locale).map((p) => p.phrase);
+}
+/**
+ * Designators whose refs the pooled fallback is built from.
+ *
+ * `platform` and `station` are excluded deliberately even though they are by far the largest buckets (GB alone has
+ * 25,109 platform digits): a platform ref is a network identifier, and its `other` bucket is 7,549 rows of
+ * `kntgwdgj`-style source codes. The three kept here are the ones whose refs are what a person reads off a sign.
+ */
+const POOLED_IDENTIFIER_DESIGNATORS = ["gate", "terminal", "campus"];
+/**
+ * Minimum usable observations before a (region, designator) uses its OWN identifier distribution.
+ *
+ * Below this the sample is noise — ES `terminal` has 5 usable refs — so the leg falls back to the region's pooled
+ * gate+terminal+campus distribution, which is what the lexicon measured at volume (452–655 refs per region). The
+ * fallback keeps the axis that matters (the REGION) and drops only the per-designator refinement.
+ */
+const MIN_OWN_SHAPE_OBSERVATIONS = 20;
+/**
+ * Build the per-region identifier model out of the lexicon's `identifierShapes`.
+ */
+export function buildIdentifierModel(lexicon, region) {
+    const byDesignator = new Map();
+    const pooled = [];
+    for (const row of lexicon.identifierShapes) {
+        if (row.region !== region)
+            continue;
+        if (!USABLE_IDENTIFIER_SHAPES.has(row.shape))
+            continue;
+        const examples = row.examples.filter((e) => isSignIdentifier(e));
+        if (!examples.length)
+            continue;
+        const bucket = { shape: row.shape, observations: row.observations, examples };
+        const list = byDesignator.get(row.designatorID);
+        if (list) {
+            list.push(bucket);
+        }
+        else {
+            byDesignator.set(row.designatorID, [bucket]);
+        }
+        if (POOLED_IDENTIFIER_DESIGNATORS.includes(row.designatorID)) {
+            pooled.push(bucket);
+        }
+    }
+    return { byDesignator, pooled };
+}
+/**
+ * Draw one identifier for `designatorID` in this region.
+ *
+ * Own distribution when it has {@link MIN_OWN_SHAPE_OBSERVATIONS} usable observations, else the region's pooled one.
+ * Shapes are weighted by observation count and an example is drawn uniformly inside the chosen shape — the lexicon
+ * ships up to eight per shape, which is the resolution available.
+ */
+export function sampleIdentifier(model, designatorID, random) {
+    const own = model.byDesignator.get(designatorID) ?? [];
+    const ownTotal = own.reduce((sum, b) => sum + b.observations, 0);
+    const buckets = ownTotal >= MIN_OWN_SHAPE_OBSERVATIONS ? own : model.pooled;
+    if (!buckets.length)
+        return null;
+    const total = buckets.reduce((sum, b) => sum + b.observations, 0);
+    let r = random() * total;
+    for (const bucket of buckets) {
+        r -= bucket.observations;
+        if (r < 0)
+            return bucket.examples[Math.floor(random() * bucket.examples.length)];
+    }
+    const last = buckets.at(-1);
+    return last.examples[Math.floor(random() * last.examples.length)];
+}
+/**
+ * The pools a source that has nothing to say contributes — en-US has no OSM extract, and DE/ES/GB are outside poi.db's
+ * four countries. Empty rather than absent so a leg's merge is unconditional.
+ */
+export const EMPTY_NAME_POOLS = {
+    venues: [],
+    attested: [],
+    rejectedVenues: [],
+    longerNames: [],
+    unpromotedShapes: [],
+};
+/**
+ * Is this name a designator sitting inside a longer proper name — the "Grand Central Terminal" class the span
+ * proposer's second structural guard already knows about, and which the corpus has to agree with?
+ */
+function isLongerProperName(low, name, designatorPhrases) {
+    return designatorPhrases.some((phrase) => containsPhrase(low, phrase) && !low.startsWith(phrase) && !/\d/.test(name));
+}
+/**
+ * Read one OSM extract and split it into name pools.
+ */
+export function readExtractPools(path, query) {
+    const rows = readSubVenueJSONL(path);
+    const venues = new Set();
+    const attested = new Set();
+    const rejectedVenues = new Set();
+    const longerNames = new Set();
+    const unpromotedShapes = new Set();
+    for (const row of rows) {
+        const name = (row.name ?? "").trim();
+        if (!name || !isCleanName(name))
+            continue;
+        const low = name.toLowerCase();
+        const venueSlot = isVenueSlotName(name);
+        if (row.tier === "venue" && venueSlot) {
+            venues.add(name);
+        }
+        for (const promoted of query.promoted) {
+            if (!containsPhrase(low, promoted.phrase))
+                continue;
+            if (!hasPromotedShape(low, promoted, query.modifiers, query.english)) {
+                if (venueSlot) {
+                    unpromotedShapes.add(name);
+                }
+            }
+            else if (matchesPromotedShape(low, promoted, query.modifiers, query.english)) {
+                attested.add(name);
+            }
+        }
+        if (!venueSlot)
+            continue;
+        if (query.rejectedPhrases.some((phrase) => containsPhrase(low, phrase))) {
+            rejectedVenues.add(name);
+        }
+        if (isLongerProperName(low, name, query.designatorPhrases)) {
+            longerNames.add(name);
+        }
+    }
+    return {
+        venues: [...venues],
+        attested: [...attested],
+        rejectedVenues: [...rejectedVenues],
+        longerNames: [...longerNames],
+        unpromotedShapes: [...unpromotedShapes],
+    };
+}
+/**
+ * Poi.db category ids this recipe reads, by category NAME (ids are assigned per build, so they are resolved at run time
+ * out of `poi_category_codes`).
+ *
+ * The venue set is the transport + institution categories whose rows name a whole venue — exactly what
+ * `overture-subvenue.ts` REJECTED as a lexicon source ("4,071 of them are the token `airport` in the aerodrome's own
+ * name") and exactly what a venue slot wants. The confound set is that file's rejection list read as a source of
+ * negatives: `shoe_store` contributes 708 hits of `wing` because Red Wing sells boots, and that is the row this shard
+ * needs to see with `wing` NOT tagged `unit`.
+ */
+const POI_VENUE_CATEGORIES = [
+    "airport",
+    "airport_terminal",
+    "train_station",
+    "hospital",
+    "college_university",
+];
+const POI_CONFOUND_CATEGORIES = [
+    "shoe_store",
+    "furniture_store",
+    "home_decor_store",
+    "town_hall",
+    "martial_arts_club",
+    "chicken_wings_restaurant",
+    "fire_station",
+];
+/**
+ * Read the venue + confound pools for a country out of `poi.db`.
+ *
+ * Poi.db is FOUR COUNTRIES — US 11,521,612 / CA 794,418 / FR 721,352 / MX 644,316 — so this is reachable for en-US and
+ * fr-FR and nothing else, and a zero here is evidence of absence in four countries rather than in the world.
+ */
+export function readPOIPools(dbPath, country, query) {
+    const db = new DatabaseSync(dbPath, { readOnly: true });
+    try {
+        const codes = db.prepare("select id, category from poi_category_codes").all();
+        const byName = new Map(codes.map((c) => [c.category, c.id]));
+        const venueIDs = POI_VENUE_CATEGORIES.map((c) => byName.get(c)).filter((id) => id != null);
+        const confoundIDs = POI_CONFOUND_CATEGORIES.map((c) => byName.get(c)).filter((id) => id != null);
+        const wanted = [...venueIDs, ...confoundIDs];
+        if (!wanted.length)
+            throw new Error(`poi.db at ${dbPath} has none of the expected categories`);
+        // One filtered full scan (measured 4.3 s over all 13,681,698 rows, 2026-08-05) rather than one
+        // query per category: `poi` is `without rowid` on (h3_cell, category_id, …), so a category
+        // predicate scans either way and scanning once is the cheaper shape.
+        const rows = db
+            .prepare(`select name, category_id from poi where country = ? and name is not null and category_id in (${wanted.map(() => "?").join(",")})`)
+            .all(country, ...wanted);
+        const venueSet = new Set(venueIDs);
+        const venues = new Set();
+        const rejectedVenues = new Set();
+        const longerNames = new Set();
+        for (const row of rows) {
+            const name = row.name.trim();
+            if (!isVenueSlotName(name))
+                continue;
+            const low = name.toLowerCase();
+            if (venueSet.has(row.category_id)) {
+                venues.add(name);
+            }
+            if (query.rejectedPhrases.some((phrase) => containsPhrase(low, phrase))) {
+                rejectedVenues.add(name);
+            }
+            if (isLongerProperName(low, name, query.designatorPhrases)) {
+                longerNames.add(name);
+            }
+        }
+        return {
+            venues: [...venues],
+            attested: [],
+            rejectedVenues: [...rejectedVenues],
+            longerNames: [...longerNames],
+            unpromotedShapes: [],
+        };
+    }
+    finally {
+        db.close();
+    }
+}
+/**
+ * Merge two sources' name pools.
+ */
+export function mergeNamePools(a, b) {
+    return {
+        venues: [...a.venues, ...b.venues],
+        attested: [...a.attested, ...b.attested],
+        rejectedVenues: [...a.rejectedVenues, ...b.rejectedVenues],
+        longerNames: [...a.longerNames, ...b.longerNames],
+        unpromotedShapes: [...a.unpromotedShapes, ...b.unpromotedShapes],
+    };
+}
+//#endregion
+//#region Address context
+/**
+ * DE + ES address context. Both read through {@link readLocaleTuples}, the `locale` recipe's own streaming + reservoir
+ * reader, so the CSV handling (quoted fields, CRLF, city-noise cleaning, the DE per-part region fallback) has exactly
+ * one implementation.
+ *
+ * DE reads `europe.zip`'s two members rather than `oa-cache/de__*.zip`: the cached per-state zips the `locale` recipe
+ * names are not materialized on this host, and the archive members are byte-identical to OA's current run (verified in
+ * `corpus/AGENTS.md`'s "a file's mtime is not its data's vintage" note).
+ */
+const CONTEXT_PARTS = {
+    DE: [
+        { zip: dataRootPath("openaddresses", "europe.zip"), csv: "de/berlin.csv", region: "Berlin" },
+        { zip: dataRootPath("openaddresses", "europe.zip"), csv: "de/sn/statewide.csv", region: "Sachsen" },
+    ],
+    ES: [{ path: dataRootPath("openaddresses", "extracted", "es", "countrywide.csv") }],
+};
+/**
+ * Load the address skeletons every leg renders onto: GB / US / FR from the house-venue v3 tuples (the same 176,519 real
+ * rows the `synth-house-venue` shard is built from, so the two shards' address halves are drawn from one pool), DE and
+ * ES streamed out of OpenAddresses.
+ */
+export async function loadContextTuples(tuplesPath, seed) {
+    const byCountry = new Map();
+    for await (const tuple of readShardTuples(tuplesPath)) {
+        const country = String(tuple.country ?? "");
+        if (!country || !tuple.locality || !tuple.street)
+            continue;
+        const mapped = {
+            house_number: String(tuple.houseNumber ?? tuple.house_number ?? ""),
+            street: String(tuple.street),
+            locality: String(tuple.locality),
+            region: String(tuple.region ?? ""),
+            postcode: String(tuple.postcode ?? ""),
+        };
+        const list = byCountry.get(country);
+        if (list) {
+            list.push(mapped);
+        }
+        else {
+            byCountry.set(country, [mapped]);
+        }
+    }
+    for (const [country, parts] of Object.entries(CONTEXT_PARTS)) {
+        if (byCountry.has(country))
+            continue;
+        const pooled = [];
+        for (const [index, part] of parts.entries()) {
+            // A dedicated stream PRNG, seeded per part, so the input sample is reproducible without
+            // perturbing the emit loop's draws (the `locale` recipe's rule, kept).
+            const streamRandom = makeMulberry32(seed + index);
+            for (const tuple of await readLocaleTuples(part, streamRandom)) {
+                pooled.push(tuple);
+            }
+        }
+        byCountry.set(country, pooled);
+    }
+    return byCountry;
+}
+/**
+ * Shortest token that can carry a `-gate` street suffix and still be a NAME rather than the bare word: `gate` itself is
+ * four characters, so the class starts at five (`Highgate`, `Moorgate`, `Stonegate`).
+ */
+const MIN_GATE_SUFFIX_TOKEN_LENGTH = 5;
+export function buildStreetNegatives(context, designatorPhrases, modifiers, country) {
+    const designatorSet = new Set(designatorPhrases);
+    const modifierSet = new Set(modifiers);
+    const designator = [];
+    const modifierDesignator = [];
+    const gateSuffix = [];
+    for (const tuple of context) {
+        const tokens = tuple.street.toLowerCase().match(/[\p{L}]+/gu) ?? [];
+        let isDesignator = false;
+        let isPair = false;
+        for (const [index, token] of tokens.entries()) {
+            if (designatorSet.has(token)) {
+                isDesignator = true;
+            }
+            if (index + 1 < tokens.length && modifierSet.has(token) && designatorSet.has(tokens[index + 1])) {
+                isPair = true;
+            }
+            if (country === "GB" && token.length >= MIN_GATE_SUFFIX_TOKEN_LENGTH && token.endsWith("gate")) {
+                gateSuffix.push(tuple);
+            }
+        }
+        if (isPair) {
+            modifierDesignator.push(tuple);
+        }
+        else if (isDesignator) {
+            designator.push(tuple);
+        }
+    }
+    return { designator, modifierDesignator, gateSuffix };
+}
+//#endregion
+//# sourceMappingURL=sub-venue-sources.js.map