@mailwoman/corpus 9.4.0 → 10.1.0

package/lib/subvenue/promotions.ts

@@ -0,0 +1,293 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   The hand-authored ledger of per-locale decisions on sub-venue designator surfaces.
+ */
+
+/**
+ * One curation decision about a designator phrase in one locale.
+ *
+ * Decisions are per locale because the same token can be a designator in one language
+ * and a common confound in another.
+ */
+export interface SubVenuePromotion {
+	/**
+	 * The {@link SubVenueDesignator.id} that the phrase belongs to.
+	 */
+	designatorID: string
+	/**
+	 * The exact normalized phrase, such as `halle` for the German form of `hall`.
+	 */
+	phrase: string
+	/**
+	 * A `<lang>-<region>` locale.
+	 *
+	 * The language part matches a surface's `lang`, and the region part matches its `region`.
+	 */
+	locale: string
+	decision: "promote" | "reject"
+	/**
+	 * A shape restriction on a promotion.
+	 *
+	 * `identifier-required` promotes the phrase only as `<phrase> <identifier>`, such as `Halle 8`.
+	 * Consumers must honor it, because the bare German `Halle` is also a city.
+	 * Without this field, the promotion holds in any shape.
+	 */
+	shape?: "identifier-required"
+	/**
+	 * The count of census hits that are real venue-interior structure.
+	 */
+	real: number
+	/**
+	 * The count of census hits that are confounds.
+	 */
+	confound: number
+	/**
+	 * A description of what the confound hits are.
+	 */
+	confoundNote: string
+	/**
+	 * The extract or dataset that the census counted, plus the census date.
+	 */
+	census: string
+}
+
+/**
+ * Every curation decision, grouped by designator.
+ *
+ * This ledger is the only input that sets `curated: true` on a machine-derived surface.
+ * Rejections stay in the ledger so the system can reject duplicate proposals until new evidence arrives.
+ *
+ * The span proposer in `packages/neural/lib/venue-structure.ts` ships its English designators
+ * without a locale gate, so a rejection here guides recipes only and does not stop the proposer.
+ */
+export const SUBVENUE_PROMOTIONS: readonly SubVenuePromotion[] = [
+	// wing
+	{
+		designatorID: "wing",
+		phrase: "wing",
+		locale: "en-GB",
+		decision: "promote",
+		real: 23,
+		confound: 6,
+		confoundNote:
+			"Wing Close, Wing Road and Wing Tollgate House — the Buckinghamshire village of Wing — plus Cambian Wing " +
+			"College. All six are the <wing> <word> shape; every real hit is a hospital wing (Bexley Wing, Sorby Wing, " +
+			"Maternity Wing, West Wing, Bedford Hospital North Wing). CAVEAT, and it limits how far this generalizes: " +
+			"24 of the 29 hits sit on a public_transport=platform, because British bus stops are named after the " +
+			"hospital wing they serve. The extractor maps no building wings at all — `aile` in France is 0 hits, `ala` " +
+			"in Spain 0, `flügel` in Germany 0 — so this evidence is a property of British stop naming, not of a source " +
+			"that has wings in it.",
+		census: "great-britain.jsonl 2026-08-05, 29 name hits across name + name:<lang>",
+	},
+	{
+		designatorID: "wing",
+		phrase: "wing",
+		locale: "en-US",
+		decision: "reject",
+		real: 4,
+		confound: 3354,
+		confoundNote:
+			"Red Wing, the boot brand, is 676 shoe_store rows; chicken wings are another 759 " +
+			"(chicken_wings_restaurant 394, chicken_restaurant 365) plus 166 general restaurants. The four real hits are " +
+			"campus_building rows. This is the same word as en-GB's clean 23 and it is unusable here.",
+		census: "poi.db 2026-05-20.0, full 13,681,698-row scan 2026-08-05",
+	},
+	{
+		designatorID: "wing",
+		phrase: "wing",
+		locale: "fr-FR",
+		decision: "reject",
+		real: 0,
+		confound: 26,
+		confoundNote:
+			"Wing Chun and Wing Tsun martial-arts clubs, 15 of 26. Nothing in the French partition of the layer is a " +
+			"building wing.",
+		census: "poi.db 2026-05-20.0, full scan 2026-08-05 (FR partition, 721,352 rows)",
+	},
+
+	// hall
+	{
+		designatorID: "hall",
+		phrase: "hall",
+		locale: "en-GB",
+		decision: "reject",
+		real: 0,
+		confound: 3273,
+		confoundNote:
+			"3,204 of 3,273 hits sit on a public_transport=platform: a British bus stop named after the village hall it " +
+			"stands outside. The distribution is Village Hall, Town Hall, British Legion Hall, and then Hall Lane / Hall " +
+			"Road / SPEKE HALL ROAD — streets. The 52 that take a <hall> <identifier> shape are bus-stop position codes " +
+			"(Village Hall (o/s), Town Hall (3)), and all 16 <modifier> <hall> hits are streets (East Hall Lane, North " +
+			"Hall Road).",
+		census: "great-britain.jsonl 2026-08-05, 3,273 name hits",
+	},
+	{
+		designatorID: "hall",
+		phrase: "hall",
+		locale: "en-US",
+		decision: "reject",
+		real: 2095,
+		confound: 27_081,
+		confoundNote:
+			"City Hall (5,066 town_hall rows), Kingdom Hall of Jehovah's Witnesses (2,312), event halls (2,622) and " +
+			"dormitory halls (1,989). Every one names a whole VENUE, not interior structure, so promoting the surface " +
+			"would tag them `unit`. The 2,095 real hits are US campus buildings (UAA Cuddy Hall) and they are outnumbered " +
+			"13 to 1.",
+		census: "poi.db 2026-05-20.0, full 13,681,698-row scan 2026-08-05",
+	},
+	{
+		designatorID: "hall",
+		phrase: "halle",
+		locale: "de-DE",
+		decision: "promote",
+		shape: "identifier-required",
+		real: 32,
+		confound: 168,
+		confoundNote:
+			"The confound is 97 hits of the CITY Halle — Halle (Saale), population 240,000, plus Halle (Westf) — and 68 " +
+			"village and event halls (PHOENIX Halle, Urexweiler Halle), which are whole venues. That is a worse " +
+			"confound than wave 1 assumed when it called `Halle 2` the German counter-example to en-GB: the word is " +
+			"also a locality. It is promoted anyway because the confound is SEPARABLE BY SHAPE, and that was checked " +
+			"rather than assumed — dumping all 32 <halle> <identifier> hits returns numbered factory, trade-fair and " +
+			"airport halls (VW Halle 42, Audi GVZ Halle G, Messe West Halle 8, Flughafen Halle 7, Speyer Halle 101) and " +
+			"not one instance of the city. Never modifier-eligible and never proposed bare.",
+		census: "germany.jsonl 2026-08-05 (403,863 rows), 200 hits; all 32 shape hits enumerated",
+	},
+	{
+		designatorID: "hall",
+		phrase: "hall",
+		locale: "fr-FR",
+		decision: "promote",
+		real: 35,
+		confound: 5,
+		confoundNote:
+			"Five: an English `Town Hall` and `Multipurpose hall` on French name:en tags, plus `Manning Hall`. French " +
+			"uses `hall` as the airport designator directly — 19 hits are Charles de Gaulle (Terminal 2E - Hall K/L/M, " +
+			"Terminal 1 - Hall A/B, Hall D) and 16 more are building halls (Hall des Sources, Hall du Marché, Hôpital " +
+			"Hall Sud). The word that is poison in en-GB is the ordinary designator one country away.",
+		census: "france.jsonl 2026-08-05 (251,260 rows), 40 hits; all 19 shape hits enumerated",
+	},
+
+	// gate: the English surface ships without modifier eligibility.
+	// These entries cover its localized forms.
+	{
+		designatorID: "gate",
+		phrase: "flugsteig",
+		locale: "de-DE",
+		decision: "promote",
+		real: 19,
+		confound: 0,
+		confoundNote:
+			"None, at full-country scale. 18 of 19 hits are <flugsteig> <identifier> (Flugsteig 10, Flugsteig 9/9A, " +
+			"Terminal 1 Flugsteig B) and the last is bare; context is 12 terminal features and 7 gates. The word is a " +
+			"pure aviation term with no street, locality or business collision in German. Frankfurt's own name:en " +
+			"renders Flugsteig J as `Pier J`, which is a second attestation of the same feature under a second " +
+			"designator. A 11-of-11 read on the Hessen extract was the hypothesis; this is the count that earned it.",
+		census: "germany.jsonl 2026-08-05 (403,863 rows), 19 hits across name + name:<lang>",
+	},
+	{
+		designatorID: "gate",
+		phrase: "porte",
+		locale: "fr-FR",
+		decision: "reject",
+		real: 19,
+		confound: 927,
+		confoundNote:
+			"946 hits, of which 894 are Paris city gates and the Métro stations named after them — Porte d'Orléans, " +
+			"Porte Dauphine, Porte de Champerret, Porte de Saint-Cloud. Unlike `halle`, the confound REACHES the " +
+			"designator+identifier shape: 17 of the 36 hits in that bucket are Porte Saint-Martin, Porte Saint-Denis, " +
+			"Porte Notre-Dame. Shape cannot separate it, which is the same test that killed `pier` for en-US. The 19 " +
+			"real hits are ferry and cruise gates (Porte 4 : Ferrys, Berliet Porte C).",
+		census: "france.jsonl 2026-08-05 (251,260 rows), 946 hits; all 36 shape hits enumerated",
+	},
+
+	// pier
+	{
+		designatorID: "pier",
+		phrase: "pier",
+		locale: "en-GB",
+		decision: "promote",
+		real: 120,
+		confound: 44,
+		confoundNote:
+			"Pier Road, Pier Street, Pier Avenue, Pier Terrace — 44 street names, all of them the <pier> <street-type> " +
+			"shape, which the designator+identifier rule cannot reach. The real side is 98 named piers (Eastbourne Pier, " +
+			"Cadogan Pier), 10 in <pier> <identifier> shape (Pier 1, Pier 2, `Terminal 3, Pier 6`), 10 bare and 2 with a " +
+			"modifier. Promoted as a SURFACE only; not modifier-eligible, on the same reasoning that excludes `gate`.",
+		census: "great-britain.jsonl 2026-08-05, 164 name hits",
+	},
+	{
+		designatorID: "pier",
+		phrase: "pier",
+		locale: "en-US",
+		decision: "reject",
+		real: 278,
+		confound: 2330,
+		confoundNote:
+			"Pier 1 Imports and its franchises are 341 home_decor_store plus 217 furniture_store rows, and they carry the " +
+			"<pier> <identifier> shape EXACTLY — so unlike en-GB the confound is not separable by shape. Another 159 are " +
+			"seafood restaurants and 144 are parks.",
+		census: "poi.db 2026-05-20.0, full 13,681,698-row scan 2026-08-05",
+	},
+
+	// terminal: the English surface ships.
+	// These entries cover its localized forms.
+	{
+		designatorID: "terminal",
+		phrase: "terminal",
+		locale: "es-ES",
+		decision: "promote",
+		real: 190,
+		confound: 0,
+		confoundNote:
+			"None. Spanish borrows the English word unchanged and uses it in every shape the task targets: Terminal A, " +
+			"Terminal T1, Terminal 4S, plus named terminals (Terminal Marítima de Santander). No Spanish street type " +
+			"collides with it. The Wikidata label `terminal aeroportuaria` is the encyclopaedic form and stays uncurated; " +
+			"this promotes the derived head noun.",
+		census: "spain.jsonl 2026-08-05, 190 hits across name + name:<lang>",
+	},
+	{
+		designatorID: "terminal",
+		phrase: "terminal",
+		locale: "ca-ES",
+		decision: "promote",
+		real: 15,
+		confound: 0,
+		confoundNote:
+			"None. Catalan is the same borrowing (Terminal 1, Terminal Sortides, Terminal Arribades) and needs its own " +
+			"entry because a surface's language subtag is `ca`, which an es-ES decision does not reach.",
+		census: "spain.jsonl 2026-08-05, 15 name:ca hits",
+	},
+	{
+		designatorID: "terminal",
+		phrase: "terminal",
+		locale: "fr-FR",
+		decision: "promote",
+		real: 169,
+		confound: 0,
+		confoundNote:
+			"None. 99 hits are <terminal> <identifier> (Aéroport CDG Terminal 2F, Terminal 2G, Lyon Saint-Exupéry " +
+			"Terminal 1), 51 are named terminals (Terminal de Grande-Bretagne, Terminal Croisières) and 13 trail a " +
+			"venue name. French borrows the word unchanged and no French street type collides with it.",
+		census: "france.jsonl 2026-08-05 (251,260 rows), 169 hits across name + name:<lang>",
+	},
+	{
+		designatorID: "terminal",
+		phrase: "ターミナル",
+		locale: "ja-JP",
+		decision: "promote",
+		real: 1213,
+		confound: 2,
+		confoundNote:
+			"Two hits are `ターミナル前` — a bus stop named 'in front of the terminal', which is arguably still real. " +
+			"Japanese has no street-name class containing the word. 246 hits take the Japanese target shape " +
+			"<identifier><designator> (第2ターミナル, 羽田空港第1ターミナル); 957 are compounds naming a terminal " +
+			"(バスターミナル, フェリーターミナル, 成田空港第一ターミナル); 10 are <designator> <identifier>. This is the " +
+			"surface the corpus task named as missing, and it exists only because the head-noun derivation produced it " +
+			"from five compound Wikidata labels first.",
+		census: "japan.jsonl 2026-08-05, 1,215 hits across name + name:<lang>",
+	},
+]

package/lib/subvenue/surfaces.ts

@@ -0,0 +1,94 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ * @file Surface-form primitives for the sub-venue lexicon. They normalize phrases and index records by phrase.
+ * They also decide whether a feature name contains a known designator.
+ *
+ * The two matching rules depend on script. Both narrowings are required.
+ */
+
+import { normalizeWhitespace } from "@mailwoman/core/strings/format"
+
+import type { SubVenueSurface } from "#subvenue/table"
+
+/**
+ * Scripts whose case is meaningful to fold.
+ * Everything else is left as written.
+ */
+const CASE_FOLDING_SCRIPT = /^[\p{Script=Latin}\p{Script=Cyrillic}\p{Script=Greek}\d\s\p{P}]+$/u
+
+/**
+ * Scripts written without spaces between words, where a token split cannot find a
+ * designator and a substring test is the correct operator.
+ *
+ * Han, Hiragana and Katakana.
+ * Hangul is excluded because Korean does space its words.
+ *
+ * The Germanic-compound argument that keeps {@link nameContainsSurfaces}
+ * token-bounded for Latin does not transfer here.
+ */
+const NON_SPACING_SCRIPT = /[\p{Script=Han}\p{Script=Hiragana}\p{Script=Katakana}]/u
+
+/**
+ * Normalize a surface for the table by trimming, collapsing internal whitespace,
+ * then lowercasing only when the string is entirely in a bicameral script.
+ */
+export function normalizeSurface(text: string): string {
+	const trimmed = normalizeWhitespace(text)
+
+	return CASE_FOLDING_SCRIPT.test(trimmed) ? trimmed.toLowerCase() : trimmed
+}
+
+/**
+ * A phrase-to-record index keyed by the normalized phrase.
+ * It attributes a matched phrase to its record.
+ */
+export type SurfaceIndex = ReadonlyMap<string, { recordID: string; recordKind: "designator" | "modifier" }>
+
+/**
+ * Index the surfaces accumulated so far by phrase.
+ *
+ * First writer wins, so a seed record beats a Wikidata alias that happens to collide.
+ */
+export function buildSurfaceIndex(surfaces: readonly SubVenueSurface[]): SurfaceIndex {
+	const index = new Map<string, { recordID: string; recordKind: "designator" | "modifier" }>()
+
+	for (const surface of surfaces) {
+		if (index.has(surface.phrase)) continue
+		index.set(surface.phrase, { recordID: surface.recordID, recordKind: surface.recordKind })
+	}
+
+	return index
+}
+
+/**
+ * Every known phrase found in `name`, as whole-token runs for spacing scripts
+ * and as substrings for non-spacing ones.
+ *
+ * Latin matching is token-bounded rather than substring, so `Nordterminal` is a deliberate miss.
+ * Han and Kana have no word boundaries, so the longest known substring is the correct operator there.
+ */
+export function nameContainsSurfaces(name: string, index: SurfaceIndex): string[] {
+	const normalized = normalizeSurface(name)
+	const hits = new Set<string>()
+
+	for (const token of normalized.split(/[\s,()/]+/u)) {
+		const stripped = token.replaceAll(/^[^\p{L}\p{N}]+|[^\p{L}\p{N}]+$/gu, "")
+
+		if (stripped && index.has(stripped)) {
+			hits.add(stripped)
+		}
+	}
+
+	if (NON_SPACING_SCRIPT.test(normalized)) {
+		for (const [phrase] of index) {
+			if (NON_SPACING_SCRIPT.test(phrase) && normalized.includes(phrase)) {
+				hits.add(phrase)
+			}
+		}
+	}
+
+	return [...hits]
+}

package/lib/subvenue/table.ts

@@ -0,0 +1,264 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ * @file The sub-venue lexicon's record schema and the shipped vocabulary it is seeded from.
+ *
+ * The seeds mirror `neural/venue-structure.ts`. `@mailwoman/corpus` has no dependency on `@mailwoman/neural`,
+ * so this file declares the shipped vocabulary locally. `sub-venue-lexicon.test.ts` pins the seed values
+ * and fails when either copy changes.
+ */
+
+import type { SubVenuePromotion } from "#subvenue/promotions"
+
+/**
+ * This table's own data version.
+ *
+ * Bump when the source vintages or the build semantics change.
+ */
+export const SUBVENUE_LEXICON_VERSION = "0.2.0"
+
+/**
+ * Which side of the containment relation a designator names.
+ *
+ * It mirrors `@mailwoman/osm/sdk`'s `SubVenueTier`, re-declared for the same
+ * dependency-direction reason as the seed.
+ */
+export const LexiconTier = {
+	SubVenue: "subvenue",
+	Venue: "venue",
+} as const
+
+export type LexiconTier = (typeof LexiconTier)[keyof typeof LexiconTier]
+
+/**
+ * One designator record — a venue-interior (or containing-venue) structural noun.
+ */
+export interface SubVenueDesignator {
+	/**
+	 * Canonical id, lowercase English.
+	 *
+	 * It matches `neural/venue-structure.ts`'s `VENUE_STRUCTURE_DESIGNATORS` wherever the two overlap.
+	 */
+	id: string
+	tier: LexiconTier
+	/**
+	 * Whether this designator may be preceded by a {@link SubVenueModifier} — the `North Terminal` shape;
+	 * `gate` and `building` are excluded because `<modifier> <id>` is also an ordinary street
+	 * name ("East Gate"), so admitting them turns a correct street parse into a sub-venue one.
+	 */
+	modifierEligible: boolean
+	/**
+	 * Whether the shipped span proposer recognizes this designator; `false` means no consumer reads it yet.
+	 */
+	shipped: boolean
+	/**
+	 * Where the term comes from, one entry per attesting source (`wof:placetype`, `osm:aeroway=terminal`,
+	 * `wikidata:Q849706`, `overture:airport_terminal`), sorted so a regenerate is stable.
+	 */
+	provenance: string[]
+}
+
+/**
+ * One positional modifier — the `North`/`Upper`/`Main` half of `North Terminal`.
+ */
+export interface SubVenueModifier {
+	id: string
+	shipped: boolean
+	provenance: string[]
+}
+
+/**
+ * One surface form stores a phrase.
+ *
+ * It records the corresponding entity and where the source attested it.
+ */
+export interface SubVenueSurface {
+	/**
+	 * The phrase, lowercased for Latin script and left as written otherwise.
+	 *
+	 * The Turkish `İ` (U+0130) is `\p{Script=Latin}` and folds to `i` plus a combining dot above, so it
+	 * round-trips through a form its own locale would not write (no shipped code depends on that today).
+	 */
+	phrase: string
+	recordID: string
+	recordKind: "designator" | "modifier"
+	/**
+	 * BCP-47-ish language subtag as the source wrote it (`en`, `ja`, `zh-Hant`, `pt-BR`),
+	 * or `und` when the source gave an untagged default name.
+	 */
+	lang: string
+	/**
+	 * ISO 3166-1 alpha-2 of the data the phrase was attested in, `""` for vocabulary
+	 * sources that attest a term's existence rather than its use.
+	 *
+	 * This is the axis promotion is decided on (`hall` is attested 3,274 times in `GB`), never a global census.
+	 */
+	region: string
+	/**
+	 * `wikidata:label`, `wikidata:alt`, `osm:name`, `osm:name:<lang>`, `overture:name`,
+	 * `derived:head-noun`, or `seed`.
+	 */
+	source: string
+	/**
+	 * Whether a human has approved this surface for parsing use in its region.
+	 *
+	 * Machine-derived entries start `false`; only a matching {@link SubVenuePromotion}
+	 * flips it, so a consumer that gates a parse must filter on this.
+	 */
+	curated: boolean
+	/**
+	 * How many source features attested this exact phrase when the source counts (OSM, Overture);
+	 * Vocabulary sources attest phrase existence and use a count of `0`.
+	 * OSM and Overture sources count occurrences.
+	 */
+	observations: number
+	/**
+	 * The rule-assigned designator of the features that had this phrase, with a count each
+	 * (`platform:3205 campus:49` for GB's `hall`); empty for vocabulary sources.
+	 *
+	 * Without it an `observations` count is a magnitude with no sign (`hall` on a `platform` row is a
+	 * British bus stop called after a village hall. On a `terminal` row it is a real German departure hall).
+	 */
+	context: Record<string, number>
+}
+
+/**
+ * The measured shape of a designator's identifier half — what follows `Gate`/`Terminal` in real data.
+ *
+ * Derived from OSM `ref` values.
+ * Each of Berlin's 26 `aeroway=gate` features has a `ref` and no name.
+ *
+ * The extractor renders `Gate A12` from `<designator> <ref>`.
+ * Generation requires the identifier distribution.
+ */
+export interface IdentifierShape {
+	designatorID: string
+	/**
+	 * ISO 3166-1 alpha-2 of the extract this distribution was measured in.
+	 *
+	 * It is per-region because the shapes differ (GB gates are 70% bare digits, Japanese platform
+	 * refs a different range), so a recipe for a French address should sample France's distribution.
+	 */
+	region: string
+	/**
+	 * A coarse class: `digit` (`5`), `letter` (`B`), `letter-digit` (`A12`),
+	 * `digit-letter` (`2F`), `range` (`16-18`, `0/1`), or `other`.
+	 */
+	shape: string
+	observations: number
+	/**
+	 * Up to eight real values, sorted.
+	 */
+	examples: string[]
+}
+
+/**
+ * One input source's provenance, copied off its fetch manifest.
+ */
+export interface SubVenueLexiconSource {
+	id: string
+	origin: string
+	license: string
+	retrieved: string
+	rows: number
+}
+
+/**
+ * The committed table.
+ */
+export interface SubVenueLexiconTable {
+	version: string
+	sources: SubVenueLexiconSource[]
+	designators: SubVenueDesignator[]
+	modifiers: SubVenueModifier[]
+	surfaces: SubVenueSurface[]
+	identifierShapes: IdentifierShape[]
+	/**
+	 * Every curation decision taken against this table, promotion and rejection,
+	 * each with the census that backs it.
+	 *
+	 * A rejection is as required as a promotion because it stops the next reader
+	 * re-proposing `hall` for en-GB.
+	 */
+	promotions: SubVenuePromotion[]
+}
+
+/**
+ * The vocabulary that already ships in `neural/venue-structure.ts`, re-declared
+ * below for the module docstring's dependency-direction reason.
+ *
+ * `tier` is added here: the seven WOF placetypes plus `terminal`/`gate` are all venue-interior,
+ * while `campus` and `building` name a whole venue as often as a part of one
+ * but are marked `subvenue` because that is the role the span proposer uses them in.
+ */
+export const SHIPPED_DESIGNATOR_SEED: ReadonlyArray<{
+	id: string
+	modifierEligible: boolean
+	provenance: string[]
+}> = [
+	{ id: "arcade", modifierEligible: true, provenance: ["wof:placetype"] },
+	{ id: "building", modifierEligible: false, provenance: ["wof:placetype"] },
+	{ id: "campus", modifierEligible: true, provenance: ["wof:placetype"] },
+	{ id: "concourse", modifierEligible: true, provenance: ["wof:placetype"] },
+	{ id: "enclosure", modifierEligible: false, provenance: ["wof:placetype"] },
+	{ id: "gate", modifierEligible: false, provenance: ["osm:aeroway=gate"] },
+	{ id: "installation", modifierEligible: false, provenance: ["wof:placetype"] },
+	{ id: "terminal", modifierEligible: true, provenance: ["osm:aeroway=terminal"] },
+	{ id: "wing", modifierEligible: true, provenance: ["wof:placetype"] },
+]
+
+/**
+ * The shipped positional modifiers, re-declared from `neural/venue-structure.ts`'s `VENUE_STRUCTURE_MODIFIERS`.
+ */
+export const SHIPPED_MODIFIER_SEED: readonly string[] = [
+	"central",
+	"east",
+	"front",
+	"inner",
+	"lower",
+	"main",
+	"north",
+	"outer",
+	"rear",
+	"south",
+	"upper",
+	"west",
+]
+
+/**
+ * Designators the lexicon adds beyond what ships, each with the source that attests it.
+ *
+ * None is `modifierEligible` because that claim requires a confound board for each term and locale.
+ * A promotion marks a surface usable while preserving the modifier grammar.
+ */
+export const PROPOSED_DESIGNATORS: ReadonlyArray<{
+	id: string
+	tier: LexiconTier
+	provenance: string[]
+}> = [
+	{ id: "airport", tier: LexiconTier.Venue, provenance: ["osm:aeroway=aerodrome"] },
+	{ id: "hall", tier: LexiconTier.SubVenue, provenance: ["wikidata:Q240854"] },
+	{ id: "pier", tier: LexiconTier.SubVenue, provenance: ["overture:pier"] },
+	{ id: "platform", tier: LexiconTier.SubVenue, provenance: ["osm:public_transport=platform", "osm:railway=platform"] },
+	{ id: "satellite", tier: LexiconTier.SubVenue, provenance: ["wikidata:Q15990706"] },
+	{ id: "station", tier: LexiconTier.Venue, provenance: ["osm:railway=station"] },
+]
+
+/**
+ * `designatorID` → Wikidata QID, matching `fetch/wikidata-subvenue.ts`'s `SUBVENUE_CONCEPTS`.
+ *
+ * The builder remains a pure function over parsed input.
+ * The test compares both maps.
+ */
+export const CONCEPT_QIDS: Readonly<Record<string, string>> = {
+	terminal: "Q849706",
+	gate: "Q247739",
+	concourse: "Q862212",
+	campus: "Q209465",
+	building: "Q41176",
+	arcade: "Q186637",
+	hall: "Q240854",
+	satellite: "Q15990706",
+}