@mailwoman/corpus 9.1.0 → 9.3.0

package/lib/recipes/sub-venue-sources.ts

@@ -0,0 +1,877 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   The READ half of the `sub-venue` slice 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 split — 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 { dataRootPath } from "@mailwoman/core/data-root"
+import { readLocalJSONFile } from "@mailwoman/core/fs/readers"
+import { resolveModulePath } from "@mailwoman/core/module/resolvers"
+import type { POIDatabase } from "@mailwoman/resolver-wof-sqlite/poi"
+import { DatabaseClient } from "@mailwoman/sqlite/client"
+import type { PathBuilderLike } from "path-ts"
+
+import { readTuples as readLocaleTuples, type LocalePart } from "#recipes/locale"
+import { makeMulberry32, readTuples as readSliceTuples } from "#recipes/scaffold"
+import type { LocaleBaseTuple } from "#synthesizers/german"
+import { pick, weightedPick } from "#synthesizers/utils"
+import {
+	classifyIdentifier,
+	readSubVenueJSONL,
+	type SubVenueLexiconTable,
+	SUBVENUE_PROMOTIONS,
+	type SubVenuePromotion,
+} from "#tools"
+
+//#region Lexicon
+
+/**
+ * The committed lexicon, resolved through the package manifest.
+ */
+export function defaultLexiconPath(): string {
+	return resolveModulePath("@mailwoman/corpus/data/sub-venue-lexicon.json")
+}
+
+/**
+ * Read and parse the lexicon. Strict: a corrupt lexicon is a build failure, not a fallback.
+ */
+export async function readSubVenueLexicon(path: string = defaultLexiconPath()): Promise<SubVenueLexiconTable> {
+	return await readLocalJSONFile<SubVenueLexiconTable>(path)
+}
+
+//#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: string): boolean {
+	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: ReadonlySet<string> = 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: ReadonlySet<string> = 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: string): boolean {
+	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)
+}
+
+//#endregion
+
+//#region Promotions
+
+/**
+ * A (designator, surface) pair usable in one locale, with the shape constraint the ledger attached to it.
+ */
+export interface PromotedSurface {
+	designatorID: string
+	phrase: string
+	/**
+	 * Rendering form — the phrase title-cased, which is how a sign writes it.
+	 */
+	surface: string
+	identifierRequired: boolean
+	modifierEligible: boolean
+}
+
+/**
+ * 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: string): string {
+	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: string, phrase: string): boolean {
+	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: ReadonlySet<string> = 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 required 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: string): boolean {
+	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 slice 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: string,
+	promoted: PromotedSurface,
+	modifiers: readonly string[] = [],
+	english = false
+): boolean {
+	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+${escaped}(?![\\p{L}\\p{N}])`, "u").test(lowerName)
+	)
+}
+
+/**
+ * {@link hasPromotedShape} plus the length cap that makes a string usable AS a sub-venue span.
+ *
+ * The two are separate because the difference decides which POOL a name lands in. `navette n2 vers terminal 2g` carries
+ * the promoted shape and is five tokens: too long to be a sub-venue string, and disqualified from being a NEGATIVE
+ * precisely because it does contain the shape. It belongs to neither pool, and only splitting the test says so.
+ */
+export function matchesPromotedShape(
+	lowerName: string,
+	promoted: PromotedSurface,
+	modifiers: readonly string[] = [],
+	english = false
+): boolean {
+	if ((lowerName.match(/\S+/g) ?? []).length > MAX_ATTESTED_TOKENS) return false
+
+	return hasPromotedShape(lowerName, promoted, modifiers, english)
+}
+
+/**
+ * The surfaces a locale may emit as `unit`.
+ *
+ * Two inputs, and the difference between them is the advisory/binding split the ledger's docstring names:
+ *
+ * - The SHIPPED English vocabulary (`neural/venue-structure.ts`, re-declared in the lexicon as `shipped: true`) is a flat
+ *   English list with no locale gate. It is promoted-by-shipping for the English legs, because the span proposer fires
+ *   on it there today and the eval board's target cases are drawn from it.
+ * - {@link SUBVENUE_PROMOTIONS} adds the localized surfaces and SUBTRACTS the rejections. A rejection of a shipped
+ *   designator cannot un-ship it (nothing here stops the proposer firing on "Red Wing"), but it absolutely stops this
+ *   recipe generating a positive: en-US `wing` produces negatives instead.
+ */
+export function promotedSurfacesFor(
+	locale: string,
+	lexicon: SubVenueLexiconTable,
+	promotions: readonly SubVenuePromotion[] = SUBVENUE_PROMOTIONS
+): PromotedSurface[] {
+	const language = locale.split("-")[0]!
+
+	const rejected = new Set(
+		promotions.filter((p) => p.decision === "reject" && p.locale === locale).map((p) => `${p.designatorID}|${p.phrase}`)
+	)
+
+	const out = new Map<string, PromotedSurface>()
+
+	if (language === "en") {
+		for (const designator of lexicon.designators) {
+			if (!designator.shipped) continue
+			const key = `${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: string,
+	promotions: readonly SubVenuePromotion[] = SUBVENUE_PROMOTIONS
+): string[] {
+	return promotions.filter((p) => p.decision === "reject" && p.locale === locale).map((p) => p.phrase)
+}
+
+//#endregion
+
+//#region Identifier sampling
+
+interface ShapeBucket {
+	shape: string
+	observations: number
+	examples: string[]
+}
+
+/**
+ * A region's identifier distributions: per designator, plus the pooled fallback.
+ */
+export interface IdentifierModel {
+	byDesignator: Map<string, ShapeBucket[]>
+	pooled: ShapeBucket[]
+}
+
+/**
+ * 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: readonly string[] = ["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: SubVenueLexiconTable, region: string): IdentifierModel {
+	const byDesignator = new Map<string, ShapeBucket[]>()
+	const pooled: ShapeBucket[] = []
+
+	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: ShapeBucket = { 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: IdentifierModel, designatorID: string, random: () => number): string | null {
+	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
+
+	// `inclusive: false` keeps this draw's original strict `r < 0` boundary, so the bucket stream is unchanged.
+	const bucket = weightedPick(buckets, random, (b) => b.observations, { inclusive: false })
+
+	return pick(bucket.examples, random)
+}
+
+//#endregion
+
+//#region Pools
+
+/**
+ * Per-leg pools read off disk once.
+ */
+export interface LegPools {
+	context: LocaleBaseTuple[]
+	/**
+	 * Real venue names for the venue slot (stations, airports, campuses; US: airports, terminals, hospitals, rail).
+	 */
+	venues: string[]
+	/**
+	 * Real sub-venue strings, already filtered to this locale's promoted surfaces and their shape constraint.
+	 */
+	attested: string[]
+	/**
+	 * Real names carrying a surface REJECTED in this locale, for the venue slot of a negative row.
+	 */
+	rejectedVenues: string[]
+	/**
+	 * Real names that contain a designator inside a longer proper name — "Lochaline Ferry Terminal", "Kingdom Hall". The
+	 * whole string is `venue`; nothing in it is `unit`.
+	 */
+	longerNames: string[]
+	/**
+	 * Real names carrying a PROMOTED phrase in a shape the promotion does NOT cover — `Halle Rosengarten`, `PHOENIX
+	 * Halle`, `Halle-Südstadt`. The other half of an `identifier-required` ruling, and the only thing that teaches the
+	 * shape boundary rather than the word: de-DE has no `reject` row at all, so without this class its 168-hit confound
+	 * (97 of them the CITY Halle) would go untaught while its 32-hit promotion got 11,000 rows.
+	 */
+	unpromotedShapes: string[]
+}
+
+/**
+ * The name pools a source contributes. `attested` and `unpromotedShapes` come only from an extract — poi.db carries no
+ * `tier` and no localized names, so it cannot say which side of a shape boundary a name sits on.
+ */
+export type NamePools = Pick<LegPools, "venues" | "attested" | "rejectedVenues" | "longerNames" | "unpromotedShapes">
+
+/**
+ * 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: NamePools = {
+	venues: [],
+	attested: [],
+	rejectedVenues: [],
+	longerNames: [],
+	unpromotedShapes: [],
+}
+
+/**
+ * What every pool reader needs to know about the leg it is reading for.
+ */
+export interface PoolQuery {
+	promoted: readonly PromotedSurface[]
+	rejectedPhrases: readonly string[]
+	designatorPhrases: readonly string[]
+	modifiers: readonly string[]
+	english: boolean
+}
+
+/**
+ * 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: string, name: string, designatorPhrases: readonly string[]): boolean {
+	return designatorPhrases.some((phrase) => containsPhrase(low, phrase) && !low.startsWith(phrase) && !/\d/.test(name))
+}
+
+/**
+ * Read one OSM extract and split it into name pools.
+ */
+export async function readExtractPools(path: string, query: PoolQuery): Promise<NamePools> {
+	const rows = await readSubVenueJSONL(path)
+	const venues = new Set<string>()
+	const attested = new Set<string>()
+	const rejectedVenues = new Set<string>()
+	const longerNames = new Set<string>()
+	const unpromotedShapes = new Set<string>()
+
+	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 slice
+ * needs to see with `wing` NOT tagged `unit`.
+ */
+const POI_VENUE_CATEGORIES: readonly string[] = [
+	"airport",
+	"airport_terminal",
+	"train_station",
+	"hospital",
+	"college_university",
+]
+
+const POI_CONFOUND_CATEGORIES: readonly string[] = [
+	"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: PathBuilderLike, country: string, query: PoolQuery): NamePools {
+	using db = new DatabaseClient<POIDatabase>(dbPath, { readOnly: true })
+
+	const codes = db.prepare("select id, category from poi_category_codes").all() as Array<{
+		id: number
+		category: string
+	}>
+
+	const byName = new Map(codes.map((c) => [c.category, c.id]))
+	const venueIDs = POI_VENUE_CATEGORIES.map((c) => byName.get(c)).filter((id): id is number => id != null)
+	const confoundIDs = POI_CONFOUND_CATEGORIES.map((c) => byName.get(c)).filter((id): id is number => 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) as Array<{ name: string; category_id: number }>
+
+	const venueSet = new Set(venueIDs)
+	const venues = new Set<string>()
+	const rejectedVenues = new Set<string>()
+	const longerNames = new Set<string>()
+
+	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: [],
+	}
+}
+
+/**
+ * Merge two sources' name pools.
+ */
+export function mergeNamePools(a: NamePools, b: NamePools): NamePools {
+	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: Readonly<Record<string, readonly LocalePart[]>> = {
+	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` slice is built from, so the two slices' address halves are drawn from one pool), DE and
+ * ES streamed out of OpenAddresses.
+ */
+export async function loadContextTuples(
+	tuplesPath: PathBuilderLike,
+	seed: number
+): Promise<Map<string, LocaleBaseTuple[]>> {
+	const byCountry = new Map<string, LocaleBaseTuple[]>()
+
+	for await (const tuple of readSliceTuples(tuplesPath)) {
+		const country = String(tuple.country ?? "")
+
+		if (!country || !tuple.locality || !tuple.street) continue
+
+		const mapped: LocaleBaseTuple = {
+			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: LocaleBaseTuple[] = []
+
+		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
+}
+
+/**
+ * The street-side confound classes, mined from the leg's OWN address tuples.
+ *
+ * Real streets, not invented ones. The 176,519-row context pool carries 195 GB `hall` streets, 114 GB `gate` streets,
+ * 134 distinct GB `-gate` single tokens and a two-figure `<modifier> <designator>` population in both GB and US — small
+ * absolute numbers, but every one of them a street somebody lives on, which is the property an invented list cannot
+ * have.
+ */
+export interface StreetNegatives {
+	designator: LocaleBaseTuple[]
+	modifierDesignator: LocaleBaseTuple[]
+	gateSuffix: LocaleBaseTuple[]
+}
+
+/**
+ * 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: readonly LocaleBaseTuple[],
+	designatorPhrases: readonly string[],
+	modifiers: readonly string[],
+	country: string
+): StreetNegatives {
+	const designatorSet = new Set(designatorPhrases)
+	const modifierSet = new Set(modifiers)
+	const designator: LocaleBaseTuple[] = []
+	const modifierDesignator: LocaleBaseTuple[] = []
+	const gateSuffix: LocaleBaseTuple[] = []
+
+	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