@mailwoman/corpus 9.2.0 → 9.3.0

package/lib/recipes/fr-order.ts

@@ -0,0 +1,245 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `fr-order` slice recipe — French reversed-order coverage (#560). Reads REAL OpenAddresses FR
+ *   tuples (`fr/countrywide.csv` from the cached zip), then for each row picks — by
+ *   `--reversed-fraction` (default 0.5) — whether to render in CANONICAL French order
+ *   (number-street, postcode-city) or one of four REVERSED / postcode-first variants the
+ *   v4.4.0→v0.5.0 regression exposed (the model misses house_number in every reversed one):
+ *
+ *   - A: "47110 Sainte-Livrade-sur-Lot, 69 Allée du Bugatel" (postcode city, HN street)
+ *   - B: "Sainte-Livrade-sur-Lot, 47110, 619 Impasse de la Rose" (city, postcode, HN street)
+ *   - C: "Sainte-Livrade-sur-Lot 59 bis Rue des Ecuries 47110" (city HN street postcode — NO commas)
+ *   - D: "47110, 6 rue de la république, Sainte-Livrade-sur-Lot" (postcode, HN street, city)
+ *
+ *   Sub-modes ride alongside order: `bis`/`ter`/`quater` ordinal suffixes in house_number, and
+ *   ALL-CAPS locality. `--golden` emits a held-out reversed-order eval slice with a different
+ *   seed.
+ *
+ *   The inline synthesis (the OA-CSV reader, the ordinal/all-caps tables, the canonical + reversed
+ *   renderers) is ported faithfully from scripts/build-fr-order-slice.mjs. This is a
+ *   `generate`-mode recipe that still reads REAL tuples off disk — `--count` bounds the OUTPUT, not
+ *   the input. The passed `random` (the framework LCG) is consumed in the exact call order the
+ *   legacy script used.
+ *
+ *   NOT ported (diagnostic-only, no effect on emitted bytes): the post-run `runSpanCheck` self-check
+ *   (it reads the finished file back with a separate PRNG and prints to stderr; the recipe's output
+ *   stream is still open during `run`), and the dead `renderReversed` helper (the legacy `main`
+ *   inlined the variant logic and never called it).
+ */
+
+import { dataRootPath } from "@mailwoman/core/data-root"
+import { isPresent } from "@mailwoman/core/objects"
+import type { ComponentTag } from "@mailwoman/core/types"
+
+import { stableSourceID } from "#adapters/utils"
+import { makeMulberry32, readOATuples, type CorpusRecipe } from "#recipes/scaffold"
+import { pick } from "#synthesizers/utils"
+import type { CanonicalRow } from "#types"
+import { alignRow } from "#utils"
+
+const SOURCE = { zip: dataRootPath("oa-cache", "fr__countrywide.zip"), csv: "fr/countrywide.csv" }
+
+/**
+ * Ordinal suffixes used in French house numbers (BAN corpus), to cover the "8 bis" sub-mode.
+ */
+const ORDINAL_SUFFIXES: readonly string[] = ["bis", "ter", "quater"]
+/**
+ * Probability that a row gets an ordinal suffix injected (matches the golden's ~10-15% rate).
+ */
+const ORDINAL_PROB = 0.12
+/**
+ * Probability that a locality renders ALL-CAPS (another sub-mode: "SAINTE-LIVRADE-SUR-LOT").
+ */
+const ALLCAPS_PROB = 0.1
+
+/**
+ * A real FR tuple read out of the cached OA zip.
+ */
+interface FrTuple {
+	house_number: string
+	street: string
+	locality: string
+	postcode: string
+}
+
+/**
+ * Stream FR tuples out of the cached OA zip. The countrywide extract is GB-scale, so this reads only as far as `limit`
+ * distinct tuples — the `break` closes the reader and releases the archive. Only keeps rows with a house_number (the
+ * slice's core signal) and a postcode (required for reversed-order rendering to be meaningful; it is also part of this
+ * recipe's dedup key).
+ */
+async function readTuples(limit: number): Promise<FrTuple[]> {
+	return readOATuples(SOURCE, {
+		limit,
+		requirePostcode: true,
+		dedupIncludesPostcode: true,
+		extra: (fields) => fields,
+	})
+}
+
+/**
+ * Optionally augment a house_number with a French ordinal suffix ("59 bis", "4 ter"). Appended with a space so it forms
+ * one multi-token house_number string that alignRow can still locate verbatim.
+ */
+function maybeAddOrdinal(random: () => number, house_number: string): string {
+	if (random() >= ORDINAL_PROB) return house_number
+	const suffix = pick(ORDINAL_SUFFIXES, random)
+
+	// Vary suffix case: "bis" (lower) vs "BIS" (upper) — a real-world split in the golden.
+	return `${house_number} ${random() < 0.5 ? suffix : suffix.toUpperCase()}`
+}
+
+/**
+ * Render a tuple in CANONICAL French order: "9 Rue de la Promenade, 01200 Villes".
+ */
+function renderCanonical(
+	hn: string,
+	street: string,
+	postcode: string,
+	locality: string
+): { raw: string; components: Partial<Record<ComponentTag, string>> } {
+	const raw = `${hn} ${street}, ${postcode} ${locality}`
+
+	return { raw, components: { house_number: hn, street, postcode, locality } }
+}
+
+// Reversed layouts, a quarter each: A postcode+city then HN+street; B city, postcode, HN+street; C run-together;
+// D postcode, HN+street, city.
+const REVERSED_VARIANT_A_CUTOFF = 0.25
+const REVERSED_VARIANT_B_CUTOFF = 0.5
+const REVERSED_VARIANT_C_CUTOFF = 0.75
+
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export const frOrderRecipe: CorpusRecipe = {
+	name: "fr-order",
+	description: "French reversed-order rows (#560): real OA FR tuples rendered canonical + 4 postcode-first variants",
+	mode: "generate",
+	options: [
+		{ flag: "--reversed-fraction <p>", description: "Fraction rendered reversed-order. Default 0.5" },
+		{ flag: "--golden", description: "Emit the held-out reversed-order eval slice" },
+	],
+	async run(opts, write) {
+		if (opts.count == null) throw new Error("fr-order recipe requires --count <N>")
+		const count = opts.count
+		// Legacy build-fr-order-slice.mjs seeded mulberry32 with the raw seed: `const random = mulberry32(opts.seed)`.
+		// (The omitted diagnostic runSpanCheck used a separate mulberry32(opts.seed + 1) — not part of generation.)
+		const random = makeMulberry32(opts.seed)
+		const source = opts.sourceName ?? "synth-fr-order"
+		const reversedFraction = opts.reversedFraction ?? 0.5
+
+		// Over-read from the CSV so the dedup + filter pass can fill `count` rows.
+		const poolLimit = Math.max(count * 8, 40_000)
+		const pool = await readTuples(poolLimit)
+
+		console.error(`  ${SOURCE.csv}: ${pool.length} unique tuples (capped read)`)
+
+		if (!pool.length) {
+			throw new Error(`No FR tuples found — is ${SOURCE.zip} present?`)
+		}
+
+		let emitted = 0
+		let skipped = 0
+		let guard = 0
+		const N = pool.length
+
+		while (emitted < count && guard++ < count * 10) {
+			const base = pool[Math.floor(random() * N)]!
+			const { street, postcode } = base
+			const locality = random() < ALLCAPS_PROB ? base.locality.toUpperCase() : base.locality
+			const house_number = maybeAddOrdinal(random, base.house_number)
+
+			// Pick canonical vs reversed by --reversed-fraction.
+			const isReversed = random() < reversedFraction
+
+			let rendered: { raw: string; components: Partial<Record<ComponentTag, string>> }
+
+			if (isReversed) {
+				const variantRoll = random()
+				let raw: string
+
+				if (variantRoll < REVERSED_VARIANT_A_CUTOFF) {
+					// Variant A: postcode+city as a unit, then HN+street
+					raw = `${postcode} ${locality}, ${house_number} ${street}`
+				} else if (variantRoll < REVERSED_VARIANT_B_CUTOFF) {
+					// Variant B: city, then postcode, then HN+street (comma-separated, postcode isolated)
+					raw = `${locality}, ${postcode}, ${house_number} ${street}`
+				} else if (variantRoll < REVERSED_VARIANT_C_CUTOFF) {
+					// Variant C: no commas — locality HN street postcode (the "run-together" format)
+					raw = `${locality} ${house_number} ${street} ${postcode}`
+				} else {
+					// Variant D: postcode, HN+street, city (reversed top-to-bottom)
+					raw = `${postcode}, ${house_number} ${street}, ${locality}`
+				}
+
+				rendered = { raw, components: { house_number, street, postcode, locality } }
+			} else {
+				rendered = renderCanonical(house_number, street, postcode, locality)
+			}
+
+			const { raw, components } = rendered
+
+			// Safety check: every component must appear verbatim in raw (alignment precondition).
+			const componentValues = Object.values(components).filter(isPresent)
+
+			if (!componentValues.every((v) => raw.includes(v))) {
+				skipped++
+
+				continue
+			}
+
+			// --golden: emit per-locale-f1 eval rows ({raw, components, country:"FR"}).
+			if (opts.golden) {
+				write(JSON.stringify({ raw, components, country: "FR" }) + "\n")
+
+				emitted++
+
+				continue
+			}
+
+			const sourceID = stableSourceID(source, {
+				street: components.street,
+				house_number: components.house_number,
+				locality: components.locality,
+				postcode: components.postcode,
+			})
+
+			const canonical: CanonicalRow = {
+				raw,
+				components,
+				country: "FR",
+				locale: "fr-FR",
+				source,
+				source_id: sourceID,
+				corpus_version: "0.5.0",
+				license: "OpenAddresses FR countrywide tuples, rendered canonical + reversed-order — see ingest SOURCE",
+			}
+
+			const aligned = alignRow(canonical)
+
+			if (aligned.kind !== "labeled" || !aligned.row) {
+				skipped++
+
+				continue
+			}
+
+			write(
+				JSON.stringify({
+					...aligned.row,
+					synth_method: "fr-order",
+					synth_order: isReversed ? "reversed" : "canonical",
+					synth_base_id: null,
+				}) + "\n"
+			)
+
+			emitted++
+		}
+
+		return { emitted, skipped }
+	},
+}

package/lib/recipes/german.ts

@@ -0,0 +1,359 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `german` slice recipe — German coverage rows from REAL OpenAddresses tuples (Berlin + Saxony,
+ *   cached zips). Each sampled tuple is rendered via {@link synthesizeGermanRow} in BOTH orders —
+ *   `--intl-fraction` (default 0.4) in international order (house-first / postcode-after-city), the
+ *   rest in idiomatic German order — then aligned to BIO. Generate-mode: it builds a tuple pool
+ *   from the cached zips, then draws `--count` rows from it with the passed `random` (so the emit
+ *   stream matches the legacy reservoir-sample loop). Ported from scripts/build-german-slice.mjs.
+ *
+ *   ORDER ROBUSTNESS (2026-06-06): mixing the two renderings stops a native-only slice from teaching
+ *   German order so well it reads the US/feed-order eval as a "collapse". See
+ *   docs/articles/evals/resolver-geo/2026-06-06-anchor-pilot.md (the order-artifact correction).
+ *
+ *   TWO REGISTERS THE OA TUPLES DO NOT CARRY (#1946). A comma-free single line — `Neusser Str. 12 Nippes
+ *   50733 Köln`, the dictation / one-field-form register — segments as ONE unit at stage 2, and with one
+ *   segment the placetype-pair prior never fires, so the model reads `Nippes` as a second street and the
+ *   one-value-per-tag projection deletes it. `--comma-free-fraction` renders that many native-order rows
+ *   with `" "` as the line separator. And OA rows have no district at all, so `--ortsteil-fraction` rows
+ *   borrow a WOF Ortsteil of the tuple's own locality as `dependent_locality` — 67,532 DE neighbourhoods
+ *   carry a locality ancestor in the admin database, `Köln-Nippes` among them. The German spelling is
+ *   recovered from the `names` table ({@link ortsteilSurface}): WOF's `spr.name` for DE neighbourhoods is
+ *   the ASCII-folded label (`Bocklemuend`), which no German ever types.
+ */
+
+import { dataRootPath } from "@mailwoman/core/data-root"
+import { pathExists } from "@mailwoman/core/fs/readers"
+import type { WOFDatabase } from "@mailwoman/resolver-wof-sqlite/schema"
+import { DatabaseClient } from "@mailwoman/sqlite/client"
+import type { PathBuilderLike } from "path-ts"
+
+import { stableSourceID } from "#adapters/utils"
+import { makeMulberry32, readZippedCSVRecords, type CorpusRecipe } from "#recipes/scaffold"
+import { synthesizeGermanRow, type LocaleBaseTuple } from "#synthesizers/german"
+import { alignRow } from "#utils"
+
+/**
+ * A German OA source (cached zip) + the Bundesland the file covers (OA's REGION column is empty for DE).
+ */
+interface GermanSource {
+	zip: PathBuilderLike
+	csv: string
+	region: string
+}
+
+/**
+ * `region` is the Bundesland the source covers. OA's REGION column is empty for DE, but the region is implied by the
+ * per-state file — the international order needs it for the "City, Region Postcode" tail (v0.9.3 / #327). berlin.csv →
+ * Berlin (a city-state, region==locality); sn/statewide → Sachsen.
+ */
+const SOURCES: GermanSource[] = [
+	{ zip: dataRootPath("oa-cache", "de__berlin.zip"), csv: "de/berlin.csv", region: "Berlin" },
+	{ zip: dataRootPath("oa-cache", "de__sn__statewide.zip"), csv: "de/sn/statewide.csv", region: "Sachsen" },
+]
+
+/**
+ * The two ASCII spellings WOF uses for one German label, so a `names.deu` row can be matched to the `spr.name` it
+ * spells. WOF folds `Bocklemünd` to `Bocklemuend` in one record and `Schöneberg` to `Schoneberg` in another — the
+ * transliteration (`ö` → `oe`, `ß` → `ss`) and the plain diacritic strip (`ö` → `o`) both occur — so a match is against
+ * either form, lower-cased.
+ */
+function foldGerman(surface: string): [transliterated: string, stripped: string] {
+	const lower = surface.toLowerCase()
+
+	const stripped = lower
+		.normalize("NFD")
+		.replaceAll(/[\u0300-\u036F]/g, "")
+		.replaceAll("ß", "ss")
+
+	const transliterated = lower
+		.replaceAll("ä", "ae")
+		.replaceAll("ö", "oe")
+		.replaceAll("ü", "ue")
+		.replaceAll("ß", "ss")
+		.normalize("NFD")
+		.replaceAll(/[\u0300-\u036F]/g, "")
+
+	return [transliterated, stripped]
+}
+
+/**
+ * Whether two surfaces spell the same German label under either of WOF's folds.
+ */
+function sameGermanLabel(left: string, right: string): boolean {
+	const [leftT, leftS] = foldGerman(left)
+	const [rightT, rightS] = foldGerman(right)
+
+	return leftT === rightT || leftS === rightS || leftT === rightS || leftS === rightT
+}
+
+/**
+ * The surface an Ortsteil is written with in an address, from WOF's rows for it.
+ *
+ * `spr.name` for a DE neighbourhood is the ASCII-folded label; the `names` rows in `deu` carry the German spelling
+ * beside unrelated labels for co-located features (`Bocklemuend` → `Bocklemünd`, `Jüdischer Friedhof Bocklemünd`,
+ * `Menara-Garten`). The German name is the one that spells the same label as `spr.name` under either of WOF's ASCII
+ * folds ({@link foldGerman}); with none, `spr.name` stands. WOF also prefixes some Ortsteile with their city
+ * (`Köln-Nippes`), a form no envelope carries once the city is its own line, so a leading `<locality>-` is dropped when
+ * something is left after it.
+ */
+export function ortsteilSurface(sprName: string, deuNames: readonly string[], locality: string): string {
+	const german = deuNames.find((name) => sameGermanLabel(name, sprName)) ?? sprName
+	const prefix = `${locality}-`
+
+	return german.toLowerCase().startsWith(prefix.toLowerCase()) && german.length > prefix.length
+		? german.slice(prefix.length)
+		: german
+}
+
+/**
+ * Every current DE neighbourhood with a locality ancestor, keyed by the folded locality name → Ortsteil surfaces. Empty
+ * when the admin database is not readable; the recipe then emits no Ortsteil rows and says so, rather than failing a
+ * build over an optional register.
+ */
+async function readOrtsteilPool(adminDB: string): Promise<Map<string, string[]>> {
+	const pool = new Map<string, string[]>()
+
+	if (!(await pathExists(adminDB))) return pool
+
+	using db = new DatabaseClient<WOFDatabase>(adminDB, { readOnly: true })
+
+	const rows = await db
+		.selectFrom("spr as s")
+		.innerJoin("ancestors as a", "a.id", "s.id")
+		.innerJoin("spr as p", "p.id", "a.ancestor_id")
+		.select(["s.id as id", "s.name as name", "p.name as locality"])
+		.where("s.country", "=", "DE")
+		.where("s.placetype", "=", "neighbourhood")
+		.where("s.is_current", "=", 1)
+		.where("s.is_deprecated", "=", 0)
+		.where("a.ancestor_placetype", "=", "locality")
+		.execute()
+
+	const deuNames = new Map<number, string[]>()
+
+	for (const row of await db
+		.selectFrom("names")
+		.select(["id", "name"])
+		.where("country", "=", "DE")
+		.where("placetype", "=", "neighbourhood")
+		.where("language", "=", "deu")
+		.execute()) {
+		const list = deuNames.get(row.id) ?? []
+
+		list.push(row.name)
+		deuNames.set(row.id, list)
+	}
+
+	for (const row of rows) {
+		if (!row.name || !row.locality) continue
+
+		const surface = ortsteilSurface(row.name, deuNames.get(row.id) ?? [], row.locality)
+		const key = row.locality.toLowerCase()
+		const list = pool.get(key) ?? []
+
+		if (!list.includes(surface)) {
+			list.push(surface)
+		}
+
+		pool.set(key, list)
+	}
+
+	return pool
+}
+
+/**
+ * Stream real German tuples out of a cached OA zip.
+ */
+async function readGermanTuples(source: GermanSource): Promise<LocaleBaseTuple[]> {
+	const tuples: LocaleBaseTuple[] = []
+	const seen = new Set<string>()
+
+	for await (const row of readZippedCSVRecords(source.zip, source.csv)) {
+		const street = row.street ?? ""
+		const locality = row.city ?? ""
+
+		if (!street || !locality) continue
+		const house_number = row.number ?? ""
+		const postcode = row.postcode ?? ""
+		// OA's REGION column is empty for DE — fall back to the source's Bundesland (set per file).
+		const region = row.region || source.region || ""
+		const key = `${house_number}|${street}|${locality}|${postcode}`.toLowerCase()
+
+		if (seen.has(key)) continue
+		seen.add(key)
+		tuples.push({ house_number, street, locality, region, postcode })
+	}
+
+	return tuples
+}
+
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export const germanRecipe: CorpusRecipe = {
+	name: "german",
+	description: "German coverage rows from real OA tuples (Berlin/Saxony), both orders → synthesizeGermanRow",
+	mode: "generate",
+	options: [
+		{ flag: "--intl-fraction <f>", description: "Fraction rendered international order. Default 0.4" },
+		{
+			flag: "--comma-free-fraction <f>",
+			description:
+				"Fraction of native-order rows rendered with no commas (the single-line register, #1946). Default 0.3",
+		},
+		{
+			flag: "--ortsteil-fraction <f>",
+			description: "Fraction of rows carrying a WOF Ortsteil of the locality as dependent_locality. Default 0.3",
+		},
+		{
+			flag: "--admin-db <path>",
+			description:
+				"WOF admin database for the Ortsteil pool. Default $MAILWOMAN_DATA_ROOT/wof/admin-global-priority-importance.db",
+		},
+	],
+	async run(opts, write) {
+		// Emit PRNG: the legacy build-german-slice.mjs seeded mulberry32(opts.seed).
+		const random = makeMulberry32(opts.seed)
+		const source = opts.sourceName ?? "synth-german"
+		const intlFraction = opts.intlFraction ?? 0.4
+		const commaFreeFraction = opts.commaFreeFraction ?? 0.3
+		const ortsteilFraction = opts.ortsteilFraction ?? 0.3
+
+		for (const [flag, value] of [
+			["--intl-fraction", intlFraction],
+			["--comma-free-fraction", commaFreeFraction],
+			["--ortsteil-fraction", ortsteilFraction],
+		] as const) {
+			if (!(value >= 0 && value <= 1)) {
+				throw new Error(`${flag} must be in [0, 1], got ${value}`)
+			}
+		}
+
+		const adminDB = opts.adminDB ?? String(dataRootPath("wof", "admin-global-priority-importance.db"))
+		const ortsteile = ortsteilFraction > 0 ? await readOrtsteilPool(adminDB) : new Map<string, string[]>()
+
+		if (ortsteilFraction > 0) {
+			console.error(
+				ortsteile.size
+					? `  Ortsteil pool: ${[...ortsteile.values()].reduce((n, list) => n + list.length, 0)} surfaces under ${ortsteile.size} localities (${adminDB})`
+					: `  Ortsteil pool: none — ${adminDB} is not readable, so no row carries a dependent_locality`
+			)
+		}
+
+		const count = opts.count ?? 4000
+
+		// Pool real tuples from every German source, then sample `count` rows from it.
+		const pool: LocaleBaseTuple[] = []
+
+		for (const s of SOURCES) {
+			const t = await readGermanTuples(s)
+
+			console.error(`  ${s.csv}: ${t.length} unique tuples`)
+
+			for (const x of t) {
+				pool.push(x)
+			} // NOT pool.push(...t) — spreading ~840K args overflows the stack
+		}
+
+		if (!pool.length) {
+			throw new Error(`No German tuples found — are the cached zips present in ${dataRootPath("oa-cache")}?`)
+		}
+
+		let emitted = 0
+		let skipped = 0
+		let guard = 0
+		let withOrtsteil = 0
+		let commaFree = 0
+		const N = pool.length
+
+		while (emitted < count && guard++ < count * 6) {
+			const drawn = pool[Math.floor(random() * N)]!
+			// Per-row order: `--intl-fraction` of rows render house-first / postcode-after-city (the US/feed
+			// layout), the rest in idiomatic German order. Same components either way.
+			const order = random() < intlFraction ? "international" : "native"
+			// The two registers OA never wrote (#1946), each drawn independently of the order so every combination
+			// occurs: an Ortsteil borrowed from the tuple's own locality, and a native line with no commas.
+			const localOrtsteile = ortsteile.get(drawn.locality.toLowerCase())
+
+			const ortsteil =
+				localOrtsteile && random() < ortsteilFraction
+					? localOrtsteile[Math.floor(random() * localOrtsteile.length)]
+					: undefined
+
+			const base: LocaleBaseTuple = ortsteil ? { ...drawn, dependent_locality: ortsteil } : drawn
+			const separator = order === "native" && random() < commaFreeFraction ? " " : ", "
+			const synth = synthesizeGermanRow(base, { random, order, separator })
+
+			if (!synth) {
+				skipped++
+
+				continue
+			}
+
+			// --golden: emit per-locale-f1 eval rows ({raw, components}) instead of aligned BIO. `order`
+			// rides along so the eval can stratify native vs international.
+			if (opts.golden) {
+				write(JSON.stringify({ raw: synth.raw, components: synth.components, country: "DE", order }) + "\n")
+
+				emitted++
+
+				continue
+			}
+
+			const sourceID = stableSourceID(source, {
+				street: synth.components.street,
+				house_number: synth.components.house_number,
+				locality: synth.components.locality,
+				postcode: synth.components.postcode,
+			})
+
+			const canonical = {
+				raw: synth.raw,
+				components: synth.components,
+				country: "DE",
+				locale: synth.locale,
+				source,
+				source_id: sourceID,
+				corpus_version: "0.4.0",
+				license: `OpenAddresses DE (Berlin/Saxony) tuples, rendered ${order}-order — see ingest SOURCES`,
+			}
+
+			const aligned = alignRow(canonical as Parameters<typeof alignRow>[0])
+
+			if (aligned.kind !== "labeled" || !aligned.row) {
+				skipped++
+
+				continue
+			}
+
+			write(
+				JSON.stringify({
+					...aligned.row,
+					synth_method: "german",
+					synth_order: order,
+					synth_separator: separator === " " ? "space" : "comma",
+					synth_base_id: null,
+				}) + "\n"
+			)
+
+			emitted++
+
+			if (ortsteil) {
+				withOrtsteil++
+			}
+
+			if (separator === " ") {
+				commaFree++
+			}
+		}
+
+		console.error(`  emitted ${emitted}: ${withOrtsteil} with an Ortsteil, ${commaFree} comma-free`)
+
+		return { emitted, skipped }
+	},
+}

package/lib/recipes/house-venue.ts

@@ -0,0 +1,92 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `house-venue` slice recipe — synthetic house_number+venue+street co-occurrence rows: tuples →
+ *   {@link synthesizeHouseVenueRow} → aligned LabeledRow. The v0.6.3 corrective companion to the
+ *   no-street slice: every row carries BOTH house_number AND venue, restoring the house_number
+ *   signal that no-street's distributional shift cost the model. Ported from
+ *   scripts/build-house-venue-slice.mjs.
+ */
+
+import {
+	alignAndWrite,
+	makeLcg,
+	readTuples,
+	sliceSourceID,
+	SYNTHETIC_TUPLE_LICENSE as LICENSE,
+	type CorpusRecipe,
+} from "#recipes/scaffold"
+import { synthesizeHouseVenueRow, type HouseVenueBaseTuple } from "#synthesizers/house-venue"
+
+/**
+ * Slice recipe registered with the corpus builder — see the file header for the parse behaviour it exists to exercise,
+ * and `description` below for the surface form it generates.
+ */
+export const houseVenueRecipe: CorpusRecipe = {
+	name: "house-venue",
+	description: "House_number+venue co-occurrence rows: tuples → synthesizeHouseVenueRow → aligned LabeledRow",
+	mode: "tuples",
+	async run(opts, write) {
+		if (!opts.input) throw new Error("house-venue recipe requires --input <tuples.jsonl>")
+		// Legacy build-house-venue-slice.mjs seeded the LCG via makeRandom(opts.seed) (s = seed).
+		const random = makeLcg(opts.seed)
+		const source = opts.sourceName ?? "synth-house-venue"
+		let read = 0
+		let emitted = 0
+		let skipped = 0
+
+		for await (const tuple of readTuples(opts.input)) {
+			read++
+
+			// FR renders without a region (postcode-before-locality tail — the run-2 contingency), so
+			// an empty region is valid there and stays required everywhere else.
+			if (!tuple.locality || !tuple.postcode || !tuple.country || (!tuple.region && tuple.country !== "FR")) {
+				skipped++
+
+				continue
+			}
+
+			for (let v = 0; v < opts.variants; v++) {
+				const synth = synthesizeHouseVenueRow(tuple as HouseVenueBaseTuple, { random })
+
+				if (!synth) {
+					skipped++
+
+					continue
+				}
+
+				const ok = alignAndWrite(
+					write,
+					{
+						raw: synth.raw,
+						components: synth.components,
+						country: tuple.country,
+						locale: synth.locale,
+						source,
+						source_id: sliceSourceID(source, {
+							locality: tuple.locality,
+							region: tuple.region,
+							postcode: tuple.postcode,
+							country: tuple.country,
+							template: synth.template,
+							v: String(v),
+						}),
+						corpus_version: "0.4.0",
+						license: LICENSE,
+					},
+					synth.template
+				)
+
+				if (ok) {
+					emitted++
+				} else {
+					skipped++
+				}
+			}
+		}
+
+		return { read, emitted, skipped }
+	},
+}