@mailwoman/corpus 9.2.0 → 9.4.0

package/lib/recipes/street-affix.ts

@@ -0,0 +1,805 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   `street-affix` slice recipe — the US street-affix coverage slice (the v0-parity `street_prefix` /
+ *   `street_suffix` gap — both ~0% F1 in the #15 assessment, collapsed into `street`). Raises
+ *   PREVALENCE of affix-split streets with format diversity so the model learns to split "N Main
+ *   St" → street_prefix="N" + street="Main" + street_suffix="St", and (negative space) sharpens
+ *   `street` itself. Ported from scripts/build-street-affix-slice.mjs.
+ *
+ *   Reads REAL US OpenAddresses tuples and SPLITS the OA `street` field via the codex:
+ *   `matchLeadingDirectional` (USPS Pub-28 C1) for the prefix, `matchTrailingSuffix` (Pub-28 C2
+ *   street suffixes) for the suffix. OA streets nearly all carry a suffix; only ~10-20% carry a
+ *   directional, so we INJECT a directional prefix onto a fraction of prefix-less streets to give
+ *   `street_prefix` real signal. Each row varies surface form per affix — abbreviated ("N", "St")
+ *   vs expanded ("North", "Street") — and varies the layout (full address / bare / street-only /
+ *   venue-prefixed).
+ *
+ *   LEAKAGE-SAFE EVAL (`--golden`): held-out eval uses the VERMONT source only (the corpus
+ *   defaultHoldout), a different seed, and emits {raw, components} for per-locale-f1. Train uses
+ *   every NON-Vermont US source.
+ *
+ *   Multi-locale BALANCE (`--multilocale-count`, opts.multilocaleCount > 0): appends NO-affix
+ *   native-order rows (FR/DE/IT/NL) AFTER the US affix rows, riding the same source weight, purely
+ *   to keep the postcode-ORDER distribution multi-locale so a US-heavy affix slice doesn't dilute
+ *   FR/DE postcode (the v0.9.8 blemish).
+ */
+
+import {
+	DirectionalAbbreviation,
+	lookupDirectional,
+	matchCase,
+	matchLeadingDirectional,
+	matchTrailingSuffix,
+	renderDirectional,
+	US_STREET_SUFFIX_PREFERRED_ABBR,
+	NAME_PRONE_US_SUFFIXES,
+} from "@mailwoman/codex/us"
+import { dataRootPath } from "@mailwoman/core/data-root"
+import type { ComponentTag } from "@mailwoman/core/types"
+import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils"
+import type { PathBuilderLike } from "path-ts"
+
+import { stableSourceID } from "#adapters/utils"
+import { readCSVRecords, readOATuples, sliceSourceID, type CorpusRecipe } from "#recipes/scaffold"
+import { pick } from "#synthesizers/utils"
+import type { CanonicalRow } from "#types"
+import { alignRow } from "#utils"
+
+// Same OA cache as the unit slice. Train = every NON-Vermont state; eval = Vermont (the holdout).
+
+interface USSource {
+	zip: PathBuilderLike
+	csv: string
+	region: string
+}
+
+const TRAIN_SOURCES: readonly USSource[] = [
+	{ zip: dataRootPath("oa-cache", "us__ca__berkeley.zip"), csv: "us/ca/berkeley.csv", region: "CA" },
+	{ zip: dataRootPath("oa-cache", "us__ca__marin.zip"), csv: "us/ca/marin.csv", region: "CA" },
+	{ zip: dataRootPath("oa-cache", "us__dc__statewide.zip"), csv: "us/dc/statewide.csv", region: "DC" },
+	{ zip: dataRootPath("oa-cache", "us__ia__statewide.zip"), csv: "us/ia/statewide.csv", region: "IA" },
+	{ zip: dataRootPath("oa-cache", "us__il__cook.zip"), csv: "us/il/cook.csv", region: "IL" },
+	{ zip: dataRootPath("oa-cache", "us__mt__statewide.zip"), csv: "us/mt/statewide.csv", region: "MT" },
+	{ zip: dataRootPath("oa-cache", "us__sd__statewide.zip"), csv: "us/sd/statewide.csv", region: "SD" },
+]
+
+const EVAL_SOURCE: USSource = {
+	zip: dataRootPath("oa-cache", "us__vt__statewide.zip"),
+	csv: "us/vt/statewide.csv",
+	region: "VT",
+}
+
+// Multi-locale BALANCE sources (--multilocale-count > 0). These rows carry NO affix split — they exist
+// only to keep the postcode-ORDER distribution multi-locale. Native-order rendering mirrors
+// build-country-slice-balanced.mjs: FR = number-street, postcode-city; DE/IT/NL = street-number,
+// postcode-city. `order` drives the body.
+interface BalanceSource {
+	zip: PathBuilderLike
+	csv: string
+	iso2: string
+	region: string
+	order: string
+}
+
+const MULTILOCALE_SOURCES: readonly BalanceSource[] = [
+	{
+		zip: dataRootPath("oa-cache", "de__sn__statewide.zip"),
+		csv: "de/sn/statewide.csv",
+		iso2: "DE",
+		region: "",
+		order: "eu",
+	},
+	{
+		zip: dataRootPath("oa-cache", "fr__countrywide.zip"),
+		csv: "fr/countrywide.csv",
+		iso2: "FR",
+		region: "",
+		order: "fr",
+	},
+	{
+		zip: dataRootPath("oa-cache", "it__countrywide.zip"),
+		csv: "it/countrywide.csv",
+		iso2: "IT",
+		region: "",
+		order: "eu",
+	},
+	{
+		zip: dataRootPath("oa-cache", "nl__countrywide.zip"),
+		csv: "nl/countrywide.csv",
+		iso2: "NL",
+		region: "",
+		order: "eu",
+	},
+]
+
+const MULTILOCALE_EVAL_SOURCES: readonly BalanceSource[] = [
+	{ zip: dataRootPath("oa-cache", "de__berlin.zip"), csv: "de/berlin.csv", iso2: "DE", region: "", order: "eu" },
+]
+
+/**
+ * ["N","E","S","W","NE","NW","SE","SW"].
+ */
+const DIRECTIONAL_ABBRS = Object.values(DirectionalAbbreviation)
+/**
+ * Fraction of prefix-less streets that get a synthetic directional.
+ */
+const INJECT_PREFIX_PROB = 0.3
+
+/**
+ * A real US skeleton tuple read from a cached OA zip.
+ */
+interface USTuple {
+	house_number: string
+	street: string
+	locality: string
+	region: string
+	postcode: string
+	base_source_id: string
+}
+
+/**
+ * A non-US BALANCE tuple (carries a postcode + native order).
+ */
+interface BalanceTuple {
+	house_number: string
+	street: string
+	locality: string
+	region: string
+	postcode: string
+	iso2: string
+	order: string
+}
+
+/**
+ * Prefix carried through render — the (canonical, abbreviation) pair `renderDirectional` consumes.
+ */
+type Prefix = Pick<NonNullable<ReturnType<typeof matchLeadingDirectional>>, "canonical" | "abbreviation">
+
+/**
+ * Stream real US tuples (number/street/city/postcode) out of a cached OA zip.
+ */
+async function readTuples(source: USSource): Promise<USTuple[]> {
+	return readOATuples(source, {
+		extra: (fields, _row, key) => ({
+			...fields,
+			region: source.region,
+			base_source_id: `openaddresses:${source.csv}:${key}`,
+		}),
+	})
+}
+
+const title = (s: string): string =>
+	s
+		.toLowerCase()
+		.split(/\s+/)
+		.map((w) => (w ? w.charAt(0).toUpperCase() + w.slice(1) : w))
+		.join(" ")
+
+const isSuffixOrDirectional = (word: string): boolean =>
+	matchTrailingSuffix(word) !== null || matchLeadingDirectional(word) !== null
+
+/**
+ * Split an OA street into { prefix?, name, suffix } using the codex. Requires a trailing suffix and a non-empty name
+ * that isn't itself an affix token. Returns null when the street has no usable suffix.
+ */
+function parseStreet(
+	street: string,
+	opts: { allowNameProneTail?: boolean } = {}
+): { prefix: Prefix | null; name: string; suffix: string } | null {
+	let words = street.trim().split(/\s+/)
+
+	if (words.length < 2) return null
+	let prefix: Prefix | null = null
+	// Leading directional — only if it leaves ≥2 words behind (room for a name + suffix).
+	const lead = matchLeadingDirectional(street)
+
+	if (lead && words.length > 2) {
+		prefix = { canonical: lead.canonical, abbreviation: lead.abbreviation }
+		words = words.slice(1)
+	}
+
+	// Trailing USPS suffix — only if it leaves ≥1 word for the name.
+	const trail = matchTrailingSuffix(words.join(" "))
+
+	if (!trail || words.length < 2) return null
+	const suffix = trail.canonical
+	const name = words.slice(0, -1).join(" ")
+
+	if (!name) return null
+
+	if (isSuffixOrDirectional(name)) {
+		const nameTail = matchTrailingSuffix(name)
+
+		if (!opts.allowNameProneTail || !nameTail || !NAME_PRONE_US_SUFFIXES.has(nameTail.canonical)) return null
+	}
+
+	return { prefix, name, suffix }
+}
+
+export type SuffixBoundaryClass = "terminal-only" | "terminal-contrast"
+
+// A terminal-only surface needs a name word before the penultimate suffix (`Blue Hill Rd`): three words at least.
+const TERMINAL_ONLY_MIN_WORDS = 3
+
+/**
+ * Classify a real street surface for #1569. `terminal-only` has an ambiguous suffix-eligible name word immediately
+ * before a different terminal type (`Blue Hill Rd`); `terminal-contrast` ends at the ambiguous word itself (`Sutton
+ * Hollow`). The contrast check intentionally wins when both final words are name-prone: only the LAST token is the
+ * suffix under the canonical rule.
+ */
+export function classifySuffixBoundaryStreet(street: string): SuffixBoundaryClass | null {
+	const words = street.trim().split(/\s+/)
+
+	if (words.length < 2) return null
+	const terminal = matchTrailingSuffix(words.at(-1)!)
+
+	if (!terminal) return null
+
+	if (NAME_PRONE_US_SUFFIXES.has(terminal.canonical)) return "terminal-contrast"
+
+	if (words.length < TERMINAL_ONLY_MIN_WORDS) return null
+
+	const penultimate = matchTrailingSuffix(words.at(-2)!)
+
+	return penultimate && NAME_PRONE_US_SUFFIXES.has(penultimate.canonical) ? "terminal-only" : null
+}
+
+/**
+ * Render the affix-split street in random surface forms (abbrev vs expanded per affix), Title-cased.
+ */
+function renderStreet(
+	random: () => number,
+	parsed: { prefix: Prefix | null; name: string; suffix: string }
+): { street: string; components: Partial<Record<ComponentTag, string>> } {
+	const name = title(parsed.name)
+	const parts: string[] = []
+	const components: Partial<Record<ComponentTag, string>> = { street: name }
+
+	// Prefix: natural (from parse) or injected onto a prefix-less street to boost street_prefix signal.
+	let prefix = parsed.prefix
+
+	if (!prefix && random() < INJECT_PREFIX_PROB) {
+		const m = lookupDirectional(pick(DIRECTIONAL_ABBRS, random))!
+		prefix = { canonical: m.directional, abbreviation: m.abbreviation }
+	}
+
+	if (prefix) {
+		const rendered = renderDirectional(prefix, random() < 0.5 ? "abbr" : "full", "Aa") // "Aa" → Title-case
+		components.street_prefix = rendered
+		parts.push(rendered)
+	}
+
+	parts.push(name)
+
+	// Suffix: abbreviated ("St") vs expanded ("Street"), Title-cased to match the name.
+	const full = title(parsed.suffix)
+
+	// canonical is uppercase word → "Street"
+	const abbr = matchCase(
+		US_STREET_SUFFIX_PREFERRED_ABBR[parsed.suffix as keyof typeof US_STREET_SUFFIX_PREFERRED_ABBR],
+		"Aa"
+	)
+
+	// "AVE" → "Ave"
+	const renderedSuffix = random() < 0.5 ? abbr : full
+	components.street_suffix = renderedSuffix
+	parts.push(renderedSuffix)
+
+	return { street: parts.join(" "), components }
+}
+
+/**
+ * Synthetic recipient/venue prefixes — the arena's "JOHN DOE, ACME INC, …" pattern.
+ */
+const VENUES = ["John Doe", "Jane Smith", "Acme Inc", "Wayne Enterprises", "Maria Garcia", "Riverside Clinic"]
+
+/**
+ * Real-venue pool for the suffix-boundary v2 venue shell (corpus 0.19.0). The 2026-08-10 recipe review found the
+ * v0.18.x slice's venue-led rows detectable as templates (six fixed venue strings), and the frozen B1 board's failures
+ * concentrate in exactly that shell (rich rows 5/43 vs bare 53/65 at v4.3.3 step-40k). HRSA's health-center site file
+ * supplies thousands of REAL US facility names ('Alburg Health Center'-register), US-government public domain, already
+ * a corpus source (`usgov-hrsa-fqhc`). Kept verbatim (including the ~11% all-caps names — real register diversity);
+ * comma-carrying names are dropped because the venue layout uses commas as its field delimiter.
+ */
+const VENUE_POOL_CSV = dataRootPath(
+	"corpus",
+	"sources",
+	"usgov-hrsa-fqhc",
+	"Health_Center_Service_Delivery_and_LookAlike_Sites.csv"
+)
+
+// Facility-name length window: shorter reads as an initialism, longer as a sentence.
+const VENUE_NAME_MIN_LENGTH = 3
+const VENUE_NAME_MAX_LENGTH = 60
+
+async function readVenuePool(csvPath: PathBuilderLike): Promise<string[]> {
+	const seen = new Set<string>()
+	const pool: string[] = []
+
+	for await (const record of readCSVRecords(csvPath)) {
+		const name = (record["site name"] ?? record["Site Name"] ?? "").trim().replaceAll(/\s+/gu, " ")
+
+		if (
+			name.length < VENUE_NAME_MIN_LENGTH ||
+			name.length > VENUE_NAME_MAX_LENGTH ||
+			name.includes(",") ||
+			!/\p{L}/u.test(name)
+		)
+			continue
+
+		const key = name.toLowerCase()
+
+		if (seen.has(key)) continue
+		seen.add(key)
+		pool.push(name)
+	}
+
+	return pool
+}
+
+const tail = (loc: string, reg: string, pc: string): string => (pc ? `${loc}, ${reg} ${pc}` : `${loc}, ${reg}`)
+
+/**
+ * Layout-shell options for {@link renderRow}. `cutoffs` are the cumulative random() boundaries for [full, bare,
+ * street-only] — the remainder is the venue shell. Defaults reproduce the original street-affix distribution
+ * (40/25/20/15) with the six template venues.
+ */
+interface RenderRowOpts {
+	venues?: readonly string[]
+	cutoffs?: readonly [number, number, number]
+}
+
+/**
+ * Embed the rendered street in a RANDOM layout so the model recognizes affixes wherever the street sits: full address,
+ * bare house+street, street-only (pure affix parse), or venue-prefixed.
+ */
+export function renderRow(
+	random: () => number,
+	base: USTuple,
+	street: string,
+	streetComponents: Partial<Record<ComponentTag, string>>,
+	opts: RenderRowOpts = {}
+): { fmt: string; raw: string; components: Partial<Record<ComponentTag, string>> } {
+	const venues = opts.venues ?? VENUES
+	const [fullCutoff, bareCutoff, streetOnlyCutoff] = opts.cutoffs ?? [0.4, 0.65, 0.85]
+
+	const hn = base.house_number,
+		loc = base.locality,
+		reg = base.region,
+		pc = base.postcode
+
+	const road = `${hn} ${street}`
+	const withRoad: Partial<Record<ComponentTag, string>> = { house_number: hn, ...streetComponents }
+	const r = random()
+
+	if (r < fullCutoff)
+		return {
+			fmt: "full",
+			raw: `${road}, ${tail(loc, reg, pc)}`,
+			components: { ...withRoad, locality: loc, region: reg, ...(pc ? { postcode: pc } : {}) },
+		}
+
+	if (r < bareCutoff) return { fmt: "bare", raw: road, components: withRoad }
+
+	if (r < streetOnlyCutoff) return { fmt: "street-only", raw: street, components: { ...streetComponents } }
+	const v = pick(venues, random)
+
+	return {
+		fmt: "venue",
+		raw: `${v}, ${road}, ${tail(loc, reg, pc)}`,
+		components: { venue: v, ...withRoad, locality: loc, region: reg, ...(pc ? { postcode: pc } : {}) },
+	}
+}
+
+/**
+ * Capped reader for the multi-locale BALANCE sources. The FR/IT/NL countrywide extracts are GB-scale, so this reads
+ * only as far as `limit` distinct tuples — the `break` closes the reader and releases the archive. Only keeps tuples
+ * that carry a POSTCODE.
+ */
+async function readBalanceTuples(source: BalanceSource, limit: number): Promise<BalanceTuple[]> {
+	return readOATuples(source, {
+		limit,
+		requirePostcode: true,
+		extra: (fields, row) => ({
+			...fields,
+			region: row.region || source.region,
+			iso2: source.iso2,
+			order: source.order,
+		}),
+	})
+}
+
+/**
+ * Render a non-US BALANCE row in native order — NO affix split, NO country token. `street` is the OA value verbatim.
+ * The sole job is to put a postcode in its native position so the slice doesn't pull the model US-ward.
+ */
+function renderBalanceRow(t: BalanceTuple): { raw: string; components: Partial<Record<ComponentTag, string>> } {
+	const { house_number: hn, street, locality: loc, postcode: pc, order } = t
+	// region is intentionally omitted — it isn't rendered in `raw`, so labeling it would fail alignment.
+	const components: Partial<Record<ComponentTag, string>> = { house_number: hn, street, locality: loc, postcode: pc }
+
+	const raw =
+		order === "fr"
+			? `${hn} ${street}, ${pc} ${loc}` // French: number-street, postcode-city
+			: `${street} ${hn}, ${pc} ${loc}`
+
+	// DE/IT/NL: street-number, postcode-city
+	return { raw, components }
+}
+
+/**
+ * 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 streetAffixRecipe: CorpusRecipe = {
+	name: "street-affix",
+	description: "US street-affix rows: OA streets split into street_prefix/street/street_suffix (+ multilocale balance)",
+	mode: "generate",
+	options: [
+		{
+			flag: "--multilocale-count <N>",
+			description: "Append N no-affix native-order balance rows (FR/DE/IT/NL). Default 0",
+		},
+	],
+	async run(opts, write) {
+		// Legacy build-street-affix-slice.mjs seeded `mulberry32(opts.seed)`.
+		const random = makeMulberry32(opts.seed)
+		const count = opts.count ?? 50_000
+		const source = opts.sourceName ?? "synth-affix"
+		const multilocaleCount = opts.multilocaleCount ?? 0
+		const sources = opts.golden ? [EVAL_SOURCE] : TRAIN_SOURCES
+
+		const pool: USTuple[] = []
+
+		for (const s of sources) {
+			const t = await readTuples(s)
+
+			console.error(`  ${s.csv}: ${t.length} unique tuples`)
+
+			for (const x of t) {
+				pool.push(x)
+			}
+		}
+
+		if (!pool.length) {
+			throw new Error(`No US tuples found — are the cached OA zips present in ${dataRootPath("oa-cache")}?`)
+		}
+
+		let emitted = 0
+		let skipped = 0
+		let noAffix = 0
+		let guard = 0
+		const formatCounts: Record<string, number> = {}
+		const affixCounts = { prefix: 0, suffix: 0, both: 0 }
+		const N = pool.length
+
+		while (emitted < count && guard++ < count * 10) {
+			const base = pool[Math.floor(random() * N)]!
+			const parsed = parseStreet(base.street)
+
+			if (!parsed) {
+				noAffix++
+
+				continue
+			}
+
+			const { street, components: streetComponents } = renderStreet(random, parsed)
+			const { fmt, raw, components } = renderRow(random, base, street, streetComponents)
+
+			// Every affix surface form must survive verbatim in raw, else alignment can't label it.
+			const surfaces = [streetComponents.street_prefix, streetComponents.street, streetComponents.street_suffix].filter(
+				(s): s is string => Boolean(s)
+			)
+
+			if (!surfaces.every((s) => raw.includes(s))) {
+				skipped++
+
+				continue
+			}
+
+			formatCounts[fmt] = (formatCounts[fmt] ?? 0) + 1
+			const hasP = !!streetComponents.street_prefix
+
+			if (hasP && streetComponents.street_suffix) {
+				affixCounts.both++
+			} else if (hasP) {
+				affixCounts.prefix++
+			} else {
+				affixCounts.suffix++
+			}
+
+			if (opts.golden) {
+				write(JSON.stringify({ raw, components, country: "US" }) + "\n")
+
+				emitted++
+
+				continue
+			}
+
+			const canonical: CanonicalRow = {
+				raw,
+				components,
+				country: "US",
+				locale: "en-US",
+				source,
+				source_id: stableSourceID(source, components),
+				corpus_version: "0.4.0",
+				license: "OpenAddresses US (non-VT) skeletons, street split via USPS Pub-28 C1/C2 (codex)",
+			}
+
+			const aligned = alignRow(canonical)
+
+			if (aligned.kind !== "labeled" || !aligned.row) {
+				skipped++
+
+				continue
+			}
+
+			write(JSON.stringify({ ...aligned.row, synth_method: "affix", synth_base_id: null }) + "\n")
+
+			emitted++
+		}
+
+		// ── Multi-locale balance rows (--multilocale-count) ─────────────────────────────────────────────
+		// Appended AFTER the US affix rows so the US affix signal is unchanged (same `--count`), and the
+		// non-US rows ride the SAME source weight. Native-order postcodes, no affix labels.
+		let balanceEmitted = 0
+		let balanceSkipped = 0
+		const balanceISO: Record<string, number> = {}
+
+		if (multilocaleCount > 0) {
+			const mlSources = opts.golden ? MULTILOCALE_EVAL_SOURCES : MULTILOCALE_SOURCES
+			const perSource = Math.ceil((multilocaleCount * 3) / mlSources.length) // over-read; balance locales
+			const mlPool: BalanceTuple[] = []
+
+			for (const s of mlSources) {
+				const t = await readBalanceTuples(s, perSource)
+
+				console.error(`  balance ${s.csv} (${s.iso2}): ${t.length} tuples`)
+
+				for (const x of t) {
+					mlPool.push(x)
+				}
+			}
+
+			const M = mlPool.length
+			let mlGuard = 0
+
+			// oxlint-disable-next-line eslint/no-unmodified-loop-condition -- `M > 0` is an invariant guard against an empty pool, not a progress condition
+			while (M > 0 && balanceEmitted < multilocaleCount && mlGuard++ < multilocaleCount * 10) {
+				const t = mlPool[Math.floor(random() * M)]!
+				const { raw, components } = renderBalanceRow(t)
+
+				// Every component surface must survive in raw, else alignment can't label it.
+				if (![components.street, components.locality, components.postcode].every((s) => !!s && raw.includes(s))) {
+					balanceSkipped++
+
+					continue
+				}
+
+				balanceISO[t.iso2] = (balanceISO[t.iso2] ?? 0) + 1
+				const locale = `${t.iso2.toLowerCase()}-${t.iso2}`
+
+				if (opts.golden) {
+					write(JSON.stringify({ raw, components, country: t.iso2 }) + "\n")
+
+					balanceEmitted++
+
+					continue
+				}
+
+				const canonical: CanonicalRow = {
+					raw,
+					components,
+					country: t.iso2,
+					locale,
+					source,
+					source_id: stableSourceID(source, components),
+					corpus_version: "0.4.0",
+					license: "OpenAddresses non-US skeletons (native-order postcode balance for the affix slice)",
+				}
+
+				const aligned = alignRow(canonical)
+
+				if (aligned.kind !== "labeled" || !aligned.row) {
+					balanceSkipped++
+
+					continue
+				}
+
+				write(JSON.stringify({ ...aligned.row, synth_method: "affix-balance", synth_base_id: null }) + "\n")
+
+				balanceEmitted++
+			}
+		}
+
+		console.error(
+			`Done: emitted ${emitted} affix rows, skipped ${skipped}, no-affix ${noAffix} (pool ${pool.length}).\n` +
+				`  formats: ${JSON.stringify(formatCounts)}\n` +
+				`  affix mix: ${JSON.stringify(affixCounts)}` +
+				(multilocaleCount > 0
+					? `\n  balance: emitted ${balanceEmitted}, skipped ${balanceSkipped}, iso ${JSON.stringify(balanceISO)}`
+					: "")
+		)
+
+		return { emitted: emitted + balanceEmitted, skipped: skipped + balanceSkipped }
+	},
+}
+
+// The venue shell needs a real register: fewer names than this and the rows read as templates again.
+const VENUE_POOL_MIN_SIZE = 500
+// While both classes are open: 80% terminal-only, 20% terminal-contrast.
+const TERMINAL_ONLY_SHARE = 0.8
+
+/**
+ * #1569 root-fix slice. Both classes come from real non-Vermont OA streets and use the affix recipe's existing layout
+ * diversity. v4.3.1 makes terminal-only 80% of the mix: the first 40/60 run moved a 100-row TRAIN sample only 4→11
+ * while contrast was already 95/100 before training (93/100 after). Post-run audit found that the global affix relabel
+ * pass corrupts many already-decomposed target rows into double suffixes; do not retrain this recipe until relabel is
+ * idempotent over a decomposed street family. The 20% contrast leg remains explicit, additive to the already-strong
+ * base distribution, and B2 still checks it unchanged.
+ *
+ * V2 (corpus 0.19.0, 2026-08-10 recipe review): the v4.3.3 board split (rich venue-led rows 5/43 vs bare 53/65) showed
+ * the model separating template rows from real ones, and the venue shell was the giveaway — six fixed venue strings. v2
+ * draws the venue shell from thousands of REAL HRSA facility names and raises its share (venue 30%, full 35%, bare 20%,
+ * street-only 15%). Dose policy moved to the recipe's config side: weight ≤4 effective passes per run (Muennighoff
+ * 2023's repetition knee) and the source is excluded from the augmentation pool — see the v4.4.0 config.
+ */
+export const suffixBoundaryRecipe: CorpusRecipe = {
+	name: "suffix-boundary",
+	description: "US #1569 suffix boundary: OA terminal-only positives + over-sampled terminal contrasts",
+	mode: "generate",
+	async run(opts, write) {
+		const random = makeMulberry32(opts.seed)
+		const count = opts.count ?? 30_000
+		const source = opts.sourceName ?? "synth-suffix-boundary"
+		const sources = opts.golden ? [EVAL_SOURCE] : TRAIN_SOURCES
+
+		// v2 venue shell: real facility names, loud when the source file is missing or thin.
+		const venuePool = await readVenuePool(VENUE_POOL_CSV)
+
+		if (venuePool.length < VENUE_POOL_MIN_SIZE) {
+			throw new Error(
+				`suffix-boundary v2 venue pool too thin: ${venuePool.length} names from ${VENUE_POOL_CSV} — ` +
+					"is the usgov-hrsa-fqhc source present?"
+			)
+		}
+
+		const shell: RenderRowOpts = { venues: venuePool, cutoffs: [0.35, 0.55, 0.7] }
+
+		const pool: Record<SuffixBoundaryClass, USTuple[]> = {
+			"terminal-only": [],
+			"terminal-contrast": [],
+		}
+
+		for (const s of sources) {
+			for (const tuple of await readTuples(s)) {
+				const rowClass = classifySuffixBoundaryStreet(tuple.street)
+
+				if (rowClass) {
+					pool[rowClass].push(tuple)
+				}
+			}
+		}
+
+		if (!pool["terminal-only"].length || !pool["terminal-contrast"].length) {
+			throw new Error(
+				`Suffix-boundary source class missing: terminal-only=${pool["terminal-only"].length}, ` +
+					`terminal-contrast=${pool["terminal-contrast"].length}`
+			)
+		}
+
+		let emitted = 0
+		let skipped = 0
+		const classCounts: Record<SuffixBoundaryClass, number> = { "terminal-only": 0, "terminal-contrast": 0 }
+		const terminalOnlyTarget = Math.round(count * 0.8)
+
+		const classTargets: Record<SuffixBoundaryClass, number> = {
+			"terminal-only": terminalOnlyTarget,
+			"terminal-contrast": count - terminalOnlyTarget,
+		}
+
+		const formatCounts: Record<string, number> = {}
+		let stemPairs = 0
+
+		const emitOne = (
+			rowClass: SuffixBoundaryClass,
+			base: USTuple,
+			parsed: NonNullable<ReturnType<typeof parseStreet>>,
+			method: string
+		): boolean => {
+			const { street, components: streetComponents } = renderStreet(random, parsed)
+			const { fmt, raw, components } = renderRow(random, base, street, streetComponents, shell)
+
+			const canonical: CanonicalRow = {
+				raw,
+				components,
+				country: "US",
+				locale: "en-US",
+				source,
+				source_id: sliceSourceID(source, { ...components, class: rowClass, n: String(emitted) }),
+				corpus_version: "0.19.0",
+				license:
+					"OpenAddresses US (non-VT) skeletons; terminal suffix labels via USPS Pub-28 codex; venue shell from HRSA site names (US public domain)",
+			}
+
+			const aligned = alignRow(canonical)
+
+			if (aligned.kind !== "labeled" || !aligned.row) {
+				skipped++
+
+				return false
+			}
+
+			write(JSON.stringify({ ...aligned.row, synth_method: method, synth_base_id: base.base_source_id }) + "\n")
+
+			classCounts[rowClass]++
+			formatCounts[fmt] = (formatCounts[fmt] ?? 0) + 1
+
+			emitted++
+
+			return true
+		}
+
+		// oxlint-disable-next-line eslint/no-unmodified-loop-condition -- emitted/skipped advance inside emitOne (a closure); the rule cannot see through the call
+		while (emitted < count && emitted + skipped < count * 10) {
+			const terminalOnlyOpen = classCounts["terminal-only"] < classTargets["terminal-only"]
+			const terminalContrastOpen = classCounts["terminal-contrast"] < classTargets["terminal-contrast"]
+
+			const rowClass: SuffixBoundaryClass =
+				terminalOnlyOpen && terminalContrastOpen
+					? random() < TERMINAL_ONLY_SHARE
+						? "terminal-only"
+						: "terminal-contrast"
+					: terminalContrastOpen
+						? "terminal-contrast"
+						: "terminal-only"
+
+			const classPool = pool[rowClass]
+			const base = pick(classPool, random)
+			const parsed = parseStreet(base.street, { allowNameProneTail: rowClass === "terminal-only" })
+
+			if (!parsed) {
+				skipped++
+
+				continue
+			}
+
+			if (!emitOne(rowClass, base, parsed, `suffix-boundary:${rowClass}`)) continue
+
+			// Stem-pair hard negative (v2, 2026-08-10 recipe review): after a terminal-only row
+			// ('Menlo Park' + 'Road'), also emit the SAME stem without its true suffix as a
+			// contrast row ('Menlo' + 'Park' under the canonical last-token rule). Sharing the stem
+			// forces the model to key on the licensing evidence — the trailing true suffix — rather
+			// than on name identity. Dose-neutral: these fill the existing 20% contrast target.
+			if (
+				rowClass === "terminal-only" &&
+				classCounts["terminal-contrast"] < classTargets["terminal-contrast"] &&
+				emitted < count &&
+				random() < 0.5
+			) {
+				const stemParsed = parseStreet(parsed.name)
+
+				if (
+					stemParsed &&
+					emitOne("terminal-contrast", base, stemParsed, "suffix-boundary:terminal-contrast:stem-pair")
+				) {
+					stemPairs++
+				}
+			}
+		}
+
+		console.error(
+			`Done: emitted ${emitted}, skipped ${skipped}; source pools ${JSON.stringify({
+				terminalOnly: pool["terminal-only"].length,
+				terminalContrast: pool["terminal-contrast"].length,
+			})}; classes ${JSON.stringify(classCounts)}; stem pairs ${stemPairs}; ` +
+				`venue pool ${venuePool.length}; formats ${JSON.stringify(formatCounts)}`
+		)
+
+		return { emitted, skipped }
+	},
+}