@mailwoman/corpus 10.5.0 → 10.7.0

package/lib/recipes/po/box/cedex/recipe.ts

@@ -20,8 +20,6 @@ import {
 	readFrTuples,
 	readPostalTuples,
 	readUsTuples,
-	US_EVAL_SOURCE,
-	US_TRAIN_SOURCES,
 } from "#recipes/po/box/cedex/sources"
 import type { Rendered, USTuple } from "#recipes/po/box/cedex/types"
 import { renderPmbUs, renderPoBoxUs } from "#recipes/po/box/cedex/us"
@@ -30,6 +28,11 @@ import { recipeSourceID, type CanonicalRecipeRow, type CorpusRecipe } from "#rec
 import { defaultRecipeSource } from "#recipes/sources"
 import { synthesizeMilitaryPoBoxRow } from "#synthesizers/po-box"
 import { SurfaceOrigin } from "#types"
+import {
+	NAD_SKELETON_DIRECTORY,
+	US_SKELETON_EVAL_SOURCE,
+	US_SKELETON_TRAIN_SOURCES,
+} from "#us/recipes/skeleton-sources"
 import { alignRow } from "#utils"
 
 const COMPONENT_ORDER = ["house_number", "street", "po_box", "venue", "locality", "postcode", "region", "cedex"]
@@ -63,7 +66,7 @@ export const poBoxCedexRecipe: CorpusRecipe = {
 		const source = opts.sourceName ?? defaultRecipeSource("synth-po-box-cedex")
 		const usPool: USTuple[] = []
 
-		for (const sourceEntry of opts.golden ? [US_EVAL_SOURCE] : US_TRAIN_SOURCES) {
+		for (const sourceEntry of opts.golden ? [US_SKELETON_EVAL_SOURCE] : US_SKELETON_TRAIN_SOURCES) {
 			const tuples = await readUsTuples(sourceEntry)
 
 			console.error(`  ${sourceEntry.csv}: ${tuples.length} tuples`)
@@ -96,7 +99,7 @@ export const poBoxCedexRecipe: CorpusRecipe = {
 
 		if (!usPool.length || !frPool.length || !qcPool.length || !onPool.length || !auPool.length || !nzPool.length)
 			throw new Error(
-				`A base pool is empty — check ${dataRootPath("oa-cache")}, ${GEONAMES_CA}, ${GEONAMES_POSTAL_AU.zip}, and ${GEONAMES_POSTAL_NZ.zip}.`
+				`A base pool is empty — check ${dataRootPath("oa-cache")}, ${dataRootPath(NAD_SKELETON_DIRECTORY)}, ${GEONAMES_CA}, ${GEONAMES_POSTAL_AU.zip}, and ${GEONAMES_POSTAL_NZ.zip}.`
 			)
 
 		let emitted = 0,
@@ -185,7 +188,7 @@ export const poBoxCedexRecipe: CorpusRecipe = {
 								? "GeoNames AU postal dump (CC-BY 4.0) locality/state/postcode tails + Australia Post Postal Delivery Type designators (@mailwoman/codex/au)"
 								: country === "NZ"
 									? "GeoNames NZ postal dump (CC-BY 4.0) locality/postcode tails + NZ Post ADV358 Delivery Service Types (@mailwoman/codex/nz)"
-									: "OpenAddresses US (non-VT) skeletons + USPS Pub-28 §29 PO-box designators (codex/corpus templates)",
+									: "OpenAddresses and National Address Database US (non-VT) skeletons + USPS Pub-28 §29 PO-box designators (codex/corpus templates)",
 			}
 
 			const aligned = alignRow(canonical as Parameters<typeof alignRow>[0])

package/lib/recipes/po/box/cedex/sources.ts

@@ -12,28 +12,7 @@ import { TextSpliterator, TSVSpliterator } from "spliterator"
 
 import type { AUTuple, FRTuple, NZTuple, USTuple } from "#recipes/po/box/cedex/types"
 import { readCSVRecords, readZippedCSVRecords } from "#recipes/scaffold"
-
-/**
- * Non-Vermont OpenAddresses files used for training rows.
- */
-export const US_TRAIN_SOURCES = [
-	{ 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" },
-]
-
-/**
- * Vermont OpenAddresses file reserved for golden rows.
- */
-export const US_EVAL_SOURCE = {
-	zip: dataRootPath("oa-cache", "us__vt__statewide.zip"),
-	csv: "us/vt/statewide.csv",
-	region: "VT",
-}
+import type { USSkeletonSource } from "#us/recipes/skeleton-sources"
 
 /**
  * Countrywide French OpenAddresses source.
@@ -72,7 +51,7 @@ export const isHoldoutLocality = (name: string): boolean => localityHash(name) %
 const cleanLocality = (locality: string) =>
 	locality && locality.length <= MAX_LOCALITY_LENGTH && !/\d|,/.test(locality) && !/cedex/i.test(locality)
 
-export async function readUsTuples(source: { zip: PathBuilderLike; csv: string; region: string }): Promise<USTuple[]> {
+export async function readUsTuples(source: USSkeletonSource): Promise<USTuple[]> {
 	const tuples: USTuple[] = []
 	const seen = new Set<string>()
 

package/lib/recipes/scaffold.ts

@@ -11,12 +11,13 @@
 
 import { pathExists } from "@mailwoman/core/fs/readers/stat"
 import { readZipEntry } from "@mailwoman/core/fs/zip"
-import { tryParsingJSON, stringifyJSON } from "@mailwoman/core/json"
+import { parseJSONStrict, tryParsingJSON, stringifyJSON } from "@mailwoman/core/json"
 import type { PathBuilderLike } from "path-ts"
 import type { AsyncChunkIterator, AsyncDataResource } from "spliterator"
 import { AsyncSequence, CSVSpliterator, TextSpliterator } from "spliterator"
 
 import { stableSourceIDFromParts } from "#adapters/source-id"
+import type { SourceRightsAdmission } from "#source-register/corpus-inputs"
 import type { SurfaceOrigin } from "#types"
 import { alignRow } from "#utils"
 
@@ -284,6 +285,29 @@ export function createRecipeLineWriter(sink: RecipeLineSink): WriteRecipeLine {
 	}
 }
 
+/**
+ * Wraps a recipe's writer so every line's `source` passes the source-rights check
+ * before the line is written.
+ *
+ * A recipe's `source` is its table default or whatever `--source-name` supplied,
+ * so the line itself is the only place the value is certain.
+ * The check caches each answer, so the cost after the first line of a source is one parse.
+ *
+ * @throws From the returned writer, when a line carries no string `source` or its source is refused.
+ */
+export function admittingRecipeLineWriter(write: WriteRecipeLine, rights: SourceRightsAdmission): WriteRecipeLine {
+	return (line) => {
+		const { source } = parseJSONStrict(line) as { source?: unknown }
+
+		if (typeof source !== "string") {
+			throw new TypeError(`a recipe wrote a row with no string \`source\`, so its rights cannot be checked: ${line}`)
+		}
+
+		rights.admit(source)
+		write(line)
+	}
+}
+
 /**
  * What a recipe records about the rows it writes.
  *
@@ -478,8 +502,9 @@ export interface RecipeOptions {
 	/**
 	 * `sub-venue`: GB/US/FR address-context tuples jsonl.
 	 *
-	 * @defaultValue the house-venue v3 tuples (`$MAILWOMAN_DATA_ROOT/corpus/intermediate/house-venue-tuples-v3.jsonl`);
-	 * DE and ES read OpenAddresses directly.
+	 * @defaultValue the house-venue v6 tuples (`$MAILWOMAN_DATA_ROOT/corpus/intermediate/house-venue-tuples-v6.jsonl`),
+	 * whose GB rows come from OS OpenData and whose US rows come from NAD and from the bodies whose own decisions
+	 * permit training. DE and ES read OpenAddresses directly.
 	 */
 	subVenueTuples?: string
 	/**

package/lib/recipes/sources.ts

@@ -135,7 +135,18 @@ export const RECIPE_SOURCES: ReadonlyArray<RecipeSource> = [
 		["es", LOCALE_RECIPE],
 		["es-pedania", LOCALE_RECIPE, "The ES source run with `--district-as-locality`, under `--source-name`."],
 		["nz", LOCALE_RECIPE],
-		["gb", LOCALE_RECIPE],
+		[
+			"gb",
+			LOCALE_RECIPE,
+			"Rendered from HM Land Registry Price Paid Data address fields, which Royal Mail and Ordnance Survey " +
+				"license for non-commercial use only. No recipe writes it now; `rendered-gb-os-open` replaces it (#2635).",
+		],
+		[
+			"gb-os-open",
+			LOCALE_RECIPE,
+			"The GB part reads the `gb-os-open` adapter's canonical rows (OS OpenData, OGL v3.0). " +
+				"No corpus stores the retired spelling.",
+		],
 		["fr-order", "fr/recipes/order.ts"],
 		["fr-admin-split", "fr/recipes/admin-split.ts", "The département is derived from the record's own postcode."],
 		["fr-lieudit", "fr/recipes/lieudit.ts"],
@@ -283,6 +294,106 @@ export const CARRIED_SOURCES: ReadonlyArray<string> = [
 	"osm-vn",
 	"osm-pk",
 	"osm-bd",
+	// The 2026-10-16 and 2026-10-17 extracts for the countries the corpus held under 1,000 street rows of,
+	// where no national address register with an electable grant was downloadable (country-gap-v1).
+	"osm-eg",
+	"osm-rs",
+	"osm-ar",
+	"osm-ua",
+	"osm-kz",
+	"osm-by",
+	"osm-md",
+	"osm-cl",
+	"osm-uz",
+	"osm-pe",
+	"osm-my",
+	"osm-bn",
+	"osm-ba",
+	"osm-ae",
+	"osm-kw",
+	"osm-ge",
+	"osm-am",
+	"osm-pr",
+	"osm-kg",
+	"osm-ph",
+	"osm-ci",
+	// The 2026-10-16 to 2026-10-18 extracts for the second set of countries under 1,000 street rows,
+	// where no national address register with an electable grant was downloadable (country-gap-v2).
+	// The Bahamas, Azerbaijan and French Polynesia also carry a national source that covers part
+	// of the country (`gbpa-bs` Freeport, `opendata-az` three cities, `ban-pf` three communes).
+	"osm-al",
+	"osm-ao",
+	"osm-aw",
+	"osm-az",
+	"osm-bb",
+	"osm-bh",
+	"osm-bm",
+	"osm-bs",
+	"osm-bt",
+	"osm-bw",
+	"osm-bz",
+	"osm-cd",
+	"osm-co",
+	"osm-cr",
+	"osm-cu",
+	"osm-cw",
+	"osm-do",
+	"osm-dz",
+	"osm-ec",
+	"osm-gg",
+	"osm-gh",
+	"osm-gl",
+	"osm-gt",
+	"osm-hk",
+	"osm-ht",
+	"osm-iq",
+	"osm-ir",
+	"osm-je",
+	"osm-jm",
+	"osm-jo",
+	"osm-ke",
+	"osm-kh",
+	"osm-lb",
+	"osm-lk",
+	"osm-lr",
+	"osm-ma",
+	"osm-me",
+	"osm-mg",
+	"osm-mk",
+	"osm-mm",
+	"osm-mn",
+	"osm-mo",
+	"osm-mt",
+	"osm-mv",
+	"osm-mz",
+	"osm-na",
+	"osm-np",
+	"osm-om",
+	"osm-pa",
+	"osm-pf",
+	"osm-py",
+	"osm-qa",
+	"osm-rw",
+	"osm-sa",
+	"osm-sm",
+	"osm-sn",
+	"osm-sr",
+	"osm-sv",
+	"osm-th",
+	"osm-tj",
+	"osm-tm",
+	"osm-tn",
+	"osm-tt",
+	"osm-tz",
+	"osm-ug",
+	"osm-ve",
+	"osm-vi",
+	"osm-zm",
+	"osm-zw",
+	// One id per dependency a national register covers beside its state, written by
+	// `corpus run <adapter> --country <cc> --source-name`, so the dependency's rows hold their own weight.
+	"ryhti-ax",
+	"matrikkelen-sj",
 	// One id per French overseas jurisdiction for the BAN department files,
 	// written by `corpus run ban --source-name`.
 	// Metropolitan rows keep `ban`, so its weight does not decide the rate at which a territory is drawn.

package/lib/recipes/street/affix.ts

@@ -26,30 +26,14 @@ import { stableSourceID } from "#adapters/source-id"
 import { readCSVRecords, readOATuples, recipeSourceID, type CorpusRecipe } from "#recipes/scaffold"
 import { defaultRecipeSource } from "#recipes/sources"
 import type { CanonicalRow } from "#types"
+import {
+	NAD_SKELETON_DIRECTORY,
+	US_SKELETON_EVAL_SOURCE,
+	US_SKELETON_TRAIN_SOURCES,
+	type USSkeletonSource,
+} from "#us/recipes/skeleton-sources"
 import { alignRow } from "#utils"
 
-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",
-}
-
 interface BalanceSource {
 	zip: PathBuilderLike
 	csv: string
@@ -118,12 +102,12 @@ interface BalanceTuple {
 
 type Prefix = Pick<NonNullable<ReturnType<typeof matchLeadingDirectional>>, "canonical" | "abbreviation">
 
-async function readTuples(source: USSource): Promise<USTuple[]> {
+async function readTuples(source: USSkeletonSource): Promise<USTuple[]> {
 	return readOATuples(source, {
 		extra: (fields, _row, key) => ({
 			...fields,
 			region: source.region,
-			base_source_id: `openaddresses:${source.csv}:${key}`,
+			base_source_id: `${source.register}:${source.csv}:${key}`,
 		}),
 	})
 }
@@ -372,7 +356,7 @@ export const streetAffixRecipe: CorpusRecipe = {
 		const count = opts.count ?? 50_000
 		const source = opts.sourceName ?? defaultRecipeSource("synth-affix")
 		const multilocaleCount = opts.multilocaleCount ?? 0
-		const sources = opts.golden ? [EVAL_SOURCE] : TRAIN_SOURCES
+		const sources = opts.golden ? [US_SKELETON_EVAL_SOURCE] : US_SKELETON_TRAIN_SOURCES
 
 		const pool: USTuple[] = []
 
@@ -387,7 +371,10 @@ export const streetAffixRecipe: CorpusRecipe = {
 		}
 
 		if (!pool.length) {
-			throw new Error(`No US tuples found — are the cached OA zips present in ${dataRootPath("oa-cache")}?`)
+			throw new Error(
+				`No US tuples found — are the cached OA zips present in ${dataRootPath("oa-cache")} and the NAD ` +
+					`skeletons in ${dataRootPath(NAD_SKELETON_DIRECTORY)}?`
+			)
 		}
 
 		let emitted = 0
@@ -448,7 +435,8 @@ export const streetAffixRecipe: CorpusRecipe = {
 				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)",
+				license:
+					"OpenAddresses and National Address Database US (non-VT) skeletons, street split via USPS Pub-28 C1/C2 (codex)",
 			}
 
 			const aligned = alignRow(canonical)
@@ -564,7 +552,7 @@ export const suffixBoundaryRecipe: CorpusRecipe = {
 		const random = makeMulberry32(opts.seed)
 		const count = opts.count ?? 30_000
 		const source = opts.sourceName ?? defaultRecipeSource("synth-suffix-boundary")
-		const sources = opts.golden ? [EVAL_SOURCE] : TRAIN_SOURCES
+		const sources = opts.golden ? [US_SKELETON_EVAL_SOURCE] : US_SKELETON_TRAIN_SOURCES
 
 		const venuePool = await readVenuePool(VENUE_POOL_CSV)
 
@@ -630,7 +618,7 @@ export const suffixBoundaryRecipe: CorpusRecipe = {
 				source_id: recipeSourceID(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)",
+					"OpenAddresses and National Address Database 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)

package/lib/recipes/sub/venue/context.ts

@@ -6,6 +6,7 @@
  */
 
 import { dataRootPath } from "@mailwoman/core/data-root"
+import { stringifyJSON } from "@mailwoman/core/json"
 import { mulberry32 as makeMulberry32 } from "@mailwoman/core/utils"
 import type { PathBuilderLike } from "path-ts"
 
@@ -22,24 +23,37 @@ import type { LocaleBaseTuple } from "#surfaces/locale"
  * so the CSV handling (quoted fields, crlf, city-noise cleaning, the DE per-part region fallback)
  * has exactly one implementation.
  *
- * The DE adapter reads two members from `europe.zip` instead of `oa-cache/de__*.zip`.
- * The cached per-state zips used by the `locale` recipe are absent on this host.
+ * The DE parts read the extracted `de/berlin.csv` and `de/sn/statewide.csv`.
+ * They are the `europe.zip` members the recipe read before that archive left the data root,
+ * with the md5s `9903d438…` and `623b099b…` that the note "a file's mtime is not
+ * its data's vintage" in `corpus/agents.md` records.
  *
- * The archive members match OA's current run byte for byte.
- * The note "a file's mtime is not its data's vintage" in `corpus/agents.md` records this comparison.
+ * A missing part throws in {@link readLocaleTuples}, so the DE leg cannot emit zero rows silently.
  */
 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" },
+		{ path: dataRootPath("openaddresses", "extracted", "de", "berlin.csv"), region: "Berlin" },
+		{ path: dataRootPath("openaddresses", "extracted", "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 `invented-house-venue` recipe output is built from, so the two
- * recipes' address halves are drawn from one pool), DE and ES streamed out of OpenAddresses.
+ * The `source` a Great Britain house-venue tuple must carry.
+ *
+ * The v3 tuples took their GB rows from HM Land Registry Price Paid Data, whose address
+ * fields Royal Mail and Ordnance Survey license for non-commercial use only (#2635).
+ * Their GB rows carry no `source`.
+ */
+export const GB_CONTEXT_TUPLE_SOURCE = "gb-os-open"
+
+/**
+ * Load the address skeletons every leg renders onto: GB / US / FR from the house-venue tuples
+ * (`corpora/house_venue_tuples.py`, whose GB rows come from the `gb-os-open` adapter's OS OpenData rows),
+ * DE and ES streamed out of OpenAddresses.
+ *
+ * @throws When a GB tuple does not carry {@linkcode GB_CONTEXT_TUPLE_SOURCE},
+ * so a tuples file built from Price Paid Data is refused rather than rendered.
  */
 export async function loadContextTuples(
 	tuplesPath: PathBuilderLike,
@@ -50,6 +64,16 @@ export async function loadContextTuples(
 	for await (const tuple of readRecipeTuples(tuplesPath)) {
 		const country = String(tuple.country ?? "")
 
+		// country-branch: GB is the one country whose earlier tuples came from a refused source (#2635).
+		// The tuples file records provenance for GB rows only, so the check reads GB rows only.
+		if (country === "GB" && tuple.source !== GB_CONTEXT_TUPLE_SOURCE) {
+			throw new Error(
+				`${tuplesPath.toString()} holds a GB tuple from source ${stringifyJSON(tuple.source ?? null)}; ` +
+					`GB address skeletons must come from ${GB_CONTEXT_TUPLE_SOURCE}. Rebuild the tuples with ` +
+					"`python -m mailwoman_train.corpora.house_venue_tuples`."
+			)
+		}
+
 		if (!country || !tuple.locality || !tuple.street) continue
 
 		const mapped: LocaleBaseTuple = {

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

@@ -12,6 +12,7 @@ import { upperFirst } from "@mailwoman/normalize/case"
 import type { PathBuilderLike } from "path-ts"
 
 import { escapeSQLString, openDuckDB } from "#parquet/duckdb"
+import { SourceRegister } from "#registers"
 import { classifyIdentifier, readSubVenueJSONL } from "#subvenue/harvest"
 import { SUBVENUE_PROMOTIONS, type SubVenuePromotion } from "#subvenue/promotions"
 import type { SubVenueLexiconTable } from "#subvenue/table"
@@ -378,11 +379,34 @@ export function sampleIdentifier(model: IdentifierModel, designatorID: string, r
 }
 
 /**
- * The address context and name pools for one recipe leg.
+ * The publication a pooled name was read from.
+ *
+ * Each row records the sources of the names it carries, because the two publications are
+ * licensed differently: OpenStreetMap under ODbL 1.0, Overture Places under CDLA-Permissive-2.0.
  */
-export interface LegPools {
-	context: LocaleBaseTuple[]
+export const NameSource = {
+	OpenStreetMap: SourceRegister.OpenStreetMap,
+	OverturePlaces: "overture-places",
+} as const
 
+export type NameSource = (typeof NameSource)[keyof typeof NameSource]
+
+/**
+ * A pooled name with the publication it was read from.
+ *
+ * A name both sources hold appears once per source, so its draw still records the pool it came from.
+ */
+export interface SourcedName {
+	name: string
+	source: NameSource
+}
+
+/**
+ * The name pools that one source contributes to a recipe leg.
+ *
+ * Only OSM extracts fill `attested` and `unpromotedShapes`, because POI data has no tier or localized names.
+ */
+export interface NamePools {
 	/**
 	 * Real venue-tier names that fit the venue slot, such as stations, airports, campuses and hospitals.
 	 */
@@ -410,11 +434,19 @@ export interface LegPools {
 }
 
 /**
- * The name pools that one source contributes to a recipe leg.
+ * The address context and name pools for one recipe leg.
  *
- * Only OSM extracts fill `attested` and `unpromotedShapes`, because POI data has no tier or localized names.
+ * `attested` and `unpromotedShapes` hold OpenStreetMap names only, as {@linkcode NamePools} states.
+ * The other three pools merge both sources and keep each name's source.
  */
-export type NamePools = Pick<LegPools, "venues" | "attested" | "rejectedVenues" | "longerNames" | "unpromotedShapes">
+export interface LegPools {
+	context: LocaleBaseTuple[]
+	venues: SourcedName[]
+	attested: string[]
+	rejectedVenues: SourcedName[]
+	longerNames: SourcedName[]
+	unpromotedShapes: string[]
+}
 
 /**
  * Empty name pools for a source that a leg lacks.
@@ -571,14 +603,23 @@ export async function readOvertureNamePools(
 }
 
 /**
- * Concatenates two sources' name pools.
+ * Concatenates an OpenStreetMap extract's pools and the Overture Places pools,
+ * OpenStreetMap first, and tags each merged name with its source.
+ *
+ * The order and length of every pool match a plain concatenation, so a seeded run
+ * draws the same names as one that recorded no source.
  */
-export function mergeNamePools(a: NamePools, b: NamePools): NamePools {
+export function mergeNamePools(osm: NamePools, overture: NamePools): Omit<LegPools, "context"> {
+	const tag = (a: readonly string[], b: readonly string[]): SourcedName[] => [
+		...a.map((name) => ({ name, source: NameSource.OpenStreetMap })),
+		...b.map((name) => ({ name, source: NameSource.OverturePlaces })),
+	]
+
 	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],
+		venues: tag(osm.venues, overture.venues),
+		attested: [...osm.attested, ...overture.attested],
+		rejectedVenues: tag(osm.rejectedVenues, overture.rejectedVenues),
+		longerNames: tag(osm.longerNames, overture.longerNames),
+		unpromotedShapes: [...osm.unpromotedShapes, ...overture.unpromotedShapes],
 	}
 }