@mailwoman/corpus 8.1.0 → 8.3.0

package/src/adapters/usgov-irs-bmf/adapter.ts

@@ -27,11 +27,21 @@ import { stableSourceID } from "../../adapter.ts"
 import { reconcileComponents } from "../../format.ts"
 import type { AdapterOptions, CanonicalRow, CorpusAdapter } from "../../types.ts"
 
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const USGOV_IRS_BMF_ADAPTER_ID = "usgov-irs-bmf"
+/**
+ * License carried by this source (Public Domain), attached to each row so downstream consumers inherit the terms rather
+ * than having to look them up.
+ */
 export const USGOV_IRS_BMF_DEFAULT_LICENSE = "Public Domain"
 
 const HOUSE_NUMBER_PREFIX = /^(\d+(?:-\d+)?[A-Za-z]?)\s+(.+)$/
-// PO box in its many written forms: "PO BOX 12", "P.O. BOX 12", "P O BOX 12", "POB 12", "BOX 12".
+/**
+ * PO box in its many written forms: "PO BOX 12", "P.O. BOX 12", "P O BOX 12", "POB 12", "BOX 12".
+ */
 const PO_BOX = /^\s*(?:P\.?\s?O\.?\s*BOX|POB|BOX)\s+\w/i
 
 interface IrsBmfRow {
@@ -43,7 +53,9 @@ interface IrsBmfRow {
 	ZIP: string
 }
 
-/** Classify the street line into a `po_box` or a `{house_number?, street}` split. */
+/**
+ * Classify the street line into a `po_box` or a `{house_number?, street}` split.
+ */
 function splitStreetLine(street: string): { po_box: string } | { house_number?: string; street: string } | null {
 	const trimmed = street.trim()
 
@@ -82,6 +94,7 @@ export function createUsgovIrsBmfAdapter(): CorpusAdapter {
 			}
 
 			const stream = createReadStream(opts.inputPath, { encoding: "utf8" })
+
 			const parser = stream.pipe(
 				csvParse({ columns: true, skip_empty_lines: true, relax_quotes: true, relax_column_count: true, trim: true })
 			)
@@ -143,6 +156,7 @@ export function createUsgovIrsBmfAdapter(): CorpusAdapter {
 						corpus_version: "",
 						license: USGOV_IRS_BMF_DEFAULT_LICENSE,
 					}
+
 					emitted++
 				}
 			} finally {
@@ -152,4 +166,7 @@ export function createUsgovIrsBmfAdapter(): CorpusAdapter {
 	}
 }
 
+/**
+ * The configured adapter instance registered with the corpus builder.
+ */
 export const usgovIrsBmfAdapter = createUsgovIrsBmfAdapter()

package/src/adapters/usgov-nad/adapter.ts

@@ -38,7 +38,15 @@ import { TextSpliterator } from "spliterator"
 import { reconcileComponents } from "../../format.ts"
 import type { AdapterOptions, CanonicalRow, CorpusAdapter } from "../../types.ts"
 
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const USGOV_NAD_ADAPTER_ID = "usgov-nad"
+/**
+ * License carried by this source (Public Domain), attached to each row so downstream consumers inherit the terms rather
+ * than having to look them up.
+ */
 export const USGOV_NAD_DEFAULT_LICENSE = "Public Domain"
 
 interface NADRecord {
@@ -192,6 +200,7 @@ function decomposeNADStreet(r: NADRecord): DecomposedNADStreet | undefined {
 
 		return { prefix, street: name, suffix, full }
 	}
+
 	const full = (r.StNam_Full ?? "").toString().trim()
 
 	if (full) return { full, street: full }
@@ -243,7 +252,7 @@ export function createUsgovNADAdapter(): CorpusAdapter {
 			// Single-file inputs (e.g. a bulk-extracted CSV) are not currently supported — the
 			// featureserver shard pattern is the primary distribution.
 			const entries = await readdir(opts.inputPath)
-			const shards = entries.filter((n) => n.endsWith(".ndjson")).sort()
+			const shards = entries.filter((n) => n.endsWith(".ndjson")).toSorted()
 
 			let emitted = 0
 			outer: for (const shard of shards) {
@@ -328,6 +337,7 @@ export function createUsgovNADAdapter(): CorpusAdapter {
 						corpus_version: "",
 						license: USGOV_NAD_DEFAULT_LICENSE,
 					}
+
 					emitted++
 				}
 			}
@@ -335,4 +345,7 @@ export function createUsgovNADAdapter(): CorpusAdapter {
 	}
 }
 
+/**
+ * The configured adapter instance registered with the corpus builder.
+ */
 export const usgovNADAdapter = createUsgovNADAdapter()

package/src/adapters/usgov-nppes/adapter.ts

@@ -30,7 +30,15 @@ import { lookupStateAbbreviation } from "../../codex/us-fips-state.ts"
 import { reconcileComponents } from "../../format.ts"
 import type { AdapterOptions, CanonicalRow, CorpusAdapter } from "../../types.ts"
 
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const USGOV_NPPES_ADAPTER_ID = "usgov-nppes"
+/**
+ * License carried by this source (Public Domain), attached to each row so downstream consumers inherit the terms rather
+ * than having to look them up.
+ */
 export const USGOV_NPPES_DEFAULT_LICENSE = "Public Domain"
 
 const HOUSE_NUMBER_PREFIX = /^(\d+(?:-\d+)?[A-Za-z]?)\s+(.+)$/
@@ -86,6 +94,7 @@ export function createUsgovNPPESAdapter(): CorpusAdapter {
 			}
 
 			const stream = createReadStream(opts.inputPath, { encoding: "utf8" })
+
 			const parser = stream.pipe(
 				csvParse({
 					columns: true,
@@ -157,6 +166,7 @@ export function createUsgovNPPESAdapter(): CorpusAdapter {
 						corpus_version: "",
 						license: USGOV_NPPES_DEFAULT_LICENSE,
 					}
+
 					emitted++
 				}
 			} finally {
@@ -166,4 +176,7 @@ export function createUsgovNPPESAdapter(): CorpusAdapter {
 	}
 }
 
+/**
+ * The configured adapter instance registered with the corpus builder.
+ */
 export const usgovNPPESAdapter = createUsgovNPPESAdapter()

package/src/adapters/usgov-samhsa-treatment-locator/adapter.ts

@@ -37,7 +37,15 @@ import { lookupStateAbbreviation } from "../../codex/us-fips-state.ts"
 import { reconcileComponents } from "../../format.ts"
 import type { AdapterOptions, CanonicalRow, CorpusAdapter } from "../../types.ts"
 
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const USGOV_SAMHSA_ADAPTER_ID = "usgov-samhsa-treatment-locator"
+/**
+ * License carried by this source (Public Domain), attached to each row so downstream consumers inherit the terms rather
+ * than having to look them up.
+ */
 export const USGOV_SAMHSA_DEFAULT_LICENSE = "Public Domain"
 
 /**
@@ -52,7 +60,9 @@ interface SamhsaSiteRow {
 	city: string
 	state: string
 	zip: string
-	/** Optional. Falls back to `stableSourceID` derived from components when missing. */
+	/**
+	 * Optional. Falls back to `stableSourceID` derived from components when missing.
+	 */
 	frid?: string
 }
 
@@ -109,7 +119,9 @@ function composeVenue(name1: string, name2: string | undefined): string {
 	return `${n1} - ${n2}`
 }
 
-/** Same envelope-style format as HRSA: venue prefix, street body, city/state/zip suffix. */
+/**
+ * Same envelope-style format as HRSA: venue prefix, street body, city/state/zip suffix.
+ */
 function composeRaw(
 	venue: string,
 	house: string | undefined,
@@ -137,6 +149,7 @@ export function createUsgovSamhsaTreatmentLocatorAdapter(): CorpusAdapter {
 			}
 
 			const stream = createReadStream(opts.inputPath, { encoding: "utf8" })
+
 			const parser = stream.pipe(
 				csvParse({
 					columns: true,
@@ -182,9 +195,10 @@ export function createUsgovSamhsaTreatmentLocatorAdapter(): CorpusAdapter {
 
 					const aligned = reconcileComponents(components, raw)
 
-					if (Object.keys(aligned).length === 0) continue
+					if (!Object.keys(aligned).length) continue
 
 					const frID = (record.frid ?? "").trim()
+
 					const sourceID = frID
 						? `${USGOV_SAMHSA_ADAPTER_ID}-${frID}`
 						: stableSourceID(USGOV_SAMHSA_ADAPTER_ID, aligned)
@@ -199,6 +213,7 @@ export function createUsgovSamhsaTreatmentLocatorAdapter(): CorpusAdapter {
 						corpus_version: "",
 						license: USGOV_SAMHSA_DEFAULT_LICENSE,
 					}
+
 					emitted++
 				}
 			} finally {
@@ -208,4 +223,7 @@ export function createUsgovSamhsaTreatmentLocatorAdapter(): CorpusAdapter {
 	}
 }
 
+/**
+ * The configured adapter instance registered with the corpus builder.
+ */
 export const usgovSamhsaTreatmentLocatorAdapter = createUsgovSamhsaTreatmentLocatorAdapter()

package/src/adapters/wof-admin-jp/adapter.ts

@@ -33,6 +33,15 @@ import { DatabaseSync } from "node:sqlite"
 
 import type { AdapterOptions, CanonicalRow, CorpusAdapter } from "../../types.ts"
 
+/**
+ * Ancestry hops walked upward from a WOF record before giving up — deeper than any real JP admin chain.
+ */
+const MAX_ANCESTRY_DEPTH = 6
+
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const WOF_ADMIN_JP_ADAPTER_ID = "wof-admin-jp"
 
 interface PlaceRow {
@@ -49,13 +58,15 @@ interface NameRow {
 	language: string
 }
 
-/** Walk parent chain up to 6 levels. */
+/**
+ * Walk parent chain up to 6 levels.
+ */
 function chainOf(db: DatabaseSync, startID: number, _jpnNames: Map<number, string>): PlaceRow[] {
 	const stmt = db.prepare(`SELECT id, name, placetype, parent_id, country FROM spr WHERE id = ?`)
 	const out: PlaceRow[] = []
 	let id = startID
 
-	for (let i = 0; i < 6 && id > 0; i++) {
+	for (let i = 0; i < MAX_ANCESTRY_DEPTH && id > 0; i++) {
 		const row = stmt.get(id) as PlaceRow | undefined
 
 		if (!row) break
@@ -66,7 +77,9 @@ function chainOf(db: DatabaseSync, startID: number, _jpnNames: Map<number, strin
 	return out
 }
 
-/** Pick the best display name for a place: prefer Japanese variant, fall back to English. */
+/**
+ * Pick the best display name for a place: prefer Japanese variant, fall back to English.
+ */
 function pickName(row: PlaceRow, jpnNames: Map<number, string>): string {
 	return jpnNames.get(row.id) ?? row.name
 }
@@ -164,6 +177,7 @@ export function createWOFAdminJpAdapter(): CorpusAdapter {
 						corpus_version: "",
 						license: "CC-BY-4.0",
 					}
+
 					emitted++
 				}
 			} finally {
@@ -173,4 +187,7 @@ export function createWOFAdminJpAdapter(): CorpusAdapter {
 	}
 }
 
+/**
+ * The configured adapter instance registered with the corpus builder.
+ */
 export const wofAdminJpAdapter = createWOFAdminJpAdapter()

package/src/adapters/wof-admin-json/adapter.ts

@@ -61,13 +61,17 @@ const COUNTRY_DISPLAY_NAME: Record<string, string> = {
 	FR: "France",
 }
 
-/** BCP-47 locale defaulting for the corpus row's `locale` field. Defaulted by country. */
+/**
+ * BCP-47 locale defaulting for the corpus row's `locale` field. Defaulted by country.
+ */
 const LOCALE_BY_COUNTRY: Record<string, string> = {
 	US: "en-US",
 	FR: "fr-FR",
 }
 
-/** Map a WOF placetype to a Mailwoman `ComponentTag`, or `undefined` to skip. */
+/**
+ * Map a WOF placetype to a Mailwoman `ComponentTag`, or `undefined` to skip.
+ */
 function placetypeToTag(placetype: WhosOnFirstPlacetype | string): ComponentTag | undefined {
 	switch (placetype) {
 		case "country":
@@ -93,10 +97,14 @@ function placetypeToTag(placetype: WhosOnFirstPlacetype | string): ComponentTag
 }
 
 interface VariantSpec {
-	/** Hierarchy-variant id appended to `source_id`. */
+	/**
+	 * Hierarchy-variant id appended to `source_id`.
+	 */
 	suffix: string
 
-	/** Component tag → display string the adapter will hand to the runner. */
+	/**
+	 * Component tag → display string the adapter will hand to the runner.
+	 */
 	components: Partial<Record<ComponentTag, string>>
 }
 
@@ -202,6 +210,10 @@ export function nameSlotsFor(rec: WOFRecord): Array<{ key: string; value: string
 	return slots
 }
 
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const WOF_ADMIN_ADAPTER_ID = "wof-admin"
 
 /**
@@ -236,7 +248,7 @@ export function createWOFAdminAdapter(): CorpusAdapter {
 			const ancestry = buildAncestryIndex(byID)
 
 			// Pass 2: emit rows in sorted-id order for deterministic JSONL.
-			const ids = [...byID.keys()].sort((a, b) => a - b)
+			const ids = [...byID.keys()].toSorted((a, b) => a - b)
 			let emitted = 0
 
 			for (const id of ids) {
@@ -256,7 +268,7 @@ export function createWOFAdminAdapter(): CorpusAdapter {
 						if (!raw) continue
 						const aligned = reconcileComponents(variant.components, raw)
 
-						if (Object.keys(aligned).length === 0) continue
+						if (!Object.keys(aligned).length) continue
 
 						yield {
 							raw,
@@ -268,6 +280,7 @@ export function createWOFAdminAdapter(): CorpusAdapter {
 							corpus_version: "",
 							license: "CC0-1.0",
 						}
+
 						emitted++
 					}
 				}
@@ -276,5 +289,7 @@ export function createWOFAdminAdapter(): CorpusAdapter {
 	}
 }
 
-/** Single shared instance, suitable for `defaultAdapterRegistry`. */
+/**
+ * Single shared instance, suitable for `defaultAdapterRegistry`.
+ */
 export const wofAdminAdapter = createWOFAdminAdapter()

package/src/adapters/wof-postalcode-json/adapter.ts

@@ -136,6 +136,10 @@ export function nameSlotsFor(rec: WOFRecord): Array<{ key: string; value: string
 	return slots
 }
 
+/**
+ * Registry id for this adapter. Stamped into every row it emits, so a corpus record can be traced back to the dataset
+ * it came from.
+ */
 export const WOF_POSTALCODE_ADAPTER_ID = "wof-postalcode"
 
 export function createWOFPostalcodeAdapter(): CorpusAdapter {
@@ -163,7 +167,7 @@ export function createWOFPostalcodeAdapter(): CorpusAdapter {
 			const ancestry = buildAncestryIndex(byID)
 
 			// Pass 2: emit postcode rows only, sorted by id for determinism.
-			const ids = [...byID.keys()].sort((a, b) => a - b)
+			const ids = [...byID.keys()].toSorted((a, b) => a - b)
 			let emitted = 0
 
 			for (const id of ids) {
@@ -186,7 +190,7 @@ export function createWOFPostalcodeAdapter(): CorpusAdapter {
 						if (!raw) continue
 						const aligned = reconcileComponents(variant.components, raw)
 
-						if (Object.keys(aligned).length === 0) continue
+						if (!Object.keys(aligned).length) continue
 
 						yield {
 							raw,
@@ -198,6 +202,7 @@ export function createWOFPostalcodeAdapter(): CorpusAdapter {
 							corpus_version: "",
 							license: "CC0-1.0",
 						}
+
 						emitted++
 					}
 				}
@@ -206,4 +211,7 @@ export function createWOFPostalcodeAdapter(): CorpusAdapter {
 	}
 }
 
+/**
+ * The configured adapter instance registered with the corpus builder.
+ */
 export const wofPostalcodeAdapter = createWOFPostalcodeAdapter()

package/src/align.ts

@@ -39,9 +39,13 @@ import { distance as levenshteinDistance } from "fastest-levenshtein"
 import { whitespaceTokenizer, type TokenSpan, type Tokenizer } from "./tokenize.ts"
 import type { CanonicalRow, LabeledRow, QuarantinedRow } from "./types.ts"
 
-/** Options for `alignRow`. */
+/**
+ * Options for `alignRow`.
+ */
 export interface AlignOptions {
-	/** Tokenizer to use. Defaults to `whitespaceTokenizer()`. */
+	/**
+	 * Tokenizer to use. Defaults to `whitespaceTokenizer()`.
+	 */
 	tokenizer?: Tokenizer
 
 	/**
@@ -60,7 +64,9 @@ export interface AlignOptions {
 	caseInsensitive?: boolean
 }
 
-/** Either a successful labeled row or a quarantined one. */
+/**
+ * Either a successful labeled row or a quarantined one.
+ */
 export type AlignmentResult = { kind: "labeled"; row: LabeledRow } | { kind: "quarantined"; row: QuarantinedRow }
 
 /**
@@ -73,7 +79,9 @@ export interface ComponentSpan {
 	end: number
 }
 
-/** Align a single row. */
+/**
+ * Align a single row.
+ */
 export function alignRow(row: CanonicalRow, opts: AlignOptions = {}): AlignmentResult {
 	const tokenizer = opts.tokenizer ?? whitespaceTokenizer()
 	const maxEditDistance = opts.maxEditDistance ?? 2
@@ -107,7 +115,7 @@ export function alignRow(row: CanonicalRow, opts: AlignOptions = {}): AlignmentR
 	// component owns ("Alaska Regional Dr, Alaska" — region "Alaska" stealing the street's first
 	// word quarantined the street; pilot2's residual class). Emit order is unaffected — spans are
 	// re-sorted by start below.
-	const entries = (Object.entries(components) as Array<[ComponentTag, string | undefined]>).sort(
+	const entries = (Object.entries(components) as Array<[ComponentTag, string | undefined]>).toSorted(
 		(a, b) => (b[1]?.length ?? 0) - (a[1]?.length ?? 0)
 	)
 
@@ -220,7 +228,7 @@ function locateSpan(args: {
 }): { start: number; end: number } | undefined {
 	const { haystack, needle, claimed, maxEditDistance } = args
 
-	if (needle.length === 0) return undefined
+	if (!needle.length) return undefined
 
 	// Pass 1: verbatim substring. Word-boundary-aligned matches are PREFERRED over intra-word ones
 	// — leftmost-substring alone let a short value claim the inside of an earlier word (region "AK"
@@ -234,13 +242,14 @@ function locateSpan(args: {
 	while (true) {
 		const idx = haystack.indexOf(needle, from)
 
-		if (idx < 0) break
+		if (idx === -1) break
 		const end = idx + needle.length
 
 		if (!overlapsClaimed(idx, end, claimed)) {
 			if (isBoundaryAligned(haystack, idx, end)) return { start: idx, end }
 			intraWord ??= { start: idx, end }
 		}
+
 		from = idx + 1
 	}
 
@@ -267,7 +276,9 @@ function locateSpan(args: {
 
 const WORD_CHAR = /[\p{L}\p{N}]/u
 
-/** Both needle edges sit on word boundaries of the haystack (string edges count as boundaries). */
+/**
+ * Both needle edges sit on word boundaries of the haystack (string edges count as boundaries).
+ */
 function isBoundaryAligned(haystack: string, start: number, end: number): boolean {
 	const before = start === 0 || !WORD_CHAR.test(haystack[start - 1]!)
 	const after = end === haystack.length || !WORD_CHAR.test(haystack[end]!)
@@ -307,9 +318,11 @@ function labelTokens(tokens: readonly TokenSpan[], spans: readonly ComponentSpan
 				} else {
 					assigned = `I-${s.tag}` as BIOLabel
 				}
+
 				break
 			}
 		}
+
 		out.push(assigned)
 	}
 

package/src/build.ts

@@ -69,15 +69,23 @@ import {
 import { defaultAugmentationsForCountry, synthesizeRow } from "./synthesize.ts"
 import type { AdapterOptions, CanonicalRow, CorpusAdapter, LabeledRow } from "./types.ts"
 
-/** Stage tags surfaced to `onProgress`. */
+/**
+ * Stage tags surfaced to `onProgress`.
+ */
 export type BuildStage = "adapter-run" | "align" | "split" | "shard" | "manifest"
 
-/** Per-invocation options for `buildCorpus`. */
+/**
+ * Per-invocation options for `buildCorpus`.
+ */
 export interface BuildCorpusOptions {
-	/** Root output directory. All build artifacts land beneath it. */
+	/**
+	 * Root output directory. All build artifacts land beneath it.
+	 */
 	outputDir: string
 
-	/** Corpus version (e.g. `"0.1.0"`). Stamped onto every row + into the output dir name. */
+	/**
+	 * Corpus version (e.g. `"0.1.0"`). Stamped onto every row + into the output dir name.
+	 */
 	corpusVersion: string
 
 	/**
@@ -92,13 +100,19 @@ export interface BuildCorpusOptions {
 	 */
 	adapterInputs: Record<string, AdapterOptions>
 
-	/** Enable synthesis pass. Default `true`. Set `false` for fixture-driven smoke tests. */
+	/**
+	 * Enable synthesis pass. Default `true`. Set `false` for fixture-driven smoke tests.
+	 */
 	synthesize?: boolean
 
-	/** Forwarded to `writeShards`. Default 1_000_000. */
+	/**
+	 * Forwarded to `writeShards`. Default 1_000_000.
+	 */
 	rowsPerShard?: number
 
-	/** Progress hook. Errors thrown abort the build. */
+	/**
+	 * Progress hook. Errors thrown abort the build.
+	 */
 	onProgress?: (stage: BuildStage, message: string) => void
 
 	/**
@@ -110,7 +124,9 @@ export interface BuildCorpusOptions {
 	excludeLicenses?: readonly RegExp[]
 }
 
-/** Top-level manifest tying every stage together. */
+/**
+ * Top-level manifest tying every stage together.
+ */
 export interface BuildCorpusManifest {
 	corpus_version: string
 	built_at: string
@@ -155,8 +171,10 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 		if (!adapterOptions) {
 			skipped.push(adapter.id)
 			opts.onProgress?.("adapter-run", `skipped ${adapter.id} (no input configured)`)
+
 			continue
 		}
+
 		// Opt-in resume (MAILWOMAN_RESUME=1): if a complete per-adapter canonical.jsonl + MANIFEST.json
 		// already exist, reuse them instead of re-emitting. The MANIFEST is written only after the
 		// canonical is fully flushed, so its presence guarantees completeness; row order is identical,
@@ -173,15 +191,19 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 			const cached = JSON.parse(readFileSync(cachedManifest, "utf8")) as AdapterRunManifest
 			opts.onProgress?.("adapter-run", `resumed ${adapter.id} (reused ${cached.yielded} canonical rows)`)
 			adapterRuns.push(cached)
+
 			continue
 		}
+
 		opts.onProgress?.("adapter-run", `running ${adapter.id}`)
+
 		const m = await runAdapter({
 			adapter,
 			adapterOptions,
 			outputDir: intermediateDir,
 			corpusVersion: opts.corpusVersion,
 		})
+
 		adapterRuns.push(m)
 	}
 
@@ -195,11 +217,13 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 		val: join(intermediateDir, "labeled-val.jsonl"),
 		test: join(intermediateDir, "labeled-test.jsonl"),
 	}
+
 	const labeledStreams: Record<SplitName, WriteStream> = {
 		train: createWriteStream(labeledPaths.train, { encoding: "utf8" }),
 		val: createWriteStream(labeledPaths.val, { encoding: "utf8" }),
 		test: createWriteStream(labeledPaths.test, { encoding: "utf8" }),
 	}
+
 	const quarantinePath = join(intermediateDir, "quarantine.jsonl")
 	const quarantineStream = createWriteStream(quarantinePath, { encoding: "utf8" })
 
@@ -230,8 +254,10 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 			// set reflects what the corpus actually CONTAINED, and `excluded_by_license` what was cut.
 			if (licenseExcluded(row.license, excludeLicenses)) {
 				excludedByLicense++
+
 				continue
 			}
+
 			const fanned: CanonicalRow[] = [row]
 
 			if (synthesize) {
@@ -245,23 +271,28 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 
 				try {
 					result = alignRow(r)
-				} catch (err) {
+				} catch (error) {
 					// Last-resort robustness (2026-06-12): no single row may crash a multi-hour build.
 					// alignRow's targeted paths normalize/quarantine known issues with specific reasons;
 					// this catches any UNKNOWN throw (e.g. assertSpanInvariants on an unforeseen span
 					// shape) → quarantine + continue. A spike in `align-threw` reasons is a finding.
-					writeQuarantine(r, `align-threw:${(err as Error).message.slice(0, 160)}`)
+					writeQuarantine(r, `align-threw:${(error as Error).message.slice(0, 160)}`)
+
 					quarantined++
+
 					continue
 				}
 
 				if (result.kind === "labeled") {
 					const split = splitForRow(result.row, holdouts)
 					labeledStreams[split].write(`${JSON.stringify(result.row)}\n`)
+
 					counts[split]++
+
 					aligned++
 				} else {
 					writeQuarantine(r, result.row.reason)
+
 					quarantined++
 				}
 			}
@@ -271,6 +302,7 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 	for (const s of Object.values(labeledStreams)) {
 		s.end()
 	}
+
 	quarantineStream.end()
 	await Promise.all([...Object.values(labeledStreams).map(streamEnd), streamEnd(quarantineStream)])
 
@@ -279,6 +311,7 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 	// manifests with disk spill for splits that exceed in-memory thresholds.
 	opts.onProgress?.("split", `splitting ${aligned} aligned rows`)
 	const splitsDir = join(opts.outputDir, "splits")
+
 	const splitCounts = await writeSplitManifestsFromLabeledFiles({
 		labeledPaths,
 		outputDir: splitsDir,
@@ -290,6 +323,7 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 	// 5. Parquet shards — per-split labeled JSONL streams in, sharded `.parquet` out. The prior
 	// `splitFor(source_id)` callback (and the `Map<source_id, SplitName>` behind it) is gone.
 	opts.onProgress?.("shard", "writing parquet shards")
+
 	const shardManifest = await writeShards(
 		{
 			train: streamJSONL<LabeledRow>(labeledPaths.train),
@@ -305,7 +339,8 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 
 	// License-set visibility (#26): loudly report the resolved license set so a build is an obvious
 	// deliberate act — especially a proprietary-weights build (did you pass --exclude-share-alike?).
-	const licenseSummary = [...licenseCounts.entries()].sort((a, b) => b[1] - a[1])
+	const licenseSummary = [...licenseCounts.entries()].toSorted((a, b) => b[1] - a[1])
+
 	opts.onProgress?.(
 		"manifest",
 		`license set: ${licenseSummary.map(([l, c]) => `${l}=${c}`).join(", ")}` +
@@ -316,6 +351,7 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 
 	// 6. Top-level manifest.
 	opts.onProgress?.("manifest", "writing top-level MANIFEST.json")
+
 	const manifest: BuildCorpusManifest = {
 		corpus_version: opts.corpusVersion,
 		built_at,
@@ -328,6 +364,7 @@ export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpus
 		licenses: Object.fromEntries(licenseSummary),
 		excluded_by_license: excludedByLicense,
 	}
+
 	await writeFile(join(opts.outputDir, "MANIFEST.json"), `${JSON.stringify(manifest, null, 2)}\n`, "utf8")
 
 	return manifest