@mailwoman/corpus 9.1.0 → 9.3.0

package/src/build.ts

@@ -1,386 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   End-to-end corpus build (Phase 1 task #10 in the plan).
- *
- *   `buildCorpus(opts)` orchestrates every stage of the pipeline:
- *
- *   1. **Adapter runs** — drives every adapter in turn (via `runAdapter`), writing
- *        `<intermediate>/<adapter.id>/canonical.jsonl` shards.
- *   2. **Synthesis** — optional. For each canonical row, every applicable augmentation in the row's
- *        country-default policy emits an augmented row alongside the original.
- *   3. **Alignment** — every row (original + augmented) is aligned via `alignRow`. Successes go to
- *        `labeled.jsonl`; quarantines are appended to `quarantine.jsonl` with reasons.
- *   4. **Splits** — `splitRows` partitions labeled `source_id`s into train/val/test by locality holdout.
- *        Manifest written to `splits/SPLIT_MANIFEST.json` + per-split `train.txt` / `val.txt` /
- *        `test.txt`.
- *   5. **Parquet shards** — `writeShards` streams labeled rows into 1M-row `.parquet` shards per split
- *        under `corpus-v<version>/{train,val,test}/part-NNNN.parquet` (SNAPPY-compressed, 50k-row
- *        row groups), with per-shard checksums + per-stage manifest in
- *        `corpus-v<version>/MANIFEST.json`.
- *   6. **Top-level manifest** — `<outputDir>/MANIFEST.json` ties every per-stage manifest together with
- *        a top-level corpus_version, built_at, and aggregate counts.
- *
- *   Output layout:
- *
- *   ```
- *   <outputDir>/
- *   MANIFEST.json
- *   intermediate/
- *     <adapter.id>/canonical.jsonl   # one per adapter
- *     labeled.jsonl                  # post-alignment, pre-shard
- *     quarantine.jsonl               # rows that failed alignment
- *   splits/
- *     SPLIT_MANIFEST.json
- *     train.txt / val.txt / test.txt
- *   corpus-v<version>/
- *     MANIFEST.json
- *     train/part-NNNN.parquet
- *     val/part-NNNN.parquet
- *     test/part-NNNN.parquet
- * ```
- *
- *   The intermediate files live alongside the final shards for reproducibility + debugging. Operators
- *   can `rm -rf intermediate/` after the build if disk is tight; the final `corpus-v<version>/` is
- *   self-contained.
- */
-
-import { createWriteStream, existsSync, readFileSync, type WriteStream } from "node:fs"
-import { mkdir, writeFile } from "node:fs/promises"
-import { join } from "node:path"
-
-import { $public } from "@mailwoman/core/env"
-import { parseJSONStrict } from "@mailwoman/core/objects"
-import { defaultAdapterRegistry } from "@mailwoman/corpus/adapters/utils"
-import { runAdapter, type AdapterRunManifest } from "@mailwoman/corpus/runner"
-import { defaultAugmentationsForCountry, synthesizeRow } from "@mailwoman/corpus/synthesizers/utils"
-import type { AdapterOptions, CanonicalRow, CorpusAdapter, LabeledRow } from "@mailwoman/corpus/types"
-import {
-	alignRow,
-	licenseExcluded,
-	writeShards,
-	defaultHoldouts,
-	splitForRow,
-	writeSplitManifestsFromLabeledFiles,
-	type ShardManifest,
-	type SplitManifest,
-	type SplitName,
-} from "@mailwoman/corpus/utils"
-import { JSONSpliterator } from "spliterator"
-
-/**
- * Stage tags surfaced to `onProgress`.
- */
-export type BuildStage = "adapter-run" | "align" | "split" | "shard" | "manifest"
-
-/**
- * Per-invocation options for `buildCorpus`.
- */
-export interface BuildCorpusOptions {
-	/**
-	 * 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.
-	 */
-	corpusVersion: string
-
-	/**
-	 * Adapters to drive, in order. Defaults to `defaultAdapterRegistry.list()`. Pass an explicit list to filter (e.g.
-	 * `[wofAdminAdapter]` for a smoke run).
-	 */
-	adapters?: readonly CorpusAdapter[]
-
-	/**
-	 * Per-adapter `AdapterOptions` — looked up by adapter id. Adapters whose id is missing from this map are skipped (and
-	 * noted in the manifest).
-	 */
-	adapterInputs: Record<string, AdapterOptions>
-
-	/**
-	 * Enable synthesis pass. Default `true`. Set `false` for fixture-driven smoke tests.
-	 */
-	synthesize?: boolean
-
-	/**
-	 * Forwarded to `writeShards`. Default 1_000_000.
-	 */
-	rowsPerShard?: number
-
-	/**
-	 * Progress hook. Errors thrown abort the build.
-	 */
-	onProgress?: (stage: BuildStage, message: string) => void
-
-	/**
-	 * License kinds to PURPOSELY exclude from this build (#26). Compiled patterns (see `compileLicenseExcludes` /
-	 * `SHARE_ALIKE_PATTERN` in `license.ts`); a row whose `license` matches any is dropped at ingest. Default (omitted)
-	 * includes EVERYTHING — exclusion is a deliberate act, not a silent default. A proprietary-weights build passes the
-	 * share-alike set (`--exclude-share-alike`).
-	 */
-	excludeLicenses?: readonly RegExp[]
-}
-
-/**
- * Top-level manifest tying every stage together.
- */
-export interface BuildCorpusManifest {
-	corpus_version: string
-	built_at: string
-	adapters: AdapterRunManifest[]
-	skipped_adapters: string[]
-	splits: { counts: SplitManifest["counts"]; holdouts: SplitManifest["holdouts"] }
-	shards: { counts: ShardManifest["counts"]; total_rows: number }
-	quarantine_count: number
-	total_aligned_rows: number
-	/**
-	 * Resolved license set across all INCLUDED rows (license string → row count), + the count dropped by
-	 * `excludeLicenses` (#26). The model card derives its data-attribution table from `licenses`.
-	 */
-	licenses: Record<string, number>
-	excluded_by_license: number
-}
-
-/**
- * Drive the full corpus build to completion.
- *
- * Memory profile: the function maintains an in-memory `Map<source_id, SplitName>` to bridge the align → shard hand-off.
- * For Phase 1 fixture-scale runs (≤ 10⁴ rows) this is trivial. For real 5M+ runs, the map fits comfortably in a few
- * hundred MB; the canonical.jsonl and labeled.jsonl payloads stream and never sit in memory.
- */
-export async function buildCorpus(opts: BuildCorpusOptions): Promise<BuildCorpusManifest> {
-	const adapters = opts.adapters ?? defaultAdapterRegistry.list()
-	const synthesize = opts.synthesize ?? true
-	const rowsPerShard = opts.rowsPerShard ?? 1_000_000
-	const built_at = new Date().toISOString()
-
-	await mkdir(opts.outputDir, { recursive: true })
-	const intermediateDir = join(opts.outputDir, "intermediate")
-	await mkdir(intermediateDir, { recursive: true })
-
-	// 1. Adapter runs.
-	const adapterRuns: AdapterRunManifest[] = []
-	const skipped: string[] = []
-
-	for (const adapter of adapters) {
-		const adapterOptions = opts.adapterInputs[adapter.id]
-
-		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,
-		// so downstream holdout-split determinism is preserved. Recovers an align-phase crash without
-		// redoing the (expensive) emit phase. Default (unset) re-emits, preserving correctness. (2026-06-12.)
-		const adapterDir = join(intermediateDir, adapter.id)
-		const cachedManifest = join(adapterDir, "MANIFEST.json")
-
-		if (
-			$public.MAILWOMAN_RESUME === "1" &&
-			existsSync(cachedManifest) &&
-			existsSync(join(adapterDir, "canonical.jsonl"))
-		) {
-			const cached = parseJSONStrict<AdapterRunManifest>(readFileSync(cachedManifest, "utf8"))
-			opts.onProgress?.("adapter-run", `resumed ${adapter.id} (reused ${cached.yielded} canonical rows)`)
-			adapterRuns.push(cached)
-
-			continue
-		}
-
-		opts.onProgress?.("adapter-run", `running ${adapter.id}`)
-
-		const m = await runAdapter({
-			adapter,
-			adapterOptions,
-			outputDir: intermediateDir,
-			corpusVersion: opts.corpusVersion,
-		})
-
-		adapterRuns.push(m)
-	}
-
-	// 2 + 3. Synthesis + alignment: stream every canonical.jsonl, optionally augment, align,
-	// and route each labeled row directly to its split-specific JSONL (`labeled-{train,val,test}.
-	// jsonl`). Memory cost is O(1) — the prior in-memory `splitInputs` array + `splitByIDMap`
-	// + `SplitManifest.{train,val,test}` arrays are gone; per-row split is decided inline via
-	// `splitForRow` (a pure function of source_id + region + holdout policy).
-	const labeledPaths: Record<SplitName, string> = {
-		train: join(intermediateDir, "labeled-train.jsonl"),
-		val: join(intermediateDir, "labeled-val.jsonl"),
-		test: join(intermediateDir, "labeled-test.jsonl"),
-	}
-
-	const labeledStreams: Record<SplitName, WriteStream> = {
-		train: createWriteStream(labeledPaths.train, { encoding: "utf8" }),
-		val: createWriteStream(labeledPaths.val, { encoding: "utf8" }),
-		test: createWriteStream(labeledPaths.test, { encoding: "utf8" }),
-	}
-
-	const quarantinePath = join(intermediateDir, "quarantine.jsonl")
-	const quarantineStream = createWriteStream(quarantinePath, { encoding: "utf8" })
-
-	let aligned = 0
-	let quarantined = 0
-	const counts: Record<SplitName, number> = { train: 0, val: 0, test: 0 }
-	const holdouts = defaultHoldouts()
-	// License accounting + the deliberate exclusion filter (#26). `licenseCounts` is the resolved
-	// license set (→ manifest + model-card attribution); `excludeLicenses` (empty by default → include
-	// everything) is the operator's PURPOSEFUL exclusion, never a silent drop.
-	const excludeLicenses = opts.excludeLicenses ?? []
-	const licenseCounts = new Map<string, number>()
-	let excludedByLicense = 0
-
-	const writeQuarantine = (row: CanonicalRow, reason: string): void => {
-		quarantineStream.write(`${JSON.stringify({ row, reason })}\n`)
-	}
-
-	for (const adapterRun of adapterRuns) {
-		opts.onProgress?.("align", `aligning ${adapterRun.adapter_id}`)
-
-		for await (const row of streamJSONL<CanonicalRow>(adapterRun.jsonl_path)) {
-			licenseCounts.set(row.license, (licenseCounts.get(row.license) ?? 0) + 1)
-
-			// Deliberate license exclusion (#26): drop a row ONLY when the operator named its license
-			// kind via `excludeLicenses`. Default (no patterns) keeps everything — exclusion is a
-			// purposeful act, not a silent default. Counted BEFORE the drop so the manifest's license
-			// set reflects what the corpus actually CONTAINED, and `excluded_by_license` what was cut.
-			if (licenseExcluded(row.license, excludeLicenses)) {
-				excludedByLicense++
-
-				continue
-			}
-
-			const fanned: CanonicalRow[] = [row]
-
-			if (synthesize) {
-				for (const aug of synthesizeRow(row, defaultAugmentationsForCountry(row.country))) {
-					fanned.push(aug)
-				}
-			}
-
-			for (const r of fanned) {
-				let result: ReturnType<typeof alignRow>
-
-				try {
-					result = alignRow(r)
-				} catch (error) {
-					// Last-resort robustness (2026-06-12): no single row may crash a multi-hour build.
-					// alignRow's targeted paths normalize/quarantine known issues with specific reasons;
-					// this catches any UNKNOWN throw (e.g. assertSpanInvariants on an unforeseen span
-					// shape) → quarantine + continue. A spike in `align-threw` reasons is a finding.
-					writeQuarantine(r, `align-threw:${(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++
-				}
-			}
-		}
-	}
-
-	for (const s of Object.values(labeledStreams)) {
-		s.end()
-	}
-
-	quarantineStream.end()
-	await Promise.all([...Object.values(labeledStreams).map(streamEnd), streamEnd(quarantineStream)])
-
-	// 4. Splits — manifest derived by streaming the per-split labeled files; no in-memory
-	// source-id arrays. `sort(1)` from coreutils produces the deterministic per-split .txt
-	// 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,
-		corpusVersion: opts.corpusVersion,
-		counts,
-		holdouts,
-	})
-
-	// 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),
-			val: streamJSONL<LabeledRow>(labeledPaths.val),
-			test: streamJSONL<LabeledRow>(labeledPaths.test),
-		},
-		{
-			outputDir: opts.outputDir,
-			corpusVersion: opts.corpusVersion,
-			rowsPerShard,
-		}
-	)
-
-	// 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()].toSorted((a, b) => b[1] - a[1])
-
-	opts.onProgress?.(
-		"manifest",
-		`license set: ${licenseSummary.map(([l, c]) => `${l}=${c}`).join(", ")}` +
-			(excludedByLicense > 0
-				? ` | EXCLUDED ${excludedByLicense} rows by --exclude-licenses`
-				: " | NO license exclusion applied (all rows kept)")
-	)
-
-	// 6. Top-level manifest.
-	opts.onProgress?.("manifest", "writing top-level MANIFEST.json")
-
-	const manifest: BuildCorpusManifest = {
-		corpus_version: opts.corpusVersion,
-		built_at,
-		adapters: adapterRuns,
-		skipped_adapters: skipped,
-		splits: { counts: splitCounts, holdouts },
-		shards: { counts: shardManifest.counts, total_rows: shardManifest.total_rows },
-		quarantine_count: quarantined,
-		total_aligned_rows: aligned,
-		licenses: Object.fromEntries(licenseSummary),
-		excluded_by_license: excludedByLicense,
-	}
-
-	await writeFile(join(opts.outputDir, "MANIFEST.json"), `${JSON.stringify(manifest, null, 2)}\n`, "utf8")
-
-	return manifest
-}
-
-async function* streamJSONL<T>(path: string): AsyncIterable<T> {
-	// JSONSpliterator yields already-parsed rows (skipEmpty is on by default, so blank
-	// lines are dropped at the row level) and throws SyntaxError on a malformed row —
-	// same fail-loud behavior as the prior readline + bare `JSON.parse`.
-	yield* JSONSpliterator.fromAsync<T>(path)
-}
-
-function streamEnd(s: WriteStream): Promise<void> {
-	return new Promise((resolve, reject) => {
-		s.once("close", resolve)
-		s.once("error", reject)
-	})
-}

package/src/index.ts

@@ -1,14 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- */
-
-export * from "./adapters/utils/index.ts"
-export * from "./adapters/gnaf/assemble.ts"
-export * from "./adapters/index.ts"
-export * from "./build.ts"
-export * from "./runner.ts"
-export * from "./parquet-wrapper/index.ts"
-export * from "./shard-recipes/index.ts"
-export * from "@mailwoman/corpus/types"

package/src/name-prone-us-suffixes.ts

@@ -1,12 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- * Re-export shim: the name-prone set moved to `@mailwoman/codex/us` (data in
- * `codex/us/street-suffix.json`, 2026-08-10) so the golden relabel flags, the #1569 shard
- * recipe, AND the Python relabel pass (via the `gazetteer affix-relabel` v2 lexicon) all read
- * one record. Import from `@mailwoman/codex/us` in new code.
- */
-
-export { NAME_PRONE_US_SUFFIXES } from "@mailwoman/codex/us"

package/src/parquet-wrapper/index.ts

@@ -1,9 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- */
-
-export * from "./reader.ts"
-export * from "./schema.ts"
-export * from "./writer.ts"

package/src/parquet-wrapper/reader.ts

@@ -1,88 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   Typed wrapper around `@dsnp/parquetjs`'s `ParquetReader` that narrows the row-iterator generic to
- *   a user-supplied record type and adds `AsyncDisposable` support so `await using` cleans up the
- *   envelope reader without an explicit `close()`.
- */
-
-import { ParquetReader as BaseParquetReader } from "@dsnp/parquetjs"
-import type { BufferReaderOptions } from "@dsnp/parquetjs/dist/lib/bufferReader.js"
-import { ParquetEnvelopeReader } from "@dsnp/parquetjs/dist/lib/reader.js"
-
-import type { ParquetSchema, ParquetRecordLike } from "./schema.ts"
-
-/**
- * A typed Parquet reader, wrapping the base Parquet reader.
- */
-export class ParquetReader<T extends ParquetRecordLike> extends BaseParquetReader implements AsyncDisposable {
-	declare schema: ParquetSchema<T>
-
-	static override async openFile<T extends ParquetRecordLike>(
-		filePath: string | URL,
-		options?: BufferReaderOptions
-	): Promise<ParquetReader<T>> {
-		const envelopeReader = await ParquetEnvelopeReader.openFile(filePath.toString(), options)
-
-		return ParquetReader.openEnvelopeReader<T>(envelopeReader, options)
-	}
-
-	static override async openBuffer<T extends ParquetRecordLike>(buffer: Buffer, options?: BufferReaderOptions) {
-		const envelopeReader = await ParquetEnvelopeReader.openBuffer(buffer, options)
-
-		return this.openEnvelopeReader<T>(envelopeReader, options)
-	}
-
-	static override async openEnvelopeReader<T extends ParquetRecordLike>(
-		envelopeReader: ParquetEnvelopeReader,
-		opts?: BufferReaderOptions
-	) {
-		if (opts?.metadata) {
-			return new ParquetReader<T>(opts.metadata, envelopeReader, opts)
-		}
-
-		try {
-			await envelopeReader.readHeader()
-
-			const metadata = await envelopeReader.readFooter()
-
-			return new ParquetReader<T>(metadata, envelopeReader, opts)
-		} catch (error) {
-			await envelopeReader.close()
-			throw error
-		}
-	}
-
-	public override [Symbol.asyncIterator](): AsyncGenerator<T, void, unknown> {
-		return super[Symbol.asyncIterator]() as AsyncGenerator<T, void, unknown>
-	}
-
-	/**
-	 * Iterate a subset of the columns, narrowed to the keys asked for.
-	 *
-	 * Parquet is columnar, so a projection is read avoided rather than read-then-discarded: on a 14-column corpus shard,
-	 * two columns cost 93 ms against 253 ms for the whole row. The default iterator reads every column, which is what a
-	 * caller wanting the whole record should use.
-	 */
-	public async *project<K extends keyof T>(...columns: K[]): AsyncGenerator<Pick<T, K>, void, unknown> {
-		const cursor = this.getCursor(columns.map((column) => [column as string]))
-
-		for (;;) {
-			const row = (await cursor.next()) as Pick<T, K> | null
-
-			if (!row) return
-
-			yield row
-		}
-	}
-
-	public async [Symbol.asyncDispose]() {
-		return this.close()
-	}
-
-	public async dispose() {
-		return this[Symbol.asyncDispose]()
-	}
-}

package/src/parquet-wrapper/schema.ts

@@ -1,88 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   Typed wrapper around `@dsnp/parquetjs`'s schema definition. Adds:
- *
- *   - `ParquetSchema<T>`: a generic class narrowing the base schema's `schema` property to a
- *       field-by-field typed dict.
- *   - `ParquetSchemaDefinitionCache`: an LRU lookup so hot paths that compute the same schema
- *       repeatedly pay the cost once. Implements `Disposable` so `using` works.
- *   - `createBloomFilters<T>`: helper that takes a schema and a list of columns and returns the
- *       `@dsnp/parquetjs`-shaped bloom-filter spec array.
- *
- *   `Symbol.dispose` is sync (the original was async, but `Disposable`'s contract is sync — async
- *   cleanup belongs on `AsyncDisposable`).
- */
-
-import { ParquetSchema as BaseParquetSchema } from "@dsnp/parquetjs"
-import type { createSBBFParams as BloomFilterCreation } from "@dsnp/parquetjs/dist/lib/bloomFilterIO/bloomFilterWriter.js"
-import type { FieldDefinition } from "@dsnp/parquetjs/dist/lib/declare.js"
-import { LRUCache } from "lru-cache"
-
-/**
- * A Parquet record-like object, i.e. a record with string keys and JSON-serializable values.
- */
-/**
- * Shape a row type must satisfy. Declare yours as a `type` alias — TypeScript gives an alias an implicit index
- * signature and an `interface` none, so an otherwise-identical interface does not satisfy this.
- */
-export type ParquetRecordLike = Record<string, unknown | undefined>
-
-/**
- * Typed Parquet schema definition.
- */
-export type ParquetSchemaDefinition<T = ParquetRecordLike> = Record<Extract<keyof T, string>, FieldDefinition>
-
-/**
- * Typed Parquet schema.
- */
-export class ParquetSchema<T> extends BaseParquetSchema {
-	declare schema: ParquetSchemaDefinition<T>
-}
-
-/**
- * Given a Parquet schema and a list of columns, create a list of Bloom filters for those columns.
- */
-export function createBloomFilters<T>(
-	parquetSchemaDef: ParquetSchemaDefinition<T>,
-	columns: Extract<keyof T, string>[]
-) {
-	const bloomFilters: BloomFilterCreation[] = []
-
-	for (const column of columns) {
-		if (!parquetSchemaDef[column]) {
-			throw new Error(`Bloom filter column ${column} not found in Parquet schema`)
-		}
-
-		bloomFilters.push({ column })
-	}
-
-	return bloomFilters
-}
-
-export class ParquetSchemaDefinitionCache
-	extends LRUCache<ParquetSchemaDefinition<ParquetRecordLike>, ParquetSchema<ParquetRecordLike>>
-	implements Disposable
-{
-	constructor(max = 1000) {
-		super({ max })
-	}
-
-	public findOrCreateSchema<T extends ParquetRecordLike>(schemaDef: ParquetSchemaDefinition<T>): ParquetSchema<T> {
-		const key = schemaDef as ParquetSchemaDefinition<ParquetRecordLike>
-		let schema = this.get(key) as ParquetSchema<T> | undefined
-
-		if (!schema) {
-			schema = new ParquetSchema<T>(schemaDef)
-			this.set(key, schema as ParquetSchema<ParquetRecordLike>)
-		}
-
-		return schema
-	}
-
-	public [Symbol.dispose]() {
-		this.clear()
-	}
-}