@mailwoman/corpus 9.2.0 → 9.3.0

package/out/src/utils/parquet.d.ts

@@ -1,162 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   Final output sharder for the corpus pipeline.
- *
- *   Phase 1 (#9) shipped JSONL shards + a Python (PyArrow) converter as the path to binary Parquet —
- *   bridging until the JS toolchain caught up. Phase 1.5 (#18 §4) replaced that with a native JS
- *   writer. The build pipeline no longer touches Python at all in its hot path; the only remaining
- *   Python is the one-shot `train_tokenizer.py` SentencePiece step.
- *
- *   Compression: `SNAPPY`. The plan in #18 §4 specified `zstd`, but `@dsnp/parquetjs` 1.7.0 only
- *   supports UNCOMPRESSED / GZIP / SNAPPY / BROTLI (see `node_modules/@dsnp/parquetjs/dist/lib/
- *   compression.js`). SNAPPY is the standard ML-corpus default (PyArrow's default too) and is the
- *   closest substitute on speed; revisit if @dsnp/parquetjs gains zstd support. Documented in
- *   `DECISIONS.md`.
- *
- *   Layout under `<outputDir>`:
- *
- *   ```
- *   corpus-v<version>/
- *   MANIFEST.json
- *   train/
- *     part-0000.parquet
- *     part-0001.parquet
- *     ...
- *   val/
- *     part-0000.parquet
- *   test/
- *     part-0000.parquet
- * ```
- *
- *   Each shard caps at `rowsPerShard` (default 1_000_000); within a shard, parquetjs flushes row
- *   groups every `ROW_GROUP_SIZE` (50_000) rows per the issue spec. The MANIFEST captures every
- *   shard's path, row count, byte size, and SHA-256 (computed by re-reading the shard once after
- *   close — cheap relative to writing it).
- */
-import { type ParquetSchemaDefinition } from "#parquet-wrapper";
-import type { LabeledRow } from "#types";
-import type { SplitName } from "./split.ts";
-/**
- * Row groups flush at this many rows (parquetjs internal cadence within a shard).
- */
-export declare const ROW_GROUP_SIZE = 50000;
-/**
- * Snappy is the only zstd-equivalent codec available in @dsnp/parquetjs 1.7.0.
- */
-export declare const SHARD_COMPRESSION: "SNAPPY";
-/**
- * A single Parquet-style row shape. The `[key: string]: unknown` index signature is required for compatibility with
- * `ParquetRecordLike` in the wrapper — parquetjs accepts any string key on rows.
- *
- * The three columns the schema marks `optional: true` are optional here too: `appendRow` receives the row with them
- * OMITTED rather than null (see {@link appendShape}), and a reader gets back whichever of null/absent parquetjs
- * surfaces.
- */
-export interface ParquetRow {
-    raw: string;
-    tokens: readonly string[];
-    labels: readonly string[];
-    span_starts: readonly number[];
-    span_ends: readonly number[];
-    span_tags: readonly string[];
-    country: string;
-    locale?: string | null;
-    source: string;
-    source_id: string;
-    corpus_version: string;
-    license: string;
-    synth_method?: string | null;
-    synth_base_id?: string | null;
-    [key: string]: unknown;
-}
-/**
- * Column names emitted into every shard. Matches `ParquetRow`.
- */
-export declare const PARQUET_COLUMNS: readonly ["raw", "tokens", "labels", "span_starts", "span_ends", "span_tags", "country", "locale", "source", "source_id", "corpus_version", "license", "synth_method", "synth_base_id"];
-/**
- * Parquet schema for `LabeledRow` per #18 §4. Optional fields use `optional: true`; repeated UTF8 columns capture
- * tokens/labels arrays. Compression is per-column SNAPPY.
- */
-export declare const LABELED_ROW_SCHEMA: ParquetSchemaDefinition<ParquetRow>;
-/**
- * Per-shard metadata captured in `MANIFEST.json`.
- */
-export interface ShardDescriptor {
-    split: SplitName;
-    path: string;
-    format: "parquet";
-    compression: typeof SHARD_COMPRESSION;
-    rows: number;
-    bytes: number;
-    sha256: string;
-    first_source_id: string;
-    last_source_id: string;
-    /**
-     * The shard's corpus source slug, when the writer knows it. `audit.ts` prefers this over inferring the source from
-     * `first_source_id`'s prefix; `writeShards` itself writes multi-source shards and leaves it unset.
-     */
-    source?: string;
-}
-export interface ShardManifest {
-    corpus_version: string;
-    schema: readonly string[];
-    rows_per_shard: number;
-    row_group_size: number;
-    shards: ShardDescriptor[];
-    counts: Record<SplitName, number>;
-    total_rows: number;
-}
-export interface WriteShardsOptions {
-    /**
-     * Root output directory; corpus version dir is created beneath.
-     */
-    outputDir: string;
-    /**
-     * Corpus version stamped onto rows + into the output directory name.
-     */
-    corpusVersion: string;
-    /**
-     * Max rows per `.parquet` shard. Default 1_000_000 per the Phase 1 plan.
-     */
-    rowsPerShard?: number;
-}
-/**
- * Pre-partitioned labeled-row streams, one per split. Callers (`buildCorpus`) decide each row's split inline at align
- * time via `splitForRow` and route rows to the matching stream, eliminating the prior `Map<source_id, SplitName>` O(n)
- * lookup table.
- *
- * Splits with no rows can be omitted (or passed as an empty iterable); `writeShards` skips them.
- */
-export type PerSplitRows = Partial<Record<SplitName, AsyncIterable<LabeledRow>>>;
-/**
- * Project a labeled row to the Parquet schema.
- *
- * The span triple is REQUIRED here (#519): `alignRow` emits it on every labeled row, so a row arriving without it came
- * from a producer that hasn't migrated — writing it would silently drop the v0.5.0 labels from the shard (the "builders
- * before parquet = silent loss" hazard). Loud failure, naming the row, instead.
- */
-export declare function rowToParquet(row: LabeledRow): ParquetRow;
-/**
- * Project a `ParquetRow` for `appendRow`. parquetjs treats `null` as "skip" for `optional` columns; passing it
- * explicitly is fine, but cleaner to omit so the on-disk Definition Levels match what PyArrow / DuckDB / etc. produce
- * for the same logical row.
- *
- * EXPORTED because it was not, and a second producer hand-rolled its own copy that predates the v0.5.0 span triple
- * (#519). `LABELED_ROW_SCHEMA` declares those three columns `repeated` and NOT optional, so the copy wrote empty arrays
- * under a schema promising values. Every new column has to arrive here once, not once per writer.
- */
-export declare function appendShape(row: ParquetRow): ParquetRow;
-/**
- * Stream labeled rows into `.parquet` shards, one set of shards per split. Splits are processed sequentially so that
- * only one shard writer is open at a time — memory cost is bounded by the parquetjs row-group buffer (~`ROW_GROUP_SIZE
- * × row_size`), not by the labeled-row count.
- *
- * Callers pass per-split `AsyncIterable<LabeledRow>` (`PerSplitRows`); the prior `splitFor(sourceID)` callback is gone
- * because pre-partitioning at the caller eliminates the O(n) `Map<source_id, SplitName>` it required. See `buildCorpus`
- * for the new wire-up.
- */
-export declare function writeShards(perSplit: PerSplitRows, opts: WriteShardsOptions): Promise<ShardManifest>;
-//# sourceMappingURL=parquet.d.ts.map

package/out/src/utils/parquet.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"parquet.d.ts","sourceRoot":"","sources":["../../../src/utils/parquet.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AAOH,OAAO,EAAiB,KAAK,uBAAuB,EAAE,MAAM,kBAAkB,CAAA;AAC9E,OAAO,KAAK,EAAE,UAAU,EAAE,MAAM,QAAQ,CAAA;AAExC,OAAO,KAAK,EAAE,SAAS,EAAE,MAAM,YAAY,CAAA;AAE3C;;GAEG;AACH,eAAO,MAAM,cAAc,QAAS,CAAA;AAEpC;;GAEG;AACH,eAAO,MAAM,iBAAiB,EAAG,QAAiB,CAAA;AAElD;;;;;;;GAOG;AACH,MAAM,WAAW,UAAU;IAC1B,GAAG,EAAE,MAAM,CAAA;IACX,MAAM,EAAE,SAAS,MAAM,EAAE,CAAA;IACzB,MAAM,EAAE,SAAS,MAAM,EAAE,CAAA;IACzB,WAAW,EAAE,SAAS,MAAM,EAAE,CAAA;IAC9B,SAAS,EAAE,SAAS,MAAM,EAAE,CAAA;IAC5B,SAAS,EAAE,SAAS,MAAM,EAAE,CAAA;IAC5B,OAAO,EAAE,MAAM,CAAA;IACf,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACtB,MAAM,EAAE,MAAM,CAAA;IACd,SAAS,EAAE,MAAM,CAAA;IACjB,cAAc,EAAE,MAAM,CAAA;IACtB,OAAO,EAAE,MAAM,CAAA;IACf,YAAY,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IAC5B,aAAa,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IAC7B,CAAC,GAAG,EAAE,MAAM,GAAG,OAAO,CAAA;CACtB;AAED;;GAEG;AACH,eAAO,MAAM,eAAe,yLAelB,CAAA;AAMV;;;GAGG;AACH,eAAO,MAAM,kBAAkB,EAAE,uBAAuB,CAAC,UAAU,CAkBlE,CAAA;AAED;;GAEG;AACH,MAAM,WAAW,eAAe;IAC/B,KAAK,EAAE,SAAS,CAAA;IAChB,IAAI,EAAE,MAAM,CAAA;IACZ,MAAM,EAAE,SAAS,CAAA;IACjB,WAAW,EAAE,OAAO,iBAAiB,CAAA;IACrC,IAAI,EAAE,MAAM,CAAA;IACZ,KAAK,EAAE,MAAM,CAAA;IACb,MAAM,EAAE,MAAM,CAAA;IACd,eAAe,EAAE,MAAM,CAAA;IACvB,cAAc,EAAE,MAAM,CAAA;IACtB;;;OAGG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;CACf;AAED,MAAM,WAAW,aAAa;IAC7B,cAAc,EAAE,MAAM,CAAA;IACtB,MAAM,EAAE,SAAS,MAAM,EAAE,CAAA;IACzB,cAAc,EAAE,MAAM,CAAA;IACtB,cAAc,EAAE,MAAM,CAAA;IACtB,MAAM,EAAE,eAAe,EAAE,CAAA;IACzB,MAAM,EAAE,MAAM,CAAC,SAAS,EAAE,MAAM,CAAC,CAAA;IACjC,UAAU,EAAE,MAAM,CAAA;CAClB;AAED,MAAM,WAAW,kBAAkB;IAClC;;OAEG;IACH,SAAS,EAAE,MAAM,CAAA;IAEjB;;OAEG;IACH,aAAa,EAAE,MAAM,CAAA;IAErB;;OAEG;IACH,YAAY,CAAC,EAAE,MAAM,CAAA;CACrB;AAED;;;;;;GAMG;AACH,MAAM,MAAM,YAAY,GAAG,OAAO,CAAC,MAAM,CAAC,SAAS,EAAE,aAAa,CAAC,UAAU,CAAC,CAAC,CAAC,CAAA;AAEhF;;;;;;GAMG;AACH,wBAAgB,YAAY,CAAC,GAAG,EAAE,UAAU,GAAG,UAAU,CAqCxD;AAED;;;;;;;;GAQG;AACH,wBAAgB,WAAW,CAAC,GAAG,EAAE,UAAU,GAAG,UAAU,CA4BvD;AAED;;;;;;;;GAQG;AACH,wBAAsB,WAAW,CAAC,QAAQ,EAAE,YAAY,EAAE,IAAI,EAAE,kBAAkB,GAAG,OAAO,CAAC,aAAa,CAAC,CA2G1G"}

package/out/src/utils/parquet.js

@@ -1,262 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   Final output sharder for the corpus pipeline.
- *
- *   Phase 1 (#9) shipped JSONL shards + a Python (PyArrow) converter as the path to binary Parquet —
- *   bridging until the JS toolchain caught up. Phase 1.5 (#18 §4) replaced that with a native JS
- *   writer. The build pipeline no longer touches Python at all in its hot path; the only remaining
- *   Python is the one-shot `train_tokenizer.py` SentencePiece step.
- *
- *   Compression: `SNAPPY`. The plan in #18 §4 specified `zstd`, but `@dsnp/parquetjs` 1.7.0 only
- *   supports UNCOMPRESSED / GZIP / SNAPPY / BROTLI (see `node_modules/@dsnp/parquetjs/dist/lib/
- *   compression.js`). SNAPPY is the standard ML-corpus default (PyArrow's default too) and is the
- *   closest substitute on speed; revisit if @dsnp/parquetjs gains zstd support. Documented in
- *   `DECISIONS.md`.
- *
- *   Layout under `<outputDir>`:
- *
- *   ```
- *   corpus-v<version>/
- *   MANIFEST.json
- *   train/
- *     part-0000.parquet
- *     part-0001.parquet
- *     ...
- *   val/
- *     part-0000.parquet
- *   test/
- *     part-0000.parquet
- * ```
- *
- *   Each shard caps at `rowsPerShard` (default 1_000_000); within a shard, parquetjs flushes row
- *   groups every `ROW_GROUP_SIZE` (50_000) rows per the issue spec. The MANIFEST captures every
- *   shard's path, row count, byte size, and SHA-256 (computed by re-reading the shard once after
- *   close — cheap relative to writing it).
- */
-import { mkdir, stat, writeFile } from "node:fs/promises";
-import { join } from "node:path";
-import { sha256File } from "@mailwoman/core/utils";
-import { ParquetWriter } from "#parquet-wrapper";
-/**
- * Row groups flush at this many rows (parquetjs internal cadence within a shard).
- */
-export const ROW_GROUP_SIZE = 50_000;
-/**
- * Snappy is the only zstd-equivalent codec available in @dsnp/parquetjs 1.7.0.
- */
-export const SHARD_COMPRESSION = "SNAPPY";
-/**
- * Column names emitted into every shard. Matches `ParquetRow`.
- */
-export const PARQUET_COLUMNS = [
-    "raw",
-    "tokens",
-    "labels",
-    "span_starts",
-    "span_ends",
-    "span_tags",
-    "country",
-    "locale",
-    "source",
-    "source_id",
-    "corpus_version",
-    "license",
-    "synth_method",
-    "synth_base_id",
-];
-/* oxlint-disable unicorn/text-encoding-identifier-case -- `"UTF8"` below is a ParquetType enum member,
-   not a text-encoding identifier. Lowercasing it does not type-check against ParquetSchemaDefinition,
-   and the rule has no way to tell the two apart. */
-/**
- * Parquet schema for `LabeledRow` per #18 §4. Optional fields use `optional: true`; repeated UTF8 columns capture
- * tokens/labels arrays. Compression is per-column SNAPPY.
- */
-export const LABELED_ROW_SCHEMA = {
-    raw: { type: "UTF8", compression: SHARD_COMPRESSION },
-    tokens: { type: "UTF8", repeated: true, compression: SHARD_COMPRESSION },
-    labels: { type: "UTF8", repeated: true, compression: SHARD_COMPRESSION },
-    // v0.5.0 char-offset label spans (#519): parallel arrays over `raw` (UTF-16 code units,
-    // [start, end) exclusive-end, sorted, non-overlapping). INT32 — raw is a short address string,
-    // and INT32 round-trips as `number` where parquetjs INT64 would surface bigint.
-    span_starts: { type: "INT32", repeated: true, compression: SHARD_COMPRESSION },
-    span_ends: { type: "INT32", repeated: true, compression: SHARD_COMPRESSION },
-    span_tags: { type: "UTF8", repeated: true, compression: SHARD_COMPRESSION },
-    country: { type: "UTF8", compression: SHARD_COMPRESSION },
-    locale: { type: "UTF8", compression: SHARD_COMPRESSION, optional: true },
-    source: { type: "UTF8", compression: SHARD_COMPRESSION },
-    source_id: { type: "UTF8", compression: SHARD_COMPRESSION },
-    corpus_version: { type: "UTF8", compression: SHARD_COMPRESSION },
-    license: { type: "UTF8", compression: SHARD_COMPRESSION },
-    synth_method: { type: "UTF8", compression: SHARD_COMPRESSION, optional: true },
-    synth_base_id: { type: "UTF8", compression: SHARD_COMPRESSION, optional: true },
-};
-/**
- * Project a labeled row to the Parquet schema.
- *
- * The span triple is REQUIRED here (#519): `alignRow` emits it on every labeled row, so a row arriving without it came
- * from a producer that hasn't migrated — writing it would silently drop the v0.5.0 labels from the shard (the "builders
- * before parquet = silent loss" hazard). Loud failure, naming the row, instead.
- */
-export function rowToParquet(row) {
-    const { span_starts, span_ends, span_tags } = row;
-    if (span_starts === undefined || span_ends === undefined || span_tags === undefined) {
-        throw new Error(`rowToParquet: row is missing the char-offset span triple (#519) — ` +
-            `span_starts=${span_starts !== undefined} span_ends=${span_ends !== undefined} span_tags=${span_tags !== undefined} ` +
-            `(source=${row.source}, source_id=${row.source_id}). ` +
-            `Every parquet-bound row must carry span_starts/span_ends/span_tags; ` +
-            `producers that emit tokens/labels only have not migrated to the v0.5.0 format.`);
-    }
-    if (span_starts.length !== span_ends.length || span_starts.length !== span_tags.length) {
-        throw new Error(`rowToParquet: span triple arrays are not parallel — ` +
-            `starts=${span_starts.length} ends=${span_ends.length} tags=${span_tags.length} ` +
-            `(source=${row.source}, source_id=${row.source_id})`);
-    }
-    return {
-        raw: row.raw,
-        tokens: row.tokens,
-        labels: row.labels,
-        span_starts,
-        span_ends,
-        span_tags,
-        country: row.country,
-        locale: row.locale ?? null,
-        source: row.source,
-        source_id: row.source_id,
-        corpus_version: row.corpus_version,
-        license: row.license,
-        synth_method: row.synth?.method ?? null,
-        synth_base_id: row.synth?.base_source_id ?? null,
-    };
-}
-/**
- * Project a `ParquetRow` for `appendRow`. parquetjs treats `null` as "skip" for `optional` columns; passing it
- * explicitly is fine, but cleaner to omit so the on-disk Definition Levels match what PyArrow / DuckDB / etc. produce
- * for the same logical row.
- *
- * EXPORTED because it was not, and a second producer hand-rolled its own copy that predates the v0.5.0 span triple
- * (#519). `LABELED_ROW_SCHEMA` declares those three columns `repeated` and NOT optional, so the copy wrote empty arrays
- * under a schema promising values. Every new column has to arrive here once, not once per writer.
- */
-export function appendShape(row) {
-    const out = {
-        raw: row.raw,
-        tokens: row.tokens,
-        labels: row.labels,
-        span_starts: row.span_starts,
-        span_ends: row.span_ends,
-        span_tags: row.span_tags,
-        country: row.country,
-        source: row.source,
-        source_id: row.source_id,
-        corpus_version: row.corpus_version,
-        license: row.license,
-    };
-    if (row.locale !== null) {
-        out.locale = row.locale;
-    }
-    if (row.synth_method !== null) {
-        out.synth_method = row.synth_method;
-    }
-    if (row.synth_base_id !== null) {
-        out.synth_base_id = row.synth_base_id;
-    }
-    return out;
-}
-/**
- * Stream labeled rows into `.parquet` shards, one set of shards per split. Splits are processed sequentially so that
- * only one shard writer is open at a time — memory cost is bounded by the parquetjs row-group buffer (~`ROW_GROUP_SIZE
- * × row_size`), not by the labeled-row count.
- *
- * Callers pass per-split `AsyncIterable<LabeledRow>` (`PerSplitRows`); the prior `splitFor(sourceID)` callback is gone
- * because pre-partitioning at the caller eliminates the O(n) `Map<source_id, SplitName>` it required. See `buildCorpus`
- * for the new wire-up.
- */
-export async function writeShards(perSplit, opts) {
-    const rowsPerShard = opts.rowsPerShard ?? 1_000_000;
-    const corpusDir = join(opts.outputDir, `corpus-v${opts.corpusVersion}`);
-    await mkdir(corpusDir, { recursive: true });
-    const shards = [];
-    const counts = { train: 0, val: 0, test: 0 };
-    let totalRows = 0;
-    for (const split of ["train", "val", "test"]) {
-        const rows = perSplit[split];
-        if (!rows)
-            continue;
-        let shardIndex = 0;
-        let writer = null;
-        let path = "";
-        let shardRows = 0;
-        let firstSourceID = "";
-        let lastSourceID = "";
-        const openShard = async () => {
-            const splitDir = join(corpusDir, split);
-            await mkdir(splitDir, { recursive: true });
-            path = join(splitDir, `part-${String(shardIndex).padStart(4, "0")}.parquet`);
-            writer = await ParquetWriter.openFile(LABELED_ROW_SCHEMA, path, {
-                rowGroupSize: ROW_GROUP_SIZE,
-            });
-            writer.setMetadata("mailwoman.corpus_version", opts.corpusVersion);
-            writer.setMetadata("mailwoman.split", split);
-            writer.setMetadata("mailwoman.shard_index", String(shardIndex));
-            shardRows = 0;
-            firstSourceID = "";
-            lastSourceID = "";
-        };
-        const closeShard = async () => {
-            if (!writer)
-                return;
-            await writer.close();
-            if (shardRows > 0) {
-                const fileStat = await stat(path);
-                const sha256 = await sha256File(path);
-                shards.push({
-                    split,
-                    path,
-                    format: "parquet",
-                    compression: SHARD_COMPRESSION,
-                    rows: shardRows,
-                    bytes: fileStat.size,
-                    sha256,
-                    first_source_id: firstSourceID,
-                    last_source_id: lastSourceID,
-                });
-            }
-            writer = null;
-        };
-        for await (const row of rows) {
-            if (!writer) {
-                await openShard();
-            }
-            const pq = rowToParquet(row);
-            await writer.appendRow(appendShape(pq));
-            if (shardRows === 0) {
-                firstSourceID = row.source_id;
-            }
-            lastSourceID = row.source_id;
-            shardRows++;
-            counts[split]++;
-            totalRows++;
-            if (shardRows >= rowsPerShard) {
-                await closeShard();
-                shardIndex++;
-            }
-        }
-        await closeShard();
-    }
-    shards.sort((a, b) => (a.split === b.split ? a.path.localeCompare(b.path) : a.split.localeCompare(b.split)));
-    const manifest = {
-        corpus_version: opts.corpusVersion,
-        schema: PARQUET_COLUMNS,
-        rows_per_shard: rowsPerShard,
-        row_group_size: ROW_GROUP_SIZE,
-        shards,
-        counts,
-        total_rows: totalRows,
-    };
-    await writeFile(join(corpusDir, "MANIFEST.json"), `${JSON.stringify(manifest, null, 2)}\n`, "utf8");
-    return manifest;
-}
-//# sourceMappingURL=parquet.js.map

package/out/src/utils/parquet.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"parquet.js","sourceRoot":"","sources":["../../../src/utils/parquet.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GAqCG;AAEH,OAAO,EAAE,KAAK,EAAE,IAAI,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAA;AACzD,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAA;AAEhC,OAAO,EAAE,UAAU,EAAE,MAAM,uBAAuB,CAAA;AAElD,OAAO,EAAE,aAAa,EAAgC,MAAM,kBAAkB,CAAA;AAK9E;;GAEG;AACH,MAAM,CAAC,MAAM,cAAc,GAAG,MAAM,CAAA;AAEpC;;GAEG;AACH,MAAM,CAAC,MAAM,iBAAiB,GAAG,QAAiB,CAAA;AA4BlD;;GAEG;AACH,MAAM,CAAC,MAAM,eAAe,GAAG;IAC9B,KAAK;IACL,QAAQ;IACR,QAAQ;IACR,aAAa;IACb,WAAW;IACX,WAAW;IACX,SAAS;IACT,QAAQ;IACR,QAAQ;IACR,WAAW;IACX,gBAAgB;IAChB,SAAS;IACT,cAAc;IACd,eAAe;CACN,CAAA;AAEV;;oDAEoD;AAEpD;;;GAGG;AACH,MAAM,CAAC,MAAM,kBAAkB,GAAwC;IACtE,GAAG,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE;IACrD,MAAM,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,IAAI,EAAE,WAAW,EAAE,iBAAiB,EAAE;IACxE,MAAM,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,IAAI,EAAE,WAAW,EAAE,iBAAiB,EAAE;IACxE,wFAAwF;IACxF,+FAA+F;IAC/F,gFAAgF;IAChF,WAAW,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,WAAW,EAAE,iBAAiB,EAAE;IAC9E,SAAS,EAAE,EAAE,IAAI,EAAE,OAAO,EAAE,QAAQ,EAAE,IAAI,EAAE,WAAW,EAAE,iBAAiB,EAAE;IAC5E,SAAS,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,QAAQ,EAAE,IAAI,EAAE,WAAW,EAAE,iBAAiB,EAAE;IAC3E,OAAO,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE;IACzD,MAAM,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE,QAAQ,EAAE,IAAI,EAAE;IACxE,MAAM,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE;IACxD,SAAS,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE;IAC3D,cAAc,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE;IAChE,OAAO,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE;IACzD,YAAY,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE,QAAQ,EAAE,IAAI,EAAE;IAC9E,aAAa,EAAE,EAAE,IAAI,EAAE,MAAM,EAAE,WAAW,EAAE,iBAAiB,EAAE,QAAQ,EAAE,IAAI,EAAE;CAC/E,CAAA;AA0DD;;;;;;GAMG;AACH,MAAM,UAAU,YAAY,CAAC,GAAe;IAC3C,MAAM,EAAE,WAAW,EAAE,SAAS,EAAE,SAAS,EAAE,GAAG,GAAG,CAAA;IAEjD,IAAI,WAAW,KAAK,SAAS,IAAI,SAAS,KAAK,SAAS,IAAI,SAAS,KAAK,SAAS,EAAE,CAAC;QACrF,MAAM,IAAI,KAAK,CACd,oEAAoE;YACnE,eAAe,WAAW,KAAK,SAAS,cAAc,SAAS,KAAK,SAAS,cAAc,SAAS,KAAK,SAAS,GAAG;YACrH,WAAW,GAAG,CAAC,MAAM,eAAe,GAAG,CAAC,SAAS,KAAK;YACtD,sEAAsE;YACtE,gFAAgF,CACjF,CAAA;IACF,CAAC;IAED,IAAI,WAAW,CAAC,MAAM,KAAK,SAAS,CAAC,MAAM,IAAI,WAAW,CAAC,MAAM,KAAK,SAAS,CAAC,MAAM,EAAE,CAAC;QACxF,MAAM,IAAI,KAAK,CACd,sDAAsD;YACrD,UAAU,WAAW,CAAC,MAAM,SAAS,SAAS,CAAC,MAAM,SAAS,SAAS,CAAC,MAAM,GAAG;YACjF,WAAW,GAAG,CAAC,MAAM,eAAe,GAAG,CAAC,SAAS,GAAG,CACrD,CAAA;IACF,CAAC;IAED,OAAO;QACN,GAAG,EAAE,GAAG,CAAC,GAAG;QACZ,MAAM,EAAE,GAAG,CAAC,MAAM;QAClB,MAAM,EAAE,GAAG,CAAC,MAAM;QAClB,WAAW;QACX,SAAS;QACT,SAAS;QACT,OAAO,EAAE,GAAG,CAAC,OAAO;QACpB,MAAM,EAAE,GAAG,CAAC,MAAM,IAAI,IAAI;QAC1B,MAAM,EAAE,GAAG,CAAC,MAAM;QAClB,SAAS,EAAE,GAAG,CAAC,SAAS;QACxB,cAAc,EAAE,GAAG,CAAC,cAAc;QAClC,OAAO,EAAE,GAAG,CAAC,OAAO;QACpB,YAAY,EAAE,GAAG,CAAC,KAAK,EAAE,MAAM,IAAI,IAAI;QACvC,aAAa,EAAE,GAAG,CAAC,KAAK,EAAE,cAAc,IAAI,IAAI;KAChD,CAAA;AACF,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,UAAU,WAAW,CAAC,GAAe;IAC1C,MAAM,GAAG,GAAe;QACvB,GAAG,EAAE,GAAG,CAAC,GAAG;QACZ,MAAM,EAAE,GAAG,CAAC,MAAM;QAClB,MAAM,EAAE,GAAG,CAAC,MAAM;QAClB,WAAW,EAAE,GAAG,CAAC,WAAW;QAC5B,SAAS,EAAE,GAAG,CAAC,SAAS;QACxB,SAAS,EAAE,GAAG,CAAC,SAAS;QACxB,OAAO,EAAE,GAAG,CAAC,OAAO;QACpB,MAAM,EAAE,GAAG,CAAC,MAAM;QAClB,SAAS,EAAE,GAAG,CAAC,SAAS;QACxB,cAAc,EAAE,GAAG,CAAC,cAAc;QAClC,OAAO,EAAE,GAAG,CAAC,OAAO;KACpB,CAAA;IAED,IAAI,GAAG,CAAC,MAAM,KAAK,IAAI,EAAE,CAAC;QACzB,GAAG,CAAC,MAAM,GAAG,GAAG,CAAC,MAAM,CAAA;IACxB,CAAC;IAED,IAAI,GAAG,CAAC,YAAY,KAAK,IAAI,EAAE,CAAC;QAC/B,GAAG,CAAC,YAAY,GAAG,GAAG,CAAC,YAAY,CAAA;IACpC,CAAC;IAED,IAAI,GAAG,CAAC,aAAa,KAAK,IAAI,EAAE,CAAC;QAChC,GAAG,CAAC,aAAa,GAAG,GAAG,CAAC,aAAa,CAAA;IACtC,CAAC;IAED,OAAO,GAAG,CAAA;AACX,CAAC;AAED;;;;;;;;GAQG;AACH,MAAM,CAAC,KAAK,UAAU,WAAW,CAAC,QAAsB,EAAE,IAAwB;IACjF,MAAM,YAAY,GAAG,IAAI,CAAC,YAAY,IAAI,SAAS,CAAA;IACnD,MAAM,SAAS,GAAG,IAAI,CAAC,IAAI,CAAC,SAAS,EAAE,WAAW,IAAI,CAAC,aAAa,EAAE,CAAC,CAAA;IACvE,MAAM,KAAK,CAAC,SAAS,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAA;IAE3C,MAAM,MAAM,GAAsB,EAAE,CAAA;IACpC,MAAM,MAAM,GAA8B,EAAE,KAAK,EAAE,CAAC,EAAE,GAAG,EAAE,CAAC,EAAE,IAAI,EAAE,CAAC,EAAE,CAAA;IACvE,IAAI,SAAS,GAAG,CAAC,CAAA;IAEjB,KAAK,MAAM,KAAK,IAAI,CAAC,OAAO,EAAE,KAAK,EAAE,MAAM,CAAU,EAAE,CAAC;QACvD,MAAM,IAAI,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAA;QAE5B,IAAI,CAAC,IAAI;YAAE,SAAQ;QAEnB,IAAI,UAAU,GAAG,CAAC,CAAA;QAClB,IAAI,MAAM,GAAqC,IAAI,CAAA;QACnD,IAAI,IAAI,GAAG,EAAE,CAAA;QACb,IAAI,SAAS,GAAG,CAAC,CAAA;QACjB,IAAI,aAAa,GAAG,EAAE,CAAA;QACtB,IAAI,YAAY,GAAG,EAAE,CAAA;QAErB,MAAM,SAAS,GAAG,KAAK,IAAmB,EAAE;YAC3C,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,KAAK,CAAC,CAAA;YACvC,MAAM,KAAK,CAAC,QAAQ,EAAE,EAAE,SAAS,EAAE,IAAI,EAAE,CAAC,CAAA;YAC1C,IAAI,GAAG,IAAI,CAAC,QAAQ,EAAE,QAAQ,MAAM,CAAC,UAAU,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,UAAU,CAAC,CAAA;YAE5E,MAAM,GAAG,MAAM,aAAa,CAAC,QAAQ,CAAa,kBAAkB,EAAE,IAAI,EAAE;gBAC3E,YAAY,EAAE,cAAc;aAC5B,CAAC,CAAA;YAEF,MAAM,CAAC,WAAW,CAAC,0BAA0B,EAAE,IAAI,CAAC,aAAa,CAAC,CAAA;YAClE,MAAM,CAAC,WAAW,CAAC,iBAAiB,EAAE,KAAK,CAAC,CAAA;YAC5C,MAAM,CAAC,WAAW,CAAC,uBAAuB,EAAE,MAAM,CAAC,UAAU,CAAC,CAAC,CAAA;YAC/D,SAAS,GAAG,CAAC,CAAA;YACb,aAAa,GAAG,EAAE,CAAA;YAClB,YAAY,GAAG,EAAE,CAAA;QAClB,CAAC,CAAA;QAED,MAAM,UAAU,GAAG,KAAK,IAAmB,EAAE;YAC5C,IAAI,CAAC,MAAM;gBAAE,OAAM;YACnB,MAAM,MAAM,CAAC,KAAK,EAAE,CAAA;YAEpB,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;gBACnB,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,IAAI,CAAC,CAAA;gBACjC,MAAM,MAAM,GAAG,MAAM,UAAU,CAAC,IAAI,CAAC,CAAA;gBAErC,MAAM,CAAC,IAAI,CAAC;oBACX,KAAK;oBACL,IAAI;oBACJ,MAAM,EAAE,SAAS;oBACjB,WAAW,EAAE,iBAAiB;oBAC9B,IAAI,EAAE,SAAS;oBACf,KAAK,EAAE,QAAQ,CAAC,IAAI;oBACpB,MAAM;oBACN,eAAe,EAAE,aAAa;oBAC9B,cAAc,EAAE,YAAY;iBAC5B,CAAC,CAAA;YACH,CAAC;YAED,MAAM,GAAG,IAAI,CAAA;QACd,CAAC,CAAA;QAED,IAAI,KAAK,EAAE,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;YAC9B,IAAI,CAAC,MAAM,EAAE,CAAC;gBACb,MAAM,SAAS,EAAE,CAAA;YAClB,CAAC;YAED,MAAM,EAAE,GAAG,YAAY,CAAC,GAAG,CAAC,CAAA;YAC5B,MAAM,MAAO,CAAC,SAAS,CAAC,WAAW,CAAC,EAAE,CAAC,CAAC,CAAA;YAExC,IAAI,SAAS,KAAK,CAAC,EAAE,CAAC;gBACrB,aAAa,GAAG,GAAG,CAAC,SAAS,CAAA;YAC9B,CAAC;YAED,YAAY,GAAG,GAAG,CAAC,SAAS,CAAA;YAE5B,SAAS,EAAE,CAAA;YAEX,MAAM,CAAC,KAAK,CAAC,EAAE,CAAA;YAEf,SAAS,EAAE,CAAA;YAEX,IAAI,SAAS,IAAI,YAAY,EAAE,CAAC;gBAC/B,MAAM,UAAU,EAAE,CAAA;gBAElB,UAAU,EAAE,CAAA;YACb,CAAC;QACF,CAAC;QAED,MAAM,UAAU,EAAE,CAAA;IACnB,CAAC;IAED,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,CAAC,KAAK,KAAK,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAC,CAAC,IAAI,CAAC,aAAa,CAAC,CAAC,CAAC,IAAI,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,aAAa,CAAC,CAAC,CAAC,KAAK,CAAC,CAAC,CAAC,CAAA;IAE5G,MAAM,QAAQ,GAAkB;QAC/B,cAAc,EAAE,IAAI,CAAC,aAAa;QAClC,MAAM,EAAE,eAAe;QACvB,cAAc,EAAE,YAAY;QAC5B,cAAc,EAAE,cAAc;QAC9B,MAAM;QACN,MAAM;QACN,UAAU,EAAE,SAAS;KACrB,CAAA;IAED,MAAM,SAAS,CAAC,IAAI,CAAC,SAAS,EAAE,eAAe,CAAC,EAAE,GAAG,IAAI,CAAC,SAAS,CAAC,QAAQ,EAAE,IAAI,EAAE,CAAC,CAAC,IAAI,EAAE,MAAM,CAAC,CAAA;IAEnG,OAAO,QAAQ,CAAA;AAChB,CAAC"}

package/out/src/utils/split.d.ts

@@ -1,127 +0,0 @@
-/**
- * @copyright Sister Software
- * @license AGPL-3.0
- * @author Teffen Ellis, et al.
- *
- *   Train / val / test split with **locality holdout** per the Phase 1 plan.
- *
- *   The corpus's val + test sets are not randomly sampled rows — they're entire low-density regions
- *   held out so the model cannot memorize them at training time. Rationale (per the plan's "Common
- *   pitfalls" section): random splits leak by neighborhood — a model fed "13 Main St, Springfield,
- *   IL" in train and "15 Main St, Springfield, IL" in test generalizes via region/locality
- *   memorization, not by learning the underlying schema.
- *
- *   Phase 1 holdouts (chosen for low data density + administrative isolation):
- *
- *   - **US**: Vermont, Wyoming, North Dakota
- *   - **FR**: Corse, Lozère, Creuse
- *
- *   Held-out rows are deterministically split 50/50 between val and test by hashing the row's
- *   `source_id`. Non-held-out rows go to train. The 90/5/5 ratio is approximate — what matters is
- *   the locality boundary, not the exact split percentages.
- *
- *   The output is a `SplitManifest`: three `string[]` arrays of `source_id`. Manifests live in git
- *   (under `corpus/splits/<version>/`) so reruns are reproducible bit-for-bit.
- */
-import type { CanonicalRow, LabeledRow } from "#types";
-export type SplitName = "train" | "val" | "test";
-export interface SplitOptions {
-    /**
-     * Region-name → holdout policy, keyed by ISO 3166-1 alpha-2 country. The values are the region-component strings the
-     * splitter looks for in `row.components.region`. Override to change the holdout for an experiment; defaults to
-     * `defaultHoldouts()`.
-     */
-    holdouts?: Record<string, readonly string[]>;
-}
-/**
- * Output manifest: source_id lists per split.
- */
-export interface SplitManifest {
-    train: string[];
-    val: string[];
-    test: string[];
-    /**
-     * Echoes the holdouts used, so the manifest is self-describing.
-     */
-    holdouts: Record<string, readonly string[]>;
-    /**
-     * Corpus version stamped onto the manifest. Read from the first row.
-     */
-    corpus_version: string;
-    /**
-     * Counts for quick sanity checks.
-     */
-    counts: {
-        train: number;
-        val: number;
-        test: number;
-        total: number;
-    };
-}
-/**
- * Phase 1 default holdouts (per plan).
- *
- * - US: Vermont, Wyoming, North Dakota (low density, easy to identify in WOF/admin sources).
- * - FR: Corse, Lozère, Creuse (small departments / regions).
- * - DE (added 2026-06-11, night-11): Saarland + Mecklenburg-Vorpommern — small Länder so the training cost is low while
- *   the slice clears the honest-eval 1000-row trust floor. DE has had NO trustable honest-eval slice since the harness
- *   shipped (flagged 2026-06-08); this takes effect at the NEXT base corpus rebuild — existing versioned corpora keep
- *   their committed SPLIT_MANIFESTs (a holdout added after a corpus is built is leakage-laundering, not a holdout).
- */
-export declare function defaultHoldouts(): Record<string, readonly string[]>;
-type SplitInputRow = Pick<CanonicalRow, "source_id" | "country" | "corpus_version" | "components">;
-/**
- * Pure per-row split decision. Used by both the in-memory `splitRows` and by the streaming `buildCorpus` align loop
- * (`build.ts`) to decide each row's split without retaining the row in heap. Identical hash bucketing semantics to the
- * array-based path so the decision is stable regardless of caller.
- */
-export declare function splitForRow(row: Pick<SplitInputRow, "source_id" | "country" | "components">, holdouts?: Record<string, readonly string[]>): SplitName;
-/**
- * Compute a `SplitManifest` from an iterable of labeled (or canonical) rows. Both shapes are accepted — only
- * `source_id`, `country`, `corpus_version`, and `components.region` are consulted.
- *
- * Retained for in-memory callers (tests; small-scale fixture runs). Real-data builds via `buildCorpus` use the
- * streaming path (`splitForRow` + `writeSplitManifestsFromLabeledFiles`) to avoid materializing every aligned row's
- * split membership in heap.
- */
-export declare function splitRows(rows: Iterable<SplitInputRow>, opts?: SplitOptions): SplitManifest;
-/**
- * Lightweight deterministic 0..(n-1) bucket from a string id.
- */
-/**
- * Deterministic bucket for a stable id.
- *
- * Stays on raw `createHash` rather than `sha256Hex` from `@mailwoman/core/utils`: it needs the digest BYTES, and the
- * shared helper returns hex. Re-parsing hex back into bytes to reach the same four octets would cost more than the one
- * line it saves.
- */
-export declare function hashBucket(id: string, n: number): number;
-/**
- * Write a `SplitManifest` to `<outputDir>/{train,val,test}.json`. The manifests are line-separated source_id lists (one
- * id per line) so they diff cleanly in git. Also writes `<outputDir>/MANIFEST.json` with the full structured manifest
- * including holdouts + counts + corpus version.
- *
- * Reruns produce byte-identical files (the underlying `splitRows` is deterministic).
- */
-export declare function writeSplitManifests(manifest: SplitManifest, outputDir: string): Promise<void>;
-/**
- * Type re-export for callers that want to ingest LabeledRow specifically.
- */
-export type SplitInputLabeledRow = Pick<LabeledRow, "source_id" | "country" | "corpus_version" | "components">;
-/**
- * Streaming variant of `writeSplitManifests`: derives the per-split source-id .txt manifests + `SPLIT_MANIFEST.json` by
- * streaming three per-split labeled-row JSONL files (one per split). Memory cost is O(1) — `sort(1)` from coreutils
- * handles the deterministic sort with disk spill for files that exceed in-memory thresholds.
- *
- * Used by `buildCorpus` after the align loop has already partitioned labeled rows into `labeled-{train,val,test}.jsonl`
- * via `splitForRow`. Counts are pre-computed by the align loop and passed in (zero re-scan).
- */
-export declare function writeSplitManifestsFromLabeledFiles(opts: {
-    labeledPaths: Record<SplitName, string>;
-    outputDir: string;
-    corpusVersion: string;
-    counts: Record<SplitName, number>;
-    holdouts?: Record<string, readonly string[]>;
-}): Promise<SplitManifest["counts"]>;
-export {};
-//# sourceMappingURL=split.d.ts.map

package/out/src/utils/split.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"split.d.ts","sourceRoot":"","sources":["../../../src/utils/split.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AAWH,OAAO,KAAK,EAAE,YAAY,EAAE,UAAU,EAAE,MAAM,QAAQ,CAAA;AAEtD,MAAM,MAAM,SAAS,GAAG,OAAO,GAAG,KAAK,GAAG,MAAM,CAAA;AAEhD,MAAM,WAAW,YAAY;IAC5B;;;;OAIG;IACH,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,CAAC,CAAA;CAC5C;AAED;;GAEG;AACH,MAAM,WAAW,aAAa;IAC7B,KAAK,EAAE,MAAM,EAAE,CAAA;IACf,GAAG,EAAE,MAAM,EAAE,CAAA;IACb,IAAI,EAAE,MAAM,EAAE,CAAA;IACd;;OAEG;IACH,QAAQ,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,CAAC,CAAA;IAC3C;;OAEG;IACH,cAAc,EAAE,MAAM,CAAA;IACtB;;OAEG;IACH,MAAM,EAAE;QAAE,KAAK,EAAE,MAAM,CAAC;QAAC,GAAG,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAA;CACnE;AAED;;;;;;;;;GASG;AACH,wBAAgB,eAAe,IAAI,MAAM,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,CAAC,CAMnE;AAED,KAAK,aAAa,GAAG,IAAI,CAAC,YAAY,EAAE,WAAW,GAAG,SAAS,GAAG,gBAAgB,GAAG,YAAY,CAAC,CAAA;AAElG;;;;GAIG;AACH,wBAAgB,WAAW,CAC1B,GAAG,EAAE,IAAI,CAAC,aAAa,EAAE,WAAW,GAAG,SAAS,GAAG,YAAY,CAAC,EAChE,QAAQ,GAAE,MAAM,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,CAAqB,GAC7D,SAAS,CASX;AAED;;;;;;;GAOG;AACH,wBAAgB,SAAS,CAAC,IAAI,EAAE,QAAQ,CAAC,aAAa,CAAC,EAAE,IAAI,GAAE,YAAiB,GAAG,aAAa,CAiC/F;AAED;;GAEG;AACH;;;;;;GAMG;AACH,wBAAgB,UAAU,CAAC,EAAE,EAAE,MAAM,EAAE,CAAC,EAAE,MAAM,GAAG,MAAM,CAMxD;AAED;;;;;;GAMG;AACH,wBAAsB,mBAAmB,CAAC,QAAQ,EAAE,aAAa,EAAE,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,IAAI,CAAC,CAenG;AAED;;GAEG;AACH,MAAM,MAAM,oBAAoB,GAAG,IAAI,CAAC,UAAU,EAAE,WAAW,GAAG,SAAS,GAAG,gBAAgB,GAAG,YAAY,CAAC,CAAA;AAE9G;;;;;;;GAOG;AACH,wBAAsB,mCAAmC,CAAC,IAAI,EAAE;IAC/D,YAAY,EAAE,MAAM,CAAC,SAAS,EAAE,MAAM,CAAC,CAAA;IACvC,SAAS,EAAE,MAAM,CAAA;IACjB,aAAa,EAAE,MAAM,CAAA;IACrB,MAAM,EAAE,MAAM,CAAC,SAAS,EAAE,MAAM,CAAC,CAAA;IACjC,QAAQ,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,SAAS,MAAM,EAAE,CAAC,CAAA;CAC5C,GAAG,OAAO,CAAC,aAAa,CAAC,QAAQ,CAAC,CAAC,CAqBnC"}