@mailwoman/corpus 9.2.0 → 9.3.0

package/out/tools/slice-kryptonite.d.ts

@@ -0,0 +1,32 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Build a parquet slice from the DeepSeek-generated kryptonite JSONL and emit the corpus-v0.4.0
+ *   MANIFEST. corpus-v0.4.0 is a pure adapter-addition revision: it points at every slice from
+ *   v0.3.0 plus the new kryptonite slice(s). No v0.3.0 bytes are touched or re-shuffled.
+ *
+ *   See docs/engineering/reference/CORPUS_V0_4_0_GENERATION.md for the why; that doc also pins the
+ *   DeepSeek model version + prompt versions used to produce the JSONL.
+ *
+ *   Invoke via `mailwoman corpus slice kryptonite \
+ *   --jsonl /data/corpus/versioned/v0.4.0/kryptonite/canonical-kryptonite.jsonl \
+ *   --base-manifest /data/corpus/versioned/v0.3.0/corpus-v0.3.0/MANIFEST.json \
+ *   --out-dir /data/corpus/versioned/v0.4.0`
+ */
+export interface SliceKryptoniteOptions {
+    jsonl: string;
+    baseManifest: string;
+    outDir: string;
+    /**
+     * Default `"0.4.0"`.
+     */
+    corpusVersion?: string;
+    /**
+     * Default `"deepseek-kryptonite"`.
+     */
+    source?: string;
+}
+export declare function buildKryptoniteSlice(options: SliceKryptoniteOptions, report?: (line: string) => void): Promise<void>;
+//# sourceMappingURL=slice-kryptonite.d.ts.map

package/out/tools/slice-kryptonite.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"slice-kryptonite.d.ts","sourceRoot":"","sources":["../../lib/tools/slice-kryptonite.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAWH,MAAM,WAAW,sBAAsB;IACtC,KAAK,EAAE,MAAM,CAAA;IACb,YAAY,EAAE,MAAM,CAAA;IACpB,MAAM,EAAE,MAAM,CAAA;IACd;;OAEG;IACH,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;OAEG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;CACf;AAiCD,wBAAsB,oBAAoB,CACzC,OAAO,EAAE,sBAAsB,EAC/B,MAAM,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,GAC7B,OAAO,CAAC,IAAI,CAAC,CA6Df"}

package/out/tools/slice-kryptonite.js

@@ -0,0 +1,99 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Build a parquet slice from the DeepSeek-generated kryptonite JSONL and emit the corpus-v0.4.0
+ *   MANIFEST. corpus-v0.4.0 is a pure adapter-addition revision: it points at every slice from
+ *   v0.3.0 plus the new kryptonite slice(s). No v0.3.0 bytes are touched or re-shuffled.
+ *
+ *   See docs/engineering/reference/CORPUS_V0_4_0_GENERATION.md for the why; that doc also pins the
+ *   DeepSeek model version + prompt versions used to produce the JSONL.
+ *
+ *   Invoke via `mailwoman corpus slice kryptonite \
+ *   --jsonl /data/corpus/versioned/v0.4.0/kryptonite/canonical-kryptonite.jsonl \
+ *   --base-manifest /data/corpus/versioned/v0.3.0/corpus-v0.3.0/MANIFEST.json \
+ *   --out-dir /data/corpus/versioned/v0.4.0`
+ */
+import { pathExists, readLocalJSONFile } from "@mailwoman/core/fs/readers";
+import { writeLocalJSONFile, writeLocalTextFile, makeDirectories } from "@mailwoman/core/fs/writers";
+import { join } from "path-ts";
+import { JSONSpliterator } from "spliterator";
+import { alignRow, PARQUET_COLUMNS, ROW_GROUP_SIZE, SLICE_COMPRESSION, writeSlices } from "#utils";
+async function* canonicalRows(jsonl, corpusVersion) {
+    for await (const raw of JSONSpliterator.fromAsync(jsonl)) {
+        // Strip sidecar underscore-prefixed fields the generator left behind for debugging.
+        const components = raw["components"];
+        yield {
+            raw: raw["raw"],
+            components,
+            country: raw["country"] ?? "US",
+            locale: raw["locale"] ?? undefined,
+            source: raw["source"] ?? "deepseek-kryptonite",
+            source_id: raw["source_id"],
+            corpus_version: corpusVersion,
+            license: raw["license"] ?? "Synthetic (DeepSeek-v4-flash, AGPL-compatible)",
+            synth: raw["synth"],
+        };
+    }
+}
+async function* labeledRows(jsonl, corpusVersion, quarantineLog) {
+    for await (const row of canonicalRows(jsonl, corpusVersion)) {
+        const result = alignRow(row);
+        if (result.kind === "labeled") {
+            yield result.row;
+        }
+        else {
+            quarantineLog.push(`${row.source_id}\t${result.row.reason}`);
+        }
+    }
+}
+export async function buildKryptoniteSlice(options, report) {
+    const corpusVersion = options.corpusVersion ?? "0.4.0";
+    const source = options.source ?? "deepseek-kryptonite";
+    if (!(await pathExists(options.jsonl)))
+        throw new Error(`jsonl not found: ${options.jsonl}`);
+    if (!(await pathExists(options.baseManifest)))
+        throw new Error(`base-manifest not found: ${options.baseManifest}`);
+    await makeDirectories(options.outDir);
+    const quarantine = [];
+    const newManifest = await writeSlices({ train: labeledRows(options.jsonl, corpusVersion, quarantine) }, { outputDir: options.outDir, corpusVersion });
+    report?.(`wrote ${newManifest.total_rows} rows into ${newManifest.slices.length} slice(s); ` +
+        `quarantined ${quarantine.length}`);
+    if (quarantine.length) {
+        const qPath = join(options.outDir, `corpus-v${corpusVersion}`, "quarantine-kryptonite.tsv");
+        await writeLocalTextFile(quarantine.join("\n") + "\n", qPath);
+        report?.(`quarantine log → ${qPath}`);
+    }
+    // Stamp the new slice's source field for audit.ts (which prefers slice.source over
+    // first_source_id-prefix inference). Without this, deepseek-kryptonite IDs would have
+    // to match a prefix in KNOWN_SOURCE_PREFIXES — we add it there too as a belt-and-braces.
+    for (const sh of newManifest.slices) {
+        sh.source = source;
+    }
+    // Compose the final corpus-v0.4.0 manifest: every slice from base + the new slice(s).
+    const base = await readLocalJSONFile(options.baseManifest);
+    const combined = {
+        corpus_version: corpusVersion,
+        schema: PARQUET_COLUMNS,
+        rows_per_slice: base.rows_per_slice,
+        row_group_size: base.row_group_size ?? ROW_GROUP_SIZE,
+        slices: [...base.slices, ...newManifest.slices],
+        counts: {
+            train: base.counts.train + (newManifest.counts.train ?? 0),
+            val: base.counts.val,
+            test: base.counts.test,
+        },
+        total_rows: base.total_rows + newManifest.total_rows,
+    };
+    // Stamp source on the legacy v0.3.0 slices too, so audit's slice.source path is the
+    // authoritative one. v0.3.0 slices mix sources; we use the first_source_id-prefix
+    // inference for them (audit.ts will re-derive on its own when slice.source is absent).
+    const combinedPath = join(options.outDir, `corpus-v${corpusVersion}`, "MANIFEST.json");
+    await writeLocalJSONFile(combined, combinedPath);
+    report?.(`wrote combined manifest → ${combinedPath}`);
+    report?.(`  total_rows=${combined.total_rows} (base=${base.total_rows}, added=${newManifest.total_rows})`);
+    report?.(`  slices=${combined.slices.length} (base=${base.slices.length}, added=${newManifest.slices.length})`);
+    report?.(`  compression=${SLICE_COMPRESSION}`);
+}
+//# sourceMappingURL=slice-kryptonite.js.map

package/out/tools/slice-kryptonite.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"slice-kryptonite.js","sourceRoot":"","sources":["../../lib/tools/slice-kryptonite.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,OAAO,EAAE,UAAU,EAAE,iBAAiB,EAAE,MAAM,4BAA4B,CAAA;AAC1E,OAAO,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,4BAA4B,CAAA;AACpG,OAAO,EAAE,IAAI,EAAE,MAAM,SAAS,CAAA;AAC9B,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAA;AAG7C,OAAO,EAAE,QAAQ,EAAE,eAAe,EAAE,cAAc,EAAE,iBAAiB,EAAE,WAAW,EAAE,MAAM,QAAQ,CAAA;AAiBlG,KAAK,SAAS,CAAC,CAAC,aAAa,CAAC,KAAa,EAAE,aAAqB;IACjE,IAAI,KAAK,EAAE,MAAM,GAAG,IAAI,eAAe,CAAC,SAAS,CAA0B,KAAK,CAAC,EAAE,CAAC;QACnF,oFAAoF;QACpF,MAAM,UAAU,GAAG,GAAG,CAAC,YAAY,CAA2B,CAAA;QAE9D,MAAM;YACL,GAAG,EAAE,GAAG,CAAC,KAAK,CAAW;YACzB,UAAU;YACV,OAAO,EAAG,GAAG,CAAC,SAAS,CAAY,IAAI,IAAI;YAC3C,MAAM,EAAG,GAAG,CAAC,QAAQ,CAAY,IAAI,SAAS;YAC9C,MAAM,EAAG,GAAG,CAAC,QAAQ,CAAY,IAAI,qBAAqB;YAC1D,SAAS,EAAE,GAAG,CAAC,WAAW,CAAW;YACrC,cAAc,EAAE,aAAa;YAC7B,OAAO,EAAG,GAAG,CAAC,SAAS,CAAY,IAAI,gDAAgD;YACvF,KAAK,EAAE,GAAG,CAAC,OAAO,CAA0B;SAC5C,CAAA;IACF,CAAC;AACF,CAAC;AAED,KAAK,SAAS,CAAC,CAAC,WAAW,CAAC,KAAa,EAAE,aAAqB,EAAE,aAAuB;IACxF,IAAI,KAAK,EAAE,MAAM,GAAG,IAAI,aAAa,CAAC,KAAK,EAAE,aAAa,CAAC,EAAE,CAAC;QAC7D,MAAM,MAAM,GAAG,QAAQ,CAAC,GAAG,CAAC,CAAA;QAE5B,IAAI,MAAM,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;YAC/B,MAAM,MAAM,CAAC,GAAG,CAAA;QACjB,CAAC;aAAM,CAAC;YACP,aAAa,CAAC,IAAI,CAAC,GAAG,GAAG,CAAC,SAAS,KAAK,MAAM,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,CAAA;QAC7D,CAAC;IACF,CAAC;AACF,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,oBAAoB,CACzC,OAA+B,EAC/B,MAA+B;IAE/B,MAAM,aAAa,GAAG,OAAO,CAAC,aAAa,IAAI,OAAO,CAAA;IACtD,MAAM,MAAM,GAAG,OAAO,CAAC,MAAM,IAAI,qBAAqB,CAAA;IAEtD,IAAI,CAAC,CAAC,MAAM,UAAU,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,oBAAoB,OAAO,CAAC,KAAK,EAAE,CAAC,CAAA;IAE5F,IAAI,CAAC,CAAC,MAAM,UAAU,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,4BAA4B,OAAO,CAAC,YAAY,EAAE,CAAC,CAAA;IAElH,MAAM,eAAe,CAAC,OAAO,CAAC,MAAM,CAAC,CAAA;IAErC,MAAM,UAAU,GAAa,EAAE,CAAA;IAE/B,MAAM,WAAW,GAAG,MAAM,WAAW,CACpC,EAAE,KAAK,EAAE,WAAW,CAAC,OAAO,CAAC,KAAK,EAAE,aAAa,EAAE,UAAU,CAAC,EAAE,EAChE,EAAE,SAAS,EAAE,OAAO,CAAC,MAAM,EAAE,aAAa,EAAE,CAC5C,CAAA;IAED,MAAM,EAAE,CACP,SAAS,WAAW,CAAC,UAAU,cAAc,WAAW,CAAC,MAAM,CAAC,MAAM,aAAa;QAClF,eAAe,UAAU,CAAC,MAAM,EAAE,CACnC,CAAA;IAED,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC;QACvB,MAAM,KAAK,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,WAAW,aAAa,EAAE,EAAE,2BAA2B,CAAC,CAAA;QAC3F,MAAM,kBAAkB,CAAC,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,IAAI,EAAE,KAAK,CAAC,CAAA;QAC7D,MAAM,EAAE,CAAC,oBAAoB,KAAK,EAAE,CAAC,CAAA;IACtC,CAAC;IAED,mFAAmF;IACnF,sFAAsF;IACtF,yFAAyF;IACzF,KAAK,MAAM,EAAE,IAAI,WAAW,CAAC,MAAM,EAAE,CAAC;QACrC,EAAE,CAAC,MAAM,GAAG,MAAM,CAAA;IACnB,CAAC;IAED,sFAAsF;IACtF,MAAM,IAAI,GAAG,MAAM,iBAAiB,CAAgB,OAAO,CAAC,YAAY,CAAC,CAAA;IAEzE,MAAM,QAAQ,GAAkB;QAC/B,cAAc,EAAE,aAAa;QAC7B,MAAM,EAAE,eAAe;QACvB,cAAc,EAAE,IAAI,CAAC,cAAc;QACnC,cAAc,EAAE,IAAI,CAAC,cAAc,IAAI,cAAc;QACrD,MAAM,EAAE,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,GAAG,WAAW,CAAC,MAAM,CAAC;QAC/C,MAAM,EAAE;YACP,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,KAAK,GAAG,CAAC,WAAW,CAAC,MAAM,CAAC,KAAK,IAAI,CAAC,CAAC;YAC1D,GAAG,EAAE,IAAI,CAAC,MAAM,CAAC,GAAG;YACpB,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,IAAI;SACtB;QACD,UAAU,EAAE,IAAI,CAAC,UAAU,GAAG,WAAW,CAAC,UAAU;KACpD,CAAA;IAED,oFAAoF;IACpF,kFAAkF;IAClF,uFAAuF;IACvF,MAAM,YAAY,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,WAAW,aAAa,EAAE,EAAE,eAAe,CAAC,CAAA;IACtF,MAAM,kBAAkB,CAAC,QAAQ,EAAE,YAAY,CAAC,CAAA;IAChD,MAAM,EAAE,CAAC,6BAA6B,YAAY,EAAE,CAAC,CAAA;IACrD,MAAM,EAAE,CAAC,gBAAgB,QAAQ,CAAC,UAAU,UAAU,IAAI,CAAC,UAAU,WAAW,WAAW,CAAC,UAAU,GAAG,CAAC,CAAA;IAC1G,MAAM,EAAE,CAAC,YAAY,QAAQ,CAAC,MAAM,CAAC,MAAM,UAAU,IAAI,CAAC,MAAM,CAAC,MAAM,WAAW,WAAW,CAAC,MAAM,CAAC,MAAM,GAAG,CAAC,CAAA;IAC/G,MAAM,EAAE,CAAC,iBAAiB,iBAAiB,EAAE,CAAC,CAAA;AAC/C,CAAC"}

package/out/tools/slice-translit.d.ts

@@ -0,0 +1,51 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Build per-script parquet slices from the DeepSeek-generated transliteration JSONL and emit the
+ *   corpus-v0.4.0 MANIFEST that combines them with the existing kryptonite + v0.3.0 slices.
+ *
+ *   Sibling to `slice-kryptonite.ts`. The two modules share the same composition pattern: take a
+ *   base MANIFEST, append new slices, write a combined MANIFEST. Differences specific to
+ *   transliteration:
+ *
+ *   - One JSONL contains rows from N target scripts (source = `deepseek-translit-<slug>`). We bucket by
+ *       `source` and write one slice per script so `audit.ts` can attribute each slice to its
+ *       synthetic source without relying on filename-prefix inference.
+ *   - Each slice is written to `train/part-translit-<slug>.parquet` (distinct from kryptonite's
+ *       `part-0000.parquet`, which v0.4.0's first builder already produced).
+ *   - Inherits the path-canonicalization fix flagged in Thread B's postmortem: v0.3.0 slice paths are
+ *       rewritten from `$MAILWOMAN_DATA_ROOT/...` to `/data/...` in the combined MANIFEST so
+ *       all paths share one container-friendly form.
+ *
+ *   See docs/engineering/reference/CORPUS_V0_4_0_GENERATION.md for prompts, model, and the
+ *   reproducibility contract.
+ *
+ *   Invoke via `mailwoman corpus slice translit \
+ *   --jsonl /data/corpus/versioned/v0.4.0/transliteration/canonical-transliteration.jsonl \
+ *   --base-manifest /data/corpus/versioned/v0.4.0/corpus-v0.4.0/MANIFEST.json \
+ *   --out-dir /data/corpus/versioned/v0.4.0`
+ */
+export interface SliceTranslitOptions {
+    jsonl: string;
+    baseManifest: string;
+    outDir: string;
+    /**
+     * Default `"0.4.0"`.
+     */
+    corpusVersion?: string;
+    /**
+     * Default `"/data/"`.
+     */
+    canonicalPathPrefix?: string;
+    /**
+     * Prefix the base manifest's slice paths currently carry, to be rewritten to
+     * {@link SliceTranslitOptions.canonicalPathPrefix}. Defaults to `mailwomanDataRoot()` with a trailing slash — the
+     * root that WROTE those paths. Pass it explicitly when translating a manifest generated under a different
+     * `$MAILWOMAN_DATA_ROOT` than the one you are running with.
+     */
+    legacyPathPrefix?: string;
+}
+export declare function buildTranslitSlice(options: SliceTranslitOptions, report?: (line: string) => void): Promise<void>;
+//# sourceMappingURL=slice-translit.d.ts.map

package/out/tools/slice-translit.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"slice-translit.d.ts","sourceRoot":"","sources":["../../lib/tools/slice-translit.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAaH,MAAM,WAAW,oBAAoB;IACpC,KAAK,EAAE,MAAM,CAAA;IACb,YAAY,EAAE,MAAM,CAAA;IACpB,MAAM,EAAE,MAAM,CAAA;IACd;;OAEG;IACH,aAAa,CAAC,EAAE,MAAM,CAAA;IACtB;;OAEG;IACH,mBAAmB,CAAC,EAAE,MAAM,CAAA;IAC5B;;;;;OAKG;IACH,gBAAgB,CAAC,EAAE,MAAM,CAAA;CACzB;AA0DD,wBAAsB,kBAAkB,CACvC,OAAO,EAAE,oBAAoB,EAC7B,MAAM,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,KAAK,IAAI,GAC7B,OAAO,CAAC,IAAI,CAAC,CAgGf"}

package/out/tools/slice-translit.js

@@ -0,0 +1,166 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Build per-script parquet slices from the DeepSeek-generated transliteration JSONL and emit the
+ *   corpus-v0.4.0 MANIFEST that combines them with the existing kryptonite + v0.3.0 slices.
+ *
+ *   Sibling to `slice-kryptonite.ts`. The two modules share the same composition pattern: take a
+ *   base MANIFEST, append new slices, write a combined MANIFEST. Differences specific to
+ *   transliteration:
+ *
+ *   - One JSONL contains rows from N target scripts (source = `deepseek-translit-<slug>`). We bucket by
+ *       `source` and write one slice per script so `audit.ts` can attribute each slice to its
+ *       synthetic source without relying on filename-prefix inference.
+ *   - Each slice is written to `train/part-translit-<slug>.parquet` (distinct from kryptonite's
+ *       `part-0000.parquet`, which v0.4.0's first builder already produced).
+ *   - Inherits the path-canonicalization fix flagged in Thread B's postmortem: v0.3.0 slice paths are
+ *       rewritten from `$MAILWOMAN_DATA_ROOT/...` to `/data/...` in the combined MANIFEST so
+ *       all paths share one container-friendly form.
+ *
+ *   See docs/engineering/reference/CORPUS_V0_4_0_GENERATION.md for prompts, model, and the
+ *   reproducibility contract.
+ *
+ *   Invoke via `mailwoman corpus slice translit \
+ *   --jsonl /data/corpus/versioned/v0.4.0/transliteration/canonical-transliteration.jsonl \
+ *   --base-manifest /data/corpus/versioned/v0.4.0/corpus-v0.4.0/MANIFEST.json \
+ *   --out-dir /data/corpus/versioned/v0.4.0`
+ */
+import { mailwomanDataRoot } from "@mailwoman/core/data-root";
+import { pathExists, readLocalJSONFile, tryStat } from "@mailwoman/core/fs/readers";
+import { writeLocalJSONFile, writeLocalTextFile, makeDirectories } from "@mailwoman/core/fs/writers";
+import { sha256File } from "@mailwoman/core/hash";
+import { join } from "path-ts";
+import { JSONSpliterator } from "spliterator";
+import { alignRow, PARQUET_COLUMNS, ROW_GROUP_SIZE, rowToParquet, SLICE_COMPRESSION, writeParquetRows } from "#utils";
+function toCanonicalRow(raw, corpusVersion) {
+    return {
+        raw: raw["raw"],
+        components: raw["components"],
+        country: raw["country"] ?? "US",
+        locale: raw["locale"] ?? undefined,
+        source: raw["source"],
+        source_id: raw["source_id"],
+        corpus_version: corpusVersion,
+        license: raw["license"] ?? "Synthetic (DeepSeek-v4-flash, AGPL-compatible)",
+        synth: raw["synth"],
+    };
+}
+/**
+ * Write one slice for a single source slug. Returns the populated SliceDescriptor + a list of quarantine reasons for
+ * rows that failed alignment.
+ */
+async function writeOneSlice(rows, outPath, source) {
+    let firstSourceID = "";
+    let lastSourceID = "";
+    for (const row of rows) {
+        if (firstSourceID === "") {
+            firstSourceID = row.source_id;
+        }
+        lastSourceID = row.source_id;
+    }
+    await writeParquetRows(rows.map(rowToParquet), outPath);
+    const fileStat = await tryStat(outPath);
+    const sha256 = await sha256File(outPath);
+    return {
+        split: "train",
+        path: outPath,
+        format: "parquet",
+        compression: SLICE_COMPRESSION,
+        rows: rows.length,
+        bytes: fileStat?.size ?? 0,
+        sha256,
+        first_source_id: firstSourceID,
+        last_source_id: lastSourceID,
+        // Stamp source so audit.ts attributes the slice without falling back to filename-prefix inference.
+        source,
+    };
+}
+function canonicalizeSlicePath(path, legacyPrefix, canonicalPrefix) {
+    if (path.startsWith(legacyPrefix))
+        return canonicalPrefix + path.slice(legacyPrefix.length);
+    return path;
+}
+export async function buildTranslitSlice(options, report) {
+    const corpusVersion = options.corpusVersion ?? "0.4.0";
+    // This is a portable manifest namespace, not a host filesystem default.
+    const canonicalPathPrefix = options.canonicalPathPrefix ?? "/data/";
+    const legacyPathPrefix = options.legacyPathPrefix ?? `${mailwomanDataRoot()}/`;
+    if (!(await pathExists(options.jsonl)))
+        throw new Error(`jsonl not found: ${options.jsonl}`);
+    if (!(await pathExists(options.baseManifest)))
+        throw new Error(`base-manifest not found: ${options.baseManifest}`);
+    const corpusDir = join(options.outDir, `corpus-v${corpusVersion}`);
+    const trainDir = join(corpusDir, "train");
+    await makeDirectories(trainDir);
+    // Bucket canonical rows by source. Quarantined rows are logged.
+    const buckets = new Map();
+    const quarantine = [];
+    let totalIn = 0;
+    for await (const raw of JSONSpliterator.fromAsync(options.jsonl)) {
+        totalIn++;
+        const canon = toCanonicalRow(raw, corpusVersion);
+        const result = alignRow(canon);
+        if (result.kind !== "labeled") {
+            quarantine.push(`${canon.source_id}\t${result.row.reason}`);
+            continue;
+        }
+        const bucket = buckets.get(canon.source);
+        if (bucket) {
+            bucket.push(result.row);
+        }
+        else {
+            buckets.set(canon.source, [result.row]);
+        }
+    }
+    report?.(`read ${totalIn} rows; ${quarantine.length} quarantined; ${buckets.size} script buckets`);
+    const newSlices = [];
+    const sortedKeys = [...buckets.keys()].toSorted();
+    for (const source of sortedKeys) {
+        const rows = buckets.get(source);
+        const slug = source.startsWith("deepseek-translit-") ? source.slice("deepseek-translit-".length) : source;
+        const outPath = join(trainDir, `part-translit-${slug}.parquet`);
+        const descriptor = await writeOneSlice(rows, outPath, source);
+        newSlices.push(descriptor);
+        report?.(`  ${source}: ${descriptor.rows} rows → ${outPath} (${descriptor.bytes} bytes)`);
+    }
+    if (quarantine.length) {
+        const qPath = join(corpusDir, "quarantine-transliteration.tsv");
+        await writeLocalTextFile(quarantine.join("\n") + "\n", qPath);
+        report?.(`quarantine log → ${qPath} (${quarantine.length} rows)`);
+    }
+    // Compose final MANIFEST: rewrite base.slices paths from /mnt/playpen/... → /data/... and append
+    // the new translit slices. Kryptonite slice already lives in the base manifest (it was written
+    // there by Thread B).
+    const base = await readLocalJSONFile(options.baseManifest);
+    const rewrittenBase = base.slices.map((sh) => ({
+        ...sh,
+        path: canonicalizeSlicePath(sh.path, legacyPathPrefix, canonicalPathPrefix),
+    }));
+    const newTrainRows = newSlices.reduce((sum, sh) => sum + sh.rows, 0);
+    const combined = {
+        corpus_version: corpusVersion,
+        schema: PARQUET_COLUMNS,
+        rows_per_slice: base.rows_per_slice,
+        row_group_size: base.row_group_size ?? ROW_GROUP_SIZE,
+        slices: [...rewrittenBase, ...newSlices],
+        counts: {
+            train: base.counts.train + newTrainRows,
+            val: base.counts.val,
+            test: base.counts.test,
+        },
+        total_rows: base.total_rows + newTrainRows,
+    };
+    const combinedPath = join(corpusDir, "MANIFEST.json");
+    await writeLocalJSONFile(combined, combinedPath);
+    report?.(`wrote combined manifest → ${combinedPath}`);
+    report?.(`  total_rows=${combined.total_rows} (base=${base.total_rows}, added=${newTrainRows})`);
+    report?.(`  slices=${combined.slices.length} (base=${base.slices.length}, added=${newSlices.length})`);
+    report?.(`  compression=${SLICE_COMPRESSION}`);
+    const pathFix = rewrittenBase.filter((s, i) => s.path !== base.slices[i].path).length;
+    if (pathFix > 0) {
+        report?.(`  path-canonicalized base slices: ${pathFix} (legacy '${legacyPathPrefix}' → '${canonicalPathPrefix}')`);
+    }
+}
+//# sourceMappingURL=slice-translit.js.map

package/out/tools/slice-translit.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"slice-translit.js","sourceRoot":"","sources":["../../lib/tools/slice-translit.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA4BG;AAEH,OAAO,EAAE,iBAAiB,EAAE,MAAM,2BAA2B,CAAA;AAC7D,OAAO,EAAE,UAAU,EAAE,iBAAiB,EAAE,OAAO,EAAE,MAAM,4BAA4B,CAAA;AACnF,OAAO,EAAE,kBAAkB,EAAE,kBAAkB,EAAE,eAAe,EAAE,MAAM,4BAA4B,CAAA;AACpG,OAAO,EAAE,UAAU,EAAE,MAAM,sBAAsB,CAAA;AACjD,OAAO,EAAE,IAAI,EAAE,MAAM,SAAS,CAAA;AAC9B,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAA;AAG7C,OAAO,EAAE,QAAQ,EAAE,eAAe,EAAE,cAAc,EAAE,YAAY,EAAE,iBAAiB,EAAE,gBAAgB,EAAE,MAAM,QAAQ,CAAA;AAwBrH,SAAS,cAAc,CAAC,GAA4B,EAAE,aAAqB;IAC1E,OAAO;QACN,GAAG,EAAE,GAAG,CAAC,KAAK,CAAW;QACzB,UAAU,EAAE,GAAG,CAAC,YAAY,CAA2B;QACvD,OAAO,EAAG,GAAG,CAAC,SAAS,CAAY,IAAI,IAAI;QAC3C,MAAM,EAAG,GAAG,CAAC,QAAQ,CAAY,IAAI,SAAS;QAC9C,MAAM,EAAE,GAAG,CAAC,QAAQ,CAAW;QAC/B,SAAS,EAAE,GAAG,CAAC,WAAW,CAAW;QACrC,cAAc,EAAE,aAAa;QAC7B,OAAO,EAAG,GAAG,CAAC,SAAS,CAAY,IAAI,gDAAgD;QACvF,KAAK,EAAE,GAAG,CAAC,OAAO,CAA0B;KAC5C,CAAA;AACF,CAAC;AAED;;;GAGG;AACH,KAAK,UAAU,aAAa,CAAC,IAA2B,EAAE,OAAe,EAAE,MAAc;IACxF,IAAI,aAAa,GAAG,EAAE,CAAA;IACtB,IAAI,YAAY,GAAG,EAAE,CAAA;IAErB,KAAK,MAAM,GAAG,IAAI,IAAI,EAAE,CAAC;QACxB,IAAI,aAAa,KAAK,EAAE,EAAE,CAAC;YAC1B,aAAa,GAAG,GAAG,CAAC,SAAS,CAAA;QAC9B,CAAC;QAED,YAAY,GAAG,GAAG,CAAC,SAAS,CAAA;IAC7B,CAAC;IAED,MAAM,gBAAgB,CAAC,IAAI,CAAC,GAAG,CAAC,YAAY,CAAC,EAAE,OAAO,CAAC,CAAA;IAEvD,MAAM,QAAQ,GAAG,MAAM,OAAO,CAAC,OAAO,CAAC,CAAA;IACvC,MAAM,MAAM,GAAG,MAAM,UAAU,CAAC,OAAO,CAAC,CAAA;IAExC,OAAO;QACN,KAAK,EAAE,OAAO;QACd,IAAI,EAAE,OAAO;QACb,MAAM,EAAE,SAAS;QACjB,WAAW,EAAE,iBAAiB;QAC9B,IAAI,EAAE,IAAI,CAAC,MAAM;QACjB,KAAK,EAAE,QAAQ,EAAE,IAAI,IAAI,CAAC;QAC1B,MAAM;QACN,eAAe,EAAE,aAAa;QAC9B,cAAc,EAAE,YAAY;QAC5B,mGAAmG;QACnG,MAAM;KACN,CAAA;AACF,CAAC;AAED,SAAS,qBAAqB,CAAC,IAAY,EAAE,YAAoB,EAAE,eAAuB;IACzF,IAAI,IAAI,CAAC,UAAU,CAAC,YAAY,CAAC;QAAE,OAAO,eAAe,GAAG,IAAI,CAAC,KAAK,CAAC,YAAY,CAAC,MAAM,CAAC,CAAA;IAE3F,OAAO,IAAI,CAAA;AACZ,CAAC;AAED,MAAM,CAAC,KAAK,UAAU,kBAAkB,CACvC,OAA6B,EAC7B,MAA+B;IAE/B,MAAM,aAAa,GAAG,OAAO,CAAC,aAAa,IAAI,OAAO,CAAA;IACtD,wEAAwE;IACxE,MAAM,mBAAmB,GAAG,OAAO,CAAC,mBAAmB,IAAI,QAAQ,CAAA;IACnE,MAAM,gBAAgB,GAAG,OAAO,CAAC,gBAAgB,IAAI,GAAG,iBAAiB,EAAE,GAAG,CAAA;IAE9E,IAAI,CAAC,CAAC,MAAM,UAAU,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,oBAAoB,OAAO,CAAC,KAAK,EAAE,CAAC,CAAA;IAE5F,IAAI,CAAC,CAAC,MAAM,UAAU,CAAC,OAAO,CAAC,YAAY,CAAC,CAAC;QAAE,MAAM,IAAI,KAAK,CAAC,4BAA4B,OAAO,CAAC,YAAY,EAAE,CAAC,CAAA;IAElH,MAAM,SAAS,GAAG,IAAI,CAAC,OAAO,CAAC,MAAM,EAAE,WAAW,aAAa,EAAE,CAAC,CAAA;IAClE,MAAM,QAAQ,GAAG,IAAI,CAAC,SAAS,EAAE,OAAO,CAAC,CAAA;IACzC,MAAM,eAAe,CAAC,QAAQ,CAAC,CAAA;IAE/B,gEAAgE;IAChE,MAAM,OAAO,GAAG,IAAI,GAAG,EAAwB,CAAA;IAC/C,MAAM,UAAU,GAAa,EAAE,CAAA;IAC/B,IAAI,OAAO,GAAG,CAAC,CAAA;IAEf,IAAI,KAAK,EAAE,MAAM,GAAG,IAAI,eAAe,CAAC,SAAS,CAA0B,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QAC3F,OAAO,EAAE,CAAA;QACT,MAAM,KAAK,GAAG,cAAc,CAAC,GAAG,EAAE,aAAa,CAAC,CAAA;QAChD,MAAM,MAAM,GAAG,QAAQ,CAAC,KAAK,CAAC,CAAA;QAE9B,IAAI,MAAM,CAAC,IAAI,KAAK,SAAS,EAAE,CAAC;YAC/B,UAAU,CAAC,IAAI,CAAC,GAAG,KAAK,CAAC,SAAS,KAAK,MAAM,CAAC,GAAG,CAAC,MAAM,EAAE,CAAC,CAAA;YAE3D,SAAQ;QACT,CAAC;QAED,MAAM,MAAM,GAAG,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,CAAC,CAAA;QAExC,IAAI,MAAM,EAAE,CAAC;YACZ,MAAM,CAAC,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAA;QACxB,CAAC;aAAM,CAAC;YACP,OAAO,CAAC,GAAG,CAAC,KAAK,CAAC,MAAM,EAAE,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,CAAA;QACxC,CAAC;IACF,CAAC;IAED,MAAM,EAAE,CAAC,QAAQ,OAAO,UAAU,UAAU,CAAC,MAAM,iBAAiB,OAAO,CAAC,IAAI,iBAAiB,CAAC,CAAA;IAElG,MAAM,SAAS,GAAsB,EAAE,CAAA;IACvC,MAAM,UAAU,GAAG,CAAC,GAAG,OAAO,CAAC,IAAI,EAAE,CAAC,CAAC,QAAQ,EAAE,CAAA;IAEjD,KAAK,MAAM,MAAM,IAAI,UAAU,EAAE,CAAC;QACjC,MAAM,IAAI,GAAG,OAAO,CAAC,GAAG,CAAC,MAAM,CAAE,CAAA;QACjC,MAAM,IAAI,GAAG,MAAM,CAAC,UAAU,CAAC,oBAAoB,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,KAAK,CAAC,oBAAoB,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAA;QACzG,MAAM,OAAO,GAAG,IAAI,CAAC,QAAQ,EAAE,iBAAiB,IAAI,UAAU,CAAC,CAAA;QAC/D,MAAM,UAAU,GAAG,MAAM,aAAa,CAAC,IAAI,EAAE,OAAO,EAAE,MAAM,CAAC,CAAA;QAC7D,SAAS,CAAC,IAAI,CAAC,UAAU,CAAC,CAAA;QAC1B,MAAM,EAAE,CAAC,KAAK,MAAM,KAAK,UAAU,CAAC,IAAI,WAAW,OAAO,KAAK,UAAU,CAAC,KAAK,SAAS,CAAC,CAAA;IAC1F,CAAC;IAED,IAAI,UAAU,CAAC,MAAM,EAAE,CAAC;QACvB,MAAM,KAAK,GAAG,IAAI,CAAC,SAAS,EAAE,gCAAgC,CAAC,CAAA;QAC/D,MAAM,kBAAkB,CAAC,UAAU,CAAC,IAAI,CAAC,IAAI,CAAC,GAAG,IAAI,EAAE,KAAK,CAAC,CAAA;QAC7D,MAAM,EAAE,CAAC,oBAAoB,KAAK,KAAK,UAAU,CAAC,MAAM,QAAQ,CAAC,CAAA;IAClE,CAAC;IAED,iGAAiG;IACjG,+FAA+F;IAC/F,sBAAsB;IACtB,MAAM,IAAI,GAAG,MAAM,iBAAiB,CAAgB,OAAO,CAAC,YAAY,CAAC,CAAA;IAEzE,MAAM,aAAa,GAAG,IAAI,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,CAAC;QAC9C,GAAG,EAAE;QACL,IAAI,EAAE,qBAAqB,CAAC,EAAE,CAAC,IAAI,EAAE,gBAAgB,EAAE,mBAAmB,CAAC;KAC3E,CAAC,CAAC,CAAA;IAEH,MAAM,YAAY,GAAG,SAAS,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,EAAE,EAAE,EAAE,CAAC,GAAG,GAAG,EAAE,CAAC,IAAI,EAAE,CAAC,CAAC,CAAA;IAEpE,MAAM,QAAQ,GAAkB;QAC/B,cAAc,EAAE,aAAa;QAC7B,MAAM,EAAE,eAAe;QACvB,cAAc,EAAE,IAAI,CAAC,cAAc;QACnC,cAAc,EAAE,IAAI,CAAC,cAAc,IAAI,cAAc;QACrD,MAAM,EAAE,CAAC,GAAG,aAAa,EAAE,GAAG,SAAS,CAAC;QACxC,MAAM,EAAE;YACP,KAAK,EAAE,IAAI,CAAC,MAAM,CAAC,KAAK,GAAG,YAAY;YACvC,GAAG,EAAE,IAAI,CAAC,MAAM,CAAC,GAAG;YACpB,IAAI,EAAE,IAAI,CAAC,MAAM,CAAC,IAAI;SACtB;QACD,UAAU,EAAE,IAAI,CAAC,UAAU,GAAG,YAAY;KAC1C,CAAA;IAED,MAAM,YAAY,GAAG,IAAI,CAAC,SAAS,EAAE,eAAe,CAAC,CAAA;IACrD,MAAM,kBAAkB,CAAC,QAAQ,EAAE,YAAY,CAAC,CAAA;IAChD,MAAM,EAAE,CAAC,6BAA6B,YAAY,EAAE,CAAC,CAAA;IACrD,MAAM,EAAE,CAAC,gBAAgB,QAAQ,CAAC,UAAU,UAAU,IAAI,CAAC,UAAU,WAAW,YAAY,GAAG,CAAC,CAAA;IAChG,MAAM,EAAE,CAAC,YAAY,QAAQ,CAAC,MAAM,CAAC,MAAM,UAAU,IAAI,CAAC,MAAM,CAAC,MAAM,WAAW,SAAS,CAAC,MAAM,GAAG,CAAC,CAAA;IACtG,MAAM,EAAE,CAAC,iBAAiB,iBAAiB,EAAE,CAAC,CAAA;IAC9C,MAAM,OAAO,GAAG,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,IAAI,KAAK,IAAI,CAAC,MAAM,CAAC,CAAC,CAAE,CAAC,IAAI,CAAC,CAAC,MAAM,CAAA;IAEtF,IAAI,OAAO,GAAG,CAAC,EAAE,CAAC;QACjB,MAAM,EAAE,CAAC,qCAAqC,OAAO,aAAa,gBAAgB,QAAQ,mBAAmB,IAAI,CAAC,CAAA;IACnH,CAAC;AACF,CAAC"}

package/out/tools/sub-venue/harvest.d.ts

@@ -0,0 +1,100 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   @file The harvest side of the sub-venue lexicon: what one row of a harvestable source looks like,
+ *   how a file of them is read back, and what a pass over them extracts — attested surfaces plus the
+ *   measured shape of each designator's identifier half.
+ *
+ *   {@link SubVenueHarvestRow} is SOURCE-NEUTRAL, and verified so: an Overture Places row from the
+ *   `airport_terminal` category is `{ designatorID, name }` and fits unchanged. It is declared here
+ *   rather than imported so the builder does not take a dependency on `@mailwoman/osm` (which
+ *   `@mailwoman/corpus` does not depend on) just to name a shape it reads from a file.
+ *
+ *   Two measurements shape {@link extractAttestedPhrases}, and both are the reason it is safe to run
+ *   over raw OSM at all:
+ *
+ *   **A feature's `name` is usually the VENUE's name, not a sub-venue phrase.** On the Berlin extract
+ *   (2,060 matched features) a `railway=platform` is named `Stendaler Straße` and an `amenity=university`
+ *   is named `Hertie School`; across 250,116 named Great Britain features only 6,003 (2.40%) contain a
+ *   designator token at all. So names are not harvested wholesale — a name contributes only when it
+ *   CONTAINS a phrase already in the surface index, which is what makes `Terminal E (Untere Ebene)`
+ *   evidence and `Otto Lilienthal Flughafen Berlin Tegel` not.
+ *
+ *   **A matched phrase belongs to the record the PHRASE names, not the record the ROW carries.**
+ *   Attributing every hit to `row.designatorID` — the rule that matched the FEATURE — produced `west →
+ *   platform`, `hall → platform` and `biggin → platform` on the GB extract, because a bus stop tagged
+ *   `public_transport=platform` is named "Village Hall" or "West Kensington"; 108 of 133 OSM-derived
+ *   surfaces named a different record than the one they pointed at. Attribution therefore runs through
+ *   the index, and the row's own designator is kept as `context` — exactly the axis a confound board
+ *   needs, since a `hall` seen on a platform is a confound and a `hall` seen on a terminal is evidence.
+ */
+import { type SurfaceIndex } from "#tools/sub-venue/surfaces";
+import type { IdentifierShape, SubVenueSurface } from "#tools/sub-venue/table";
+/**
+ * One row of a harvestable source, as read back off JSONL or out of a layer database.
+ *
+ * SOURCE-NEUTRAL by design, and verified so in wave 2: an Overture Places row from the `airport_terminal` category is
+ * `{ designatorID, name }` and fits unchanged. What did NOT fit was the harvest function's hardcoded `osm:name` source
+ * stamp — see `overture-subvenue.ts`'s docstring. Declared locally so the builder does not import `@mailwoman/osm`
+ * (which `@mailwoman/corpus` does not depend on) just to name a shape it reads from a file.
+ */
+export interface SubVenueHarvestRow {
+    /**
+     * The designator the SOURCE's rule assigned to the FEATURE. Not necessarily the record a matched phrase names — see
+     * this file's header on phrase attribution. Carried into {@link SubVenueSurface.context}.
+     */
+    designatorID: string;
+    /**
+     * `venue` (station, airport, campus) or `sub_venue`, when the source's rule assigns one. The harvest itself does not
+     * read it; the sub-venue slice recipe fills its venue slot from it.
+     */
+    tier?: string;
+    name?: string | null;
+    ref?: string | null;
+    localizedNames?: Record<string, string>;
+}
+/**
+ * Classify an OSM `ref` into an {@link IdentifierShape} class.
+ *
+ * The classes are the ones Berlin's gates actually produced, plus the two aviation forms the corpus task names
+ * (`Terminal 2F` is digit-letter, `Concourse B` is letter). `range` covers both separators OSM uses for a gate serving
+ * more than one stand: `16-18` and `0/1`.
+ */
+export declare function classifyIdentifier(ref: string): string;
+/**
+ * Options for one harvest pass — which source stamp its surfaces carry and which region they were attested in.
+ *
+ * Both default to the OSM/unknown-region values wave 1 hardcoded, so an existing caller is unchanged.
+ */
+export interface HarvestOptions {
+    /**
+     * Source family: `osm` yields `osm:name` / `osm:name:<lang>`, `overture` yields `overture:name`.
+     */
+    source?: string;
+    /**
+     * ISO 3166-1 alpha-2 of the extract or partition. `""` when unknown.
+     */
+    region?: string;
+}
+/**
+ * Harvest attested phrases and identifier shapes out of a source's rows.
+ *
+ * `index` gates the name harvest — a name contributes only when it CONTAINS a phrase already in the table. That filter
+ * is the whole reason this function is safe to run over raw OSM: see this file's header for the Berlin measurement that
+ * motivated it. The index also decides ATTRIBUTION: a hit is a surface of the record the PHRASE names, and the row's
+ * own designator is recorded as `context`.
+ *
+ * Returns surfaces with real `observations` counts, so the lexicon can rank `terminal` above a phrase attested once.
+ */
+export declare function extractAttestedPhrases(rows: Iterable<SubVenueHarvestRow>, index: SurfaceIndex, options?: HarvestOptions): {
+    surfaces: SubVenueSurface[];
+    identifierShapes: IdentifierShape[];
+};
+/**
+ * Read a JSONL file of {@link SubVenueHarvestRow}s. Blank lines and unparseable rows are skipped rather than fatal — an
+ * extract is a build output, and one malformed line should not cost the whole lexicon.
+ */
+export declare function readSubVenueJSONL(path: string): Promise<SubVenueHarvestRow[]>;
+//# sourceMappingURL=harvest.d.ts.map

package/out/tools/sub-venue/harvest.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"harvest.d.ts","sourceRoot":"","sources":["../../../lib/tools/sub-venue/harvest.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAMH,OAAO,EAAwB,KAAK,YAAY,EAAE,MAAM,2BAA2B,CAAA;AACnF,OAAO,KAAK,EAAE,eAAe,EAAE,eAAe,EAAE,MAAM,wBAAwB,CAAA;AAE9E;;;;;;;GAOG;AACH,MAAM,WAAW,kBAAkB;IAClC;;;OAGG;IACH,YAAY,EAAE,MAAM,CAAA;IACpB;;;OAGG;IACH,IAAI,CAAC,EAAE,MAAM,CAAA;IACb,IAAI,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACpB,GAAG,CAAC,EAAE,MAAM,GAAG,IAAI,CAAA;IACnB,cAAc,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAA;CACvC;AAED;;;;;;GAMG;AACH,wBAAgB,kBAAkB,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,CActD;AAYD;;;;GAIG;AACH,MAAM,WAAW,cAAc;IAC9B;;OAEG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;IACf;;OAEG;IACH,MAAM,CAAC,EAAE,MAAM,CAAA;CACf;AAED;;;;;;;;;GASG;AACH,wBAAgB,sBAAsB,CACrC,IAAI,EAAE,QAAQ,CAAC,kBAAkB,CAAC,EAClC,KAAK,EAAE,YAAY,EACnB,OAAO,GAAE,cAAmB,GAC1B;IAAE,QAAQ,EAAE,eAAe,EAAE,CAAC;IAAC,gBAAgB,EAAE,eAAe,EAAE,CAAA;CAAE,CAiFtE;AAED;;;GAGG;AACH,wBAAsB,iBAAiB,CAAC,IAAI,EAAE,MAAM,GAAG,OAAO,CAAC,kBAAkB,EAAE,CAAC,CAmBnF"}

package/out/tools/sub-venue/harvest.js

@@ -0,0 +1,168 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   @file The harvest side of the sub-venue lexicon: what one row of a harvestable source looks like,
+ *   how a file of them is read back, and what a pass over them extracts — attested surfaces plus the
+ *   measured shape of each designator's identifier half.
+ *
+ *   {@link SubVenueHarvestRow} is SOURCE-NEUTRAL, and verified so: an Overture Places row from the
+ *   `airport_terminal` category is `{ designatorID, name }` and fits unchanged. It is declared here
+ *   rather than imported so the builder does not take a dependency on `@mailwoman/osm` (which
+ *   `@mailwoman/corpus` does not depend on) just to name a shape it reads from a file.
+ *
+ *   Two measurements shape {@link extractAttestedPhrases}, and both are the reason it is safe to run
+ *   over raw OSM at all:
+ *
+ *   **A feature's `name` is usually the VENUE's name, not a sub-venue phrase.** On the Berlin extract
+ *   (2,060 matched features) a `railway=platform` is named `Stendaler Straße` and an `amenity=university`
+ *   is named `Hertie School`; across 250,116 named Great Britain features only 6,003 (2.40%) contain a
+ *   designator token at all. So names are not harvested wholesale — a name contributes only when it
+ *   CONTAINS a phrase already in the surface index, which is what makes `Terminal E (Untere Ebene)`
+ *   evidence and `Otto Lilienthal Flughafen Berlin Tegel` not.
+ *
+ *   **A matched phrase belongs to the record the PHRASE names, not the record the ROW carries.**
+ *   Attributing every hit to `row.designatorID` — the rule that matched the FEATURE — produced `west →
+ *   platform`, `hall → platform` and `biggin → platform` on the GB extract, because a bus stop tagged
+ *   `public_transport=platform` is named "Village Hall" or "West Kensington"; 108 of 133 OSM-derived
+ *   surfaces named a different record than the one they pointed at. Attribution therefore runs through
+ *   the index, and the row's own designator is kept as `context` — exactly the axis a confound board
+ *   needs, since a `hall` seen on a platform is a confound and a `hall` seen on a terminal is evidence.
+ */
+import { readLocalTextFile } from "@mailwoman/core/fs/readers";
+import { parseJSONStrict } from "@mailwoman/core/json";
+import { TextSpliterator } from "spliterator";
+import { nameContainsSurfaces } from "#tools/sub-venue/surfaces";
+/**
+ * Classify an OSM `ref` into an {@link IdentifierShape} class.
+ *
+ * The classes are the ones Berlin's gates actually produced, plus the two aviation forms the corpus task names
+ * (`Terminal 2F` is digit-letter, `Concourse B` is letter). `range` covers both separators OSM uses for a gate serving
+ * more than one stand: `16-18` and `0/1`.
+ */
+export function classifyIdentifier(ref) {
+    const value = ref.trim();
+    if (/^[0-9]+$/.test(value))
+        return "digit";
+    if (/^[A-Za-z]$/.test(value))
+        return "letter";
+    if (/^[A-Za-z]+[0-9]+$/.test(value))
+        return "letter-digit";
+    if (/^[0-9]+[A-Za-z]+$/.test(value))
+        return "digit-letter";
+    if (/^[0-9A-Za-z]+\s*[-/]\s*[0-9A-Za-z]+$/.test(value))
+        return "range";
+    return "other";
+}
+/**
+ * How many real `ref` values each {@link IdentifierShape} keeps.
+ *
+ * Eight, not "all" and not one. The field exists so a slice author can see what a class actually CONTAINS — GB's
+ * `other` class turned out to be semicolon multi-values (`1;2;3`, `13;14`), which one example would have hidden and
+ * which the class name does not say. Eight fits a terminal line and covers the variety inside every class the GB
+ * extract produced. The COUNT lives in `observations`; this is a sample, not a census.
+ */
+const IDENTIFIER_EXAMPLES_PER_SHAPE = 8;
+/**
+ * Harvest attested phrases and identifier shapes out of a source's rows.
+ *
+ * `index` gates the name harvest — a name contributes only when it CONTAINS a phrase already in the table. That filter
+ * is the whole reason this function is safe to run over raw OSM: see this file's header for the Berlin measurement that
+ * motivated it. The index also decides ATTRIBUTION: a hit is a surface of the record the PHRASE names, and the row's
+ * own designator is recorded as `context`.
+ *
+ * Returns surfaces with real `observations` counts, so the lexicon can rank `terminal` above a phrase attested once.
+ */
+export function extractAttestedPhrases(rows, index, options = {}) {
+    const source = options.source ?? "osm";
+    const region = options.region ?? "";
+    /**
+     * `phrase\0lang\0source` → { count, context }.
+     */
+    const surfaceCounts = new Map();
+    /**
+     * `designatorID\0shape` → { count, examples }.
+     */
+    const shapes = new Map();
+    const note = (phrase, lang, sourceTag, context) => {
+        const key = `${phrase}\0${lang}\0${sourceTag}`;
+        const entry = surfaceCounts.get(key) ?? { count: 0, context: new Map() };
+        entry.count++;
+        entry.context.set(context, (entry.context.get(context) ?? 0) + 1);
+        surfaceCounts.set(key, entry);
+    };
+    for (const row of rows) {
+        if (row.name) {
+            for (const hit of nameContainsSurfaces(row.name, index)) {
+                // `und` — the default `name` tag carries no language. Overture's `name` is the same: a
+                // primary name in whatever language the place uses, untagged.
+                note(hit, "und", `${source}:name`, row.designatorID);
+            }
+        }
+        for (const [lang, localized] of Object.entries(row.localizedNames ?? {})) {
+            for (const hit of nameContainsSurfaces(localized, index)) {
+                note(hit, lang, `${source}:name:${lang}`, row.designatorID);
+            }
+        }
+        if (row.ref) {
+            const shape = classifyIdentifier(row.ref);
+            const key = `${row.designatorID}\0${shape}`;
+            const entry = shapes.get(key) ?? { count: 0, examples: new Set() };
+            entry.count++;
+            if (entry.examples.size < IDENTIFIER_EXAMPLES_PER_SHAPE) {
+                entry.examples.add(row.ref.trim());
+            }
+            shapes.set(key, entry);
+        }
+    }
+    const surfaces = [...surfaceCounts].map(([key, entry]) => {
+        const [phrase, lang, sourceTag] = key.split("\0");
+        const record = index.get(phrase);
+        return {
+            phrase,
+            recordID: record.recordID,
+            recordKind: record.recordKind,
+            lang,
+            region,
+            source: sourceTag,
+            curated: false,
+            observations: entry.count,
+            context: Object.fromEntries([...entry.context].toSorted((a, b) => b[1] - a[1] || a[0].localeCompare(b[0]))),
+        };
+    });
+    const identifierShapes = [...shapes].map(([key, entry]) => {
+        const [designatorID, shape] = key.split("\0");
+        return {
+            designatorID,
+            region,
+            shape,
+            observations: entry.count,
+            examples: [...entry.examples].toSorted((a, b) => a.localeCompare(b)),
+        };
+    });
+    return { surfaces, identifierShapes };
+}
+/**
+ * Read a JSONL file of {@link SubVenueHarvestRow}s. Blank lines and unparseable rows are skipped rather than fatal — an
+ * extract is a build output, and one malformed line should not cost the whole lexicon.
+ */
+export async function readSubVenueJSONL(path) {
+    const out = [];
+    // `TextSpliterator` rather than `split("\n")` — a whole-country extract runs to 250,000 lines
+    // (52 MB for Great Britain), and materializing every segment before reading the first is exactly
+    // what the repo lint rule exists to prevent.
+    for (const line of TextSpliterator.from(await readLocalTextFile(path))) {
+        const trimmed = line.trim();
+        if (!trimmed)
+            continue;
+        try {
+            out.push(parseJSONStrict(trimmed));
+        }
+        catch {
+            continue;
+        }
+    }
+    return out;
+}
+//# sourceMappingURL=harvest.js.map