@mailwoman/corpus 9.3.0 → 10.0.0

package/lib/tools/golden/relabel-street/row.ts

@@ -0,0 +1,361 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Golden-set street-suffix relabel — v0.1.2 → v0.1.3.
+ *
+ *   ## Why this exists
+ *
+ *   The golden answer key and the training corpus disagreed about ONE thing, and the v9.0.0
+ *   promotion eval read the disagreement as a model regression (`us.street` 87.4 vs a floor of
+ *   87.8). The corpus SPLITS a US street into `street` + `street_suffix` — TIGER's adapter
+ *   decomposes at `corpus/src/adapters/tiger/street-decompose.ts`, the `street-affix` slice recipe
+ *   teaches it from USPS Pub-28, and `ComponentTag` carries `street_suffix` as a first-class tag.
+ *   The golden set FOLDED it: 2,216 US rows carry a `street`, and exactly 2 of them label a
+ *   `street_suffix`. Operator ruling, 2026-08-06: **the split is canonical**; the golden is the
+ *   stale side. This tool moves the answer key onto the corpus convention.
+ *
+ *   ## The instrument
+ *
+ *   `matchTrailingSuffix` from `@mailwoman/codex/us` — the USPS Pub-28 Appendix C table, which is
+ *   also what the corpus slice recipe splits on. The table is NOT re-implemented here, and the
+ *   libpostal dictionary TIGER reads is deliberately not used: measured on this golden set the two
+ *   disagree on 51 US rows, and the disagreements run in the codex table's favour (libpostal's
+ *   `directionals.txt` lists `center|c`, so TIGER reads the `C` of "C STREET" as a directional
+ *   prefix and then emits no suffix at all).
+ *
+ *   ## What it changes, and what it refuses to
+ *
+ *   Applied only to rows whose `country` is `US`. Three branches, mirroring the SHAPE of TIGER's
+ *   `decomposeStreet` on codex tables:
+ *
+ *   - **street type** — the last whitespace-separated word is a Pub-28 suffix, and something is left
+ *       over: "Main St" → `street: "Main"`, `street_suffix: "St"` (1,559 rows).
+ *   - **street type + post-directional** — the last word is a directional AND the one before it is a
+ *       Pub-28 suffix: "Pennsylvania Avenue NW" → `street: "Pennsylvania"`,
+ *       `street_suffix: "Avenue NW"` (347 rows). The post-directional joins the suffix rather than
+ *       becoming a tag of its own, because that is what the corpus adapter emits; there is no
+ *       `street_postfix` tag to move it to.
+ *   - **everything else is left folded** and reported. In particular a BARE post-directional tail
+ *       ("Seymour East", "BROADWAY N" — 16 rows) is NOT split: a directional is not a Pub-28 suffix,
+ *       and the observed rows in that class are unit-contaminated ("1ST AVE SW BOX E", where the
+ *       trailing "E" is a box letter).
+ *
+ *   FR rows are untouched, deliberately and permanently as far as this tool is concerned. French
+ *   street typology puts the type FIRST ("Rue de la Paix") and the golden labels only 7 of 665 FR
+ *   street rows with a `street_prefix`; whether FR should split at all is a different question with
+ *   a different table behind it, and nothing here should be read as having answered it.
+ *
+ *   ## Surface bytes
+ *
+ *   The split is a break at a whitespace run in the ORIGINAL string — no trimming, no case
+ *   normalization, no re-joining of tokens. `street + gap + street_suffix` reconstructs the input
+ *   byte-for-byte, so the whitespace between them belongs to neither span (the same shape the corpus
+ *   adapter's spans have). The tool asserts this per row and refuses to write a file if it ever
+ *   fails.
+ */
+
+import {
+	isStreetDirectionalToken,
+	matchTrailingSuffix,
+	type USStreetSuffix,
+	NAME_PRONE_US_SUFFIXES,
+} from "@mailwoman/codex/us"
+import { escapeRegExp } from "@mailwoman/core/strings/regexp"
+
+// ── Types ──────────────────────────────────────────────────────────────────
+
+/**
+ * A golden-set row, as stored one-per-line in `us.jsonl` / `fr.jsonl` / `adversarial.jsonl`. Only the fields this tool
+ * reads are modeled; every other key rides through untouched.
+ */
+export interface GoldenStreetRow {
+	raw: string
+	components: Record<string, string>
+	country?: string
+	source?: string
+	notes?: string
+	[key: string]: unknown
+}
+
+/**
+ * What the tool decided about one row. Every value except the two `split-*` classes means "left folded".
+ */
+export type GoldenRelabelClass =
+	| "split-suffix"
+	| "split-suffix-postdirectional"
+	| "split-prefix-only"
+	| "already-split"
+	| "single-token"
+	| "suffix-only-street"
+	| "postdirectional-tail-only"
+	| "no-suffix-match"
+	| "no-street"
+	| "not-us"
+	| "untrimmed-street"
+
+/**
+ * A review trigger on a row the tool DID change. A flag is never an adjudication — it marks the row for the operator's
+ * deck, and the split is applied either way.
+ */
+export interface GoldenRelabelFlag {
+	kind: "name-prone-suffix" | "venue-context" | "remainder-is-affix"
+	detail: string
+}
+
+/**
+ * Row-level relabel outcome.
+ */
+export interface GoldenRelabelResult {
+	/**
+	 * The row to write. Identical object reference when nothing changed.
+	 */
+	row: GoldenStreetRow
+	changed: boolean
+	rowClass: GoldenRelabelClass
+	flags: GoldenRelabelFlag[]
+	/**
+	 * A leading directional was lifted into `street_prefix` on this row.
+	 */
+	prefixSplit: boolean
+	/**
+	 * The street span as it stood in the parent version — recorded for the review deck.
+	 */
+	beforeStreet?: string
+}
+
+// ── The name-prone suffix set ──────────────────────────────────────────────
+
+/**
+ * Pub-28 canonicals that are also ordinary head nouns of PROPER NAMES — "Lincoln Park", "Boston Common", "Willow
+ * Brook". A split on one of these is still applied (the table is the table), but the row lands in the review deck
+ * because the trailing word may belong to the name rather than to the street type.
+ *
+ * Chosen against the surfaces this golden set actually carries (park 6, green 5, hill 7, heights 3, hollow 3, brook 3,
+ * pass 3 — the whole flagged class is 60 rows of 1,906), not from the whole 200-entry table: flagging every possible
+ * name-head would mark a third of the corrections and stop being a review artifact.
+ */
+// ── Byte-exact tail split ──────────────────────────────────────────────────
+
+interface TailSplit {
+	head: string
+	gap: string
+	tail: string
+}
+
+/**
+ * Split `s` at its LAST whitespace run, returning the three pieces verbatim. Null when there is no interior whitespace,
+ * when the head would be empty, or when `s` carries leading/trailing whitespace (a golden row is stored trimmed; an
+ * untrimmed one is reported rather than silently normalized).
+ */
+function splitLastWord(s: string): TailSplit | null {
+	if (s !== s.trim() || !s) return null
+	const match = /^(.*\S)(\s+)(\S+)$/.exec(s)
+
+	if (!match) return null
+
+	return { head: match[1]!, gap: match[2]!, tail: match[3]! }
+}
+
+/**
+ * Split `s` at its FIRST whitespace run — the leading-directional counterpart of {@link splitLastWord}. `head` is the
+ * first word, `tail` the rest, both verbatim.
+ */
+function splitFirstWord(s: string): TailSplit | null {
+	if (s !== s.trim() || !s) return null
+	const match = /^(\S+)(\s+)(\S.*)$/.exec(s)
+
+	if (!match) return null
+
+	return { head: match[1]!, gap: match[2]!, tail: match[3]! }
+}
+
+// ── Row-level relabel ──────────────────────────────────────────────────────
+
+/**
+ * Rebuild `components` with `street_suffix` inserted immediately after `street`, so the written row reads in address
+ * order rather than with the new tag appended at the end.
+ */
+function withStreetSpans(
+	components: Record<string, string>,
+	spans: { prefix?: string; street: string; suffix?: string }
+): Record<string, string> {
+	const out: Record<string, string> = {}
+
+	for (const [key, value] of Object.entries(components)) {
+		if (key === "street") {
+			if (spans.prefix) {
+				out.street_prefix = spans.prefix
+			}
+
+			out.street = spans.street
+
+			if (spans.suffix) {
+				out.street_suffix = spans.suffix
+			}
+
+			continue
+		}
+
+		if (key === "street_prefix" || key === "street_suffix") continue
+		out[key] = value
+	}
+
+	return out
+}
+
+/**
+ * Options for {@linkcode relabelGoldenStreetRow}.
+ */
+export interface RelabelStreetRowOptions {
+	/**
+	 * Also lift a folded LEADING directional out into `street_prefix`. Default true.
+	 *
+	 * On by default because the fold applies both ways and the answer key has to be corrected on both, or the correction
+	 * is not a correction: 207 of the 1,682 split dev rows (12.3%) still opened with a directional after the suffix move
+	 * — "N Desmet Avenue" would have graded `street: "N Desmet"` against a model that says `street_prefix: "N", street:
+	 * "Desmet"`. Turn it OFF only to measure what the prefix fold alone costs.
+	 */
+	splitPrefix?: boolean
+}
+
+/**
+ * Decide, and apply, the US street-span split for ONE golden row. Pure: never mutates its argument, and returns the
+ * same object reference when the row is left alone.
+ */
+export function relabelGoldenStreetRow(
+	row: GoldenStreetRow,
+	options: RelabelStreetRowOptions = {}
+): GoldenRelabelResult {
+	const splitPrefix = options.splitPrefix ?? true
+
+	const keep = (rowClass: GoldenRelabelClass): GoldenRelabelResult => ({
+		row,
+		changed: false,
+		rowClass,
+		flags: [],
+		prefixSplit: false,
+	})
+
+	if ((row.country ?? "").toUpperCase() !== "US") return keep("not-us")
+	const street = row.components.street
+
+	if (!street) return keep("no-street")
+
+	if (row.components.street_suffix) return keep("already-split")
+
+	if (street !== street.trim()) return keep("untrimmed-street")
+
+	const flags: GoldenRelabelFlag[] = []
+	let rowClass: GoldenRelabelClass
+	let name = street
+	let suffix: string | undefined
+	let suffixGap = ""
+	let canonical: USStreetSuffix | undefined
+
+	const split = splitLastWord(street)
+
+	if (!split) {
+		rowClass = matchTrailingSuffix(street) ? "suffix-only-street" : "single-token"
+	} else if (isStreetDirectionalToken(split.tail)) {
+		// Street type + post-directional ("Pennsylvania Avenue NW"). The corpus adapter emits the pair as
+		// ONE suffix span, and there is no post-directional tag to move it to.
+		const inner = splitLastWord(split.head)
+		const typeMatch = inner ? matchTrailingSuffix(inner.tail) : null
+
+		if (!inner || !typeMatch) {
+			rowClass = "postdirectional-tail-only"
+		} else {
+			rowClass = "split-suffix-postdirectional"
+			name = inner.head
+			suffix = `${inner.tail}${split.gap}${split.tail}`
+			suffixGap = inner.gap
+			canonical = typeMatch.canonical
+		}
+	} else {
+		const typeMatch = matchTrailingSuffix(street)
+
+		if (!typeMatch) {
+			rowClass = "no-suffix-match"
+		} else {
+			rowClass = "split-suffix"
+			name = split.head
+			suffix = split.tail
+			suffixGap = split.gap
+			canonical = typeMatch.canonical
+		}
+	}
+
+	// Leading directional → street_prefix, on whatever name survived the suffix move. Independent of the
+	// suffix branch, because "N Main" is as folded as "N Main St" is.
+	let prefix: string | undefined
+	let prefixGap = ""
+
+	if (splitPrefix && !row.components.street_prefix) {
+		const lead = splitFirstWord(name)
+
+		if (lead && isStreetDirectionalToken(lead.head)) {
+			prefix = lead.head
+			prefixGap = lead.gap
+			name = lead.tail
+		}
+	}
+
+	if (!suffix && !prefix) return keep(rowClass)
+
+	if (!suffix) {
+		rowClass = "split-prefix-only"
+	}
+
+	// The invariant this tool exists to keep: the spans plus the whitespace between them ARE the original
+	// span, byte for byte. Anything else means a token was rewritten.
+	const rebuilt = `${prefix ? prefix + prefixGap : ""}${name}${suffix ? suffixGap + suffix : ""}`
+
+	if (rebuilt !== street) {
+		throw new Error(`golden-relabel: span reconstruction failed for ${JSON.stringify(street)}`)
+	}
+
+	if (canonical && NAME_PRONE_US_SUFFIXES.has(canonical)) {
+		flags.push({
+			kind: "name-prone-suffix",
+			detail: `${canonical} heads proper names as often as street types`,
+		})
+	}
+
+	const venue = row.components.venue
+
+	if (suffix && venue && new RegExp(`(^|\\W)${escapeRegExp(suffix)}(\\W|$)`, "i").test(venue)) {
+		flags.push({
+			kind: "venue-context",
+			detail: `venue ${JSON.stringify(venue)} also carries ${JSON.stringify(suffix)}`,
+		})
+	}
+
+	// Narrow on purpose: a name that happens to be a Pub-28 canonical is NOT interesting ("Mountain Rd",
+	// "Valley Dr", "Mills Ln" are ordinary streets, and flagging them buried the deck — 108 rows of noise
+	// on the first run). A name that is a bare DIRECTIONAL is: "East Rd" leaves `street: "East"`, which is
+	// a direction, not a name.
+	if (isStreetDirectionalToken(name)) {
+		flags.push({
+			kind: "remainder-is-affix",
+			detail: `street would become the bare directional ${JSON.stringify(name)}`,
+		})
+	}
+
+	return {
+		row: {
+			...row,
+			components: withStreetSpans(row.components, {
+				...(prefix ? { prefix } : row.components.street_prefix ? { prefix: row.components.street_prefix } : {}),
+				street: name,
+				...(suffix ? { suffix } : {}),
+			}),
+		},
+		changed: true,
+		rowClass,
+		flags,
+		prefixSplit: Boolean(prefix),
+		beforeStreet: street,
+	}
+}

package/lib/tools/index.ts

@@ -12,17 +12,17 @@ export * from "#tools/audit"
 export * from "#tools/corpus-stats"
 export * from "#tools/ingest-csv"
 export * from "#tools/overlay-manifest"
-export * from "#tools/fetch/download"
+export * from "#tools/fetch/download/index"
 export * from "#tools/fetch/index"
-export * from "#tools/golden-expand"
-export * from "#tools/golden-promote"
-export * from "#tools/golden-relabel-street"
+export * from "#tools/golden/expand"
+export * from "#tools/golden/promote"
+export * from "#tools/golden/relabel-street/index"
 export * from "#tools/jsonl-to-parquet"
-export * from "#tools/lint-slice"
-export * from "#tools/lint-slice-vocab"
-export * from "#tools/slice-kryptonite"
+export * from "#tools/lint/slice/index"
+export * from "#tools/lint/slice/vocab"
+export * from "#tools/slice/kryptonite"
 export * from "#tools/overture-subvenue"
 export * from "#tools/postcode-triples"
-export * from "#tools/sub-venue-lexicon"
-export * from "#tools/sub-venue-promotions"
-export * from "#tools/slice-translit"
+export * from "#tools/sub/venue/lexicon"
+export * from "#tools/sub/venue/promotions"
+export * from "#tools/slice/translit"

package/lib/tools/ingest-csv.ts

@@ -138,14 +138,12 @@ async function runIngest(opts: IngestOptions): Promise<void> {
 	// file descriptor.
 	const rows = (): AsyncIterable<string[]> =>
 		CSVSpliterator.fromAsync<string[]>(opts.inputPath, {
-			mode: "array",
 			header: false,
 			columnDelimiter: opts.separator,
 			drop: opts.skipLines,
-			enableQuoteHandling: true,
 		})
 
-	// --- Pass 1: read header + sample rows for type inference ---
+	// Read the header and sample rows before inferring the SQLite schema.
 	process.stderr.write(`Reading ${opts.inputPath} for schema inference...\n`)
 
 	let headerRow: string[] | null = null
@@ -169,7 +167,7 @@ async function runIngest(opts: IngestOptions): Promise<void> {
 		throw new Error("No header line found in CSV")
 	}
 
-	// --- Determine column names ---
+	// Normalize and deduplicate the input column names.
 	let rawHeaders: string[]
 
 	if (opts.hasHeader && headerRow) {
@@ -182,7 +180,7 @@ async function runIngest(opts: IngestOptions): Promise<void> {
 
 	const colNames = dedupColumns(rawHeaders.map(normalizeColumnName))
 
-	// --- Infer column types ---
+	// Infer each column's SQLite type from its sample values.
 	const columns: ColumnInfo[] = colNames.map((name, i) => {
 		const samples = sampleRows.map((row) => {
 			const raw = row[i] ?? ""
@@ -196,7 +194,7 @@ async function runIngest(opts: IngestOptions): Promise<void> {
 		return info
 	})
 
-	// --- Generate SQL ---
+	// Generate the table definition and insert statement from the inferred schema.
 	const colDefs = columns.map((c) => `"${c.name}" ${c.type}`).join(",\n  ")
 	// Raw DDL by design: the column set + types are INFERRED from the CSV at runtime (colDefs above),
 	// so a Kysely builder loop would just wrap the same dynamic strings with ceremony and no type safety.
@@ -213,7 +211,7 @@ async function runIngest(opts: IngestOptions): Promise<void> {
 		return
 	}
 
-	// --- Create database + import ---
+	// Create the database, then import every input row.
 	const { DatabaseClient } = await import("@mailwoman/sqlite/client")
 	await makeDirectories(dirname(opts.outputPath))