mailwoman 9.2.0 → 9.3.0

package/README.md

@@ -154,7 +154,7 @@ await parse("350 5TH AVE, NEW YORK, NY 10118", {
 
 Resolution turns parsed components into a Who's On First place ID and coordinate. It needs a
 gazetteer SQLite database — build one with `mailwoman gazetteer build admin` +
-`mailwoman gazetteer build fts`, or point at a prebuilt shard. The resolver is
+`mailwoman gazetteer build fts`, or point at a prebuilt database. The resolver is
 administrative/postcode-level, not rooftop: it returns place centroids (locality, region,
 postcode), not delivery-point coordinates.
 
@@ -181,7 +181,7 @@ Germany                            →   postcode, locality, dependent_locality,
 everything else                    →   locality, dependent_locality, postcode, region, country
 ```
 
-Both routes to the first arm live in `@mailwoman/codex` and are read by `@mailwoman/resolver`'s `adminLadderForNodes`. The second arm is the #945 convention, adopted for France where one _code postal_ often spans several communes.
+Both routes to the first arm live in `@mailwoman/codex` and are read by `@mailwoman/resolver`'s `adminLadderForNodes`. The second arm is the #945 convention, adopted for France where one _code postal_ can span several communes.
 
 The distinction is a fact about a country's **administrative** geography rather than its postal system, and code length does not predict it — France and Germany are both five digits and disagree. Germany's Gemeinden are large enough that Berlin is a single locality, so its PLZ is finer by a wide margin: on 2,997 OpenAddresses rows the coordinate p50 goes 5.84 km → 1.24 km, better on every percentile including p99, and the within-5-km population doubles.
 
@@ -192,7 +192,7 @@ Membership is earned by a full-panel measurement, never by a shape that looks ti
 Mailwoman's runtime pipeline is a staged coordinator:
 
 ```
-normalize → query-shape → locale-gate → kind-classifier → phrase-grouper → classifier → decoder
+normalize → query-shape → locale-hint → kind-classifier → phrase-grouper → classifier → decoder
 ```
 
 Each stage is published as its own `@mailwoman/*` package. The `mailwoman`
@@ -209,7 +209,7 @@ package is the umbrella that wires them together as a single `npm install`.
 | `@mailwoman/neural-weights-fr-fr` | Trained model bundle (fr-FR)                       |
 | `@mailwoman/normalize`            | Stage 1: input preprocessing                       |
 | `@mailwoman/query-shape`          | Stage 1.5: structural priors                       |
-| `@mailwoman/locale-gate`          | Stage 2: locale detection                          |
+| `@mailwoman/locale-hint`          | Stage 2: locale detection                          |
 | `@mailwoman/kind-classifier`      | Stage 2.5: query kind classification               |
 | `@mailwoman/phrase-grouper`       | Stage 2.7: phrase boundary discovery               |
 | `@mailwoman/codex`                | Postal reference data                              |

package/lib/admin-coherence.ts

@@ -0,0 +1,319 @@
+/**
+ * @copyright Sister Software
+ * @license AGPL-3.0
+ * @author Teffen Ellis, et al.
+ *
+ *   Admin-coherence verdicts (#1717 stage 1) — after a geocode resolves, compare the PARSED admin
+ *   qualifiers (`region`, `country`) against what the winning candidate's resolved ancestry actually
+ *   says, and report a per-component verdict. FLAG-ONLY: nothing reads these verdicts to rank,
+ *   re-pick, or check — they exist so a board run can count how often the resolver's answer ignores a
+ *   qualifier the parse got right (`Weimar, Thüringen` → Weimar TX), and how often the winner
+ *   carries no ancestry to check at all (the `unverifiable` count).
+ *
+ *   Per-component verdicts, never one scalar (the Google-confirmation-levels / USPS-DPV pattern):
+ *
+ *   - `confirmed` — the qualifier matches a winner-ancestry node of its class under the shared
+ *     name fold ({@link normalizeLocalityForKey}, the same fold candidate.db's `name_key` is built
+ *     with — build side and check side agree by construction), or the winner IS that qualifier's
+ *     own resolution (a region-tagged winner confirms the region qualifier by identity —
+ *     containment degenerates to self).
+ *   - `contradicted` — the winner's ancestry HAS a node of that class and none of them fold-match
+ *     the parsed value.
+ *   - `unstated` — the parse produced no such qualifier. The common case, and not a problem: most
+ *     queries simply don't name a region or country.
+ *   - `unverifiable` — the parse produced the qualifier but the winner carries NO ancestry of that
+ *     class to check against. Report it faithfully; folding it into either decided verdict would
+ *     hide exactly the gap #1717 wants measured.
+ *
+ *   STATED BOUNDS (v1 is fold-equality only — do not read more into a verdict than this):
+ *
+ *   - Cross-language variant forms are NOT bridged: `Thüringen` folds to `thuringen`, the stored
+ *     exonym `Thuringia` to `thuringia`, so a variant-form match the gazetteer could vouch for
+ *     still reads `contradicted`. Bridging it needs a candidate.db alias probe, which would pull
+ *     the SQLite lookup implementation into this pure module — deliberately skipped.
+ *   - The only normalizers consulted beyond the fold are the codex tables, because they are pure
+ *     and already in-house: {@link matchCountry} (surface form / alpha-2 / alpha-3 → country) and
+ *     {@link matchSubdivision} (US state + CA province code ↔ name). So `IL` confirms against
+ *     `Illinois` and `Deutschland` against a DE winner, but an uncurated endonym (`Alemania`)
+ *     against a DE winner reads `contradicted` — the module never silently over-claims a match it
+ *     cannot derive.
+ *   - The region verdict additionally carries the MISLABEL BRIDGE: a region slot holding a COUNTRY
+ *     name ("Batumi, Georgia" parses region="Georgia") confirms against the winner's country-class
+ *     evidence, because containment holds and `contradicted` would misdescribe the geography. It
+ *     runs after the region band, is monotone (`contradicted`/`unverifiable` → `confirmed` is the
+ *     only movement it can cause), and inherits the pure-codex bound above — parsed "Russia"
+ *     against an RU winner still reads `contradicted`, because codex holds only "Russian
+ *     Federation" for RU and the module never over-claims.
+ *
+ *   The winner's checkable ancestry arrives as the resolver's `metadata.ancestors` stamp (#404 —
+ *   the geocode path opts in by default, and both backends serve it when their artifact carries an
+ *   ancestors table: the FTS database's `ancestors`, candidate.db's `candidate_ancestor` sidecar),
+ *   while country-class ancestry is nearly always available via the `resolver_country` stamp.
+ *   `unverifiable` remains the faithful verdict wherever the stamp is absent — an artifact
+ *   predating the sidecar, a database-fed winner with no recorded ancestry, or a caller that opted
+ *   out.
+ */
+
+import { countrySurfaceForms, ISO2_TO_NAME, matchCountry } from "@mailwoman/codex/country"
+import { walkNodes } from "@mailwoman/core/decoder"
+import { REGION_CLASS_PLACETYPES, regionKeys } from "@mailwoman/resolver-wof-sqlite/region-keys"
+import { normalizeLocalityForKey } from "@mailwoman/resolver-wof-sqlite/street"
+
+/**
+ * One admin-coherence verdict. See the module docstring for the exact meaning of each — in particular, `unverifiable`
+ * is an absence-of-evidence claim about the WINNER, never about the parse.
+ */
+type AdminCoherenceVerdict = "confirmed" | "contradicted" | "unstated" | "unverifiable"
+
+/**
+ * The per-component verdicts. Both members are always present when the report exists (the `intent_markers` discipline:
+ * state the empty case — `unstated` IS the explicit "no qualifier" claim, so an optional member would be a second way
+ * to say the same thing). The report as a whole is what's optional: absent means the geocode resolved no winner to
+ * check against.
+ */
+export interface AdminCoherenceReport {
+	region: AdminCoherenceVerdict
+	country: AdminCoherenceVerdict
+}
+
+/**
+ * The parsed admin qualifiers — the raw spans off the address tree's `region` / `country` nodes (the parse view, not
+ * the resolved view). Empty / whitespace-only reads as absent.
+ */
+export interface ParsedAdminQualifiers {
+	region?: string | undefined
+	country?: string | undefined
+}
+
+/**
+ * One link of the winner's containment lineage, as the resolver stamped it (`metadata.ancestors`, the #404 opt-in) — a
+ * structural slice of `@mailwoman/core`'s `Ancestor`.
+ */
+interface AdminAncestor {
+	placetype: string
+	name: string
+}
+
+/**
+ * The winning candidate, reduced to what the check reads: its own component tag (for the self-confirmation case), the
+ * resolver-stamped ISO 3166-1 alpha-2 (`resolver_country` — the one piece of country-class ancestry the candidate
+ * backend always carries), and the stamped ancestor chain when a backend supplied one.
+ */
+export interface AdminCoherenceWinner {
+	tag: string
+	countryCode?: string | undefined
+	ancestry?: readonly AdminAncestor[] | undefined
+}
+
+/**
+ * Fold both sides of every name comparison through the shared candidate.db `name_key` normalizer — one function, both
+ * sides, so the check can never disagree with the index it's checking against.
+ *
+ * The region-side expansion ({@link regionKeys}) and the region band ({@link REGION_CLASS_PLACETYPES}) moved DOWN to
+ * `@mailwoman/resolver-wof-sqlite/region-keys` when the #1717 stage-2 containment re-rank became their second consumer
+ * — the dependency points that way, and the #861 rule wants one function, not a mirrored copy.
+ */
+function foldKey(name: string): string {
+	return normalizeLocalityForKey(name)
+}
+
+/**
+ * The comparable keys a country-class value expands to: its own fold, plus — when {@link matchCountry} recognizes it —
+ * an `iso2:` channel key and the folds of the canonical name and curated surface forms, so `Deutschland`, `DEU`, `DE`
+ * and `Germany` all meet in one key set regardless of which side spelled which.
+ */
+function countryKeys(value: string): Set<string> {
+	const keys = new Set([foldKey(value)])
+	const match = matchCountry(value)
+
+	if (match) {
+		keys.add(`iso2:${match.iso2}`)
+
+		if (match.canonical) {
+			keys.add(foldKey(match.canonical))
+		}
+
+		for (const form of countrySurfaceForms(match.iso2)) {
+			keys.add(foldKey(form))
+		}
+	}
+
+	return keys
+}
+
+function intersects(a: ReadonlySet<string>, b: ReadonlySet<string>): boolean {
+	for (const key of a) {
+		if (b.has(key)) return true
+	}
+
+	return false
+}
+
+/**
+ * The winner's COUNTRY-class evidence keys: the resolver-stamped alpha-2 expanded through the codex tables into every
+ * spelling the check can vouch for, plus any country-placetype ancestors. One assembly, two consumers — the country
+ * verdict compares against it, and the region verdict's mislabel bridge (below) does too, so the two verdicts can never
+ * disagree about what counts as country-class evidence.
+ */
+function winnerCountryKeys(winner: AdminCoherenceWinner): Set<string> {
+	const winnerKeys = new Set<string>()
+	const iso = winner.countryCode?.trim().toUpperCase() || undefined
+
+	if (iso) {
+		winnerKeys.add(`iso2:${iso}`)
+		winnerKeys.add(foldKey(iso))
+		const canonical = ISO2_TO_NAME.get(iso)
+
+		if (canonical) {
+			winnerKeys.add(foldKey(canonical))
+		}
+
+		for (const form of countrySurfaceForms(iso)) {
+			winnerKeys.add(foldKey(form))
+		}
+	}
+
+	for (const ancestor of winner.ancestry ?? []) {
+		if (ancestor.placetype !== "country") continue
+
+		for (const key of countryKeys(ancestor.name)) {
+			winnerKeys.add(key)
+		}
+	}
+
+	return winnerKeys
+}
+
+function regionVerdict(parsedRegion: string | undefined, winner: AdminCoherenceWinner): AdminCoherenceVerdict {
+	const parsed = parsedRegion?.trim()
+
+	if (!parsed) return "unstated"
+
+	// The winner IS a region resolution — the qualifier is the thing that resolved, so containment
+	// degenerates to identity. The resolver's own binding (alias-aware, unlike the fold) is the
+	// match evidence here; re-checking it under fold-equality would misread every alias hit as a
+	// contradiction.
+	if (winner.tag === "region") return "confirmed"
+
+	const regionAncestors = (winner.ancestry ?? []).filter((a) => REGION_CLASS_PLACETYPES.has(a.placetype))
+	const iso = winner.countryCode?.trim().toUpperCase() || undefined
+	const parsedKeys = regionKeys(parsed, iso)
+
+	for (const ancestor of regionAncestors) {
+		if (intersects(parsedKeys, regionKeys(ancestor.name, iso))) return "confirmed"
+	}
+
+	// The mislabel bridge: the region SLOT sometimes holds a COUNTRY name — "Moscow, Russia" parses
+	// region="Russia", "Batumi, Georgia" parses region="Georgia" (the shape the flag's own first
+	// triage counted at ~4 of 16 contradictions). Containment still holds when the winner's
+	// country-class evidence matches the qualifier, so `contradicted` would be the wrong claim about
+	// the geography. Checked AFTER the region band (a genuine region match never depends on it) and
+	// monotone by construction: it can only move `contradicted`/`unverifiable` → `confirmed`.
+	if (intersects(countryKeys(parsed), winnerCountryKeys(winner))) return "confirmed"
+
+	return regionAncestors.length ? "contradicted" : "unverifiable"
+}
+
+function countryVerdict(parsedCountry: string | undefined, winner: AdminCoherenceWinner): AdminCoherenceVerdict {
+	const parsed = parsedCountry?.trim()
+
+	if (!parsed) return "unstated"
+
+	if (winner.tag === "country") return "confirmed"
+
+	const winnerKeys = winnerCountryKeys(winner)
+
+	if (!winnerKeys.size) return "unverifiable"
+
+	return intersects(countryKeys(parsed), winnerKeys) ? "confirmed" : "contradicted"
+}
+
+/**
+ * Assess the parsed admin qualifiers against the winning candidate. Pure — no I/O, no lookup, no side effects; call it
+ * once at result assembly, only when a winner exists (no winner → no report, absence meaning "nothing resolved to check
+ * against").
+ */
+export function assessAdminCoherence(
+	parsed: ParsedAdminQualifiers,
+	winner: AdminCoherenceWinner
+): AdminCoherenceReport {
+	return {
+		region: regionVerdict(parsed.region, winner),
+		country: countryVerdict(parsed.country, winner),
+	}
+}
+
+/**
+ * A resolved-tree node slice the assembly adapter reads — structurally satisfied by `@mailwoman/core`'s `AddressNode`,
+ * declared locally so the pure module carries no decoder import.
+ */
+export interface AdminCoherenceSourceNode {
+	tag: string
+	value: string
+	metadata?: Record<string, unknown> | undefined
+}
+
+/**
+ * The assembly-point adapter: derive the parsed qualifiers (the `region` / `country` node spans — the parse view) and
+ * the winner's checkable ancestry (the `resolver_country` stamp + any `metadata.ancestors` chain) off the resolved
+ * tree's nodes, and return a spreadable result fragment. `winner` is the admin-ladder pick; `fallbackWinner` is the
+ * primary resolved node the street-backed tiers report instead (the resolution context the coordinate was scoped by).
+ * No winner at all → an empty fragment: the `admin_coherence` field stays ABSENT, which is a different claim from
+ * `unverifiable` (nothing resolved, so there was no candidate to check).
+ */
+export function adminCoherenceField(
+	nodes: readonly AdminCoherenceSourceNode[],
+	winner: AdminCoherenceSourceNode | undefined,
+	fallbackWinner: AdminCoherenceSourceNode | undefined
+): { admin_coherence?: AdminCoherenceReport } {
+	const picked = winner ?? fallbackWinner
+
+	if (!picked) return {}
+
+	const report = assessAdminCoherence(
+		{
+			region: nodes.find((n) => n.tag === "region")?.value?.trim() || undefined,
+			country: nodes.find((n) => n.tag === "country")?.value?.trim() || undefined,
+		},
+		{
+			tag: picked.tag,
+			countryCode: (picked.metadata?.["resolver_country"] as string | undefined)?.trim() || undefined,
+			// The resolver's #404 stamp — present when the geocode path opted in AND the backend's
+			// artifact carries an ancestors table; its absence is what the verdicts report as
+			// `unverifiable`.
+			ancestry: picked.metadata?.["ancestors"] as readonly AdminAncestor[] | undefined,
+		}
+	)
+
+	return { admin_coherence: report }
+}
+
+/**
+ * A node tree shaped like the decoder's `AddressNode` — structural rather than imported, so this module stays free of
+ * the decoder dependency.
+ */
+export interface AdminCoherenceTreeNode extends AdminCoherenceSourceNode {
+	children: readonly AdminCoherenceTreeNode[]
+}
+
+/**
+ * Coherence for a fork-to-entity answer (#1724): a forked answer carries a verdict like any other resolved answer --
+ * absence means "nothing resolved to check", and something did. The entity offers a country and no ancestor chain, so a
+ * stated region grades `unverifiable` rather than going silently unchecked.
+ */
+export function forkedEntityCoherenceField(
+	roots: readonly AdminCoherenceTreeNode[],
+	entity: { name: string; country: string }
+): { admin_coherence?: AdminCoherenceReport } {
+	const nodes: AdminCoherenceSourceNode[] = [...walkNodes(roots)].map((n) => ({
+		tag: n.tag,
+		value: n.value,
+		metadata: n.metadata,
+	}))
+
+	return adminCoherenceField(
+		nodes,
+		{ tag: "venue", value: entity.name, metadata: { resolver_country: entity.country } },
+		undefined
+	)
+}