fix(vinpin): prefer exact model over wrong-market catalog in matcher

The Vinpin catalog matcher could route a European Renault VIN
(VF1RFE00653633190, decoded "KADJAR") onto the China-market catalog
"KADJAR ÇİN" (source XZH, 22 categories) instead of the correct European
"KADJAR" (XFE, 44 categories), serving wrong parts. Root cause: the
Turkish "ÇİN" lost its Ç/İ to the ASCII token strip and collapsed to a
dropped 1-char "N", so "KADJAR ÇİN" tokenized identically to "KADJAR" —
the market qualifier was invisible to the scorer.

Fixes:
- modelTokens now folds diacritics (NFD + combining-mark strip) so
  "ÇİN" survives as the ASCII token "CIN".
- Scorer prefers an EXACT normalized model match (no extra tokens) over
  a superset, via a bonus kept smaller than the year-range swing so
  multi-generation routing (2022 TIPO-EGEA → MCA) is unaffected.
- New MARKET_QUALIFIERS set (ÇİN/CIN, CHINA, CHINE, RUSYA, ... grounded
  in the real Renault/Dacia catalog rows) heavily penalizes candidates
  whose EXTRA tokens are region qualifiers; generation tokens are not
  penalized, so generations stay matchable.
- Richer-catalog (categoryCount) tiebreak among equal-score candidates;
  match() query selects the category count via a correlated subquery.

Adds unit tests incl. the KADJAR case, CLIO/DUSTER generation cases, and
diacritic-folding. Fiat behavior unchanged; all 39 vinpin tests green.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

apps/api/src/integrations/vinpin/vinpin.matcher.spec.ts

@@ -85,4 +85,103 @@ describe("pickBestCatalogMatch", () => {
     // Identical tokens + no year → equal score → recency decides.
     expect(pickBestCatalogMatch({ model: "PUNTO", modelYear: null }, puntos)).toBe("punto-new");
   });
+
+  it("breaks a score tie toward the richer catalog (more categories)", () => {
+    const koleos: CatalogCandidate[] = [
+      { id: "koleos-thin", model: "KOLEOS", year: null, categoryCount: 12 },
+      { id: "koleos-full", model: "KOLEOS", year: null, categoryCount: 44 },
+    ];
+    // Identical tokens + no year → equal score → category count decides.
+    expect(pickBestCatalogMatch({ model: "KOLEOS", modelYear: null }, koleos)).toBe("koleos-full");
+  });
+});
+
+describe("pickBestCatalogMatch — market-qualifier precision (Renault/Dacia)", () => {
+  it("prefers the EXACT model over a wrong-market superset (KADJAR beats KADJAR ÇİN)", () => {
+    // The confirmed prod bug: a European VF1 Renault decoded as "KADJAR" landed
+    // on the China-market "KADJAR ÇİN" catalog (fewer categories, wrong parts).
+    const candidates: CatalogCandidate[] = [
+      { id: "kadjar-cn", model: "KADJAR ÇİN", year: null, categoryCount: 22 },
+      { id: "kadjar-eu", model: "KADJAR", year: null, categoryCount: 44 },
+    ];
+    expect(pickBestCatalogMatch({ model: "KADJAR", modelYear: "2018" }, candidates)).toBe(
+      "kadjar-eu",
+    );
+  });
+
+  it("still picks the exact model even when the market candidate is richer", () => {
+    // Market penalty must dominate — an exact match wins regardless of category
+    // count or ordering.
+    const candidates: CatalogCandidate[] = [
+      { id: "kadjar-cn", model: "KADJAR ÇİN", year: null, categoryCount: 999 },
+      { id: "kadjar-eu", model: "KADJAR", year: null, categoryCount: 1 },
+    ];
+    expect(pickBestCatalogMatch({ model: "KADJAR", modelYear: null }, candidates)).toBe(
+      "kadjar-eu",
+    );
+  });
+
+  it("penalizes a market qualifier but NOT a generation token (CLIO)", () => {
+    const candidates: CatalogCandidate[] = [
+      { id: "clio-cn", model: "CLIO ÇİN", year: null },
+      { id: "clio", model: "CLIO", year: null },
+    ];
+    // Exact "CLIO" beats the China-market superset.
+    expect(pickBestCatalogMatch({ model: "CLIO", modelYear: null }, candidates)).toBe("clio");
+  });
+
+  it("keeps generations matchable — a generation extra token is not a market penalty", () => {
+    // Decoded "CLIO" with only generation-qualified catalogs (no exact bare CLIO)
+    // and one China catalog. The generation candidate must win over the market one,
+    // proving the market penalty doesn't collateral-damage generations.
+    const candidates: CatalogCandidate[] = [
+      { id: "clio-cn", model: "CLIO ÇİN", year: null },
+      {
+        id: "clio-4",
+        model: "CLIO 4 / LUTECIA 4 (2012-2019)",
+        year: "2012-2019",
+        categoryCount: 40,
+      },
+      {
+        id: "clio-3",
+        model: "CLIO 3 / LUTECIA 3 (2005-2014)",
+        year: "2005-2014",
+        categoryCount: 30,
+      },
+    ];
+    const id = pickBestCatalogMatch({ model: "CLIO", modelYear: "2015" }, candidates);
+    // Year-overlap picks the 2012-2019 gen-4; the China catalog is penalized out.
+    expect(id).toBe("clio-4");
+  });
+
+  it("matches an exact generation model and penalizes only the market superset (DUSTER II)", () => {
+    // Roman-numeral generations survive tokenization (single digits are dropped by
+    // the length>=2 filter, an existing behavior). Decoded "DUSTER II" is exact on
+    // duster-2; "DUSTER II ÇİN" carries an extra market token and is penalized out.
+    const candidates: CatalogCandidate[] = [
+      { id: "duster-2", model: "DUSTER II", year: null, categoryCount: 30 },
+      { id: "duster-cn", model: "DUSTER II ÇİN", year: null, categoryCount: 30 },
+    ];
+    expect(pickBestCatalogMatch({ model: "DUSTER II", modelYear: null }, candidates)).toBe(
+      "duster-2",
+    );
+  });
+
+  it("falls back to a market catalog only when it is the sole candidate", () => {
+    // If the ONLY catalog for a decoded model is a regional one, still match it —
+    // the penalty lowers the score but does not disqualify.
+    const candidates: CatalogCandidate[] = [{ id: "x62-cn", model: "X62 CHINE", year: null }];
+    expect(pickBestCatalogMatch({ model: "X62", modelYear: null }, candidates)).toBe("x62-cn");
+  });
+});
+
+describe("modelTokens — diacritic folding", () => {
+  it("folds Turkish diacritics so market qualifiers survive as ASCII tokens", () => {
+    // Without folding, ÇİN loses Ç/İ to the ASCII strip and collapses to a
+    // dropped 1-char "N", making KADJAR ÇİN tokenize identically to KADJAR.
+    expect(modelTokens("KADJAR ÇİN")).toEqual(["KADJAR", "CIN"]);
+    expect(modelTokens("ARKANA RUSYA")).toEqual(["ARKANA", "RUSYA"]);
+    // The single-digit "2" is dropped by the length>=2 filter (existing behavior).
+    expect(modelTokens("KOLEOS 2 - ÇİN")).toEqual(["KOLEOS", "CIN"]);
+  });
 });

apps/api/src/integrations/vinpin/vinpin.matcher.ts

@@ -18,8 +18,62 @@ export interface CatalogCandidate {
   id: string;
   model: string | null;
   year: string | null;
+  /** How many catalog categories this vehicle has fetched. Used only as a final
+   *  tiebreak among otherwise-equal candidates — the fuller catalog is the
+   *  mainstream one. Optional so the pure picker can be unit-tested without it. */
+  categoryCount?: number | null;
 }
 
+/**
+ * Market/region qualifier tokens (diacritic-folded, upper-case). A catalog model
+ * that carries one of these tokens BEYOND the decoded model targets a specific
+ * regional market — e.g. "KADJAR ÇİN" is the China-market Kadjar catalog (source
+ * XZH, 22 categories) vs the mainstream European "KADJAR" (XFE, 44 categories).
+ * For a European (VF1…) VIN the regional catalog serves the wrong parts, so a
+ * candidate whose EXTRA tokens are region qualifiers is heavily penalized — but
+ * only when the token is extra (the decode itself doesn't carry a market today).
+ *
+ * Generation tokens (roman numerals, digits, platform codes like "II"/"3"/"XM3")
+ * are deliberately NOT here, so multi-generation models stay fully matchable;
+ * only MARKET qualifiers get the penalty.
+ *
+ * Grounded in prod data — the real Renault/Dacia catalog models that carry a
+ * region qualifier are: KADJAR ÇİN, CAPTUR II ÇİN, KOLEOS 2 - ÇİN, ARKANA RUSYA
+ * and X62 CHINE. ("EUROPE" appears too — ARKANA EUROPE, CAPTUR II EUROPE — but
+ * that is the mainstream market and is never penalized.)
+ */
+export const MARKET_QUALIFIERS = new Set<string>([
+  // China — "ÇİN" folds to CIN via diacritic normalization; CHINE is the French
+  // spelling seen in "X62 CHINE".
+  "CIN",
+  "CHINA",
+  "CHINE",
+  "CN",
+  // Russia
+  "RUSYA",
+  "RUSSIA",
+  "RU",
+  // Brazil
+  "BREZILYA",
+  "BRAZIL",
+  "BR",
+  // India
+  "HINDISTAN",
+  "INDIA",
+  "IN",
+  // Korea
+  "KORE",
+  "KOREA",
+  // Mexico / Mercosur
+  "MEKSIKA",
+  "MEXICO",
+  "MERCOSUR",
+  // Other regional
+  "GCC",
+  "USA",
+  "AMERIKA",
+]);
+
 export interface DecodedForMatch {
   model: string | null;
   modelYear: string | null;
@@ -37,14 +91,23 @@ export interface DecodedForMatch {
  *  tiebreak over the modern Egea. */
 export function modelTokens(s: string | null | undefined): string[] {
   if (!s) return [];
-  return s
-    .toUpperCase()
-    .replace(/\([^)]*\)/g, " ") // drop parenthetical year ranges
-    .replace(/\b(19[5-9]\d|20[0-3]\d)\b/g, " ") // drop bare YEARS 1950-2039; keep displacements
-    .replace(/[^A-Z0-9]+/g, " ")
-    .split(" ")
-    .map((t) => t.trim())
-    .filter((t) => t.length >= 2);
+  return (
+    s
+      .toUpperCase()
+      // Fold diacritics to ASCII so market qualifiers survive tokenization. Without
+      // this the Turkish "ÇİN" (China) loses its Ç/İ to the [^A-Z0-9] strip below
+      // and collapses to a dropped 1-char "N" — making "KADJAR ÇİN" tokenize
+      // IDENTICALLY to "KADJAR" and defeating both the specificity tiebreak and the
+      // market penalty. NFD + combining-mark removal maps ÇİN→CIN, Ş→S, Ğ→G, Ö→O …
+      .normalize("NFD")
+      .replace(/\p{M}/gu, "") // strip the combining marks NFD split off
+      .replace(/\([^)]*\)/g, " ") // drop parenthetical year ranges
+      .replace(/\b(19[5-9]\d|20[0-3]\d)\b/g, " ") // drop bare YEARS 1950-2039; keep displacements
+      .replace(/[^A-Z0-9]+/g, " ")
+      .split(" ")
+      .map((t) => t.trim())
+      .filter((t) => t.length >= 2)
+  );
 }
 
 /** Parse a year range from free text. Returns {start, end} where end===null
@@ -77,8 +140,18 @@ function yearInRange(year: number, range: { start: number; end: number | null })
 /**
  * Pick the best catalog_vehicles row for a decoded model + year. A candidate
  * qualifies only when EVERY decoded model token appears in the candidate's
- * model tokens (subset match). Among qualifiers, prefer year-range overlap,
- * then the most specific (fewest extra tokens). Returns the id or null.
+ * model tokens (subset match). Scoring, strongest signal first:
+ *   1. Year-range overlap — the strongest signal; routes multi-generation models
+ *      (a 2022 "TIPO-EGEA" → the 2020+ MCA catalog, not the 2015-2021 one).
+ *   2. EXACT normalized model match (candidate tokens ≡ decoded tokens) — a
+ *      smaller bonus that breaks the tie in favour of a plain "KADJAR" over the
+ *      superset "KADJAR ÇİN" when no year distinguishes them.
+ *   3. Market-qualifier penalty — extra tokens that are region qualifiers (ÇİN,
+ *      RUSYA, …) make this a WRONG-market catalog; penalized hard so it only ever
+ *      wins as the sole candidate. Generation / other extra tokens keep only the
+ *      mild "fewer extras = more specific" penalty, so generations stay matchable.
+ *   4. Tiebreaks on equal score: richer catalog (more categories), then newer.
+ * Returns the id or null.
  */
 export function pickBestCatalogMatch(
   decoded: DecodedForMatch,
@@ -88,7 +161,12 @@ export function pickBestCatalogMatch(
   if (decTokens.length === 0) return null;
   const decYear = decoded.modelYear ? Number.parseInt(decoded.modelYear, 10) : null;
 
-  let best: { id: string; score: number; startYear: number } | null = null;
+  let best: {
+    id: string;
+    score: number;
+    categoryCount: number;
+    startYear: number;
+  } | null = null;
 
   for (const cand of candidates) {
     const candText = `${cand.model ?? ""} ${cand.year ?? ""}`;
@@ -98,21 +176,47 @@ export function pickBestCatalogMatch(
     // Subset requirement — all decoded model tokens must be present.
     if (!decTokens.every((t) => candSet.has(t))) continue;
 
+    const extraTokens = candTokens.filter((t) => !decTokens.includes(t));
+
     let score = 1000; // base for any qualifying candidate
+
+    // (1) EXACT normalized model-name match: candidate carries NO tokens beyond
+    // the decoded model. This breaks the tie in favour of the plain model over a
+    // superset (e.g. "KADJAR" over "KADJAR ÇİN" when neither has a distinguishing
+    // year). Kept SMALLER than the year-range swing (±700) on purpose, so a
+    // decisive model-year still routes generations correctly — a 2022 "TIPO-EGEA"
+    // must still land on the 2020+ MCA catalog, not the exact-named 2015-2021 one.
+    if (extraTokens.length === 0) score += 300;
+
+    // (2) Year-range overlap.
     const range = parseYearRange(candText);
     if (decYear && Number.isFinite(decYear) && range) {
       score += yearInRange(decYear, range) ? 500 : -200;
     }
-    // Tiebreak: fewer extra tokens = more specific match.
-    const extra = candTokens.filter((t) => !decTokens.includes(t)).length;
-    score -= extra;
 
-    // Final tiebreak on equal score: prefer the newer catalog. Without a year
-    // signal an ambiguous model (e.g. bare "TIPO") must never fall back onto an
-    // ancient generation — modern is the overwhelmingly likelier intent.
+    // (3) Extra-token penalties. A region-qualifier extra token means the
+    // candidate is a wrong-market catalog for a decode that carries no market —
+    // penalize hard so it can only win when it is the ONLY candidate. Everything
+    // else (generation numerals, platform codes) keeps the mild specificity
+    // penalty, leaving multi-generation models fully matchable.
+    const marketExtras = extraTokens.filter((t) => MARKET_QUALIFIERS.has(t)).length;
+    const otherExtras = extraTokens.length - marketExtras;
+    score -= otherExtras;
+    score -= marketExtras * 5000;
+
+    // (4) Tiebreaks on equal score: prefer the richer catalog (more categories —
+    // the fuller catalog is the mainstream one), then the newer generation. A
+    // bare, year-less model (e.g. "TIPO") must never fall back onto an ancient
+    // generation — modern is the overwhelmingly likelier intent.
+    const categoryCount = cand.categoryCount ?? 0;
     const startYear = range?.start ?? 0;
-    if (!best || score > best.score || (score === best.score && startYear > best.startYear)) {
-      best = { id: cand.id, score, startYear };
+    const better =
+      !best ||
+      score > best.score ||
+      (score === best.score && categoryCount > best.categoryCount) ||
+      (score === best.score && categoryCount === best.categoryCount && startYear > best.startYear);
+    if (better) {
+      best = { id: cand.id, score, categoryCount, startYear };
     }
   }
 
@@ -137,6 +241,12 @@ export class VinpinCatalogMatcher {
         id: catalogVehicles.id,
         model: catalogVehicles.model,
         year: catalogVehicles.year,
+        // Category count feeds the richer-catalog tiebreak (more categories = the
+        // mainstream catalog). Correlated subquery keeps the row shape flat.
+        categoryCount: sql<number>`(
+          SELECT count(*)::int FROM categories
+          WHERE categories.catalog_vehicle_id = ${catalogVehicles.id}
+        )`,
       })
       .from(catalogVehicles)
       .where(