feat(canonical-lexicon): phrase-override katmanı — adjective+head yanlış-kovalar

Skordan önce yüksek-güven phrase kuralları: "<sistem> yağı$" → Yağlar (end-
anchor, "motor yağı filtresi/pompası" hariç), "(cam|silecek|kalorifer) motoru"
→ Elektrik, "zamanlama (dişli|zincir|kayış)" → Triger. ~1400 motor/şanzıman
yanlış-kovası + ~230 silecek-motoru + ~790 zamanlama düzelir. 29/29 test.

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

apps/api/src/jobs/__tests__/canonical-lexicon.spec.ts

@@ -225,6 +225,20 @@ describe("mapToCanonical", () => {
     expect(mapToCanonical("burkma cubugu").slug).toBe("suspansiyon");
   });
 
+  it("phrase overrides fix adjective+head mis-buckets", () => {
+    // "<system> yağı" = fluid, not the system
+    expect(mapToCanonical("motor yagi").slug).toBe("yaglar-sivilar");
+    expect(mapToCanonical("sanziman yagi").slug).toBe("yaglar-sivilar");
+    // but the PARTS keep their bucket (end-anchor excludes them)
+    expect(mapToCanonical("motor yagi filtresi").slug).toBe("filtreler");
+    expect(mapToCanonical("motor yagi pompasi").slug).toBe("motor");
+    // body electric motors → elektrik, not engine
+    expect(mapToCanonical("cam silecek motoru").slug).toBe("elektrik");
+    expect(mapToCanonical("cam motoru").slug).toBe("elektrik");
+    // timing gear → triger, not şanzıman
+    expect(mapToCanonical("zamanlama dislileri").slug).toBe("triger-zincir-rulman");
+  });
+
   it("returns null when no keyword matches", () => {
     expect(mapToCanonical("romork duzenegi cikarilabilir").slug).toBeNull();
     // genuinely generic Mercedes names stay unmapped (better than wrong bucket)

apps/api/src/jobs/canonical-lexicon.ts

@@ -317,7 +317,25 @@ export interface CanonicalMatch {
  * score; a runner-up within 1 point (no strong keyword decisively separated the
  * two) is flagged ambiguous. Confidence scales with the winning margin.
  */
+// High-confidence PHRASE overrides, checked before the bag-of-words score.
+// They fix "adjective + head-noun" cases where a strong keyword wins on the
+// modifier instead of the head: "motor yağı" scores motor(3) but the head noun
+// is the oil. Order doesn't matter (first match wins); keep each tight.
+const OVERRIDES: [RegExp, string][] = [
+  // "<system> yağı/yağları" (head = oil) → fluid. END-ANCHORED so the PARTS
+  // "motor yağı filtresi / pompası / soğutucusu" are NOT captured (they keep
+  // scoring to filtre / motor / soğutma).
+  [/(motor|sanziman|diferansiyel|direksiyon|hidrolik|fren|torpido) yag(i|lari)$/, "yaglar-sivilar"],
+  // body/electric motors (window/wiper/blower), not the engine.
+  [/(cam|silecek|kalorifer|kaput) motor/, "elektrik"],
+  // timing gear/chain/belt → triger, not şanzıman via the weak "dişli".
+  [/zamanlama (disli|zincir|kayis|seti|gergi|kasnak)/, "triger-zincir-rulman"],
+];
+
 export function mapToCanonical(folded: string): CanonicalMatch {
+  for (const [re, slug] of OVERRIDES) {
+    if (re.test(folded)) return { slug, ambiguous: false, confidence: 0.95 };
+  }
   const toks = tokens(folded);
   let bestSlug: string | null = null;
   let best = 0;