Teach local AI from user category corrections

- Add MerchantCorrection model: upsert by merchantName, Category enum
- Check corrections DB first in suggestCategoryForMerchant (source: "learned",
  no confirmation required); falls through to rules then Ollama if no match
- Inject recent corrections as few-shot examples in the Ollama prompt so the
  model improves even for merchants not yet explicitly corrected
- Add POST /categories/correct route to persist corrections
- Detect category override on form save (suggestedCategory !== chosen category)
  and silently fire a correction — no extra UX required
- Fix test isolation: beforeEach re-applies vi.fn() defaults after restoreAllMocks

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

prisma/schema.prisma

@@ -42,6 +42,14 @@ model PaySchedule {
   createdAt   DateTime @default(now())
 }
 
+model MerchantCorrection {
+  id           String   @id @default(cuid())
+  merchantName String   @unique
+  category     Category
+  createdAt    DateTime @default(now())
+  updatedAt    DateTime @updatedAt
+}
+
 model MonthlyInsight {
   id              String   @id @default(cuid())
   month           String   @unique