Let manual corrections outrank the model's own precedent

History rows now carry a source label: manual edits and merchant rules
are the user's decisions, ranked ahead of equally similar rows the
model classified itself and guaranteed slots in a full history window.
Without the distinction, precedent fed the model its own uncorrected
answers as majority evidence, so a correction never won against the
rows it was meant to fix. Both system prompts state that user entries
outrank ai entries. Alias write-back on manual merchant links and the
per-merchant usual category already learned locally; this closes the
loop for categories and tags.

internal/classification/candidates.go

@@ -160,6 +160,10 @@ type promptHistory struct {
 	CategoryID   string   `json:"category_id"`
 	MerchantID   string   `json:"merchant_id,omitempty"`
 	TagIDs       []string `json:"tag_ids"`
+	// Source separates the user's own decisions ("user") from earlier model
+	// output ("ai"): without the distinction, precedent feeds the model its
+	// own past answers as evidence and a manual correction never wins.
+	Source string `json:"source"`
 }
 type candidateSet struct {
 	categories  []categoryPrompt
@@ -314,10 +318,16 @@ func answerSchema(d domain.Dataset, kind string) map[string]any {
 	return retrieve("", kind, d, nil, nil).schema()
 }
 
+// history selects precedent for the prompt: the nearest rows by word overlap,
+// filled out with the most recent. The user's own decisions — manual edits
+// and locally applied merchant rules — outrank rows the model classified
+// itself, so one correction beats any number of uncorrected AI answers for
+// the same payee.
 func history(f domain.Facts, d domain.Dataset, clean func(string) string, limit int) []promptHistory {
 	type row struct {
 		tx    domain.Transaction
 		score int
+		user  bool
 	}
 	rows := []row{}
 	for _, tx := range d.Transactions {
@@ -325,19 +335,50 @@ func history(f domain.Facts, d domain.Dataset, clean func(string) string, limit
 		if tx.Facts.ID == f.ID || e.Kind == "transfer" || e.CategoryID == "" || e.CategoryID == domain.ExpenseFallback || e.CategoryID == domain.IncomeFallback {
 			continue
 		}
-		rows = append(rows, row{tx: tx, score: similarity(f.RawDescription+" "+f.Counterparty, tx.Facts.RawDescription+" "+tx.Facts.Counterparty)})
+		source := tx.Enrichment.Classification.Source
+		rows = append(rows, row{
+			tx:    tx,
+			score: similarity(f.RawDescription+" "+f.Counterparty, tx.Facts.RawDescription+" "+tx.Facts.Counterparty),
+			user:  source == "manual" || source == "rule",
+		})
 	}
 	sort.Slice(rows, func(i, j int) bool {
 		if rows[i].score != rows[j].score {
 			return rows[i].score > rows[j].score
 		}
+		if rows[i].user != rows[j].user {
+			return rows[i].user
+		}
 		if rows[i].tx.Facts.BookingDate != rows[j].tx.Facts.BookingDate {
 			return rows[i].tx.Facts.BookingDate > rows[j].tx.Facts.BookingDate
 		}
 		return rows[i].tx.Facts.ID < rows[j].tx.Facts.ID
 	})
 	if limit > 0 && len(rows) > limit {
-		rows = rows[:limit]
+		// Never let recent AI output crowd every correction out of a full
+		// window: user rows keep their slots ahead of equally similar AI rows.
+		kept := make([]row, 0, limit)
+		users := 0
+		for _, r := range rows {
+			if r.user {
+				users++
+			}
+		}
+		userBudget := min(users, limit/2)
+		aiBudget := limit - userBudget
+		for _, r := range rows {
+			if r.user && userBudget > 0 {
+				kept = append(kept, r)
+				userBudget--
+			} else if !r.user && aiBudget > 0 {
+				kept = append(kept, r)
+				aiBudget--
+			} else if r.user && aiBudget > 0 {
+				kept = append(kept, r)
+				aiBudget--
+			}
+		}
+		rows = kept
 	}
 	out := make([]promptHistory, 0, len(rows))
 	for _, row := range rows {
@@ -345,11 +386,16 @@ func history(f domain.Facts, d domain.Dataset, clean func(string) string, limit
 		if tags == nil {
 			tags = []string{}
 		}
+		source := "ai"
+		if row.user {
+			source = "user"
+		}
 		out = append(out, promptHistory{
 			Date: row.tx.Facts.BookingDate, Amount: string(row.tx.Facts.Amount),
 			Description: clean(row.tx.Facts.RawDescription), Counterparty: clean(row.tx.Facts.Counterparty),
 			CategoryID: row.tx.Enrichment.CategoryID, MerchantID: row.tx.Enrichment.MerchantID,
 			TagIDs: append([]string{}, tags...),
+			Source: source,
 		})
 	}
 	return out