fix(memory): use rune-based truncation for UTF-8 safety

truncateSnippet sliced bytes directly, which could split multi-byte
UTF-8 characters. Switch to []rune so truncation always respects
character boundaries.

internal/memory/provider/builtin.go

@@ -380,10 +380,11 @@ func (p *BuiltinProvider) Usage(ctx context.Context, filters map[string]any) (Us
 
 func truncateSnippet(s string, n int) string {
 	trimmed := strings.TrimSpace(s)
-	if len(trimmed) <= n {
+	runes := []rune(trimmed)
+	if len(runes) <= n {
 		return trimmed
 	}
-	return strings.TrimSpace(trimmed[:n]) + "..."
+	return strings.TrimSpace(string(runes[:n])) + "..."
 }
 
 func deduplicateItems(items []MemoryItem) []MemoryItem {

internal/memory/provider/builtin_test.go

@@ -0,0 +1,54 @@
+package provider
+
+import (
+	"testing"
+	"unicode/utf8"
+)
+
+func TestTruncateSnippet_ASCII(t *testing.T) {
+	t.Parallel()
+	got := truncateSnippet("hello world", 5)
+	if got != "hello..." {
+		t.Fatalf("expected %q, got %q", "hello...", got)
+	}
+}
+
+func TestTruncateSnippet_NoTruncation(t *testing.T) {
+	t.Parallel()
+	got := truncateSnippet("short", 100)
+	if got != "short" {
+		t.Fatalf("expected %q, got %q", "short", got)
+	}
+}
+
+func TestTruncateSnippet_CJK(t *testing.T) {
+	t.Parallel()
+	// 5 CJK characters (15 bytes in UTF-8), truncate to 3 runes.
+	got := truncateSnippet("你好世界啊", 3)
+	if !utf8.ValidString(got) {
+		t.Fatalf("result is not valid UTF-8: %q", got)
+	}
+	if got != "你好世..." {
+		t.Fatalf("expected %q, got %q", "你好世...", got)
+	}
+}
+
+func TestTruncateSnippet_Emoji(t *testing.T) {
+	t.Parallel()
+	// Emoji are 4 bytes each in UTF-8.
+	got := truncateSnippet("😀😁😂🤣😃", 2)
+	if !utf8.ValidString(got) {
+		t.Fatalf("result is not valid UTF-8: %q", got)
+	}
+	if got != "😀😁..." {
+		t.Fatalf("expected %q, got %q", "😀😁...", got)
+	}
+}
+
+func TestTruncateSnippet_TrimWhitespace(t *testing.T) {
+	t.Parallel()
+	got := truncateSnippet("  hello  ", 100)
+	if got != "hello" {
+		t.Fatalf("expected %q, got %q", "hello", got)
+	}
+}