fix: address final review findings

- Comment: "~4 characters" → "~4 bytes" (len() counts bytes, not runes)
- Use utf8.RuneStart from stdlib instead of custom isUTF8Start helper
- Skip diff block entirely when Diff is empty (handles edge cases:
  draft→ready with no delta, force-push matching base, etc.)

budget/budget.go

@@ -8,6 +8,7 @@ package budget
 import (
 	"fmt"
 	"strings"
+	"unicode/utf8"
 )
 
 // modelLimit pairs a model name prefix with its context window size.
@@ -38,7 +39,7 @@ const diffTooLargeMarker = "... [diff too large for context window — review ma
 const userMetaTruncMarker = "\n... [description truncated] ..."
 
 // EstimateTokens estimates the number of tokens in a string.
-// Uses the rough heuristic of ~4 characters per token, which is
+// Uses the rough heuristic of ~4 bytes per token, which is
 // conservative for English text and code.
 func EstimateTokens(s string) int {
 	return len(s) / 4
@@ -188,9 +189,11 @@ func buildResult(s Sections, trimmed []string, estTokens int) Result {
 		usr.WriteString(s.FileContext)
 		usr.WriteString("\n")
 	}
-	usr.WriteString("\n### Diff (changes to review)\n\n```diff\n")
-	usr.WriteString(s.Diff)
-	usr.WriteString("\n```\n")
+	if s.Diff != "" {
+		usr.WriteString("\n### Diff (changes to review)\n\n```diff\n")
+		usr.WriteString(s.Diff)
+		usr.WriteString("\n```\n")
+	}
 
 	if len(trimmed) > 0 {
 		usr.WriteString("\n⚠️ Note: Context was trimmed to fit model limits. Dropped: ")
@@ -212,15 +215,8 @@ func truncateUTF8(s string, maxBytes int) string {
 	if len(s) <= maxBytes {
 		return s
 	}
-	// Walk backwards from maxBytes to find a valid UTF-8 boundary
-	for maxBytes > 0 && !isUTF8Start(s[maxBytes]) {
+	for maxBytes > 0 && !utf8.RuneStart(s[maxBytes]) {
 		maxBytes--
 	}
 	return s[:maxBytes]
 }
-
-// isUTF8Start returns true if b is a valid start byte for a UTF-8 sequence
-// (single-byte ASCII or multi-byte lead byte, not a continuation byte).
-func isUTF8Start(b byte) bool {
-	return b&0xC0 != 0x80
-}