fix: utf8 decoder, finish_reason override, and chunk boundary handling

Fix: streaming proxy
- Use utf8.decoder.bind() instead of per-chunk utf8.decode() to avoid
  FormatException on incomplete UTF-8 sequences across TCP chunks

Fix: finish_reason consistency
- Override finish_reason to tool_calls when toolCalls is non-empty
  even if upstream reports finishReason stop
- Prevents OpenCode from treating a tool-call response as conversation
  end and starting extra rounds

Fix: non-streaming handler
- Add leftover buffer for chunk boundary handling, matching streaming
  handler behavior

Author: Khip01

lib/src/server/proxy.dart

@@ -189,8 +189,7 @@ class ProxyServer {
         bool sentAssistantRole = false;
         final toolCalls = <Map<String, dynamic>>[];
 
-        await for (final raw in upstreamRes) {
-          final chunk = utf8.decode(raw);
+        await for (final chunk in utf8.decoder.bind(upstreamRes)) {
           final parts = (leftover + chunk).split('\n');
           leftover = parts.removeLast();
 
@@ -280,6 +279,10 @@ class ProxyServer {
           }
         }
 
+        if (toolCalls.isNotEmpty && finishReason == 'stop') {
+          finishReason = 'tool_calls';
+        }
+
         _sendSSE(response, {
           'id': 'cmpl-${DateTime.now().millisecondsSinceEpoch}',
           'object': 'chat.completion.chunk',
@@ -394,10 +397,14 @@ class ProxyServer {
       int inputTokens = 0;
       int outputTokens = 0;
       final toolCalls = <Map<String, dynamic>>[];
+      String leftover = '';
 
       await for (final raw in upstreamRes) {
         final chunk = utf8.decode(raw);
-        for (final line in chunk.split('\n')) {
+        final parts = (leftover + chunk).split('\n');
+        leftover = parts.removeLast();
+
+        for (final line in parts) {
           if (line.trim().isEmpty) continue;
           try {
             final event = jsonDecode(line) as Map<String, dynamic>;
@@ -424,6 +431,10 @@ class ProxyServer {
         }
       }
 
+      if (toolCalls.isNotEmpty && finishReason == 'stop') {
+        finishReason = 'tool_calls';
+      }
+
       final choice = <String, dynamic>{
         'index': 0,
         'message': {