fix: AI 复习变式请求字段与校验一致；转写后删除不保留的录音

- generateReviewVariant 的指令要求 stimulus/answer/acceptedAnswers 等字段，
  而 decodeGeneratedReviewVariant 只接受 schemaVersion、variantId、
  targetItemId、prompt、expectedAnswer 五个字段，导致 AI 变式始终校验失败。
  现在指令只要求这五个字段，并补充了请求与解码一致性的测试。
- finishVoiceInput(keepAudio: false) 转写完成（或页面已关闭）后删除音频，
  补练页和测评页的语音输入不再在本机残留录音文件。

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

kouyu_english/lib/widgets/voice_answer.dart

@@ -56,7 +56,7 @@ mixin VoiceAnswerMixin<T extends StatefulWidget> on State<T> {
   /// [onTranscript] runs inside `setState` with the trimmed, non-empty text.
   /// [afterTranscribe] runs once transcription has finished, whether or not
   /// anything was recognized. With [keepAudio] the answer audio also becomes
-  /// the playback [recordingPath].
+  /// the playback [recordingPath]; otherwise it is deleted once transcribed.
   Future<void> finishVoiceInput({
     required void Function(String text) onTranscript,
     VoidCallback? afterTranscribe,
@@ -64,7 +64,10 @@ mixin VoiceAnswerMixin<T extends StatefulWidget> on State<T> {
     String noSpeech = noSpeechMessage,
   }) async {
     final path = await VoiceService.instance.stopRecording();
-    if (!mounted) return;
+    if (!mounted) {
+      if (!keepAudio) await VoiceService.instance.deleteRecording(path);
+      return;
+    }
     setState(() {
       aiVoiceRecording = false;
       listening = false;
@@ -79,6 +82,8 @@ mixin VoiceAnswerMixin<T extends StatefulWidget> on State<T> {
       endpoint: config.endpoint,
       model: config.model,
     );
+    // Without [keepAudio] the audio only served transcription.
+    if (!keepAudio) await VoiceService.instance.deleteRecording(path);
     if (!mounted) return;
     final text = transcribed?.trim() ?? '';
     setState(() {