fix: restore default streaming params and use minimal silence flush

- QwenASRRecognizer: revert chunk_sec/rollback/unfixed_chunks to C API
  defaults (2.0s/5/2) — custom values (1.5/3/1) caused decoder
  degeneracy on long audio
- ASREngine.flush: push 0.1s silence + finalize instead of reading
  stable+unfixed directly, ensuring rollback tokens are committed

Streaming CER improved from 0.0968 to 0.0519 on 67-sample benchmark.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: ZhaoChaoqun

Sources/ASREngine.swift

@@ -201,11 +201,13 @@ class QwenASREngine: ASREngine {
                 return
             }
 
-            // 直接取 stable + unfixed 作为最终结果，不推送 silence
-            // 推送 silence + finalize 会导致 decoder 在静音上 hallucinate（重复之前的内容）
-            let stableText = self.recognizer.getResult()
-            let unfixedText = self.recognizer.getUnfixed() ?? ""
-            let result = stableText + unfixedText
+            // 推送极少量 silence（0.1s）+ finalize，让 Rust 处理尾部不足一个 chunk 的音频
+            // 并 commit rollback 窗口内的 token。0.1s silence 不会导致 decoder hallucinate，
+            // 而之前的 1s silence 在长音频上导致 decoder 重复之前的内容。
+            let minimalSilence = [Float](repeating: 0.0, count: 1600)
+            _ = self.recognizer.pushAudio(samples: minimalSilence, finalize: true)
+
+            let result = self.recognizer.getResult()
             self.recognizer.reset()
 
             DispatchQueue.main.async {

Sources/QwenASRRecognizer.swift

@@ -27,11 +27,10 @@ class QwenASRStreamRecognizer {
         // 强制简体中文输出，避免自动检测导致输出繁体字
         qwen_asr_set_language(engine, "chinese")
 
-        // 流式参数：chunk_sec 和 max_new_tokens 保持默认以获得更好的准确度
-        // rollback 和 unfixed_chunks 适度降低以减少首字延迟
-        qwen_asr_stream_set_chunk_sec(engine, 1.5)
-        qwen_asr_stream_set_rollback(engine, 3)
-        qwen_asr_stream_set_unfixed_chunks(engine, 1)
+        // 流式参数：使用 C API 默认值
+        qwen_asr_stream_set_chunk_sec(engine, 2.0)
+        qwen_asr_stream_set_rollback(engine, 5)
+        qwen_asr_stream_set_unfixed_chunks(engine, 2)
         qwen_asr_stream_set_max_new_tokens(engine, 32)
 
         streamState = qwen_asr_stream_new()