翻译复读治理: 初译批次4→2、逐条isEcho复读检测(归一化+二元组Jaccard)拒写抄原文、二次翻译同规则、提示词显式禁抄并允许空串

app/src/main/java/com/k2fsa/sherpa/onnx/MainActivity.kt

@@ -309,7 +309,7 @@ class MainActivity : AppCompatActivity() {
     // KV 污染 workaround(2026-09-14,不改 .so):nativeGenerate 不清 KV,
     // 同引擎连续 generate 会复读上一条 → 每"批"generate 前重建引擎。
     // 初译从逐句改为小批量编号翻译(凑 4 句或 2.5s 定时封批)。
-    private val initialBatchMax = 4
+    private val initialBatchMax = 2   // 4 句批后期句常退化成抄原文;2 句批明显更稳
     private val initialBatchDelayMs = 2_500L
     private val initialPending = ArrayList<Pair<Long, String>>()
     private var initialBatchTimer: Runnable? = null
@@ -1935,7 +1935,9 @@ class MainActivity : AppCompatActivity() {
             uiHandler.post {
                 val touched = HashSet<SubtitleLine>()
                 refs.forEach { s2 ->
-                    map[s2.id.toString()]?.takeIf { it.isNotBlank() }?.let { tr ->
+                    map[s2.id.toString()]?.takeIf {
+                        it.isNotBlank() && !isEcho(s2.text, it)
+                    }?.let { tr ->
                         if (s2.trans != tr) {
                             s2.trans = tr
                             s2.line?.let { touched.add(it) }
@@ -1995,8 +1997,10 @@ class MainActivity : AppCompatActivity() {
             if (map == null || !plausibleTranslations(map.values)) return@post
             uiHandler.post {
                 val touched = HashSet<SubtitleLine>()
-                batch.forEach { (sid, _) ->
+                batch.forEach { (sid, src) ->
-                    map[sid.toString()]?.takeIf { it.isNotBlank() }?.let { tr ->
+                    map[sid.toString()]?.takeIf {
+                        it.isNotBlank() && !isEcho(src, it)
+                    }?.let { tr ->
                         findSentence(sid)?.let { s2 ->
                             if (s2.trans != tr) {
                                 s2.trans = tr
@@ -2028,7 +2032,9 @@ class MainActivity : AppCompatActivity() {
                 systemPrompt = "你是专业翻译。逐句翻译下面编号句子为" + targetName +
                         "。各句相互独立,必须逐句对应翻译,不得合并、不得重复、不得解释。" +
                         "只输出JSON对象,键是编号字符串,值是译文," +
-                        "例如 {\"1\":\"…\",\"2\":\"…\"}。不得增删编号。",
+                        "例如 {\"1\":\"…\",\"2\":\"…\"}。不得增删编号。" +
+                        "严禁把原句原样作为译文输出;确实无法翻译的句子," +
+                        "值输出空字符串。",
                 userPrompt = "【句子】\n$numbered",
                 maxNewTokens = (160 * items.size).coerceAtMost(900),
                 temperature = 0f,
@@ -2050,6 +2056,30 @@ class MainActivity : AppCompatActivity() {
         return v.size < 2 || v.distinct().size * 2 >= v.size
     }
 
+    // 复读检测:译文与原文归一化后全等/长串包含/字符二元组 Jaccard≥0.72
+    // → 判为"把原文抄进译文",该条丢弃(维持原状),不写入字幕
+    private fun normalizeForEcho(x: String): String =
+        x.lowercase().filter { it.isLetterOrDigit() || it.code >= 0x3040 }
+
+    private fun isEcho(src: String, tr: String): Boolean {
+        val a = normalizeForEcho(src)
+        val b = normalizeForEcho(tr)
+        if (a.isEmpty() || b.isEmpty()) return false
+        if (a == b) return true
+        if (a.length >= 8 && (a.contains(b) || b.contains(a))) return true
+        if (a.length < 6 || b.length < 6) return false
+        fun bigrams(x: String): HashSet<String> {
+            val set = HashSet<String>()
+            for (i in 0 until x.length - 1) set.add(x.substring(i, i + 2))
+            return set
+        }
+        val ba = bigrams(a); val bb = bigrams(b)
+        if (ba.isEmpty() || bb.isEmpty()) return false
+        val inter = bb.count { ba.contains(it) }
+        val union = ba.size + bb.size - inter
+        return union > 0 && inter.toDouble() / union >= 0.72
+    }
+
     private val unloadRunnable = Runnable {
         val e = translateEngine ?: return@Runnable
         if (initialPending.isEmpty()) {