feat(asr): REQ-264 DTW 漂移校正生产接线（融合前校正+测量门槛）

Author: Aparencia

app/src-tauri/src/dtw_align.rs

@@ -169,6 +169,56 @@ pub fn alignment_accuracy(alignment: &Alignment, subtitles: &[SubtitleSegment],
     matched as f32 / alignment.pairs.len() as f32
 }
 
+/// 融合前漂移校正的判定结果（REQ-264 生产接线）。
+///
+/// @ai-context: corrected 恒可用（未校正时为原样克隆）——调用方无需分支；
+///              measured_ms=全程测量值；applied_ms=实际采纳的平移量
+///              （仅当 |measured| 落在 [min_apply, max_apply] 内才校正——
+///              过小=抖动噪声不值得动、过大=首尾强对齐误差/数据异常应拒动）。
+#[derive(Debug, Clone, PartialEq)]
+pub struct DriftCorrection {
+    /// 校正后的字幕流（未校正 = 原样克隆）。
+    pub corrected: Vec<SubtitleSegment>,
+    /// 测量的漂移（ms；负=字幕超前）。数据不足 → None。
+    pub measured_ms: Option<i64>,
+    /// 实际采纳的平移量（None=未校正）。
+    pub applied_ms: Option<i64>,
+}
+
+/// 生产接线包装（纯函数）：字幕流太少不估；漂移在合理带内才校正。
+pub fn correct_drift_if_any(
+    subtitles: &[SubtitleSegment],
+    asr_segments: &[TranscriptSegment],
+    min_segs: usize,
+    min_apply_ms: i64,
+    max_apply_ms: i64,
+) -> DriftCorrection {
+    let identity = || DriftCorrection {
+        corrected: subtitles.to_vec(),
+        measured_ms: None,
+        applied_ms: None,
+    };
+    if subtitles.len() < min_segs || asr_segments.len() < min_segs {
+        return identity();
+    }
+    let alignment = align_sequences(subtitles, asr_segments);
+    let Some(drift) = estimate_drift_ms(subtitles, asr_segments, &alignment) else {
+        return identity();
+    };
+    if drift.abs() < min_apply_ms || drift.abs() > max_apply_ms {
+        return DriftCorrection {
+            corrected: subtitles.to_vec(),
+            measured_ms: Some(drift),
+            applied_ms: None,
+        };
+    }
+    DriftCorrection {
+        corrected: correct_subtitles(subtitles, drift),
+        measured_ms: Some(drift),
+        applied_ms: Some(drift),
+    }
+}
+
 /// 单测独立文件（保持本文件 ≤300 行，AGENTS.md §3）。
 #[cfg(test)]
 #[path = "dtw_align_tests.rs"]

app/src-tauri/src/dtw_align_tests.rs

@@ -154,3 +154,56 @@ fn correct_subtitles_clamps_negative() {
     assert_eq!(corrected[0].end_ms, 0);
     assert_eq!(corrected[1].start_ms, 6_000);
 }
+
+// ── correct_drift_if_any（REQ-264 生产接线包装）──
+
+#[test]
+fn correction_applies_within_band() {
+    // Arrange：字幕超前 800ms（在 [200, 15000] 带内）
+    let (subs, asrs) = synthetic_session(6, 800, 0);
+    // Act
+    let r = correct_drift_if_any(&subs, &asrs, 3, 200, 15_000);
+    // Assert：测量 ≈-800 且已采纳；回校后误差归零
+    let m = r.measured_ms.unwrap();
+    assert!((m + 800).abs() <= 100, "测得 {}", m);
+    assert_eq!(r.applied_ms, Some(m));
+    for (i, s) in r.corrected.iter().enumerate() {
+        let err = (s.start_ms as i64 - asrs[i].start_ms as i64).abs();
+        assert!(err <= 100, "第 {} 段误差 {}ms", i, err);
+    }
+}
+
+#[test]
+fn correction_skips_small_jitter() {
+    // Arrange：仅 50ms 偏移（抖动噪声，不值得动）
+    let (subs, asrs) = synthetic_session(6, 50, 0);
+    // Act
+    let r = correct_drift_if_any(&subs, &asrs, 3, 200, 15_000);
+    // Assert：测量存在但不采纳（原样克隆）
+    assert!(r.measured_ms.is_some());
+    assert_eq!(r.applied_ms, None);
+    assert_eq!(r.corrected, subs);
+}
+
+#[test]
+fn correction_rejects_absurd_drift() {
+    // Arrange：超大步进（-40s，首尾强对齐误差/异常数据）
+    let (subs, asrs) = synthetic_session(6, -40_000, 0);
+    // Act
+    let r = correct_drift_if_any(&subs, &asrs, 3, 200, 15_000);
+    // Assert：不采纳（但保留测量，供统计）
+    assert!(r.measured_ms.is_some());
+    assert_eq!(r.applied_ms, None);
+}
+
+#[test]
+fn correction_skips_when_too_few_segments() {
+    // Arrange：字幕 2 段 < min_segs=3（残片不足以估漂移）
+    let (subs, asrs) = synthetic_session(2, 800, 0);
+    // Act
+    let r = correct_drift_if_any(&subs, &asrs, 3, 200, 15_000);
+    // Assert：不测量不校正（原样）
+    assert_eq!(r.measured_ms, None);
+    assert_eq!(r.applied_ms, None);
+    assert_eq!(r.corrected, subs);
+}

app/src-tauri/src/live_keyframes.rs

@@ -325,7 +325,20 @@ pub fn rewrite_with_fusion(
         eprintln!("[Fusion] 会话 {} 无字幕段，短路跳过融合（ASR 段原样保留）", session_id);
         return Ok(());
     }
-    let fused = merge_transcript(subtitles, asr_segments, 0);
+    // REQ-264（v0.20.1）：融合前做字幕漂移校正——字幕 OCR 流时间轴滞后/超前
+    // 是"字幕权威路线错位"主因；测量带内才采纳（过小=抖动噪声、过大=首尾强
+    // 对齐误差/异常拒动；段数不足不估——与 harness 会话信道门槛同口径）
+    const MIN_FUSION_DTW_SEGS: usize = 3;
+    const MIN_APPLY_DRIFT_MS: i64 = 200;
+    const MAX_APPLY_DRIFT_MS: i64 = 15_000;
+    let correction =
+        crate::dtw_align::correct_drift_if_any(subtitles, asr_segments, MIN_FUSION_DTW_SEGS, MIN_APPLY_DRIFT_MS, MAX_APPLY_DRIFT_MS);
+    if let Some(d) = correction.applied_ms {
+        eprintln!("[Fusion] 会话 {} 字幕漂移校正 {:+}ms（{} 段）", session_id, d, subtitles.len());
+    } else if correction.measured_ms.is_some() {
+        eprintln!("[Fusion] 会话 {} 字幕漂移测量 {:+}ms 未采纳（带外或段数不足）", session_id, correction.measured_ms.unwrap());
+    }
+    let fused = merge_transcript(&correction.corrected, asr_segments, 0);
     if fused.is_empty() {
         return Ok(());
     }