dj sets + lyrics: a mix that names its own tracks, and words pulled off the separated voice

src-tauri/src/lyrics.rs

@@ -1410,6 +1410,7 @@ pub async fn lyrics_get(
         let payload = payload_from(user, offset, duration_secs);
         if payload.status != "none" {
             return Ok(finish_synced(
+                &app,
                 payload,
                 entry.auto_sync_off,
                 path.as_deref(),
@@ -1514,6 +1515,7 @@ pub async fn lyrics_get(
         let payload = payload_from(doc, offset, duration_secs);
         if payload.status == "synced" {
             return Ok(finish_synced(
+                &app,
                 payload,
                 entry.auto_sync_off,
                 path.as_deref(),
@@ -1603,6 +1605,7 @@ pub async fn lyrics_get(
         // local copy may be the corrected one).
         if payload.status == "synced" || payload.status == "instrumental" {
             return Ok(finish_synced(
+                &app,
                 payload,
                 entry.auto_sync_off,
                 path.as_deref(),
@@ -1638,6 +1641,7 @@ pub async fn lyrics_get(
 
 /// Stamp the auto-sync fields onto a winning synced payload.
 async fn finish_synced(
+    app: &tauri::AppHandle,
     mut payload: LyricsPayload,
     auto_sync_off: bool,
     path: Option<&str>,
@@ -1649,6 +1653,35 @@ async fn finish_synced(
         let first = payload.synced.first().map(|l| l.time_ms);
         payload.auto_offset_ms =
             compute_auto_offset(path, duration_secs, source_duration_secs, first).await;
+        // Second evidence source, consulted only when the corroborated
+        // duration/silence probe found nothing: the word-timing artifact.
+        // When forced alignment has already measured this file against this
+        // doc and the sung words consistently start N ms from their cues,
+        // that median is the doc's shift. Cache read only — never computed
+        // here. (Aligned WORD times themselves are absolute and exempt from
+        // this offset; only line cues move. See ArtifactLine::ctc.)
+        if payload.auto_offset_ms == 0 {
+            if let Some(p) = path {
+                // The exact shape the word-timing request sends: original
+                // indexes, empties dropped — the artifact key depends on it.
+                let lines: Vec<crate::lyric_prep::PrepLine> = payload
+                    .synced
+                    .iter()
+                    .enumerate()
+                    .map(|(i, l)| crate::lyric_prep::PrepLine {
+                        index: i as u32,
+                        time_ms: l.time_ms as i64,
+                        text: l.text.clone(),
+                    })
+                    .filter(|l| !l.text.trim().is_empty())
+                    .collect();
+                if let Some(shift) =
+                    crate::lyric_prep::cue_shift_evidence(app, Path::new(p), &lines)
+                {
+                    payload.auto_offset_ms = shift;
+                }
+            }
+        }
     }
     payload
 }