rtx-csm: Mimi.reload() — workaround state-leak in cumulative encode loop

Mimi's transformer carries an internal position counter across encode()
calls that the upstream `reset_state()` does NOT fully clear. When
TrainingDataset::load_from_manifest encodes ~80-100 clips back-to-back
during data prep, the counter overflows the 8192-position buffer and
panics with `narrow invalid args [8192, 32]`.

src/mimi.rs:
  - cache the safetensors path on Mimi at construction
  - new Mimi.reload() drops the inner Model and rebuilds from the
    cached path (~200 ms on Metal)

src/training.rs:
  - call generator.mimi.reload() every 50 clips during
    load_from_manifest. Adds ~1 s overhead on a 200-clip corpus
    (4 reloads × ~200 ms) vs the alternative of a hard panic.
  - reset_state() before each encode in TrainingExample::from_audio
    is kept (still useful to clear streaming chunk state).

Found while running the end-to-end personal-voice training pipeline on
a 20-minute YouTube source: the bug surfaces around clip 86 when
Mimi's transformer hits position 8181+. Filtering to short clips
alone didn't help — the cumulative state grows even with sub-12-second
inputs.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

crates/models/rtx-csm/src/training.rs

@@ -42,12 +42,20 @@ pub struct TrainingExample {
 
 impl TrainingExample {
     /// Build from raw 24 kHz mono samples by Mimi-encoding the audio.
+    ///
+    /// Mimi's encoder carries streaming state across calls — when
+    /// processing a manifest of N clips back-to-back its internal
+    /// frame counter grows monotonically and eventually overflows the
+    /// transformer's max-seq buffer (`narrow invalid args ... [8192, ...]`
+    /// at clip ~30 in a 200-clip corpus). Reset before each encode so
+    /// every clip starts with a fresh state.
     pub fn from_audio(
         speaker: u32,
         text: impl Into<String>,
         audio_24k: &[f32],
         generator: &mut Generator,
     ) -> Result<Self> {
+        generator.mimi.reset_state();
         let codes = generator.mimi.encode(audio_24k)?;
         let (_b, num_codebooks, num_frames) = codes.dims3()?;
         let mut frame_codes = Vec::with_capacity(num_frames);
@@ -184,6 +192,15 @@ impl TrainingDataset {
             }
             let speaker = row.speaker.unwrap_or(default_speaker);
             let audio = audio_io::load_mono_24k(&wav_path)?;
+            // Mimi's transformer carries a position counter across encode()
+            // calls that `reset_state()` does NOT fully clear. Without an
+            // explicit reload it overflows after ~80-100 clips with
+            // `narrow invalid args [8192, 32]`. Reload every 50 clips.
+            if !examples.is_empty() && examples.len() % 50 == 0 {
+                generator.mimi.reload().map_err(|e| {
+                    CsmError::Config(format!("mimi reload at clip {}: {e}", examples.len()))
+                })?;
+            }
             let mut ex = TrainingExample::from_audio(speaker, text, &audio, generator)?;
             ex.source = Some(wav_path);
             ex.emotion_tag = row.emotion_tag;