rtx-csm: --decoder-steering-layers — layer subset fixes repetition

Mirror of --steering-layers for the decoder: comma-separated layer
indices to actually steer (decoder has 4 layers; useful subsets are
[3], [2,3], [1,2]).

Sweep at seed 7, [email protected]:
  [0,1,2,3]  cos 0.86  WER 0.93  "I'm not that tall ×3" ← repetition
  [3]        cos 0.58  WER 0.86  "I'll be off and offense..."
  [2,3]      cos 0.80  WER 1.43  "My daughter, Penny Ryan, and I have…"
  [0,1]      cos 0.82  WER 1.57  "And I'll check on them..."
  [1,2]      cos 0.81  WER 1.43  "I'm going to call him an X-Man..."

The repetition is specific to all-layers-at-once steering. Any 2-layer
subset eliminates it while preserving most of the cosine boost. Same
pattern as the backbone's [8,10,12] finding: partial perturbation
lets the unsteered layers act as a stabilizing prior.

[2,3] (decoder last 2) is the new recommended recipe — best cosine
of the no-repetition subsets and the longest fluent transcript.
Documented in docs/perf_history.md.

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

crates/models/rtx-csm/docs/perf_history.md

@@ -209,6 +209,30 @@ Both fail differently; neither produces production-quality emotional
 speech in single-shot. The N-seed picker (`emotional_speech_n.sh`)
 is still the right consumer interface — it doesn't care which
 failure mode produced the bad samples, just sorts by WER.
+
+**Decoder layer subset fixes the repetition.** The "I'm not that
+tall ×3" failure at all-4-layers @ scale 1.0 disappears entirely
+when steering is restricted to a 2-layer subset. Same pattern as
+the backbone's `[8,10,12]` finding: perturbing all layers
+simultaneously is the bad regime; partial perturbation lets the
+unsteered layers act as a stabilizing prior.
+
+  layer subset    cos    WER    repetition?
+  [0,1,2,3] (all) 0.86   0.93   yes ("I'm not that tall ×3")
+  [3] only        0.58   0.86   no
+  [2,3] last 2    0.80   1.43   no — "My daughter, Penny Ryan…"
+  [0,1] first 2   0.82   1.57   no
+  [1,2] middle 2  0.81   1.43   no
+
+`[2,3]` (decoder's last two layers, the analogue of backbone's
+`[8,10,12]` mid-deep recipe) is the best balance — preserves the
+cosine boost while eliminating the repetition. New decoder recipe:
+
+```
+--decoder-steering-vec /path/<emotion>.safetensors
+--decoder-steering-scale 1.0
+--decoder-steering-layers 2,3
+```
 - *RAVDESS corpus shape biases output length*: clips are ~3 s
   ("Kids are talking by the door") so the steering biases the
   model toward terse outputs (one or two words). Longer-utterance

crates/models/rtx-csm/examples/generate.rs

@@ -150,6 +150,13 @@ struct Cli {
     #[arg(long, default_value_t = 1.0)]
     decoder_steering_scale: f32,
 
+    /// Comma-separated layer indices to actually steer on the decoder
+    /// (e.g. "2,3"). The decoder has 4 layers (CSM-1B); useful subsets
+    /// to try: `3` (deepest), `2,3` (last two), `1,2` (middle two).
+    /// Defaults to all 4 layers.
+    #[arg(long, value_delimiter = ',')]
+    decoder_steering_layers: Option<Vec<usize>>,
+
     /// Per-frame CFG scale schedule (Selective CFG, arXiv 2509.19668).
     /// Forms:
     ///   - `const:<scale>`               (constant scale, equivalent to --cfg-scale)
@@ -247,11 +254,14 @@ fn main() -> Result<()> {
         let mut steering =
             rtx_csm::steering::LayerSteering::load_safetensors(dec_path, n, &device)?;
         steering.set_scale(cli.decoder_steering_scale);
+        if let Some(layers) = cli.decoder_steering_layers.as_ref() {
+            steering.restrict_to_layers(layers);
+        }
         tracing::info!(
-            "loaded decoder steering from {} (scale={}, layers={})",
+            "loaded decoder steering from {} (scale={}, active_layers={:?})",
             dec_path.display(),
             cli.decoder_steering_scale,
-            n,
+            steering.active_layers(),
         );
         generator.set_decoder_steering(Some(steering))?;
     }