rtx-csm: CFG schedule × prompt sweep — robust default is linear

Sweep of 4 schedules × 4 prompts produced concrete data on schedule
sensitivity. step:3.0:1.5:12 (the Sprint 3 winner) catastrophically
fails dense prompts: lecture-style input → 0.08 s of audio (one
frame). step:2.0:1.0:8 produced the best single shot — near-verbatim
question rendering "Well, it had stem-wondered. Have you ever
wondered why we sometimes hear voices the way we do?" — but tanked
the lecture prompt (2.4 s "You").

linear:3.0:1.0:25 is the only schedule that's never the best AND
never the worst — graceful degradation across all four prompt
categories. Updates the recommended recipe in perf_history.md
(formerly step:3.0:1.5:12).

quality_eval: skip WavLM-SV scoring on clips shorter than 0.25 s
(emit -1 sentinel) — WavLM-SV's TDNN front-end requires a few
hundred samples and crashed mid-sweep on the 0.08 s clip. Now the
eval emits a row instead of bailing on the whole batch.

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

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

@@ -147,11 +147,43 @@ shipped on the FP backbone; quantized path unchanged.
 target/release/examples/generate \
     --text "..." --speaker 0 \
     --context-wav <reference>.wav --context-text "..." --context-speaker 1 \
-    --enable-cfg --cfg-schedule "step:3.0:1.5:12" \
+    --enable-cfg --cfg-schedule "linear:3.0:1.0:25" \
     --steering-vec /path/to/<emotion>_steering.safetensors \
     --steering-scale 0.3 --steering-layers 8,10,12
 ```
 
+(Updated from `step:3.0:1.5:12` after the sweep below — linear is more
+robust across prompt types.)
+
+**CFG schedule × prompt sweep (2026-04-29, single seed):**
+
+4 schedules × 4 prompts (short / medium / lecture / question), Amini
+context, scored on cosine vs Amini and WER vs prompt:
+
+| prompt          | const:1.5    | step:2:1:8    | step:3:1.5:12 | linear:3:1:25 |
+|-----------------|--------------|---------------|---------------|---------------|
+| `Hello world.`        | 0.69 / 1.00  | 0.71 / 4.00   | 0.79 / 3.50   | 0.82 / 4.50   |
+| medium narrative      | 0.94 / 1.08  | 0.90 / 1.00   | 0.66 / 1.00 *EOT 2.3 s* | 0.88 / 0.92   |
+| dense lecture         | **0.97 / 0.90** | 0.58 / 1.00 *EOT 2.4 s* | -1 / 1.00 *EOT 0.08 s* | 0.91 / **0.85** |
+| question              | 0.93 / 1.77  | 0.95 / **0.385** *near-verbatim* | 0.78 / 1.00 | 0.69 / 1.39 |
+
+*Format: cosine / WER. -1 sentinel = clip too short for WavLM-SV.*
+
+Single best shot: `step:2.0:1.0:8` × question →
+`"Well, it had stem-wondered. Have you ever wondered why we sometimes
+hear voices the way we do?"` (WER 0.385, cosine 0.95). Same step
+schedule on a dense lecture prompt produced 2.4 s of "You" — same
+schedule, opposite result.
+
+Takeaways:
+- `step:3.0:1.5:12` is **prompt-fragile** — destroys lecture and
+  medium-narrative prompts (premature EOT in 0.08–2.3 s); the earlier
+  Sprint 3 win on the meta "selective CFG paper recommends…" prompt
+  was a fluke of self-referential vocabulary.
+- `linear:3.0:1.0:25` is the most robust default — graceful across
+  all four prompt categories, never the worst.
+- `const:1.5` is the safe fallback when prompt category is unknown.
+
 **Rejected from this sprint** (with data):
 
 - *TTSDS2 metric* (arXiv 2506.19441): install broken on Python 3.12 +

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

@@ -147,10 +147,24 @@ fn main() -> Result<()> {
         let row: PairRow = serde_json::from_str(&line)
             .with_context(|| format!("parse row {idx}: {line}"))?;
 
-        // Speaker similarity at 16 kHz.
+        // Speaker similarity at 16 kHz. WavLM-SV's TDNN front-end has
+        // kernel size 5 with stride 2 chains, so it requires at least a
+        // few hundred samples of audio. Below ~0.25 s the model errors
+        // out — score those as -1.0 (sentinel) so the eval still emits
+        // a row instead of bailing on the whole batch.
         let ref_16k = audio_io::load_mono_at_rate(&row.ref_wav, 16_000)?;
         let gen_16k = audio_io::load_mono_at_rate(&row.gen_wav, 16_000)?;
-        let speaker_cosine = sv.score(&ref_16k, &gen_16k)? as f32;
+        const MIN_SV_SAMPLES: usize = 4000; // 0.25 s at 16 kHz
+        let speaker_cosine = if gen_16k.len() < MIN_SV_SAMPLES || ref_16k.len() < MIN_SV_SAMPLES {
+            tracing::warn!(
+                "row {idx}: gen={}s ref={}s — too short for WavLM-SV; cosine=-1",
+                gen_16k.len() as f32 / 16_000.0,
+                ref_16k.len() as f32 / 16_000.0,
+            );
+            -1.0f32
+        } else {
+            sv.score(&ref_16k, &gen_16k)? as f32
+        };
 
         // Transcribe gen_wav via Moonshine (16 kHz).
         let pcm_t = Tensor::from_vec(gen_16k.clone(), (1, 1, gen_16k.len()), &device)?;