rtx-interpret: fix decoder transpose in SAE compute_and_apply_gradients

The encoder-gradient path through the decoder was transposing the decoder
before the matmul, producing `[batch, d_model] × [d_sae, d_model]` — a
shape mismatch for every batch > 1. The decoder is stored as
`[d_model, d_sae]`, so `recon_grad @ decoder` is already the right shape
(and matches the comment at the call site, which reads
"recon_grad @ decoder @ d_relu").

All 9 existing `sae::tests` still pass. Omni-Cortex's `LatentDictionary`
now trains correctly on batches larger than 1.

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

crates/core/rtx-interpret/src/sae/training.rs

@@ -355,9 +355,13 @@ impl SAETrainer {
         // Decoder bias gradient: sum of recon_grad along batch dimension
         let decoder_bias_grad = recon_grad.sum(Some(0))?;
 
-        // Encoder gradient (through decoder): recon_grad @ decoder @ d_relu
+        // Encoder gradient (through decoder): recon_grad @ decoder @ d_relu.
-        let decoder_t = self.sae.decoder().transpose(0, 1)?;
+        // `decoder` has shape [d_model, d_sae] and `recon_grad` has shape
-        let pre_encoder_grad = recon_grad.matmul(&decoder_t)?;
+        // [batch, d_model], so `recon_grad @ decoder = [batch, d_sae]` —
+        // no transpose needed. (The previous transpose here was an
+        // upstream bug: it produced a shape-mismatched matmul for every
+        // batch size > 1.)
+        let pre_encoder_grad = recon_grad.matmul(self.sae.decoder())?;
 
         // ReLU gradient: mask where features > 0
         // Create a mask by computing (features > 0) as 1.0 or 0.0