fix smoke.

src/projected_grpo/extract_vhack.py

@@ -43,7 +43,7 @@ def collect_last_token_hidden(
         ids = tokenizer(p, return_tensors="pt").to(device)
         out = model(**ids, output_hidden_states=True)
         # out.hidden_states is tuple of (n_layers+1,) tensors of shape (1, seq, d)
-        h = out.hidden_states[layer_idx][0, -1, :].cpu()  # "d"
+        h = out.hidden_states[layer_idx][0, -1, :].float().cpu()  # "d" — fp32 for stable v_hack
         hs.append(h)
     return torch.stack(hs, dim=0)
 
@@ -67,8 +67,10 @@ def extract_vhack(
 
     logger.info(
         f"v_hack extracted layer={layer_idx} n_train={len(h_hack_train)} "
-        f"n_val={len(h_hack_val)} val_acc={correct:.3f} "
-        f"SHOULD val_acc>0.9 ELSE pair quality or layer is wrong"
+        f"n_val={len(h_hack_val)} val_acc={correct:.3f}  "
+        f"SHOULD>0.9 on a trained model: v_hack should separate hack from clean. "
+        f"On tiny-random/untrained models val_acc~0.5 (no semantic structure yet), "
+        f"which is fine for smoke -- the projection mechanism is what we test there."
     )
 
     return VHackResult(