amygdala: bump subspace-k default to 512

k=20 was far too aggressive a truncation — it discards per-attention-head
discriminability entirely. At hidden_dim=5120, 40 heads × head_dim=128 each
contribute their own 128-dim block to the residual stream via W_o columns.
To resolve 'this concept lives in head H', per-story SVD needs enough rank
to separate head contributions, which means k on the order of hundreds.

512 is a reasonable default: clamped to n_tokens per story so short stories
use their full natural rank. The eigenvalue spectrum of M_pos - M_base
should become sharper (larger λ_0/λ_1 gap) as we stop averaging across
nuisance-shared directions.

Co-Authored-By: Proof of Concept <poc@bcachefs.org>

training/amygdala_training/train_steering_vectors.py

@@ -850,8 +850,12 @@ def main() -> None:
     ap.add_argument(
         "--subspace-k",
         type=int,
-        default=20,
-        help="Top-k right singular vectors per story for subspace method",
+        default=512,
+        help="Max top-k right singular vectors per story for subspace method "
+             "(clamped to n_tokens per story). Default 512 is enough to span "
+             "each story's full natural subspace including per-attention-head "
+             "contributions on a hidden_dim=5120 residual stream. Smaller "
+             "values (e.g. 20) discard per-head discriminability.",
     )
     ap.add_argument(
         "--quality-report",