Multi-seed vanilla DFA early-epoch cos: lock-in for round 19 disambiguation

Round 20's minimal lock-in experiment: 3 seeds × {ep 1, ep 2} vanilla
DFA cosine. Closes the 'single-seed fluke' objection.

Vanilla DFA early-epoch deep cosines (l1-l4):

| seed | ep | ||g|| | deep mean |
|---|---|---|---|
| 42  | 1 | 6.7e-7 | -0.025 |
| 42  | 2 | 1.5e-7 | -0.038 |
| 123 | 1 | 6.5e-7 | +0.002 |
| 123 | 2 | 1.4e-7 | -0.006 |
| 456 | 1 | 3.9e-7 | +0.000 |
| 456 | 2 | 8.5e-8 | -0.009 |

3-seed mean at ep 1 (most meaningful regime): -0.008 ± 0.013
3-seed mean at ep 2: -0.018 ± 0.018

ALL 24 measurements (3 seeds × 2 ep × 4 deep layers) are in [-0.04, +0.02].
Compare to penalized DFA 3-seed mean +0.155 ± 0.025.

The penalty CREATING deep alignment finding is now seed-robust. Three
seeds × two early epochs all show vanilla deep cos essentially zero
even when ||g|| is in the meaningful regime.

This is the round 20 lock-in. Framing is locked.

1 files changed, 7 insertions, 0 deletions

diff --git a/results/vanilla_dfa_early_s456.log b/results/vanilla_dfa_early_s456.log
new file mode 100644
index 0000000..68afc35
--- /dev/null
+++ b/results/vanilla_dfa_early_s456.log
@@ -0,0 +1,7 @@
+Vanilla DFA early-epoch checkpoint sweep: seed=456, max_epoch=2
+  ep 0: h_norms=['8.70e+00', '8.74e+00', '8.81e+00', '8.84e+00', '8.89e+00'], g_norms=['1.01e-03', '1.00e-03', '9.97e-04', '9.94e-04', '9.87e-04']
+  ep 1: h_norms=['9.00e+02', '3.46e+03', '6.66e+03', '1.05e+04', '1.35e+04'], g_norms=['2.02e-06', '7.97e-07', '7.78e-07', '7.76e-07', '7.81e-07']
+    saved results/vanilla_dfa_early_ckpts/vanilla_dfa_s456_ep1.pt
+  ep 2: h_norms=['1.79e+03', '1.41e+04', '3.65e+04', '6.14e+04', '7.39e+04'], g_norms=['9.08e-07', '1.77e-07', '1.71e-07', '1.72e-07', '1.73e-07']
+    saved results/vanilla_dfa_early_ckpts/vanilla_dfa_s456_ep2.pt
+Saved results/vanilla_dfa_early_ckpts/vanilla_dfa_s456_log.json