holofast: exact halved-jvp AEP track — 1.55x nudged phase, gradient-gate passed

holo_a_track computed the doubled-batch jvp/vjp on [v0; -v0] at a shared
anchor zbar — exact antisymmetric redundancy (phase deviations from the
common mode are exact negatives). holo_a_track_fast computes at batch B and
mirrors: single-eval parity 6e-7 (exact); trajectory-level 45% divergence
SHARED with the original's own FD noise floor (1e-6 state noise -> 49%
self-divergence — the 2r=0.04 finite difference amplifies fp noise; training
averages it via pema/momentum). Ship gate: cos(EP,BPTT) orig vs fast
indistinguishable (0.907/0.912, 0.853/0.853, 0.918/0.918). Timing 6.43->4.16s
on the T2=40 nudged phase (contended GPU, relative). --holofast flag,
default off; queued ablation arms deliberately stay on orig for fidelity.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 25 insertions, 0 deletions

diff --git a/ep_run/holo_fast_gate.py b/ep_run/holo_fast_gate.py
new file mode 100644
index 0000000..2a5f39b
--- /dev/null
+++ b/ep_run/holo_fast_gate.py
@@ -0,0 +1,25 @@
+"""Ship-gate for --holofast: gradient-level equivalence. cos(EP,BPTT) with orig vs fast track on the
+same batches (s2000, track path, holo=2 t2sel=40). Ship iff the two cos columns are statistically
+indistinguishable (the a-level 45% parity gap is FD noise; what matters is the gradient direction)."""
+import torch
+import lt_ep_train as L
+from diag_cos import cos_ep_bptt
+
+torch.manual_seed(0)
+blk = L.EQBlock(512, 16, 256, 256, c=1.0, attn_mode='thick'); blk.qknorm = True
+ck = torch.load('runs/redx_traj/s2000.pt', map_location=L.dev)
+with torch.no_grad():
+    for p, w in zip(blk.allp, ck['allp']):
+        p.copy_(w.to(L.dev))
+blk.track = True
+
+print(f"{'batch':>5} {'cos_orig':>9} {'cos_fast':>9} {'res':>9}", flush=True)
+torch.manual_seed(11)
+batches = [L.get_batch('train', 24, 256) for _ in range(3)]
+for b, (idx, y) in enumerate(batches):
+    blk.holofast = False
+    c0, r0 = cos_ep_bptt(blk, idx, y, 150, 20, 0.1, 0.02, holo=2, hr=0.02, t2sel=40)
+    blk.holofast = True
+    c1, r1 = cos_ep_bptt(blk, idx, y, 150, 20, 0.1, 0.02, holo=2, hr=0.02, t2sel=40)
+    print(f"{b:>5} {c0:>9.4f} {c1:>9.4f} {r0:>9.2e}", flush=True)
+print("HOLO_FAST_GATE_DONE", flush=True)