magic-s2000 study: reg_delay/noadaptc flags, 4-arm queue v2, redx trajectory audit

- lt_ep_train: --reg_delay N (reg-free early phase: resreg/jr/floss/adaptc off
  for first N steps) + --noadaptc (kill hidden jacreg==0 damping feedback that
  would pollute single-reg ablation arms)
- queue v2: 4 arms delay-first (abl_delay = reg-free 2k -> proven pair)
- eig_traj/2/3: ARPACK audit of redx_traj — the run crossed the edge EARLY and
  oscillated (s1000 rotating-unstable, s1400 excursion mu=+2.1 self-recovered,
  s2000 the ONLY stable snapshot mu=-0.02, s2100/s2200 already back out) =>
  s2000 is a post-excursion STABILITY-DIP capture, dip width <100 steps;
  learning survives mild instability (val fell through unstable stretches).
  lead_rho cold-40 under-reads clusters — NOT a classifier; ARPACK for audits.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 39 insertions, 0 deletions

diff --git a/ep_run/eig_traj2.py b/ep_run/eig_traj2.py
new file mode 100644
index 0000000..bae22b7
--- /dev/null
+++ b/ep_run/eig_traj2.py
@@ -0,0 +1,39 @@
+"""E1b: ARPACK gold-standard re-measurement of the redx trajectory at 6 key ckpts (lead_rho's cold 2-D
+iteration under-reads near-unity clusters by ~0.02-0.03, so absolute mu from eig_traj.py is suspect).
+Top-3 |lam| of the forward map M = I + eps*J_F at the 400-step deep state, same seed-42 batch."""
+import numpy as np, torch, scipy.sparse.linalg as sla
+from torch.autograd.functional import jvp
+import lt_ep_train as L
+
+EPS, B, C = 0.1, 6, 1.0
+KEY = [1000, 1400, 2000, 2400, 2800, 3200]
+
+for s in KEY:
+    torch.manual_seed(0)
+    blk = L.EQBlock(512, 16, 256, 256, c=C, attn_mode='thick'); blk.qknorm = True
+    ck = torch.load(f'runs/redx_traj/s{s}.pt', map_location=L.dev)
+    with torch.no_grad():
+        for p, w in zip(blk.allp, ck['allp']):
+            p.copy_(w.to(L.dev))
+    torch.manual_seed(42)
+    idx, _ = L.get_batch('train', B, 256)
+    xin = blk.embed(idx).detach()
+    z = L.relax(blk, xin.clone(), xin, 400, EPS)
+    sh, n = z.shape, z.numel()
+    kk = 1.0 - EPS * (1.0 + C)
+
+    def mv(x, z=z, sh=sh):
+        v = torch.from_numpy(np.asarray(x, dtype=np.float32)).to(L.dev).view(sh)
+        with torch.no_grad():
+            Mv = kk * v + EPS * jvp(blk.nc_force, z, v)[1]
+        return Mv.reshape(-1).double().cpu().numpy()
+
+    A = sla.LinearOperator((n, n), matvec=mv, dtype=np.float64)
+    try:
+        vals = sorted(sla.eigs(A, k=3, which='LM', return_eigenvectors=False, maxiter=2000, tol=1e-4),
+                      key=lambda x: -abs(x))
+        out = "  ".join(f"|l|={abs(l):.5f}(mu={(l.real-1)/EPS:+.4f}{l.imag/EPS:+.3f}j)" for l in vals)
+    except Exception as e:
+        out = f"ARPACK-fail {type(e).__name__}"
+    print(f"s{s:<5} {out}", flush=True)
+print("EIG_TRAJ2_DONE", flush=True)