SPECTRAL GOVERNOR (--governor): the from-scratch recipe candidate, dip-map-calibrated

Dip screening (4/4 seeds have dips; first-dip cluster 1200-1700, width
100-300 steps, depth |lam|~0.998 = s2000-class; s2 shows excursion->dip in
sequence): reg-free until a scout (warm lead_rho, every gov_k=100, deep-400
subbatch) flags rho<0.985 after gov_min=1000, then ARPACK-certify all top-3
|lam|<1 -> leash ON (pair regs). Post-engagement: sustained rho_scout>1.02
x2 -> rescue boost (resreg x3 for 500 steps; the proven hr2 maneuver).
abl_delay's fixed-2000 death explained: it engaged AFTER the dip cluster,
mid-excursion. Two governor seeds launched on 107 (trained-state engagement
is Pascal-safe per hr2 precedent).

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

1 files changed, 38 insertions, 2 deletions

diff --git a/ep_run/lt_ep_train.py b/ep_run/lt_ep_train.py
index 0fd9180..2e555f9 100644
--- a/ep_run/lt_ep_train.py
+++ b/ep_run/lt_ep_train.py
@@ -491,6 +491,11 @@ def main():
     ap.add_argument('--holoavg', action='store_true')          # trend-aware stop + plateau-avg track (gate: 0.913->0.936 @t2sel160)
     ap.add_argument('--fastfp', action='store_true')           # Anderson-accelerated free phase (pure-speed opt-in; alters resreg semantics)
     ap.add_argument('--bf16polish', type=int, default=0)       # bf16 bulk relax + fp32 last-K polish (0=off; gate before use)
+    ap.add_argument('--governor', action='store_true')         # spectral governor: reg-free until a certified stability DIP, then leash; rescue-boost on breakout
+    ap.add_argument('--gov_min', type=int, default=1000)       # no engagement before this step (dip cluster: 1200-1700 per dipfarm)
+    ap.add_argument('--gov_k', type=int, default=100)          # scout cadence (steps)
+    ap.add_argument('--gov_scout', type=float, default=0.985)  # scout rho threshold to trigger ARPACK confirmation
+    ap.add_argument('--gov_boost', type=float, default=3.0)    # resreg multiplier during rescue mode
     ap.add_argument('--rt_final', type=float, default=0.0)     # anneal res_target to this (0=off), 25%-75% of run
     ap.add_argument('--nudge_brake', type=float, default=0.0)  # kappa: anchor spring during nudge (Tikhonov adjoint)
     ap.add_argument('--init_ckpt', type=str, default='')       # warm-start weights from a saved ckpt
@@ -634,15 +639,46 @@ def main():
         print(f"[fingerprint] ckpt={cfg.init_ckpt or 'scratch'} | res={fp['res']:.2e} cos(EP,BPTT)={fp['cos']:.4f} "
               f"rho={fp['rho']:.5f} Re_mu={fp['mu_re']:+.4f} val={fp['val']:.4f}", flush=True)
         return
+    gov = {'engaged': not cfg.governor, 'boost_until': 0, 'hi': 0, 'cache': {}}
     for step in range(start_step, cfg.steps + 1):
+        if cfg.governor and step % cfg.gov_k == 0 and step > 0:
+            from eig_control import lead_rho
+            idxg, _ = get_batch('train', 4, cfg.T)
+            xing = blk.embed(idxg).detach()
+            zg = relax(blk, xing.clone(), xing, 400, cfg.eps)
+            _, rho_s, _mu_s = lead_rho(blk, zg, cfg.eps, blk.c, gov['cache'], iters=8)
+            if not gov['engaged'] and step >= cfg.gov_min and rho_s < cfg.gov_scout:
+                import numpy as _np, scipy.sparse.linalg as _sla   # ARPACK dip confirmation
+                _sh, _n = zg.shape, zg.numel()
+                _kk = 1.0 - cfg.eps * (1.0 + blk.c)
+                def _mv(x, z=zg, sh=_sh):
+                    v = torch.from_numpy(_np.asarray(x, dtype=_np.float32)).to(dev).view(sh)
+                    with torch.no_grad():
+                        Mv = _kk * v + cfg.eps * torch.autograd.functional.jvp(blk.nc_force, z, v)[1]
+                    return Mv.reshape(-1).double().cpu().numpy()
+                try:
+                    _vals = _sla.eigs(_sla.LinearOperator((_n, _n), matvec=_mv, dtype=_np.float64),
+                                      k=3, which='LM', return_eigenvectors=False, maxiter=1500, tol=2e-4)
+                    if all(abs(l) < 1.0 for l in _vals):
+                        gov['engaged'] = True
+                        print(f"[governor] DIP CERTIFIED @step {step}: " +
+                              " ".join(f"{abs(l):.5f}" for l in _vals) + " -> leash ON", flush=True)
+                except Exception as _e:
+                    print(f"[governor] confirm failed {type(_e).__name__}", flush=True)
+            elif gov['engaged']:
+                gov['hi'] = gov['hi'] + 1 if rho_s > 1.02 else 0
+                if gov['hi'] >= 2 and step >= gov['boost_until']:
+                    gov['boost_until'] = step + 500
+                    print(f"[governor] breakout (rho_scout {rho_s:.4f}) -> rescue boost x{cfg.gov_boost} until {gov['boost_until']}", flush=True)
         idx, y = get_batch('train', cfg.B, cfg.T)
         if cfg.mode == 'ep':
             sw = hw_swap() if hw_on else None
-            dly = step < cfg.reg_delay            # reg-free early phase (magic-ckpt hypothesis: reach the edge unleashed)
+            dly = (step < cfg.reg_delay) or (cfg.governor and not gov['engaged'])
+            _rr = cfg.resreg * (cfg.gov_boost if step < gov['boost_until'] else 1.0)
             grads, res = ep_step(blk, idx, y, cfg.T1, cfg.T2, cfg.eps, cfg.beta, 0.0 if dly else jr,
                                  cfg.holo, cfg.hr,
                                  cfg.t1max, cfg.res_est, cfg.t2sel, cfg.corr_every, cfg.res_gate,
-                                 0.0 if dly else cfg.resreg,
+                                 0.0 if dly else _rr,
                                  cfg.eigreg, cfg.eig_margin, 0.0 if dly else cfg.floss,
                                  cfg.floss_q, cfg.floss_rho, cfg.floss_bsub)
             if sw is not None: