cascade: retire zil per user directive (EP route = equilibrium two-phase); probe gains --init_sweep (state warm-start, EP-clean readout) and --sopt adam; B2 solver sweep

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

2 files changed, 19 insertions, 4 deletions

diff --git a/docs/campaign/CASCADE_ABLATION_PLAN.md b/docs/campaign/CASCADE_ABLATION_PLAN.md
index c480c11..03e3b52 100644
--- a/docs/campaign/CASCADE_ABLATION_PLAN.md
+++ b/docs/campaign/CASCADE_ABLATION_PLAN.md
@@ -43,8 +43,14 @@ The looped-EP precedent multiplier was ~230× BP; the K-frontier decides whether
   reads (no global backward graph); the EQUILIBRIUM mode (jacobi/CG to convergence) remains the
   physically-meaningful EP column — priced expensive by the sweep, CG/preconditioning is the B2 job,
   and it is the analog-hardware rung (E-tier).
-- **C1 money run LAUNCHED:** casc_ep6 (zil, seed 0) vs casc_bp6 (BP twin, best 2.9746@4k) —
-  step-0 CE bit-identical (true twin). 107 GPU3.
+- **C1 (zil) ran and is RETIRED with zil itself:** casc_ep6 best 3.3236 vs BP twin 2.9746 (gap 0.35
+  — single-sided zil top-read carries an O(β) shift on the readout term; moot now).
+  **USER DIRECTIVE (2026-07-09 night): zil is NOT the route — it is BP in disguise; the project
+  stays on TRUE EP = equilibrium-mode two-phase relaxation.** zil survives only as (a) a diagnostic
+  upper bound, (b) optionally a numerical STATE-INIT trick for GPU simulation (`--init_sweep`:
+  readout still taken at the relaxed equilibrium = clean EP semantics; hardware needs no init trick
+  — physics settles). **Critical path = B2: make the equilibrium solver cheap** (Adam-on-states /
+  init-sweep warm start / GS-multi-sweep / λ_l preconditioning), then rerun C1 in equilibrium mode.
 
 ## Tier 0 — gate hardening (probe-scale, hours, no training)  → K1
 
diff --git a/ep_run/cascade_probe.py b/ep_run/cascade_probe.py
index da3d380..dda4059 100644
--- a/ep_run/cascade_probe.py
+++ b/ep_run/cascade_probe.py
@@ -24,6 +24,8 @@ ap.add_argument('--include_io', action='store_true')   # gate emb/pos/tied-reado
 ap.add_argument('--seed', type=int, default=0); ap.add_argument('--warp', type=float, default=1.0)
 ap.add_argument('--ckpt', type=str, default='')         # gate at a casc_bp_train checkpoint
 ap.add_argument('--sumscale', action='store_true')      # relax in SUM units (gamma=1 Z-IL semantics)
+ap.add_argument('--init_sweep', action='store_true')    # one reverse gamma=1 sweep as state INIT (readout still at equilibrium = clean EP)
+ap.add_argument('--sopt', choices=['sgd', 'adam'], default='sgd')   # state-relaxation optimizer
 args = ap.parse_args()
 torch.manual_seed(args.seed)
 dev = 'cuda' if torch.cuda.is_available() else 'cpu'
@@ -89,9 +91,16 @@ def relax(z0, y, beta):
     """minimize F_beta over states; free init = forward pass. Returns relaxed states."""
     with torch.no_grad():
         zs = [z.detach().clone() for z in fwd_states(z0)[1:]]
+    if args.init_sweep:                                  # numerical warm-start of the STATE only
+        sv = args.sumscale; args.sumscale = True
+        for l in range(args.L - 1, -1, -1):
+            g = local_grad(zs, z0, l, beta, y)
+            zs[l] = (zs[l] - g).detach()
+        args.sumscale = sv
     if args.scheme == 'jacobi':
         for z in zs: z.requires_grad_(True)
-        opt = torch.optim.SGD(zs, lr=args.eta, momentum=args.mom)
+        opt = (torch.optim.Adam(zs, lr=args.eta) if args.sopt == 'adam'
+               else torch.optim.SGD(zs, lr=args.eta, momentum=args.mom))
         for _ in range(args.K):
             opt.zero_grad()
             E = 0.0; prev = z0
@@ -180,6 +189,6 @@ print('per-block cos: ' + '  '.join(f'{l}:{np.mean(v):.4f}' for l, v in enumerat
 if args.include_io:
     idx = [i for i, n in enumerate(names) if n.startswith('io.')]
     print(f'io cos (last batch): {F.cosine_similarity(flat([gep[i] for i in idx]), flat([gbp[i] for i in idx]), dim=0).item():.4f}')
-print(f'SUMMARY scheme={args.scheme} L={args.L} C={args.C} K={args.K} eta={args.eta} beta={args.beta} '
+print(f'SUMMARY scheme={args.scheme}{"+init" if args.init_sweep else ""}+{args.sopt} L={args.L} C={args.C} K={args.K} eta={args.eta} beta={args.beta} '
       f'io={int(args.include_io)} warp={args.warp} ckpt={args.ckpt or "-"} '
       f'cos={cm:.4f} cosmin={cmin:.4f} shrink={sm:.3f} t={time.time()-t0:.1f}s')