File size: 24,731 Bytes
504c02d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
"""
cell_g_class_probe_v3.py β€” three-way geometric probe

Tests the same geometric battery of metrics on three batteries:
  H2a:     Q-rank02 (D=4, V=32, 40K params, 1000 batches Adam)
  G-Cand:  Q-rank09 (D=3, V=32, 29K params, 1000 batches Adam)
  h2-64:   single-noise gaussian battery (D=8, V=64, 57K params, 10 epochs)

Key question: is the antipodal+rotational structure found in G-Cand a
property of D=3 specifically, or a property of LOW-band attractors at
ANY D? h2-64 has D=8 which sits in LOW band naturally (CV ~0.21).

Predicted outcomes:
  - h2-64 looks like H2 (uniform sphere, stable rows): G-class is D=3-specific
  - h2-64 looks like G (antipodal pairs, rotating frame): G-class is the
    universal LOW-band character; H2a is the OUTLIER for being so static
  - h2-64 looks like neither (some third pattern): D=8 has its own
    geometric character we haven't seen yet

Loading h2-64 from `loaded` if defined in session, else fetches from HF.
"""

import json
import math
import sys
from pathlib import Path

import numpy as np
import torch
import torch.nn.functional as F
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D  # noqa
from sklearn.cluster import KMeans
from sklearn.metrics import silhouette_score

CKPT_DIR = Path("/content/phaseQ_reports")
RANK09_CKPT = CKPT_DIR / "Q_rank09_h64_V32_D3_dp0_nx0_adam" / "epoch_1_checkpoint.pt"
RANK02_CKPT = CKPT_DIR / "Q_rank02_h64_V32_D4_dp0_nx0_adam" / "epoch_1_checkpoint.pt"
OUTPUT_PLOT = CKPT_DIR / "g_class_probe_v3.png"
OUTPUT_JSON = CKPT_DIR / "g_class_probe_v3.json"


# ════════════════════════════════════════════════════════════════════
# Loading
# ════════════════════════════════════════════════════════════════════

def load_qsweep_model(variant_str, ckpt_path):
    """Load Q-sweep model (rank02 or rank09)."""
    cfgs = get_phaseQ_configs()
    cfg_dict = next(c for c in cfgs if variant_str in c['variant'])
    cfg = build_run_config(cfg_dict)
    overrides = cfg_dict['overrides']

    model = PatchSVAE_F_Ablation(
        matrix_v=cfg.matrix_v, D=cfg.D, patch_size=cfg.patch_size,
        hidden=cfg.hidden, depth=cfg.depth,
        n_cross_layers=cfg.n_cross_layers, n_heads=cfg.n_heads,
        max_alpha=overrides.get('max_alpha', cfg.max_alpha),
        alpha_init=cfg.alpha_init,
        activation=overrides.get('activation', 'gelu'),
        row_norm=overrides.get('row_norm', 'sphere'),
        svd_mode=overrides.get('svd', 'fp64'),
        linear_readout=overrides.get('linear_readout', False),
        match_params=overrides.get('match_params', True),
        init_scheme=overrides.get('init', 'orthogonal'),
    )

    ckpt = torch.load(ckpt_path, map_location='cpu', weights_only=False)
    state_dict = (
        ckpt.get('model_state')
        or ckpt.get('model_state_dict')
        or ckpt.get('state_dict')
        or ckpt
    )
    model.load_state_dict(state_dict)
    model.eval()
    return model, cfg


def load_h2_64_battery(battery_idx=0, phase='final'):
    """Get one battery from the h2-64 array.

    Tries `loaded` from globals first (already in Colab session),
    falls back to AutoModel.from_pretrained.

    Returns (bank_module, V, D, patch_size, img_size).
    """
    array_model = globals().get('loaded')

    if array_model is None:
        print(f"  `loaded` not found, fetching from HF...")
        # Importing geolip_svae.arrays auto-registers BatteryArrayConfig
        # with HF Auto* β€” without this, model_type='battery_array' is unknown.
        import geolip_svae.arrays  # noqa: F401
        from transformers import AutoModel
        array_model = AutoModel.from_pretrained(
            "AbstractPhil/geolip-svae-h2-64")
        print(f"  Loaded h2-64 from HF")
    else:
        print(f"  Using `loaded` from global session")

    # Get the specific battery bank
    bank = array_model.bank(battery_idx, phase)
    bank.eval()

    # Get architecture from config
    cfg_dict = array_model.config.batteries[battery_idx]
    print(f"  Battery {battery_idx} ({phase}): "
          f"subgroup={cfg_dict.get('subgroup')}, "
          f"variant={cfg_dict.get('variant')}, "
          f"noise_types={cfg_dict.get('noise_types')}")

    # Architecture is uniform across h2-64 batteries
    V = 64
    D = 8
    patch_size = 2
    img_size = 64

    return bank, V, D, patch_size, img_size


# ════════════════════════════════════════════════════════════════════
# Collect M rows
# ════════════════════════════════════════════════════════════════════

def collect_per_sample_M(model, V, D, patch_size, img_size,
                          n_batches=8, batch_size=64,
                          is_h2_64_bank=False):
    """Collect [n_samples, V, D] M tensors from gaussian inputs."""
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model = model.to(device)

    ds = OmegaNoiseDataset(
        size=n_batches * batch_size,
        img_size=img_size,
        allowed_types=[0])  # gaussian
    loader = torch.utils.data.DataLoader(
        ds, batch_size=batch_size, shuffle=False)

    all_M = []
    with torch.no_grad():
        for imgs, _ in loader:
            imgs = imgs.to(device)
            out = model(imgs)
            # Both PatchSVAE and h2-64 banks return dict with 'svd' or
            # similar β€” the M tensor is at out['svd']['M'][:, 0]
            if 'svd' in out and 'M' in out['svd']:
                M_patch0 = out['svd']['M'][:, 0]  # [B, V, D]
            elif 'M' in out:
                M_patch0 = out['M'][:, 0]
            else:
                # Fall back: try to access via internal encode_patches
                from johanna_F_trainer import extract_patches
                patches = extract_patches(imgs, patch_size)
                enc = model.encode_patches(patches)
                M_patch0 = enc['M'][:, 0]
            all_M.append(M_patch0.cpu())

    return torch.cat(all_M, dim=0).numpy()  # [n_samples, V, D]


# ════════════════════════════════════════════════════════════════════
# Tests (carry over from v2)
# ════════════════════════════════════════════════════════════════════

def test_sphere_norm(all_M):
    row_norms = np.linalg.norm(all_M, axis=2)
    return {
        'min': float(row_norms.min()),
        'max': float(row_norms.max()),
        'mean': float(row_norms.mean()),
        'std': float(row_norms.std()),
        'sphere_normed': bool(
            abs(row_norms.mean() - 1.0) < 0.05 and row_norms.std() < 0.05),
    }


def test_row_stability(all_M):
    mean_dirs = all_M.mean(axis=0)
    mean_dir_norms = np.linalg.norm(mean_dirs, axis=1)
    return {
        'mean': float(mean_dir_norms.mean()),
        'min': float(mean_dir_norms.min()),
        'max': float(mean_dir_norms.max()),
        'std': float(mean_dir_norms.std()),
        'mean_dir_norms': mean_dir_norms.tolist(),
    }


def test_per_sample_clustering(all_M, k_test=5, n_samples=20):
    silhouettes = []
    for i in range(min(n_samples, all_M.shape[0])):
        M = all_M[i]
        try:
            km = KMeans(n_clusters=k_test, n_init=10, random_state=42)
            labels = km.fit_predict(M)
            if len(set(labels)) >= 2:
                sil = silhouette_score(M, labels)
                silhouettes.append(sil)
        except Exception:
            pass
    silhouettes = np.array(silhouettes)
    return {
        'k_tested': k_test,
        'mean': float(silhouettes.mean()) if len(silhouettes) else None,
        'std': float(silhouettes.std()) if len(silhouettes) else None,
        'silhouettes_per_sample': silhouettes.tolist(),
    }


def test_angular_distribution(all_M):
    all_rows = all_M.reshape(-1, all_M.shape[-1])
    norms = np.linalg.norm(all_rows, axis=1, keepdims=True)
    unit_rows = all_rows / np.clip(norms, 1e-12, None)

    n_subset = min(500, unit_rows.shape[0])
    idx = np.random.RandomState(42).choice(
        unit_rows.shape[0], n_subset, replace=False)
    subset = unit_rows[idx]

    cosines = subset @ subset.T
    triu_idx = np.triu_indices(n_subset, k=1)
    pairwise_cos = cosines[triu_idx]
    pairwise_angles = np.arccos(np.clip(pairwise_cos, -1, 1))

    return {
        'mean_angle': float(pairwise_angles.mean()),
        'median_angle': float(np.median(pairwise_angles)),
        'fraction_near_zero': float((pairwise_angles < 0.5).mean()),
        'fraction_near_pi': float((pairwise_angles > math.pi - 0.5).mean()),
        'fraction_near_perp': float(
            ((pairwise_angles > math.pi/2 - 0.3) &
             (pairwise_angles < math.pi/2 + 0.3)).mean()),
        'pairwise_angles_subset': pairwise_angles[:200].tolist(),
    }


def test_antipodal(all_M):
    mean_dirs = all_M.mean(axis=0)
    norms = np.linalg.norm(mean_dirs, axis=1, keepdims=True)
    unit_dirs = mean_dirs / np.clip(norms, 1e-12, None)

    cosines = unit_dirs @ unit_dirs.T
    np.fill_diagonal(cosines, 1.0)
    most_anti_cos = cosines.min(axis=1)
    n_pairs = (most_anti_cos < -0.9).sum() // 2

    return {
        'min_cos': float(most_anti_cos.min()),
        'mean_cos': float(most_anti_cos.mean()),
        'fraction_with_antipode': float((most_anti_cos < -0.9).mean()),
        'estimated_pairs': int(n_pairs),
        'max_possible_pairs': all_M.shape[1] // 2,
    }


def test_effective_rank(all_M):
    M_avg = all_M.mean(axis=0)
    sv = np.linalg.svd(M_avg, compute_uv=False)
    sv_norm = sv / sv.sum()
    erank = math.exp(-(sv_norm * np.log(sv_norm + 1e-12)).sum())
    return {
        'singular_values': sv.tolist(),
        'normalized_SV': sv_norm.tolist(),
        'effective_rank': float(erank),
        'D': int(all_M.shape[2]),
        'utilization': float(erank / all_M.shape[2]),
        'top1_share': float(sv_norm[0]),
    }


def run_all_tests(all_M, label):
    print(f"\n[{label}]")
    print(f"  Shape: {all_M.shape}")

    sphere = test_sphere_norm(all_M)
    print(f"  Sphere-norm: mean={sphere['mean']:.4f}, "
          f"std={sphere['std']:.4f} β†’ {'YES' if sphere['sphere_normed'] else 'NO'}")

    stability = test_row_stability(all_M)
    print(f"  Row stability: mean={stability['mean']:.3f}, "
          f"range=[{stability['min']:.3f}, {stability['max']:.3f}]")

    cluster = test_per_sample_clustering(all_M)
    if cluster['mean'] is not None:
        print(f"  Cluster (k=5): silhouette mean={cluster['mean']:.3f}, "
              f"std={cluster['std']:.3f}")

    angular = test_angular_distribution(all_M)
    print(f"  Angular: mean={angular['mean_angle']:.3f} "
          f"(uniform=Ο€/2={math.pi/2:.3f})")
    print(f"           near-perp: {angular['fraction_near_perp']:.3f}, "
          f"near-Ο€: {angular['fraction_near_pi']:.3f}")

    antipodal = test_antipodal(all_M)
    print(f"  Antipodal: {antipodal['estimated_pairs']}/"
          f"{antipodal['max_possible_pairs']} pairs, "
          f"frac with antipode={antipodal['fraction_with_antipode']:.3f}")

    erank = test_effective_rank(all_M)
    print(f"  Effective rank: {erank['effective_rank']:.2f} of {erank['D']} "
          f"({erank['utilization']*100:.0f}% utilization)")

    return {
        'sphere_norm': sphere,
        'stability': stability,
        'clustering': cluster,
        'angular': angular,
        'antipodal': antipodal,
        'rank': erank,
    }


# ════════════════════════════════════════════════════════════════════
# Composite character classification
# ════════════════════════════════════════════════════════════════════

def classify_battery_character(results):
    """Determine if battery is H2-like (sphere-solver) or G-like
    (rotating-antipodal) or something else."""
    stab = results['stability']['mean']
    antipodal_frac = results['antipodal']['fraction_with_antipode']
    cluster_sil = results['clustering']['mean']
    rank_util = results['rank']['utilization']

    # H2-like: high stability, low antipodal fraction, full rank
    is_h2_like = (
        stab > 0.85 and
        antipodal_frac < 0.55 and
        rank_util > 0.95
    )
    # G-like: low stability, high antipodal fraction
    is_g_like = (
        stab < 0.65 and
        antipodal_frac > 0.80
    )
    # Hybrid: somewhere in between

    if is_h2_like:
        return f"H2-LIKE (static sphere-solver)"
    elif is_g_like:
        return f"G-LIKE (rotating antipodal frame)"
    elif stab < 0.65 and antipodal_frac < 0.55:
        return f"DIFFUSE (low stability, no antipodal structure)"
    else:
        return (f"HYBRID (stab={stab:.2f}, antipodal_frac="
                f"{antipodal_frac:.2f})")


# ════════════════════════════════════════════════════════════════════
# Main
# ════════════════════════════════════════════════════════════════════

def main():
    print("=" * 70)
    print("Loading three batteries for comparative analysis")
    print("=" * 70)

    print("\n[1/3] H2a (Q-rank02, D=4, 1000-batch Adam)")
    h2_model, h2_cfg = load_qsweep_model('rank02', RANK02_CKPT)
    print(f"  V={h2_cfg.matrix_v}, D={h2_cfg.D}, "
          f"params={sum(p.numel() for p in h2_model.parameters()):,}")

    print("\n[2/3] G-Class candidate (Q-rank09, D=3, 1000-batch Adam)")
    g_model, g_cfg = load_qsweep_model('rank09', RANK09_CKPT)
    print(f"  V={g_cfg.matrix_v}, D={g_cfg.D}, "
          f"params={sum(p.numel() for p in g_model.parameters()):,}")

    print("\n[3/3] h2-64 single-noise gaussian battery (D=8, 10 epochs converged)")
    h264_bank, h264_V, h264_D, h264_ps, h264_img = load_h2_64_battery(
        battery_idx=0, phase='final')
    print(f"  V={h264_V}, D={h264_D}, patch_size={h264_ps}, img_size={h264_img}")

    # ════════════════════════════════════════════════════════════════
    # Collect M rows
    # ════════════════════════════════════════════════════════════════

    print("\n" + "=" * 70)
    print("Collecting M rows (gaussian inputs, 512 samples each)")
    print("=" * 70)

    print("\n  H2a...")
    all_M_h2 = collect_per_sample_M(
        h2_model, h2_cfg.matrix_v, h2_cfg.D,
        h2_cfg.patch_size, h2_cfg.img_size)

    print("  G-Cand...")
    all_M_g = collect_per_sample_M(
        g_model, g_cfg.matrix_v, g_cfg.D,
        g_cfg.patch_size, g_cfg.img_size)

    print("  h2-64 gaussian...")
    all_M_h264 = collect_per_sample_M(
        h264_bank, h264_V, h264_D, h264_ps, h264_img,
        is_h2_64_bank=True)

    # ════════════════════════════════════════════════════════════════
    # Run tests on each
    # ════════════════════════════════════════════════════════════════

    print("\n" + "=" * 70)
    print("GEOMETRIC ANALYSIS")
    print("=" * 70)

    results_h2 = run_all_tests(all_M_h2, "H2a (D=4, 1000-batch Adam)")
    results_g = run_all_tests(all_M_g, "G-Cand (D=3, 1000-batch Adam)")
    results_h264 = run_all_tests(
        all_M_h264, "h2-64 gaussian (D=8, 10 epochs)")

    # ════════════════════════════════════════════════════════════════
    # Side-by-side comparison
    # ════════════════════════════════════════════════════════════════

    print("\n" + "=" * 70)
    print("THREE-WAY COMPARISON")
    print("=" * 70)

    headers = f"{'Metric':<32} {'H2a (D=4)':>12} {'G-Cand (D=3)':>14} {'h2-64 (D=8)':>14}"
    print(f"\n  {headers}")
    print("  " + "-" * len(headers))

    rows = [
        ('Effective rank',
         results_h2['rank']['effective_rank'],
         results_g['rank']['effective_rank'],
         results_h264['rank']['effective_rank'],
         '.2f'),
        ('Dim utilization (%)',
         results_h2['rank']['utilization'] * 100,
         results_g['rank']['utilization'] * 100,
         results_h264['rank']['utilization'] * 100,
         '.0f'),
        ('Row stability',
         results_h2['stability']['mean'],
         results_g['stability']['mean'],
         results_h264['stability']['mean'],
         '.3f'),
        ('Per-sample silhouette (k=5)',
         results_h2['clustering']['mean'] or 0,
         results_g['clustering']['mean'] or 0,
         results_h264['clustering']['mean'] or 0,
         '.3f'),
        ('Mean pairwise angle (rad)',
         results_h2['angular']['mean_angle'],
         results_g['angular']['mean_angle'],
         results_h264['angular']['mean_angle'],
         '.3f'),
        ('Antipodal pair fraction',
         results_h2['antipodal']['fraction_with_antipode'],
         results_g['antipodal']['fraction_with_antipode'],
         results_h264['antipodal']['fraction_with_antipode'],
         '.3f'),
        ('Estimated antipodal pairs',
         results_h2['antipodal']['estimated_pairs'],
         results_g['antipodal']['estimated_pairs'],
         results_h264['antipodal']['estimated_pairs'],
         'd'),
    ]

    for row in rows:
        name, h2v, gv, h264v, fmt = row
        if fmt == 'd':
            print(f"  {name:<32} {h2v:>12d} {gv:>14d} {h264v:>14d}")
        else:
            print(f"  {name:<32} {h2v:>12{fmt}} {gv:>14{fmt}} {h264v:>14{fmt}}")

    print()
    char_h2 = classify_battery_character(results_h2)
    char_g = classify_battery_character(results_g)
    char_h264 = classify_battery_character(results_h264)

    print(f"  Character verdict:")
    print(f"    H2a:    {char_h2}")
    print(f"    G-Cand: {char_g}")
    print(f"    h2-64:  {char_h264}")

    # Headline conclusion
    print("\n" + "=" * 70)
    print("CONCLUSION")
    print("=" * 70)

    if "G-LIKE" in char_h264:
        print("  h2-64 (D=8, fully converged) shows G-CLASS character.")
        print("  β†’ The antipodal+rotational structure is NOT D=3-specific.")
        print("  β†’ It's the LOW-band attractor's natural geometry.")
        print("  β†’ H2a (D=4 at HIGH band) is the OUTLIER β€” its sphere-solver")
        print("    rigidity is HIGH-band-specific, not the universal pattern.")
    elif "H2-LIKE" in char_h264:
        print("  h2-64 (D=8, fully converged) shows H2 sphere-solver character.")
        print("  β†’ G-Class at D=3 is genuinely different from sphere-solvers.")
        print("  β†’ D=3 specifically can't form a stable static 32-row arrangement,")
        print("    so it falls into the rotating-antipodal regime.")
        print("  β†’ Higher D recovers static sphere-solver behavior even in LOW band.")
    elif "HYBRID" in char_h264 or "DIFFUSE" in char_h264:
        print("  h2-64 (D=8) shows mixed character β€” partial G-like features.")
        print("  β†’ Possible spectrum: HIGH-band β†’ static sphere (H2),")
        print("    LOW-band β†’ progressively more antipodal as D decreases.")
        print("  β†’ D=8 sits in transition; D=3 is fully G-class; D=4 HIGH is fully H2.")

    all_results = {
        'h2a': results_h2,
        'g_class_candidate': results_g,
        'h2_64_gaussian': results_h264,
        'characters': {
            'h2a': char_h2,
            'g_class': char_g,
            'h2_64': char_h264,
        },
    }

    with open(OUTPUT_JSON, 'w') as f:
        json.dump(all_results, f, indent=2, default=str)
    print(f"\n  Saved: {OUTPUT_JSON}")

    # Plot
    plot_three_way(all_M_h2, all_M_g, all_M_h264,
                    results_h2, results_g, results_h264, OUTPUT_PLOT)
    print(f"  Saved: {OUTPUT_PLOT}")

    return all_results


def plot_three_way(M_h2, M_g, M_h264, r_h2, r_g, r_h264, output_path):
    """6-panel comparison figure: 3 batteries Γ— 2 metrics each."""
    fig = plt.figure(figsize=(18, 14))

    # Row 1: Single-sample row scatters (project to first 3 dims)
    ax1 = fig.add_subplot(3, 3, 1, projection='3d')
    s = M_h2[0]
    ax1.scatter(s[:, 0], s[:, 1], s[:, 2], c=np.arange(len(s)),
                 cmap='viridis', s=80, edgecolors='black', linewidths=0.5)
    ax1.set_title(f'H2a (D=4) β€” single sample\nrows projected to first 3 dims')

    ax2 = fig.add_subplot(3, 3, 2, projection='3d')
    s = M_g[0]
    ax2.scatter(s[:, 0], s[:, 1], s[:, 2], c=np.arange(len(s)),
                 cmap='viridis', s=80, edgecolors='black', linewidths=0.5)
    ax2.set_title(f'G-Cand (D=3) β€” single sample\nfull native dims')

    ax3 = fig.add_subplot(3, 3, 3, projection='3d')
    s = M_h264[0]
    ax3.scatter(s[:, 0], s[:, 1], s[:, 2], c=np.arange(len(s)),
                 cmap='viridis', s=80, edgecolors='black', linewidths=0.5)
    ax3.set_title(f'h2-64 gaussian (D=8) β€” single sample\nrows projected to first 3 dims')

    # Row 2: Per-row stability sorted (descending)
    ax4 = fig.add_subplot(3, 3, 4)
    ax4.plot(sorted(r_h2['stability']['mean_dir_norms'], reverse=True),
              'o-', color='blue', markersize=4)
    ax4.set_title(f"H2a row stability\nmean={r_h2['stability']['mean']:.3f}")
    ax4.set_xlabel('Row index (sorted)')
    ax4.set_ylabel('Mean direction norm')
    ax4.set_ylim([0, 1.05])
    ax4.grid(alpha=0.3)

    ax5 = fig.add_subplot(3, 3, 5)
    ax5.plot(sorted(r_g['stability']['mean_dir_norms'], reverse=True),
              'o-', color='red', markersize=4)
    ax5.set_title(f"G-Cand row stability\nmean={r_g['stability']['mean']:.3f}")
    ax5.set_xlabel('Row index (sorted)')
    ax5.set_ylabel('Mean direction norm')
    ax5.set_ylim([0, 1.05])
    ax5.grid(alpha=0.3)

    ax6 = fig.add_subplot(3, 3, 6)
    ax6.plot(sorted(r_h264['stability']['mean_dir_norms'], reverse=True),
              'o-', color='green', markersize=4)
    ax6.set_title(f"h2-64 row stability\nmean={r_h264['stability']['mean']:.3f}")
    ax6.set_xlabel('Row index (sorted)')
    ax6.set_ylabel('Mean direction norm')
    ax6.set_ylim([0, 1.05])
    ax6.grid(alpha=0.3)

    # Row 3: Pairwise angle distributions
    ax7 = fig.add_subplot(3, 3, 7)
    ax7.hist(r_h2['angular']['pairwise_angles_subset'], bins=30,
              color='blue', alpha=0.7, density=True)
    ax7.axvline(math.pi/2, color='black', linestyle='--', alpha=0.5)
    ax7.set_title(f"H2a pairwise angles\nmean={r_h2['angular']['mean_angle']:.3f}")
    ax7.set_xlabel('Angle (radians)')

    ax8 = fig.add_subplot(3, 3, 8)
    ax8.hist(r_g['angular']['pairwise_angles_subset'], bins=30,
              color='red', alpha=0.7, density=True)
    ax8.axvline(math.pi/2, color='black', linestyle='--', alpha=0.5)
    ax8.set_title(f"G-Cand pairwise angles\nmean={r_g['angular']['mean_angle']:.3f}")
    ax8.set_xlabel('Angle (radians)')

    ax9 = fig.add_subplot(3, 3, 9)
    ax9.hist(r_h264['angular']['pairwise_angles_subset'], bins=30,
              color='green', alpha=0.7, density=True)
    ax9.axvline(math.pi/2, color='black', linestyle='--', alpha=0.5)
    ax9.set_title(f"h2-64 pairwise angles\nmean={r_h264['angular']['mean_angle']:.3f}")
    ax9.set_xlabel('Angle (radians)')

    plt.tight_layout()
    plt.savefig(output_path, dpi=120, bbox_inches='tight')
    plt.show()


if __name__ == '__main__':
    results = main()