stage1_analysis_trainer.py

"""
Geometric Transformer — CIFAR-100 Training with CM-Validated Analysis

Changes from previous version:
  - CM gate diagnostics per layer: active anchors, gate_mean, cm_positive_frac
  - CM quality in geometric residual analysis (replaces blind gate)
  - Geometric regularization losses (CV target + anchor spread) in training loop
  - Anchor diagnostics via model.anchor_diagnostics()
  - CM quality trajectory alongside CV and bridge KL for cooperation analysis

TensorBoard logging of every geometric feature element:
  - CV (coefficient of variation) per layer — the pentachoron band metric
  - CM gate: active anchors, gate mean, cm_positive_frac, quality per position
  - Stream agreement/divergence per layer
  - Anchor utilization, entropy, spread
  - Patchwork activation statistics (from CM-validated triangulation)
  - Bridge vs assignment consistency
  - Triangulation distance distributions
  - SVD spectrum, entropy, novelty
  - Quaternion arm norms and composition statistics
  - Cayley rotation ‖R-I‖ per layer
  - FiLM gamma/beta deviation from identity
  - Gate activation statistics
  - Gradient norms per component type (including cm_gate)
  - Weight norms per component type
  - Geometric regularization: CV loss, spread loss per epoch

!pip install geolip-core torchvision tqdm tensorboard
"""

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import time, json, math
from pathlib import Path
from tqdm.auto import tqdm
from torch.utils.tensorboard import SummaryWriter

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Device: {device}")
if device.type == 'cuda':
    print(f"  GPU: {torch.cuda.get_device_name()}")
    print(f"  VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB")

# ═══════════════════════════════════════════════════════════════════════════════
# IMPORT TRANSFORMER
# ═══════════════════════════════════════════════════════════════════════════════

# Try geolip_core installed package first, fall back to local file
try:
    from geolip_core.pipeline.components.geometric_transformer import (
        GeometricTransformer, GeometricTransformerLayer,
        CayleyOrthogonal, QuaternionCompose, FiLMLayer,
        ContentAttention, GeometricAttention, CMValidatedGate,
        TorchComponent, BaseTower,
        anchor_neighborhood_cm,
    )
    print("  Imported from geolip_core (installed)")
except ImportError:
    try:
        from geometric_transformer import (
            GeometricTransformer, GeometricTransformerLayer,
            CayleyOrthogonal, QuaternionCompose, FiLMLayer,
            ContentAttention, GeometricAttention, CMValidatedGate,
            TorchComponent, BaseTower,
            anchor_neighborhood_cm,
        )
        print("  Imported from local geometric_transformer.py")
    except ImportError:
        raise ImportError(
            "Cannot find geometric_transformer. Place geometric_transformer.py "
            "in the working directory or install geolip-core.")

torch.set_float32_matmul_precision('high')


# ═══════════════════════════════════════════════════════════════════════════════
# CONFIG
# ═══════════════════════════════════════════════════════════════════════════════

CONFIG = {
    # Model
    'd_model': 256,
    'n_heads': 8,
    'n_layers': 8,
    'n_anchors': 128,
    'manifold_dim': 128,
    'n_comp': 4,
    'd_comp': 16,
    'context_dim': 64,
    'quat_dim': 32,
    'dropout': 0.1,
    'cm_neighbors': 3,       # CM simplex neighbors

    # Input stage
    'patch_size': 4,
    'img_size': 32,
    'in_channels': 3,
    'conv_channels': 64,
    'svd_rank': 16,

    # Training
    'epochs': 100,
    'batch_size': 1024,
    'lr': 1e-3,
    'weight_decay': 0.05,
    'warmup_epochs': 5,
    'label_smoothing': 0.1,
    'num_workers': 8,

    # Geometric regularization
    'cv_target': 0.215,      # pentachoron band center
    'cv_weight': 0.1,        # CV loss weight
    'spread_weight': 0.01,   # anchor spread loss weight

    # Augmentation — tuned for CM gate training
    'cutmix_alpha': 1.0,     # CutMix beta distribution α (1.0 = uniform box sizes)
    'cutmix_prob': 0.5,      # probability of applying CutMix per batch
    'random_erasing_p': 0.25,  # probability of erasing per image

    # InfoNCE memory bank on geometric residual
    'nce_bank_size': 4096,   # queue size (0 to disable)
    'nce_temperature': 0.1,  # InfoNCE temperature
    'nce_weight': 0.1,       # loss weight

    # Data
    'num_classes': 100,

    # Logging
    'log_geo_every': 5,      # full geometric analysis every N epochs
    'log_grads_every': 10,   # gradient norms every N epochs
    'log_dir': 'runs/geo_cifar100',
}


# ═══════════════════════════════════════════════════════════════════════════════
# INPUT STAGE
# ═══════════════════════════════════════════════════════════════════════════════

try:
    from geolip_core.core.input.svd import SVDObserver
    _HAS_SVD = True
except ImportError:
    _HAS_SVD = False

    class SVDObserver(nn.Module):
        """Fallback SVDObserver."""
        def __init__(self, in_channels, svd_rank=24):
            super().__init__()
            self.svd_rank = svd_rank
            self.to_svd = nn.Conv2d(in_channels, svd_rank, 1, bias=False)
            self.register_buffer('ema_s', torch.ones(svd_rank))
            self.register_buffer('ema_vh_flat', torch.eye(svd_rank).reshape(-1))
            self.ema_momentum = 0.99

        def extract_features(self, S, Vh):
            B, k = S.shape
            S_safe = S.clamp(min=1e-6)
            s_norm = S_safe / (S_safe.sum(dim=-1, keepdim=True) + 1e-8)
            vh_diag = Vh.diagonal(dim1=-2, dim2=-1)
            vh_offdiag = (Vh.pow(2).sum((-2, -1)) - vh_diag.pow(2).sum(-1)).unsqueeze(-1).clamp(min=0)
            s_ent = -(s_norm * torch.log(s_norm.clamp(min=1e-8))).sum(-1, keepdim=True)
            out = torch.cat([s_norm, vh_diag, vh_offdiag, s_ent], dim=-1)
            return torch.where(torch.isfinite(out), out, torch.zeros_like(out))

        def compute_novelty(self, S):
            return S - self.ema_s.clone().unsqueeze(0)

        def forward(self, x):
            B, C, H, W = x.shape
            h = self.to_svd(x)
            h_flat = h.permute(0, 2, 3, 1).reshape(B, H * W, self.svd_rank)
            with torch.amp.autocast('cuda', enabled=False):
                with torch.no_grad():
                    gram = torch.bmm(h_flat.float().transpose(1, 2), h_flat.float())
                    evals, evecs = torch.linalg.eigh(gram)
                    evals = evals.flip(-1).clamp(min=1e-12)
                    S = evals.sqrt()
                    Vh = evecs.flip(-1).transpose(-2, -1)
                    S = torch.where(torch.isfinite(S), S, torch.ones_like(S))
                    Vh = torch.where(torch.isfinite(Vh), Vh, torch.zeros_like(Vh))
            features = self.extract_features(S, Vh)
            novelty = self.compute_novelty(S)
            return S, Vh, features, novelty

        @torch.no_grad()
        def update_ema(self, S, Vh):
            m = self.ema_momentum
            self.ema_s.mul_(m).add_(S.detach().mean(0), alpha=1-m)
            self.ema_vh_flat.mul_(m).add_(Vh.detach().mean(0).reshape(-1), alpha=1-m)

        @property
        def feature_dim(self):
            return 2 * self.svd_rank + 2


class ConvSVDPatchEmbedding(TorchComponent):
    """Input stage: conv frontend → SVDObserver → patch tokens."""
    def __init__(self, name, img_size=32, patch_size=4, in_channels=3,
                 conv_channels=64, d_model=256, svd_rank=16):
        super().__init__(name)
        self.patch_size = patch_size
        self.n_patches = (img_size // patch_size) ** 2
        self.d_model = d_model
        self.svd_rank = svd_rank

        self.conv_frontend = nn.Sequential(
            nn.Conv2d(in_channels, conv_channels, 3, padding=1, bias=False),
            nn.BatchNorm2d(conv_channels), nn.GELU(),
            nn.Conv2d(conv_channels, conv_channels, 3, padding=1, bias=False),
            nn.BatchNorm2d(conv_channels), nn.GELU(),
        )
        self.svd_observer = SVDObserver(conv_channels, svd_rank)
        self.patch_proj = nn.Conv2d(
            conv_channels, d_model, kernel_size=patch_size,
            stride=patch_size, bias=False)
        self.patch_norm = nn.LayerNorm(d_model)

        svd_feat_dim = self.svd_observer.feature_dim
        self.svd_to_gamma = nn.Linear(svd_feat_dim, d_model)
        self.svd_to_beta = nn.Linear(svd_feat_dim, d_model)
        nn.init.zeros_(self.svd_to_gamma.weight); nn.init.ones_(self.svd_to_gamma.bias)
        nn.init.zeros_(self.svd_to_beta.weight); nn.init.zeros_(self.svd_to_beta.bias)

        self.cls_token = nn.Parameter(torch.randn(1, 1, d_model) * 0.02)
        self.pos_embed = nn.Parameter(
            torch.randn(1, self.n_patches + 1, d_model) * 0.02)

    def forward(self, x):
        B = x.shape[0]
        feat = self.conv_frontend(x)
        S, Vh, svd_features, novelty = self.svd_observer(feat)
        tokens = self.patch_proj(feat)
        tokens = tokens.flatten(2).transpose(1, 2)
        tokens = self.patch_norm(tokens)
        gamma = self.svd_to_gamma(svd_features).unsqueeze(1)
        beta = self.svd_to_beta(svd_features).unsqueeze(1)
        tokens = gamma * tokens + beta
        cls = self.cls_token.expand(B, -1, -1)
        tokens = torch.cat([cls, tokens], dim=1)
        tokens = tokens + self.pos_embed
        svd_state = {
            'singular_values': S, 'Vh': Vh,
            'svd_features': svd_features, 'novelty': novelty,
        }
        if self.training:
            self.svd_observer.update_ema(S, Vh)
        return tokens, svd_state


# ═══════════════════════════════════════════════════════════════════════════════
# CLASSIFIER ( uses GeometricTransformer with CM gates)
# ═══════════════════════════════════════════════════════════════════════════════

class GeoViTClassifier(BaseTower):
    """Geometric Vision Transformer for classification.

    Wraps ConvSVDPatchEmbedding + GeometricTransformer + task head.
    Exposes geometric_losses() for regularization during training.
    """
    def __init__(self, name, config):
        super().__init__(name)
        self.config = config

        self.attach('patch_embed', ConvSVDPatchEmbedding(
            'patch_embed', img_size=config['img_size'],
            patch_size=config['patch_size'], in_channels=config['in_channels'],
            conv_channels=config['conv_channels'], d_model=config['d_model'],
            svd_rank=config['svd_rank'],
        ))
        self.attach('transformer', GeometricTransformer(
            'geo_cifar', d_model=config['d_model'], n_heads=config['n_heads'],
            n_layers=config['n_layers'], n_anchors=config['n_anchors'],
            manifold_dim=config['manifold_dim'], n_comp=config['n_comp'],
            d_comp=config['d_comp'], context_dim=config['context_dim'],
            quat_dim=config['quat_dim'], dropout=config['dropout'],
            cm_neighbors=config.get('cm_neighbors', 3),
            nce_bank_size=config.get('nce_bank_size', 4096),
            nce_temperature=config.get('nce_temperature', 0.1),
        ))
        self.attach('head', nn.Sequential(
            nn.LayerNorm(config['d_model']),
            nn.Linear(config['d_model'], config['d_model']),
            nn.GELU(), nn.Dropout(config['dropout']),
            nn.Linear(config['d_model'], config['num_classes']),
        ))

    def forward(self, x, return_geo_state=False):
        tokens, svd_state = self['patch_embed'](x)
        if return_geo_state:
            features, geo_states = self['transformer'](tokens, return_geo_state=True)
        else:
            features = self['transformer'](tokens)
        cls_out = features[:, 0]
        logits = self['head'](cls_out)
        if return_geo_state:
            return logits, geo_states, svd_state
        return logits

    def geometric_losses(self):
        """Delegate to transformer's built-in geometric regularization."""
        return self['transformer'].geometric_losses(
            cv_target=self.config.get('cv_target', 0.215),
            cv_weight=self.config.get('cv_weight', 0.1),
            spread_weight=self.config.get('spread_weight', 0.01),
        )

    def infonce_loss(self):
        """InfoNCE contrastive loss on CLS token's geometric residual.
        Uses cached residual from last forward pass."""
        return self['transformer'].infonce_loss()

    def update_nce_bank(self):
        """Enqueue current batch's residuals. Call AFTER backward."""
        self['transformer'].update_nce_bank()

    def anchor_diagnostics(self):
        """Delegate to transformer's anchor diagnostics."""
        return self['transformer'].anchor_diagnostics()


# ═══════════════════════════════════════════════════════════════════════════════
# GEOMETRIC ANALYSIS BATTERY ( includes CM diagnostics)
# ═══════════════════════════════════════════════════════════════════════════════

@torch.no_grad()
def compute_cv(points):
    """Coefficient of variation on S^(d-1).
    CV = std(pairwise_cosine_distances) / mean(pairwise_cosine_distances)
    Pentachoron band: CV ∈ [0.20, 0.23].
    """
    points = F.normalize(points.float(), dim=-1)
    cos_sim = points @ points.T
    n = points.shape[0]
    idx = torch.triu_indices(n, n, offset=1, device=points.device)
    pairwise_dist = 1.0 - cos_sim[idx[0], idx[1]]
    mean_d = pairwise_dist.mean()
    std_d = pairwise_dist.std()
    cv = (std_d / (mean_d + 1e-8)).item()
    return cv, mean_d.item(), std_d.item()


@torch.no_grad()
def log_geometric_analysis(model, writer, epoch, test_loader, device, config):
    """Full geometric analysis battery with CM diagnostics."""
    model.eval()

    images, labels = next(iter(test_loader))
    images = images[:min(64, images.shape[0])].to(device)
    labels = labels[:min(64, labels.shape[0])].to(device)

    logits, geo_states, svd_state = model(images, return_geo_state=True)

    n_layers = len(geo_states)
    pred = logits.argmax(1)
    batch_acc = (pred == labels).float().mean().item()
    writer.add_scalar('analysis/batch_accuracy', batch_acc, epoch)

    # ─── SVD Input Stage ───
    S = svd_state['singular_values']
    s_norm = S / (S.sum(dim=-1, keepdim=True) + 1e-8)
    s_ent = -(s_norm * torch.log(s_norm.clamp(min=1e-8))).sum(-1)
    novelty = svd_state['novelty']

    writer.add_scalar('svd/entropy_mean', s_ent.mean().item(), epoch)
    writer.add_scalar('svd/entropy_std', s_ent.std().item(), epoch)
    writer.add_scalar('svd/novelty_norm', novelty.norm(dim=-1).mean().item(), epoch)
    writer.add_scalar('svd/top1_ratio', (S[:, 0] / (S.sum(-1) + 1e-8)).mean().item(), epoch)
    writer.add_scalar('svd/condition_number',
        (S[:, 0] / (S[:, -1].clamp(min=1e-8))).mean().item(), epoch)
    for k in range(min(S.shape[1], 5)):
        writer.add_scalar(f'svd/S_{k}', S[:, k].mean().item(), epoch)

    # SVD FiLM deviation
    pe = model['patch_embed']
    writer.add_scalar('svd_film/gamma_weight_norm', pe.svd_to_gamma.weight.data.norm().item(), epoch)
    writer.add_scalar('svd_film/gamma_bias_dev_from_1',
        (pe.svd_to_gamma.bias.data - 1.0).abs().mean().item(), epoch)
    writer.add_scalar('svd_film/beta_weight_norm', pe.svd_to_beta.weight.data.norm().item(), epoch)
    writer.add_scalar('svd_film/beta_bias_norm', pe.svd_to_beta.bias.data.abs().mean().item(), epoch)

    # ─── Anchor Diagnostics (built-in) ───
    anchor_diag = model.anchor_diagnostics()
    for layer_name, d in anchor_diag.items():
        for k, v in d.items():
            writer.add_scalar(f'anchor_diag/{layer_name}_{k}', v, epoch)

    # ─── Per-Layer Geometric Analysis ───
    for i, gs in enumerate(geo_states):
        prefix = f'layer_{i}'

        # === CV — pentachoron band metric ===
        emb = gs['embedding']
        # Anchor CV
        transformer = model['transformer']
        layer = transformer[f'layer_{i}']
        anchors = F.normalize(
            layer['observer'].association.constellation.anchors, dim=-1)
        cv_anchors, mean_d_anchors, std_d_anchors = compute_cv(anchors)
        writer.add_scalar(f'{prefix}/cv_anchors', cv_anchors, epoch)
        writer.add_scalar(f'{prefix}/anchor_mean_dist', mean_d_anchors, epoch)
        writer.add_scalar(f'{prefix}/anchor_std_dist', std_d_anchors, epoch)

        # Embedding CV
        emb_flat = emb.reshape(-1, emb.shape[-1])
        n_sample = min(512, emb_flat.shape[0])
        idx = torch.randperm(emb_flat.shape[0], device=device)[:n_sample]
        cv_emb, mean_d_emb, std_d_emb = compute_cv(emb_flat[idx])
        writer.add_scalar(f'{prefix}/cv_embeddings', cv_emb, epoch)
        writer.add_scalar(f'{prefix}/embedding_mean_dist', mean_d_emb, epoch)

        # === CM Gate Diagnostics ===
        gate_info = gs.get('gate_info', {})
        gate_values = gs.get('gate_values')
        cm_quality = gs.get('cm_quality')

        if gate_info:
            writer.add_scalar(f'{prefix}/cm_active_anchors',
                gate_info.get('active', 0), epoch)
            writer.add_scalar(f'{prefix}/cm_gate_mean',
                gate_info.get('gate_mean', 0), epoch)
            writer.add_scalar(f'{prefix}/cm_positive_frac',
                gate_info.get('cm_positive_frac', 0), epoch)

        if gate_values is not None:
            gv = gate_values
            writer.add_scalar(f'{prefix}/gate_values_min', gv.min().item(), epoch)
            writer.add_scalar(f'{prefix}/gate_values_max', gv.max().item(), epoch)
            writer.add_scalar(f'{prefix}/gate_values_std', gv.std().item(), epoch)
            # Per-anchor gate mean (which anchors are consistently open/closed)
            gv_per_anchor = gv.mean(dim=0).mean(dim=0)  # average over B and L
            writer.add_scalar(f'{prefix}/gate_anchor_spread',
                gv_per_anchor.std().item(), epoch)
            # Fraction of positions with >50% anchors open
            if gv.dim() == 3:
                pos_open_frac = (gv.mean(dim=-1) > 0.5).float().mean().item()
            else:
                pos_open_frac = (gv > 0.5).float().mean().item()
            writer.add_scalar(f'{prefix}/gate_positions_open_frac', pos_open_frac, epoch)

        if cm_quality is not None:
            writer.add_scalar(f'{prefix}/cm_quality_mean', cm_quality.mean().item(), epoch)
            writer.add_scalar(f'{prefix}/cm_quality_std', cm_quality.std().item(), epoch)
            writer.add_scalar(f'{prefix}/cm_quality_min', cm_quality.min().item(), epoch)

        # === Stream Agreement ===
        content = gs['content']
        geometric = gs['geometric']
        agreement = F.cosine_similarity(
            content.reshape(-1, content.shape[-1]),
            geometric.reshape(-1, geometric.shape[-1]), dim=-1)
        writer.add_scalar(f'{prefix}/stream_agreement_mean', agreement.mean().item(), epoch)
        writer.add_scalar(f'{prefix}/stream_agreement_std', agreement.std().item(), epoch)

        writer.add_scalar(f'{prefix}/content_norm', content.norm(dim=-1).mean().item(), epoch)
        writer.add_scalar(f'{prefix}/geometric_norm', geometric.norm(dim=-1).mean().item(), epoch)

        # === Disagreement arm analysis ===
        disagree = content - geometric
        agree = content * geometric
        writer.add_scalar(f'{prefix}/disagree_norm', disagree.norm(dim=-1).mean().item(), epoch)
        writer.add_scalar(f'{prefix}/agree_norm', agree.norm(dim=-1).mean().item(), epoch)

        # === Anchor Utilization ===
        tri = gs['triangulation']
        assignment = gs['assignment']
        nearest = gs['nearest']
        n_anchors = tri.shape[-1]

        nearest_flat = nearest.reshape(-1)
        counts = torch.bincount(nearest_flat, minlength=n_anchors).float()
        total_assignments = counts.sum()

        probs = counts / (total_assignments + 1e-8)
        anchor_entropy = -(probs * torch.log(probs.clamp(min=1e-8))).sum().item()
        max_entropy = math.log(n_anchors)
        writer.add_scalar(f'{prefix}/anchor_entropy_normalized',
            anchor_entropy / (max_entropy + 1e-8), epoch)
        active = (counts > 0).sum().item()
        writer.add_scalar(f'{prefix}/anchors_active', active, epoch)
        writer.add_scalar(f'{prefix}/anchors_active_frac', active / n_anchors, epoch)
        dead = (counts == 0).sum().item()
        writer.add_scalar(f'{prefix}/anchors_dead', dead, epoch)

        # === Triangulation Statistics ===
        writer.add_scalar(f'{prefix}/tri_mean', tri.mean().item(), epoch)
        writer.add_scalar(f'{prefix}/tri_std', tri.std().item(), epoch)

        # === Soft Assignment Statistics ===
        assign_ent = -(assignment * torch.log(assignment.clamp(min=1e-8))).sum(-1)
        writer.add_scalar(f'{prefix}/assignment_entropy_mean', assign_ent.mean().item(), epoch)
        writer.add_scalar(f'{prefix}/assignment_max_prob',
            assignment.max(dim=-1).values.mean().item(), epoch)

        # === Patchwork Statistics (now from CM-validated triangulation) ===
        pw = gs['patchwork']
        writer.add_scalar(f'{prefix}/patchwork_norm', pw.norm(dim=-1).mean().item(), epoch)
        writer.add_scalar(f'{prefix}/patchwork_std', pw.std().item(), epoch)
        pw_sparsity = (pw.abs() < 0.01).float().mean().item()
        writer.add_scalar(f'{prefix}/patchwork_sparsity', pw_sparsity, epoch)

        # === Bridge Consistency ===
        bridge = gs['bridge']
        bridge_soft = F.softmax(bridge, dim=-1)
        bridge_assign_kl = F.kl_div(
            bridge_soft.log().reshape(-1, n_anchors),
            assignment.reshape(-1, n_anchors),
            reduction='batchmean', log_target=False)
        writer.add_scalar(f'{prefix}/bridge_assignment_kl', bridge_assign_kl.item(), epoch)

        # === Quaternion Composition ===
        composed = gs['composed']
        writer.add_scalar(f'{prefix}/composed_norm', composed.norm(dim=-1).mean().item(), epoch)

        # === Geo Context ===
        geo_ctx = gs['geo_ctx']
        writer.add_scalar(f'{prefix}/geo_ctx_norm', geo_ctx.norm(dim=-1).mean().item(), epoch)

        # === Geometric Residual Stream (CM-conditioned) ===
        geo_res = gs.get('geo_residual')
        if geo_res is not None:
            res_norms = geo_res.norm(dim=-1)
            writer.add_scalar(f'{prefix}/geo_res_norm', res_norms.mean().item(), epoch)
            writer.add_scalar(f'{prefix}/geo_res_std', geo_res.std().item(), epoch)
            writer.add_scalar(f'{prefix}/geo_res_sparsity',
                (geo_res.abs() < 0.01).float().mean().item(), epoch)
            # Cross-position consistency
            geo_res_flat = geo_res.reshape(-1, geo_res.shape[-1])
            n_s = min(256, geo_res_flat.shape[0])
            idx_s = torch.randperm(geo_res_flat.shape[0], device=geo_res.device)[:n_s]
            sampled = F.normalize(geo_res_flat[idx_s], dim=-1)
            cos_mat = sampled @ sampled.T
            triu = torch.triu_indices(n_s, n_s, offset=1, device=geo_res.device)
            writer.add_scalar(f'{prefix}/geo_res_consistency',
                cos_mat[triu[0], triu[1]].mean().item(), epoch)

    # ─── Cayley Rotation Analysis ───
    for name, mod in model.named_modules():
        if isinstance(mod, CayleyOrthogonal):
            R = mod.get_rotation()
            I = torch.eye(R.shape[0], device=R.device)
            r_dist = (R - I).norm().item()
            clean_name = name.replace('.', '_')
            writer.add_scalar(f'cayley/{clean_name}_R_minus_I', r_dist, epoch)

    # ─── FiLM Layer Analysis ───
    film_idx = 0
    for name, mod in model.named_modules():
        if isinstance(mod, FiLMLayer):
            g_b = mod.to_gamma.bias.data
            b_b = mod.to_beta.bias.data
            writer.add_scalar(f'film/{film_idx}_gamma_dev',
                (g_b - 1.0).abs().mean().item(), epoch)
            writer.add_scalar(f'film/{film_idx}_beta_dev',
                b_b.abs().mean().item(), epoch)
            film_idx += 1

    # ─── Cross-Layer Trajectories ───
    cv_trajectory = []
    cm_quality_trajectory = []
    res_norms = []
    bridge_kls = []

    for i, gs in enumerate(geo_states):
        # CV
        emb = gs['embedding']
        emb_flat = emb.reshape(-1, emb.shape[-1])
        n_sample = min(512, emb_flat.shape[0])
        idx = torch.randperm(emb_flat.shape[0], device=device)[:n_sample]
        cv, _, _ = compute_cv(emb_flat[idx])
        cv_trajectory.append(cv)

        # CM quality
        cm_q = gs.get('cm_quality')
        if cm_q is not None:
            cm_quality_trajectory.append(cm_q.mean().item())

        # Geo residual norms
        geo_res = gs.get('geo_residual')
        if geo_res is not None:
            res_norms.append(geo_res.norm(dim=-1).mean().item())

        # Bridge KL
        n_anchors = gs['assignment'].shape[-1]
        bridge_soft = F.softmax(gs['bridge'], dim=-1)
        bkl = F.kl_div(
            bridge_soft.log().reshape(-1, n_anchors),
            gs['assignment'].reshape(-1, n_anchors),
            reduction='batchmean', log_target=False).item()
        bridge_kls.append(bkl)

    # CV trajectory
    writer.add_scalar('cv/trajectory_mean', np.mean(cv_trajectory), epoch)
    writer.add_scalar('cv/trajectory_std', np.std(cv_trajectory), epoch)
    in_band = sum(1 for cv in cv_trajectory if 0.20 <= cv <= 0.23)
    writer.add_scalar('cv/layers_in_pentachoron_band', in_band, epoch)
    writer.add_scalar('cv/layers_in_band_frac', in_band / len(cv_trajectory), epoch)

    # CM quality trajectory
    if cm_quality_trajectory:
        writer.add_scalar('cm/quality_trajectory_mean',
            np.mean(cm_quality_trajectory), epoch)
        writer.add_scalar('cm/quality_trajectory_std',
            np.std(cm_quality_trajectory), epoch)
        writer.add_scalar('cm/quality_min_layer',
            np.min(cm_quality_trajectory), epoch)
        writer.add_scalar('cm/quality_max_layer',
            np.max(cm_quality_trajectory), epoch)

    # Geometric residual trajectory
    if res_norms:
        writer.add_scalar('geo_res/trajectory_start', res_norms[0], epoch)
        writer.add_scalar('geo_res/trajectory_end', res_norms[-1], epoch)
        writer.add_scalar('geo_res/accumulation_ratio',
            res_norms[-1] / (res_norms[0] + 1e-8), epoch)
        growth = [res_norms[j+1] - res_norms[j] for j in range(len(res_norms)-1)]
        writer.add_scalar('geo_res/growth_mean', np.mean(growth), epoch)
        writer.add_scalar('geo_res/growth_std', np.std(growth), epoch)

        # Cooperation analysis (includes CM quality)
        if len(res_norms) >= 4:
            cv_corr = float(np.corrcoef(res_norms, cv_trajectory)[0, 1])
            bkl_corr = float(np.corrcoef(res_norms, bridge_kls)[0, 1])
            writer.add_scalar('cooperation/geo_res_vs_cv', cv_corr, epoch)
            writer.add_scalar('cooperation/geo_res_vs_bridge_kl', bkl_corr, epoch)

            if len(cm_quality_trajectory) == len(res_norms):
                cm_corr = float(np.corrcoef(
                    res_norms, cm_quality_trajectory)[0, 1])
                writer.add_scalar('cooperation/geo_res_vs_cm_quality', cm_corr, epoch)
                # CM vs CV: do layers with better CM quality also have better CV?
                cm_cv_corr = float(np.corrcoef(
                    cm_quality_trajectory, cv_trajectory)[0, 1])
                writer.add_scalar('cooperation/cm_quality_vs_cv', cm_cv_corr, epoch)

    return {
        'batch_acc': batch_acc,
        'cv_trajectory': cv_trajectory,
        'cm_quality_trajectory': cm_quality_trajectory,
        'res_norms': res_norms,
        'bridge_kls': bridge_kls,
    }


@torch.no_grad()
def log_gradient_norms(model, writer, epoch):
    """Log gradient norms per component type (includes cm_gate)."""
    type_grads = {}
    for name, param in model.named_parameters():
        if param.grad is not None:
            grad_norm = param.grad.norm().item()
            if 'projection' in name and 'proj' in name:
                key = 'manifold_proj'
            elif 'cm_gate' in name:
                key = 'cm_gate'
            elif 'observer' in name or 'constellation' in name or 'anchor' in name:
                key = 'constellation'
            elif 'context' in name:
                key = 'geo_context'
            elif 'content' in name:
                key = 'content_attn'
            elif 'geometric' in name and 'film' not in name:
                key = 'geo_attn'
            elif 'film' in name:
                key = 'film'
            elif 'rotation' in name or 'cayley' in name or 'A_upper' in name:
                key = 'cayley'
            elif 'compose' in name or 'quat' in name or 'proj_w' in name:
                key = 'quaternion'
            elif 'decode' in name:
                key = 'decode'
            elif 'gate' in name:
                key = 'gate'
            elif 'conv' in name or 'patch' in name:
                key = 'input_stage'
            elif 'head' in name:
                key = 'head'
            elif 'svd' in name:
                key = 'svd'
            elif 'geo_proj' in name:
                key = 'geo_residual_proj'
            else:
                key = 'other'

            if key not in type_grads:
                type_grads[key] = []
            type_grads[key].append(grad_norm)

    for key, norms in type_grads.items():
        writer.add_scalar(f'grad_norm/{key}_mean', np.mean(norms), epoch)
        writer.add_scalar(f'grad_norm/{key}_max', np.max(norms), epoch)

    total = sum(p.grad.norm().item() ** 2
                for p in model.parameters() if p.grad is not None) ** 0.5
    writer.add_scalar('grad_norm/total', total, epoch)


@torch.no_grad()
def log_weight_norms(model, writer, epoch):
    """Log weight norms per component type."""
    for name, param in model.named_parameters():
        if 'A_upper' in name:
            clean = name.replace('.', '_')
            writer.add_scalar(f'weights/{clean}_norm', param.norm().item(), epoch)


# ═══════════════════════════════════════════════════════════════════════════════
# DATA
# ═══════════════════════════════════════════════════════════════════════════════

def get_dataloaders(config):
    import torchvision
    import torchvision.transforms as T

    # Augmentation pipeline tuned for geometric transformer:
    #   TrivialAugmentWide: continuous severity spectrum of geometric + photometric
    #     transforms. Exercises CM gate across full quality range — mild distortion
    #     keeps CM high, severe distortion creates partially-degenerate simplices.
    #   RandomErasing: creates degenerate manifold projections (zero-volume CM simplices).
    #     Trains CM gate to close on corrupted regions.
    #   CutMix applied at batch level in train_epoch (not here).
    train_transform = T.Compose([
        T.RandomCrop(32, padding=4),
        T.RandomHorizontalFlip(),
        T.TrivialAugmentWide(),
        T.ToTensor(),
        T.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
        T.RandomErasing(p=config.get('random_erasing_p', 0.25),
                        scale=(0.02, 0.33), ratio=(0.3, 3.3)),
    ])
    test_transform = T.Compose([
        T.ToTensor(),
        T.Normalize((0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)),
    ])

    train_ds = torchvision.datasets.CIFAR100(
        root='./data', train=True, download=True, transform=train_transform)
    test_ds = torchvision.datasets.CIFAR100(
        root='./data', train=False, download=True, transform=test_transform)

    train_loader = torch.utils.data.DataLoader(
        train_ds, batch_size=config['batch_size'], shuffle=True,
        num_workers=config['num_workers'], pin_memory=True, drop_last=True)
    test_loader = torch.utils.data.DataLoader(
        test_ds, batch_size=config['batch_size'], shuffle=False,
        num_workers=config['num_workers'], pin_memory=True)

    return train_loader, test_loader


# ═══════════════════════════════════════════════════════════════════════════════
# CUTMIX — batch-level augmentation for CM gate boundary training
# ═══════════════════════════════════════════════════════════════════════════════

def cutmix_batch(images, labels, alpha=1.0):
    """Apply CutMix to a batch. Returns mixed images + label pairs + lambda.

    CutMix replaces a rectangular region of image A with image B.
    Positions inside each region have coherent geometry — valid CM simplices.
    The boundary between regions has mixed geometric context — the CM gate
    should learn to suppress these positions.

    Args:
        images: (B, C, H, W) batch
        labels: (B,) integer labels
        alpha: Beta distribution parameter (1.0 = uniform box sizes)

    Returns:
        images: (B, C, H, W) mixed batch (modified in-place)
        labels_a: (B,) labels for region A
        labels_b: (B,) labels for region B
        lam: float, fraction of image A remaining
    """
    lam = np.random.beta(alpha, alpha)
    B = images.size(0)
    idx = torch.randperm(B, device=images.device)

    H, W = images.shape[2], images.shape[3]
    cut_ratio = (1.0 - lam) ** 0.5
    cw = int(W * cut_ratio)
    ch = int(H * cut_ratio)
    cx = np.random.randint(W)
    cy = np.random.randint(H)
    x1 = max(cx - cw // 2, 0); x2 = min(cx + cw // 2, W)
    y1 = max(cy - ch // 2, 0); y2 = min(cy + ch // 2, H)

    images[:, :, y1:y2, x1:x2] = images[idx, :, y1:y2, x1:x2]
    lam_actual = 1.0 - (x2 - x1) * (y2 - y1) / (W * H)
    return images, labels, labels[idx], lam_actual


# ═══════════════════════════════════════════════════════════════════════════════
# TRAINING (geometric losses + CutMix integrated)
# ═══════════════════════════════════════════════════════════════════════════════

def train_epoch(model, loader, optimizer, scheduler, epoch, config, writer):
    model.train()
    total_loss = 0
    total_geo_loss = 0
    total_nce_loss = 0
    correct = 0
    total = 0

    cutmix_alpha = config.get('cutmix_alpha', 1.0)
    cutmix_prob = config.get('cutmix_prob', 0.5)
    label_smoothing = config.get('label_smoothing', 0.1)
    nce_weight = config.get('nce_weight', 0.1)
    criterion = nn.CrossEntropyLoss(label_smoothing=label_smoothing)

    for batch_idx, (images, labels) in enumerate(loader):
        images = images.to(device)
        labels = labels.to(device)

        # CutMix: applied probabilistically per batch
        use_cutmix = np.random.rand() < cutmix_prob
        if use_cutmix:
            images, labels_a, labels_b, lam = cutmix_batch(
                images, labels, alpha=cutmix_alpha)
            logits = model(images)
            ce_loss = lam * criterion(logits, labels_a) + \
                      (1.0 - lam) * criterion(logits, labels_b)
            # Accuracy: count correct if matches either label
            pred = logits.argmax(1)
            correct += (lam * (pred == labels_a).float() +
                        (1.0 - lam) * (pred == labels_b).float()).sum().item()
        else:
            logits = model(images)
            ce_loss = criterion(logits, labels)
            correct += (logits.argmax(1) == labels).sum().item()

        # Geometric regularization — CV target + anchor spread
        geo_losses = model.geometric_losses()
        geo_loss = geo_losses.get('geo_total', torch.tensor(0.0, device=device))

        # InfoNCE on geometric residual — discriminative pressure
        nce_losses = model.infonce_loss()
        nce_loss = nce_losses.get('nce', torch.tensor(0.0, device=device))

        loss = ce_loss + geo_loss + nce_weight * nce_loss

        optimizer.zero_grad()
        loss.backward()

        # Enqueue AFTER backward — detached residuals go into bank
        model.update_nce_bank()

        # Log gradient norms periodically
        if epoch % config['log_grads_every'] == 0 and batch_idx == 0:
            log_gradient_norms(model, writer, epoch)

        torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
        optimizer.step()
        if scheduler is not None:
            scheduler.step()

        total_loss += ce_loss.item() * images.size(0)
        total_geo_loss += geo_loss.item() * images.size(0)
        total_nce_loss += nce_loss.item() * images.size(0)
        total += images.size(0)

    avg_ce = total_loss / total
    avg_geo = total_geo_loss / total
    avg_nce = total_nce_loss / total
    return avg_ce, avg_geo, avg_nce, correct / total


@torch.no_grad()
def evaluate(model, loader):
    model.eval()
    correct = 0
    total = 0
    for images, labels in loader:
        images = images.to(device)
        labels = labels.to(device)
        logits = model(images)
        correct += (logits.argmax(1) == labels).sum().item()
        total += images.size(0)
    return correct / total


def main():
    config = CONFIG.copy()

    print("=" * 60)
    print("  Geometric Transformer — CIFAR-100 (CM-Validated)")
    print(f"  Input: conv({config['in_channels']}→{config['conv_channels']}) + "
          f"SVD(rank={config['svd_rank']}) + "
          f"{config['patch_size']}×{config['patch_size']} patches = "
          f"{(config['img_size']//config['patch_size'])**2} tokens + CLS")
    print(f"  Model: d={config['d_model']}, heads={config['n_heads']}, "
          f"layers={config['n_layers']}, anchors={config['n_anchors']}")
    print(f"  CM: neighbors={config['cm_neighbors']}, "
          f"cv_target={config['cv_target']}, "
          f"cv_weight={config['cv_weight']}, "
          f"spread_weight={config['spread_weight']}")
    print(f"  Aug: TrivialAugmentWide + CutMix(α={config['cutmix_alpha']}, "
          f"p={config['cutmix_prob']}) + "
          f"RandomErasing(p={config['random_erasing_p']})")
    print(f"  NCE: bank={config['nce_bank_size']}, "
          f"temp={config['nce_temperature']}, "
          f"weight={config['nce_weight']}")
    print("=" * 60)

    writer = SummaryWriter(config['log_dir'])
    writer.add_text('config', json.dumps(config, indent=2))

    print("\nLoading CIFAR-100...")
    train_loader, test_loader = get_dataloaders(config)
    print(f"  Train: {len(train_loader.dataset):,} | Test: {len(test_loader.dataset):,}")

    model = GeoViTClassifier('geo_vit_cifar100', config)
    if hasattr(model, 'network_to'):
        model.network_to(device=device, strict=False)
    else:
        model = model.to(device)

    n_params = sum(p.numel() for p in model.parameters())
    n_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    print(f"\n  Total params:     {n_params:,}")
    print(f"  Trainable params: {n_trainable:,}")

    for name, module in model.named_children():
        n = sum(p.numel() for p in module.parameters())
        if n > 0:
            print(f"    {name:<20s}: {n:,}")

    writer.add_scalar('model/total_params', n_params, 0)

    # Initial anchor diagnostics
    print(f"\n  Initial anchor diagnostics:")
    diag = model.anchor_diagnostics()
    for layer_name, d in diag.items():
        print(f"    {layer_name}: cv={d['anchor_cv']:.4f}, "
              f"cm_pos={d['cm_positive_frac']:.3f}, "
              f"min_dist={d['min_pairwise_dist']:.4f}")

    # Optimizer + scheduler
    optimizer = torch.optim.AdamW(
        model.parameters(), lr=config['lr'], weight_decay=config['weight_decay'])

    total_steps = config['epochs'] * len(train_loader)
    warmup_steps = config['warmup_epochs'] * len(train_loader)

    def lr_lambda(step):
        if step < warmup_steps:
            return step / warmup_steps
        progress = (step - warmup_steps) / (total_steps - warmup_steps)
        return 0.5 * (1 + np.cos(np.pi * progress))

    scheduler = torch.optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)

    print(f"\n{'━'*60}")
    print(f"  Training for {config['epochs']} epochs")
    print(f"  Warmup: {config['warmup_epochs']} epochs, "
          f"LR: {config['lr']}, WD: {config['weight_decay']}")
    print(f"  Geo reg: cv_w={config['cv_weight']}, spread_w={config['spread_weight']}")
    print(f"  NCE bank: size={config['nce_bank_size']}, "
          f"temp={config['nce_temperature']}, weight={config['nce_weight']}")
    print(f"  Aug: TrivialAugmentWide + CutMix(p={config['cutmix_prob']}) + "
          f"RandomErasing(p={config['random_erasing_p']})")
    print(f"  TensorBoard: {config['log_dir']}")
    print(f"  Geo analysis every {config['log_geo_every']} epochs")
    print(f"{'━'*60}\n")

    best_acc = 0
    save_dir = Path('geo_cifar100'); save_dir.mkdir(exist_ok=True)

    for epoch in tqdm(range(config['epochs']), desc="Epochs"):
        t0 = time.time()

        ce_loss, geo_loss, nce_loss, train_acc = train_epoch(
            model, train_loader, optimizer, scheduler, epoch, config, writer)

        test_acc = evaluate(model, test_loader)
        elapsed = time.time() - t0

        lr = optimizer.param_groups[0]['lr']
        writer.add_scalar('train/ce_loss', ce_loss, epoch)
        writer.add_scalar('train/geo_loss', geo_loss, epoch)
        writer.add_scalar('train/nce_loss', nce_loss, epoch)
        writer.add_scalar('train/total_loss', ce_loss + geo_loss + nce_loss, epoch)
        writer.add_scalar('train/accuracy', train_acc, epoch)
        writer.add_scalar('test/accuracy', test_acc, epoch)
        writer.add_scalar('train/lr', lr, epoch)
        writer.add_scalar('train/epoch_time', elapsed, epoch)
        writer.add_scalar('gap/train_test', train_acc - test_acc, epoch)

        log_weight_norms(model, writer, epoch)

        if test_acc > best_acc:
            best_acc = test_acc
            torch.save({
                'state_dict': {k: v.cpu() for k, v in model.state_dict().items()},
                'epoch': epoch,
                'test_acc': test_acc,
                'config': config,
            }, save_dir / 'best.pt')

        # Full geometric analysis periodically
        if epoch % config['log_geo_every'] == 0 or epoch == config['epochs'] - 1:
            geo_info = log_geometric_analysis(
                model, writer, epoch, test_loader, device, config)

            cv_str = ', '.join(f'{cv:.3f}' for cv in geo_info['cv_trajectory'])
            cm_str = ', '.join(f'{q:.3f}' for q in geo_info.get('cm_quality_trajectory', []))
            res_str = ', '.join(f'{r:.3f}' for r in geo_info.get('res_norms', []))
            tqdm.write(
                f"  E{epoch:>3d}  ce={ce_loss:.4f}  geo={geo_loss:.4f}  "
                f"nce={nce_loss:.4f}  "
                f"train={train_acc:.4f}  test={test_acc:.4f}  "
                f"best={best_acc:.4f}  {elapsed:.1f}s"
                f"\n        CV=[{cv_str}]"
                f"\n        CM=[{cm_str}]"
                f"\n        GR=[{res_str}]")
        elif epoch % 5 == 0:
            tqdm.write(
                f"  E{epoch:>3d}  ce={ce_loss:.4f}  geo={geo_loss:.4f}  "
                f"nce={nce_loss:.4f}  "
                f"train={train_acc:.4f}  test={test_acc:.4f}  "
                f"best={best_acc:.4f}  {elapsed:.1f}s")

    # Final summary
    print(f"\n{'═'*60}")
    print(f"  CIFAR-100 RESULTS (CM-Validated)")
    print(f"{'═'*60}")
    print(f"  Best test accuracy: {best_acc:.4f} ({best_acc*100:.2f}%)")
    print(f"  Parameters: {n_params:,}")
    print(f"  Checkpoint: {save_dir}/best.pt")
    print(f"  TensorBoard: {config['log_dir']}")

    # Final geometric state + anchor diagnostics
    print(f"\n  Final geometric state:")
    geo_info = log_geometric_analysis(
        model, writer, config['epochs'], test_loader, device, config)

    print(f"\n  Final anchor diagnostics:")
    diag = model.anchor_diagnostics()
    for layer_name, d in diag.items():
        print(f"    {layer_name}: cv={d['anchor_cv']:.4f}, "
              f"cm_pos={d['cm_positive_frac']:.3f}, "
              f"cm_mean={d['cm_mean']:.4f}")

    writer.close()
    print(f"\nDone.")


if __name__ == '__main__':
    main()