File size: 30,427 Bytes
e841b45
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
"""
SRT Router: Statistical Routing Theory implementation.

Two modes:
  --srt_metric_mode hard:
    Whitening + ZCA + L2 (equivalent to Pooled Mahalanobis per Theorem 4).
    Matches routing_analysis experiment: whitening fitted on train, applied always.
    Single fixed metric β†’ argmin is always valid.

  --srt_metric_mode dynamics:
    No whitening. SRM global metric selection on synthetic data.
    Bug fixed: SRM runs AFTER add_task so ALL tasks get SRM-assigned metric.
    β†’ All tasks use same metric β†’ argmin is valid.

Zero-rehearsal compliant: only sufficient statistics, no raw data.
"""

import math
import numpy as np
import torch
from typing import Dict, List, Tuple, Optional, Union


# ─────────────────────────────────────────────────────────────────────────────
#  METRICS
# ─────────────────────────────────────────────────────────────────────────────

def metric_l2(
    h: np.ndarray,    # (n, d) or (d,)
    mu: np.ndarray,
) -> np.ndarray:
    """L2 distance from centroid. Use when isotropic (PaR β‰ˆ d)."""
    if h.ndim == 1:
        h = h.reshape(1, -1)
    diff = h - mu
    return np.sqrt(np.einsum('nd,nd->n', diff, diff))


def metric_mahalanobis(
    h: np.ndarray,
    mu: np.ndarray,
    Sinv: np.ndarray,
) -> np.ndarray:
    """Mahalanobis distance. Use when anisotropic (PaR β‰ͺ d)."""
    if h.ndim == 1:
        h = h.reshape(1, -1)
    diff = h - mu
    return np.einsum('nd,dg,ng->n', diff, Sinv, diff)


def metric_psr(
    h: np.ndarray,
    mu: np.ndarray,
    eigvecs: np.ndarray,   # (d, k)
    eigvals: np.ndarray,   # (k,)
    d_total: int,
) -> np.ndarray:
    """
    Probabilistic Subspace Routing (PSR) metric from C1 Β§3.3.

    d_PSR(h, t) = ||U_tα΅€(h - ΞΌ_t)||Β² + (r/(d-r)) Β· ||(I - U_t U_tα΅€)(h - ΞΌ_t)||Β²

    Use when low-rank anisotropic (PaR β‰ͺ d, effective dims low).
    """
    if h.ndim == 1:
        h = h.reshape(1, -1)
    diff = h - mu          # (n, d)

    proj = diff @ eigvecs                       # (n, k)
    in_sub = np.sum(proj ** 2, axis=-1)         # (n,)

    v_norm_sq = np.einsum('nd,nd->n', diff, diff)             # (n,)
    out_sub = v_norm_sq - in_sub                               # (n,)

    k = eigvecs.shape[1]
    scale = k / max(d_total - k, 1)
    return in_sub + scale * out_sub


def pinv_ridge(Sigma: np.ndarray, rcond: float = 1e-6) -> np.ndarray:
    """Stable pseudo-inverse of a symmetric covariance matrix."""
    eigvals, eigvecs = np.linalg.eigh(Sigma)
    eigvals = np.maximum(eigvals, rcond * np.max(eigvals))
    idx = np.argsort(eigvals)[::-1]
    eigvals = eigvals[idx]
    eigvecs = eigvecs[:, idx]
    return eigvecs @ np.diag(1.0 / eigvals) @ eigvecs.T


# ─────────────────────────────────────────────────────────────────────────────
#  ZCA WHITENING
# ─────────────────────────────────────────────────────────────────────────────

def compute_whitening(task_embs: Dict[int, np.ndarray]) -> Tuple[np.ndarray, np.ndarray]:
    """
    Compute ZCA whitening transform from pooled embeddings.

    ZCA: W such that (h - ΞΌ)α΅€ Wα΅€ W (h - ΞΌ) is isotropic.

    Fit on ALL task embeddings pooled together (per experiment design
    in compare_routing.py: fit on train, apply to test).

    Returns: (mu_global, W_zca) where W_zca @ (h - mu_global) whitens h.
    """
    all_embs = np.vstack(list(task_embs.values()))
    mu_global = all_embs.mean(0)
    Xc = all_embs - mu_global
    cov_global = np.cov(Xc, rowvar=False)
    eigvals, eigvecs = np.linalg.eigh(cov_global)
    eigvals = np.maximum(eigvals, 1e-8)
    idx = np.argsort(eigvals)[::-1]
    eigvals = eigvals[idx]
    eigvecs = eigvecs[:, idx]
    W_zca = eigvecs @ np.diag(1.0 / np.sqrt(eigvals)) @ eigvecs.T
    return mu_global, W_zca


def apply_whitening(
    h: np.ndarray,
    mu_global: np.ndarray,
    W_zca: np.ndarray,
) -> np.ndarray:
    """Apply ZCA whitening: h_whitened = (h - mu_global) @ W_zca.T."""
    if h.ndim == 1:
        h = h.reshape(1, -1)
    return (h - mu_global) @ W_zca.T


# ─────────────────────────────────────────────────────────────────────────────
#  PARTICIPATION RATIO
# ─────────────────────────────────────────────────────────────────────────────

def participation_ratio(Sigma: np.ndarray) -> float:
    """PaR = (Σλ)² / Σλ² ∈ [1, d]."""
    eigvals = np.linalg.eigvalsh(Sigma)
    eigvals = np.maximum(eigvals, 1e-10)
    return (eigvals.sum() ** 2) / (eigvals ** 2).sum()


def ledoit_wolf_shrinkage(Sigma: np.ndarray, factor: float = 0.1) -> np.ndarray:
    """
    Ledoit-Wolf shrinkage toward scalar identity.

    Ξ£_shrunk = (1 - Ξ΄) Β· Ξ£ + Ξ΄ Β· Ξ»Μ„ Β· I
    """
    d = Sigma.shape[0]
    trace = np.trace(Sigma)
    sigma_mean = trace / d
    target = sigma_mean * np.eye(d)
    return (1 - factor) * Sigma + factor * target


# ─────────────────────────────────────────────────────────────────────────────
#  POOLED COVARIANCE UPDATE  (Welford–Hart)
# ─────────────────────────────────────────────────────────────────────────────

def update_pooled_cov(
    mu_old: np.ndarray,
    cov_old: np.ndarray,
    n_old: int,
    mu_new: np.ndarray,
    cov_new: np.ndarray,
    n_new: int,
) -> Tuple[np.ndarray, np.ndarray, int]:
    """
    Welford–Hart compact update for pooled mean and covariance.
    """
    total = n_old + n_new
    if total <= 1:
        raise ValueError(f"total={total} must be > 1")

    mu_pool = (n_old * mu_old + n_new * mu_new) / total

    delta = mu_new - mu_old
    C = (n_old * n_new / total) * np.outer(delta, delta)

    cov_pool = (
        (n_old - 1) * cov_old
        + (n_new - 1) * cov_new
        + C
    ) / (total - 1)

    return mu_pool, cov_pool, total


# ─────────────────────────────────────────────────────────────────────────────
#  POOLED SHRINKAGE  (Theorem 5)
# ─────────────────────────────────────────────────────────────────────────────

def pooled_shrinkage_target(
    Sigma_t: np.ndarray,
    Sigma_pool: np.ndarray,
    n_t: int,
    n_pool: int,
    alpha_min: float = 0.01,
    alpha_max: float = 0.99,
) -> Tuple[np.ndarray, float]:
    """
    Compute optimal pooled shrinkage target for task t (Theorem 5 from C1).

    Optimal:  Ξ£_t* = (1 - Ξ±_t*) Ξ£Μ‚_t + Ξ±_t* Ξ£Μ‚_pool
    where Ξ±_t* β‰ˆ n_pool / (n_pool + n_t)
    """
    alpha_opt = n_pool / (n_pool + n_t)
    alpha_opt = max(alpha_min, min(alpha_max, alpha_opt))
    Sigma_shrunk = (1 - alpha_opt) * Sigma_t + alpha_opt * Sigma_pool
    return Sigma_shrunk, alpha_opt


# ─────────────────────────────────────────────────────────────────────────────
#  SRM METRIC SELECTION  (Theorem 7 β€” dynamics mode only)
# ─────────────────────────────────────────────────────────────────────────────

def srm_metric_selection(
    signatures: Dict[int, 'TaskSignature'],
    n_synthetic: int = 500,
    random_seed: int = 42,
) -> Tuple[Dict[int, str], Dict[str, float]]:
    """
    Structural Risk Minimization (SRM) from C1 Theorem 7.

    Generates synthetic validation points from each task's Gaussian model,
    evaluates each metric globally, selects the metric minimizing routing error.

    CRITICAL FIX: SRM must run AFTER all tasks (including current) are added
    to signatures. Otherwise current task uses PaR fallback β†’ mixed metrics β†’
    argmin across incomparable distances β†’ misroute.

    Returns: ({task_id: metric}, {metric_name: error_rate})
    """
    rng = np.random.RandomState(random_seed)
    task_list = sorted(signatures.keys())
    T = len(task_list)

    if T < 2:
        return {t: 'l2' for t in task_list}, {}

    # Generate synthetic validation data from each task
    h_val = []
    labels = []
    n_per = n_synthetic // T

    for t_id in task_list:
        sig = signatures[t_id]
        L = np.linalg.cholesky(sig.Sigma + 1e-6 * np.eye(sig.d))
        z = rng.randn(n_per, sig.d)
        h_syn = sig.mu + (L @ z.T).T
        h_val.append(h_syn)
        labels.extend([t_id] * n_per)

    h_val = np.vstack(h_val)
    labels = np.array(labels)

    # Evaluate each metric globally
    errors = {}
    for metric_name in ['l2', 'mahalanobis', 'psr']:
        n_correct = 0
        for i, h_i in enumerate(h_val):
            true_t = labels[i]
            best_t = None
            best_dist = np.inf
            for t_id in task_list:
                sig = signatures[t_id]
                if metric_name == 'l2':
                    d = metric_l2(h_i, sig.mu)
                elif metric_name == 'mahalanobis':
                    d = metric_mahalanobis(h_i, sig.mu, sig.Sinv)
                else:  # psr
                    k = max(1, int(np.sum(sig.eigvals > 1e-6 * sig.eigvals[0])))
                    d = metric_psr(h_i, sig.mu, sig.eigvecs[:, :k], sig.eigvals[:k], sig.d)

                if d < best_dist:
                    best_dist = d
                    best_t = t_id

            if best_t == true_t:
                n_correct += 1

        errors[metric_name] = 1.0 - n_correct / len(labels)

    # Global selection: one metric for ALL tasks
    best_metric = min(errors, key=errors.get)
    print(f"  [SRM] Routing errors: {errors}, selected: {best_metric}")

    # Same metric for all tasks
    return {t_id: best_metric for t_id in task_list}, errors


# ─────────────────────────────────────────────────────────────────────────────
#  TASK SIGNATURE
# ─────────────────────────────────────────────────────────────────────────────

class TaskSignature:
    """
    Statistical signature for one task.
    Stores: ΞΌ_t, Ξ£_t, n_t, eigenvalues, eigenvectors, metric type.
    Zero-rehearsal compliant: only sufficient statistics, no raw data.
    """

    def __init__(
        self,
        task_id: Union[int, str],
        mu: np.ndarray,
        Sigma: np.ndarray,
        n: int,
        metric: str = 'l2',
        alpha: float = 0.0,
        Sigma_raw: Optional[np.ndarray] = None,
        h_train: Optional[np.ndarray] = None,  # kept for backward compat on load; IGNORED
        mu_raw: Optional[np.ndarray] = None,
    ):
        self.task_id = task_id
        self.mu = mu.astype(np.float64)
        self.d = mu.shape[0]
        self.Sigma_raw = (Sigma_raw if Sigma_raw is not None else Sigma).astype(np.float64)
        self.Sigma = Sigma.astype(np.float64)
        self.n = n
        self.alpha = alpha
        self.mu_raw = mu_raw.astype(np.float64) if mu_raw is not None else mu.astype(np.float64).copy()
        self._metric = metric

        # Cached decompositions
        self._eigvals: Optional[np.ndarray] = None
        self._eigvecs: Optional[np.ndarray] = None
        self._Sinv: Optional[np.ndarray] = None
        self._par: Optional[float] = None

    @property
    def eigvals(self) -> np.ndarray:
        if self._eigvals is None:
            self._eigvals, self._eigvecs = np.linalg.eigh(self.Sigma)
            self._eigvals = np.maximum(self._eigvals, 1e-10)
            idx = np.argsort(self._eigvals)[::-1]
            self._eigvals = self._eigvals[idx]
            self._eigvecs = self._eigvecs[:, idx]
        return self._eigvals

    @property
    def eigvecs(self) -> np.ndarray:
        if self._eigvecs is None:
            _ = self.eigvals
        return self._eigvecs

    @property
    def Sinv(self) -> np.ndarray:
        if self._Sinv is None:
            self._Sinv = pinv_ridge(self.Sigma)
        return self._Sinv

    @property
    def par(self) -> float:
        if self._par is None:
            self._par = participation_ratio(self.Sigma)
        return self._par

    @property
    def metric(self) -> str:
        return self._metric

    def reshrink(self, Sigma_pool: np.ndarray, n_pool: int, alpha: float):
        """
        Re-shrink toward pooled covariance using raw Ξ£ as base
        (to avoid compounding shrinkage across re-shrink rounds).
        """
        self.alpha = alpha
        self.Sigma = (1 - alpha) * self.Sigma_raw + alpha * Sigma_pool
        self.Sigma = self.Sigma.astype(np.float64)
        self._eigvals = None
        self._eigvecs = None
        self._Sinv = None
        self._par = None

    def distance(self, h: np.ndarray) -> np.ndarray:
        """
        Compute distance from h to this task's centroid using this task's metric.
        """
        if h.ndim == 1:
            h = h.reshape(1, -1)

        m = self.metric
        if m == 'l2':
            return metric_l2(h, self.mu)
        elif m == 'mahalanobis':
            return metric_mahalanobis(h, self.mu, self.Sinv)
        elif m == 'psr':
            k = max(1, int(np.sum(self.eigvals > 1e-6 * self.eigvals[0])))
            return metric_psr(h, self.mu, self.eigvecs[:, :k], self.eigvals[:k], self.d)
        else:
            # Fallback: PaR-based
            if self.par >= 0.9 * self.d:
                return metric_l2(h, self.mu)
            elif self.par >= 0.3 * self.d:
                return metric_mahalanobis(h, self.mu, self.Sinv)
            else:
                k = max(1, int(np.sum(self.eigvals > 1e-6 * self.eigvals[0])))
                return metric_psr(h, self.mu, self.eigvecs[:, :k], self.eigvals[:k], self.d)

    def to_dict(self) -> dict:
        d = {
            'task_id': self.task_id,
            'mu': self.mu,
            'Sigma': self.Sigma,
            'Sigma_raw': self.Sigma_raw,
            'mu_raw': self.mu_raw,
            'n': self.n,
            'metric': self.metric,
            'alpha': self.alpha,
        }
        # Zero-rehearsal compliant: only sufficient statistics, NO raw data.
        # h_train is NOT needed: Sigma_raw + mu_raw fully encode the distribution.
        # W_zca is reconstructed from Sigma_pool (already stored separately).
        return d

    @classmethod
    def from_dict(cls, d: dict) -> 'TaskSignature':
        return cls(
            task_id=d['task_id'],
            mu=d['mu'],
            Sigma=d['Sigma'],
            n=d['n'],
            metric=d.get('metric', 'l2'),
            alpha=d.get('alpha', 0.0),
            Sigma_raw=d.get('Sigma_raw', d['Sigma']),
            h_train=None,   # h_train intentionally dropped: zero-rehearsal
            mu_raw=d.get('mu_raw'),
        )


# ─────────────────────────────────────────────────────────────────────────────
#  SRT ROUTER
# ─────────────────────────────────────────────────────────────────────────────

class SRTRouter:
    """
    Statistical Routing Theory Router.

    Two modes (set via srt_metric_mode):
      'hard'    : whitening + L2. No SRM. Matches routing_analysis experiment.
                  Whitening fitted on train embeddings, applied always.
                  Single fixed metric (L2 in whitened space = Pooled Mahalanobis).
                  argmin is always valid because all tasks use L2 in same space.

      'dynamics': no whitening. SRM global metric selection on synthetic data.
                  All tasks use same SRM-assigned metric β†’ argmin is valid.
                  Bug fixed: SRM runs AFTER add_task so ALL tasks get metric.

    Zero-drift: no learnable parameters.
    """

    def __init__(
        self,
        srt_metric_mode: str = 'hard',
        use_shrink: bool = True,
        shrink_factor: float = 0.1,
    ):
        """
        Args:
            srt_metric_mode: 'hard' | 'dynamics'
                'hard'     β†’ ZCA whitening + L2 (matches experiment)
                'dynamics' β†’ SRM metric selection (matches contribution_UNIFIED)
            use_shrink: apply Ledoit-Wolf shrinkage to covariance
            shrink_factor: shrinkage intensity
        """
        assert srt_metric_mode in ('hard', 'dynamics'), \
            f"Unknown srt_metric_mode: {srt_metric_mode}"

        self.signatures: Dict[int, TaskSignature] = {}
        self._mu_pool: Optional[np.ndarray] = None
        self._Sigma_pool: Optional[np.ndarray] = None
        self._n_pool: int = 0

        self.srt_metric_mode = srt_metric_mode
        self.use_shrink = use_shrink
        self.shrink_factor = shrink_factor

        # ── ZCA whitening (hard mode) ────────────────────────────────────
        self._mu_global: Optional[np.ndarray] = None
        self._W_zca: Optional[np.ndarray] = None
        self._zca_fitted: bool = False   # True after first ZCA fit; NEVER refit

        # ── SRM state (dynamics mode) ────────────────────────────────────
        self._srm_metrics: Dict[int, str] = {}

    # ── Pooled statistics ───────────────────────────────────────────────

    def _update_pooled(self, mu_t: np.ndarray, Sigma_t: np.ndarray, n_t: int):
        """Update running pooled mean and covariance."""
        if self._n_pool == 0:
            self._mu_pool = mu_t.copy()
            self._Sigma_pool = Sigma_t.copy()
            self._n_pool = n_t
        else:
            self._mu_pool, self._Sigma_pool, self._n_pool = update_pooled_cov(
                self._mu_pool, self._Sigma_pool, self._n_pool,
                mu_t, Sigma_t, n_t,
            )

    def _reshrink_all(self):
        """Re-shrink ALL tasks toward updated pooled covariance (Theorem 5)."""
        if self._n_pool <= 1:
            return
        for t_id, sig in self.signatures.items():
            alpha_opt = self._n_pool / (self._n_pool + sig.n)
            alpha_opt = max(0.01, min(0.99, alpha_opt))
            sig.reshrink(self._Sigma_pool, self._n_pool, alpha_opt)

    # ── Add task ─────────────────────────────────────────────────────

    def add_task(
        self,
        task_id: Union[int, str],
        h_train: np.ndarray,
    ) -> TaskSignature:
        """
        Add a new task's statistical signature.

        HARD MODE β€” ZCA fit-once (the key fix):
          - Task 1: store raw stats only, ZCA NOT fitted yet
          - Task 2+: fit ZCA ONCE from pooled covariance of ALL seen tasks,
            WHITEN ALL tasks (including previous) with this fixed W.
            NEVER refit W again.
          - This matches the offline experiment (compute_whitening on all train
            embeddings once) and avoids the catastrophic incremental refitting
            bug where W changes every add_task() β†’ centroids non-comparable.

        DYNAMICS MODE β€” SRM (unchanged logic):
          - Pooled update + re-shrink + SRM per task.

        Args:
            task_id: integer or string task ID
            h_train: (n_t, d) embeddings from FROZEN backbone

        Returns:
            TaskSignature for this task
        """
        n_t, d = h_train.shape

        # ── Sufficient statistics (always in raw space) ────────────────
        mu_t = h_train.mean(axis=0)
        Sigma_t = np.cov(h_train, rowvar=False, ddof=1)

        # Optional pre-whitening LW shrinkage (on raw covariance)
        if self.use_shrink:
            Sigma_t_shrunk = ledoit_wolf_shrinkage(Sigma_t, factor=self.shrink_factor)
        else:
            Sigma_t_shrunk = Sigma_t.copy()

        # ── Pooled statistics update (Welford–Hart) ───────────────────
        self._update_pooled(mu_t, Sigma_t_shrunk, n_t)

        # ── Dynamics mode ────────────────────────────────────────────────
        if self.srt_metric_mode == 'dynamics':
            # Re-shrink previous tasks toward updated pool
            if len(self.signatures) > 0:
                self._reshrink_all()

            sig = TaskSignature(
                task_id, mu_t, Sigma_t, n_t,
                metric='l2',
                Sigma_raw=Sigma_t,
                h_train=h_train,
                mu_raw=mu_t,
            )
            self.signatures[task_id] = sig

            if len(self.signatures) >= 2:
                srm_results, _ = srm_metric_selection(self.signatures)
                self._srm_metrics = srm_results
                for t_id, m in srm_results.items():
                    self.signatures[t_id]._metric = m
            return sig

        # ── Hard mode: ZCA fit-once ───────────────────────────────────
        # Re-shrink previous tasks toward updated pool (raw Ξ£, pre-whitening)
        if len(self.signatures) > 0:
            self._reshrink_all()

        if not self._zca_fitted:
            # FIRST TIME: fit ZCA from pooled mean + pooled covariance.
            # Uses Ξ£_pool which already aggregates ALL seen tasks via Welford–Hart.
            # This is the single, FINAL whitening transform β€” never refitted.
            self._mu_global = self._mu_pool.copy()
            cov_pool = self._Sigma_pool.copy()

            # Eigendecomposition of pooled covariance β†’ ZCA transform
            eigvals, eigvecs = np.linalg.eigh(cov_pool)
            eigvals = np.maximum(eigvals, 1e-8)
            idx = np.argsort(eigvals)[::-1]
            eigvals = eigvals[idx]
            eigvecs = eigvecs[:, idx]
            self._W_zca = eigvecs @ np.diag(1.0 / np.sqrt(eigvals)) @ eigvecs.T
            self._zca_fitted = True

            # Re-whiten ALL previous tasks (from their raw centroids)
            for sig in self.signatures.values():
                # sig.Sigma was reshrunk-but-unwhitened by _reshrink_all()
                # sig.mu_raw is the unwhitened centroid
                sig.mu = apply_whitening(
                    sig.mu_raw.reshape(1, -1),
                    self._mu_global, self._W_zca,
                ).ravel()
                sig.Sigma = self._W_zca @ sig.Sigma @ self._W_zca.T
                sig._eigvals = None; sig._eigvecs = None
                sig._Sinv = None; sig._par = None
                sig._metric = 'l2'

            n_d = self._n_pool / d
            print(f"  [SRT] ZCA fitted once: n_pool={self._n_pool}, d={d}, n/d={n_d:.2f}")

        # Whiten current task with the FIXED (already-fitted) ZCA
        mu_t_w = apply_whitening(
            mu_t.reshape(1, -1), self._mu_global, self._W_zca,
        ).ravel()
        Sigma_t_w = self._W_zca @ Sigma_t_shrunk @ self._W_zca.T

        # NOTE: NO Ledoit-Wolf here β€” space is already whitened.
        # Post-whitening LW with lambda-bar-I target would distort isotropy.

        sig = TaskSignature(
            task_id, mu_t_w, Sigma_t_w, n_t,
            metric='l2',
            Sigma_raw=Sigma_t_shrunk,
            h_train=h_train,
            mu_raw=mu_t,
        )
        self.signatures[task_id] = sig
        return sig

    # ── Route ────────────────────────────────────────────────────────

    def route(self, h: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
        """
        Route embeddings h to the nearest stored task.

        Args:
            h: (n, d) or (d,) β€” embeddings from frozen backbone

        Returns:
            task_ids: (n,) β€” predicted task ID for each embedding
            dists: (n, T) β€” distance to each task
        """
        if h.ndim == 1:
            h = h.reshape(1, -1)

        # [hard mode] Apply ZCA whitening to query embeddings
        if self.srt_metric_mode == 'hard' and self._W_zca is not None:
            h = apply_whitening(h, self._mu_global, self._W_zca)

        n = h.shape[0]
        task_list = sorted(self.signatures.keys())
        T = len(task_list)

        if T == 0:
            raise RuntimeError("SRT Router: no tasks registered.")

        dists = np.zeros((n, T), dtype=np.float64)
        for i, t_id in enumerate(task_list):
            dists[:, i] = self.signatures[t_id].distance(h)

        nearest_idx = np.argmin(dists, axis=1)
        nearest_task = np.array(task_list)[nearest_idx]

        # ── DEBUG ────────────────────────────────────────────────
        if not hasattr(self, '_route_debug_count'):
            self._route_debug_count = 0
        self._route_debug_count += 1
        if self._route_debug_count % 1000 == 0 and n > 0:
            print(f"[SRT-ROUTE] mode={self.srt_metric_mode} n_tasks={T} "
                  f"task_list={task_list} whiten={'Yes' if self._W_zca is not None else 'No'}")
            print(f"[SRT-ROUTE] Sample 0: dists={dists[0,:].tolist()} "
                  f"argmin={nearest_idx[0]} pred={nearest_task[0]}")
            if hasattr(self, '_srm_metrics'):
                print(f"[SRT-ROUTE] SRM metrics: {self._srm_metrics}")

        return nearest_task, dists

    # ── Save / Load ───────────────────────────────────────────────────

    def save(self, path: str):
        """Save all signatures to disk."""
        sigs_data = {k: v.to_dict() for k, v in self.signatures.items()}
        np.savez_compressed(
            path,
            signatures=sigs_data,
            mu_pool=self._mu_pool if self._mu_pool is not None else np.array([]),
            Sigma_pool=self._Sigma_pool if self._Sigma_pool is not None else np.array([]),
            n_pool=np.array([self._n_pool]),
            srt_metric_mode=self.srt_metric_mode,
            use_shrink=np.array([self.use_shrink]),
            shrink_factor=np.array([self.shrink_factor]),
            mu_global=self._mu_global if self._mu_global is not None else np.array([]),
            W_zca=self._W_zca if self._W_zca is not None else np.array([]),
            zca_fitted=np.array([self._zca_fitted]),
        )

    def load(self, path: str):
        """Load signatures from disk."""
        data = np.load(path, allow_pickle=True)
        self.srt_metric_mode = str(data.get('srt_metric_mode', 'hard'))

        # Restore pooled statistics (empty array = None sentinel)
        mu_p = data['mu_pool']
        self._mu_pool = mu_p if mu_p.size > 0 else None
        Sigma_p = data['Sigma_pool']
        self._Sigma_pool = Sigma_p if Sigma_p.size > 0 else None
        self._n_pool = int(data['n_pool'][0])

        # Restore shrink settings
        if 'use_shrink' in data:
            self.use_shrink = bool(data['use_shrink'][0])
        if 'shrink_factor' in data:
            self.shrink_factor = float(data['shrink_factor'][0])

        # ZCA whitening (empty array = None sentinel)
        mu_g = data.get('mu_global', np.array([]))
        W_z = data.get('W_zca', np.array([]))
        self._mu_global = mu_g if mu_g.size > 0 else None
        self._W_zca = W_z if W_z.size > 0 else None
        # ZCA was fitted once; restore flag (backward compat: infer from W if not saved)
        if 'zca_fitted' in data:
            self._zca_fitted = bool(data['zca_fitted'][0])
        else:
            self._zca_fitted = (self._mu_global is not None and self._W_zca is not None)

        for k, v in data['signatures'].item().items():
            self.signatures[k] = TaskSignature.from_dict(v)

    # ── Summary ───────────────────────────────────────────────────────

    def summary(self) -> dict:
        """Return routing statistics."""
        task_list = sorted(self.signatures.keys())
        metrics = [self.signatures[t].metric for t in task_list]
        pars = [self.signatures[t].par for t in task_list]
        return {
            'n_tasks': len(self.signatures),
            'task_ids': task_list,
            'metrics': metrics,
            'pars': [f"{p:.1f}" for p in pars],
            'avg_par': float(np.mean(pars)) if pars else 0.0,
            'pool_n': self._n_pool,
            'mode': self.srt_metric_mode,
        }