File size: 29,673 Bytes
8c74f19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Read-only integrity audit for a pooled-v2 compact dataset.

The audit is intentionally stricter than a reader smoke test.  For every
graph selected in a v2 dataset it proves four links:

1. the manifest/source index maps the dataset index to one unique discovered
   raw pose path;
2. the reconstructed graph node rows, atom-static columns, coordinates and
   coordinate-error labels match that raw protein/native/predicted pose;
3. PP edges and pose-specific non-PP edges are exactly the cutoff graph of
   those raw coordinates; and
4. when a matching v1 compact staging shard is supplied, all serialized graph
   tensors are bitwise equal for every common source graph index.

It only reads input datasets.  A JSON report is optional and is written only
to a path that does not already exist.
"""

from __future__ import annotations

import argparse
import json
import os
import sys
import time
from collections import Counter
from dataclasses import dataclass
from datetime import datetime, timezone
from pathlib import Path
from typing import Any, Dict, Iterable, List, Mapping, Sequence, Tuple

import numpy as np
import torch
from scipy.spatial.distance import cdist


THIS_DIR = Path(__file__).resolve().parent
SYSTEM_SPLIT = THIS_DIR.parent / "system_split_code"
for directory in (THIS_DIR, SYSTEM_SPLIT):
    if str(directory) not in sys.path:
        sys.path.insert(0, str(directory))

import build_pooled_v2 as v2  # noqa: E402
from compact_graph_dataset import CompactGraphDataset  # noqa: E402


@dataclass(frozen=True)
class Arguments:
    data_dir: Path
    v2_root: Path
    v1_shard: Path | None
    report: Path | None
    cutoff: float


def parse_args(argv: Sequence[str] | None = None) -> Arguments:
    parser = argparse.ArgumentParser(description=__doc__)
    parser.add_argument("--data-dir", required=True, type=Path)
    parser.add_argument("--v2-root", required=True, type=Path)
    parser.add_argument(
        "--v1-shard",
        type=Path,
        default=None,
        help="Optional old gnncp_compact_v1 shard containing the same source indices.",
    )
    parser.add_argument(
        "--report",
        type=Path,
        default=None,
        help="Optional new JSON report path; existing files are never overwritten.",
    )
    parser.add_argument("--cutoff", type=float, default=6.0)
    raw = parser.parse_args(argv)
    return Arguments(
        data_dir=raw.data_dir.expanduser().resolve(),
        v2_root=raw.v2_root.expanduser().resolve(),
        v1_shard=raw.v1_shard.expanduser().resolve() if raw.v1_shard else None,
        report=raw.report.expanduser().resolve() if raw.report else None,
        cutoff=float(raw.cutoff),
    )


def _atomic_json_new(path: Path, payload: Mapping[str, Any]) -> None:
    if path.exists():
        raise FileExistsError(f"refusing to overwrite audit report: {path}")
    path.parent.mkdir(parents=True, exist_ok=True)
    temporary = path.with_name(path.name + f".tmp.{os.getpid()}")
    with temporary.open("w", encoding="utf-8") as handle:
        json.dump(payload, handle, indent=2, ensure_ascii=False)
        handle.write("\n")
        handle.flush()
        os.fsync(handle.fileno())
    if path.exists():
        temporary.unlink(missing_ok=True)
        raise FileExistsError(f"audit report appeared concurrently: {path}")
    os.replace(temporary, path)


def _load_json(path: Path) -> Dict[str, Any]:
    with path.open("r", encoding="utf-8") as handle:
        value = json.load(handle)
    if not isinstance(value, dict):
        raise TypeError(f"expected object in {path}")
    return value


def _bounds(pointer: torch.Tensor, index: int, name: str) -> tuple[int, int]:
    begin = int(pointer[index].item())
    end = int(pointer[index + 1].item())
    if begin < 0 or end < begin:
        raise AssertionError(f"invalid {name}[{index}] = [{begin}, {end})")
    return begin, end


def _slice_system(shard: Mapping[str, torch.Tensor], system_index: int) -> Dict[str, torch.Tensor]:
    static_start, static_end = _bounds(shard["system_node_ptr"], system_index, "system_node_ptr")
    protein_start, protein_end = _bounds(shard["protein_ptr"], system_index, "protein_ptr")
    native_start, native_end = _bounds(
        shard["native_ligand_ptr"], system_index, "native_ligand_ptr"
    )
    pp_start, pp_end = _bounds(shard["pp_edge_ptr"], system_index, "pp_edge_ptr")
    return {
        "n_protein": shard["n_protein"][system_index : system_index + 1],
        "x_static": shard["x_static"][static_start:static_end],
        "protein_pos": shard["protein_pos"][protein_start:protein_end],
        "native_ligand_pos": shard["native_ligand_pos"][native_start:native_end],
        "pp_edge_upper": shard["pp_edge_upper"][:, pp_start:pp_end],
    }


def _slice_pose(
    shard: Mapping[str, torch.Tensor], pose_index: int
) -> Dict[str, torch.Tensor]:
    dynamic_start, dynamic_end = _bounds(shard["pose_node_ptr"], pose_index, "pose_node_ptr")
    ligand_start, ligand_end = _bounds(shard["pose_ligand_ptr"], pose_index, "pose_ligand_ptr")
    nonpp_start, nonpp_end = _bounds(shard["nonpp_edge_ptr"], pose_index, "nonpp_edge_ptr")
    return {
        "x_dynamic": shard["x_dynamic"][dynamic_start:dynamic_end],
        "ligand_pos": shard["ligand_pos"][ligand_start:ligand_end],
        "nonpp_edge_upper": shard["nonpp_edge_upper"][:, nonpp_start:nonpp_end],
    }


def _equal(name: str, actual: torch.Tensor, expected: torch.Tensor, context: str) -> None:
    if actual.dtype != expected.dtype or tuple(actual.shape) != tuple(expected.shape):
        raise AssertionError(
            f"{context}: {name} shape/dtype differs: "
            f"{actual.dtype}{tuple(actual.shape)} vs {expected.dtype}{tuple(expected.shape)}"
        )
    if not torch.equal(actual, expected):
        difference = (
            float((actual.to(torch.float64) - expected.to(torch.float64)).abs().max().item())
            if actual.numel() and actual.is_floating_point()
            else None
        )
        raise AssertionError(f"{context}: {name} differs; max_abs={difference}")


def _allclose(
    name: str,
    actual: torch.Tensor,
    expected: torch.Tensor,
    context: str,
    *,
    rtol: float = 1e-6,
    atol: float = 1e-6,
) -> float:
    """Return max absolute difference after a deliberately stated tolerance."""
    if actual.dtype != expected.dtype or tuple(actual.shape) != tuple(expected.shape):
        raise AssertionError(
            f"{context}: {name} shape/dtype differs: "
            f"{actual.dtype}{tuple(actual.shape)} vs {expected.dtype}{tuple(expected.shape)}"
        )
    difference = (
        float((actual.to(torch.float64) - expected.to(torch.float64)).abs().max().item())
        if actual.numel()
        else 0.0
    )
    if not torch.allclose(actual, expected, rtol=rtol, atol=atol):
        raise AssertionError(
            f"{context}: {name} differs beyond rtol={rtol}, atol={atol}; max_abs={difference}"
        )
    return difference


def _as_tensor(array: np.ndarray) -> torch.Tensor:
    return torch.from_numpy(np.ascontiguousarray(array))


def _expected_pp_upper(coords_protein: np.ndarray, cutoff: float) -> torch.Tensor:
    distance = cdist(coords_protein, coords_protein)
    mask = (distance <= cutoff) & (~np.eye(coords_protein.shape[0], dtype=bool))
    src, dst = np.where(mask)
    keep = src < dst
    return _as_tensor(np.vstack([src[keep], dst[keep]]).astype(np.int32))


def _expected_nonpp_upper(
    coords_protein: np.ndarray,
    coords_ligand: np.ndarray,
    cutoff: float,
) -> torch.Tensor:
    """Recreate the row-major upper-edge order used by the original builder."""
    n_protein = coords_protein.shape[0]
    protein_ligand = cdist(coords_protein, coords_ligand)
    protein_src, ligand_local = np.where(protein_ligand <= cutoff)
    protein_ligand_edges = np.vstack(
        [protein_src, n_protein + ligand_local]
    ).astype(np.int32)
    ligand_ligand = cdist(coords_ligand, coords_ligand)
    ligand_mask = (ligand_ligand <= cutoff) & (~np.eye(coords_ligand.shape[0], dtype=bool))
    ligand_src, ligand_dst = np.where(ligand_mask)
    keep = ligand_src < ligand_dst
    ligand_ligand_edges = np.vstack(
        [n_protein + ligand_src[keep], n_protein + ligand_dst[keep]]
    ).astype(np.int32)
    return _as_tensor(np.concatenate([protein_ligand_edges, ligand_ligand_edges], axis=1))


def _full_static(protein_atoms: Any, ligand_atoms: Any) -> np.ndarray:
    protein_static, _ = v2._static_features(protein_atoms, protein=True)
    ligand_static, _ = v2._static_features(ligand_atoms, protein=False)
    return np.concatenate([protein_static, ligand_static], axis=0)


def _discover_expected(data_dir: Path, method: str, n_source_systems: int) -> List[v2.SystemSpec]:
    config = v2.BuildConfig(
        data_dir=data_dir,
        output_dir=Path("/tmp/unused_audit_output"),
        method=method,
        cutoff=6.0,
        target_shard_mib=1,
        system_workers=1,
        max_systems=n_source_systems,
        max_poses_per_system=None,
        include_systems=(),
        on_error="abort",
        verify_reference=False,
        verify_reference_systems=1,
        verify_reference_poses=1,
        reader_smoke_graphs=0,
    )
    return v2.discover_systems(config)


def _tensor_storage_bytes(shard: Mapping[str, torch.Tensor]) -> int:
    return sum(value.numel() * value.element_size() for value in shard.values() if torch.is_tensor(value))


def _v2_shards(root: Path, manifest: Mapping[str, Any]) -> List[Mapping[str, torch.Tensor]]:
    result = []
    for entry in manifest["shards"]:
        path = root / str(entry["path"])
        result.append(torch.load(path, map_location="cpu", mmap=True, weights_only=True))
    return result


def _check_v2_index_and_manifest(
    root: Path,
    manifest: Mapping[str, Any],
    source_index: Mapping[str, Any],
    expected_by_source: Mapping[int, v2.PoseSpec],
) -> tuple[Dict[int, tuple[int, int, int]], Dict[str, int]]:
    """Return source -> (shard, local pose, local storage system)."""
    shards = _v2_shards(root, manifest)
    source_locations: Dict[int, tuple[int, int, int]] = {}
    pointer_systems_checked = 0
    records_checked = 0
    for shard_index, (entry, shard) in enumerate(zip(manifest["shards"], shards)):
        n_poses = int(shard["pose_system"].numel())
        if n_poses != int(entry["num_graphs"]):
            raise AssertionError(f"shard {shard_index}: manifest pose count differs")
        if int(shard["source_graph_index"].numel()) != n_poses:
            raise AssertionError(f"shard {shard_index}: source_graph_index length differs")
        if int(shard["n_protein"].numel()) != int(entry["num_systems"]):
            raise AssertionError(f"shard {shard_index}: manifest storage-system count differs")
        for local_pose, source_value in enumerate(shard["source_graph_index"].tolist()):
            source = int(source_value)
            if source in source_locations:
                raise AssertionError(f"source graph index appears twice: {source}")
            if source not in expected_by_source:
                raise AssertionError(f"unexpected source graph index in v2: {source}")
            storage_system = int(shard["pose_system"][local_pose].item())
            source_locations[source] = (shard_index, local_pose, storage_system)
        for storage_system, record in enumerate(entry["systems"]):
            pose_start, pose_end = _bounds(
                shard["system_graph_ptr"], storage_system, "system_graph_ptr"
            )
            actual_sources = [
                int(value)
                for value in shard["source_graph_index"][pose_start:pose_end].tolist()
            ]
            declared_sources = [int(value) for value in record["source_graph_indices"]]
            if actual_sources != declared_sources:
                raise AssertionError(
                    f"shard {shard_index} storage system {storage_system}: "
                    "manifest source indices differ from tensor pointers"
                )
            if int(record["num_graphs"]) != pose_end - pose_start:
                raise AssertionError(f"storage group graph count mismatch for {record['system_id']}")
            expected_poses = [expected_by_source[source] for source in actual_sources]
            expected_systems = {pose.system_id for pose in expected_poses}
            if expected_systems != {str(record["source_system_id"])}:
                raise AssertionError(f"storage group source system mismatch for {record['system_id']}")
            expected_paths = [
                str(pose.ligand_pred.relative_to(next(iter(expected_poses)).ligand_pred.parents[1]))
                for pose in expected_poses
            ]
            # Dataset paths are relative to the data root, rather than their
            # immediate system directory.  Recompute below with a stable root.
            del expected_paths
            pointer_systems_checked += 1
            records_checked += 1

    declared_sources = [int(value) for value in source_index["source_graph_indices"]]
    if len(declared_sources) != len(set(declared_sources)):
        raise AssertionError("source_index has duplicate source graph indices")
    if set(declared_sources) != set(source_locations):
        missing = sorted(set(declared_sources).symmetric_difference(source_locations))[:10]
        raise AssertionError(f"source_index/tensor source set mismatch: {missing}")
    graph_map = manifest["graph_map"]
    if len(graph_map) != len(declared_sources):
        raise AssertionError("manifest graph_map length differs from source_index")
    for dataset_index, source in enumerate(declared_sources):
        location = [int(value) for value in graph_map[dataset_index]]
        actual = source_locations[source]
        if location != list(actual[:2]):
            raise AssertionError(
                f"dataset index {dataset_index}: graph_map {location} != source location {actual[:2]}"
            )
        expected = expected_by_source[source]
        if str(source_index["graph_to_system"][dataset_index]) != expected.system_id:
            raise AssertionError(f"dataset index {dataset_index}: graph_to_system mismatch")
    return source_locations, {"storage_systems": pointer_systems_checked, "records": records_checked}


def _check_declared_pose_paths(
    data_dir: Path,
    manifest: Mapping[str, Any],
    expected_by_source: Mapping[int, v2.PoseSpec],
) -> int:
    checked = 0
    for shard_entry in manifest["shards"]:
        for record in shard_entry["systems"]:
            declared_sources = [int(value) for value in record["source_graph_indices"]]
            declared_paths = [str(value) for value in record["source_pose_paths"]]
            expected_paths = [
                str(expected_by_source[source].ligand_pred.relative_to(data_dir))
                for source in declared_sources
            ]
            if declared_paths != expected_paths:
                raise AssertionError(
                    f"{record['system_id']}: declared pose paths do not match source graph indices"
                )
            checked += len(declared_sources)
    return checked


def _check_raw_and_edges(
    arguments: Arguments,
    manifest: Mapping[str, Any],
    source_index: Mapping[str, Any],
    expected_by_source: Mapping[int, v2.PoseSpec],
    source_locations: Mapping[int, tuple[int, int, int]],
) -> Dict[str, int]:
    """Verify every row/label against the raw pose and every stored upper edge."""
    dataset = CompactGraphDataset(arguments.v2_root, strict=True)
    shards = _v2_shards(arguments.v2_root, manifest)
    protein_cache: Dict[str, tuple[np.ndarray, np.ndarray, Any, Any]] = {}
    checked_pp_systems: set[tuple[int, int]] = set()
    counts = Counter()
    static_index = torch.tensor(v2.STATIC_COLUMNS, dtype=torch.int64)

    for dataset_index, source_value in enumerate(source_index["source_graph_indices"]):
        source = int(source_value)
        pose = expected_by_source[source]
        shard_index, local_pose, storage_system = source_locations[source]
        shard = shards[shard_index]
        graph = dataset[dataset_index]
        if int(shard["source_graph_index"][local_pose].item()) != source:
            raise AssertionError(f"dataset index {dataset_index}: source graph index changed")
        if int(shard["pose_system"][local_pose].item()) != storage_system:
            raise AssertionError(f"dataset index {dataset_index}: storage system pointer changed")

        if pose.system_id not in protein_cache:
            protein_universe = v2.load_pdb_clean_models(str(pose.protein))
            native_universe = v2.load_pdb_clean_models(str(pose.ligand_native))
            protein_atoms = protein_universe.select_atoms("not name H*")
            native_atoms = native_universe.select_atoms("not name H*")
            protein_cache[pose.system_id] = (
                protein_atoms.positions.astype(np.float32),
                native_atoms.positions.astype(np.float32),
                protein_atoms,
                native_atoms,
            )
        coords_protein, coords_native, protein_atoms, _ = protein_cache[pose.system_id]
        ligand_universe = v2.load_pdb_clean_models(str(pose.ligand_pred))
        ligand_atoms = ligand_universe.select_atoms("not name H*")
        coords_ligand = ligand_atoms.positions.astype(np.float32)
        if coords_ligand.shape[0] != coords_native.shape[0]:
            raise AssertionError(f"{pose.ligand_pred}: raw pred/native ligand atom count differs")
        expected_pos = _as_tensor(np.vstack([coords_protein, coords_ligand]))
        expected_y_grt = _as_tensor(np.vstack([coords_protein, coords_native]))
        expected_static = _as_tensor(_full_static(protein_atoms, ligand_atoms))
        # The original graph builder uses NumPy norm, whereas
        # CompactGraphDataset deliberately reconstructs y_true with Torch from
        # stored float32 coordinates.  These are mathematically identical but
        # can differ by one float32 ULP.  Verify the reader formula bitwise and
        # independently verify legacy/raw semantics within one ULP tolerance.
        expected_error_legacy = _as_tensor(
            np.concatenate(
                [
                    np.zeros(coords_protein.shape[0], dtype=np.float32),
                    np.linalg.norm(coords_ligand - coords_native, axis=1).astype(np.float32),
                ]
            )
        ).unsqueeze(-1)
        context = f"source={source} dataset={dataset_index} pose={pose.ligand_pred.name}"
        _equal("raw pos", graph.pos, expected_pos, context)
        _equal("raw y_pred", graph.y_pred, expected_pos, context)
        _equal("raw y_grt", graph.y_grt, expected_y_grt, context)
        expected_error_reader = torch.zeros_like(graph.y_true)
        ligand_delta = expected_pos[coords_protein.shape[0] :] - expected_y_grt[
            coords_protein.shape[0] :
        ]
        expected_error_reader[coords_protein.shape[0] :, 0] = torch.sqrt(
            torch.sum(ligand_delta * ligand_delta, dim=1)
        )
        _equal("reader-reconstructed y_true", graph.y_true, expected_error_reader, context)
        raw_y_true_difference = _allclose(
            "raw legacy y_true",
            graph.y_true,
            expected_error_legacy,
            context,
            rtol=1e-6,
            atol=1e-6,
        )
        counts["max_raw_y_true_abs"] = max(
            raw_y_true_difference,
            float(counts.get("max_raw_y_true_abs", 0.0)),
        )
        _equal("raw static atom features", graph.x.index_select(1, static_index), expected_static, context)
        expected_is_protein = torch.zeros((expected_pos.shape[0], 1), dtype=torch.float32)
        expected_is_protein[: coords_protein.shape[0]] = 1.0
        _equal("protein/ligand node partition", graph.is_protein, expected_is_protein, context)

        # The reader emits exactly one reverse edge per stored upper edge.
        stored_pose = _slice_pose(shard, local_pose)
        stored_system = _slice_system(shard, storage_system)
        expected_edge_count = 2 * (
            stored_system["pp_edge_upper"].shape[1]
            + stored_pose["nonpp_edge_upper"].shape[1]
        )
        if int(graph.edge_index.shape[1]) != expected_edge_count:
            raise AssertionError(f"{context}: reconstructed edge count differs from compact storage")

        system_key = (shard_index, storage_system)
        if system_key not in checked_pp_systems:
            _equal(
                "raw PP upper edges",
                stored_system["pp_edge_upper"],
                _expected_pp_upper(coords_protein, arguments.cutoff),
                context,
            )
            checked_pp_systems.add(system_key)
            counts["pp_systems"] += 1
        _equal(
            "raw non-PP upper edges",
            stored_pose["nonpp_edge_upper"],
            _expected_nonpp_upper(coords_protein, coords_ligand, arguments.cutoff),
            context,
        )

        # Independently verify every reconstructed edge attribute from the raw
        # coordinate rows, including both directed orientations.
        src, dst = graph.edge_index
        distance = torch.sqrt(
            torch.sum(
                (expected_pos[src].to(torch.float64) - expected_pos[dst].to(torch.float64)) ** 2,
                dim=1,
            )
        )
        expected_attr = torch.stack(
            [
                (distance / arguments.cutoff).to(torch.float32),
                torch.exp(-distance / 3.0).to(torch.float32),
                (src < coords_protein.shape[0]).to(torch.float32),
                (dst < coords_protein.shape[0]).to(torch.float32),
            ],
            dim=1,
        )
        _equal("reconstructed edge attributes", graph.edge_attr, expected_attr, context)
        counts["raw_poses"] += 1
        if counts["raw_poses"] % 50 == 0:
            print(f"[raw] checked {counts['raw_poses']}/{len(dataset)} poses", flush=True)
    return dict(counts)


def _check_v1_tensor_parity(
    v1_shard_path: Path,
    v2_root: Path,
    manifest: Mapping[str, Any],
    source_locations: Mapping[int, tuple[int, int, int]],
) -> Dict[str, int]:
    """Bitwise-compare every v2 source pose against its same-source v1 record."""
    v1 = torch.load(v1_shard_path, map_location="cpu", mmap=True, weights_only=True)
    v2_shards = _v2_shards(v2_root, manifest)
    v1_locations: Dict[int, tuple[int, int]] = {}
    for local_pose, source_value in enumerate(v1["source_graph_index"].tolist()):
        source = int(source_value)
        if source in v1_locations:
            raise AssertionError(f"v1 shard has duplicate source graph index {source}")
        v1_locations[source] = (local_pose, int(v1["pose_system"][local_pose].item()))
    missing = sorted(set(source_locations).difference(v1_locations))
    if missing:
        raise AssertionError(f"v1 shard lacks v2 source graph indices; first: {missing[:10]}")

    compared_system_pairs: set[tuple[int, int]] = set()
    compared_poses = 0
    for source in sorted(source_locations):
        shard_index, v2_pose_index, v2_system_index = source_locations[source]
        v1_pose_index, v1_system_index = v1_locations[source]
        pair = (v1_system_index, v2_system_index)
        if pair not in compared_system_pairs:
            left = _slice_system(v1, v1_system_index)
            right = _slice_system(v2_shards[shard_index], v2_system_index)
            context = f"source={source} v1system={v1_system_index} v2system={v2_system_index}"
            for name in ("n_protein", "x_static", "protein_pos", "native_ligand_pos", "pp_edge_upper"):
                _equal(f"v1/v2 {name}", right[name], left[name], context)
            compared_system_pairs.add(pair)
        left_pose = _slice_pose(v1, v1_pose_index)
        right_pose = _slice_pose(v2_shards[shard_index], v2_pose_index)
        context = f"source={source} v1pose={v1_pose_index} v2pose={v2_pose_index}"
        for name in ("x_dynamic", "ligand_pos", "nonpp_edge_upper"):
            _equal(f"v1/v2 {name}", right_pose[name], left_pose[name], context)
        compared_poses += 1
    return {
        "v1_v2_storage_system_pairs": len(compared_system_pairs),
        "v1_v2_poses_bitwise_compared": compared_poses,
        "v1_v2_v1_shard_source_graphs": len(v1_locations),
    }


def run(arguments: Arguments) -> Dict[str, Any]:
    if not arguments.data_dir.is_dir():
        raise FileNotFoundError(arguments.data_dir)
    if not arguments.v2_root.is_dir():
        raise FileNotFoundError(arguments.v2_root)
    if arguments.v1_shard is not None and not arguments.v1_shard.is_file():
        raise FileNotFoundError(arguments.v1_shard)
    if arguments.report is not None and arguments.report.exists():
        raise FileExistsError(arguments.report)

    started = time.perf_counter()
    manifest = _load_json(arguments.v2_root / "manifest.json")
    source_index = _load_json(arguments.v2_root / "source_index.json")
    if manifest.get("format") != "gnncp_compact_v1" or int(manifest.get("schema_version", -1)) != 1:
        raise AssertionError("v2 output is not compact_v1-reader compatible")
    if manifest.get("status") != "complete":
        raise AssertionError("v2 manifest is not complete")
    expected_systems = _discover_expected(
        arguments.data_dir,
        str(manifest["method"]),
        int(manifest["source"]["discovered_source_systems"]),
    )
    expected_poses = [pose for system in expected_systems for pose in system.poses]
    expected_by_source = {pose.source_graph_index: pose for pose in expected_poses}
    if len(expected_by_source) != len(expected_poses):
        raise AssertionError("source discovery unexpectedly assigned duplicate indices")
    if int(manifest["n_graphs"]) != len(expected_poses):
        raise AssertionError(
            f"manifest graphs={manifest['n_graphs']} vs raw discovery={len(expected_poses)}"
        )
    if int(source_index["n_graphs"]) != len(expected_poses):
        raise AssertionError("source_index graph count differs from raw discovery")
    if list(source_index["graph_to_system"]) != [pose.system_id for pose in expected_poses]:
        raise AssertionError("source_index graph_to_system differs from raw discovery ordering")

    print(f"[index] auditing {len(expected_systems)} systems / {len(expected_poses)} poses", flush=True)
    source_locations, index_counts = _check_v2_index_and_manifest(
        arguments.v2_root, manifest, source_index, expected_by_source
    )
    declared_paths_checked = _check_declared_pose_paths(
        arguments.data_dir, manifest, expected_by_source
    )
    print("[raw] verifying raw node rows, labels, static atoms, and cutoff edges", flush=True)
    raw_counts = _check_raw_and_edges(
        arguments, manifest, source_index, expected_by_source, source_locations
    )
    parity_counts: Dict[str, int] = {}
    if arguments.v1_shard is not None:
        print("[v1] bitwise-comparing all common compact tensors", flush=True)
        parity_counts = _check_v1_tensor_parity(
            arguments.v1_shard, arguments.v2_root, manifest, source_locations
        )
    v2_shards = _v2_shards(arguments.v2_root, manifest)
    tensor_bytes = sum(_tensor_storage_bytes(shard) for shard in v2_shards)
    report: Dict[str, Any] = {
        "status": "passed",
        "created_utc": datetime.now(timezone.utc).isoformat(),
        "elapsed_seconds": time.perf_counter() - started,
        "inputs": {
            "data_dir": str(arguments.data_dir),
            "v2_root": str(arguments.v2_root),
            "v1_shard": str(arguments.v1_shard) if arguments.v1_shard else None,
            "cutoff": arguments.cutoff,
        },
        "counts": {
            "raw_discovered_systems": len(expected_systems),
            "raw_discovered_poses": len(expected_poses),
            "manifest_graphs": int(manifest["n_graphs"]),
            "manifest_storage_groups": int(manifest["n_systems"]),
            "declared_pose_paths_checked": declared_paths_checked,
            "v2_tensor_storage_bytes": tensor_bytes,
            **index_counts,
            **raw_counts,
            **parity_counts,
        },
        "guarantees_checked": [
            "unique source_graph_index and graph_map placement",
            "source-index order and source-system labels against deterministic raw discovery",
            "manifest pose paths against raw discovered pose paths",
            (
                "all raw heavy-atom node coordinates, atom-static features, partition flags, "
                "and y_pred/y_grt; y_true exact under the reader's float32 formula and "
                "within 1e-6 of the legacy NumPy formula"
            ),
            "all PP and all pose-specific non-PP cutoff upper edges against raw coordinates",
            "all reconstructed edge attributes against raw coordinate rows",
            "all common v1/v2 serialized static, dynamic, coordinate, and edge tensors bitwise equal",
        ],
    }
    if arguments.report is not None:
        _atomic_json_new(arguments.report, report)
        print(f"[report] wrote {arguments.report}", flush=True)
    print(
        f"PASS: {len(expected_poses)} poses, {raw_counts['pp_systems']} storage systems, "
        f"{raw_counts['raw_poses']} raw node/edge checks in {report['elapsed_seconds']:.1f}s",
        flush=True,
    )
    return report


def main(argv: Sequence[str] | None = None) -> int:
    run(parse_args(argv))
    return 0


if __name__ == "__main__":
    raise SystemExit(main())