aprender-core 0.60.0

Next-generation machine learning library in pure Rust
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
use super::*;

#[test]
fn tcov_conversion_path_display_with_intermediates() {
    let path = ConversionPath::chain(
        FormatType::Gguf,
        vec![FormatType::Apr],
        FormatType::SafeTensors,
    );
    let display = format!("{path}");
    assert_eq!(display, "GGUF → APR → SafeTensors");
}

#[test]
fn tcov_verify_roundtrip_safetensors() {
    let source = create_safetensors_fixture();
    let rosetta = RosettaStone::new();
    let report = rosetta.verify_roundtrip(&source, FormatType::Apr);
    // Round-trip may succeed or fail based on implementation details
    // The key is that it doesn't panic
    match report {
        Ok(vr) => {
            // If it succeeds, verify report structure
            assert!(vr.max_diff >= 0.0);
        }
        Err(_) => {
            // Also acceptable - some round-trip failures expected
        }
    }
    let _ = std::fs::remove_file(source);
}

// ========================================================================
// Section 23: Coverage Gap Tests (P050+)
// Targets uncovered branches in mod.rs identified by coverage analysis.
// ========================================================================

// ========================================================================
// P050-P059: FormatType::from_extension error paths (lines 90-116)
// ========================================================================

/// P050: from_extension with a real directory path (is_dir = true, no model files found)
/// Covers: line 90 (is_dir check), lines 93-103 (directory with no candidates)
#[test]
fn p050_from_extension_real_directory_no_candidates() {
    // /tmp is a real directory that won't contain model.gguf, model.apr, model.safetensors
    let path = Path::new("/tmp");
    let result = FormatType::from_extension(path);
    assert!(result.is_err(), "Directory path should fail from_extension");
    let err = format!("{:?}", result.unwrap_err());
    assert!(
        err.contains("is a directory"),
        "Error should mention directory: {err}"
    );
    assert!(
        err.contains(".gguf") || err.contains(".apr") || err.contains(".safetensors"),
        "Error should mention expected extensions: {err}"
    );
}

/// P051b: from_extension with a nonexistent path that has no extension
/// Covers: lines 112-116 (not a directory, no extension -> "No file extension found")
#[test]
fn p051b_from_extension_nonexistent_no_extension() {
    let path = Path::new("/nonexistent_dir_12345/no_extension_file");
    let result = FormatType::from_extension(path);
    assert!(
        result.is_err(),
        "Nonexistent path with no extension should fail"
    );
    let err = format!("{:?}", result.unwrap_err());
    assert!(
        err.contains("No file extension"),
        "Error should mention missing extension: {err}"
    );
}

/// P052b: from_extension with a directory containing a candidate model file
/// Covers: lines 104-109 (directory with found candidates -> "Did you mean" suggestion)
#[test]
fn p052b_from_extension_directory_with_candidate() {
    // Create a temp directory with a model file inside
    let temp_dir = std::env::temp_dir().join("rosetta_test_dir_p052b");
    let _ = std::fs::create_dir_all(&temp_dir);
    let model_file = temp_dir.join("model.gguf");
    let _ = std::fs::write(&model_file, b"dummy");

    let result = FormatType::from_extension(&temp_dir);
    assert!(
        result.is_err(),
        "Directory path should fail even with candidates"
    );
    let err = format!("{:?}", result.unwrap_err());
    assert!(
        err.contains("Did you mean"),
        "Error should suggest candidate file: {err}"
    );
    assert!(
        err.contains("model.gguf"),
        "Suggestion should include model.gguf: {err}"
    );

    let _ = std::fs::remove_dir_all(temp_dir);
}

/// P053b: from_extension with a directory containing multiple candidate model files
/// The code returns the first found candidate from `[model.gguf, model.apr, model.safetensors]`
#[test]
fn p053b_from_extension_directory_with_multiple_candidates() {
    let temp_dir = std::env::temp_dir().join("rosetta_test_dir_p053b");
    let _ = std::fs::create_dir_all(&temp_dir);
    let _ = std::fs::write(temp_dir.join("model.apr"), b"dummy");
    let _ = std::fs::write(temp_dir.join("model.safetensors"), b"dummy");

    let result = FormatType::from_extension(&temp_dir);
    assert!(result.is_err());
    let err = format!("{:?}", result.unwrap_err());
    assert!(
        err.contains("Did you mean"),
        "Should suggest a candidate: {err}"
    );

    let _ = std::fs::remove_dir_all(temp_dir);
}

/// P054b: from_extension with a nonexistent directory (is_dir returns false)
/// Covers: the else branch at line 112 when path doesn't exist and has no extension
#[test]
fn p054b_from_extension_nonexistent_directory_path() {
    let path = Path::new("/nonexistent_dir_12345");
    let result = FormatType::from_extension(path);
    assert!(result.is_err());
    let err = format!("{:?}", result.unwrap_err());
    // is_dir() returns false for nonexistent paths, so we get "No file extension"
    assert!(
        err.contains("No file extension"),
        "Nonexistent path without extension: {err}"
    );
}

// ========================================================================
// P060-P069: compute_tensor_validation thorough testing (lines 1036-1176)
// Tests the core validation logic directly via RosettaStone instance
// ========================================================================

/// P060: compute_tensor_validation with empty data
/// Covers: lines 1038-1052 (early return for empty data)
#[test]
fn p060_compute_validation_empty_data() {
    let rosetta = RosettaStone::new();
    let tv = rosetta.compute_tensor_validation("empty.weight", &[]);
    assert!(tv.is_valid, "Empty tensor should be valid");
    assert_eq!(tv.element_count, 0);
    assert_eq!(tv.nan_count, 0);
    assert_eq!(tv.inf_count, 0);
    assert_eq!(tv.zero_count, 0);
    assert_eq!(tv.min, 0.0);
    assert_eq!(tv.max, 0.0);
    assert_eq!(tv.mean, 0.0);
    assert_eq!(tv.std, 0.0);
    assert!(tv.failures.is_empty());
}

/// P061b: compute_tensor_validation with all NaN values
/// Covers: lines 1062-1064 (NaN counting), 1082-1087 (valid_count=0 path), 1105-1108
#[test]
fn p061b_compute_validation_all_nan() {
    let rosetta = RosettaStone::new();
    let data = [f32::NAN, f32::NAN, f32::NAN, f32::NAN];
    let tv = rosetta.compute_tensor_validation("corrupted.weight", &data);
    assert!(!tv.is_valid, "All-NaN tensor should be invalid");
    assert_eq!(tv.nan_count, 4);
    assert_eq!(tv.element_count, 4);
    assert_eq!(tv.mean, 0.0, "Mean should be 0 when no valid values");
    assert_eq!(tv.std, 0.0, "Std should be 0 when no valid values");
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("NaN") && f.contains("F-DATA-QUALITY-002")),
        "Should report NaN failure"
    );
}

/// P062b: compute_tensor_validation with all Inf values
/// Covers: lines 1066-1068 (Inf counting), 1110-1113
#[test]
fn p062b_compute_validation_all_inf() {
    let rosetta = RosettaStone::new();
    let data = [f32::INFINITY, f32::NEG_INFINITY, f32::INFINITY];
    let tv = rosetta.compute_tensor_validation("overflow.weight", &data);
    assert!(!tv.is_valid, "All-Inf tensor should be invalid");
    assert_eq!(tv.inf_count, 3);
    assert_eq!(tv.nan_count, 0);
    assert_eq!(tv.mean, 0.0, "Mean should be 0 when no valid values");
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("Inf") && f.contains("F-DATA-QUALITY-002")),
        "Should report Inf failure"
    );
}

/// P063b: compute_tensor_validation with mixed NaN and Inf
/// Covers: both NaN and Inf branches simultaneously, valid_count = 0
#[test]
fn p063b_compute_validation_mixed_nan_inf() {
    let rosetta = RosettaStone::new();
    let data = [f32::NAN, f32::INFINITY, f32::NAN, f32::NEG_INFINITY];
    let tv = rosetta.compute_tensor_validation("broken.weight", &data);
    assert!(!tv.is_valid);
    assert_eq!(tv.nan_count, 2);
    assert_eq!(tv.inf_count, 2);
    assert!(
        tv.failures.len() >= 2,
        "Should have both NaN and Inf failures"
    );
}

/// P064: compute_tensor_validation with all zeros
/// Covers: lines 1070-1072 (zero counting), 1115-1117 (all zeros failure)
#[test]
fn p064_compute_validation_all_zeros() {
    let rosetta = RosettaStone::new();
    let data = [0.0_f32; 100];
    let tv = rosetta.compute_tensor_validation("uninitialized.weight", &data);
    assert!(!tv.is_valid, "All-zero tensor should be invalid");
    assert_eq!(tv.zero_count, 100);
    assert_eq!(tv.element_count, 100);
    assert!(tv.is_all_zeros());
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("All values are zero") && f.contains("F-DATA-QUALITY-001")),
        "Should report all-zeros failure: {:?}",
        tv.failures
    );
}

/// P065: compute_tensor_validation density gate for embedding tensors (>50% zeros)
/// Covers: lines 1119-1134 (density gate with embedding name threshold=50%)
#[test]
fn p065_compute_validation_embedding_high_zeros() {
    let rosetta = RosettaStone::new();
    // 60% zeros in an embedding tensor (threshold is 50%)
    let mut data = vec![0.0_f32; 60];
    data.extend(vec![0.1_f32; 40]);
    let tv = rosetta.compute_tensor_validation("model.embed_tokens.weight", &data);
    assert!(!tv.is_valid, "Embedding with >50% zeros should be invalid");
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("DENSITY") && f.contains("F-DATA-QUALITY-001")),
        "Should report density failure: {:?}",
        tv.failures
    );
}

/// P066: compute_tensor_validation density gate for weight tensors (>80% zeros)
/// Covers: lines 1129 (non-embedding path, threshold=80%)
#[test]
fn p066_compute_validation_weight_high_zeros() {
    let rosetta = RosettaStone::new();
    // 85% zeros in a non-embedding weight tensor (threshold is 80%)
    let mut data = vec![0.0_f32; 85];
    data.extend(vec![0.5_f32; 15]);
    let tv = rosetta.compute_tensor_validation("model.layers.0.self_attn.q_proj.weight", &data);
    assert!(!tv.is_valid, "Weight with >80% zeros should be invalid");
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("DENSITY") && f.contains("80")),
        "Should report 80% density threshold: {:?}",
        tv.failures
    );
}

/// P067: compute_tensor_validation density gate for weight under threshold
/// Covers: density gate path where zero_pct <= threshold (no failure)
#[test]
fn p067_compute_validation_weight_acceptable_zeros() {
    let rosetta = RosettaStone::new();
    // 50% zeros in a non-embedding weight tensor (threshold is 80%) -- should pass density
    let mut data = vec![0.0_f32; 50];
    data.extend(vec![0.5_f32; 50]);
    let tv = rosetta.compute_tensor_validation("model.layers.0.weight", &data);
    assert!(
        !tv.failures.iter().any(|f| f.contains("DENSITY")),
        "50% zeros in weight should not trigger 80% density gate: {:?}",
        tv.failures
    );
}

/// P068: compute_tensor_validation L2 norm gate (near-zero L2)
/// Covers: lines 1136-1149 (L2 norm < 1e-6)
#[test]
fn p068_compute_validation_low_l2_norm() {
    let rosetta = RosettaStone::new();
    // Very small values that produce L2 norm < 1e-6
    let data = [1e-7_f32, -1e-7, 1e-7, -1e-7];
    let tv = rosetta.compute_tensor_validation("tiny.weight", &data);
    assert!(!tv.is_valid, "Near-zero L2 norm should be invalid");
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("L2 norm") && f.contains("F-DATA-QUALITY-003")),
        "Should report L2 norm failure: {:?}",
        tv.failures
    );
}

/// P069: compute_tensor_validation constant value gate (all identical, non-zero)
/// Covers: lines 1151-1159 (max - min < 1e-10 for non-norm tensors)
#[test]
fn p069_compute_validation_constant_values() {
    let rosetta = RosettaStone::new();
    let data = [0.5_f32; 100];
    let tv = rosetta.compute_tensor_validation("model.layers.0.weight", &data);
    assert!(!tv.is_valid, "Constant tensor should be invalid");
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("constant") && f.contains("F-DATA-QUALITY-003")),
        "Should report constant value failure: {:?}",
        tv.failures
    );
}

/// P070b: compute_tensor_validation constant value exemption for norm tensors
/// Covers: lines 1153-1155 (is_norm_or_bias exemption)
#[test]
fn p070b_compute_validation_constant_norm_exempt() {
    let rosetta = RosettaStone::new();
    // All 1.0 values in a norm tensor (e.g., RMS norm weight initialized to 1.0)
    let data = [1.0_f32; 100];
    let tv = rosetta.compute_tensor_validation("model.layers.0.input_layernorm.weight", &data);
    // Norm tensors are exempt from the constant-value check
    assert!(
        !tv.failures.iter().any(|f| f.contains("constant")),
        "Norm tensor should be exempt from constant check: {:?}",
        tv.failures
    );
}

/// P070c: compute_tensor_validation constant value exemption for bias tensors
/// Covers: lines 1153-1155 (is_norm_or_bias with "bias" in name)
#[test]
fn p070c_compute_validation_constant_bias_exempt() {
    let rosetta = RosettaStone::new();
    let data = [0.0_f32; 50]; // All zeros in a bias
                              // Even though all zeros, bias is exempt from constant check.
                              // However, all-zeros still triggers the all-zeros failure.
    let tv = rosetta.compute_tensor_validation("model.layers.0.bias", &data);
    assert!(
        !tv.failures.iter().any(|f| f.contains("constant")),
        "Bias tensor should be exempt from constant check: {:?}",
        tv.failures
    );
}

/// P070d: compute_tensor_validation constant value exemption for ln_ tensors
/// Covers: lines 1153-1155 (is_norm_or_bias with "ln_" in name)
#[test]
fn p070d_compute_validation_constant_ln_exempt() {
    let rosetta = RosettaStone::new();
    let data = [1.0_f32; 100];
    let tv = rosetta.compute_tensor_validation("model.ln_f.weight", &data);
    assert!(
        !tv.failures.iter().any(|f| f.contains("constant")),
        "ln_ tensor should be exempt from constant check: {:?}",
        tv.failures
    );
}

/// P070e: compute_tensor_validation with single valid element (valid_count=1, std=0)
/// Covers: line 1097-1101 (valid_count <= 1 path for std computation)
#[test]
fn p070e_compute_validation_single_element() {
    let rosetta = RosettaStone::new();
    let data = [0.5_f32];
    let tv = rosetta.compute_tensor_validation("single.weight", &data);
    assert_eq!(tv.element_count, 1);
    assert_eq!(tv.std, 0.0, "Std should be 0 for single element");
    assert!((tv.mean - 0.5).abs() < 1e-6);
    assert!((tv.min - 0.5).abs() < 1e-6);
    assert!((tv.max - 0.5).abs() < 1e-6);
}

/// P070f: compute_tensor_validation with NaN mixed with valid data
/// Covers: lines 1062-1064 NaN skip, 1082-1084 mean with valid_count > 0
/// Also covers line 1092 (NaN skip in variance computation)
#[test]
fn p070f_compute_validation_nan_mixed_with_valid() {
    let rosetta = RosettaStone::new();
    let data = [1.0_f32, f32::NAN, 3.0, f32::NAN, 5.0];
    let tv = rosetta.compute_tensor_validation("mixed.weight", &data);
    assert!(!tv.is_valid, "Mixed NaN tensor should be invalid");
    assert_eq!(tv.nan_count, 2);
    assert_eq!(tv.element_count, 5);
    // mean of valid values [1.0, 3.0, 5.0] = 3.0
    assert!(
        (tv.mean - 3.0).abs() < 1e-5,
        "Mean should be ~3.0, got {}",
        tv.mean
    );
    assert!((tv.min - 1.0).abs() < 1e-6);
    assert!((tv.max - 5.0).abs() < 1e-6);
    assert!(tv.std > 0.0, "Std should be non-zero for varied values");
}

/// P070g: compute_tensor_validation min/max clamping when only NaN/Inf present
/// Covers: lines 1170-1171 (min/max clamping when they remain INFINITY/NEG_INFINITY)
#[test]
fn p070g_compute_validation_minmax_clamped() {
    let rosetta = RosettaStone::new();
    // Only NaN and Inf, no finite values -> min stays INFINITY, max stays NEG_INFINITY
    let data = [f32::NAN, f32::INFINITY];
    let tv = rosetta.compute_tensor_validation("bad.weight", &data);
    // min was initialized to INFINITY (never updated), max to NEG_INFINITY
    // The code clamps: if min.is_infinite() { 0.0 } and if max.is_infinite() { 0.0 }
    assert_eq!(tv.min, 0.0, "min should be clamped to 0.0");
    assert_eq!(tv.max, 0.0, "max should be clamped to 0.0");
}

/// P070h: compute_tensor_validation with Inf mixed with valid data
/// Covers: Inf skip in main loop and variance loop
#[test]
fn p070h_compute_validation_inf_mixed_with_valid() {
    let rosetta = RosettaStone::new();
    let data = [2.0_f32, f32::INFINITY, 4.0, f32::NEG_INFINITY, 6.0];
    let tv = rosetta.compute_tensor_validation("mixed_inf.weight", &data);
    assert!(!tv.is_valid);
    assert_eq!(tv.inf_count, 2);
    // mean of valid values [2.0, 4.0, 6.0] = 4.0
    assert!(
        (tv.mean - 4.0).abs() < 1e-5,
        "Mean should be ~4.0, got {}",
        tv.mean
    );
    assert!((tv.min - 2.0).abs() < 1e-6);
    assert!((tv.max - 6.0).abs() < 1e-6);
}

// ========================================================================
// PMAT-889: F-DATA-QUALITY-007 — DEAD OUTPUT ROW
// ========================================================================
//
// Pillar-4 fail-closed BEAT. A weight tensor whose individual entries are all
// finite, whose whole-tensor zero density is well under the 50%/80% gate, and
// whose global L2 norm is large can STILL be semantically corrupt if one
// OUTPUT ROW is entirely zero. For an lm_head `[vocab, hidden]` that means one
// token's logit is structurally a constant (its dot-product against any hidden
// state is 0) -> that token can never be emitted with positive evidence (dead
// token). For an embedding `[vocab, hidden]` a zero row is a dead/unreachable
// token vector. The whole-tensor gates (F-DATA-QUALITY-001..005) all MISS this
// because a single zero row is a tiny fraction of the tensor. llama.cpp / Ollama
// load and run such a model silently. apr must fail closed.

/// Build a healthy `[out, in]` row-major weight: dense, finite, varied, L2 >> 0,
/// no exact zeros, mean ~ 0.
fn healthy_rowmajor(out: usize, in_dim: usize) -> Vec<f32> {
    (0..out * in_dim)
        .map(|i| {
            let x = ((i % 17) as f32) - 8.0;
            if x == 0.0 {
                0.5
            } else {
                x * 0.01
            }
        })
        .collect()
}

/// RED (PMAT-889): an lm_head `[256, 64]` that is healthy EXCEPT one fully-zero
/// output row must be REJECTED with rule id `F-DATA-QUALITY-007`. On current
/// main this fails — the whole-tensor gates do not scan rows, so the dead row
/// passes clean. The fix is the shape-aware `compute_tensor_validation_with_shape`.
#[test]
fn pmat889_red_dead_output_row_lm_head_rejected() {
    let rosetta = RosettaStone::new();
    let (out, in_dim) = (256usize, 64usize);
    let mut data = healthy_rowmajor(out, in_dim);
    // Force output row 7 fully zero (one dead vocab token).
    let dead = 7usize;
    for v in data.iter_mut().skip(dead * in_dim).take(in_dim) {
        *v = 0.0;
    }

    let tv = rosetta.compute_tensor_validation_with_shape("lm_head.weight", &data, &[out, in_dim]);
    assert!(
        !tv.is_valid,
        "FAIL-CLOSED VIOLATION: apr ACCEPTED an lm_head with a fully-zero output row \
         (dead token logit -> -inf). This is the garbage llama.cpp/Ollama ship silently. \
         failures={:?}",
        tv.failures
    );
    assert!(
        tv.failures
            .iter()
            .any(|f| f.contains("F-DATA-QUALITY-007")),
        "Dead output row must be reported as F-DATA-QUALITY-007: {:?}",
        tv.failures
    );
}

/// RED (PMAT-889): same dead-row gate for `embed_tokens.weight`.
#[test]
fn pmat889_red_dead_output_row_embed_rejected() {
    let rosetta = RosettaStone::new();
    let (vocab, hidden) = (200usize, 48usize);
    let mut data = healthy_rowmajor(vocab, hidden);
    let dead = 3usize;
    for v in data.iter_mut().skip(dead * hidden).take(hidden) {
        *v = 0.0;
    }
    let tv =
        rosetta.compute_tensor_validation_with_shape("model.embed_tokens.weight", &data, &[vocab, hidden]);
    assert!(
        !tv.is_valid,
        "FAIL-CLOSED VIOLATION: apr ACCEPTED an embedding with a dead token row. failures={:?}",
        tv.failures
    );
    assert!(
        tv.failures.iter().any(|f| f.contains("F-DATA-QUALITY-007")),
        "Dead embedding row must be F-DATA-QUALITY-007: {:?}",
        tv.failures
    );
}

/// FALSE-POSITIVE BOUND (PMAT-889): a fully healthy lm_head with NO zero rows
/// must still validate clean — fail-closed must not block legitimate models.
#[test]
fn pmat889_fp_bound_healthy_lm_head_accepted() {
    let rosetta = RosettaStone::new();
    let (out, in_dim) = (256usize, 64usize);
    let data = healthy_rowmajor(out, in_dim);
    let tv = rosetta.compute_tensor_validation_with_shape("lm_head.weight", &data, &[out, in_dim]);
    assert!(
        tv.is_valid,
        "FALSE POSITIVE: apr rejected a healthy lm_head (no dead rows). failures={:?}",
        tv.failures
    );
    assert!(
        !tv.failures.iter().any(|f| f.contains("F-DATA-QUALITY-007")),
        "Healthy lm_head must not trigger the dead-row gate: {:?}",
        tv.failures
    );
}

/// FALSE-POSITIVE BOUND (PMAT-889): a non-output-role tensor (e.g. a generic
/// intermediate buffer) with a zero row must NOT trip the dead-row gate — the
/// gate is intentionally scoped to lm_head/embed/output roles to avoid FPs on
/// roles where a structurally-zero row is not unambiguously corrupt.
#[test]
fn pmat889_fp_bound_non_output_role_zero_row_accepted_by_dead_row_gate() {
    let rosetta = RosettaStone::new();
    let (out, in_dim) = (64usize, 32usize);
    let mut data = healthy_rowmajor(out, in_dim);
    for v in data.iter_mut().skip(5 * in_dim).take(in_dim) {
        *v = 0.0;
    }
    // A name that does NOT match the output-projection role allow-list.
    let tv = rosetta.compute_tensor_validation_with_shape(
        "model.layers.0.mlp.intermediate_buffer",
        &data,
        &[out, in_dim],
    );
    assert!(
        !tv.failures.iter().any(|f| f.contains("F-DATA-QUALITY-007")),
        "Dead-row gate must be scoped to output roles, not fire on a generic tensor: {:?}",
        tv.failures
    );
}