khive-quant 0.12.0

SQ8 scalar quantization codecs (global-scale symmetric + per-dim affine) shared by HNSW and Vamana indexes.
Documentation
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
use super::*;

fn rand_vecs(n: usize, dims: usize, seed: u64) -> Vec<Vec<f32>> {
    let mut h = seed;
    (0..n)
        .map(|_| {
            (0..dims)
                .map(|_| {
                    h = h
                        .wrapping_mul(0x6c62_272e_07bb_0142)
                        .wrapping_add(0x62b8_2175_62d9_6b1a);
                    let bits = (h >> 33) as u32;
                    (bits as f32) / (u32::MAX as f32) * 2.0 - 1.0
                })
                .collect()
        })
        .collect()
}

fn dot_f32(a: &[f32], b: &[f32]) -> f32 {
    a.iter().zip(b).map(|(x, y)| x * y).sum()
}

fn l2_sq_f32(a: &[f32], b: &[f32]) -> f32 {
    a.iter().zip(b).map(|(x, y)| (x - y) * (x - y)).sum()
}

// ── QUANT-AUD-002: validation regression tests ──────────────────────────

#[test]
fn sq8_try_train_ragged_rows_returns_error_not_panic() {
    let vecs = vec![vec![0.0], vec![1.0, 2.0]];
    let err = Sq8Codec::try_train(&vecs).expect_err("ragged rows must be rejected");
    assert_eq!(
        err,
        QuantError::RaggedRow {
            row: 1,
            expected: 1,
            got: 2
        }
    );
}

#[test]
fn gs_try_train_ragged_rows_returns_error_not_panic() {
    let vecs = vec![vec![0.0], vec![1.0, 2.0]];
    let err = GsSq8Codec::try_train(&vecs).expect_err("ragged rows must be rejected");
    assert_eq!(
        err,
        QuantError::RaggedRow {
            row: 1,
            expected: 1,
            got: 2
        }
    );
}

#[test]
fn sq8_try_train_empty_corpus_returns_error() {
    let vecs: Vec<Vec<f32>> = vec![];
    assert_eq!(
        Sq8Codec::try_train(&vecs).unwrap_err(),
        QuantError::EmptyCorpus
    );
}

#[test]
fn sq8_try_train_flat_zero_dims_returns_error() {
    assert_eq!(
        Sq8Codec::try_train_flat(&[1.0, 2.0], 0).unwrap_err(),
        QuantError::ZeroDims
    );
}

#[test]
fn gs_try_train_flat_zero_dims_returns_error() {
    assert_eq!(
        GsSq8Codec::try_train_flat(&[1.0, 2.0], 0).unwrap_err(),
        QuantError::ZeroDims
    );
}

#[test]
fn sq8_try_train_flat_remainder_returns_error() {
    // 5 elements is not a multiple of dims=2.
    let err = Sq8Codec::try_train_flat(&[1.0, 2.0, 3.0, 4.0, 5.0], 2).unwrap_err();
    assert_eq!(err, QuantError::FlatLengthNotDivisible { len: 5, dims: 2 });
}

#[test]
fn gs_try_train_flat_remainder_returns_error() {
    let err = GsSq8Codec::try_train_flat(&[1.0, 2.0, 3.0, 4.0, 5.0], 2).unwrap_err();
    assert_eq!(err, QuantError::FlatLengthNotDivisible { len: 5, dims: 2 });
}

#[test]
fn sq8_try_encode_shorter_input_returns_error_not_malformed_vector() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let err = codec.try_encode(&[0.0, 1.0]).unwrap_err();
    assert_eq!(
        err,
        QuantError::EncodeLengthMismatch {
            expected: 4,
            got: 2
        }
    );
}

#[test]
fn sq8_try_encode_longer_input_returns_error() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let err = codec.try_encode(&[0.0, 1.0, 2.0, 3.0, 4.0]).unwrap_err();
    assert_eq!(
        err,
        QuantError::EncodeLengthMismatch {
            expected: 4,
            got: 5
        }
    );
}

#[test]
fn gs_try_encode_empty_input_returns_error_not_malformed_vector() {
    // QUANT-AUD-002: previously encode(&[]) on a trained (dims=1) codec
    // silently returned an EMPTY code vector in release builds (the
    // length check was debug_assert-only), which is_in_distribution
    // vacuously accepted and l2_sq scored as 0.0. Must now be a typed
    // error, never a malformed zero-length code vector.
    let codec = GsSq8Codec::train_flat(&[1.0, 2.0, 3.0, 4.0], 1);
    let err = codec.try_encode(&[]).unwrap_err();
    assert_eq!(
        err,
        QuantError::EncodeLengthMismatch {
            expected: 1,
            got: 0
        }
    );
}

#[test]
fn sq8_try_encode_flat_par_zero_dims_returns_error() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    assert_eq!(
        codec.try_encode_flat_par(&[1.0, 2.0], 0).unwrap_err(),
        QuantError::ZeroDims
    );
}

#[test]
fn gs_try_encode_flat_par_zero_dims_returns_error() {
    let codec = GsSq8Codec::train(&rand_vecs(10, 4, 1));
    assert_eq!(
        codec.try_encode_flat_par(&[1.0, 2.0], 0).unwrap_err(),
        QuantError::ZeroDims
    );
}

#[test]
fn sq8_try_encode_flat_par_remainder_returns_error() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let flat: Vec<f32> = (0..9).map(|i| i as f32).collect(); // 9 not a multiple of 4
    assert_eq!(
        codec.try_encode_flat_par(&flat, 4).unwrap_err(),
        QuantError::FlatLengthNotDivisible { len: 9, dims: 4 }
    );
}

#[test]
fn sq8_try_encode_flat_par_dims_mismatch_returns_error() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let flat: Vec<f32> = (0..6).map(|i| i as f32).collect();
    let err = codec.try_encode_flat_par(&flat, 3).unwrap_err();
    assert_eq!(
        err,
        QuantError::EncodeLengthMismatch {
            expected: 4,
            got: 3
        }
    );
}

#[test]
#[should_panic(expected = "row 1 has length 2, expected 1")]
fn sq8_train_still_panics_with_typed_message_on_ragged_rows() {
    // The panicking convenience wrapper is preserved for existing callers,
    // but must now surface the typed QuantError message rather than
    // panicking from a raw out-of-bounds index.
    let _ = Sq8Codec::train(&[vec![0.0], vec![1.0, 2.0]]);
}

#[test]
fn sq8_try_train_and_encode_roundtrip_matches_panicking_api() {
    let vecs = rand_vecs(20, 8, 7);
    let a = Sq8Codec::train(&vecs);
    let b = Sq8Codec::try_train(&vecs).expect("valid corpus must train");
    assert_eq!(a.min, b.min);
    assert_eq!(a.scale, b.scale);
    let ea = a.encode(&vecs[0]);
    let eb = b.try_encode(&vecs[0]).expect("valid vector must encode");
    assert_eq!(ea.codes, eb.codes);
}

// ── Sq8Codec tests ──────────────────────────────────────────────────────

#[test]
fn encode_decode_roundtrip_is_bounded() {
    let vecs = rand_vecs(100, 32, 42);
    let codec = Sq8Codec::train(&vecs);
    for v in &vecs {
        let ev = codec.encode(v);
        assert_eq!(ev.codes.len(), v.len());
        for (d, &code) in ev.codes.iter().enumerate() {
            let decoded = code as f32 * codec.scale[d] + codec.min[d];
            let err = (decoded - v[d]).abs();
            assert!(
                err <= codec.scale[d] + 1e-5,
                "dim {d}: err={err} scale={}",
                codec.scale[d]
            );
        }
    }
}

#[test]
fn approx_dot_relative_error_bounded() {
    let vecs = rand_vecs(200, 64, 77);
    let codec = Sq8Codec::train(&vecs);
    let encoded: Vec<EncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();

    let mut max_rel_err = 0.0f32;
    for i in 0..vecs.len() {
        for j in (i + 1)..vecs.len().min(i + 10) {
            let true_dot = dot_f32(&vecs[i], &vecs[j]);
            let approx = codec.approx_dot(&encoded[i], &encoded[j]);
            let denom = true_dot.abs().max(1e-3);
            let rel = (approx - true_dot).abs() / denom;
            if rel > max_rel_err {
                max_rel_err = rel;
            }
        }
    }
    assert!(
        max_rel_err < 0.15,
        "max relative dot error {max_rel_err:.4} >= 0.15"
    );
}

#[test]
fn approx_l2_sq_relative_error_bounded() {
    let vecs = rand_vecs(200, 64, 88);
    let codec = Sq8Codec::train(&vecs);
    let encoded: Vec<EncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();

    let mut max_rel_err = 0.0f32;
    for i in 0..vecs.len() {
        for j in (i + 1)..vecs.len().min(i + 10) {
            let true_l2 = l2_sq_f32(&vecs[i], &vecs[j]);
            let approx = codec.approx_l2_sq(&encoded[i], &encoded[j]);
            let denom = true_l2.max(1e-6);
            let rel = (approx - true_l2).abs() / denom;
            if rel > max_rel_err {
                max_rel_err = rel;
            }
        }
    }
    assert!(
        max_rel_err < 0.15,
        "max relative L2² error {max_rel_err:.4} >= 0.15"
    );
}

#[test]
fn narrow_sq8_dimensions_survive_a_wide_training_dimension() {
    for dims in [2, 17, 384, 768, 1536, 3072] {
        let zero = vec![0.0_f32; dims];
        let mut upper = vec![255.0_f32 / 16384.0; dims];
        upper[0] = 255.0;
        let mut narrow = upper.clone();
        narrow[0] = 0.0;
        let codec = Sq8Codec::train(&[zero.clone(), upper]);
        let q = codec.encode(&zero);
        let b = codec.encode(&narrow);
        assert!(q.codes.iter().all(|&code| code == 0));
        assert_eq!(b.codes[0], 0);
        assert!(b.codes[1..].iter().all(|&code| code == 255));

        let expected = ((dims - 1) as f64 * 65025.0 / 268435456.0) as f32;
        let distance = codec.approx_l2_sq(&q, &b);
        assert_eq!(distance, expected, "dims={dims}");
        assert!(distance > codec.approx_l2_sq(&q, &q), "dims={dims}");
        assert_eq!(codec.approx_dot(&b, &b), expected, "dims={dims}");
        if dims == 2 {
            assert!(codec.approx_cosine_dist(&b, &b) < 1e-6);
        }
    }
}

#[test]
fn order_preservation_triplets_cosine() {
    let vecs = rand_vecs(300, 64, 99);
    let codec = Sq8Codec::train(&vecs);
    let encoded: Vec<EncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();

    let n = vecs.len();
    let mut agree = 0usize;
    let mut total = 0usize;

    for anchor in 0..50 {
        let a = &vecs[anchor];
        let ea = &encoded[anchor];
        for b_idx in 0..n {
            for c_idx in (b_idx + 1)..n.min(b_idx + 5) {
                let norm_a: f32 = a.iter().map(|x| x * x).sum::<f32>().sqrt();
                let norm_b: f32 = vecs[b_idx].iter().map(|x| x * x).sum::<f32>().sqrt();
                let norm_c: f32 = vecs[c_idx].iter().map(|x| x * x).sum::<f32>().sqrt();

                let cos_ab = dot_f32(a, &vecs[b_idx]) / (norm_a * norm_b).max(1e-9);
                let cos_ac = dot_f32(a, &vecs[c_idx]) / (norm_a * norm_c).max(1e-9);
                let dist_ab_true = 1.0 - cos_ab;
                let dist_ac_true = 1.0 - cos_ac;

                let dist_ab_approx = codec.approx_cosine_dist(ea, &encoded[b_idx]);
                let dist_ac_approx = codec.approx_cosine_dist(ea, &encoded[c_idx]);

                if (dist_ab_true - dist_ac_true).abs() < 0.01 {
                    continue;
                }

                let true_closer_b = dist_ab_true < dist_ac_true;
                let approx_closer_b = dist_ab_approx < dist_ac_approx;
                if true_closer_b == approx_closer_b {
                    agree += 1;
                }
                total += 1;
            }
        }
    }

    let rate = agree as f64 / total.max(1) as f64;
    assert!(
        rate >= 0.95,
        "order preservation {rate:.3} < 0.95 ({agree}/{total})"
    );
}

#[test]
fn order_preservation_triplets_l2() {
    let vecs = rand_vecs(300, 64, 101);
    let codec = Sq8Codec::train(&vecs);
    let encoded: Vec<EncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();

    let n = vecs.len();
    let mut agree = 0usize;
    let mut total = 0usize;

    for anchor in 0..50 {
        let a = &vecs[anchor];
        let ea = &encoded[anchor];
        for b_idx in 0..n {
            for c_idx in (b_idx + 1)..n.min(b_idx + 5) {
                let dist_ab_true = l2_sq_f32(a, &vecs[b_idx]);
                let dist_ac_true = l2_sq_f32(a, &vecs[c_idx]);

                let dist_ab_approx = codec.approx_l2_sq(ea, &encoded[b_idx]);
                let dist_ac_approx = codec.approx_l2_sq(ea, &encoded[c_idx]);

                if (dist_ab_true - dist_ac_true).abs() < 0.001 {
                    continue;
                }

                let true_closer_b = dist_ab_true < dist_ac_true;
                let approx_closer_b = dist_ab_approx < dist_ac_approx;
                if true_closer_b == approx_closer_b {
                    agree += 1;
                }
                total += 1;
            }
        }
    }

    let rate = agree as f64 / total.max(1) as f64;
    assert!(
        rate >= 0.95,
        "L2 order preservation {rate:.3} < 0.95 ({agree}/{total})"
    );
}

#[test]
fn train_flat_matches_train_rows() {
    let vecs = rand_vecs(50, 16, 123);
    let flat: Vec<f32> = vecs.iter().flatten().copied().collect();

    let codec_rows = Sq8Codec::train(&vecs);
    let codec_flat = Sq8Codec::train_flat(&flat, 16);

    for d in 0..16 {
        assert!((codec_rows.min[d] - codec_flat.min[d]).abs() < 1e-6);
        assert!((codec_rows.scale[d] - codec_flat.scale[d]).abs() < 1e-6);
    }
}

#[test]
fn encode_par_matches_sequential() {
    let vecs = rand_vecs(50, 32, 555);
    let codec = Sq8Codec::train(&vecs);

    let seq: Vec<EncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();
    let par = codec.encode_par(&vecs);

    assert_eq!(seq.len(), par.len());
    for (s, p) in seq.iter().zip(par.iter()) {
        assert_eq!(s.codes, p.codes);
        assert!((s.soc_sum - p.soc_sum).abs() < 1e-5);
    }
}

#[test]
fn sq8_try_encode_par_short_row_returns_error_not_panic() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let mut rows = rand_vecs(5, 4, 2);
    rows[3] = vec![0.0, 1.0];
    let err = codec.try_encode_par(&rows).unwrap_err();
    assert_eq!(
        err,
        QuantError::EncodeLengthMismatch {
            expected: 4,
            got: 2
        }
    );
}

#[test]
fn sq8_try_encode_par_long_row_returns_error_not_panic() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let mut rows = rand_vecs(5, 4, 2);
    rows[3] = vec![0.0, 1.0, 2.0, 3.0, 4.0, 5.0];
    let err = codec.try_encode_par(&rows).unwrap_err();
    assert_eq!(
        err,
        QuantError::EncodeLengthMismatch {
            expected: 4,
            got: 6
        }
    );
}

#[test]
#[should_panic(expected = "vector length 2 does not match codec dims 4")]
fn sq8_encode_par_still_panics_with_typed_message_on_short_row() {
    let codec = Sq8Codec::train(&rand_vecs(10, 4, 1));
    let mut rows = rand_vecs(5, 4, 2);
    rows[3] = vec![0.0, 1.0];
    let _ = codec.encode_par(&rows);
}

#[test]
fn u8_dot_u32_matches_scalar() {
    let a: Vec<u8> = (0u8..=255).take(384).collect();
    let b: Vec<u8> = (0u8..=255).rev().take(384).collect();
    let scalar: u32 = a
        .iter()
        .zip(b.iter())
        .map(|(&x, &y)| x as u32 * y as u32)
        .sum();
    assert_eq!(u8_dot_u32(&a, &b), scalar, "u8_dot_u32 mismatch");
}

#[test]
#[should_panic(expected = "u8_dot_u32 inputs must have equal length")]
fn u8_dot_u32_rejects_shorter_second_slice() {
    let a = [1u8; 16];
    let b = [2u8; 1];

    let _ = u8_dot_u32(&a, &b);
}

#[test]
#[should_panic(expected = "u8_dot_u32 inputs must have equal length")]
fn u8_dot_u32_rejects_longer_second_slice() {
    let a = [1u8; 1];
    let b = [2u8; 16];

    let _ = u8_dot_u32(&a, &b);
}

#[test]
#[should_panic(expected = "approx_dot input codes must match codec dims")]
fn approx_dot_rejects_mismatched_codes() {
    let vectors = rand_vecs(2, 16, 1);
    let codec = Sq8Codec::train(&vectors);
    let a = codec.encode(&vectors[0]);
    let mut b = codec.encode(&vectors[1]);
    b.codes.pop();

    let _ = codec.approx_dot(&a, &b);
}

#[test]
#[should_panic(expected = "approx_dot input codes must match codec dims")]
fn approx_dot_rejects_longer_first_codes() {
    let vectors = rand_vecs(2, 16, 1);
    let codec = Sq8Codec::train(&vectors);
    let mut a = codec.encode(&vectors[0]);
    let b = codec.encode(&vectors[1]);
    a.codes.push(0);

    let _ = codec.approx_dot(&a, &b);
}

#[test]
#[should_panic(expected = "approx_l2_sq input codes must match codec dims")]
fn approx_l2_sq_rejects_mismatched_codes() {
    let vectors = rand_vecs(2, 16, 1);
    let codec = Sq8Codec::train(&vectors);
    let a = codec.encode(&vectors[0]);
    let mut b = codec.encode(&vectors[1]);
    b.codes.pop();

    let _ = codec.approx_l2_sq(&a, &b);
}

#[test]
#[should_panic(expected = "approx_l2_sq input codes must match codec dims")]
fn approx_l2_sq_rejects_longer_first_codes() {
    let vectors = rand_vecs(2, 16, 1);
    let codec = Sq8Codec::train(&vectors);
    let mut a = codec.encode(&vectors[0]);
    let b = codec.encode(&vectors[1]);
    a.codes.push(0);

    let _ = codec.approx_l2_sq(&a, &b);
}

#[test]
fn u8_helpers_tail_path_max_diff() {
    for len in [1usize, 7, 15, 17, 100, 383] {
        let a = vec![255u8; len];
        let b = vec![0u8; len];
        assert_eq!(u8_l2sq_u32(&a, &b), len as u32 * 255 * 255, "l2 len={len}");
        assert_eq!(u8_dot_u32(&a, &a), len as u32 * 255 * 255, "dot len={len}");
    }
}

#[test]
fn u8_l2sq_u32_matches_scalar() {
    let a: Vec<u8> = (0u8..=255).take(384).collect();
    let b: Vec<u8> = (0u8..=255).rev().take(384).collect();
    let scalar: u32 = a
        .iter()
        .zip(b.iter())
        .map(|(&x, &y)| {
            let d = (x as i32) - (y as i32);
            (d * d) as u32
        })
        .sum();
    assert_eq!(u8_l2sq_u32(&a, &b), scalar, "u8_l2sq_u32 mismatch");
}

#[test]
#[should_panic(expected = "u8_l2sq_u32 inputs must have equal length")]
fn u8_l2sq_u32_rejects_shorter_second_slice() {
    let a = [1u8; 16];
    let b = [2u8; 1];

    let _ = u8_l2sq_u32(&a, &b);
}

// ── GsSq8Codec tests ────────────────────────────────────────────────────

/// Regression counterexample (2026-06-12): ranges [0,1] and [0,1e6].
///
/// Without global-scale, the per-dim fast path reversed near/far ordering by >6 OOM.
/// With GsSq8Codec the global scale is dominated by the wide dim; the narrow dim
/// loses code resolution but contributes proportionally little to L2 — ordering is preserved.
#[test]
fn gs_l2_sq_anisotropic_ordering_preserved() {
    let corpus = vec![
        vec![0.0f32, 0.0f32],    // origin
        vec![1.0f32, 1.0f32],    // near: exact L2² = 2.0
        vec![1.0f32, 4001.0f32], // far: exact L2² ~ 16_000_002
    ];
    let codec = GsSq8Codec::train(&corpus);

    let enc_origin = codec.encode(&corpus[0]);
    let enc_near = codec.encode(&corpus[1]);
    let enc_far = codec.encode(&corpus[2]);

    let d_near = codec.l2_sq(&enc_origin, &enc_near);
    let d_far = codec.l2_sq(&enc_origin, &enc_far);

    assert!(
        d_near < d_far,
        "GsSq8Codec reversed near/far on anisotropic corpus: near={d_near} far={d_far} \
             (anisotropy_ratio={:.1})",
        codec.anisotropy_ratio
    );
}

#[test]
fn gs_l2_sq_isotropic_small_error() {
    let vecs = rand_vecs(200, 64, 202);
    let codec = GsSq8Codec::train(&vecs);
    let encoded: Vec<GsEncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();

    let mut max_rel = 0.0f32;
    for i in 0..vecs.len() {
        for j in (i + 1)..vecs.len().min(i + 10) {
            let true_l2 = l2_sq_f32(&vecs[i], &vecs[j]);
            let approx = codec.l2_sq(&encoded[i], &encoded[j]);
            let denom = true_l2.max(1e-6);
            let rel = (approx - true_l2).abs() / denom;
            if rel > max_rel {
                max_rel = rel;
            }
        }
    }
    assert!(
        max_rel < 0.15,
        "GsSq8Codec max relative L2² error {max_rel:.4} >= 0.15"
    );
}

#[test]
fn gs_train_flat_matches_train_rows() {
    let vecs = rand_vecs(50, 16, 321);
    let flat: Vec<f32> = vecs.iter().flatten().copied().collect();

    let codec_rows = GsSq8Codec::train(&vecs);
    let codec_flat = GsSq8Codec::train_flat(&flat, 16);

    assert!((codec_rows.gs - codec_flat.gs).abs() < 1e-7);
    for d in 0..16 {
        assert!((codec_rows.min[d] - codec_flat.min[d]).abs() < 1e-6);
    }
}

#[test]
fn gs_l2_sq_order_preservation_triplets() {
    let vecs = rand_vecs(300, 64, 303);
    let codec = GsSq8Codec::train(&vecs);
    let encoded: Vec<GsEncodedVector> = vecs.iter().map(|v| codec.encode(v)).collect();

    let n = vecs.len();
    let mut agree = 0usize;
    let mut total = 0usize;

    for anchor in 0..50 {
        let a = &vecs[anchor];
        let ea = &encoded[anchor];
        for b_idx in 0..n {
            for c_idx in (b_idx + 1)..n.min(b_idx + 5) {
                let dist_ab_true = l2_sq_f32(a, &vecs[b_idx]);
                let dist_ac_true = l2_sq_f32(a, &vecs[c_idx]);

                let dist_ab_approx = codec.l2_sq(ea, &encoded[b_idx]);
                let dist_ac_approx = codec.l2_sq(ea, &encoded[c_idx]);

                if (dist_ab_true - dist_ac_true).abs() < 0.001 {
                    continue;
                }

                let true_closer_b = dist_ab_true < dist_ac_true;
                let approx_closer_b = dist_ab_approx < dist_ac_approx;
                if true_closer_b == approx_closer_b {
                    agree += 1;
                }
                total += 1;
            }
        }
    }

    let rate = agree as f64 / total.max(1) as f64;
    assert!(
        rate >= 0.95,
        "GsSq8Codec L2 order preservation {rate:.3} < 0.95 ({agree}/{total})"
    );
}