gras 1.1.2

A Genetic Programming framework for Neural Architecture Search in Rust
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
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
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
//! Step-race run state — replaces the generational `gen_XX.json` + per-gen
//! `robustness.csv` artifact shape with a present-tense model:
//!
//! - `engine.json` — the run header, overwritten when reproducible config
//!   changes. Still named `engine.json` so it reads as "the engine run's data."
//! - `nets/<hash>.json` — one per live net, overwritten each step the net
//!   survives. Contains everything needed to hand that net back to the stream
//!   later (topology JSON, weight-init seed, current step, lineage placeholder).
//!
//! No per-gen snapshots, no default `history.csv` (each net state file carries
//! its own metrics snapshot), no default `culled.log` (a net that stops being
//! stepped simply stops being updated; an explicit tombstone log can be added
//! later if an audit trail is wanted).

use std::collections::HashMap;
use std::path::Path;

use flodl::tensor::Result;
use serde::{Deserialize, Serialize};

use crate::engine::config::RaceConfig;
use crate::engine::fitness::{Direction, FitnessLabel, Metric};
use crate::graph::topology::{Topology, TopologyOptions};
use crate::utils::error::EngineError;
use crate::utils::seed::topo_hash;

// ── Run header (engine.json) ───────────────────────────────────────────────

/// The run's **complete** configuration, serialized into `engine.json` under
/// `"config"`.
///
/// `engine.json` is the single source of truth for a run's settings: every knob
/// that can change the trajectory is recorded here, so the file alone is enough
/// to reproduce the run — no parallel `options.csv` is written. To add a knob,
/// add it to this struct and to [`ConfigSnapshot::from_config`].
#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
#[serde(default)]
pub struct ConfigSnapshot {
    /// Population size the race held constant.
    pub pop_size: usize,
    pub checkpoint_every: usize,
    pub crossover_rolls: usize,
    pub mutate_rolls: usize,
    /// Checkpoint gate strictness (`"hard"` | `"soft"`).
    pub crossover_gate: String,
    /// Crossover gate window (last-k checkpoints; 0 = all). Replay-relevant.
    #[serde(default)]
    pub crossover_gate_window: usize,
    /// Extra gate-aware retries per crossover roll (0 = none).
    pub crossover_retries: usize,
    /// Crossover replacement policy (`"worst"` | `"random"`). Crossover-only:
    /// mutation immigrants always evict via fitness-inverse roulette.
    pub crossover_cull_policy: String,
    /// Elite guard size: top-k nets by smoothed fitness immune to all culls.
    pub elite_count: usize,
    /// Crossover operator pool (`"one_point"`/`"uniform"`); empty ⇒ both.
    pub crossover_ops_pool: Vec<String>,
    pub crossover_prob: f32,
    pub mutate_prob: f32,
    /// Stop budget as configured (`None` = no limit). Duplicated at
    /// the header root for readers that want the stop criterion without
    /// opening `config`; this copy is the authoritative one.
    #[serde(default)]
    pub max_steps: Option<usize>,
    /// Whether a pluggable `custom_stop` closure was installed. The closure is
    /// code, not data, so only its presence is recordable.
    pub has_custom_stop: bool,
    /// Per-step verbosity (`"none"` | `"summ"` | `"minimal"`).
    pub log_level: String,
    /// Whether the run wrote the `telemetry.jsonl` event sidecar
    /// (`#[serde(default)]` so pre-knob `engine.json` files still load).
    #[serde(default)]
    pub trace_file: bool,
    /// Problem-space target (`"tabular"` | `"onecimage"` | `"threecimage"` |
    /// `"nlp"` | `"rl"`).
    pub mode: String,
    /// Effective shared-stream geometry — after the trainer's `stream_shape()`
    /// override, i.e. the shape the run actually used. `None` in modes with no
    /// shared stream (RL): there is no stream and no eval batch to describe,
    /// and the trainer owns the per-step volume (`matches_per_step`, …),
    /// recorded in `engine.json` → `"trainer"`. Tabular defaults must never be
    /// stamped into a run that never draws a batch.
    pub batch_size: Option<usize>,
    pub eval_batch_size: Option<usize>,
    /// Per-net fitness smoothing window (K).
    pub smoothing_window: usize,
    /// Device the run executed on (`"cpu"` | `"cuda:0"`).
    pub device: String,
    /// Whether the lossless per-step `metrics.csv` was written.
    pub csv_export: bool,
    /// Post-race pruner: `(method, extra_steps)`. `None` = off. `#[serde(default)]`
    /// so legacy `engine.json` headers load cleanly.
    #[serde(default)]
    pub pop_pruner: Option<(String, usize)>,
    /// Human label for the experiment (`set_run_name`), verbatim. Purely
    /// informative — never affects the results folder name.
    pub run_name: Option<String>,
    /// Anti-devolution A (`set_elite_freeze`): top-k nets skip the trainer
    /// call. Changes WHICH nets train during a replay, so it must match on
    /// resume — record it.
    #[serde(default)]
    pub freeze_elites: bool,
    /// Mutation catch-up toggle (`set_mutation_catch_up`, engine.json key
    /// kept as `fresh_immigrants`): `true` = immigrants replay catch-up.
    /// Replay-relevant — a resumed run must use the same insertion semantics
    /// or the replayed population diverges. Reads the active mode arm
    /// (tabular and RL each carry their own copy).
    #[serde(default = "default_false")]
    pub fresh_immigrants: bool,
    /// Crossover catch-up toggle (`set_crossover_catch_up`): `false` = a
    /// crossover child skips replay AND the checkpoint gate. Replay-relevant
    /// for the same reason as `fresh_immigrants`.
    #[serde(default = "default_true")]
    pub crossover_catch_up: bool,
    /// How many of the initial population slots were user-supplied run
    /// topologies (the founding batch) (`set_run_topologies`). Informational (lineage reads:
    /// ordinals `< run_topology_count` are imports, the rest are random draws).
    #[serde(default)]
    pub run_topology_count: usize,
    /// Per-checkpoint elite weight snapshot flag
    /// (`set_elite_checkpoint_weights`). Informational — describes artifact
    /// cadence, not replay semantics.
    #[serde(default = "default_true")]
    pub elite_checkpoint_weights: bool,
    /// Mutation probation steps (`set_mutation_probation_steps`):
    /// cull-immunity for a net's first k clocks. Replay-relevant — it changes
    /// WHICH nets get culled, so a resumed run must use the same window.
    #[serde(default)]
    pub mutation_probation_steps: usize,
    /// Anti-plateau challenge probability (`set_run_challenge_prob`):
    /// per-step, per-net chance a challenged step fires. Replay-relevant —
    /// the trigger is re-derived from (run_seed, net_seed, step) on replay,
    /// so a resumed run must roll the same probability.
    #[serde(default)]
    pub challenge_prob: f32,
    /// One-line identity of the binary that produced the run (`gras` version,
    /// profile, exe path, build time). Diagnostics only — lets a reader tell
    /// a stale-process artifact from a logic bug.
    #[serde(default)]
    pub build: String,
}

impl ConfigSnapshot {
    /// Capture every knob from a run's config plus the effective stream shape
    /// (`None`/`None` for modes with no shared stream — see the fields).
    pub fn from_config(
        cfg: &RaceConfig,
        batch_size: Option<usize>,
        eval_batch_size: Option<usize>,
    ) -> Self {
        ConfigSnapshot {
            pop_size: cfg.pop_size,
            checkpoint_every: cfg.checkpoint_every,
            crossover_rolls: cfg.crossover_rolls,
            mutate_rolls: cfg.mutate_rolls,
            crossover_gate: format!("{:?}", cfg.crossover_gate).to_lowercase(),
            crossover_gate_window: cfg.crossover_gate_window,
            crossover_retries: cfg.crossover_retries,
            crossover_cull_policy: format!("{:?}", cfg.crossover_cull_policy).to_lowercase(),
            elite_count: cfg.elite_count,
            crossover_ops_pool: cfg.crossover_ops_pool.clone(),
            crossover_prob: cfg.crossover_prob,
            mutate_prob: cfg.mutate_prob,
            max_steps: cfg.max_steps,
            has_custom_stop: cfg.custom_stop.is_some(),
            log_level: format!("{:?}", cfg.log_level).to_lowercase(),
            trace_file: cfg.trace_file,
            mode: format!("{:?}", cfg.mode).to_lowercase(),
            batch_size,
            eval_batch_size,
            smoothing_window: cfg.smoothing_window,
            device: if cfg!(feature = "cuda") {
                "cuda:0"
            } else {
                "cpu"
            }
            .to_string(),
            csv_export: cfg.csv_export,
            pop_pruner: cfg
                .pop_pruner
                .map(|p| (format!("{:?}", p.method).to_lowercase(), p.steps)),
            run_name: cfg.run_name.clone(),
            freeze_elites: cfg.freeze_elites,
            fresh_immigrants: cfg.mode_specific.mutation_catch_up(),
            crossover_catch_up: cfg.mode_specific.crossover_catch_up(),
            run_topology_count: cfg.run_topologies.len(),
            elite_checkpoint_weights: cfg.elite_checkpoint_weights,
            mutation_probation_steps: cfg.mutation_probation_steps,
            challenge_prob: cfg.challenge_prob,
            build: crate::engine::format::build_stamp(),
        }
    }
}

/// The reproducible run header, serialized as `engine.json` in the run dir.
///
/// This is the "who are we, what rules are we running under" snapshot. It is
/// overwritten whenever the config that affects reproducibility changes — it
/// does **not** accumulate history. The live population lives in
/// `nets/<hash>.json` files, not here.
#[derive(Clone, Debug, Serialize)]
pub struct RunHeader {
    /// User-visible run id (directory name or CLI-chosen label).
    pub run_id: String,
    /// Mode discriminator (engine split, TODO.md step 6): `"tabular"` | `"rl"`.
    /// Sits at the header ROOT so a reader picks the engine before touching
    /// `config`. Written by both engines at construction and at every stop;
    /// legacy shared-shape headers have it ABSENT — `from_json` then derives
    /// it from `config.mode` (one-generation migration, documented in
    /// OPTIONS.md §0).
    #[serde(default)]
    pub engine_mode: Option<String>,
    /// Single deterministic root for the whole run. When the user passes
    /// `--seed N` this is that N; when no seed is given the engine generates
    /// one and records it here so an interrupted run can be resumed exactly.
    pub run_seed: u64,
    /// Fitness label used for ranking (e.g. `"f1"`, `"cross_entropy"`).
    pub fitness_label: FitnessLabel,
    /// Whether lower or higher fitness is better.
    pub fitness_direction: Direction,
    /// Dataset dims the run was launched with (informational; the stream
    /// derives its pool from the actual dataset on disk).
    pub input_dim: usize,
    pub output_dim: usize,
    /// Topology template + GP-pool knobs the run was launched with.
    pub topology_options: TopologyOptions,
    /// Hidden-dim sampling range.
    pub hidden_dim_pool: String,
    pub hidden_dim_stride: usize,
    /// Combine/activation/standardize pools actually used by the run.
    pub combine_op_pool: Vec<String>,
    pub activation_pool: Vec<String>,
    pub standardize_op_pool: Vec<String>,
    /// Informative (non-ranking) metrics configured for the run, if any.
    /// Their labels determine the extra columns a reader may expect in a
    /// per-net metrics snapshot.
    pub informative_metrics: Vec<String>,
    /// Stop criteria / budgets configured for the run. The scheduler honors
    /// them at runtime and `config` carries the authoritative copy of both;
    /// the stopped-by fact itself is in `history.csv` / the stop banner.
    pub max_steps: Option<usize>,
    /// When the run was started (wall clock), for human-readability.
    pub started_at: Option<String>,
    pub train_eval_split_ratio: Option<f32>,
    pub held_out_eval_rows: Option<usize>,
    /// Cumulative culls recorded at the last stop. Restored on resume so a
    /// resumed run continues the ORIGINAL cull budget instead of a fresh one.
    /// `0` on legacy engine.json files (written before this existed).
    #[serde(default)]
    pub culls: usize,
    /// Wall-clock seconds the run had accumulated at the last stop. Restored
    /// on resume as the base offset, so `elapsed_seconds` keeps counting the
    /// run's true age across stop/resume cycles. `0` on legacy files.
    #[serde(default)]
    pub run_elapsed_secs: u64,
    /// Children born per step clock at the last stop (same-clock siblings get
    /// distinct child seeds). Restored on resume; empty on legacy files.
    #[serde(default)]
    pub children_born_at_clock: std::collections::HashMap<usize, usize>,
    /// The run's complete configuration (see [`ConfigSnapshot`]). This is the
    /// authoritative settings record — `engine.json` alone reproduces a run.
    #[serde(default)]
    pub config: ConfigSnapshot,
    /// The training scheme's self-description, from `Trainer::describe()`
    /// (learning rate, grad clip, optimizer, schedule, …). Free-form JSON:
    /// the engine persists it verbatim and never interprets it. `None` when
    /// the trainer doesn't implement the hook (or on legacy `engine.json`).
    #[serde(default)]
    pub trainer: Option<serde_json::Value>,
}

/// Everything that defines a step-race run's reproducible configuration —
/// the field-by-field input to [`RunHeader::from_race_options`].
#[derive(Debug)]
pub struct RunConfig {
    pub run_id: String,
    /// Mode discriminator (`"tabular"` | `"rl"`) — see [`RunHeader::engine_mode`].
    pub engine_mode: Option<String>,
    pub run_seed: u64,
    pub fitness_label: FitnessLabel,
    pub fitness_direction: Direction,
    pub input_dim: usize,
    pub output_dim: usize,
    pub topology_options: TopologyOptions,
    pub hidden_dim_pool: std::ops::RangeInclusive<usize>,
    pub hidden_dim_stride: usize,
    pub combine_op_pool: Vec<String>,
    pub activation_pool: Vec<String>,
    pub standardize_op_pool: Vec<String>,
    pub informative_metrics: Vec<Metric>,
    pub max_steps: Option<usize>,
    pub train_eval_split_ratio: Option<f32>,
    pub held_out_eval_rows: Option<usize>,
    pub culls: usize,
    pub run_elapsed_secs: u64,
    pub children_born_at_clock: std::collections::HashMap<usize, usize>,
    /// The run's complete configuration (see [`ConfigSnapshot`]).
    pub config: ConfigSnapshot,
    /// The training scheme's self-description (see `Trainer::describe`).
    pub trainer: Option<serde_json::Value>,
}

impl RunHeader {
    /// Build the header from the run's launch config.
    pub fn from_race_options(cfg: RunConfig) -> Self {
        let RunConfig {
            run_id,
            engine_mode,
            run_seed,
            fitness_label,
            fitness_direction,
            input_dim,
            output_dim,
            topology_options,
            hidden_dim_pool,
            hidden_dim_stride,
            combine_op_pool,
            activation_pool,
            standardize_op_pool,
            informative_metrics,
            max_steps,
            train_eval_split_ratio,
            held_out_eval_rows,
            culls,
            run_elapsed_secs,
            children_born_at_clock,
            config,
            trainer,
        } = cfg;
        RunHeader {
            run_id,
            engine_mode,
            run_seed,
            fitness_label,
            fitness_direction,
            input_dim,
            output_dim,
            topology_options,
            hidden_dim_pool: format!("{}..={}", hidden_dim_pool.start(), hidden_dim_pool.end()),
            hidden_dim_stride,
            combine_op_pool,
            activation_pool,
            standardize_op_pool,
            informative_metrics: informative_metrics
                .into_iter()
                .map(|m| m.label().to_string())
                .collect(),
            max_steps,
            started_at: None,
            train_eval_split_ratio,
            held_out_eval_rows,
            culls,
            run_elapsed_secs,
            children_born_at_clock,
            config,
            trainer,
        }
    }

    /// Like `from_race_options`, but also stamps `started_at`.
    pub fn from_race_options_at(cfg: RunConfig, started_at: String) -> Self {
        let mut out = Self::from_race_options(cfg);
        out.started_at = Some(started_at);
        out
    }

    /// Serialize to pretty JSON for writing `engine.json`.
    pub fn to_json(&self) -> Result<String> {
        Ok(serde_json::to_string_pretty(self)
            .map_err(|e| EngineError::Json(format!("engine.json serialize: {e}")))?)
    }

    /// Load a previously written `engine.json`.
    pub fn from_json(raw: &str) -> Result<Self> {
        let v: serde_json::Value = serde_json::from_str(raw)
            .map_err(|e| EngineError::Json(format!("engine.json parse: {e}")))?;
        let fitness_direction = v
            .get("fitness_direction")
            .and_then(|f| f.as_str())
            .and_then(|s| match s {
                "Minimize" => Some(Direction::Minimize),
                "Maximize" => Some(Direction::Maximize),
                _ => None,
            })
            .unwrap_or(Direction::Maximize);
        let fitness_label = serde_json::from_value(
            v.get("fitness_label")
                .cloned()
                .unwrap_or_else(|| serde_json::Value::String("loss".into())),
        )
        .unwrap_or(FitnessLabel("loss".into()));
        let input_dim = v.get("input_dim").and_then(|f| f.as_u64()).unwrap_or(0) as usize;
        let output_dim = v.get("output_dim").and_then(|f| f.as_u64()).unwrap_or(0) as usize;
        let topology_options = v
            .get("topology_options")
            .map(|t| serde_json::from_value(t.clone()).unwrap_or_default())
            .unwrap_or_default();
        let hidden_dim_pool = v
            .get("hidden_dim_pool")
            .and_then(|f| f.as_str())
            .map(|s| s.to_string())
            .unwrap_or_else(|| "4..=8".into());
        let hidden_dim_stride: usize = v
            .get("hidden_dim_stride")
            .and_then(|f| f.as_u64())
            .map(|n| n as usize)
            .unwrap_or(1);
        let combine_op_pool: Vec<String> = serde_json::from_value(
            v.get("combine_op_pool")
                .cloned()
                .unwrap_or(serde_json::Value::Null),
        )
        .unwrap_or_default();
        let activation_pool: Vec<String> = serde_json::from_value(
            v.get("activation_pool")
                .cloned()
                .unwrap_or(serde_json::Value::Null),
        )
        .unwrap_or_default();
        let standardize_op_pool: Vec<String> = serde_json::from_value(
            v.get("standardize_op_pool")
                .cloned()
                .unwrap_or(serde_json::Value::Null),
        )
        .unwrap_or_default();
        let informative_metrics: Vec<String> = serde_json::from_value(
            v.get("informative_metrics")
                .cloned()
                .unwrap_or(serde_json::Value::Null),
        )
        .unwrap_or_default();
        let max_steps = v
            .get("max_steps")
            .and_then(|f| f.as_u64())
            .map(|n| n as usize);
        let started_at = v
            .get("started_at")
            .and_then(|f| f.as_str())
            .map(|s| s.to_string());
        let train_eval_split_ratio = v
            .get("train_eval_split_ratio")
            .and_then(|f| f.as_f64())
            .map(|n| n as f32);
        let held_out_eval_rows = v
            .get("held_out_eval_rows")
            .and_then(|f| f.as_u64())
            .map(|n| n as usize);
        // Counter persistence (written at stop; legacy files read as fresh).
        let culls = v.get("culls").and_then(|f| f.as_u64()).unwrap_or(0) as usize;
        let run_elapsed_secs = v
            .get("run_elapsed_secs")
            .and_then(|f| f.as_u64())
            .unwrap_or(0);
        let children_born_at_clock: std::collections::HashMap<usize, usize> = v
            .get("children_born_at_clock")
            .and_then(|c| c.as_object())
            .map(|m| {
                m.iter()
                    .filter_map(|(k, val)| {
                        let clock: usize = k.parse().ok()?;
                        let n = val.as_u64()? as usize;
                        Some((clock, n))
                    })
                    .collect()
            })
            .unwrap_or_default();
        // Missing on legacy engine.json files ⇒ Default (all-zero snapshot).
        let config: ConfigSnapshot = v
            .get("config")
            .cloned()
            .and_then(|c| serde_json::from_value(c).ok())
            .unwrap_or_default();
        // Trainer self-description (see `Trainer::describe`). Absent on legacy
        // files and on trainers that don't implement the hook.
        let trainer = v.get("trainer").cloned();
        // Migration (engine split, TODO.md step 6): legacy shared-shape
        // headers have no root `engine_mode` — derive it from `config.mode`
        // (same vocabulary, always present). New headers carry it at the
        // root so a reader picks the engine before touching `config`.
        let engine_mode = v
            .get("engine_mode")
            .and_then(|m| m.as_str())
            .map(|s| s.to_string())
            .or_else(|| {
                v.get("config")
                    .and_then(|c| c.get("mode"))
                    .and_then(|m| m.as_str())
                    .map(|s| s.to_string())
            });
        Ok(RunHeader {
            run_id: v
                .get("run_id")
                .and_then(|f| f.as_str())
                .unwrap_or("unknown")
                .to_string(),
            engine_mode,
            run_seed: v.get("run_seed").and_then(|f| f.as_u64()).unwrap_or(0),
            fitness_label,
            fitness_direction,
            input_dim,
            output_dim,
            topology_options,
            hidden_dim_pool,
            hidden_dim_stride,
            combine_op_pool,
            activation_pool,
            standardize_op_pool,
            informative_metrics,
            max_steps,
            started_at,
            train_eval_split_ratio,
            held_out_eval_rows,
            culls,
            run_elapsed_secs,
            children_born_at_clock,
            config,
            trainer,
        })
    }
}

fn default_alive() -> bool {
    true
}

fn default_false() -> bool {
    false
}

fn default_true() -> bool {
    true
}

// ── Per-net state (nets/<hash>.json) ──────────────────────────────────────

/// The last-known state of one live net. Written to `nets/<hash>.json` and
/// overwritten each step the net survives (and on cull-enter, if a tombstone
/// path is enabled later).
///
/// This is the resume anchor for that net: load it, rebuild the topology +
/// network from the embedded blueprint + weight-init seed, then replay the
/// deterministic stream to `step`.
#[derive(Clone, Debug, Serialize, Deserialize)]
pub struct NetState {
    /// Canonical xxh3 hash of this individual's topology (16 hex chars).
    pub hash: String,
    /// The topology blueprint, serialized. Self-describing: `Network::build`
    /// reads `options.topology_seed` (weight init) and `options.dropout_prob` (the
    /// regularization the run applied) straight from here.
    pub topology: String,
    /// Weight-init seed embedded in the topology's `options.topology_seed`. Stored
    /// here redundantly so a reader doesn't have to re-parse the topology to
    /// recover the seed the run used for this individual.
    pub net_seed: usize,
    /// The step this net has been trained/evaluated through. A resume replays
    /// the stream from step 0 up to (but not including) this step, then
    /// continues from here.
    pub step: usize,
    /// Whether this net is currently alive in the active population. Culled nets
    /// are written as final tombstones to disk with `is_alive: false`.
    #[serde(default = "default_alive")]
    pub is_alive: bool,
    /// Step at which this net was culled (`None` while it is alive). With
    /// `cull_reason` and `final_smoothed_fitness` this makes a tombstone a
    /// complete record of when and why a net left the population — not just
    /// the surviving frontier.
    #[serde(default)]
    pub culled_at_step: Option<usize>,
    /// Why it left (`"crossover"` | `"immigrant-slot"`).
    #[serde(default)]
    pub cull_reason: Option<String>,
    /// Its smoothed fitness at the moment of culling.
    #[serde(default)]
    pub final_smoothed_fitness: Option<f32>,
    /// When this individual entered the live population. For lifetime tracking
    /// and the robustness-equivalent (appearances / span / final score).
    pub entered_at_step: usize,
    /// Optional lineage record for Iter 5: which parent(s) / mutation produced
    /// this individual. Left as an optional string so the field exists today
    /// without inventing the crossover/mutation schema ahead of Iter 5.
    pub created_from: Option<String>,
    /// Last-seen metrics snapshot. Enough for a resume/time-travel read
    /// without walking a separate CSV. Schema: loss + fitness + any configured
    /// informative metrics. Empty until the first persisted step.
    pub last_metrics: Option<NetMetrics>,
    /// Inclusive clock ranges this net executed as FROZEN (act-and-measure:
    /// played/measured, weight update discarded). Replay-relevant: resume
    /// re-runs frozen clocks through a no-op optimizer to reproduce the
    /// recorded metrics bit-exactly (a real optimizer pass would move the
    /// frozen weights). Compact run-length form: `[(start, end), …]`,
    /// non-overlapping, extending contiguously.
    #[serde(default)]
    pub frozen_spans: Vec<(usize, usize)>,
    /// Rich per-individual metadata (params, dims, training regime, run
    /// context). Written once at construction; missing on legacy JSONs
    /// (serde default) and backfilled on first rewrite.
    #[serde(default)]
    pub meta: NetMeta,
}

/// Rich metadata for one individual, embedded in `nets/<hash>.json` under
/// the `meta` key. Everything a reader (human or tool) needs to understand
/// the net *and* the conditions it trained under, without touching code.
/// Populated once at construction; `params` is stamped right after the
/// `Network` is built (cheap, computed once — never re-serialized per step).
#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq)]
pub struct NetMeta {
    /// Trainable parameter count (elements) of the built network.
    #[serde(default)]
    pub params: Option<i64>,
    /// Number of parameter tensors.
    #[serde(default)]
    pub param_tensors: Option<usize>,
    /// Dataset dims the run was launched with.
    #[serde(default)]
    pub input_dim: Option<usize>,
    #[serde(default)]
    pub output_dim: Option<usize>,
    /// Shared training regime for the run. `None` in modes with no shared
    /// stream (RL) — learning rate, grad clip and the per-step volume are all
    /// trainer-owned and live in `engine.json` → `"trainer"`
    /// (`Trainer::describe`), never in per-net schema.
    #[serde(default)]
    pub batch_size: Option<usize>,
    /// Dropout probability stamped into the topology (also lives in the
    /// topology JSON; duplicated here for quick scanning).
    #[serde(default)]
    pub dropout_prob: Option<f32>,
    /// Ranking fitness label for this run. There is deliberately no
    /// `loss_label` beside it: the loss/update scheme is trainer-owned (the
    /// engine cannot see it), so `engine.json` → `"trainer"` is the only
    /// honest record.
    #[serde(default)]
    pub fitness_label: Option<String>,
    /// What ranking direction this run uses ("minimize"/"maximize").
    #[serde(default)]
    pub direction: Option<String>,
    /// The step this individual was born into the run (== entered_at_step;
    /// duplicated so the meta block is self-contained).
    #[serde(default)]
    pub born_at_step: Option<usize>,
    /// Population size + run seed context (inherited from the header).
    #[serde(default)]
    pub pop_size: Option<usize>,
    #[serde(default)]
    pub run_seed: Option<u64>,
}

/// One snapshot of metrics for a single net, at a single step.
#[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
pub struct NetMetrics {
    pub step: usize,
    /// Training loss for this step's train batch.
    pub train_loss: f32,
    /// Held-out loss for this step's eval batch (eval mode).
    pub eval_loss: Option<f32>,
    /// Ranking fitness for this step's eval batch (the signal the run evolves on).
    pub fitness: f32,
    /// Informative-only metric values, in the same order the run configured them.
    pub informative: Vec<f32>,
    /// True when this step was an ACT-AND-MEASURE pass (frozen elite: the net
    /// played/measured at this clock but the weight update was discarded).
    /// Replay-relevant: resume replays frozen steps through a no-op optimizer
    /// to reproduce the recorded metrics bit-exactly.
    #[serde(default)]
    pub frozen: bool,
}

impl NetState {
    /// Stamp the `meta` block from the materialized network + the run context.
    /// Called once per net right after its `Network` is built (initial seed,
    /// child finalize, resume rebuild) — never per step.
    pub fn stamp_meta(
        &mut self,
        net: &crate::graph::network::Network,
        run: &crate::engine::config::RunMetaCtx,
    ) {
        use flodl::Module;
        let params = net.parameters();
        self.meta.param_tensors = Some(params.len());
        self.meta.params = Some(params.iter().map(|p| p.variable.numel()).sum());
        self.meta.input_dim = Some(run.input_dim);
        self.meta.output_dim = Some(run.output_dim);
        self.meta.batch_size = run.batch_size;
        // Learning rate, grad clip and the loss scheme are trainer-owned, so
        // the engine records none of them here; the trainer's `describe()`
        // block in engine.json is the authoritative record.
        self.meta.fitness_label = Some(run.fitness_label.clone());
        self.meta.direction = Some(run.direction.clone());
        self.meta.born_at_step = Some(self.entered_at_step);
        self.meta.pop_size = Some(run.pop_size);
        self.meta.run_seed = Some(run.run_seed);
        // Dropout lives in the topology too — duplicate for quick scanning.
        self.meta.dropout_prob = Some(run.dropout_prob);
    }

    /// Build a new live individual's state from its topology and entry step.
    pub fn new(topo: &Topology, step: usize, created_from: Option<String>) -> Result<Self> {
        let topo_json = topo
            .to_json()
            .map_err(|e| EngineError::Json(format!("net state topology json: {e}")))?;
        let net_seed = topo.options.topology_seed;
        Ok(NetState {
            hash: topo_hash(&topo_json),
            topology: topo_json,
            net_seed,
            step,
            is_alive: true,
            culled_at_step: None,
            cull_reason: None,
            final_smoothed_fitness: None,
            entered_at_step: step,
            created_from,
            last_metrics: None,
            frozen_spans: Vec::new(),
            meta: NetMeta::default(),
        })
    }

    /// Record that clock `step` was an act-and-measure (frozen) step,
    /// extending the last span when contiguous.
    pub fn record_frozen_step(&mut self, step: usize) {
        match self.frozen_spans.last_mut() {
            Some((_, end)) if *end + 1 == step => *end = step,
            _ => self.frozen_spans.push((step, step)),
        }
    }

    /// Whether `step` lies inside a recorded frozen span.
    pub fn is_frozen_step(&self, step: usize) -> bool {
        self.frozen_spans.iter().any(|&(s, e)| step >= s && step <= e)
    }

    /// Serialize to pretty JSON for writing `nets/<hash>.json`.
    /// Backfill `meta` on legacy JSONs so every written record carries a
    /// self-contained metadata block. We rebuild a throwaway `Network` from
    /// the topology just to count params (cheap, once per write flight); the
    /// `RunMetaCtx::default()` is fine here because we only need the network
    /// shape, not the run regime (that comes from the callers who call
    /// `stamp_meta` directly with the real run context).
    pub fn to_json(&self) -> Result<String> {
        let mut to_write = self.clone();
        if to_write.meta.params.is_none() {
            let topo = crate::graph::topology::Topology::from_json(&to_write.topology)
                .map_err(|e| EngineError::Json(format!("net state meta backfill: {e}")))?;
            let net = crate::graph::network::Network::build(&topo, crate::Device::CPU)
                .map_err(|e| EngineError::Json(format!("net state meta backfill build: {e}")))?;
            to_write.stamp_meta(&net, &crate::engine::config::RunMetaCtx::default());
        }
        Ok(serde_json::to_string_pretty(&to_write)
            .map_err(|e| EngineError::Json(format!("net state json: {e}")))?)
    }

    /// Load a previously written `nets/<hash>.json`.
    pub fn from_json(raw: &str) -> Result<Self> {
        Ok(serde_json::from_str(raw)
            .map_err(|e| EngineError::Json(format!("net state parse: {e}")))?)
    }

    /// Update the last-metrics snapshot.
    pub fn record_metrics(&mut self, metrics: NetMetrics) {
        self.last_metrics = Some(metrics);
    }

    /// Advance the step counter (called after a net's step is fully recorded).
    pub fn advance_step(&mut self) {
        self.step += 1;
    }

    /// Recover the topology from the embedded JSON.
    pub fn topology(&self) -> Result<Topology> {
        Ok(Topology::from_json(&self.topology)
            .map_err(|e| EngineError::Json(format!("net state topology parse: {e}")))?)
    }
}

// ── Runtime population state (in memory) ──────────────────────────────────

/// The in-memory view of the live population that the scheduler reads/writes.
///
/// Backed by a hash map from canonical topology hash to `NetState`, plus a
/// small robustness-equivalent accumulator that tracks per-topology lifetime
/// and final score over the run (the replacement for the old per-generation
/// `robustness.csv` semantics — here it is one accumulator in the run, not a
/// separate CSV written at generation boundaries).
pub struct RaceState {
    /// Live nets, indexed by canonical topology hash.
    nets: HashMap<String, NetState>,
    /// Per-topology lifetime/score accumulator (robustness-equivalent).
    topology_lifetime: HashMap<String, TopologyLifetime>,
}

/// Per-topology lifetime tracking — the robustness-equivalent for the step-race
/// model. Tracks how many times a topology entered the live population, the
/// step span it covered, and its final recorded fitness.
#[derive(Clone, Debug)]
pub struct TopologyLifetime {
    /// Times this topology entered the live population.
    appearances: usize,
    /// First step this topology was live at.
    first_step: usize,
    /// Last step this topology was live at (updated as it survives).
    last_step: usize,
    /// Last recorded fitness for this topology (the final one seen in the run).
    last_fitness: Option<f32>,
}

impl TopologyLifetime {
    pub fn new(first_step: usize, fitness: Option<f32>) -> Self {
        TopologyLifetime {
            appearances: 1,
            first_step,
            last_step: first_step,
            last_fitness: fitness,
        }
    }

    pub fn record_entry(&mut self, step: usize, fitness: Option<f32>) {
        self.appearances = self.appearances.saturating_add(1);
        self.last_step = step;
        self.last_fitness = fitness;
    }

    /// Advance the liveness window: `last_step` always moves forward;
    /// fitness only overwrites when a fresh finite value is given.
    pub fn record_progress(&mut self, step: usize, fitness: Option<f32>) {
        self.last_step = self.last_step.max(step);
        if let Some(f) = fitness {
            self.last_fitness = Some(f);
        }
    }

    pub fn appearances(&self) -> usize {
        self.appearances
    }

    pub fn first_step(&self) -> usize {
        self.first_step
    }

    pub fn last_step(&self) -> usize {
        self.last_step
    }

    pub fn span(&self) -> usize {
        self.last_step.saturating_sub(self.first_step)
    }

    pub fn last_fitness(&self) -> Option<f32> {
        self.last_fitness
    }
}

impl Default for RaceState {
    fn default() -> Self {
        Self::new()
    }
}

impl RaceState {
    pub fn new() -> Self {
        RaceState {
            nets: HashMap::new(),
            topology_lifetime: HashMap::new(),
        }
    }

    /// Register a newly inserted net (after catch-up or initial insertion).
    pub fn insert(&mut self, state: NetState, initial_fitness: Option<f32>) {
        let key = state.hash.clone();
        if let Some(entry) = self.topology_lifetime.get_mut(&key) {
            entry.record_entry(state.entered_at_step, initial_fitness);
        } else {
            self.topology_lifetime.insert(
                key.clone(),
                TopologyLifetime::new(state.entered_at_step, initial_fitness),
            );
        }
        self.nets.insert(key, state);
    }

    /// Record that a live net advanced one step with the given metrics.
    pub fn record_step(&mut self, hash: &str, metrics: NetMetrics) -> Result<()> {
        let state = self.nets.get_mut(hash).ok_or_else(|| {
            EngineError::InvalidOptions(format!("race state: net {hash} not live"))
        })?;
        state.record_metrics(metrics.clone());
        state.advance_step();
        let fitness = metrics.fitness;
        let fitness = if fitness.is_finite() {
            Some(fitness)
        } else {
            None
        };
        let entry = self
            .topology_lifetime
            .entry(hash.to_string())
            .or_insert_with(|| TopologyLifetime::new(state.entered_at_step, fitness));
        entry.record_progress(metrics.step, fitness);
        Ok(())
    }

    pub fn net(&self, hash: &str) -> Option<&NetState> {
        self.nets.get(hash)
    }

    pub fn net_mut(&mut self, hash: &str) -> Option<&mut NetState> {
        self.nets.get_mut(hash)
    }

    pub fn live_hashes(&self) -> Vec<String> {
        let mut hashes: Vec<String> = self.nets.keys().cloned().collect();
        hashes.sort();
        hashes
    }

    pub fn live_count(&self) -> usize {
        self.nets.len()
    }

    /// Remove a net from the live set. The state is returned so the caller can
    /// decide whether to tombstone it (off by default today) or drop it.
    pub fn remove(&mut self, hash: &str) -> Option<NetState> {
        let state = self.nets.remove(hash);
        if let Some(state) = &state {
            self.topology_lifetime.remove(&state.hash);
        }
        state
    }

    /// Iterate over every topology lifetime entry.
    pub fn lifetimes(&self) -> Vec<(&String, &TopologyLifetime)> {
        let mut out: Vec<_> = self.topology_lifetime.iter().collect();
        out.sort_by_key(|(a, _)| *a);
        out
    }

    /// Clear the entire state (used by tests / fresh-run init).
    pub fn clear(&mut self) {
        self.nets.clear();
        self.topology_lifetime.clear();
    }
}

// ── File I/O ───────────────────────────────────────────────────────────────

/// Write the run header to `engine.json` in `dir`.
pub fn write_engine_json(dir: &Path, header: &RunHeader) -> Result<()> {
    std::fs::create_dir_all(dir).map_err(|source| EngineError::Io {
        path: dir.display().to_string(),
        source,
    })?;
    let path = dir.join("engine.json");
    let json = header.to_json()?;
    std::fs::write(&path, json).map_err(|source| EngineError::Io {
        path: path.display().to_string(),
        source,
    })?;
    Ok(())
}

/// Load the run header from `engine.json` in `dir`.
pub fn load_engine_json(dir: &Path) -> Result<RunHeader> {
    let path = dir.join("engine.json");
    let raw = std::fs::read_to_string(&path).map_err(|source| EngineError::Io {
        path: path.display().to_string(),
        source,
    })?;
    RunHeader::from_json(&raw)
}

/// Ensure the `nets/` directory exists under `dir`.
pub fn ensure_nets_dir(dir: &Path) -> Result<()> {
    std::fs::create_dir_all(dir.join("nets")).map_err(|source| EngineError::Io {
        path: dir.display().to_string(),
        source,
    })?;
    Ok(())
}

/// Write (overwrite) one net's state to `nets/<hash>.json`.
pub fn write_net_state(dir: &Path, state: &NetState) -> Result<()> {
    ensure_nets_dir(dir)?;
    let path = dir.join("nets").join(format!("{}.json", state.hash));
    let json = state.to_json()?;
    std::fs::write(&path, json).map_err(|source| EngineError::Io {
        path: path.display().to_string(),
        source,
    })?;
    Ok(())
}

/// Load one net's state from `nets/<hash>.json`.
pub fn load_net_state(dir: &Path, hash: &str) -> Result<NetState> {
    let path = dir.join("nets").join(format!("{}.json", hash));
    let raw = std::fs::read_to_string(&path).map_err(|source| EngineError::Io {
        path: path.display().to_string(),
        source,
    })?;
    NetState::from_json(&raw)
}

/// Remove one net's state file. Returns whether a file was removed.
pub fn remove_net_state(dir: &Path, hash: &str) -> Result<bool> {
    let path = dir.join("nets").join(format!("{}.json", hash));
    if path.exists() {
        std::fs::remove_file(&path).map_err(|source| EngineError::Io {
            path: path.display().to_string(),
            source,
        })?;
        Ok(true)
    } else {
        Ok(false)
    }
}

// ── Tests ──────────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    use crate::engine::fitness::{Direction, FitnessLabel};
    use crate::graph::topology::TopologyOptions;
    use std::path::PathBuf;

    fn fresh_dir() -> PathBuf {
        let dir = std::env::temp_dir().join(format!(
            "gras_race_state_test_{}",
            std::time::SystemTime::now()
                .duration_since(std::time::UNIX_EPOCH)
                .unwrap()
                .as_nanos()
        ));
        let _ = std::fs::remove_dir_all(&dir);
        dir
    }

    fn baseline_header(run_id: &str) -> RunHeader {
        RunHeader::from_race_options(RunConfig {
            run_id: run_id.to_string(),
            engine_mode: Some("tabular".into()),
            run_seed: 42,
            fitness_label: FitnessLabel("f1".to_string()),
            fitness_direction: Direction::Maximize,
            input_dim: 784,
            output_dim: 10,
            topology_options: TopologyOptions::default(),
            hidden_dim_pool: 4..=8,
            hidden_dim_stride: 1,
            combine_op_pool: vec!["Mean".into(), "Min".into()],
            activation_pool: vec!["ReLU".into(), "SELU".into()],
            standardize_op_pool: vec!["Identity".into()],
            informative_metrics: vec![Metric::custom("dummy", |_p, _y| Ok(0.0))],
            max_steps: Some(10_000),
            train_eval_split_ratio: Some(0.2),
            held_out_eval_rows: Some(256),
            culls: 0,
            run_elapsed_secs: 0,
            children_born_at_clock: std::collections::HashMap::new(),
            config: ConfigSnapshot::default(),
            trainer: None,
        })
    }

    fn tiny_topology() -> Topology {
        let mut topo = crate::graph::topology::Topology::new(
            0,
            Some(crate::graph::topology::TopologyOptions {
                topology_seed: 7,
                min_hidden_num_nodes: 1,
                max_hidden_num_nodes: 1,
                min_hidden_inputs_per_node: 1,
                max_hidden_inputs_per_node: 1,
                min_hidden_outputs_per_node: 1,
                max_hidden_outputs_per_node: 1,
                input_dim: Some(2),
                output_dim: Some(2),
                dropout_prob: 0.0,
            }),
        );
        topo.nodes.push(crate::graph::node::Node::new_input(0, 2));
        topo.nodes
            .push(crate::graph::node::Node::new_hidden(1, 2, 4));
        topo.nodes
            .push(crate::graph::node::Node::new_output(2, 4, 2));
        topo
    }

    #[test]
    fn engine_json_write_and_read_back_identical() {
        let dir = fresh_dir();
        let header = baseline_header("run-1");
        write_engine_json(&dir, &header).unwrap();
        let loaded = load_engine_json(&dir).unwrap();
        assert_eq!(loaded.run_id, "run-1");
        assert_eq!(loaded.run_seed, 42);
        assert_eq!(loaded.fitness_label.0, "f1");
        assert_eq!(loaded.fitness_direction, Direction::Maximize);
        assert_eq!(loaded.input_dim, 784);
        assert_eq!(loaded.output_dim, 10);
        assert_eq!(loaded.max_steps, Some(10_000));
        // Overwrite changes the header.
        let updated = baseline_header("run-2");
        write_engine_json(&dir, &updated).unwrap();
        let again = load_engine_json(&dir).unwrap();
        assert_eq!(again.run_id, "run-2");
    }

    #[test]
    fn net_state_round_trip_keeps_hash_and_seed_and_step() {
        let dir = fresh_dir();
        let topo = tiny_topology();
        let mut state = NetState::new(&topo, 0, Some("parent-a x parent-b".into())).unwrap();
        state.record_metrics(NetMetrics {
            step: 5,
            train_loss: 1.25,
            eval_loss: Some(1.40),
            fitness: 0.61,
            informative: vec![0.4, 0.6],
            frozen: false,
        });
        write_net_state(&dir, &state).unwrap();
        let loaded = load_net_state(&dir, &state.hash).unwrap();
        assert_eq!(loaded.hash, state.hash);
        assert_eq!(loaded.net_seed, 7);
        assert_eq!(loaded.step, 0);
        assert_eq!(loaded.entered_at_step, 0);
        assert_eq!(loaded.created_from.as_deref(), Some("parent-a x parent-b"));
        let m = loaded.last_metrics.as_ref().unwrap();
        assert_eq!(m.step, 5);
        assert_eq!(m.train_loss, 1.25);
        assert_eq!(m.eval_loss, Some(1.40));
        assert_eq!(m.fitness, 0.61);
        assert_eq!(m.informative, vec![0.4, 0.6]);
        // Topology round-trips exactly (topo_hash is canonical — same JSON ⇒ same hash).
        assert_eq!(loaded.hash, topo_hash(&loaded.topology));
        let rebuilt = loaded.topology().unwrap();
        let rebuilt_json = rebuilt.to_json().unwrap();
        assert_eq!(rebuilt_json, loaded.topology);
    }

    /// Schema separation: run-level facts live in the header's `config`, per-net
    /// facts in `nets/<hash>.json` — and neither invents a value for a mode that
    /// has none (RL has no shared stream ⇒ no batch geometry; the loss/update
    /// scheme is trainer-owned ⇒ no `loss_label` in the schema at all).
    #[test]
    fn schema_records_mode_appropriate_geometry_and_no_trainer_owned_fields() {
        // RL header: no stream, so geometry is genuinely absent (`null`), and
        // the step budget sits beside the fitness target in `config`.
        let rl_cfg = crate::engine::RaceConfig {
            mode: crate::engine::config::RunMode::Rl,
            max_steps: Some(25),
            ..Default::default()
        };
        let rl = ConfigSnapshot::from_config(&rl_cfg, None, None);
        assert_eq!(rl.batch_size, None);
        assert_eq!(rl.eval_batch_size, None);
        assert_eq!(
            rl.max_steps,
            Some(25),
            "step budget recorded with the stop criteria"
        );

        // Tabular header: the stream shape the run actually used is recorded.
        let tab =
            ConfigSnapshot::from_config(&crate::engine::RaceConfig::default(), Some(16), Some(128));
        assert_eq!(tab.batch_size, Some(16));
        assert_eq!(tab.eval_batch_size, Some(128));

        // In the written header, absent means absent — not a 16 default.
        let dir = fresh_dir();
        let mut header = baseline_header("rl-run");
        header.config = rl;
        write_engine_json(&dir, &header).unwrap();
        let v: serde_json::Value =
            serde_json::from_str(&std::fs::read_to_string(dir.join("engine.json")).unwrap())
                .unwrap();
        assert_eq!(v["config"]["batch_size"], serde_json::Value::Null);
        assert_eq!(v["config"]["max_steps"], serde_json::json!(25));

        // Per-net meta follows the same rule, and carries no trainer-owned knob.
        let topo = tiny_topology();
        let net = crate::graph::network::Network::build(&topo, crate::Device::CPU).unwrap();
        let mut state = NetState::new(&topo, 0, None).unwrap();
        state.stamp_meta(&net, &crate::engine::config::RunMetaCtx::default());
        assert_eq!(state.meta.batch_size, None);
        let net_json: serde_json::Value = serde_json::from_str(&state.to_json().unwrap()).unwrap();
        for absent in ["loss_label", "learning_rate", "grad_clip"] {
            assert!(
                net_json["meta"].get(absent).is_none(),
                "`{absent}` is trainer-owned (engine.json → trainer), not per-net schema"
            );
        }
    }

    #[test]
    fn overwrite_replaces_previous_step() {
        let dir = fresh_dir();
        let topo = tiny_topology();
        let mut state = NetState::new(&topo, 0, None).unwrap();
        state.record_metrics(NetMetrics {
            step: 0,
            train_loss: 2.0,
            eval_loss: Some(2.1),
            fitness: 0.1,
            informative: vec![],
            frozen: false,
        });
        write_net_state(&dir, &state).unwrap();

        state.advance_step();
        state.record_metrics(NetMetrics {
            step: 1,
            train_loss: 1.5,
            eval_loss: Some(1.6),
            fitness: 0.2,
            informative: vec![],
            frozen: false,
        });
        write_net_state(&dir, &state).unwrap();

        let loaded = load_net_state(&dir, &state.hash).unwrap();
        assert_eq!(loaded.step, 1);
        assert_eq!(loaded.last_metrics.as_ref().unwrap().fitness, 0.2);
    }

    #[test]
    fn race_state_insert_and_remove() {
        let mut st = RaceState::new();
        let topo = tiny_topology();
        let state = NetState::new(&topo, 0, None).unwrap();
        st.insert(state, Some(0.3));

        assert_eq!(st.live_count(), 1);
        let hashes = st.live_hashes();
        assert_eq!(hashes.len(), 1);
        assert_eq!(hashes[0].len(), 16);

        let removed = st.remove(&hashes[0]);
        assert!(removed.is_some());
        assert_eq!(st.live_count(), 0);
    }

    #[test]
    fn race_state_lifetime_tracks_appearances_and_span() {
        let mut st = RaceState::new();
        let topo = tiny_topology();
        // Insert a net at step 10 with an initial fitness — the lifetime
        // accumulator seeds from that entry call.
        let state = NetState::new(&topo, 10, None).unwrap();
        st.insert(state, Some(0.3));

        let h = st.live_hashes()[0].clone();
        // Each record_step advances the net's own step counter, and seeds or
        // updates the per-topology lifetime from the metrics' step and fitness.
        st.record_step(
            &h,
            NetMetrics {
                step: 10,
                train_loss: 0.0,
                eval_loss: None,
                fitness: 0.4,
                informative: vec![],
                frozen: false,
            },
        )
        .unwrap();
        st.record_step(
            &h,
            NetMetrics {
                step: 11,
                train_loss: 0.0,
                eval_loss: None,
                fitness: 0.5,
                informative: vec![],
                frozen: false,
            },
        )
        .unwrap();

        let (_, life) = st.lifetimes().into_iter().next().unwrap();
        assert_eq!(life.appearances(), 1);
        assert_eq!(life.first_step(), 10);
        assert_eq!(life.last_step(), 11);
        assert_eq!(life.span(), 1);
        assert_eq!(life.last_fitness(), Some(0.5));
        assert_eq!(st.live_count(), 1);
    }

    #[test]
    fn removed_net_file_is_gone() {
        let dir = fresh_dir();
        let topo = tiny_topology();
        let state = NetState::new(&topo, 0, None).unwrap();
        write_net_state(&dir, &state).unwrap();
        assert!(
            dir.join("nets")
                .join(format!("{}.json", state.hash))
                .exists()
        );
        remove_net_state(&dir, &state.hash).unwrap();
        assert!(
            !dir.join("nets")
                .join(format!("{}.json", state.hash))
                .exists()
        );
    }

    #[test]
    fn load_nonexistent_net_errors() {
        let dir = fresh_dir();
        let err = load_net_state(&dir, "deadbeef").unwrap_err();
        assert!(err.to_string().contains("cannot access"));
    }
}