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
//! The engine — the step-race NAS loop.
//!
//! One global step clock; every live net trains and evals on the same shared
//! batch each step. Divergence over the population's smoothed fitness culls
//! the worst net and births a caught-up replacement. The old generational
//! engine was removed for 1.0 — [`RaceEngine`] is the only engine.
//!
//! Module layout (one file per concern):
//! - [`config`] — `RaceConfig` knob surface, `StopReason`, divergence/stop
//! closure types, `RaceSnapshot`, `RunMetaCtx`
//! - [`population`] — initial population generation from the config's pools
//! - [`child`] — child generation (roulette → crossover → mutate) + catch-up
//! - [`smoothing`] — the rolling fitness buffer + its mean (ranking inputs)
//! - [`core`] — the `CoreEngine` itself: construction, the run loop,
//! per-net stepping, culling, persistence hooks
//! - [`logging`] — the `tracing` subscriber + its three sinks (console,
//! `Minimal` frame, opt-in `telemetry.jsonl`)
//! - [`reporting`] — the user-facing log lines (per-step rollup, `Minimal`
//! frame, stop summary)
//! - [`format`] — display/CSV formatting + the build identity stamp
//! - [`artifacts`] — champion/worst dumps + their ranking helpers
//! - [`checkpoint`] — the checkpoint ledger, gate bar, and surprise exam
//! - [`evolve`] — the crossover/immigrant rolls, cull/insert, victim selection
//! - [`pruner`] — the post-race solo-training phase
//! - [`step`] — one net's step path (trainer dispatch, freeze pass, RNG seeding)
//! - [`stop`] — stop criteria + `RaceSnapshot`
//! - [`tabular_engine`] — tabular-mode constructors (`resume`)
//! - [`rl_engine`] — RL-mode constructors (`resume_rl`)
//! - [`fitness`] — `Fitness`, `Direction`, `Metric`, `FitnessLabel`
pub use ;
pub use ;
pub use CoreEngine;
pub use ;
pub use RlEngine;
pub use ;
pub use TabularEngine;