RL CLI Helpers (API)

TorchLean's runnable RL examples (NN/Examples/Models/RL/*) share one CLI shape:

• --updates <n>: how many update iterations to run,
• --eval-every <n>: evaluate every n updates,
• --eval-episodes <n>: number of evaluation episodes per checkpoint,
• --eval-max-steps <n>: max steps per evaluation episode,
• --log <path|off|none|false>: where to write the widget-friendly TrainLog JSON.

This module centralizes that parsing so we don't duplicate the same flag boilerplate across CartPole/Pong/GridWorld examples.

structure NN.API.rl.cli.PpoFlags : Type

Parsed PPO-style training flags shared by multiple runnable examples.

• updates : ℕ
• evalEvery : ℕ
• evalEpisodes : ℕ
• evalMaxSteps : ℕ
• log : Runtime.Training.LogDestination
• logPath : System.FilePath

def NN.API.rl.cli.instReprPpoFlags.repr : PpoFlags → ℕ → Std.Format

@[implicit_reducible]

instance NN.API.rl.cli.instReprPpoFlags : Repr PpoFlags

def NN.API.rl.cli.parsePpoFlags (exeName : String) (args : List String) (defaultLogPath : System.FilePath) (defaultUpdates defaultEvalEvery defaultEvalEpisodes defaultEvalMaxSteps : ℕ) : Except String (PpoFlags × List String)

Parse PPO-style shared flags.

Notes:

• --log off|none|false disables writing the JSON artifact but still returns the resolved default logPath (useful for printing consistent banners).
• We treat 0 as invalid for the update/eval counts because a “no-op” run usually indicates a CLI mistake.