feat(self-improve): generalize ReST-EM loop to mlxlm backend + agent-task scenarios

[jayscambler]

What

The self-improving loop (autoctx self-improve, ReST-EM / expert iteration) was hardcoded to the from-scratch mlx backend and game scenarios. This generalizes it along both axes — the two follow-ons chosen after closing the recursive loop.

Backend: mlx → {mlx, mlxlm}

The loop now drives either SFT backend: mlx (from-scratch GPT) or mlxlm (LoRA on a pretrained base, --backend mlxlm). mlxlm's in-scenario assessment now collects {strategy, score} samples to a collect_samples_path (the adapter analogue of the existing mlx collect path), so ReST-EM can keep the elite and retrain. train.py permits collect_samples_path for mlx+mlxlm and rejects it for the online-RL/distillation backends (grpo/opd/trl), which have no SFT sample stream.

Scenarios: game → game + agent-task

Agent-task scenarios (scored via evaluate_output on free text) are now first-class. _assess_mlxlm collects the raw text as the sample's strategy, and records_to_completions uses a string strategy verbatim as the completion (json.dumps would quote/escape it). Game scenarios (JSON strategies) are unchanged.

CLI

autoctx self-improve gains --backend, --base-model, --fine-tune-type, --num-layers, and --batch-size (lower it for mlxlm on small seeds: mlx-lm requires the validation split to hold ≥ batch_size examples).

Verified live

autoctx self-improve --backend mlxlm on grid_ctf (cached Qwen2.5-0.5B), 2 rounds: dataset grew 24→28, per-round avg_score 0.755 → 0.878. The loop collects samples, filters elite, grows the dataset, and retrains the adapter each round — end to end through the CLI.

Tests (CI-safe, no mlx)

records_to_completions verbatim-text vs JSON-strategy; loop rejects non-SFT backends; loop threads backend + adapter params into run_training (mocked); collect_samples_path rejected for non-SFT backends. Updated the prior MLX-only guard test to the new mlx and mlxlm message. Documents the new flags + backend/scenario coverage in mlx-training.md. ruff + mypy clean.