feat: probar experiment — ML training experiment tracker with early stopping

[noahgift]

Problem

ML training exploration in bashrs (Shell Safety Classifier) consumed significant GPU time:

• 6 training runs on RTX 4090
• 3 kill criteria triggered
• Peaked at 32% accuracy
• HuggingFace model ultimately deleted
• Root cause (data distribution mismatch) identified late

Proposed Solution

probar experiment init "SSC v11 training"
probar experiment run --config train.toml --budget 2h-gpu
probar data-audit training_data.jsonl        # Pre-flight data quality check
probar experiment status                      # Live training dashboard
probar experiment compare run-1 run-6        # Compare runs

Requirements

1. Pre-flight data audit: Check label distribution, detect class imbalance, validate domain coverage before burning GPU time
2. Budget gates: Set GPU-hour or cost limits per experiment
3. Early stopping: Monitor validation loss curve, stop when plateau detected (save remaining GPU budget)
4. Experiment tracking: Store hyperparameters, metrics, and artifacts per run
5. Kill criteria evaluation: Check accuracy thresholds at intermediate checkpoints, not just at the end
6. Cost reporting: Track actual GPU-hours and estimated cost per experiment

Acceptance Criteria

• probar data-audit detects class imbalance >3:1
• probar data-audit detects domain mismatch between train/eval
• Budget gate prevents exceeding GPU-hour limit
• Early stopping triggers on validation loss plateau (patience=3)
• Kill criteria checked at each checkpoint, not just end of training
• Experiment comparison shows metric diff across runs

Cross-Reference

Filed from paiml-mcp-agent-toolkit#258 as part of sovereign stack tooling gap analysis.

[noahgift]

Triage: P3 — feature request. ML training experiment tracker with early stopping. Nice-to-have enhancement.