[TohruSu]

AdaptiveQueue can now train a tabular agent to choose FIFO, SPT, EDF, or Priority for each scheduling decision and evaluate the saved policy on held-out workloads. Existing baseline commands and the public simulate(jobs, policy) API retain their behavior.

• Extract a shared Simulation engine that exposes the current ready queue/time and executes one non-preemptive job per step. Both complete episodes and learning use the same arrival/time advancement logic.
• Encode up to 96 states from queue length, minimum slack, short-job proportion, and high-priority presence. Slack is deadline minus current time minus duration; exact thresholds are documented.
• Use each completed job's negative priority-weighted cost as reward, configurable Q-Learning with gamma=1 by default, and episode-based epsilon decay.
• Add readable, versioned JSON persistence with formulation/training metadata, plus train/evaluate CLI commands. Evaluation is greedy, leaves Q-values and RNG state unchanged, rejects overlapping training seed ranges, and reports pooled job metrics.

Validation on Python 3.12: all 74 pytest cases passed, including the 39 existing baseline tests; Ruff lint and format checks, strict mypy targeting Python 3.10, and whitespace checks passed. The documented 1,000-episode training/save/load workflow and 20 held-out evaluation episodes ran successfully. No runtime dependencies were added.

The coarse state can alias different queues, so training does not guarantee improvement over a baseline. Unseen states use FIFO through deterministic action tie-breaking. Saved models preserve evaluation predictions; exact continuation of an interrupted training RNG stream is outside this change.