[TohruSu]

Add a fixed, paired experiment comparing three independently trained Q-Learning models with FIFO, SPT, EDF, Priority, and seeded Random. Run it with adaptivequeue benchmark --output-dir benchmark-results; cached models are validated against the protocol, and --retrain regenerates them.

Models train on mixed workloads for 1,000 episodes at seeds 42, 2042, and 4042 with existing defaults. Evaluation covers stable, bursty, and mixed workloads at seeds 20000–20029, with each job set generated once and reused for all policies. RL evaluation is greedy and leaves Q-values and RNG state unchanged.

The committed reference run includes 720 per-episode metric rows, 24 summary rows with paired cost differences/win rates, 36 action-frequency rows (including contended-only frequencies), both figures, three saved Q-tables, and protocol metadata. Summary statistics average episode metrics, including episode P95 waiting time; training runs remain separate.

Observed mean primary cost J (lower is better):

Policy / training seed	Stable	Bursty	Mixed
Priority	9.852	30.867	9.212
Q-Learning / 42	9.270	28.804	9.284
Q-Learning / 2042	10.164	31.471	10.402
Q-Learning / 4042	9.210	27.130	9.900

Priority has the best classical mean cost in each workload. Two RL runs improve on it in stable and bursty workloads, but every RL run loses to Priority on mixed. SPT achieves the lowest deadline-miss rate in every workload. docs/experiments.md includes all five baselines, paired comparisons, action mixes, reproduction details, and the limitations of three training seeds at one load/deadline setting.

Validation: inspected six preliminary baseline episodes before the fixed final run; all 81 pytest tests pass, along with Ruff lint/format, strict mypy targeting Python 3.10, and whitespace checks. Fresh/cached reproducibility, identical evaluation job objects, aggregation, frozen evaluation, and figure generation are tested. Matplotlib is an optional benchmark dependency; existing baseline and RL CLI tests remain passing.