Fix publication benchmark for Python 3.13 and remove legacy PPO runtime

[vtavakkoli]

Summary

Follow-up to merged PR #66. This PR fixes the publication workflow after the historical TensorFlow PPO container failed to build and makes the benchmark run entirely on the normal Python 3.13 image.

Fixes

• switches the main Docker builder/runtime from python:3.12-slim to python:3.13-slim
• adds Python 3.13 package metadata and CI coverage
• removes Dockerfile.ppo-legacy
• removes scripts/export_original_ppo.py
• removes the original-ppo Compose service and dependency
• changes the publication chain to prepare-full-commag -> publication-test
• adds agentic_ran.publication as the clean publication entrypoint
• explicitly prevents stale original_ppo_actions.csv.gz files from being picked up by the older helper
• keeps proposed/FQI/linear-CQL/HGB, OOD generalization, shortcut diagnostics, and paired statistics unchanged

Original COMMAG PPO handling

The original PPO is no longer executed. Instead, the repository stores a literature-reference record from Bonati et al., IEEE Communications Magazine 59(10), 21-27, 2021, DOI 10.1109/MCOM.101.2001120, arXiv:2012.01263.

The paper reports:

• eMBB spectral-efficiency gain up to 20% over the best-performing static scheduler
• URLLC average-buffer reduction of 37% vs RR, 5% vs WF, and 17% vs PF

These values are written to results/publication/literature_reference.json and explicitly excluded from publication_baselines.csv and paired statistical tests because they are paper-reported results under different metrics/test conditions, not reproduced results.

Run

docker compose -f docker-compose.publication.yml up --build publication-test

The workflow now requires only the standard Agentic-RAN image and no TensorFlow/TF-Agents legacy environment.

Expected outputs

• publication_baselines.csv
• publication_decisions.csv.gz
• ood_generalization.json
• policy_shortcut_test.csv
• paired_statistics.json
• literature_reference.json
• publication_summary.json

Validation

CI now includes Python 3.13. The Docker smoke job builds from python:3.13-slim, so the same runtime family used by the publication workflow is exercised in CI.