Address reviewer baseline and communication-cost concerns

[TomasOrtega]

Summary

This PR addresses the reviewer-requested experiment, communication-cost, manuscript, and rebuttal work on the review-response-experiments branch.

Changes now include:

• Decentralized adaptive-gradient baselines: AdaGrad, RMSProp, Adam, AdamW, momentum, and Nesterov.
• Communication-cost accounting in run_simulation, including DECO-i (wealth, G) versus DECO-ii G-only messages.
• A synthetic review baseline experiment: experiments/run_review_baseline_experiment.py.
• A real-data review experiment on offline-packaged real datasets: experiments/run_review_real_data_experiment.py.
• Offline-friendly real-data environments in RealDataEnvironment for diabetes, breast_cancer, digits, wine, and linnerud, while preserving the existing LIBSVM datasets.
• Deterministic Erdos-Renyi graph generation via a seed argument.
• Workflow updates so run_all.sh and run_all.bat include both review experiments.
• Manuscript tex updates: tex/main.tex now injects focused reviewer-driven additions into the experiments section while preserving the original manuscript body as tex/main_original.tex.
• Rebuttal updates: tex/rebuttal/review_response.tex now directly answers each reviewer concern with communication accounting, adaptive baselines, real-data results, DECO-i/DECO-ii communication normalization, notation/potential-function caveats, and convex-scope limitations.

Validation performed

I ran the review experiments locally from a reconstructed working tree because the execution container could not clone from GitHub directly.

Synthetic review run:

PYTHONPATH=. python -u -m experiments.run_review_baseline_experiment --T 100 --seeds 0 --topologies cycle erdos_renyi

Real-data review runs, split per dataset to avoid the tool timeout:

PYTHONPATH=. python -u -m experiments.run_review_real_data_experiment --T 100 --seeds 0 --datasets diabetes --topology cycle
PYTHONPATH=. python -u -m experiments.run_review_real_data_experiment --T 100 --seeds 0 --datasets breast_cancer --topology cycle
PYTHONPATH=. python -u -m experiments.run_review_real_data_experiment --T 100 --seeds 0 --datasets digits --topology cycle

I also syntax-tested the manuscript wrapper mechanism with a local minimal LaTeX fixture to confirm that the added review_revision_*.tex files are inserted at the intended subsection hooks.

The results support a more careful rebuttal: the adaptive baselines are competitive or stronger in these short-horizon/tuned settings, while DECO-i generally outperforms DECO-ii at the cost of one additional scalar per directed neighbor exchange. The text therefore avoids overclaiming empirical dominance and emphasizes the parameter-free learning-rate guarantee, explicit communication accounting, and convex-online-learning scope.

Remaining note

The branch now contains tex/ as ordinary files rather than a submodule. I have not merged this PR.