Fix publication OPE support with experiment cross-fitting

[vtavakkoli]

Summary

Fix the publication blocker exposed by the corrected independent evaluation: the earlier configuration-level policy/OPE split left only ~1% seen and ~6% unseen evaluator support for proposed actions. This PR replaces that partition with experiment-level role-swapped cross-fitting and constrains all counterfactual methods to common policy/OPE action support.

Scientific fix

Experiment-level role swap

For each COMMAG experiment:

• proposed policy + FQI/CQL/HGB are fitted on tr0–tr11 from that experiment;
• the independent OPE evaluator is fitted on tr0–tr11 from the opposite experiment(s);
• validation from the policy experiment calibrates thresholds/CQL/hysteresis;
• seen/unseen final-test rows from that experiment are evaluated exactly once.

With two experiments this is:

Fold A: policy/baselines <- exp1 | OPE <- exp2
Fold B: policy/baselines <- exp2 | OPE <- exp1

This preserves all training configurations on both evidence roles instead of sacrificing action coverage by holding out whole configurations.

Positivity/common-support constraint

Before calibration/evaluation, OfflinePolicyBundle.actions_by_slice is restricted to the intersection of actions observed in policy-fit and OPE-fit data for each slice. The same support restriction therefore automatically applies to:

• proposed controller selection;
• FQI action selection;
• linear CQL action selection;
• HGB behavior baseline normalization.

The primary statistics can no longer contain counterfactual actions unsupported by the independent evaluator.

Readiness now also requires:

• =80% of policy action cells retained by common support;

• =2 common actions per slice;

• = configured evaluator-support gate on seen and unseen tests.

This prevents a trivial one-action collapse from falsely satisfying the support gate.

Correct scientific framing

The report now explicitly limits claims when evidence is weak:

• zero within-episode scheduler+PRB changes => frame as offline RAN configuration/policy selection, not demonstrated dynamic closed-loop control;
• OOD AUROC <0.70 => OOD remains a secondary limited-discrimination diagnostic;
• p95 host/container latency >500 ms => no real-time RIC claim;
• negative clustered utility CI => no utility-superiority claim; report utility-vs-churn/safety trade-off instead.

A new tradeoff_summary.json exports utility delta, churn reduction, SLA delta, and clustered effects for each baseline.

HTML report

The self-contained report.html now adds:

• experiment role-swap explanation;
• common-support/positivity table per fold;
• allowed manuscript claims and restrictions.

No external assets are introduced.

Tests

The synthetic publication integration test now executes the production cross-fit path and verifies:

• two experiment role-swap folds;
• disjoint policy/OPE experiments;
• 100% common support in the synthetic fixture;

• =95% evaluator support on seen/unseen data;

• generated cross-fit HTML report and trade-off artifact.

Run

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

The full COMMAG benchmark must be rerun after this PR. Previous numerical results must not be reused because the evidence partition, action support, and final counterfactual estimates change.