feat(aeh): MLflow logging, tool/MCP checks, and advisory regression

[GuyZivRH]

Summary

Wire Agent Eval Harness (AEH) observability and advisory checks into CI/monitoring: MLflow logging after evaluate, tool/MCP input validation, optional non-blocking regression and harness inventory, plus shared-namespace MLflow deploy defaults.

Changes

• Add post-evaluate MLflow logging (scripts/log_aeh_mlflow.py) with log-results, push-feedback, and optional sync-dataset (schema mapping inferred from input.yaml:prompt when present)
• Fall back to a minimal logger when upstream log_results.py skips because mlflow is missing; ensure mlflow is importable in the evaluate step
• Harden AEH scoring: avoid null judges mapping to reward 1.0, bake LiteLLM judge model into task packages, set reward score_range to [0, 1] for LLM judges
• Preserve Harbor tool/MCP interception artifacts; validate inputs.tools / handlers in validate.py
• Add advisory (non-blocking) AEH regression (score.py regression) and harness inventory (eval-check) via new scripts and evaluate/test task steps
• Expose pipeline params: enable-aeh-regression-check, aeh-baseline-run-id, enable-aeh-eval-check (CI); pass AEH-related params through monitoring pipelines
• Add MLflow OpenShift manifests under config/mlflow/ and retarget defaults to the shared ab-eval-flow namespace
• Sample aeh-hello-world updates (eval config, skill package) plus unit tests for the new paths

Test plan

• python -m pytest — 1100 passed
• Manual: AEH evaluate run logs MLflow results (and push-feedback / sync-dataset when mapping available)
• Manual: advisory regression / eval-check emit reports without failing the PipelineRun
• Optional: set aeh-baseline-run-id and confirm regression compares against that run

Related

• Builds on AEH engine integration (APPENG-5300)
• Harbor/AEH MLflow traces work upstream (agent-eval-harness)