feat: run classic Harbor on stock 0.20.0 + OpenShift custom env

[ikrispin]

Summary

• Migrate classic Harbor A/B (eval_engine: harbor) off the skills_eval_corrections fork onto pinned upstream Harbor (harbor==0.20.0) plus the AEH-style custom OpenShift environment (abevalflow.harbor_extensions.openshift_environment:OpenShiftEnvironment).
• Prebuilt CI now hands off trial images via task.toml docker_image and selects the env with environment.import_path / --environment-import-path (not environment.type: openshift or fork kwargs.image_ref).
• Update eval-base (Dockerfile.base + build_base_image.sh) to install stock Harbor + pinned AEH k8s env; remove harbor-fork-url / harbor-fork-revision pipeline params; refresh docs and unit tests.

Closes / tracks: APPENG-5914 (parent APPENG-4901).

Motivation

Classic Harbor CI depended on a Harbor fork for OpenShift. AEH already runs stock Harbor with a custom env plugin. This aligns classic A/B with that model so we can track upstream Harbor without maintaining the fork or adopting upstream’s oc/SCC OpenShift backend.

What changed

Area	Change
Config	generate_eval_config.py: prebuilt → import_path + docker_image; local-build → type: docker
Contract	Shared OPENSHIFT_ENVIRONMENT_IMPORT_PATH in abevalflow/harbor_extensions
Images	templates/Dockerfile.base: harbor==0.20.0, AEH SHA ff8b8301…, bake abevalflow
Pipeline	evaluate / harbor-eval: PyPI Harbor fallback + --environment-import-path; drop fork params
Docs	New Docs/harbor_custom_env.md; fork/OpenShift docs redirected
Tests	Updated test_generate_eval_config.py; added test_openshift_environment.py

Out of scope (per ticket): upstream type: openshift (oc/SCC), harbor run -d, dropping DB column harbor_fork_revision.

Test plan

• Unit tests for config generation + OpenShiftEnvironment import path
• Rebuild/push eval-base:latest and eval-base:local-env with Harbor 0.20.0
• Unbaked cluster smoke (pipeline clones feature branch; no baked sources image):
  • Namespace: itay-evalflow
  • Submission: hello-world / Harbor / prebuilt
  • Branch: feat/harbor-custom-env from https://github.com/ikrispin/ABEvalFlow.git
  • Result: Succeeded (prepare → evaluate → analyze → store → cleanup)
  • Verified in logs: Harbor 0.20.0, docker_image set, --environment-import-path, no skills_eval_corrections, treatment/control mean reward 1.0
  • Run: https://console-openshift-console.apps.cn-ai-lab.2vn8.p1.openshiftapps.com/k8s/ns/itay-evalflow/tekton.dev~v1~PipelineRun/harbor-unbaked-smoke-59p2g/
• After merge: apply updated Tasks/Pipelines in shared ab-eval-flow and confirm monitoring CI still healthy
• Optional follow-up: replace deprecated --environment-import-path with Harbor’s --env flag

Notes for reviewers

• Feature branch is based on origin/main (Harbor-only commit d6cb164).
• Smoke used namespace-local apply tweaks (litellm URL, image refs, scheduling) that are not part of this PR.

Successful run: