ci: run Strix behavioral evals through OrchestrAI

[saman-amd]

Summary

Run the existing Linux and Windows Strix Halo behavioral evals through OrchestrAI while preserving the workflow name, required evals / results check, per-skill verdicts, and existing Instinct path.

Details

• Group selected Strix cases into one Linux and one Windows OrchestrAI run.
• Test the immutable PR head commit and require enable_orchestrai_ci before using hardware.
• Use the Playbooks-compatible ORCHESTRAI_CONTROL_RUNNER and ORCHESTRAI_WAIT_RUNNER variables, with ubuntu-latest as the fallback.
• Keep routing on ubuntu-latest; unrelated label events are no-ops.
• Add bounded retries, cancellation handling, sanitized verdicts, and ReportPortal View Results links.
• Support live and allocation-free mock runs through manual dispatch.
• Add the controller, configuration, reporting helpers, tests, and documentation without changing skills, eval datasets, test-library code, or shared pipeline code.

Validation

• 54 controller, plan, reporting, and workflow-contract tests pass.
• Python compilation and diff checks pass.
• Trusted Linux and Windows live validation completed, and all 10 per-test artifacts included a ReportPortal link.

Authenticated live checks require a maintainer-owned branch because GitHub does not provide target-repository secrets to fork pull requests.

[danielholanda]

Hi @saman-amd, can you please create this PR from a branch instead of a fork, so the workflows can pick up the right credentials?

[saman-amd]

Superseded by #247, which uses a trusted branch in amd/skills so Actions can receive repository credentials.