feat: add experimental BinEval evaluation support

[Copilot]

Adds native BinEval-style evaluations to gh-aw — small, binary questions declared in workflow frontmatter, executed post-run via an LLM harness, with results aggregated and reported as CI artifacts.

Schema (evals frontmatter)

• New optional evals array with id + question fields; validated for unique IDs and non-empty questions
• Emits experimental warning at compile time

evals:
  - id: builds
    question: Does the generated code compile?
  - id: focused
    question: Is the implementation limited to the requested change?

Evaluation model

• EvalDefinition, EvalResult, EvalSummary types in frontmatter_types.go
• WorkflowData.Evals []EvalDefinition for downstream consumers

Eval job

• New eval job injected after agent + detection jobs in the compiled workflow
• JS harness (eval_harness.cjs) calls GitHub Models API (gpt-4o-mini) per question independently — no MCPs, no checkout
• Prompt generation produces per-question binary prompts with rationale; no holistic scoring
• Results aggregated (total/passed/failed/pass-rate) and uploaded as a eval artifact with a markdown step summary

Not included

Phase 8 (persisting results to a git branch, à la experiments) is deferred.

[pelikhan]

@copilot run inference in AWF and use /reflect to find a suitable inference endpoint.

[github-actions]

Hey @Copilot 👋 — great work on the BinEval evaluation support! The end-to-end shape is clean: the evals frontmatter schema, the Go compiler module, the JS harness, and the wiring into the orchestrator all land as a cohesive unit.

A few things that stand out positively:

• Well-scoped: every changed file is in service of the same feature — nothing unrelated was mixed in.
• Tests at every layer: compiler_evals_test.go (Go unit tests), eval_harness.test.cjs (JS unit tests), and evals_experimental_warning_test.go (integration guard) give solid coverage across the stack.
• Clear description: the PR body explains the schema, the evaluation model, the harness design, and explicitly calls out what was deferred (Phase 8 persistence) — exactly the context reviewers need.
• Experimental gating: emitting a compile-time warning via emitExperimentalFeatureWarnings is the right pattern for a feature that isn't production-ready yet.

This looks ready for review. 🚀

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• patchdiff.githubusercontent.com

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "patchdiff.githubusercontent.com"

See Network Configuration for more information.

Generated by ✅ Contribution Check · 293.1 AIC · ⌖ 20.4 AIC · ⊞ 6K · ◷

[pelikhan]

@copilot keep working