OpenEvolve: gate iteration comments on CI and measure fitness in CI

[Copilot]

OpenEvolve iteration comments on tsb-perf-evolve were posted with ⏳ Pending CI / Metric: pending CI (sandbox has no bun) and never reconciled, because the strategy playbook posts before CI runs and the sandbox can't install bun to measure fitness itself. CI also didn't run the benchmark, so even a green build left fitness: null on every population entry.

Fix 2 — measure fitness in CI

• Extracted ## Evaluation from .autoloop/programs/tsb-perf-evolve/program.md into an executable evaluate.sh next to it; program.md now just calls bash evaluate.sh. Single source of truth shared between agent and CI.
• New benchmark job in .github/workflows/ci.yml, gated on autoloop/*-evolve branches: Bun + Python + pandas, runs evaluate.sh, uploads bench-result.json, and creates an OpenEvolve benchmark check-run titled fitness=<num> (or fitness=null + neutral if the benchmark itself errored). Fitness/result are passed to actions/github-script via env, not expression interpolation. Workflow gets checks: write.

Fix 1 — make iteration comments reconcile with CI

• Added ### Step 6.5. Wait for CI between Steps 6 and 7 in .autoloop/strategies/openevolve/strategy.md. Resolves the PR (existing_pr → gh pr list --head fallback), blocks on gh pr checks --watch --interval 30 --fail-fast, classifies with the same awk classifier as the generic Step 5a, then reads fitness from the check-run:

  fitness=$(gh api "repos/${GITHUB_REPOSITORY}/commits/${SHA}/check-runs" \
    --jq '.check_runs[] | select(.name == "OpenEvolve benchmark") | .output.title' \
    | sed -n 's/^fitness=//p' | head -n1)

• Branching: success → record numeric fitness, post ✅ Accepted; failure → defer to the existing Gate autoloop iteration acceptance on CI green with a bounded fix-retry loop #176 fix-retry loop, on exhaustion post ❌ Rejected / ⚠️ Error with pause_reason: ci-fix-exhausted; pending (60-min wall-clock cap fired) → record status: pending-ci and leave a single reconciliation-pending comment the next iteration is allowed to overwrite.

• Step 6 reframed as a cheap pre-filter only — null sandbox metric is no longer grounds for rejection.

• Step 8 requires the iteration comment use the terminal Step 6.5 status and the Iteration History fitness come from the check-run, never from the sandbox. ⏳ Pending CI is explicitly forbidden as a permanent terminal state.

Notes

• Stale comments on [Autoloop: tsb-perf-evolve] #189 / [Autoloop] [Autoloop: tsb-perf-evolve] #190 are left as historical record per the issue.
• Firewall allowlist for releaseassets.githubusercontent.com remains a separate concern; CI runners are unrestricted, so this PR works without it.