Evergreen: skip agent and conclusion when no PR needs attention

[Copilot]

At every-5-min cadence, Evergreen runs the agent even when no open PR is sick. The agent responds with prose instead of calling noop, gh-aw treats the empty safe-outputs as failure, and a fresh false-failure issue gets filed every run.

Fix: surface the pre-step's selection as a step output and gate the expensive/noisy downstream work on it.

.github/workflows/evergreen.md (source)

• Add id: find-pr to the pre-step.
• Add an emit_selected_output(pr_number) helper invoked at every exit path so selected is always written to $GITHUB_OUTPUT (empty when nothing to do, PR number otherwise).

def emit_selected_output(pr_number):
    gh_output = os.environ.get("GITHUB_OUTPUT")
    if gh_output:
        with open(gh_output, "a") as f:
            f.write(f"selected={'' if pr_number is None else pr_number}\n")

.github/workflows/evergreen.lock.yml (mirrors source so the fix is effective without waiting for a gh-aw recompile)

• Same id + emit changes inside the embedded python.
• Promote selected to an agent job output: selected: ${{ steps.find-pr.outputs.selected }}.
• Gate the Execute GitHub Copilot CLI step with if: steps.find-pr.outputs.selected != '' — skips the Copilot invocation when idle.
• Extend the conclusion job condition with && (needs.agent.outputs.selected != '' || needs.activation.outputs.lockdown_check_failed == 'true'). The conclusion job hosts Handle Agent Failure (the step that files the false issue); gating it here is what stops the noise. Lockdown-failure escape hatch preserved.

Behavior

• Idle: Copilot step skipped, conclusion job skipped → no failure issue filed, no Copilot spend.
• Sick PR: unchanged end-to-end.
• Forced PR via workflow_dispatch: unchanged; selected is populated as today.

Note

The .lock.yml edit duplicates the .md change so the fix takes effect immediately. A subsequent gh-aw recompile will reproduce the same result from source.