safe-outputs.jobs custom jobs are not gated on threat detection (built-in safe_outputs is) #69404

Analysis

Custom safe-output jobs (safe-outputs.jobs.<name>) run even when threat detection fails; the built-in safe_outputs job does not.

buildSafeJobs in pkg/workflow/safe_jobs.go (v0.92.3) adds needs: [agent, detection] with the comment "so that the condition can gate on needs.detection.result == 'success'", but only adds that gate when IsConditionalDetection(...) is true. With ordinary always-on detection the compiled condition is:

(!cancelled()) && needs.agent.result != 'skipped' && contains(needs.agent.outputs.output_types, '<job>')

!cancelled() lets the job run after a failed detection job, so any custom write (Slack, an external API) bypasses detection. safe_outputs uses ... && needs.detection.result == 'success'.

Reproduce: compile any workflow with a safe-outputs.jobs entry and default detection; compare jobs.<name>.if with jobs.safe_outputs.if in the lock.

Implementation Plan

1. pkg/workflow/safe_jobs.go (buildSafeJobs): when threatDetectionEnabled and not IsConditionalDetection, AND the base condition with buildDetectionSuccessCondition() (same gate as safe_outputs).
2. Pass GH_AW_DETECTION_CONCLUSION: ${{ needs.detection.outputs.detection_conclusion }} to custom job steps, as the built-in job receives it.
3. Tests next to the existing safe-jobs tests: the condition contains needs.detection.result == 'success' when detection is enabled, and is unchanged when detection is disabled or expression-controlled.
4. Run make agent-finish before completing.

Acceptance Criteria

With threat detection enabled, a failed detection job skips every safe-outputs.jobs.* job, as it skips safe_outputs; the new tests pass.