
Split a single long job into a fan-out/fan-in graph: a build job that works out the version once and publishes it as a job output, three check jobs fanning out from it, a packaging job fanning them back in and reading the version through needs, and a job outside the chain that cannot see the producer's output. One nominated check fails per grade, so the failing sibling cannot be predicted.
Add the failure paths to a pipeline: a notification gated with failure() that names which check failed by reading needs.<job>.result, a job gated on one specific needs result, an always-run cleanup, and an optional job whose probe step is allowed to fail without ending the job. Which check fails is nominated per grade, so the notification cannot hardcode a name.
Assemble a release train from ten jobs: a build job publishing the version, three checks fanning out and publishing attestations, a release gated on success() that binds the version and all three attestations into a release token, a non-blocking optional job whose probe step may fail, a notification gated with failure() that names the failed check from needs results, a job gated on one specific needs result, an always-run cleanup, and a job outside the chain that cannot see the producer's output.