[EzraBrooks]

Pull requests, and by hand for now releases, build in vinca-gha's dependency stages, one job per batch of recipes. Today each distribution and platform gets one six-hour job that needs re-running. Releases use the same workflow and only add the uploads, so the build branches can retire once a release run has uploaded correctly.

How (staged-build.yaml, generated by tools/staged_build.py workflow; rs check verifies it is up to date):

• plan (one job per distribution × platform): generates the recipes once (artifact), restores the PR cache (PRs only), leaves out the recipes whose every variant is cached, and lists the jobs with vinca-gha --json (vinca-gha --json: write the batched stages instead of a workflow vinca#201). Each job carries needs, the earlier jobs building its requirements.
• schedule: puts every plan into one shared set of 64 stage slots. Slot k runs stage k of every distribution and platform as one matrix, capped at 256 jobs by merging the smallest batches. It also writes the schedule each job reads its recipes from. A run is about 20 plan jobs + schedule + 64 slots + 20 collect jobs, no longer 20 × 66.
• stage_k → stage-job.yaml: downloads only the artifacts of the jobs it needs (directly and transitively, plus the restored cache, with one download-artifact brace pattern; no REST API calls, so no rate limit and fork PRs work), indexes them as a local channel (ROBOSTACK_LOCAL_CHANNEL in rattler_build()), builds with rs build-ci, and uploads pkgs-<distro>-<platform>-<job> with a status/<job> file.
  • A job whose requirements failed fails at once. Slots don't use !failure(), so other distributions and platforms keep building.
  • "Re-run failed jobs" re-runs every later slot; jobs that already succeeded find their own artifact and skip the build.
  • In release mode the job also uploads to the channels (rs upload: prefix.dev trusted publishing, plus the distribution's anaconda token via secrets[...]).
• collect (PRs only, per distribution × platform): saves the packages, now including noarch/, as the PR cache, then runs the ABI check and the rs gha check (while the build branches remain).
• release.yaml: workflow_dispatch with optional distributions and platforms, calling the same workflow with upload: true.

Notes

• vinca is pinned to EzraBrooks/vinca@0a04e39: vinca-gha --json: write the batched stages instead of a workflow vinca#201 cherry-picked onto the current pin. Repin after it merges.
• Before a release run, prefix.dev's trusted publishing (Repository Access) must allow .github/workflows/stage-job.yaml; today it allows the build branches' build.yaml.
• Slots are a fixed ceiling (the workflow can't change once a run starts). Lyrical's full linux-64 build has 53 stages; stages past slot 63 fold into one job.
• Downloading only the needed artifacts saves less than hoped: a lyrical job needs 115 of the 135 earlier jobs on average, because almost everything depends on the core packages.
• The PR cache key changes (staged-…), so open PRs start from an empty cache once.
• tools/ci_build.py (the timeout wrapper) is no longer used by PR builds.