# setup-acceptance-model contract

This composite action only orchestrates preparation of one pinned local model and its owned daemon.
Do not call it on a shared workstation. It uses fresh `RUNNER_TEMP` storage
and the disposable hosted runner's user font/tool directories. Third-party
archives and model blobs must retain their exact source-pinned hashes and lengths.

The cache contains only the fixed manifest and model blobs. Cache hits never
bypass byte-length, SHA-256 or regular-file checks; corrupt or symlinked entries
fail before daemon startup rather than being replaced. Runtime/daemon custody
and inference records always remain in the fresh owned directory.

The model action binds an unoccupied loopback port, disables cloud operations
and implicit model acquisition, and records its owned daemon PID and birth.
Every caller of that action must perform identity-checked process-group cleanup
in an `always()` step. No shared service may be stopped or replaced.

Run `actionlint` after edits and retain genuine failed preparation diagnostics.
See [README.md](README.md) for usage.

An owned supervisor remains the process-group anchor throughout observation,
including when the server or its first runner exits. It ignores SIGTERM so
cleanup retains a verified owner until the final whole-group SIGKILL and live
member readback. Cleanup authenticates PID/birth identities and refuses reuse.

The warm-up request explicitly uses `num_ctx=4096`, matching the real-model
smoke clients. This avoids switching context profiles immediately before
timed requests; successful preparation alone grants no runtime acceptance.