Inference admission

Use MereRunAdmission to coordinate inference work without importing the CLI, HTTP server, or tensor runtimes. It uses Foundation and platform system APIs and has no Swift package dependencies.

Two admission scopes

Scope	Owner	Lifetime
Machine reservation	MachineInferenceCoordinator	One command or the full server lifetime
Request slot	RuntimeRequestAdmission	One request or model-maintenance operation
Loaded-model lease	Runtime pool	Runtime use, including streaming production

A server holds one weighted machine reservation. Its request queue enforces --max-active-requests inside that reservation. Acquiring a request slot does not acquire another machine ticket.

The machine coordinator preserves weighted FIFO ordering, disk and memory headroom checks, file locking, atomic state updates, and dead-process and previous-boot recovery. CLI adapters select the existing shared admission state directory and classify command arguments. Typed callers provide a MachineInferenceRequest; resolved model-size estimates can use its estimatedModelBytes initializer with a modality cost floor or explicit exclusive admission.

Release ownership

Use withMachineInferenceAdmission when the operation's lifetime matches the machine reservation. API serving uses this scope around server creation and execution, so startup failure and shutdown release the same reservation.

Use withRuntimeRequestAdmission when the operation finishes before its caller returns. The helper releases the slot on success, failure, or cancellation. Scoped cancellations increment the cancellation counter.

For a streaming response whose producer outlives its handler, acquire an explicit request lease and release it when production ends. The streaming owner retains both its request slot and its loaded-model lease. Do not wrap response construction in a scope that releases before streaming finishes.

Request admission checks cancellation before acquiring a slot and after suspended memory-pressure sampling. It removes cancelled queued requests and releases a lease if cancellation races with a queued grant. Its pressure and capacity checks retain FIFO ordering across actor suspension.

Runtime boundaries

RuntimeRequestProgress carries admission telemetry. CLI adapters translate Core's ChatProgress into that type. The existing status field names remain compatible.

MereRunResidency owns model generations, leases, and eviction mechanics. RuntimeModelPool and APISidecarModelPool supply runtime factories, settings, and pressure samples. Their runtime owners retain batching and KV caches. The admission library does not load checkpoints.

Validation

To compile the library and its test module without the inference targets, run:

swift build --target MereRunAdmissionTests

This command compiles the tests. Run swift test or ./scripts/check.sh to execute the package tests. Coverage includes machine FIFO and crash recovery, request cancellation, scoped cleanup, concurrent release, and the distinction between server reservations and request concurrency. Runtime-pool and API integration tests remain in MereRunCLITests.