swarmxq-model-orchestrator

Always call ModelOrchestrator.evictIncompatible() before loading any 7B-class model such as Architect, Oracle, Forge, Auditor, or Lab. This guarantees only one 7B model is inference-active at a time, since CPU-only hardware supports exactly one inference thread.

The SINGLE-7B LOCK is the platform's core invariant: only one 7B-class model may be actively inferencing at any moment. On 16 GB hosts a second model like Pilot may stay resident in RAM but must remain idle, never concurrently inferencing.

resolveCanonicalTag() maps legacy aliases like phi4-fast, deepseek-reasoner, and qwen-worker to their canonical production tags via LEGACY_ALIAS_MAP. Unknown tags throw an error, so deprecated names never reach the model registry or logs.

No. Even though 16 GB allows OLLAMA_MAX_LOADED_MODELS=2 for dual residency (Pilot warm plus one 7B), two 7B models must never be inference-active together regardless of available RAM. The lock is about serial CPU inference, not just memory.

PRESSURE_CRITICAL fires when MemAvailable drops to or below RAM_CRITICAL_MB (800 MB), the minimum headroom before the Linux OOM killer activates. This constant is protected and must never be changed; high pressure instead triggers a backoff delay and re-check.

shouldAutoEnableLowRamMode() switches the pipeline to the lite instruct-phi4-lite-q4km-prod model, sets maxLoadedModels to 1, disables Pilot keep-alive, and reduces frame budgets and resolution. Degradation skips must be logged at warn level, never silently ignored.