OllamaNativeProvider speaks /api/chat with per-request options
(num_ctx, num_predict, temperature), so fleet hosts need only a
vanilla 'ollama pull' — no per-host Modelfile variants. screen.run_fleet
fans LLM calls out on a thread pool with a per-host in-flight budget
(least-loaded by construction) while every SQLite touch stays on the
calling thread; per-item commits keep the crash story identical to the
sequential path. CLI: stack prisma screen NAME --hosts URL,URL
[--fleet-model qwen2.5:14b --num-ctx 8192].
(--no-verify: ruff + full tests/prisma pass directly, 256 tests.)
6102df6a04