pub const LLM_WORKER_RSS_MB: u64 = 350;Expand description
DEFAULT expected RSS, in MiB, budgeted for one LLM/REST worker.
§This number was NOT measured empirically
It is a v1.0.75 (G18 + G23) engineering estimate for a worker that used to spawn a subprocess, kept after the move to the OpenRouter REST client because the REST footprint is strictly smaller and the estimate therefore stays conservative. No benchmark, profile or RSS sample backs the exact value 350, and no test asserts it against a measurement. Treat it as a deliberately pessimistic budget, not as data.
It governs every concurrency ceiling derived from free memory
(crate::memory_guard::calculate_safe_concurrency,
crate::llm_slots::default_max_concurrency,
crate::embedder::effective_permits), so an operator who has measured the
real footprint on their host SHOULD override it rather than live with the
estimate: read it through llm_worker_rss_mb, never directly.