Module runtime_optimizations 

Runtime-toggleable performance optimizations.

Per coding_agent_session_search-yvv7r (rollback env vars for ifr7/lxn5) and coding_agent_session_search-waijq (CASS_F16_PRECONVERT for mng4).

Design contract

Each runtime optimization is gated by a CASS_<FEATURE> env var read once at startup and cached in a OnceLock<bool>. Operators flipping a toggle must restart cass; per-query toggling is intentionally not supported (the contract is “operator flips, restarts, measures”).

Truthful values

Bool semantics (case-insensitive):

• Unset, 1, true, on, yes → optimization ENABLED (default).
• 0, false, off, no → optimization DISABLED (rollback path).
• Any other value → log a tracing::warn! and treat as ENABLED.

Health surface

cass health --json exposes runtime_optimizations: { simd_dot, parallel_search, preconvert_f16, config_source } so operators and monitoring can confirm the flip took effect.

Why not std::env::var

Per AGENTS.md, ALL configuration must load via dotenvy::var() — it respects the project’s .env file. std::env::var would skip .env entries and produce inconsistent behavior between dev and production.

Structs

RuntimeOptimizationsSnapshot

Snapshot of the cached toggle values. This is the canonical shape exposed in cass health --json under runtime_optimizations.

Enums

ConfigSource

Source of the runtime-optimization configuration as surfaced in cass health --json under runtime_optimizations.config_source.

Functions

init_from_env

Force-resolve all toggles. Called once at startup before any query path reads them. Subsequent calls are no-ops thanks to OnceLock::set.

parallel_search_enabled

Whether parallel rayon-driven vector search is enabled.

preconvert_f16_enabled

Whether f16→f32 preconversion at vector-load time is enabled.

simd_dot_enabled

Whether SIMD dot product is enabled. Lazily initializes from env on first read if init_from_env() has not been called yet.