Module measure 

Measurement backends.

Two tiers, deliberately separated:

• Deterministic (instructions) — reproducible to ~0.02% run-to-run and ~0.035% across wildly different machine load. This is the only tier that may gate CI.
• Timing (wall_*) — recorded and charted, never gated. On a quiet 32-core host wall clock still shows 4–20% coefficient of variation; under contention the median moves ~150%.

Syscall counts and peak RSS sit awkwardly between the two: better than wall clock (~1%) but not deterministic, because they move with thread scheduling. They are recorded, and may be flagged, but must not gate at a tight threshold.

Structs

Counted

Instruction counts from repeated cachegrind runs.

Plan

Functions

instructions

Instruction count via valgrind --tool=cachegrind, repeated.

valgrind_available

Is cachegrind usable on this machine?

wall

Wall-clock statistics over plan.runs samples.