snapshot

Purpose

Answers "what does this cost, and is this machine even able to tell me?"

Read ${CLAUDE_PLUGIN_ROOT}/reference/harness-integrity.md before writing any harness here. It is not optional background: five harnesses in the source run returned confident wrong answers, and four of the five were checks written specifically to avoid being fooled.

Phase order, and why it is this order

1. Qualify the host, before measuring anything

The lib is plugin-bundled, not installed, so it needs its directory on sys.path before the import. Anchor to the plugin root rather than the working directory; a bare from spawn_noise import ... raises ModuleNotFoundError unless the caller happens to already be in lib/.

import sys
from pathlib import Path

lib = Path(__file__).resolve().parents[3] / "lib"   # <plugin-root>/lib
if str(lib) not in sys.path:
    sys.path.insert(0, str(lib))

from spawn_noise import spawn_probe, is_measurable  # noqa: E402

summary = spawn_probe()
measurable, why = is_measurable(summary)

parents[3] is correct from <plugin-root>/skills/<skill>/scripts/x.py. Count the levels for wherever the caller actually sits rather than copying the index.

is_measurable() returns a verdict and its basis. A False is a hard refusal to report a wall-clock number, subject only to the recorded override below.

The refusal names what it can still report. That matters: an unexplained refusal gets overridden reflexively. This host spread 15.7x across identical no-op spawns, and the durable result from the source run was a deterministic spawn count of 4 -> 1, not a duration.

Say plainly that this refusal is a house rule. No surveyed benchmarking tool refuses above a variance threshold: pyperf, Criterion, JMH and benchstat all warn and print the number anyway. pyperf's own thresholds (stdev >= 10% of the mean, min/max >= 50% from the mean, shortest value < 1 ms) are warnings. Presenting this refusal as consensus would be a miscitation.