Grading latency

Time to grade one correct submission, per tier. Measured sequentially (no parallel contention) by tools/bench_latency.py: the oracle solution of three tasks, five runs each after one discarded warm-up run (15 runs per tier), on Docker Desktop (Windows 11, 12 logical CPUs, 3.7 GB VM memory).

Tier	Median (s)	p95 (s)	Mean (s)	Runs
T0 naive	2.02	2.36	2.07	15
T1 typical	3.19	3.83	3.27	15
T2 compat	3.12	8.06	4.39	15
T3 strict	3.39	9.24	4.86	15

The strict tier starts three throw-away containers (agent, candidate execution, judge) and audits each one's effective configuration before starting it, and it costs about 1.4 seconds more per grade (median) than the naive single-container tier. The p95 on T2 and T3 is dominated by occasional slow container starts on a small local VM, not by the grading logic. Numbers will differ on other machines; the weekly workflow publishes a Linux-runner measurement as a build artifact.