The test-job memory instrument (#411) reported memory.peak — the high-water mark of
memory.current, which charges reclaimable page cache to the cgroup. A build does heavy
NuGet/npm/obj/bin/coverage I/O, so cache can dominate the peak, and page cache is reclaimed
under a tighter cap rather than OOM-killed. Sizing a per-job cap (server-management#604) off
memory.peak therefore inverts the decision. The OOM-forcing quantity is peak anon, which the
kernel exposes no counter for and which the end-of-job split misses (a job that peaks
mid-dotnet-test then frees reports a low anon).
New scripts/ci-peak-anon.sh: a `start` step (before Build/Test/Coverage) launches a detached
background sampler tracking the high-water mark of cgroup anon; a `report` step (last) stops it
and prints the sampled peak anon as the headline, keeping memory.peak + end-of-job split as a
cache-inflated ceiling and reference. Both continue-on-error + fail-open so they never redden a
build. Validated on bumblebee: survives step-boundary re-execs, catches a transient 2.5 GiB
anon spike the snapshot reports as 0, stops cleanly on kill, degrades gracefully.
Compiler-server A/B (swap-off, sampled peak-anon, n=2 interleaved): OFF (CI config) ~5.84 GiB
consistent; ON (defaults) 6.3-7.6 GiB, always higher, + a ~3 GiB resident VBCSCompiler.
Disabling the servers is worth it, but OFF sits right at 6 GiB for the build phase alone and the
test job adds test+coverage, so #406's "budget loosens well under 6 GiB" premise is not
supported. Size the cap off the live test-job sampler.
Docs: ci-cd.md instrument section rewritten (peak-anon headline + A/B table + premise verdict);
decisions.md entry added. No .cs touched.
fixes#412