fix(bundle): Pin urunc at the commit that sizes vCPUs from --cpus

[ananos]

Summary

Every guest the bundle boots has one vCPU, whatever brig run --cpus or the profile asks for (4 for the agent profiles). urunc never read the CPU resources in the container's OCI spec. It took the monitor's default_vcpus, which the bundle sets to 1, so cloud-hypervisor got --cpus boot=1. Memory was not affected, since urunc already follows the container's memory limit.

This moves the urunc pin to 74dd0cc9e3028ea18c5daff69e4ba5452075f21d on feat/unchanged_containers-exec-fixes. That is 0818ff1, the rc9 pin, plus one commit, feat(unikontainers): Size a container boot's vCPUs from its CPU limit:

• a CPU quota (what nerdctl run --cpus sets) gives ceil(quota/period) vCPUs;
• a cpuset gives the number of CPUs it names;
• when both are set, the smaller wins;
• the count is capped at the CPUs urunc may run on;
• only a container boot changes, and a unikernel keeps the monitor default.

default_vcpus = 1 stays as the fallback for a container with no CPU limit.

Once this is merged, bundle rc10 is cut from main, and brig's install.sh moves its runtime pin from v0.1.0-rc9 to rc10 in a separate brig PR.

Changes

• scripts/build-bundle.sh: URUNC_REF_DEFAULT moves from 0818ff1 to 74dd0cc, and the comment says what the pin carries.
• README.md, DESIGN.md, docs/variants.md: the pinned commit, plus a line in the README's pins section on --cpus.

Testing

• The urunc commit's own unit tests pass on an Ubuntu 24.04 amd64 host: TestMonitorVCPUs, TestCpusetCount, TestBuildMonitorSpecVCPUs, plus the existing monitor-spec and container-boot tests. With the line that applies the count removed, TestBuildMonitorSpecVCPUs fails. golangci-lint --new-from-rev reports 0 new issues.

• This repo: sh -n and shellcheck are clean, and tests/*.sh all pass, source-pins.sh included.

• CI built all four bundles, and each pins.env names URUNC_REF=74dd0cc… with URUNC_PINNED=true.

• Live, on an Ubuntu 24.04 amd64 host (14 CPUs) with a rootless user install. This PR's bundle-amd64-rootless CI artifact was installed over rc9 with INSTALL_BRIG_BUNDLE (upgrade in place, no sudo), with brig main 3e5ae98. The cloud-hypervisor argument is read from ps:

  case	CPU quota / period	cloud-hypervisor	guest nproc
  rc9 baseline, ubuntu	(not read)	--cpus boot=1	1
  ubuntu, profile default	200000/100000	--cpus boot=2	2
  claude-code, profile default	400000/100000	--cpus boot=4	4
  --cpus 1	100000/100000	--cpus boot=1	1
  --cpus 3	300000/100000	--cpus boot=3	3
  --cpus 64	6400000/100000	--cpus boot=14 (capped)	14, boots

  --mem 1024 still gives --memory size=1073M and a MemTotal of about 1 GB.

• No regression on that host:

  • 300 short brig sh calls: none missing.
  • 10 claude-code boots: 0 .claude refusals.
  • A forced image pull with DOCKER_CONFIG unset.
  • No leftovers.