Test against the init and builder images the checkout builds

[MayCXC]

Type of Change

• Bug fix
• New feature
• Breaking change
• Documentation update

Motivation and Context

Five defects in make integration, all of which let a run report a pass about something other than what the checkout built.

The built init image is cleared before the tests see it. make integration clears the application data directory after init-block has already loaded the freshly built init image, so a run against a dedicated APP_ROOT tested the runtime's default init image instead of the one the editable containerization checkout produced. Features that live in the guest agent then fail their tests with no hint that the guest under test was the wrong one. The saved tar stays in the repository's bin directory and the integration recipe loads it after its own system start, once the data directory clearing is behind it; init-block keeps its own load so building the init image alone still refreshes the running store.

init-block cannot answer the kernel prompt. install-init.sh starts the system to load the init image it built, and a start against a data directory with no configured kernel asks on stdin whether to install one. The integration path runs this from make with no terminal, so the question fails the read and init-block exits before a single test begins. It passes --enable-kernel-install the way the integration recipe's own start already does, so the answer is the flag instead of a prompt.

A locally built builder image exists in no registry, the position the editable init image is in, so the recipe lands bin/builder.tar into the store the tests see the same way it lands bin/init.tar, and a fixture's builder start finds the image the configuration names instead of reaching for a registry that does not carry it.

The network suite pulls mid-run. It runs a python http.server, and pulling that image from Docker Hub in the middle of the concurrent pass put a registry round trip and its rate limits inside the test's readiness window. It is warmed with the other suite images now, so the pull happens once, before the load.

The default scratch root was not ignored. The integration recipe defaults SCRATCH_ROOT to .test-scratch under the repository so the builder can reach build contexts, and nothing ignored it, so an interrupted run's scratch trees were one git add away from a commit.

Testing

• Tested locally
• Added/updated tests
• Added/updated docs

make integration against a dedicated APP_ROOT runs the built init image, starts without a terminal, and finds the built builder image. Integration suite: 397 passed. Unit suite: 772 passed. make fmt, make check clean.