tests: cover interface resolution against real Ansible

[ideaship]

The unit tests for get_rabbitmq_node_addresses() replace both subprocess and the Redis client, so what they verify is the shape of the calls the function makes. That is a poor fit for the part of it that matters most: the value it resolves, internal_interface, is frequently Jinja-valued, and its meaning is defined by Ansible's templating and by how Ansible names interface facts. A mock cannot speak for either, so those tests pass whether or not the resolution agrees with Ansible.

This adds integration tests that drive the public function with the real collaborators — a live Redis holding the cached facts, an inventory on disk, and the real ansible binaries — and deletes the unit tests they supersede:

retired unit test	replaced by
literal interface	test_literal_interface
Jinja resolved from facts	test_fact_derived_interface
interface name → fact key	test_dotted_interface_name, test_dashed_interface_name
no internal_interface	test_missing_internal_interface_yields_no_addresses
no fact for the interface	test_interface_without_matching_fact_yields_no_addresses
interface fact without ipv4	test_interface_fact_without_ipv4_yields_no_addresses
ipv4 without an address	test_ipv4_without_address_yields_no_addresses

Each pair covers the same behaviour, one against mocks and one against Ansible, so keeping both would only duplicate. The unit tests that remain are the ones with no equivalent at this level: the per-host loop, the inventory and cache error paths, and the two describing the internals of the resolver itself.

Because these target the public function rather than its internals, they keep their meaning across changes to how the resolution is implemented — which is the point. They are what will make it visible that the reimplementation in the next PR preserves the shapes that already worked.

One case is xfail(strict=True): an internal_interface pointing at an inventory variable rather than a fact, which is what was reported in the issue below. The current resolver walks dotted paths through the facts only, so it cannot resolve it. A strict marker documents the behaviour where the others are documented, and forces whatever fixes it to remove the marker — an unexpected pass fails the suite. The suite is green with the xfail present (exit code 0).

CI. The integration job installs ansible from requirements.ansible.txt, the file the container image already builds from, so the version under test is the version that ships and no second ansible pin is introduced. OSISM_REQUIRE_ANSIBLE makes a missing install fail the job rather than skip every ansible-dependent test, mirroring OSISM_REQUIRE_REDIS.

Ansible is installed for the integration job only, because that is the only job that needs it — the unit tests mock the subprocess away.

Documents the bug reported in:

• Jinja2 variables not rendered by rabbitmq3to4 issues#1425

🤖 Generated with Claude Code