Unpack the mock app store fixtures and build the zips at test time

[ClaydeCode]

The mock app store fixtures were checked in as zips, so the compose templates and app_meta.json files inside them could not be read or diffed without unpacking first. This unpacks them into the repo and builds the zips at test time instead.

Two commits: tests: drop the unused large_app fixture, then tests, meta: build the mock app store zips at test time.

What changed

Fixtures are now plain files. Each app directory holds its app_meta.json, docker-compose.yml.template and icon directly. A session-scoped fixture (mock_app_store_zips) packs them into a TemporaryDirectory at session start and removes it at the end, yielding the same <root>/<name>/<name>.zip layout the six consumers already expected.

The unpacked files came out of the zips, not from the loose copies that were sitting beside them — four of those had already drifted from what actually ran. always_on, filebrowser, large_app and mock_app lacked the version: '3.5' line the live template carried, and always_on's named the shard network where the live one names portal. Taking the zip as the source means the suite installs byte-identical apps to before, apart from the port change below.

large_app is gone. No test installed it by name; its only trace was # todo: test_large_app_does_not_start, removed with it. Its app_meta shape duplicated filebrowser's, so the migration corpus loses nothing.

ports: - 80:80 dropped from every template. Nothing consumed the host binding: Traefik reaches apps over the shared portal network at http://<container>:<port> (shard_core/service/traefik_dynamic_config.py:193), and no app in the app-repository publishes a host port. The binding only collided with whatever already held port 80, which made test_app_starts_and_stops unpassable on any host running its own reverse proxy. pause_cycle already had no ports block, so this makes the set consistent.

immich/.env needed a .gitignore negation. The repo's .env rule matched it, so unpacking it as a loose file would have left it uncommitted and the built immich zip would have silently lost it. build_mock_app_store_zips now asserts every fixture file is git-tracked, so the next such file fails the run instead of quietly producing a different app in CI than on a developer machine.

test_all_mock_app_store_metas_migrate reads the app_meta.json files directly rather than opening each zip for one entry.

Verification

• Full suite on this branch: 308 passed, 0 failed. On main, the same machine gives 1 failed — test_app_lifecycle.py::test_app_starts_and_stops, on Bind for 0.0.0.0:80 failed: port is already allocated.
• pytest --collect-only gives 308 on both this branch and main, so no test was dropped.
• Built zips compared file-by-file against the deleted ones from main: every entry byte-identical, and the six templates differ only by the removed ports block.
• The git-tracked assertion was proven to fire — a stray stray_secret.env in mock_app/ produced AssertionError: mock_app has uncommitted fixture files: ['stray_secret.env'].

Recommended reading order

1. .gitignore — the negation that lets immich/.env be committed
2. tests/mock_app_store/** — the unpacked fixtures (the template diffs are the removed ports block; version: '3.5' lines appear because the stale loose copies lacked them)
3. tests/util.py — build_mock_app_store_zips and the git-tracked assertion
4. tests/conftest.py — the session fixture, and mock_app_store taking the built root
5. tests/test_app_installation.py, tests/test_app_zip.py, tests/test_app_install_crash_recovery.py — consumers moved onto the fixture
6. tests/test_app_meta_migration.py — reads app_meta.json directly
7. tests/test_app_lifecycle.py — the removed todo

🤖 Generated with Claude Code

https://claude.ai/code/session_01AHmisdKK8deh7iEd3Shn2Q