CI runs only 4 of 272 unit-test files

[0xKT]

Problem

make test-python -- the only pytest invocation in CI -- hardcodes four file paths:

test-python:
	uv run --extra dev pytest tests/test_commit_lint.py tests/test_large_file_check.py tests/test_cli_smoke.py tests/test_litellm_setup.py -q

So CI executes 4 of 272 top-level test files (~1.5%), or 59 of 4448 collectable tests (~1.3%). The other ~268 files never run on any PR and can red-rot unnoticed. The Windows node-discovery test fixed in #189 was one symptom: it had been broken and nobody saw it, because CI never ran that file.

A second, independent defect: a bare uv run pytest does not even get to the running stage -- collection aborts with 7 errors, because 7 channel-adapter tests import optional SDKs (telegram, slack_sdk, nio, lark_oapi, botpy, dingtalk_stream, wecom_aibot_sdk) at module import time. So "just run everything in CI" fails immediately until those are guarded.

Third: testpaths = ["tests"] with no default deselection means a bare pytest also collects tests/integration/ (real LLM, real VM, real subprocess), which breaks on any machine without that environment configured.

Why it matters

A test suite that does not run is decorative. 268 files' worth of regression protection is currently inert, and contributors get no signal from it.

Proposed fix

1. Guard the 7 channel-adapter tests with pytest.importorskip so collection succeeds.
2. Make the default pytest invocation safe: register integration / e2e markers, add --strict-markers / --strict-config, exclude tests/integration via norecursedirs, and default to -m "not integration and not e2e".
3. Relocate the integration-flavoured and real-resource files that sit in the flat unit tree into tests/integration/.
4. Replace the hardcoded file list in make test-python with a bare pytest over the whole unit tree, and turn the CI python job into a blocking unit matrix.
5. Fix whatever the first full run reveals.

Step 5 is not hypothetical: the first full run surfaced 44 pre-existing failures, including a test that asserted the provider catalog has exactly four concrete backends when it actually has six, and three tests that only passed because the developer's machine had a populated ~/.raven.

Note for maintainers

After this lands, the branch-protection required status check must be updated from python to unit -- the job is renamed, so the old required check would wait on a job that no longer exists.