fix(tests): scope test_ent183_skill_packages' sys.modules stubs to one import (#1898)

[dolho]

Summary

test_ent183_skill_packages.py wrote fake modules straight into sys.modules at import time and left them there. That is collection-time code, and pytest imports every test module before running a single test — so every file collected afterwards resolved services.settings_service (plus database, services.agent_client, utils.url_validation) to a four-function fake, captured names from it at module scope, and kept those bindings for the rest of the session.

Its own autouse sys.modules-restore fixture never had a chance. By the time a fixture first runs, the damage is a binding in another module's namespace, and restoring the module table doesn't reach it.

Two symptom shapes, both of which read as unrelated bugs:

ImportError: cannot import name 'resolve_github_pat' from
'services.settings_service' (unknown location)          # at COLLECTION (#1855)

…and silent wrong behaviour when the stub happens to define the name — the 7 failures in test_1081_physical_meter.py this issue reproduces.

Fix

The stubs now live inside a mock.patch.dict scoped to the single import that needs them. Nothing survives the with; the imported module keeps its stub bindings, which is all the stubbing was ever for.

A false start worth reading

My first version also evicted cached services.skill* modules, to force a fresh import against the stubs regardless of collection order. It broke 10 sibling tests in test_ent236_skills_lifecycle and test_ent237_skill_sources: evicting and re-importing produces a second module object, so their monkeypatching lands on a module the code under test isn't using.

I caught it by running the same selection with and without my change — 435 passed without, 10 failed with — not by reading the diff. Reverted, with the reasoning left in the file: the order-dependence it addressed is a real wart, but it isn't what #1898 is about, and scope creep in a test-isolation fix is how a test-isolation fix breaks tests.

Verification

Case	Before	After
The issue's reproduction (+ test_1081_physical_meter)	7 failed	74 passed
#1855's pair (+ test_ent125_resilient_system_deploy)	collection error	79 passed
+ test_ent236_skills_lifecycle	error	124 passed
The selection my false start regressed	435 passed	436 passed, 0 failed

3 guard tests, mutation-verified — reintroducing the bare module-scope assignment fails all three. The static guard's AST walk deliberately allows sys.modules assignment inside a with: a rule banning every assignment would ban the fix along with the bug.

One correction

When I triaged this I said it explained a test_subscription_auto_switch_pingpong flake I'd seen on #2025. I can't substantiate that. The error signature matched #1855's family, but the pairing doesn't reproduce pre-fix, and neither does a broader selection. test_ent236_skills_lifecycle is a confirmed victim (error → 124 passed); the pingpong one stays unexplained.

Related

#1855 is the same root cause with a different victim and should close as a duplicate once this lands.

Closes #1898

[github-actions]

⚠️ Nightly unit-suite check skipped — merge conflict against dev.

Resolve by running git merge dev locally and pushing the result. The next nightly run will re-test once the conflict is gone.