Per-source unit test files: not run by CI, and should we discourage them by default?

[bbr111]

While reviewing PR #7248 (added tests/test_borlange_energi_se.py alongside a
parser fix), I checked two things that turned out to be connected. Posting
both together since one is the cause of the other.

1. pytest.ini silently excludes all per-source test files from CI

Root pytest.ini:

[pytest]
norecursedirs = custom_components/waste_collection_schedule/waste_collection_schedule/test
filterwarnings = ignore:.*:DeprecationWarning
python_files = test_source_components.py test_fetch_retry.py

python_files is an allowlist of exactly two filenames. Verified with
pytest tests/ --collect-only --ignore=tests/test_source_components.py:
only the 4 tests in test_fetch_retry.py are collected. None of the
following ever run in CI's bare pytest invocation (validate.yaml):

• tests/test_colchester_gov_uk.py
• tests/test_darebin_vic_gov_au.py
• tests/test_heinz_entsorgung_de.py
• tests/test_hillingdon_gov_uk.py
• tests/test_midlothian_gov_uk.py
• tests/test_olo_sk.py
• tests/test_southlanarkshire_gov_uk.py
• tests/test_stockport_gov_uk.py
• tests/test_sudestavenir_fr.py
• tests/test_borlange_energi_se.py (new, from fix(borlange_energi_se): handle "Tömning idag" / "imorgon" #7248)

Consequence: at least 4 of these already fail when run in isolation
(icon-mapping drift, a renamed exception import, a stale mock target), and
nothing ever caught it, because CI has never actually executed them. They've
been silently decorative since whenever they were added.

2. Should we even want a tests/test_<source>.py file per source by default?

Separately from the CI-wiring question: with ~600 sources, one dedicated
test file per source doesn't scale well and risks becoming unmanageable
clutter in tests/. TEST_CASES + a live run of
test_sources.py -s <name> -l is already the standard coverage path for a
source, and is what reviewers actually check today.

Proposal: keep dedicated per-source unit test files as something a
contributor may add for genuinely non-trivial logic (a tricky date/regex
parser, say — the Borlänge one is actually a good example of this done
well: pure function, mocks nothing because it needs nothing, no network),
but stop treating it as a default expectation, and say so explicitly in the
AI-assistant instruction files so agents don't start generating one for
every new source PR.

Draft wording (not yet committed anywhere) for AGENTS.md / CLAUDE.md /
.github/copilot-instructions.md:

Don't add a dedicated tests/test_<source>.py unit-test file by default.
TEST_CASES plus a live run of test_sources.py -s <name> -l is the
normal coverage for a source. Only add a standalone unit test when a
source has genuinely non-trivial logic worth isolating that TEST_CASES
can't exercise well on its own, and mock any network calls in it rather
than hitting the live API.

Decisions needed

1. CI wiring: widen python_files to actually collect these (after
   fixing the 4 currently-broken ones), or leave the allowlist as the
   deliberate CI scope and just document that per-source test files aren't
   automatically run?
2. Policy: adopt the "don't add by default" wording above in the
   AI-assistant instruction files, yes/no/amend?

Not proposing a PR yet — flagging for discussion first since it touches
project-wide contributor guidance.