test: unify a shared flaky-rerun marker; fix nwis collection-time call

[thodson-usgs]

Problem

The Python package CI on main failed (run 27969553989, on the #326 merge): lint/type-check and all Windows test jobs passed, but all three Ubuntu jobs failed with the same cause —

dataretrieval.exceptions.ServiceUnavailable: HTTP 503 (URL: https://waterservices.usgs.gov/nwis/site?...)

A transient 503 from the live legacy NWIS service hit the live tests in tests/nwis_test.py. Windows passed only by timing. Not a regression from #326 — the live tests and the offending call date to #62.

Two things made a transient outage fatal instead of self-healing:

1. No rerun protection. waterdata_test.py / ngwmn_test.py retry transient 429/5xx via pytest.mark.flaky (test(waterdata): rerun flaky transient 5xx/429 from the chunked fan-out #325); nwis_test.py had none.
2. A live call at collection time. TestTZ fetched its sites in the class body (sites, _ = what_sites(stateCd="MD")), so the 503 raised during collection (exit 2, Interrupted: 1 error during collection on 3.13). pytest-rerunfailures reruns failed tests, never collection errors.

What this does

Audited every test module for live-API calls (ran each behind a dead proxy: genuine calls fail, mocked ones pass). Then unified the rerun handling instead of adding a fourth inline copy:

• One shared marker, conftest.flaky_api, with a single transient-pattern list. Applied to every live suite:
  • nwis_test.py, waterdata_test.py, ngwmn_test.py — module-level pytestmark = flaky_api.
  • utils_test.py::Test_query — class-level @flaky_api (the only other live suite the audit found; the rest of that module is mocked).
• The status pattern now matches both message shapes — the OGC path's "<status>:" and the legacy query path's "HTTP <status>" — so one list serves all four.
• Fixes a latent coverage gap: ngwmn's old marker omitted ServiceUnavailable and the chunked QuotaExhausted/ServiceInterrupted wrappers (test(waterdata): rerun flaky transient 5xx/429 from the chunked fan-out #325 added those to waterdata but missed ngwmn), so a 503 in an NGWMN live test would have failed CI too. It now inherits the full set.
• Fixes the collection bug: TestTZ's class-body fetch moves into a class-scoped sites fixture, so it runs at test time where the marker can retry it.

Everything else is mocked, module-skipped (nadp), or — the chunking ..._on_real_transport test — a deterministic local http.server, not an external service.

Verification

• nwis_test.py collection is now network-free (--collect-only makes no live call).
• The shared marker attaches to all four suites with the unified config; ngwmn now covers ServiceUnavailable/ServiceInterrupted.
• The pattern matches ServiceUnavailable: HTTP 503 and ... 503: but not a deterministic AssertionError.
• Live Test_query + TestTZ pass; mocked suites unaffected.
• The originally-failed CI run was re-run and is now green.

🤖 Generated with Claude Code