Regression tests: cache round-trip and gcsfs mock helper

[benmsanderson]

Summary

Adds three focused regression tests targeting the class of bug where the outer cache validator's expected filename and the inner save/load filename diverge — silently triggering redundant network fetches on every call. That was the shape of the bug fixed in the sibling cache-validator PR; this hardens the boundary so a repeat lands in CI, not production.

No src/ changes. Tests only.

What's in

1. Pattern-scaling cache-path round-trip (tests/unit/test_cache_handler.py)

For each of tas / pr, asserts that
CacheHandler.get_pattern_scaling_cache_path(model, variable=v) is byte-equal to the location MeteorPatternScaling.save_model writes to when instantiated with the corresponding per-variable name (cmip6-{model}-aer-{v}). This is the smoking-gun test for the class of bug: any drift between the two sides of the write/validate/read boundary now fails at CI time.

2. Composite training data contiguity (tests/unit/test_cmip6_meteor_data_getter.py)

Uses the existing light_mock_cache fixture to verify that make_meteor_training_data_composite(["historical", "ssp370"], "CanESM5") produces 251 strictly contiguous years with no gaps, no duplicates, and no NaN gaps in either variable. Existing tests check the size but not the join — an off-by-one at the boundary would pass those but corrupt training silently.

3. gcs_zarr_mock fixture + fetch-and-cache round-trip

Introduces a small fixture that patches the two network seams the data-getter uses (gcsfs.GCSFileSystem and xarray.open_zarr) and yields a synthetic monthly CMIP6-shape dataset. Tests can now drive the fetch path without a network.

Uses it to assert that a cache-miss fetch writes an .nc file at the exact path _generate_cmip6_cache_key predicts, and a second call is served from the cache without re-consulting gcsfs. Any divergence between the write filename and the validate/read filename shows up here as either a missing file or a redundant fetch.

Why this matters for CI resilience

The existing test suite avoids the network by leaning on a pre-populated disk cache (~40 real netCDFs in tests/test-data/light_mock_cache/). That's fast but only exercises the cache-hit path. The gcs_zarr_mock fixture opens up the fetch path to unit-testing, which is where cache-boundary bugs like the one in the sibling PR live. Follow-up work could extend it to test the catalog HTTPS fallback and NaN-zstore paths without touching the network.

Verified

• All 35 tests in test_cache_handler.py + test_cmip6_meteor_data_getter.py pass.
• black, ruff, pylint 10.00/10 all clean.
• No src/ changes; zero behavior risk.

Test plan

• pytest tests/unit/test_cache_handler.py::test_pattern_scaling_cache_path_roundtrips_meteor_pattern_scaling_save_name
• pytest tests/unit/test_cmip6_meteor_data_getter.py::test_make_meteor_training_data_composite_stitches_historical_and_ssp_contiguously
• pytest tests/unit/test_cmip6_meteor_data_getter.py::test_get_single_var_mod_data_monthly_writes_cache_at_expected_path
• Full sweep: pytest tests/unit/test_cache_handler.py tests/unit/test_cmip6_meteor_data_getter.py

🤖 Generated with Claude Code

[maritsandstad]

Looks fine, I'll fix linting and merge in