Add regression tests for the bug classes that escaped 0.4.x CI

[ZachHoppinen]

Adds four targeted tests for the bug classes that leaked through to released versions during the 0.4.0-0.4.4 cycle. All run as pure unit tests — no network, no Earthdata creds — and would have flagged the regressions on PR rather than in user-facing crashes.

What's covered

Test	Bug class	Closes the gap from
`test_snow_index.py::test_calc_delta_cross_ratio_disjoint_orbits_with_singleton`	Multi-orbit + disjoint times + single-acquisition orbit. In-place `.loc[time=...]` write into broadcast-backed read-only buffer under numpy 2.x.	PR #85
`test_wetsnow.py::test_flag_wet_snow_snowcover_mask_applied_to_all_timesteps`	Stale-`ts` loop variable used outside its loop scope; snowcover mask applied only to the last timestep.	PR #79 piggyback fix
`test_generating_dataarrays.py::test_generate_forest_fraction_dataarray_outside_conus`	Non-CONUS FCF code path never exercised by existing test AOI. Mocks `download_proba_v` to a synthetic tif and runs both the no-ref and with-ref branches end-to-end.	PRs #79 (TypeError), #80 (eager-load OOM), #85 (clip_box WindowError)
`test_pipeline_synthetic.py::test_post_download_pipeline_runs_end_to_end` (new file)	Whole post-download pipeline (preprocessing → snow-index → wet-snow) against a multi-orbit + single-acquisition synthetic dataset. Offline counterpart to the integration test.	All snow-index / wet-snow regressions

What's still NOT covered

• Real download paths (asf_search, earthaccess) — unit tests still mock these. A bug in URL construction that only manifests against the real APIs would still escape.
• Numerical correctness — the synthetic pipeline test asserts the pipeline runs and outputs are in plausible ranges (non-negative snow_depth, binary wet_snow), not that the numbers are right. A snapshot fixture from a real run could harden this.
• `generate_reference_grid` y-orientation flip (Fix non-CONUS snow-depth pipeline (FCF + snow_index + VIIRS tile filter) #85) — defensive code that fires only on the degenerate single-acquisition track case. Hard to exercise without real S1 burst data; left as a known gap.

Future work for higher confidence: either run `RUN_INTEGRATION=1` in CI (needs Earthdata creds), or record a tiny real-data fixture for the post-download pipeline test.

Test plan

• `pytest tests/` — 77/77 pass (was 73; +4 new).
• Verified test_calc_delta_cross_ratio_disjoint_orbits_with_singleton fixture has exactly the shape that triggered the read-only error (3 orbits with disjoint times + a single-acquisition orbit).

🤖 Generated with Claude Code