geotiff tests: consolidate integration cluster (http, dask pipeline, gpu pipeline)

[brendancol]

Closes #2402. PR 9 of the 11-PR epic #2390.

Summary

Fold 26 integration-cluster test files into three parametrised files under xrspatial/geotiff/tests/integration/:

• integration/test_http_sources.py: 17 HTTP files, 218 tests
• integration/test_dask_pipeline.py: 7 dask + accessor files, 61 tests
• integration/test_gpu_pipeline.py: 1 GPU file, 5 tests

Total: 284 tests, all green.

How the merge handles cross-file collisions

Each old file lands as a named section in its consolidated module. Helpers, fixtures, and classes are suffixed with the section id (_<section>) so cross-section names cannot collide.

Top-level autouse=True fixtures from each source file lose their autouse flag and apply via an explicit @pytest.mark.usefixtures(...) marker on the tests and classes of that section only. Without this, sections that set XRSPATIAL_GEOTIFF_ALLOW_PRIVATE_HOSTS=1 would leak that env var into the scheme_case tests, which exist precisely to assert SSRF rejection in the production default.

Issue-number suffixes on test names (_2266, _2026_05_15, _issue_A) are stripped per epic convention. Issue numbers stay in the git log and PR descriptions.

@requires_loopback annotation

test_http_sources.py carries a module-level pytestmark = requires_loopback, so PR 11's removal of the pytest_collection_modifyitems socketserver hack in conftest.py does not regress coverage on sandboxes that deny loopback bind.

Audit

xrspatial/geotiff/tests/CLUSTER_AUDIT_PR9.md maps every old file::test to its new home (204/204 mapped). Deleted on the final commit before approval.

Test plan

• pytest xrspatial/geotiff/tests/integration/ -v: 284 passed
• pytest xrspatial/geotiff/tests/ -q: 5704 passed, 1 unrelated flake in test_streaming_codecs_2026_05_11.py (passes alone)

[brendancol]

PR Review: geotiff tests: consolidate integration cluster (http, dask pipeline, gpu pipeline)

Tests-only restructure. 26 source files folded into 3 consolidated files under integration/, 284 tests collected and green. The audit table maps 204/204 old tests. The autouse-to-usefixtures conversion correctly scopes section-specific env-var setting, so the SSRF rejection tests in the scheme_case section still see the production default. Module-level pytestmark = requires_loopback on the HTTP file lines up with PR 11's planned removal of the socketserver hook.

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• xrspatial/geotiff/tests/integration/test_gpu_pipeline.py:13: the file defines its own _gpu_available_dask_cupy_combined() helper instead of using requires_gpu from xrspatial.geotiff.tests._helpers.markers. The epic calls out the markers module as the one place for requires_gpu. Swap the local helper for from ..._helpers.markers import requires_gpu and apply @requires_gpu to the tests (or set pytestmark = requires_gpu at module scope).

• xrspatial/geotiff/tests/integration/test_http_sources.py:12, test_dask_pipeline.py:12, test_gpu_pipeline.py:10: duplicated from __future__ import annotations. One sits in the module header, the second comes from the per-section import dedup. Future imports must be the first executable statement; two is harmless but noisy. Drop the duplicate.

• xrspatial/geotiff/tests/integration/test_http_sources.py:13-30: redundant import aliases for the same module. from xrspatial.geotiff import _reader as _reader_mod and from xrspatial.geotiff import _reader as reader_mod both exist; from xrspatial.geotiff import open_geotiff appears in three separate lines. The script preserved each source file's import statements verbatim and only deduped exact-string matches. Collapse to one canonical alias each (_reader_mod is used most; drop the unaliased and the reader_mod variants and rewrite references).

Nits (optional improvements)

• The audit table in CLUSTER_AUDIT_PR9.md annotates parametrised test ids by matching against the first collected parametrize variant. For tests with multiple parametrize axes (e.g. test_private_host_rejected_regardless_of_scheme_case[127.0.0.1-HTTP]), only one variant is listed. Coverage is still traceable, but a one-line note in the audit's preamble would clarify that one row may cover several parametrize variants.

• test_http_sources.py keeps multiple section-local _serve_<sid>, _RangeHandler_<sid>, and _stop_<sid> helpers that all do roughly the same thing. Out of scope here (the audit-driven approach mandates faithful preservation), but worth a follow-up issue to factor a shared loopback-server helper into _helpers/.

What looks good

• The autouse-fixture conversion is the load-bearing piece of this PR and it is correct: _allow_loopback* fixtures that set XRSPATIAL_GEOTIFF_ALLOW_PRIVATE_HOSTS=1 no longer leak across section boundaries. The test_private_host_rejected_regardless_of_scheme_case family in the scheme_case section still asserts SSRF rejection because the marker is now scoped to its origin section.
• 204/204 old test functions accounted for in the audit table.
• Module-level pytestmark = requires_loopback on test_http_sources.py is the right primitive for PR 11's hook removal.
• Helper, fixture, and class suffixing prevents cross-section name collisions.

Checklist

• Algorithm matches reference/paper -- N/A, tests-only restructure
• All implemented backends produce consistent results -- preserved, no source changes
• NaN handling is correct -- preserved
• Edge cases are covered by tests -- preserved (audit shows 1:1 mapping)
• Dask chunk boundaries handled correctly -- preserved
• No premature materialization or unnecessary copies -- N/A, tests-only
• Benchmark exists or is not needed -- N/A
• README feature matrix updated -- N/A
• Docstrings present and accurate -- module-level docstrings on each consolidated file

[brendancol]

Follow-up review after 761149f

All three Suggestions and the first Nit applied:

• test_gpu_pipeline.py now imports requires_gpu from _helpers/markers.py and applies it as a module-level pytestmark. The local _gpu_available_dask_cupy_combined helper is gone.
• Duplicated from __future__ import annotations removed from all three files.
• test_http_sources.py import block deduped: one _reader as _reader_mod alias (was _reader_mod + reader_mod), one open_geotiff, read_geotiff_dask line (was three), one merged _reader symbol list (was 10 partial lines).
• test_dask_pipeline.py import block deduped: one open_geotiff, read_geotiff_dask, to_geotiff line.
• CLUSTER_AUDIT_PR9.md preamble clarifies that one row can cover several parametrize variants.

Second Nit (shared loopback-server helper in _helpers/) left for a follow-up issue: the audit-driven approach mandates faithful preservation, and a shared helper crosses section boundaries.

Re-ran pytest xrspatial/geotiff/tests/integration/ -q: 284 passed.