Reduce TROP test time in pure Python CI

[igerber]

Summary

• Replace file-level pytestmark = pytest.mark.slow in test_trop.py with per-class @pytest.mark.slow on the 5 most expensive test classes (38 tests)
• Remove -m '' override from pure Python CI job so the default not slow filter takes effect
• 59 fast tests continue running in pure Python CI; all 97 tests still run in Rust CI

Methodology references (required if estimator / math changes)

• N/A — no methodology changes

Validation

• Tests added/updated: No new tests; verified collection counts match expectations:
  • pytest tests/test_trop.py --collect-only -q → 59 tests (fast subset)
  • pytest tests/test_trop.py --collect-only -q -m '' → 97 tests (all)
  • pytest tests/test_trop.py --collect-only -q -m slow → 38 tests (slow only)
• DIFF_DIFF_BACKEND=python pytest tests/test_trop.py -q → 59 passed

Security / privacy

• Confirm no secrets/PII in this PR: Yes

---

Generated with Claude Code

[github-actions]

Overall Assessment

✅ Looks good

Executive Summary

• No estimator, weighting, variance/SE, identification, or default-behavior implementation changed. I cross-checked the affected methods against docs/methodology/REGISTRY.md, and this PR is confined to CI/docs/tests.
• The test_trop.py marker split is mechanically consistent: the file-level slow mark is replaced by five class-level slow marks, leaving the rest of the TROP test surface runnable under the default not slow filter.
• The main downside is coverage, not correctness: removing -m '' from the pure-Python workflow makes that job skip all slow tests, not just the re-marked TROP classes.
• That broader skip removes pure-Python CI coverage for non-TROP inference checks in Callaway-Sant’Anna and Sun-Abraham, plus all end-to-end TROP method="global"/"joint" fit tests.
• CLAUDE.md is slightly out of sync with the actual default collection behavior because it omits the slow Callaway SE verification test.
• I could not run pytest in this environment because it is not installed, so the scope/count conclusions are from static inspection of markers and workflow config.

Methodology

• No findings. Cross-check against docs/methodology/REGISTRY.md did not surface any methodology deviation because the diff only touches .github/workflows/rust-test.yml:179, tests/test_trop.py:466, and CLAUDE.md:125, not estimator code under diff_diff/.

Code Quality

• No findings.

Performance

• No findings. The workflow change should reduce pure-Python CI time as intended.

Maintainability

• No findings.

Tech Debt

• No findings.

Security

• No findings.

Documentation/Tests

• P2 Impact: Removing -m '' from .github/workflows/rust-test.yml:179 means the suite-wide default marker filter in pyproject.toml:77 now applies to the pure-Python fallback job. That skips every slow test there, not just the TROP classes changed in this PR. In practice, it drops pure-Python execution of the Callaway SE convergence test at tests/test_methodology_callaway.py:804 and the Sun-Abraham bootstrap suite at tests/test_sun_abraham.py:343. Because pure-Python mode also changes test scaling via tests/conftest.py:97, those fallback-specific inference paths are no longer covered in CI. Concrete fix: add a dedicated pure-Python step that explicitly runs the non-TROP slow inference tests with -m '', or scope the skip to TROP-only targets instead of relying on the global slow filter.
• P2 Impact: The class-level @pytest.mark.slow on tests/test_trop.py:2721 suppresses all end-to-end pure-Python CI coverage for the public TROP.fit(..., method="global"/"joint") path. That also hides cheap contract tests such as tests/test_trop.py:2847, tests/test_trop.py:2852, and tests/test_trop.py:2866 along with the expensive optimization/bootstrap cases. The remaining fast joint-related coverage at tests/test_trop.py:2602 and tests/test_trop.py:3588 is low-level helper coverage, not fit() smoke coverage. Concrete fix: split TestTROPJointMethod into a fast smoke/API class and a slow optimization/bootstrap class, and keep at least one public method="global" fit test in the non-slow set.
• P3 Impact: CLAUDE.md:125 now documents only TROP/Sun-Abraham/TROP-parity slow tests, but the default exclusion also covers the Callaway SE formula test at tests/test_methodology_callaway.py:804. Contributor guidance is therefore incomplete. Concrete fix: reword that paragraph to say “all @pytest.mark.slow tests are excluded by default” or explicitly add the Callaway test to the list.