Add a test suite (unit + regression)

[shreyasgm]

The package currently has no tests. Every refactor in the modernization tiers (#41, #42, #43) is risky without them, and users keep hitting correctness questions (#8, #19, #21, #23, #25, #38, #39) that we can't answer decisively without reproducible fixtures.

Goals

1. Lock in current behavior before we refactor, so we can prove "no behavior change" commits.
2. Cover the known numerical pitfalls so correctness regressions are caught automatically.
3. Run in CI on every PR.

Scope

Infrastructure

• Add pytest, pytest-cov, and hypothesis (for property tests) as a [test] extra.
• tests/ directory layout; conftest.py with fixtures.
• GitHub Actions workflow running on 3.10 / 3.11 / 3.12.

Fixtures

• A small synthetic trade panel (~5 countries, ~10 products, 3 years) checked in as CSV. Hand-computed or independently-computed expected values for RCA, MCP, diversity/ubiquity, proximity, and ECI/PCI.
• A larger golden-file regression fixture: run the current main on a real sample (e.g. a subset of SITC 2-digit) and snapshot the output dataframe. Compare to tolerance on every PR.

Unit tests

• calculate_rca — verify against hand-computed values on the synthetic fixture.
• calculate_rpop — including the zero-population edge case (RPOP - Handling zeros in diversity / ubiquity #22).
• calculate_mcp — discrete thresholding, presence_test="both" behavior, and manual-MCP validation error at ComplexityData.py:158-164.
• calc_discrete_proximity / calc_continuous_proximity — including whether diagonal should be zero (Proximity matrix diagonal should be set to 0 #38).
• calc_density with and without knn; regression test for the integer-output bug (Using knn option gives integer density outputs #35).
• calc_eci_pci — eigenvector sign convention, NaN reinsertion for zero-diversity rows, normalization formula (PCI calculation. #19, Possible mistake in the code of ecomplexity.py #23, PCI normalization using wrong mean and standard deviation.  #25).
• calc_coi_cog.

Integration / golden tests

• Run ecomplexity() end-to-end on the synthetic fixture; assert column set, dtypes, and values to tolerance.
• Regression test: output equals snapshot for a larger real-ish input. Snapshot regenerated only with an explicit flag.
• Reproduce the discrepancy-with-Atlas-data report (ecomplexity function's pci does not match the pci in the raw data #39, ecomplexity output does not conform with atlas dataverse results using R reticulate #21) as a documented test, so we can track whether the gap is expected vs. a bug.

Property tests (hypothesis)

• RCA is non-negative.
• MCP ∈ {0, 1}.
• ECI has mean ≈ 0 and std ≈ 1 after normalization.
• Proximity matrix is symmetric when asymmetric=False and has values in [0, 1].
• Output row count = locs × prods × years (rectangularized).

Edge cases from open issues

• Using knn option gives integer density outputs #35: knn option should return floats, not ints.
• RPOP - Handling zeros in diversity / ubiquity #22: RPOP with zero diversity/ubiquity for some locs/prods.
• Handle column label conflicts #36: column-label conflicts raise a clear error.
• Proximity matrix diagonal should be set to 0 #38: diagonal of proximity — decide the expected behavior and lock it in.

Acceptance

• pytest runs clean locally and in CI.
• Coverage ≥ 80% on ecomplexity/.
• Regression test catches at least one of the known open bugs (used as a red-green check).