Add DRAGON cross-language parity test

[marouenbg]

Summary

Add a parity test that pins netZooPy's DRAGON outputs to a fixed numeric snapshot. The same shared inputs and gold values are committed byte-identical in netZooR (companion PR linked below). If both tests pass, the two implementations agree to 1e-5 absolute tolerance.

What's covered

• lambdas (penalty parameters)
• shrunken covariance
• precision matrix
• partial correlation matrix

What's not covered

• κ estimation and p-values — these are unimplemented stubs in netZooR (estimate_kappa_dragon and estimate_p_values_dragon in DRAGON.R are empty function bodies; the public dragon() prints "not yet implemented in R; to estimate p-values, use netZooPy"). When netZooR gains a κ/p-value implementation, this parity test should be extended.

Fixtures

• Inputs: X1.csv (50×20), X2.csv (50×15) — generated deterministically by tests/dragon_parity/regenerate.py from simulate_dragon_data(seed=20250510), then column-scaled.
• Gold outputs: lambdas.txt, cov.csv, prec.csv, ggm.csv. Snapshot of netZooPy's pipeline at the time the fixtures were generated.
• The same files live byte-identical at netZooR's tests/testthat/dragon_parity/.

Local results

R lambdas : 0.8032896157, 0.8167110796
py lambdas: 0.8032890169, 0.8167107609
cov   max|diff| = 2.20e-7
prec  max|diff| = 2.14e-7
ggm   max|diff| = 1.99e-7

R and Python agree two orders of magnitude tighter than the 1e-5 tolerance.

Companion PR

• netZooR side: Add DRAGON cross-language parity test netZooR#406

Test plan

• pytest tests/test_dragon_parity.py — 2 passed locally
• CI runs the new test alongside the rest of tests/test_dragon.py

🤖 Generated with Claude Code