fix: thread PCA n_pcs and external covariates into GWAS assoc step

[dvitale199]

Summary

• Fix PipelineRunner._run_single_step (assoc branch), which threaded only build and maf_lambdas into AssocConfig and silently dropped three user-supplied GWAS args:
  • --pca N (requested PC count): PCAConfig was built without n_pcs, so PCA always ran the default 10 PCs regardless of --pca N.
  • --covars / --covar-names: never reached AssocConfig.covariates, so external covariates were silently ignored and GWAS fell back to PCA eigenvectors as covariates.
• Now threads n_pcs into PCAConfig and builds a CovariateConfig from the covar path/names when given; normalizes run_pca/run_gwas to bool. Adds CovariateConfig to the runner's config-class map.
• Extends the old-vs-new parity harness to guard both args.
• Updates REFACTOR_HARDENING.md (item 8 done) and TESTING.md.

This directly affects real-cohort GWAS: before the fix, --pca N and external --covars were ignored by the new CLI.

What changed

• genotools/cli/runner.py — assoc branch threads n_pcs + covariates; CovariateConfig added to config-class map.
• tests/unit/test_cli/test_runner_regression.py — new TestAssocStepThreadsPcaAndCovariateArgs (hermetic capture of the config passed to run_association: n_pcs + external covars reach AssocConfig, build preserved) and TestAssocStepPcaProducesRequestedPcs (end-to-end --pca 3 → 3-PC eigenvec via real PLINK2, was 10).
• tests/regression/test_parity.py — test_old_vs_new_pca_ncount_parity (--pca 20 → 20-PC eigenvec in both CLIs) and test_old_vs_new_gwas_external_covars_parity (external covars used identically → p-values agree; decision B does not apply because the PCA eigenvectors are discarded for external covars).

Test plan

• Unit + regression suite green: 397 passed (391 → 397) with .venv-stable present.
• New unit tests fail RED against the pre-fix runner, pass after (n_pcs=10 vs 3; covar_path=None).
• Both new parity tests verified to fail against the pre-fix runner (n_pcs 10 vs 20; 40500 p-mismatches + lambda 0.978 vs 1.007) and pass after.
• Parity tests skip cleanly without .venv-stable/plink2.
• Real-cohort parity (separate gate, run by dev before refactor/main → main).

Notes

• Targets refactor/main, not main.
• Decision B unaffected: the external-covars parity test intentionally exercises the path where PCA eigenvectors are discarded, so the ratified PCA region-exclusion divergence does not apply and per-variant p-values must match.