fix: Compute PCA variances and OLS TSS with a two-pass scheme

[KFilippopolitis]

Summary

PCA, PCA with transformations, and the OLS total sum of squares computed variance-like quantities from aggregated raw moments:

• PCA: variances = (sum(x^2) - n * mean^2) / (n - 1) (in FederatedPCA.fit and duplicated in the pca_with_transformations wrapper).
• OLS: tss = sum(y^2) - 2 * mean * sum(y) + n * mean^2.

Both cancel catastrophically when the mean is large relative to the spread (e.g. values around 1e8 with unit standard deviation). This PR switches them to the same two-pass scheme already used for Pearson, describe, and the t-tests:

1. Aggregate n and sums to get the global mean.
2. Aggregate sums of squared deviations from that mean.

Changes

• exaflow/algorithms/federated/decomposition/pca.py: new global_mean_std(agg_client, X) helper (round 1 sums [n, column sums] in one call, round 2 sums centered squares). FederatedPCA.fit uses it. The negative-variance clip is removed because a sum of squares cannot be negative.
• exaflow/algorithms/exareme3/decomposition/pca_with_transformations.py: reuses the helper instead of its own raw-moment copy. The n <= 1 fallback is unchanged.
• exaflow/algorithms/federated/linear_model/ols.py: drops the sum(y^2) aggregation and aggregates sum((y - mean_y)^2) after the residual sums.
• Every worker issues the same aggregation calls in the same order, with no data-dependent branches. Aggregation call counts: PCA 4 to 3, wrapper standardization 3 to 2, OLS unchanged.
• Docs: PCA.md, PCAWithTransformation.md, LinearRegression.md describe the two-round flow. federated/docs/ols.md no longer lists y'y as aggregated (the commit hook also reformatted that file's Markdown).

Tests

New large-offset regression tests (OFFSET = 1e8, 3 simulated workers via _simulate_federated_execution, rtol 1e-6):

• test_pca.py: test_global_mean_std_large_offset, test_pca_large_offset, test_pca_with_transformations_large_offset.
• test_ols.py: test_ols_tss_large_offset (TSS, R², F against statsmodels).

Against the old code, the PCA eigenvalue test fails (up to 158% relative error) and the OLS TSS test fails (42% relative error).

uv run pytest -q tests/standalone_tests/federated_algorithms/decomposition tests/standalone_tests/federated_algorithms/linear_model

128 passed. Other OLS/PCA consumers (two-way ANOVA, cross-validation, pipeline, linear regression wrapper) also pass.

Risk

• No API, request, or response shape change. Results change only by becoming more accurate; differences on normally scaled data are at floating-point noise level.
• No privacy, deployment, or proto impact.
• Rollback: revert this commit.

🤖 Generated with Claude Code