feat: Horn parallel analysis for component retention (paran oracle)

[seonghobae]

Iteration 7 of the autonomous paper-implementation loop. Stacked on #223 (base seonghobae-classification-accuracy).

Horn (1965) parallel analysis — Glorfeld (1995) centile variant

Oracle: Dinno's CRAN paran 1.5.6 R sources, read line by line (Horn 1965 and Glorfeld 1995 themselves are paywalled; cited as-implemented-by/as-cited-in paran — divergences documented in the module docs).

Contract (PCA path only): eigenvalues of the observed Pearson correlation matrix (cyclic Jacobi, Err on non-convergence) adjusted by sampling bias random_eigenvalue - 1 from n_iterations standard-normal same-shape datasets (single deterministic LCG stream); centile=0 = mean benchmark, 1..=99 = R type-7 quantile; retention = paran's left-to-right scan stopping at first adjusted <= 1 (resurgence does not count).

Deliberate divergences (module docs): PCA only (paran cfa/ginv path out of scope); single crate-LCG stream — paran-inspired, not bit-identical to any R run (paran.R line 35 references loop var k before definition, so the oracle does not faithfully implement its own seeding claim either); narrowed guards; wrapper supplies paran's 30 * n_items default.

Evidence

Check	Result
Fixture parity vs independent NumPy replication (mirrors crate LCG with np.uint64)	1e-9, first run
Mutation spot-checks	5/5 killed & restored (uncentered corr, no-interpolation quantile, bias w/o −1, count-vs-scan, no sort)
500-rep Monte Carlo (#[ignore], 2-factor n=200 p=10)	100% retained==2 (threshold 90%)
Adversarial spec-verify	GO-WITH-FIXES, 11 defects applied pre-implementation
Adversarial impl-review	2 MAJOR fixed + regressions (overflow-safe correlation guards, checked n_iterations * n_items), 1 MINOR (wrapper ValueError for negative scalars)
mlsirm-core suite	477 passed
Python wrapper tests	3 passed (fixture parity through binding, degenerate-input rejections, impl-review regressions)

[Copilot]

Pull request overview

This PR adds Horn (1965) / Glorfeld (1995) parallel analysis for PCA component retention to mlsirm-core, exposes it via the PyO3 extension and a thin Python validation/marshaling wrapper, and adds Rust + Python fixture-based tests plus changelog documentation.

Changes:

• Implement mlsirm_core::parallel::parallel_analysis (correlation → Jacobi eigenvalues → paran-style bias/adjustment → retention scan).
• Add Python API fast_mlsirm.parallel_analysis(...) -> ParallelAnalysisResult backed by the compiled Rust core and re-export it from fast_mlsirm.
• Add deterministic fixtures and degenerate-input rejection tests in both Rust and Python; document the feature in CHANGELOG.md.

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
crates/mlsirm-core/src/parallel.rs	New Rust implementation of Horn/Glorfeld parallel analysis (PCA path) with deterministic RNG and input guards.
crates/mlsirm-core/src/lib.rs	Exposes the new parallel module.
crates/fast-mlsirm-py/src/lib.rs	Adds PyO3 binding for parallel_analysis and registers it in the extension module.
python/fast_mlsirm/parallel_analysis.py	New Python wrapper: validates inputs, calls the Rust core, returns a dataclass result.
python/fast_mlsirm/__init__.py	Re-exports the new API (but currently drops ClassificationResult and omits ParallelAnalysisResult from __all__).
tests/unit/parallel_tests.rs	New Rust unit tests with pinned NumPy-derived fixtures, mutation anchors, and regression cases.
tests/test_paper_features.py	New Python-level parity/guard tests through the wrapper/binding.
CHANGELOG.md	Documents the new parallel analysis feature and its deliberate divergences from the R oracle.

Comments suppressed due to low confidence (1)

python/fast_mlsirm/init.py:178

• ParallelAnalysisResult is imported and part of the public surface, but it isn't included in __all__ alongside other *Result exports. This makes from fast_mlsirm import * omit the new result type and is inconsistent with existing exports like DetectResult/ClassificationResult.

    "rudner_classification",
    "lee_classification",
    "parallel_analysis",
    "ClassificationResult",
    "DetectResult",

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[seonghobae]

@copilot Fix the code for all comments in this review thread.

When a review comment includes a suggested change, apply the suggestion exactly.

Do not make changes beyond what is described in the linked review thread.

[Copilot]

@copilot Fix the code for all comments in this review thread.

When a review comment includes a suggested ch...

Both issues fixed in commit fix: add ClassificationResult import and optimize parallel.rs mean-path memory:

1. __init__.py — ClassificationResult added to the from .classification import (...) block so it is actually importable and consistent with __all__ (lines 45-49 now include it).

2. parallel.rs — The centile == 0 (mean) path now uses a single-pass running-sum accumulator per item position, avoiding the O(n_iterations × n_items) intermediate sim buffer entirely. The quantile path (centile != 0) is unchanged and still stores the full buffer for sorting.

[Copilot]

Pull request overview

Copilot reviewed 8 out of 8 changed files in this pull request and generated 3 comments.