Merge pull request #80 from TheHiddenObserver/codex/survival-gpu-completion

feat(survival): complete GPU Cox phase one

Author: TheHiddenObserver

.github/workflows/test.yml

@@ -89,6 +89,19 @@ jobs:
             dev/tests/test_logistic.py \
             dev/tests/test_cox.py \
             dev/tests/test_cox_cv.py \
+            dev/tests/test_cox_core_completion.py \
+            dev/tests/test_cox_phase1_completion.py \
+            dev/tests/test_penalized_cox_completion.py \
+            dev/tests/test_pr80_all_censored_loss.py \
+            dev/tests/test_pr80_post_review_fixes.py \
+            dev/tests/test_pr80_fit_boundary.py \
+            dev/tests/test_pr80_constructor_boundaries.py \
+            dev/tests/test_pr80_workspace_estimator.py \
+            dev/tests/test_pr80_cv_fit_boundary.py \
+            dev/tests/test_pr80_complete_review_cycle.py \
+            dev/tests/test_pr80_completion_contract_followup.py \
+            dev/tests/test_pr80_target_transfer_overflow_cache.py \
+            dev/tests/test_survival_risk_sets.py \
             dev/tests/test_distributions_backend.py \
             dev/tests/test_penalties_and_exports.py \
             dev/tests/test_ridge_inference.py \
@@ -176,10 +189,12 @@ jobs:
             statgpu/linear_model/cv/_ridge_cv.py \
             statgpu/linear_model/penalized/_fit_mixin.py \
             statgpu/linear_model/penalized/_inference_mixin.py \
+            statgpu/linear_model/penalized/_penalized_cox.py \
             statgpu/linear_model/penalized/_penalized_cv.py \
             statgpu/linear_model/penalized/_penalized_linear.py \
             statgpu/linear_model/wrappers/_linear.py \
             statgpu/linear_model/wrappers/_ridge.py \
+            statgpu/losses/_cox_ph.py \
             statgpu/metrics \
             statgpu/nonparametric/kernel_methods \
             statgpu/nonparametric/kernel_smoothing \
@@ -189,7 +204,15 @@ jobs:
             statgpu/penalties/_base.py \
             statgpu/semiparametric \
             statgpu/solvers/_fista_lla.py \
+            statgpu/survival/_concordance.py \
             statgpu/survival/_cox.py \
+            statgpu/survival/_cox_errors.py \
+            statgpu/survival/_cox_fit_adapter.py \
+            statgpu/survival/_cox_counting.py \
+            statgpu/survival/_cox_cv.py \
+            statgpu/survival/_numeric.py \
+            statgpu/survival/_cox_score.py \
+            statgpu/survival/_risk_sets.py \
             statgpu/unsupervised/_kmeans.py \
             statgpu/unsupervised/_nndescent.py \
             statgpu/unsupervised/_umap.py \

.gitignore

@@ -18,14 +18,28 @@ SECURITY.md
 .claude/
 
 # Results
+# Keep the final PR80 Exact benchmark source available to the PR76 dashboard
+# pipeline while continuing to ignore ad-hoc result artifacts.
 results/
+!results/
+results/*
+!results/benchmark_frontend_sources/
+results/benchmark_frontend_sources/*
+!results/benchmark_frontend_sources/coxph_exact_pr80_20260726.json
+!results/benchmark_frontend_sources/coxph_exact_strata_pr80_20260726.json
+!results/benchmark_frontend_sources/coxph_exact_delayed_entry_strata_pr80_20260727.json
+!results/benchmark_frontend_sources/coxph_exact_strata_count_pr80_20260727.json
+!results/benchmark_frontend_sources/penalized_cox_trusted_gradient_pr80_20260727.json
 
 # Dev tooling (benchmarks, scripts, plans, docs — not production code)
 dev/benchmarks/
 # Keep the PR79 evidence pipeline and its reviewed manifest versioned while
 # continuing to ignore ad-hoc benchmark artifacts elsewhere.
 !dev/benchmarks/
 dev/benchmarks/*
+!dev/benchmarks/benchmark_exact_ties_scaling.py
+!dev/benchmarks/benchmark_exact_strata_count_scaling.py
+!dev/benchmarks/benchmark_penalized_cox_trusted_gradient.py
 !dev/benchmarks/pr79/
 dev/benchmarks/pr79/*
 !dev/benchmarks/pr79/aggregate_results.py

CHANGELOG-history-through-2026-07-27.md

@@ -0,0 +1,226 @@
+# Changelog
+
+All notable changes to statgpu are documented here, organized by date and PR.
+
+## 2026-07-27
+
+### PR #80 — Cox review-fix follow-up
+- Reused Cox preprocessing across SCAD/MCP iterations and removed redundant objective, metadata, and finite-check transfers.
+- Unified CoxPH on stable risk-set objectives with a bounded ordinary suffix fast path, and restored cancellation-safe penalized moments.
+- Added Breslow Hessian workspace gates, preserved device/runtime solver errors, rejected complex high-level inputs, and unified no-pair scoring.
+- Added optimized stratified Exact and delayed-entry batching paths with maintained P100/R benchmark artifacts.
+- Hardened Cox/CV cleanup, packed-target provenance, prepared/fast-path integrity, fold-level metadata reuse, strict hazard-ratio exponentiation, backend-consistent prediction/strata validation, fixed-penalty frequentist inference, stable fit and raw-stop diagnostics, public numerical errors, robust-inference unit/PSD/rank/variance gates, shared inference/backend paths, valid eventless-stratum survival, survival-aware L1/L2/ElasticNet/SCAD/MCP CV with complete-evidence selection, fitted-backend pinning, CompositePenalty cloning, strict CV side-array/fold indices, general disjoint custom splits, promotion-safe scalar alpha-grid validation across all public penalty families, device-native GPU group metadata, scalar actual-fold and Cox evaluable-fold auto-device sizing, ElasticNet KKT alpha grids, operational auto-device fallback, non-tunable no-penalty rejection, and one-sync concordance tiling; inactive legacy kernels remain test-only.
+
+## 2026-07-26
+
+### PR #80 — Complete GPU Cox phase one
+- Added Breslow, Efron, and Exact Cox risk sets with delayed entry, start-stop rows, strata, robust inference, and subject-grouped CV across NumPy, CuPy, and Torch.
+- Hardened penalized Cox estimation, formula handling, sklearn compatibility, numerical stability, and backend-preserving prediction and scoring.
+- Added synchronized GPU and R validation artifacts for coefficients, likelihood, covariance, convergence, and performance.
+- Composed optimized Exact kernels across strata, cutting the `n=160` P100 full-fit time from 0.276/4.25/2.58 s to 0.0143/0.174/0.0747 s for NumPy/CuPy/Torch and preserving large-sample GPU acceleration.
+
+## 2026-07-25
+
+### PR #85 — Release statgpu 0.2.2
+
+- Bumped the package version from 0.2.1 to 0.2.2 in `pyproject.toml` and
+  `statgpu/__init__.py`.
+- Based the release candidate on the current `master`, including the PR #79
+  hardening work and PR #84 maintained-documentation refresh.
+- Retained the `STATGPU_NO_EXT=1` pure-Python `py3-none-any` wheel policy and sdist.
+- Validated 122 maintained documentation files, the full CPU-only suite, both
+  distribution formats, `twine check`, artifact contents, and clean installs.
+
+## 2026-07-24
+
+### PR #84 — Refresh maintained documentation contracts
+
+- Refreshed the release-facing README, documentation portals, method inventory,
+  and bilingual ANOVA, covariance, kernel-method, and PyTorch backend guides.
+- Added deterministic bilingual-link normalization and CI validation for
+  maintained relative links, release-facing content, and Python examples.
+
+### PR #79 — Exact-head review closure and documentation synchronization
+
+- Final reviewed production head `c85750d63d4e6dbc9d988847566c20f5fa862e91`
+  passed GitHub Actions Tests run #545, including Python 3.9–3.12, static contracts,
+  canonical smoke, and the full CPU suite.
+- The maintained Tesla P100 suite passed 33/33 executed checks with two expected skips;
+  ignored legacy diagnostic scripts are tracked separately in Issue #83.
+- Corrected the documented CoxPH delayed-entry contract: robust/cluster inference raises
+  when `compute_inference=True`, while `compute_inference=False` permits estimation-only
+  fits with inference fields unset.
+- Documented PooledOLS backend-preserving prediction, stable HAC `time_index` ordering,
+  effective-rank residual degrees of freedom, and rank-deficient coefficient inference as
+  `NOT_COMPARABLE` rather than `ERROR`.
+- Synchronized README, bilingual model pages, release notes, and the auditable PR79 report.
+- Removed stale hard-coded final accuracy artifacts; a new full canonical report may be
+  committed only after an exact-head full raw campaign is processed by the current
+  aggregator and renderer.
+
+## 2026-07-23
+
+### PR #79 — Complete review contract and evidence-pipeline hardening
+
+- Unified CoxPH final-KKT, line-search, termination-reason, and public fitted-state
+  contracts across CPU, CuPy, and Torch; failed CPU line searches no longer update
+  coefficients or report convergence.
+- Made delayed-entry penalty and robust-covariance limitations explicit, added
+  strict/approx robust inference with provenance fields, and introduced the
+  `statgpu[survival]` optional dependency.
+- Preserved estimator backends in Cox prediction/scoring, vectorized baseline
+  hazard risk sets, removed the affected Torch `O(n p^2)` Hessian materialization,
+  and avoided unconditional full training-data host transfers for nonrobust GPU inference.
+- Unified complex RBF rejection, Cox chi-square survival-function evaluation, and
+  CuPy Cholesky inverse solves.
+- Rebuilt PR79 diagnostic/canonical-report validation so missing, failed,
+  duplicate, non-finite, or wrong-SHA evidence fails closed; added CPU smoke CI.
+- Canonical evidence now requires clean, stable, exact-head Git provenance; stale
+  hard-coded final PASS artifacts are not authoritative and must not be regenerated
+  without a full validated campaign.
+- Added behavioral regression coverage and synchronized the English/Chinese Cox
+  support matrix.
+
+## 2026-07-21
+
+### PR #79 — Final physical GPU validation and correctness hardening
+
+- Completed GPU smoke, three-backend correctness, metamorphic, device-purity,
+  memory-leak, performance, external-validation, and full CPU/GPU gates on Tesla P100.
+- Full campaign result on `2f18e5d`: 1100 passed, 0 failed, 124 skipped, and
+  1 version-limited strict XFAIL; all 40 initial Gate B failures were eliminated or
+  formally dispositioned.
+- Completed a subsequent review-fix cycle covering backend-native `LinearRegression`,
+  PooledOLS HAC ordering and effective rank, formula-weight alignment, validator integrity,
+  weighted CPU/CuPy/Torch fitting, and degenerate GPU F-statistic semantics.
+- Exact-head physical GPU acceptance on clean SHA
+  `786af9e2eb4742a56e5203b4380b03aec63a3ac8`: **17 passed, 0 failed, 0 skipped**
+  in 7.28 seconds, with CuPy and Torch CUDA tests both executed.
+- Degenerate F tests now agree across backends: perfect non-constant fit returns
+  `(inf, 0.0)`; intercept-only and otherwise undefined overall tests return `(nan, nan)`.
+- Follow-up issues #81, #82, and #83 remain non-blocking; see
+  `dev/reviews/pr79_physical_gpu_validation.md`.
+
+## 2026-07-14
+
+### PR #79 — Third review/fix cycle
+
+- Fixed Torch vector Cholesky solves, Panel string-label/device paths, KernelPCA/RidgeCV/
+  thin-plate Torch failures, and full-design CPU fallbacks in panel array workflows.
+- Added shared finite-input validation for panel, covariance, unsupervised, KernelPCA,
+  Nystroem, and thin-plate paths plus 21 focused regressions.
+- The physical-GPU work pending at this stage was completed on 2026-07-21; see the final
+  validation entry and `dev/reviews/pr79_physical_gpu_validation.md`.
+
+## 2026-07-12
+
+### PR #79 — Second full-repository review and auto-fix
+
+- Fixed Stepwise backward selection/order/state contracts, backend-native Welch ANOVA,
+  incomplete-fold CV selection, regression diagnostics, summary-statistic edge cases,
+  Torch RBF kernels, weighted quadratic SCAD/MCP routing, resampling validation, and
+  Cox score-test duplication.
+- Hardened estimator cloning, knockoff selectors/draw validation, composite penalties,
+  effect sizes, backend factory semantics, KDE zero-density handling, and dtype/device
+  preservation; added 40+ focused regression tests and synchronized public docs.
+
+### PR #79 — Native three-backend execution follow-up
+
+- Removed complete numeric-array NumPy fallbacks from `GraphicalLasso`,
+  `GraphicalLassoCV`, `MinCovDet`, `SplineTransformer`, and `FamaMacBeth`.
+- Kept Graphical Lasso block-coordinate descent/CV, FAST-MCD C-steps and
+  reweighting, spline Cox–de Boor recurrence, and Fama–MacBeth regressions/HAC
+  covariance on the selected NumPy, CuPy, or Torch backend.
+- Kept Tukey/Bonferroni group reductions on-device; only scalar distribution
+  CDF/quantile evaluations cross the CPU boundary.
+- Added NumPy/Torch parity and backend-preservation tests plus optional CuPy CUDA
+  checks. The physical CuPy/Torch CUDA validation planned at this stage was completed
+  on 2026-07-21.
+- Synchronized README, bilingual implemented-method lists, model pages, and all
+  three changelogs with the corrected execution and validation boundaries.
+
+### PR #79 — Public module statistical-contract follow-up
+
+- Extended the repository review beyond Ridge to every top-level public module family,
+  combining full-package high-signal static analysis with targeted numerical invariants,
+  nested-model checks, and parity comparisons against established reference libraries.
+- Corrected two-way ANOVA residual and balance semantics, Welch/post-hoc degenerate cases,
+  chi-square kernels, KernelRidge/KernelRidgeCV scoring, KernelPCA embedding consistency,
+  and Nystroem normalization for indefinite kernels.
+- Corrected empirical precision estimation, Graphical Lasso block-coordinate updates,
+  MinCovDet centered semantics, panel cluster/HAC contracts, Patsy side-array alignment,
+  and rank-deficient panel regression fallbacks.
+- Implemented real spline extrapolation modes; hardened B-spline, KDE, kernel regression,
+  GAM, and binary-metric input contracts.
+- Added three focused regression suites and expanded the permanent Python 3.9–3.12,
+  full-CPU, static-contract, compilation, and complete-collection gates.
+- The physical CuPy/Torch CUDA numerical, memory, type/device, and performance validation
+  planned at this stage was completed on 2026-07-21.
+
+### PR #79 — Ridge objective and weighted-path consistency follow-up
+
+- Confirmed that statgpu Ridge uses the package-wide average-loss objective rather
+  than scikit-learn's unnormalized residual-sum-of-squares convention.
+- Preserved the exact normal equations `X'X + n*alpha*I` for unweighted fits and
+  `X'WX + sum(w)*alpha*I` for weighted fits; scikit-learn comparisons now use the
+  explicit corresponding alpha mapping.
+- Unified weighted Ridge behavior across the optimized wrapper, generic exact solver,
+  FISTA, formula fitting, CPU/CuPy/Torch exact paths, Gaussian inference, RidgeCV,
+  and `PenalizedGLM_CV(loss="squared_error", penalty="l2")`.
+- Corrected weighted centering before square-root weighting, weighted intercept and
+  residual construction for inference, and weighted default alpha-grid generation.
+- `PenalizedGLM_CV` now generates weighted alpha grids from the normalized weighted
+  null gradient and avoids building an unused host-side Gram cache for the default
+  GPU Newton Ridge route.
+- Formula evaluation now exposes retained row positions so sample weights remain
+  aligned when Patsy drops rows containing missing values.
+- GPU sample-weight validation and normalization use device-side reductions and
+  synchronize only scalar results, avoiding full weight-vector host transfers.
+- Added regression coverage for weighted closed forms, weight-rescaling invariance,
+  exact/FISTA and wrapper/generic equality, formula missing rows, inference covariance,
+  both Ridge CV implementations, and weighted scikit-learn alpha mapping.
+
+## 2026-07-11
+
+### PR #79 — Full repository review and hardening
+
+- Completed an iterative repository-wide review covering correctness, backend routing,
+  statistical/API contracts, readability, maintainability, extensibility, performance
+  risks, test quality, and compliance with `dev/AGENTS.md`.
+- Fixed backend/device validation, sklearn-style estimator parameters, Torch inference
+  routing, UMAP fuzzy-union and random-state semantics, NNDescent correctness, adaptive
+  L1 and knockoff runtime errors, CV input contracts, KMeans/UMAP edge cases, and Cox
+  Efron observed-information orientation.
+- Hardened tests so optional Torch/CuPy dependencies skip explicitly instead of failing
+  collection or swallowing unexpected errors; moved the remote GPU runner outside the
+  pytest test tree.
+- Added focused review regression suites and permanent Python 3.9–3.12, full CPU,
+  compilation, static-contract, and complete test-collection CI gates.
+- Added `dev/reviews/pr79_full_repository_review.md` with accepted fixes, deferred
+  architectural debt, and the physical-GPU validation plan.
+- The physical CuPy/Torch CUDA numerical, memory, and performance validation required at
+  this stage was completed on 2026-07-21.
+
+## 2026-07-08
+
+### v0.2.1 — Packaging / PyPI release hygiene
+
+- **Version bump** 0.2.0 → 0.2.1 (`pyproject.toml`, `statgpu/__init__.py`).
+- **Pure-Python wheel policy**: the PyPI release workflow now sets `STATGPU_NO_EXT=1`,
+  so the published wheel is tagged `py3-none-any` and installs on every OS / Python
+  version. Previously `python -m build` compiled the optional Cython extensions during
+  `bdist_wheel`, producing a platform-locked wheel that served almost no one and forced
+  everyone else onto the sdist.
+- **setup.py**: added the `STATGPU_NO_EXT` switch. The Cython extensions remain optional
+  CPU accelerators with pure-Python fallbacks.
+- **publish.yml**: added `twine check dist/*` before upload.
+
+### PR #74 — Ordered Newton-Raphson + Analytical Hessian Inference + Unified Sandwich Engine
+
+- Ordered Logit/Probit: L-BFGS replaced with Newton-Raphson + trust-region (3-backend).
+- Ordered inference: analytical Hessian, SE/z/p/CI, loglikelihood/aic/bic (CPU+GPU).
+- Sandwich engine: m-estimation inference, Fisher information, and penalty curvature API.
+- Penalized inference: sandwich (L2/EN), oracle active-set (SCAD/MCP).
+- QuantileRegression standalone class with kernel and bootstrap inference.
+- 28 bug fixes across four code-review rounds; scipy distribution calls routed through
+  the project distribution abstraction where applicable.