feat: add measured experimental feature engineering improvements by thinkall · Pull Request #9 · thinkall/featcopilot

thinkall · 2026-10-10T13:54:54Z

Summary

Add opt-in learner-aware tabular selection, fitted rules, point-in-time features and explicit classification ensembles.
Harden sklearn integration, optional-package typing and provider error handling without fabricated feature fallbacks.
Add matched, resource-bounded, resumable benchmarks with source/split provenance, dataset admission, regression averaging and training-only refit diagnostics.
Preserve complete positive and negative evidence, including failed reference cells and blocked confirmation decisions. Keep legacy defaults unchanged.

Measured improvements and limits

On the exposed development panels, three-seed classifier averaging reduced mean log loss by 6.82% versus its matched raw control. Two-seed regression replication reduced primary-panel mean RMSE by 23.79% for linear single FE and 2.97% for LightGBM averaging.

These results do not establish SOTA or justify default promotion. Individual hardware folds still suffered severe harms, and independent confirmation remains unresolved. The committed reports retain those failures and the subsequent training-only numerical investigation.

Validation

Full suite exercised 3,729 tests: initially 3,704 passed, 19 skipped and six failed. The six failures were corrected in two test modules: a functional worker test's startup allowance and relational fake estimators' prediction contract. Both affected modules then passed all 265 tests under coverage.
Package coverage: 93.48%, above the configured 85% requirement.
Package mypy, isolated wheel import/fit and mandatory pre-commit run --all-files passed.
Verified 63 archive indexes and exact staged bytes for 591 frozen benchmark artifacts/descriptors. Git attributes and formatter exclusions preserve their checksum-bound bytes.

Unrelated submissions, raw datasets and generated test scratch are excluded. This PR publishes commit 5367d34c84eb42b8580250c4e851d00406154894.

Add opt-in learner-aware tabular selection, fitted rules, point-in-time features and explicit classification ensembles while retaining legacy defaults. Harden sklearn integration, provider errors and optional-package typing without fabricated feature fallbacks. Add matched, resource-bounded, resumable benchmarks with dataset admission, source and split provenance, exact controls and preserved failed outcomes. Include regression averaging and training-only refit diagnostics. Retain the full development and confirmation evidence, including severe tail harms and blocked decisions. Replicated mean gains support these experimental improvements, not SOTA or default promotion. Preserve hash-bound evidence bytes in Git and keep generated test scratch, raw datasets and unrelated submissions out of the commit. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

chatgpt-codex-connector · 2026-10-10T13:55:06Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-10T14:00:38.411347Z	`5367d34`	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

Copilot

🟡 Changes recommended

The newly registered frequency operator cannot be executed through the registry’s public feature-creation contract because it requires an unprovided mapping argument.

1 open finding

Stateful frequency mapping violates unary registry contract · New

What changed in this PR

Adds opt-in learner-aware feature engineering, leakage-safe fitted transforms, provider hardening, and reproducible benchmark evidence while preserving legacy defaults.

Changes:

Introduces fitted, predictive, categorical, and point-in-time feature workflows.
Hardens sklearn/LLM integration, typing, validation, and error handling.
Publishes bounded benchmark protocols, diagnostics, and archived results.

File	Description
`tests/test_utils.py`	Updates model-default tests.
`tests/test_transform_rules.py`	Tests shared execution namespace.
`tests/test_sklearn_compat.py`	Tests cloning and cross-fitting.
`tests/test_selection.py`	Tests constant correlations.
`tests/test_paired_selection.py`	Tests paired selection policy.
`tests/test_litellm.py`	Adds isolated LiteLLM fixture.
`tests/test_engines.py`	Tests categorical encoding behavior.
`tests/test_core.py`	Tests registry-created execution.
`pyproject.toml`	Adds benchmark dependencies.
`mkdocs.yml`	Adds classification guide.
`featcopilot/utils/parallel.py`	Normalizes parallel results.
`featcopilot/utils/models.py`	Updates model and SDK handling.
`featcopilot/utils/frequency.py`	Adds frozen frequency mappings.
`featcopilot/utils/cache.py`	Validates metadata shape.
`featcopilot/selection/redundancy.py`	Guards missing correlations.
`featcopilot/llm/explainer.py`	Tightens client typing.
`featcopilot/llm/code_generator.py`	Centralizes code normalization.
`featcopilot/llm/_types.py`	Defines provider protocols.
`featcopilot/engines/text.py`	Improves transformers typing.
`featcopilot/engines/__init__.py`	Exports experimental engines.
`featcopilot/core/transform_rule.py`	Fixes execution scoping.
`featcopilot/core/registry.py`	Registers frequency transformation.
`featcopilot/core/feature.py`	Resolves registered callables.
`featcopilot/core/base.py`	Adds training-transform hook.
`featcopilot/core/__init__.py`	Exports fitted-rule models.
`featcopilot/__init__.py`	Exposes fitted-rule APIs.
`docs/user-guide/engines.md`	Documents new engines.
`docs/user-guide/cli.md`	Documents predictive CLI options.
`docs/index.md`	Revises benchmark claims.
`docs/getting-started/quickstart.md`	Adds leakage-safe pipeline example.
`docs/getting-started/authentication.md`	Documents provider behavior.
`docs/examples/relational-feature-engineering.md`	Adds point-in-time example.
`docs/examples/llm-powered.md`	Updates model guidance.
`benchmarks/tabular-results/2026-10-08-regression-training-audit/protocol.json`	Records audit protocol.
`benchmarks/tabular-results/2026-10-08-regression-training-audit/launcher_timing.json`	Records runtime.
`benchmarks/tabular-results/2026-10-08-regression-training-audit/auxiliary_archive.json`	Indexes audit source.
`benchmarks/tabular-results/2026-10-08-regression-training-audit/archive.json`	Indexes audit results.
`benchmarks/tabular-results/2026-10-08-regression-policy-replication/launcher_timing.json`	Records replication runtime.
`benchmarks/tabular-results/2026-10-08-regression-policy-replication/auxiliary_archive.json`	Indexes replication inputs.
`benchmarks/tabular-results/2026-10-08-regression-policy-replication/archive.json`	Indexes replication results.
`benchmarks/tabular-results/2026-10-08-regression-average-development/launcher_timing.json`	Records development runtime.
`benchmarks/tabular-results/2026-10-08-regression-average-development/controller.json`	Records controller provenance.
`benchmarks/tabular-results/2026-10-08-regression-average-development/auxiliary_archive.json`	Indexes development inputs.
`benchmarks/tabular-results/2026-10-08-regression-average-development/archive.json`	Indexes development results.
`benchmarks/tabular-results/2026-10-04-thoracic-fixed-average-confirmation/auxiliary_archive.json`	Indexes confirmation inputs.
`benchmarks/tabular-results/2026-10-04-thoracic-fixed-average-confirmation/archive.json`	Indexes confirmation results.
`benchmarks/tabular-results/2026-10-04-thoracic-fixed-average-confirmation/ANALYSIS.md`	Reports blocked confirmation.
`benchmarks/tabular-results/2026-10-04-oof-blend-development/decision.json`	Records blend decision.
`benchmarks/tabular-results/2026-10-04-oof-blend-development/auxiliary_archive.json`	Indexes blend inputs.
`benchmarks/tabular-results/2026-10-04-oof-blend-development/archive.json`	Indexes blend results.
`benchmarks/tabular-results/2026-10-04-fixed-average-development/decision.json`	Records averaging decision.
`benchmarks/tabular-results/2026-10-04-fixed-average-development/auxiliary_archive.json`	Indexes averaging inputs.
`benchmarks/tabular-results/2026-10-04-fixed-average-development/archive.json`	Indexes averaging results.
`benchmarks/tabular-results/2026-10-04-autofeat-safe-scale-diagnostic/launcher_timing.json`	Records diagnostic runtime.
`benchmarks/tabular-results/2026-10-04-autofeat-safe-scale-diagnostic/auxiliary_archive.json`	Indexes diagnostic source.
`benchmarks/tabular-results/2026-10-04-autofeat-safe-scale-diagnostic/archive.json`	Indexes diagnostic results.
`benchmarks/tabular-results/2026-10-03-target-encoding-development/decision.json`	Records encoding decision.
`benchmarks/tabular-results/2026-10-03-target-encoding-development/auxiliary_archive.json`	Indexes encoding inputs.
`benchmarks/tabular-results/2026-10-03-target-encoding-development/archive.json`	Indexes encoding results.
`benchmarks/tabular-results/2026-10-03-subsample85-development/decision.json`	Records subsampling decision.
`benchmarks/tabular-results/2026-10-03-subsample85-development/auxiliary_archive.json`	Indexes subsampling inputs.
`benchmarks/tabular-results/2026-10-03-subsample85-development/archive.json`	Indexes subsampling results.
`benchmarks/tabular-results/2026-10-03-qualified-policy-confirmation/auxiliary_archive.json`	Indexes policy inputs.
`benchmarks/tabular-results/2026-10-03-qualified-policy-confirmation/archive.json`	Indexes policy results.
`benchmarks/tabular-results/2026-10-03-classifier-extraction-parity/parity.json`	Records parity checks.
`benchmarks/tabular-results/2026-10-03-classifier-extraction-parity/auxiliary_archive.json`	Indexes parity inputs.
`benchmarks/tabular-results/2026-10-03-classifier-extraction-parity/archive.json`	Indexes parity results.
`benchmarks/tabular-results/2026-10-03-classifier-extraction-parity/ANALYSIS.md`	Reports classifier parity.
`benchmarks/tabular-results/2026-10-01-calibration-replication/REPORT.md`	Reports calibration replication.
`benchmarks/tabular-results/2026-10-01-calibration-replication/auxiliary_archive.json`	Indexes replication inputs.
`benchmarks/tabular-results/2026-10-01-calibration-replication/archive.json`	Indexes replication results.
`benchmarks/tabular-results/2026-10-01-calibration-development/REPORT.md`	Reports calibration development.
`benchmarks/tabular-results/2026-10-01-calibration-development/auxiliary_archive.json`	Indexes calibration inputs.
`benchmarks/tabular-results/2026-10-01-calibration-development/archive.json`	Indexes calibration results.
`benchmarks/tabular-results/2026-09-30-robust-confirmation/auxiliary_archive.json`	Indexes confirmation inputs.
`benchmarks/tabular-results/2026-09-30-robust-confirmation/archive.json`	Indexes confirmation results.
`benchmarks/tabular-results/2026-09-30-fixed-plan-development/decision.json`	Records fixed-plan decision.
`benchmarks/tabular-results/2026-09-30-fixed-plan-development/auxiliary_archive.json`	Indexes fixed-plan inputs.
`benchmarks/tabular-results/2026-09-30-fixed-plan-development/archive.json`	Indexes fixed-plan results.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/request-5.json`	Records audit request.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/request-4.json`	Records audit request.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/request-3.json`	Records audit request.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/request-2.json`	Records audit request.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/request-1.json`	Records audit request.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/request-0.json`	Records audit request.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/launcher_timing.json`	Records audit runtime.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/auxiliary_archive.json`	Indexes audit inputs.
`benchmarks/tabular-results/2026-09-29-training-selection-audit/archive.json`	Indexes audit results.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/resource_accounting.json`	Records resource usage.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-21.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-20.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-17.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-16.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-13.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-12.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-09.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-08.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-05.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-04.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-01.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/record-00.json`	Preserves audit result.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/launcher_timing.json`	Records audit runtime.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/archive.json`	Indexes audit results.
`benchmarks/tabular-results/2026-09-29-robust-training-audit/.gitignore`	Excludes local audit artifacts.
`benchmarks/tabular-results/2026-09-29-robust-development/phase1_checkpoint.json`	Records resumable checkpoint.
`benchmarks/tabular-results/2026-09-29-robust-development/auxiliary_archive.json`	Indexes development inputs.
`benchmarks/tabular-results/2026-09-29-robust-development/archive.json`	Indexes development results.
`benchmarks/tabular-results/2026-09-29-bagged-development/decision.json`	Records bagging decision.
`benchmarks/tabular-results/2026-09-29-bagged-development/auxiliary_archive.json`	Indexes bagging inputs.
`benchmarks/tabular-results/2026-09-29-bagged-development/archive.json`	Indexes bagging results.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/supplementary_archive.json`	Indexes oversized records.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/record-15.json`	Preserves aligned result.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/record-12.json`	Preserves aligned result.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/record-09.json`	Preserves aligned result.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/record-06.json`	Preserves aligned result.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/record-03.json`	Preserves aligned result.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/record-00.json`	Preserves aligned result.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/launcher_timing.json`	Records aligned runtime.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/auxiliary_archive.json`	Indexes aligned inputs.
`benchmarks/tabular-results/2026-09-29-aligned-training-audit/archive.json`	Indexes aligned results.
`benchmarks/tabular-results/2026-09-29-aligned-development/auxiliary_archive.json`	Indexes development inputs.
`benchmarks/tabular-results/2026-09-29-aligned-development/archive.json`	Indexes development results.
`benchmarks/tabular-results/2026-09-28-frequency-iid-confirmation/auxiliary_archive.json`	Indexes frequency inputs.
`benchmarks/tabular-results/2026-09-28-frequency-iid-confirmation/archive.json`	Indexes frequency results.
`benchmarks/tabular-results/2026-09-24-logistic-worker-efficiency/auxiliary_archive.json`	Indexes efficiency inputs.
`benchmarks/tabular-results/2026-09-24-logistic-worker-efficiency/archive.json`	Indexes efficiency results.
`benchmarks/tabular-results/2026-09-24-frozen-native-confirmation/auxiliary_archive.json`	Indexes native inputs.
`benchmarks/tabular-results/2026-09-24-frozen-native-confirmation/archive.json`	Indexes native results.
`benchmarks/tabular-results/2026-09-24-frequency-development/auxiliary_archive.json`	Indexes frequency inputs.
`benchmarks/tabular-results/2026-09-24-frequency-development/archive.json`	Indexes frequency results.
`benchmarks/tabular-results/2026-09-23-joint-importance-development/auxiliary_archive.json`	Indexes importance inputs.
`benchmarks/tabular-results/2026-09-23-joint-importance-development/archive.json`	Indexes importance results.
`benchmarks/tabular-results/2026-09-23-extended-development/auxiliary_archive.json`	Indexes extended inputs.
`benchmarks/tabular-results/2026-09-23-extended-development/archive.json`	Indexes extended results.
`benchmarks/tabular-results/2026-09-23-extended-development/ANALYSIS.md`	Reports extended comparison.
`benchmarks/tabular-results/2026-09-22-full-reference-development/auxiliary_archive.json`	Indexes reference inputs.
`benchmarks/tabular-results/2026-09-22-full-reference-development/archive.json`	Indexes reference results.
`benchmarks/tabular-results/2026-09-21-retail-relational/ANALYSIS.md`	Reports temporal evaluation.
`benchmarks/tabular-results/2026-09-21-confirmation-completion/archive.json`	Indexes completion results.
`benchmarks/tabular-results/2026-09-21-confirmation-completion/ANALYSIS.md`	Reports confirmation completion.
`benchmarks/tabular-results/2026-09-20-seoul-temporal/archive.json`	Indexes temporal results.
`benchmarks/tabular-results/2026-09-20-seoul-temporal/ANALYSIS.md`	Reports temporal findings.
`benchmarks/tabular-results/2026-09-20-locked-confirmation/archive.json`	Indexes locked results.
`benchmarks/tabular-results/2026-09-20-locked-confirmation/ADMISSION_STATUS.md`	Records admission failures.
`benchmarks/tabular-results/2026-09-20-llm-pilot/archive.json`	Indexes LLM results.
`benchmarks/tabular-results/2026-09-20-llm-pilot/ANALYSIS.md`	Reports LLM pilot.
`benchmarks/tabular-results/2026-09-17-roadmap-pilot/archive.json`	Indexes roadmap results.
`benchmarks/tabular_suite.json`	Defines development datasets.
`benchmarks/tabular_openfe.py`	Adds OpenFE compatibility.
`benchmarks/suites/seoul_inventory.json`	Registers temporal inventory.
`benchmarks/suites/regression_training_audit_20261008.json`	Freezes audit configuration.
`benchmarks/datasets.py`	Validates loader contracts.
`.pre-commit-config.yaml`	Protects frozen artifacts.
`.github/workflows/tests.yml`	Adds mypy CI check.
`.gitattributes`	Preserves benchmark bytes.

🧠 Review effort: Balanced

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

Keep fitted frequency mappings out of the public unary transformation registry while retaining their internal callable during stored-feature execution. Cover all default registry transformations and serialized frequency features. Preserve optional XGBoost constructors through the existing model protocol, convert NumPy index/count scalars before division, and annotate variable array shapes for the CI Python/NumPy matrix without suppressing errors. Leave archived benchmark evidence and experimental defaults unchanged. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

The exceptionally large experimental and checksum-bound benchmark surface warrants final human validation despite extensive testing.

0 open findings

1 resolved since last review

Stateful frequency mapping violates unary registry contract

🧠 Review effort: Balanced

Declare process-test dependencies and exercise the optional LLM adapter in CI. Keep pure mock tests independent of the SDK, support its explicit stdio transport, and report missing SDK configuration before opening data. Preserve current and legacy session cleanup without type suppression. Restore the original checksum-bound CRLF bytes of the two legacy cache descriptors and preserve them through Git and pre-commit. Their parsed contents and recorded checksums remain unchanged. Verify logging helper restoration against each handler's original stream instead of stderr only. Keep the native solver and its parity tests unchanged; the isolated Python 3.10 reproduction did not reproduce the CI native crash. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

Its 339,000-line scope combines new public ML APIs with extensive checksum-bound experimental evidence requiring final human validation.

0 open findings

🧠 Review effort: Balanced

Require SDK 0.3.0 or newer to match the existing keyword-only session APIs. Those SDK versions require Python 3.11+, while the core package remains compatible with Python 3.10. Do not install an older incompatible SDK in the Python 3.10 CI job; continue running its core and mock-provider tests. Document the optional-extra requirement and exercise supported SDK contracts in the Python 3.11-3.13 jobs. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

Its very large scope combines experimental ML behavior, provider integrations, benchmark machinery, and hundreds of immutable evidence artifacts requiring final human validation.

0 open findings

🧠 Review effort: Balanced

codecov · 2026-10-10T18:21:18Z

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 93.34%. Comparing base (bf11981) to head (e8cc44d).

Additional details and impacted files

@@            Coverage Diff             @@
##             main       #9      +/-   ##
==========================================
+ Coverage   89.34%   93.34%   +3.99%     
==========================================
  Files          38       52      +14     
  Lines        4364     7796    +3432     
==========================================
+ Hits         3899     7277    +3378     
- Misses        465      519      +54

Flag	Coverage Δ
unittests	`93.34% <ø> (+3.99%)`	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

Run the unchanged logistic worker tests and full suite under GDB on the failing Linux/Python 3.10 lane. Preserve child exit status and all test assertions so the repeated SciPy native crash can be diagnosed rather than hidden by a solver change or skip. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

The Linux native backtrace faults in SciPy's dtrsm_iunncopy_SKYLAKEX worker. Compare the unchanged tests and solver with the same numerical packages using the Haswell kernel, and log the effective BLAS implementation before both diagnostic runs. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

SDK metadata serialization changes public field names, and shipped examples still contradict the new no-mock/default-model contract.

0 open findings

🧠 Review effort: Balanced

The native crash reproduces with both SkylakeX and Haswell kernels in SciPy's bundled OpenBLAS 0.3.28. Its worker-growth path matches the missing buffer initialization fixed by OpenMathLib/OpenBLAS#4879. Preinitialize the eight workers required by the unchanged tests before Python starts. Restore automatic CPU dispatch; do not reduce thread caps, change solvers, skip assertions, or change numerical dependencies. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

Its broad API, leakage-control, CI, dependency, and checksum-bound benchmark changes warrant final human verification despite no specific unresolved defect found.

0 open findings

🧠 Review effort: Balanced

Copilot

🔵 Needs a closer look

LLM code normalization and validation incorrectly accept assignments to result that exist only in nested scopes.

0 open findings

🧠 Review effort: Balanced

The reproduced x86-64 Linux crash follows OpenBLAS 0.3.28's missing worker-buffer initialization path. SciPy 1.15.0 through 1.15.3 share the affected scipy-openblas32 0.3.28.0.2 wheel-build dependency. Exclude only those four releases on the reproduced platform. SciPy 1.14.1 remains compatible with Python 3.10 and NumPy 2.2.6, while other platforms and future patched releases remain unconstrained by this fix. Restore ordinary full-suite execution on every matrix lane and remove temporary GDB, CPU-dispatch and startup-thread diagnostics. Keep the L-BFGS-B solver, engine thread caps and all test assertions unchanged. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

The single-row target-encoding fallback leaks the training target into its own encoded feature.

0 open findings

🧠 Review effort: Balanced

thinkall · 2026-10-10T19:17:10Z

Review feedback and CI failures are addressed through 46c7db92165cb4078699c3f4cc00ae9855f3c002.

Resolved the unary-registry finding while preserving stored frequency-feature execution and serialization.
Fixed portable NumPy/optional-model typing, SDK-independent mocks, missing-SDK preflight errors, current/legacy SDK transport and cleanup, and the logging-fixture stream assumption.
Restored the original checksum-bound bytes of both legacy cache descriptors; their parsed data and recorded checksums were not changed.
Declared the Copilot SDK version actually required by the existing APIs. Core Python 3.10 remains supported; the Copilot extra requires Python 3.11+.
Isolated the Linux native crash to the OpenBLAS 0.3.28 worker-buffer path documented by OpenMathLib/OpenBLAS#4879. The final fix excludes SciPy 1.15.0-1.15.3 only on Python 3.10/Linux/x86-64, allowing SciPy 1.14.1 there. Other platforms and future patched releases remain allowed.

Final ordinary CI is green: Python 3.10, 3.11, 3.12 and 3.13 tests/type checks, pre-commit, and both Codecov checks passed in run 38077674235. The formerly crashing Python 3.10 lane used SciPy 1.14.1 and passed 3,554 tests, including both native binary/multiclass parity cases, with 199 reported skips and 93.37% coverage.

Temporary GDB, CPU-kernel and startup-thread overrides have been removed. The native solver, eight-thread assertions, full core matrix, experimental defaults and benchmark outcomes remain unchanged.

The existing review thread is resolved. GitHub reports the PR mergeable, with required review approval still outstanding.

## Summary Add the independent grok-4.7 stage-2 architecture review for PR 9. It reports four open design-contract issues: plan-provider ownership, screening versus deployment preprocessing, benchmark controller separation, and the legacy gate_n_jobs knob. ## Prompting Intent Run an independent High stage-2 Architecture and patterns review of PR 9 from base bf11981 through 46c7db9 plus uncommitted review fixes. Do not read other review artifacts, edit product code, call providers, or run real benchmarks. Write only the assigned Markdown artifact and commit it. The parent owns fixes and tests. ## Linked Sources - Review contract: session review-pr9/stage2-attempt1/review-round-2.md - Scope: session review-pr9/scope.json - Pull request: #9 - Base: bf11981 - No ADO work item id was supplied; the assigned slug is task-9 ## Rationale The artifact records findings without product-code fixes because the assignment reserves fixes and tests for the parent. Only concrete API and dependency-direction defects are included. The gate_n_jobs coupling is classified as pre-existing after checking the base commit, rather than reported as a new regression. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

## Summary Repair defects identified by the six-stage High review of PR #9 and retain all 24 independent review artifacts with their original verdicts and appended resolution history. Harden fitted schemas, categorical and target semantics, optional diagnostics, isolated provider lifecycles, resumable benchmark accounting, source identity, worker supervision and atomic publication. Remove verified avoidable allocation and history-scan costs, align optional benchmark dependencies, and correct public documentation. ## Prompting Intent The engineer asked to use the review-code skill to review PR #9, following earlier requests to address PR comments and CI failures. The engineer explicitly authorized a High two-model override using gpt-6.1-sol at max reasoning and grok-4.7 at xhigh, waiving unavailable Opus while retaining all six ordered stages and the two-attempt stage cap. Preserve legacy defaults, unrelated work and checksum-bound historical evidence; make no unsupported SOTA claim. ## Linked Sources - Pull request: #9 - Review scope base: bf11981 - Starting reviewed head: 46c7db9 - Review decisions, findings and resolution history: reviews/pr9/ - Repository contributor policy: AGENTS.md - Benchmark contract: benchmarks/TABULAR_PROTOCOL.md - Reference environment and execution instructions: benchmarks/README.md - Public behavior contracts: docs/user-guide/ and docs/examples/relational-feature-engineering.md - No Azure DevOps task was supplied; linkage is to GitHub PR #9. ## Rationale Fix concrete introduced defects at their owning boundaries rather than change defaults or regenerate historical scores. Share isolated typed-plan transport without redesigning legacy code mode. Preserve primary errors and privacy-safe diagnostics, exact source-byte identity, immutable evidence generations and training-only feature semantics. Avoid mtime-only fingerprint caches and unnecessary per-cell report work. Keep pre-existing concerns and capped dirty stage outcomes explicit: repair validation is not a new independent clean review, and this work does not establish SOTA or a general no-harm guarantee. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

thinkall · 2026-10-14T18:25:16Z

High review completed; fixes published in `08317c7`

Completed all six ordered review-code stages with gpt-6.1-sol (max)
and grok-4.7 (xhigh), using the explicitly approved two-model override.
All 24 independent review artifacts, original verdicts, driver decisions
and repair history are committed in
reviews/pr9.

The repairs cover fitted output schemas and target/categorical semantics,
isolated plan-provider lifecycle and safe diagnostics, resumable benchmark
accounting and immutable publication, worker cleanup, source-identity checks,
avoidable history scans/candidate allocations, dependency constraints and
documentation. Historical numerical evidence and legacy defaults were not
rewritten; no new real-data measurements or live provider calls were made.

This is not a blanket clean verdict. Each stage reached its two-attempt
cap with findings; subsequent repairs do not retroactively change those
verdicts. Explicit follow-ups include independently verified pre-existing
code-mode permission/parsing behavior, empty categorical input handling,
legacy documentation/loader warnings, and broader architectural refactors.
Details and limitations remain in the artifacts.

Validation: The frozen-source full suite completed with 4,217 passes,
19 skips and one stale test fault-injection failure. That test was corrected
to target serialization rather than the former combined hash helper, and all
18 related tests then passed, including strict propagation of three exception
types. Production bytes were unchanged; coverage is 93.42%. Mypy passed
52 files, wheel isolation passed, and mandatory full pre-commit passed.
All frozen-evidence checksums remain intact.

Fresh Python 3.10-3.13 CI is running on the published commit:
Tests run.
No merge or SOTA claim is implied.

Copilot

🔵 Needs a closer look

The changes require final human review because they are too complex or risky for automated approval.

0 open findings

🧠 Review effort: Balanced

## Summary Fix the two cross-environment failures in the published PR review repairs. Keep isolated-plan return typing explicit when the optional SDK is absent. Check target-encoding batch invariance exactly at the learner input while using the established numerical tolerance for final probabilities. ## Prompting Intent Complete the requested review-code review of PR #9 and address resulting CI failures without weakening runtime contracts, requiring optional dependencies, changing legacy defaults or rewriting frozen benchmark evidence. ## Linked Sources - Pull request: #9 - Failed matrix: https://github.com/thinkall/featcopilot/actions/runs/38614776818 - Review repairs: 08317c7 - Driver validation history: reviews/pr9/task-9-attempt-1-stage-attempt-1-review-1-gpt-6.1-sol.md - Existing probability tolerance: tests/test_classification.py - No ADO task was supplied; the GitHub PR carries the chore label. ## Rationale The shared plan validator already guarantees SDK message text; a local type annotation preserves that contract when mypy cannot import optional SDK types. Do not install the SDK on unsupported Python 3.10 or suppress type checking. Linux BLAS can differ in the final floating-point bits for batched prediction, so preserve exact transformed-input assertions and the existing tight learner tolerances instead of changing production arithmetic. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

thinkall · 2026-10-14T18:54:05Z

CI follow-up pushed in 8e44ad4. The first matrix found an optional-SDK typing boundary on Python 3.10 and overly strict bitwise probability assertions on Linux. The fix retains exact encoded-input equality, uses the existing tight probability tolerances, and adds no runtime fallback or mandatory SDK dependency. Missing/installed-SDK mypy, affected tests on both SDK versions, wheel build, and full pre-commit passed. Fresh matrix: https://github.com/thinkall/featcopilot/actions/runs/38618933772

Copilot

🔵 Needs a closer look

Broad experimental API and benchmark changes require final human review; project tests were not run during review.

0 open findings

🧠 Review effort: Balanced

## Summary Make the SDK-independent synchronous-bridge cancellation test assert asyncio.TimeoutError and its chained cause, which are distinct from the built-in TimeoutError on Python 3.10. ## Prompting Intent Finish the requested PR #9 review and resulting CI repairs while preserving the reviewed timeout, cancellation and exception-identity behavior. ## Linked Sources - Pull request: #9 - Matrix: https://github.com/thinkall/featcopilot/actions/runs/38618933772 - Failing Python 3.10 job: https://github.com/thinkall/featcopilot/actions/runs/38618933772/job/116041110938 - Driver history: reviews/pr9/task-9-attempt-1-stage-attempt-1-review-1-gpt-6.1-sol.md - No ADO task was supplied; the GitHub PR carries the chore label. ## Rationale asyncio.wait_for correctly propagates asyncio.TimeoutError. The test relied on its Python 3.11+ alias to the built-in class. Assert the originating API's type rather than normalize away the production exception identity or change deadline and cancellation guarantees. Full tests/types already pass on Python 3.11-3.13; this is a two-assertion test-only compatibility repair. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: fc6747ca-1d37-46b8-a1c4-fb650ac71116

Copilot

🔵 Needs a closer look

Broad experimental and integration changes require human approval; independent confirmation remains unresolved.

0 open findings

🧠 Review effort: Balanced

thinkall · 2026-10-14T19:41:46Z

Final verification is green on e8cc44d: Tests run 38622490682 passed all four Python lanes, including package mypy (52 files) in each lane. Pre-commit also passed.

Python	Passed	Optional skips	Coverage
3.10	3,963	276	92.32%
3.11	4,048	191	93.34%
3.12	4,048	191	93.34%
3.13	4,048	191	93.34%

The six-stage High review and CI follow-ups are complete. All 24 review artifacts and resolution histories are committed. Earlier capped-stage verdicts and deferred legacy/architectural concerns remain explicit; green CI is not a blanket clean review or a SOTA claim. The PR remains mergeable and awaits its required human approval; no merge was performed.

Copilot AI balanced review requested due to automatic review settings October 10, 2026 13:54

Copilot started reviewing on behalf of thinkall October 10, 2026 13:55

View session