feat: add Panel Tier-1 Stage B diagnostics

[TheHiddenObserver]

Summary

Tracks #93 and implements Stage B / PR-B of the Panel Tier-1 roadmap on top of merged Stage A (#119).

Stage-B scope includes Hausman FE-vs-RE, pooling F, Breusch-Pagan LM, parameter-based panel R², adjusted R²/model F where defined, structured diagnostics, NumPy/CuPy/Torch parity, external-definition checks, frontend evidence, and physical GPU acceptance.

Final lifecycle status

COMPLETE / ready for review.

Current PR head: cdff57179d97cd53863681cfa46301da919999ec.

The complete review/fix loop was rerun under .claude/skills/code-review.md using changed+adjacent review and auto-fix semantics. Fresh final review on this exact head found:

• CRITICAL: 0
• HIGH: 0
• relevant MEDIUM: 0
• unresolved review threads: 0

Fresh final review ID: 4891642296.

Applicable-Hausman review fix

The final P2 required the physical GPU gate to execute at least one genuinely applicable Hausman statistic/p-value/df path rather than counting only matching structured-inapplicable results.

The runner now includes a deterministic fitted FE/RE fixture with seed 20260810, 12 entities x 4 observations, one slope, nonzero entity-effect scale 0.005, noise scale 0.1, and 48 observations. Every requested GPU backend must produce at least one successful applicable=true Hausman result with finite statistic/p-value and positive df. The four existing non-PSD parameterizations remain, so both applicability branches are validated.

Accepted physical P100 evidence

Exact clean measured implementation:

2701aa9feb3796c33c94e6480fcb78c80c6a809c

Raw artifact:

results/pr122_p100/panel_stage_b_gpu_validation_2701aa9f.json

Provenance:

• artifact commit: 0d0d654d825cea872672f27d02107a58048b345f
• raw Git blob: fa3a253e6d882a4e69be29e7e3b1dce7b223b9a9
• clean working tree: true
• Tesla P100-SXM2-16GB
• Python 3.9.16
• NumPy 1.24.2
• SciPy 1.10.1
• Torch 2.0.0
• timing collected: false

Both physical GPU backends passed:

• CuPy: 17/17 estimator cases, 5/5 Hausman diagnostics, executed_backend=cupy, no CPU fallback;
• Torch: 17/17 estimator cases, 5/5 Hausman diagnostics, executed_backend=torch, no CPU fallback.

The original four Hausman parameterizations remain successful structured applicable=false / non-PSD cases.

hausman_applicable_nonzero_effect is physically applicable=true on both backends with df=1:

• CuPy statistic 1.1965942530851057, p-value 0.27400344142676447;
• Torch statistic 1.1965942530849238, p-value 0.2740034414268009;
• NumPy reference statistic 1.196594253085033, p-value 0.274003441426779.

Statistic differences versus NumPy are 7.26e-14 (CuPy) and 1.09e-13 (Torch); FE/RE coefficient and diagnostic-covariance differences are below 1e-12.

The historical focused disconnected two-way FE artifact remains supplementary evidence for the unchanged df=1 path and is explicitly identified with its original measurement SHA a57efcea29b0e87ecb89865c5a6902d5773812c6; it is not relabeled as a 2701aa9 measurement.

Canonical frontend evidence

Promoted source:

results/benchmark_frontend_sources/panel_stage_b_pr122_p100_20260809_2701aa9f.json

• SHA-256: 2056f836bfe2a708b3131becca42dbba15e519762c8bafb582000b19f81120bf
• source id: panel-stage-b-pr122-20260809-2056f836bfe2
• validation-only rows: 44 = 17 estimator cases x 2 backends + 5 Hausman diagnostics x 2 backends
• no timing or speedup claims
• applicable Hausman statistic/pvalue/df are preserved; missing/invalid numeric evidence fails closed
• previous 42-row a57efcea... source is retained as explicitly superseded historical audit evidence
• frontend/public and docs/assets generated JSON mirrors are byte-identical.

Compare 2701aa9f...cdff5717 contains no statgpu/ numerical source change and no physical-runner change. Post-measurement changes are raw/canonical evidence, contracts, generated assets, and documentation only, so the P100 evidence remains applicable to the final candidate.

Exact-final hosted validation

For current head cdff57179d97cd53863681cfa46301da919999ec:

• Tests #1657: SUCCESS — full CPU suite, static/docs contracts, Python 3.9/3.10/3.11/3.12 regression matrix, maintained Torch 2.0 CPU regressions, and executable linearmodels==7.0 Stage-B alignment;
• Maintenance compatibility #670: SUCCESS;
• Release notes validation #601: SUCCESS;
• Release package validation #607: SUCCESS, including Ubuntu/Windows/macOS wheel smoke;
• Benchmark Frontend #674: SUCCESS — Python 3.9/3.11 parser/generator contracts, deterministic staleness, frontend typecheck/build, Chromium e2e, and Chromium/Firefox/WebKit production QA.

Review conclusion

The final applicable-Hausman P2 has been answered with exact physical evidence and resolved. All review threads are resolved.

Under .claude/skills/code-review.md, the final review/fix hard exit is COMPLETE. PR #122 is ready for review/merge from correctness, inference, three-backend behavior, physical acceptance, provenance, artifacts, documentation, and hosted-CI perspectives.

Issue #93 remains open for Stage C covariance expansion/completion.

Merge is intentionally not performed as part of this lifecycle; merging remains a separate explicit action.

[TheHiddenObserver]

Fresh post-fix local review checkpoint on exact head 2701aa9feb3796c33c94e6480fcb78c80c6a809c, following .claude/skills/code-review.md changed+adjacent and auto-fix standards. This review does not inherit the earlier merge-ready conclusion.

Review/fix loop findings closed locally:

• the physical runner now contains a deterministic fitted FE/RE Hausman fixture with a nonzero entity effect that is stably applicable on the NumPy reference, and each requested CUDA backend must produce at least one successful applicable result with finite statistic/p-value/positive df;
• the dedicated fixture additionally checks FE/RE coefficient and diagnostic-covariance parity plus Hausman statistic/p-value/df parity against NumPy;
• existing structured-inapplicable Hausman fixtures are retained, so the physical matrix exercises both applicability branches;
• frontend canonicalization distinguishes the dedicated 48x1 fixture from ordinary balanced/unbalanced Hausman rows and prevents method/case identity collisions;
• future canonical applicable-Hausman evidence must preserve valid statistic/pvalue/df or the parser fails closed;
• the new parser contract is explicitly exercised under the Benchmark Frontend Python 3.9/3.11 matrix;
• physical acceptance/changelog/release-process documentation now treats the older P100/42-row evidence as historical after the validator changed.

Exact-head hosted gates are fully green:

• Tests #1645: SUCCESS (full CPU, static/docs, Python 3.9-3.12, maintained Torch 2.0 CPU, linearmodels==7.0 alignment);
• Maintenance compatibility #658: SUCCESS;
• Release notes validation #589: SUCCESS;
• Release package validation #595: SUCCESS, including Ubuntu/Windows/macOS wheel smoke;
• Benchmark Frontend #662: SUCCESS, including Python 3.9/3.11 parser/generator contracts, deterministic staleness, typecheck/build, Chromium e2e, and Chromium/Firefox/WebKit production QA.

Final tree audit from the previous merge-ready head contains only nine expected runner/test/parser/CI/review/changelog/release-process files; temporary review-fix carrier workflows are absent. Review-thread audit leaves exactly one unresolved thread: the applicable-Hausman physical-coverage P2.

Current review result:

• local CRITICAL: 0
• local HIGH: 0
• locally actionable in-scope MEDIUM: 0
• remote-only blocking finding: 1

The remaining blocker is a fresh exact-clean-head P100 CuPy/Torch execution of the changed dev/benchmarks/validate_panel_stage_b_gpu.py on this SHA. The P2 thread should remain unresolved and the PR should remain Draft until that artifact is audited, canonicalized (44 validation rows), frontend evidence is regenerated, exact-final-head hosted CI is rerun, and a fresh review finds no new issue.

Hard exit: PARTIAL_REMOTE_PENDING. No merge action is taken by this review.

[TheHiddenObserver]

Fresh independent final review on exact head cdff57179d97cd53863681cfa46301da919999ec under .claude/skills/code-review.md (auto-fix semantics). This review does not inherit any earlier READY conclusion.

Impact classification / active gates:

• three-backend behavior: active;
• inference/correctness: active;
• benchmark/physical provenance: active;
• tests/artifacts/docs: active;
• formula compatibility: verified through existing Stage-B matrix/external alignment where touched;
• loss/penalty/solver/CV: inactive because this review-fix delta does not alter those components.

Fresh changed+adjacent review result:

• CRITICAL: 0
• HIGH: 0
• relevant MEDIUM: 0
• unresolved review threads: 0 after closing the applicable-Hausman physical-coverage P2.

Physical correctness/backend gate:

• exact clean measured implementation: 2701aa9feb3796c33c94e6480fcb78c80c6a809c;
• raw artifact: results/pr122_p100/panel_stage_b_gpu_validation_2701aa9f.json;
• artifact commit: 0d0d654d825cea872672f27d02107a58048b345f;
• raw Git blob: fa3a253e6d882a4e69be29e7e3b1dce7b223b9a9;
• CuPy and Torch: 17/17 estimator cases and 5/5 Hausman diagnostics each, requested/executed backend identity, no CPU fallback;
• the four historical Hausman parameterizations remain correctly structured inapplicable/non-PSD;
• hausman_applicable_nonzero_effect physically executes the applicable statistic/p-value/df path on both backends with df=1 and statistic differences versus NumPy of 7.26e-14 (CuPy) and 1.09e-13 (Torch); FE/RE coefficient/covariance differences are below 1e-12.

Evidence/provenance review:

• promoted canonical source: results/benchmark_frontend_sources/panel_stage_b_pr122_p100_20260809_2701aa9f.json;
• SHA-256: 2056f836bfe2a708b3131becca42dbba15e519762c8bafb582000b19f81120bf;
• source id: panel-stage-b-pr122-20260809-2056f836bfe2;
• 44 validation-only rows = 17 estimator cases x2 + 5 Hausman diagnostics x2;
• statistic/pvalue/df for the applicable branch are preserved and fail closed when absent/invalid;
• no timing or speedup is claimed;
• previous 42-row source is explicitly superseded historical evidence;
• frontend and docs generated bundles are byte-identical by Git blob;
• compare 2701aa9f...cdff5717 contains no statgpu/ numerical source and no physical-runner change, so the P100 artifact remains applicable to the final candidate.

Exact-final hosted gates on cdff5717...:

• Tests #1657: SUCCESS — full CPU suite, static/docs contracts, Python 3.9/3.10/3.11/3.12 regression matrix, maintained Torch 2.0 CPU regressions, and linearmodels==7.0 Stage-B alignment;
• Maintenance compatibility #670: SUCCESS;
• Release notes validation #601: SUCCESS;
• Release package validation #607: SUCCESS, including Ubuntu/Windows/macOS wheel smoke;
• Benchmark Frontend #674: SUCCESS — Python 3.9/3.11 generator/parser gates, deterministic staleness, frontend build/typecheck, Chromium e2e, and Chromium/Firefox/WebKit production QA.

Review/fix loop hard exit: COMPLETE. PR #122 is ready for review/merge from correctness, inference, backend parity, physical acceptance, provenance, artifact, documentation, and hosted-CI perspectives. This review does not merge the PR.