longitudinal: distinguish lag-pair-average and equal-unit irregular-rate estimands

[seonghobae]

Scientific finding

Longitudinal Modeling Draft #310 contains two distinct irregular residual log-rate weighting targets that must not be collapsed into one another.

The current landing authority is TEPP#310@ba10820e0d28cc33d1b91ef37f6f6d163b3d91e9, open/Draft/mergeable/unmerged on protected main@a243f18da4a4ca8a8d068c39922537f1f8ed6ad0.

center_within_unit_event_lags forms one lagged residual per admitted consecutive event-time pair. The pair-weighted recovery therefore lets a unit with m_i admitted occasions contribute up to m_i - 1 rates. That target is a lag-pair average, not automatically an equal-unit average.

The checked-in contract names the targets separately:

• tepp.irregular_rate.lag_pair_average.v1: every admitted consecutive lag pair has equal weight; evidence reports candidate/contributing units and candidate/admitted/refused pairs.
• tepp.irregular_rate.unit_average.v1: summarize admitted pair rates within unit first, then combine contributing units with equal unit weight unless a separately versioned design-weight contract says otherwise. This target is typed but deliberately fails closed until TEPP can consume an immutable released reusable finite-mean contract from fast-mlsirm.

This issue does not claim the pair-average estimator is numerically wrong. It owns estimand identity, denominator policy, and scientific acceptance. Occasion count, follow-up duration, missingness, pair refusal, or cross-classified/multiple-membership composition must not silently become an undeclared weighting rule.

Peer-reviewed informative-cluster-size work is used only as methodological evidence for that weighting-population distinction, not as treatment-effect authority for TEPP: Wang, Kong, and Datta (2011), Huang and Leroux (2011), and Kahan et al. (2023).

Current exact-head evidence

Rust Foundation 34672270380 is terminal on ba10820e...: 1,575 tests across 281 binaries; 1,574 PASS / 1 FAIL / 0 skipped. Formatting, Clippy, repository/Python contracts, workspace contract, Rust documentation contract, SBOM/provenance, and Live PostgreSQL are GREEN. No #495 estimand, denominator, unequal-follow-up, permutation, refusal-population, or 4,096-replicate informative-missingness Monte Carlo contract failed.

The sole Rust failure is the independent numerical-owner RED mixed_sign_mean_rounding_contract::half_ulp_tail_changes_the_final_mixed_sign_rounding: actual bits 5080060379673919488, correctly-rounded expected 5080060379673919487. Exact line/branch coverage generation stops when the instrumented suite reaches that same product RED, before authored-denominator enforcement; this is not a separately invented coverage deficit.

The current head also removes the redundant private stable_irregular_rate facade and the brittle source-string shadow-authority test. Crate-public composition now points directly to canonical irregular_residual; the stale occasion_mean.rs import was repaired. The reduced test/binary counts are structural cleanup, not removal of #495 scientific acceptance.

Documentation Quality 34672270392 and SAST Semgrep 34672270367 are GREEN. Security 34672270381 is RED because authoritative dependency-review support fails before the pinned Dependency Review step executes; OSV/Scorecard/Trivy evidence is not promoted as a semantic substitute. Required CodeQL 34672270345 is RED in the central producer-after-consumer settlement lane.

The live CodeQL owner sequence is .github#2106 followed by .github#2040. The second fresh sweep in this run observed .github#2106@48c6304ffb7294c4748ddc00388e72aa2e9d49bf, open/mergeable and Ready-for-review/Proposed after an ordinary non-force restack onto protected .github/main@fb17ef556f94f673234aa557254ae52779e9a7b0; its owner body records 3,073 passed / 1 skipped / 36 subtests and the credential-fallback response-isolation repair. Ready is review admission, not merge authorization, and #2106 remains an active mutable owner whose live head/state must be re-read before acceptance rather than pinned by TEPP. After the backward-compatible codeql-scan-v2 bootstrap lands while preserving v1, canonical producer/consumer successor .github#2040@85522306949bada2b5939608dc911f6374125f1b must non-force restack and switch to v2. .github#2051/#2056 remain evidence-preserving predecessors. These central-control REDs do not change the #495 estimand result.

Current review authority remains fail-closed. All current review threads except the finite-mean owner-boundary finding are resolved. The remaining unresolved thread on scaled_compensated_mean is valid: TEPP's local generic finite-binary64 mean must not become reusable numerical authority while the public mixed-sign half-ULP contract is RED. There is no qualifying independent current-head APPROVED review.

The branch carries the following scientific authority for this issue:

• IrregularRateEstimand::{LagPairAverageV1, UnitAverageV1} and denominator-bearing IrregularRateSummary;
• deterministic unequal-follow-up, follow-up-multiplicity, permutation, refusal-denominator, balanced and highly-unbalanced known-truth fixtures;
• a 4,096-replicate informative-missingness Monte Carlo contract with attempted/recovered/failed denominators, admitted-pair-conditioned truth, bias with Monte Carlo uncertainty, RMSE/stability, conditional 95% interval coverage with Wilson uncertainty, and exact replay;
• docs/product/prd-v0.4-amendment-longitudinal-time-ownership.md, docs/TRD.md, and docs/adr/longitudinal-modeling-ownership-addendum.md naming the same weighting identities, denominator populations, fail-closed unit-average activation, and the rule that occasion/follow-up multiplicity is not an implicit design or multiple-membership weight;
• claim-specific literature authority in docs/research/irregular-rate-estimand-weighting.md and docs/research/standards-and-literature.md, enforced by tests/quality/test_longitudinal_scientific_authority_sync.py.

Canonical docs/product-technical-gap-baseline.md and docs/TRACEABILITY.md remain owned by Draft #435. They must currentize against this exact #310 head before #495 can close; this source lane does not concurrently edit that owner branch.

Remaining acceptance

Before this issue can close, the landing/successor must preserve all valid existing evidence and satisfy the following on one protected exact head:

• canonical docs(ddd): restore queue authority and bounded-context ownership #435 product-gap/TRACEABILITY authority names the same estimands, maturity, denominators, current exact head, and numerical-owner boundary;
• event-time/leakage-safe admission and time-varying multilevel, cross-classified, and multiple-membership structure remain intact; occasion multiplicity is never substituted for a declared membership/design weight;
• UnitAverageV1 remains unavailable rather than approximated with another TEPP-local generic mean until a protected immutable fast-mlsirm numerical release exists;
• the independent mixed-sign half-ULP numerical RED is repaired through the canonical numerical owner, followed by a released-contract consumer bump, not by a third TEPP summation implementation;
• one unchanged TEPP exact head satisfies current Rust/rustdoc/owned line+branch+edge, security, central controls, independent review, scientific recovery, and release gates before normal protected integration;
• if feat(longitudinal): consolidate lagged correlation and discreteDRIFTstd #310 cannot land, a verified successor inherits every valid delta, test, fixture, contract, documentation change, and exact evidence before feat(longitudinal): consolidate lagged correlation and discreteDRIFTstd #310 or this issue is closed.

Ownership boundary

Reusable finite binary64 sum/mean arithmetic belongs to ContextualWisdomLab/fast-mlsirm and may be consumed only through its immutable released contract. Current numerical implementation vehicle #1816 (432765ccf633c9802e0f796ceeb4d6d572059acf) and GPU prerequisite #1717 (0b31640928e07f4362ce27dad3d310e630ab1b5d) remain mutable Drafts, not TEPP production authority.

Fresh source inspection of #1717 narrows the GPU prerequisite beyond a generic adapter-capacity check. gpu_marginal.rs constructs the E-step layout as one uniform plus 17 storage buffers and the score layout as one uniform plus 18 storage buffers. The controlled SwiftShader adapter exposes max_storage_buffers_per_shader_stage = 10; splitting the same COMPUTE-visible resources across additional bind groups cannot repair the defect because the limit applies per shader stage. An owner-correct topology repair must make the real E-step and score pipelines instantiate within that stage limit—e.g. by packing immutable sparse/index and numeric input families behind explicit offsets while keeping mutable outputs isolated, or an equivalent causal redesign—and then prove CPU-f64 parity/recovery on the real GPU path. A lowered probe constant, CPU fallback counted as GPU evidence, or skipped parity is not acceptance.

This issue does not authorize source copying, a mutable PR-head dependency, another TEPP generic summation implementation, CPU-as-GPU substitution, skip/xfail, coverage denominator tricks, or bypass of protected gates.

References

Huang, Y., & Leroux, B. (2011). Informative cluster sizes for subcluster-level covariates and weighted generalized estimating equations. Biometrics, 67(3), 843–851. https://doi.org/10.1111/j.1541-0420.2010.01542.x

Kahan, B. C., Li, F., Blette, B., Jairath, V., Copas, A., & Harhay, M. O. (2023). Informative cluster size in cluster-randomised trials: A case study from the TRIGGER trial. Clinical Trials, 20(6), 661–669. https://doi.org/10.1177/17407745231186094

Wang, M., Kong, M., & Datta, S. (2011). Inference for marginal linear models for clustered longitudinal data with potentially informative cluster sizes. Statistical Methods in Medical Research, 20(4), 347–367. https://doi.org/10.1177/0962280209347043

[seonghobae]

#495 implementation progress on the Longitudinal Modeling landing vehicle #310:

• c7f55acb347ceac38675d9566eff767372739dba adds typed/versioned IrregularRateEstimand::{LagPairAverageV1, UnitAverageV1} plus denominator-bearing IrregularRateSummary and recover_within_unit_irregular_rate_summary.
• 2f3f0a97a67d01e0a0c0da78d54ee395b49c0317 exports that contract from longitudinal_core; 0d0d4f80f12cd8ce24d949006e9bec9278c3c3fa keeps the stable irregular-rate facade free of a second numerical implementation.
• f45e83ba5223ad9d89022482e759ad9e705b229c adds a public unequal-pair-count fixture. The pair-weighted target is bit-identical to the legacy scalar result, differs from the independently computed equal-unit comparison, is input-row-permutation invariant, and retains denominators when zero rates are admitted.
• 493b45d62e2413bea0f56509b8a379c1b152c83d / a3065f6d7589d5a310e54f85d6447b27414955ba split refusals into zero/opposite-sign versus same-sign non-representable-rate denominators, so failure population is not silently lost.
• 28e5d766e849784ca7d97fe29c3f3cfcee725a6a currentizes the research authority and records the first-release candidate boundary.

The scientific choice is deliberately asymmetric: tepp.irregular_rate.lag_pair_average.v1 is the implemented first-release candidate; tepp.irregular_rate.unit_average.v1 has a stable typed identity but fails closed. Activating equal-unit aggregation needs a second finite-mean step, and TEPP will not add another generic mean while the canonical fast-mlsirm correctly-rounded finite-mean contract is unreleased.

Do not close #495 yet. Remaining acceptance is current-head hosted Rust/rustdoc/coverage evidence, canonical PRD/TRD/TRACEABILITY currentization, a collision-free superseding estimand ADR owned by the documentation lane, known-truth/unbalanced-follow-up Monte Carlo evidence for the declared target, and protected integration/release. The existing #310 mixed-sign binary64 RED remains independent and unresolved.

[seonghobae]

#495 current-head progress — 7373eef9ed129db09556ca104c4fa988b34a490a

Added the missing follow-up-multiplicity acceptance fixture in crates/longitudinal_core/tests/irregular_rate_estimand_contract.rs without activating UnitAverageV1 or adding numerical arithmetic.

The fixture keeps unit 1 unchanged and extends unit 2 from [3, 1, -4] with -4/3, +4/3. Unit 2 therefore remains centered at zero; its admissible log-rate target stays ln(1/3), while its admitted-pair count rises from 1 to 2. Acceptance now requires:

• pair-weighted admitted pairs: 3 → 4 and candidate pairs: 5 → 7;
• explicit refused-pair denominator: 2 → 3;
• the pair-weighted estimate moves toward unit 2 by a scientifically visible amount;
• the independently computed equal-unit comparator remains bit-identical because the added admissible unit-2 rate is again ln(1/3).

This pins the issue requirement that follow-up/occasion multiplicity may change the declared pair estimand but must not silently redefine an equal-unit target. The equal-unit computation remains an oracle only; production UnitAverageV1 still fails closed pending the immutable fast-mlsirm finite-mean owner release.

Predecessor b3751f8a... hosted Rust Foundation ran 1,574 tests with 1,573 PASS / 1 FAIL / 0 skipped; the only failure was the independent mixed-sign finite-mean half-ULP RED. Current-head CI has been triggered and is not promoted before terminal evidence.

Keep #495 open. Remaining acceptance still includes canonical PRD/TRD/TRACEABILITY + collision-free estimand ADR handoff, balanced/highly-unbalanced and informative-missingness known-truth evidence, worker determinism where applicable, protected integration, and release evidence.

[seonghobae]

#495 current-head evidence update — ec718201ac8e21e2ae4f33170f18be3bccd0df25

The follow-up-multiplicity RED has now been extended with deterministic known-truth fixtures for both balanced and highly unbalanced observation schedules, and the exact hosted Rust lane reached them.

• extra_admitted_pair_changes_pair_weight_without_changing_unit_target: GREEN. The added unit-2 occasions preserve that unit's ln(1/3) target while increasing admitted pair multiplicity, so the declared pair-weighted target moves and the equal-unit oracle does not.
• balanced_follow_up_recovers_declared_pair_weighted_truth: GREEN. Two units contribute two admitted rates each (ln(1/2) and ln(1/3)), recovering the expected midpoint.
• rate_associated_highly_unbalanced_follow_up_recovers_pair_weighted_truth: GREEN. Unit 1 contributes 15 admitted ln(1/2) rates and unit 2 contributes 2 admitted ln(1/3) rates; the implementation recovers the declared 15:2 pair target and remains >0.1 away from the equal-unit comparator.

Hosted Rust Foundation 34645410937 on this exact head: formatting GREEN, Clippy GREEN, 1,577 tests / 1,576 PASS / 1 FAIL / 0 skipped. The sole failure is the independent mixed-sign finite-mean half-ULP RED (5080060379673919488 actual vs 5080060379673919487 correctly rounded), not any #495 estimand test.

These deterministic fixtures are contract/known-truth unit-integration evidence only. They do not close the remaining realistic scientific acceptance work: informative-missingness generation, repeated Monte Carlo recovery with declared failure denominator, RMSE/bias/coverage/convergence/reproducibility and Monte Carlo uncertainty, plus canonical PRD/TRD/TRACEABILITY and collision-free estimand ADR handoff. Keep #495 open.

UnitAverageV1 remains typed but fail-closed; no equal-unit production arithmetic has been added while the immutable fast-mlsirm finite-mean contract remains unreleased.

[seonghobae]

Current-head RCA and repair on the Longitudinal Modeling vehicle #310:

• e6d1d637d3d271a1e3703a7d793bb7fdc9c6b40c made the older deterministic retention helper Clippy-clean, but hosted Rust Foundation 34650015208 exposed a scientific-test defect that predecessor evidence had not reached: 1,579 tests ran, 1,577 passed, 2 failed, 0 skipped. Besides the independent mixed-sign finite-mean RED, informative_missingness_recovery_is_reproducible_and_pair_weighted failed because only 243/256 represented-data replicates fell inside its fixed 1e-12 pseudo-coverage band.
• That fixture was not the canonical realistic Monte Carlo acceptance. It mixed a noiseless generated-rate calibration with finite centering/representation error and called |error| <= 1e-12 “coverage”, while the already-present irregular_rate_monte_carlo_contract.rs is the stronger successor: 4,096 noisy informative-missingness replicates, truth conditioned on admitted pair counts, attempted/recovered/failed denominator, bias + Monte Carlo SE, RMSE + first-quarter stabilization, empirical conditional 95% coverage + Wilson interval, and exact replay. That canonical test was not one of the two failing tests on e6d1d637....
• b72f1f1e9c6cd24e273104a8c31d0e68f081faa6 therefore removes only the superseded duplicate Monte Carlo helper/test from irregular_rate_estimand_contract.rs and makes that file explicitly deterministic-contract-only. It does not change production arithmetic, estimand identity, refusal rules, tolerance in the canonical Monte Carlo fixture, or the mixed-sign RED. No skip/xfail or coverage-denominator suppression was introduced.

This is a successor-evidence consolidation, not a scientific weakening: the required informative-missingness bias/RMSE/coverage/reproducibility/Monte Carlo-uncertainty evidence remains in irregular_rate_monte_carlo_contract.rs; deterministic unequal-pair, multiplicity, balanced/unbalanced, permutation, refusal-denominator and fail-closed UnitAverage fixtures remain in irregular_rate_estimand_contract.rs.

Keep #495 open. Exact-head CI for b72f1f1e... is running; do not promote this repair until the canonical 4,096-replicate Monte Carlo test is observed GREEN on that exact head. Canonical PRD/TRD/TRACEABILITY + collision-free estimand ADR, numerical-owner release, protected integration and release evidence also remain outstanding.

[seonghobae]

Exact-head verification for b72f1f1e9c6cd24e273104a8c31d0e68f081faa6 has now reached the no-retry Rust result. cargo nextest run --workspace --all-features ran 1,578 tests across 282 binaries: 1,577 PASS / 1 FAIL / 0 skipped. The only remaining failure is the independently owned mixed_sign_mean_rounding_contract::half_ulp_tail_changes_the_final_mixed_sign_rounding RED (left bits 5080060379673919488, expected 5080060379673919487). The superseded 256-replicate pseudo-coverage failure is gone without changing a tolerance, skipping a test, or changing production arithmetic.

This also means the unchanged canonical irregular_rate_monte_carlo_contract.rs acceptance executed within the exact workspace suite without becoming a failure: the 4,096-replicate informative-missingness contract remains the active realistic evidence for attempted/recovered/failed denominators, conditional truth under admitted pair counts, bias + Monte Carlo SE, RMSE stabilization, conditional 95% coverage + Wilson interval, and exact replay.

Exact-head formatting and Clippy are GREEN; repository/Python contracts, Rust documentation contract, SBOM/provenance, and Live PostgreSQL are GREEN. Production coverage generation is still downstream RED because the unresolved mixed-sign product test aborts the instrumented suite before denominator enforcement. Keep #495 open: PRD/TRD/TRACEABILITY + collision-free estimand ADR, released numerical-owner dependency, protected integration, and release evidence remain outstanding.

[seonghobae]

Current landing authority is now TEPP#310@bf393bc4391a671466438efa53ae092e7c8aabab (open/Draft/mergeable/unmerged). This run closed the #495 PRD/TRD/ADR docs-to-code gap by ordinary forward commits without changing production arithmetic or Monte Carlo thresholds.

RED→repair lineage:

• 6a5acddf5a88ec89cbd422259c7aa1f05010786f: quality contract first required PRD/ADR to name tepp.irregular_rate.lag_pair_average.v1 and tepp.irregular_rate.unit_average.v1, denominator/refusal populations, fail-closed unit-average activation, and the rule that occasion count is not a membership weight.
• a33c0e8e7272add3bd5b6ba5f58e82da88febe5d: PRD amendment now names both estimands, their weighting populations and future design-weighted separation; status corrected from premature Approved to Proposed while the landing vehicle is Draft.
• 8f3bd98eec8aa4dc668f47dc4d67e97f5f22d537: ADR 0005 addendum records pair-vs-unit weighting as scientific identity, keeps time-varying multilevel/cross-classified/multiple-membership weights distinct from pair count, and is now Proposed rather than prematurely Accepted.
• a9420c1e80933f879b9d551dd5ae5a70c9b56150: quality contract extended to require matching TRD traceability.
• bf393bc4391a671466438efa53ae092e7c8aabab: TRD now requires explicit versioned weighting, unequal-follow-up/permutation/refusal-denominator evidence and informative-missingness known-truth Monte Carlo with attempted/recovered/failed denominators, bias/RMSE/coverage and Monte Carlo uncertainty.

Compare b72f1f1e...→bf393bc... is five ordinary-forward commits touching only PRD, TRD, ADR and the documentation quality test. The existing LagPairAverageV1 production path, typed fail-closed UnitAverageV1, denominator-bearing IrregularRateSummary, and 4,096-replicate Monte Carlo contract are unchanged.

Issue stays open. Canonical docs/product-technical-gap-baseline.md / docs/TRACEABILITY.md integration remains owned by #435, and UnitAverageV1 must remain unavailable until a protected immutable fast-mlsirm finite-mean release is actually consumable. The independent TEPP mixed-sign half-ULP numerical RED is not part of this estimand-policy repair.

[seonghobae]

Follow-up currentization: landing authority advanced by ordinary-forward research doctoring to TEPP#310@3b1c801d5fb3f1d0ea92577c696d956b9d4e7f73.

The new RED→GREEN pair is narrow and claim-specific:

• 29e44e15ed074209fc383e114aa2006210334de2 adds a quality contract requiring Wang, Kong & Datta (2011), Huang (2011), Kahan et al. (2023), and the record-multiplicity/irregular-rate estimand boundary in the canonical APA register.
• 3b1c801d5fb3f1d0ea92577c696d956b9d4e7f73 registers those sources in docs/research/standards-and-literature.md, explicitly as methodological analogies for weighting-population identity rather than imported treatment-effect estimands.

Verified compare bf393bc...→3b1c801... is exactly 2 commits / 2 files: the quality test (+12) and canonical literature register (+8), with no deletion and no production source mutation. #495 remains open for canonical #435 baseline/TRACEABILITY integration and immutable numerical-owner release.

[seonghobae]

Current landing authority advanced to TEPP#310@4f7f6ba7545e57d36c4660badf572780e96e0fed.

Research doctoring is now complete for this lane: both the claim-specific note and canonical APA register carry Huang & Leroux (2011) and Kahan et al. (2023) with complete journal metadata, and the quality contract checks those full records. During the full-file register edit I detected two unrelated citation drifts (Liu author initial, then Brown NeurIPS URL); both were repaired immediately by ordinary-forward commits rather than hidden/rebased. Net 3b1c801...→4f7f6ba... is 5 commits / 3 files and changes only the irregular-rate research note, canonical literature records, and their quality contract. Production source, estimand behavior, refusal policy, and Monte Carlo thresholds remain untouched.

The issue body has also been currentized to the corrected references and present acceptance boundary. Keep #495 open: canonical docs/product-technical-gap-baseline.md / docs/TRACEABILITY.md integration still belongs to #435, UnitAverageV1 remains fail-closed until an immutable fast-mlsirm finite-mean release exists, and #310's independent mixed-sign half-ULP RED is unresolved.

[seonghobae]

Exact-head verification update — TEPP#310@4f7f6ba7545e57d36c4660badf572780e96e0fed.

The #495 documentation/authority delta is now being exercised on the current head rather than relying on predecessor evidence. Documentation Quality 34657771609 is GREEN. Rust Foundation 34657771583 is terminal RED: formatting and Clippy are GREEN; the no-retry workspace test step is RED; exact line and branch coverage generation are also RED before their enforcement/diagnostic artifact steps. In the same run, repository/Python tooling coverage, workspace contract, Rust documentation contract, SBOM/provenance, and Live PostgreSQL are GREEN.

I am not promoting the predecessor b72f1f1e... test count onto this head. The verified b72f1f1e...→4f7f6ba... delta is 12 ordinary-forward commits limited to PRD/TRD/Longitudinal ADR/research/literature/quality-contract files, with no production arithmetic, estimand, refusal-policy, or Monte Carlo-threshold mutation; nevertheless the current no-retry failure remains a current-head RED until its exact failure evidence is available. The independent mixed-sign half-ULP numerical RED therefore remains unresolved rather than being relabeled as #495 acceptance.

Security Scan, Semgrep, and delegated CodeQL were still queued at the latest current-head read, so none is promoted. Canonical baseline/TRACEABILITY integration remains #435-owned; the 4f7f6ba... handoff is already recorded there. Keep #495 open for actual #435 integration, immutable fast-mlsirm finite-mean release, a qualifying current-head approval, protected integration, and release evidence.

[seonghobae]

Fresh exact-head authority for TEPP#310@4f7f6ba7545e57d36c4660badf572780e96e0fed:

• Rust Foundation run 34657771583, job 103453819031: cargo nextest run --workspace --all-features ran 1,578 tests across 282 binaries: 1,577 PASS / 1 FAIL / 0 skipped. The sole failure is the independent numerical-owner RED longitudinal_core::mixed_sign_mean_rounding_contract::half_ulp_tail_changes_the_final_mixed_sign_rounding, at mixed_sign_mean_rounding_contract.rs:133: actual bits 5080060379673919488, correctly-rounded expected bits 5080060379673919487. Therefore the longitudinal: distinguish lag-pair-average and equal-unit irregular-rate estimands #495 estimand/denominator/4,096-replicate Monte Carlo contracts are current-head GREEN; do not fold the remaining mixed-sign binary64 mean failure into longitudinal: distinguish lag-pair-average and equal-unit irregular-rate estimands #495 or widen its scientific acceptance.
• Formatting and Clippy are GREEN. Production line/branch coverage generation stops because that product RED is exercised under instrumentation; denominator enforcement/artifact publication therefore do not become independent coverage acceptance.
• Security run 34657771596: OSV, Trivy and Scorecard are GREEN, but authoritative dependency-review job 103454465919 fails at Check dependency review support; the pinned Dependency Review step is skipped. This remains .github#810 owner evidence, not a TEPP-local bypass opportunity.
• Required CodeQL run 34657771589 reproduces the central settlement defect: Python consumer 103455852569 terminally failed at 23:37:38Z after reading the current-head verdict; Actions consumer 103455852575 did the same at 23:40:44Z; producer dispatch 103457274496 only started at 23:40:58Z and then succeeded. This remains .github#2051 owner evidence; no fixed sleep/no-op retrigger/synthetic receipt is acceptable.
• All currently returned feat(longitudinal): consolidate lagged correlation and discreteDRIFTstd #310 inline review threads are resolved. Formal reviews are COMMENTED only; there is no qualifying current-head independent APPROVED review.

Keep #495 open. Its scientific implementation/acceptance is no longer the Rust failure frontier; remaining closure prerequisites are canonical #435 baseline/TRACEABILITY currentization, immutable finite-mean owner release before UnitAverageV1 activation, protected integration/release evidence, and the repository approval/control gates.

[seonghobae]

Fresh exact-head closure check for the scientific lane: TEPP#310@8eb95d14f43ac40b06d0d039ef4d94ad1a053924 Rust Foundation 34669350804 is terminal and reproduces the predecessor result exactly. Nextest ran 1,578 tests / 282 binaries: 1,577 PASS, 1 FAIL, 0 skipped. The only failure is mixed_sign_mean_rounding_contract::half_ulp_tail_changes_the_final_mixed_sign_rounding (actual 5080060379673919488, expected 5080060379673919487). No #495 estimand, denominator, unequal-follow-up, permutation, refusal-population or 4,096-replicate informative-missingness Monte Carlo contract failed. Therefore #495's TEPP-owned scientific contract is current-head GREEN; the remaining Rust RED is independently owned finite-binary64 arithmetic. Issue stays open because #435 canonical baseline/TRACEABILITY currentization and the immutable fast-mlsirm release/consumer-bump sequence are still required.

[seonghobae]

Current landing authority advanced ordinary-forward to TEPP#310@ba10820e0d28cc33d1b91ef37f6f6d163b3d91e9; #495's estimand identity, denominator/refusal policy, unequal-follow-up fixtures, and 4,096-replicate informative-missingness acceptance are unchanged by this review-repair sequence. The new source change removes a redundant private irregular-rate facade and exports the canonical Longitudinal implementation directly; it does not change LagPairAverageV1, the fail-closed UnitAverageV1, production numerical formulas, or Monte Carlo thresholds. Fresh exact-head hosted evidence is still being acquired: Documentation Quality and repository/Python/workspace/rustdoc/SBOM contracts are GREEN and Rust formatting is GREEN; Clippy/tests/coverage/live PostgreSQL and central/security gates remain in progress. The generic finite-binary64 mean finding is now independently reaffirmed by CodeRabbit and remains an owner-boundary RED handed to fast-mlsirm#1816, not a reason to create another TEPP summation kernel. #495 stays open until #435 currentization and immutable numerical-owner release/consumer replacement complete.