scientific(analysis): preserve evidence identity in longitudinal CWC composition #592

Scientific / DDD finding

The Draft longitudinal_cwc_v1 adapter on PR #372 dropped the opaque evidence identity that the Analysis Run boundary uses to prevent one cutoff-visible source observation from being replayed as extra scientific weight. LongitudinalClusterScore carried only snapshot_id, cluster_key, predictor/outcome coordinates, and availability time.

That omission allows one observation already available at the historical cutoff to be replayed more than once. Because the numerical owner stacks within-cluster residual rows and builds cluster means from admitted rows, replay can change row weighting, cluster means, within/between slopes, contextual effect, and row_count. This is pseudo-replication.

Leakage-safe identity population

The surviving Analysis Run vehicle #416 establishes the current historical-replay rule: availability admission precedes duplicate/domain admission. A row with AvailableTime > knowledge_cutoff was not part of the historical evidence population and must not perturb an earlier replay even if its identity duplicates a visible row. Duplicate identity is therefore rejected among cutoff-visible evidence, not across future-unavailable rows.

This corrects the first #592 test formulation, which conservatively rejected even a future-unavailable duplicate. That formulation conflicted with #416's leakage-safe census semantics and has been repaired ordinary-forward rather than hidden by history rewrite.

Required / implemented repair

Keep numerical CWC arithmetic in psychometric_core; repair only the Analysis Run adapter boundary.

• carry one bounded opaque evidence_id on every LongitudinalClusterScore;
• validate it with the existing analysis-engine identifier contract;
• exclude future-unavailable rows before evidence-identity admission;
• reject a repeated identity among cutoff-visible evidence with AnalysisEngineError::DuplicateEvidence before scientific composition;
• do not infer identity from predictor/outcome/cluster/time tuples: numerically identical but distinct evidence remains admissible;
• retain immutable snapshot_id, availability provenance, cross-snapshot refusal, cutoff binding, and causal-refusal contracts;
• preserve the raw MAX_EVIDENCE_UNITS operational bound before filtering;
• maintain public regression coverage for visible duplicate refusal, future-unavailable duplicate historical invariance, distinct-identity equal-value positive control, accessor/constructor validation;
• currentize feat(analysis): bind CWC within/between slopes to an analysis-run profile #372 / scientific: establish longitudinal CWC recovery acceptance evidence #501 scientific acceptance authority after repair; this integrity fix does not constitute full scientific: establish longitudinal CWC recovery acceptance evidence #501 Monte Carlo acceptance.

Ordinary-forward lineage

• RED e66a90f3c6be7c04ecc9a310baf955b16dbd6f76: introduces explicit evidence identity and visible duplicate/equal-value contracts; it also initially over-constrained future-unavailable duplicates.
• Repair 138be1fb8ad3ae47a18d7e05645d2ca2830cd341: adds evidence_id and duplicate detection, initially before cutoff admission.
• Fixture migration 66cffe0c88f2c8e54881f6018a776c62daaf788c → 418222f232dfc1eb8d6fa27d841e2302148ccf55.
• Doctoring 104e2b1620ee0b7d42d1990d1fa14caaf5da46af records the first repair state.
• Supplemental RED bb8cb21ac40fc9f1a2d86612e8cbaaac5ccecb67: proves a future-unavailable duplicate must not change the historical artifact or terminal result.
• Causal repair 6b07a4eafa3c009e36299aeca6e87a5b45201d56: moves identity admission inside the cutoff-visible branch while retaining cross-snapshot validation and raw cardinality bounds.

Ownership / acceptance

This is TEPP Analysis Run evidence/provenance composition, not a new static psychometric estimator. No source copy from fast-mlsirm, no LLM numerical authority, no skip/xfail, and no tuple-based heuristic deduplication.

A successor may close this issue only after the valid identity + cutoff-order delta, tests, provenance contract, and exact-head evidence are preserved on the surviving Analysis Run vehicle.