Constraint Lineage as an Auditable Control Object for Conversational Text-to-SQL

Conversational parsers can carry a predicate from one SQL turn into a later one even when the predicate's origin no longer justifies its persistence. Local executability and edit traces reveal that a clause survived, but they do not encode the evidence under which it became state. We introduce LineageGate, an auditable state-commit interface that decomposes SQL into atomic hypotheses and records each hypothesis's origin, observed support, admission, endorsement, and reuse. The interface separates runtime selection features from an independent post-hoc entitlement judgment, making it possible to ask whether a retained constraint was licensed when it entered dialogue state rather than merely whether the current query is executable. We formalize the lifecycle of an inherited constraint and define three complementary measurement objects: Constraint Propagation Rate (CPR) for unsupported hypotheses that cross the commit boundary and reach first reuse, Downstream Error Amplification (DEA) for counterfactual attribution of a later failure, and Provenance Consistency (PC) for supported active state. The design composes a learned semantic selector with deterministic authorization and SQL-safety checks, while keeping their responsibilities explicit. This formulation turns conversational SQL continuity into a reviewable control object: an evaluator can inspect the admission history of each inherited clause, distinguish prediction from entitlement measurement, and trace a later reuse back to the evidence available at commitment time.