TraceCone: Exact Dependency Reuse for Localized Repository Repair

Repository-repair agents create diagnostic findings, patch fragments, and test outcomes whose validity must survive later failures. TraceCone makes this validity question an explicit execution operation. Each generated record carries the complete identity of every executable input it consumed; a failed contract queries those identities in two directions, recovering verified diagnostic context from incoming dependencies while invalidating the outgoing descendant closure. This gives learned executable state a precise preservation rule: records outside the closure remain reusable, while affected generated records are regenerated and affected checks are re-executed. We formalize the condition under which this operation is exact, distinguish it from content retrieval and learned cone prediction, and specify an evaluation contract that fixes units, endpoints, pairing, and budget accounting before any empirical aggregate is reported. The resulting contribution is a validity-maintenance interface for repository repair: persistent records become reusable state only when a system can identify the consumed inputs that justify preserving them.