What Does a Repair Statistic Measure? A Claim-Indexed Audit of Repository-Level LLM Repair

Repository-level repair systems emit many logged events before a verifier pass. A rate over one event is easy to misread when the report omits the event's operational meaning, opportunity budget, support unit, or verifier boundary. We introduce a claim-indexed measurement model that binds each repair statistic to five coordinates: construct, opportunity design, unit of support, endpoint interpretation, and instrumentation provenance. We instantiate the model in a retrospective audit of 2,400 executions: two 30-task DeepSeek strata with disjoint task IDs and a 30-task GLM boundary stratum that reuses the clean task surface. A field-integrity audit identifies 12 raw GLM action positives that conflict with the patch-application evidence required by the governed-action construct. The harmonized table contains no verifier pass outside the governed-action state, while conditional endpoint yield is 14/199 (7.04%) and 14/204 (6.86%) in the two DeepSeek strata. Under eight retained opportunities per fixed cell, governed-action discovery reaches 40.83% and 46.67%; verifier-pass discovery reaches 4.17% and 1.67%. The 28 passing rows reduce to seven positive cells, three tasks, and two task families. Endpoint controls produce the expected result for all six gold runs, six no-op runs, and nine reference-derived mutants. These findings show that repair evidence cannot be reduced to one success scalar. Field semantics, opportunity, support, verifier scope, and instrumentation determine what a reported number measures and which comparisons it can sustain.