The evidence reducer solves concurrent writes, but operator.add also keeps earlier evidence when a rejected review returns to the planner. After a retry, the synthesizer may see results from two different plans without knowing which search produced each item.

I would attach a plan or attempt identifier to evidence and make the merge step's reuse policy explicit: keep prior evidence deliberately, or synthesize only from the current attempt. A fixture where the second plan changes the service being investigated would expose accidental carryover. Bounded retries then constrain execution while the evidence contract constrains what each attempt can conclude.