Bisection by replay needs a compatibility check before its results can identify a culprit. Swapping a tool schema while keeping the old prompt and policy may create a configuration that was never valid, even if each artifact is correctly versioned. A failed replay would then show an incompatible combination rather than establish which change caused the production regression. Your interaction tests help, but recording which combinations are supported would make those failures easier to interpret. Would you make compatibility validation a prerequisite for admitting a configuration into the replay matrix?