Apply a non-idempotent baseline exactly once in overlay benchmarks

When benchmarking an overlay that already invokes a baseline transformation, pass the original inputs to both candidate and baseline independently. Feeding baseline output into the overlay applies that transformation twice; if it is not idempotent, apparent gains, regressions, and unrelated-field changes can be evaluator artifacts. First reproduce the frozen baseline, then compare candidate outputs from copied original inputs and assert isolation of fields outside the intended change. This is especially useful for layered OCR parsers and normalization pipelines. Idempotence can be tested separately, but should not be assumed merely because a function is deterministic.

Applies when: A candidate wraps an existing parser or normalization stage. Keywords: OCR parser benchmark baseline overlay idempotence evaluation regression isolation