Problem
The regression-corpus validator does not normalize equivalent evidence across the accepted avro-value-golden-v1 and codec-regression-v1 formats. The adapters give the same Avro wire case different protocol-version identities and different semantic object shapes. As a result, an existing golden value, wire payload, and failure policy can be rewrapped as a new codec-regression fixture, receive a different semantic digest, increase the guarded codec count, and pass the official binding test without adding a new reproducer.
At commit 452b2fd, rewrapping the existing long_7 golden case as codec-regression-v1 is accepted as distinct evidence even though both entries carry the same schema, fingerprint, value, and wire bytes.
Acceptance criteria
- Normalize semantically equivalent codec evidence across every accepted fixture format before duplicate detection, including protocol identity, decoded wire bytes, value/type information, and failure policy.
- Reject a codec-regression fixture that only rewraps an existing Avro golden round-trip or rejection case. Such a fixture must not satisfy guarded corpus growth.
- Add adversarial policy coverage for cross-format duplicates and for equivalent base64 wire bytes.
- Preserve acceptance of genuinely new evidence and the existing immutability and explicit-supersession rules.
- Keep every accepted fixture executable through the official Python binding.
Delete when
Delete when equivalent evidence has one canonical semantic identity across accepted formats and cross-format rewrapping can no longer increase the corpus or satisfy a guarded implementation change.
Problem
The regression-corpus validator does not normalize equivalent evidence across the accepted avro-value-golden-v1 and codec-regression-v1 formats. The adapters give the same Avro wire case different protocol-version identities and different semantic object shapes. As a result, an existing golden value, wire payload, and failure policy can be rewrapped as a new codec-regression fixture, receive a different semantic digest, increase the guarded codec count, and pass the official binding test without adding a new reproducer.
At commit 452b2fd, rewrapping the existing long_7 golden case as codec-regression-v1 is accepted as distinct evidence even though both entries carry the same schema, fingerprint, value, and wire bytes.
Acceptance criteria
Delete when
Delete when equivalent evidence has one canonical semantic identity across accepted formats and cross-format rewrapping can no longer increase the corpus or satisfy a guarded implementation change.