Equivalence harness: regression (drift) mode for in-place COBOL changes #4513

Why. The same generated golden-master scenarios that verify a translation can guard everyday changes to the original COBOL: a CICS developer changes a program and wants proof that only the intended outputs changed.

What.

1. Generate scenarios for a program and commit the original's outputs as an approved baseline.
2. On every change, re-run them. Any changed output is flagged, with the exact field and scenario.
3. A person approves intended differences, scoped (e.g. 'only LATE-FEE, only for overdue accounts'). Build this on Declared differences: equivalence except where a person approved a difference #4051 (declared differences), not a second mechanism.
4. Anything not approved fails the check.

Note: this checks that behavior is UNCHANGED, not that it is correct. Existing bugs are baselined too. For z/OS-native shops it needs the original to run there (#4050).

Done when: a demo PR that changes one COBOL program shows the approved difference passing and an unintended side effect caught.

Related: #3804, #4051, #4050, #4493.