Account for checked auxiliary inputs in PBT experiments

[CaelmBleidd]

When relation guidance finds an input, the experiment driver previously could omit it from class coverage or accept an incomplete confirmation. This change reuses the existing original/counter replay path for the original confirmed auxiliary input and requires the configured property, sources, selected relation, exact entry/local binding, native refutation event and admitted coverage identity to agree. Reduced inputs never replace the original witness.

Original named-check faults remain HYPOTHESIS candidates and are counted only when their canonical confirmation matches a timely checked auxiliary attempt. Missing, late, mismatched or incomplete confirmations remain explicit auxiliaryUncheckedInputs; summaries retain confirmation diagnostics, relation/original-query costs and dropped/omitted observation counts. The four ordinary controls keep their existing behavior.

Based on #607. Only the experiment driver and its tests change (2 files, +351/-6).

Validation:

• 15 Python unit tests pass, including the existing ordinary-control cases.
• Actual public receipts accept complete native coverage and the early canonical auxiliary named-check fault; the 5-second confirmation timeout is omitted with its diagnostic.
• 19 adversarial variants of the actual public receipt are rejected, including mismatched registration, source pins, selection, target/context, null binding/event, wrong input/check, absent coverage and a local target without compiler retention.
• git diff --check passes.

The later original/counter replay still measures native class coverage. These receipts verify the measurement contract; they do not establish a comparative scientific advantage. CI and integrated campaign results remain unverified at publication.