[DylanCosto]

A robustness experiment could previously finish earlier scores before discovering that a later pack did not match the genotype extraction. All pack/source compatibility checks now run before scoring output is created. Preparation also fingerprints inputs before opening them and verifies those fingerprints afterward.

robustness --preflight uses the shared preparation checks and writes a separate input-check report with target/model/scenario counts, structural row bounds, selected reference members, shared target IDs, model inventories and provenance. It never emits score tables or an experiment completion event. Genotype/frequency data blocks remain lazy, so the report explicitly does not claim scoring success, common-callable coverage, runtime or memory estimates. Truncated and oversized frequency headers now return errors instead of panicking.

Validation: 187 Rust tests, the Rust 1.88 preflight CLI regression, formatting and Clippy pass. Tests cover a later incompatible pack, unavailable sample IDs, malformed frequency headers, output preservation, count overflow and the deliberate lazy-block limit. Preflight matches the completed 128-person study's fingerprints and 5,376-row inventories; the 1,745,179-term eight-person pilot preserves prior scores, attribution and reference results. The archive verifies 58 payload hashes and 54 documentation links, passes its extracted offline examples, and rebuilds byte-identically.

Stacked on #7. No release tag or publication is included.