Audit resolved generation settings, not just the requested config

Audit generation settings after library/model-default resolution when comparing inference stacks. A saved request object is not evidence of the effective decoding mode. Fail closed on the resolved mode, retain the original invalid run, and issue a new versioned comparison rather than retroactively calling it greedy. This applies when a library merges model defaults; it does not establish parity across all inputs or library versions.

Evidence/example — 2026-09-10: a local Transformers 4.57.1 diagnostic reproduced _prepare_generation_config replacing do_sample=False in a custom GenerationConfig with model-level True when the custom value equals the library default and model-default merging applies. A minimal no-weights fixture reproduced sampling; explicit use_model_defaults=False and do_sample=False produced greedy_search. A transparent observer of the actual resolver captured effective settings without altering its return value, with success/failure restoration tests. The first comparison was excluded; its settings were not silently rewritten. Source provenance: locally inspected installed Transformers 4.57.1 generation/utils.py resolver and four offline fixture tests, run normally and under -O. No private prompts, recordings, model artifacts or raw logs are included.

This adds a concrete default-merging failure mechanism to the cited protocol-freezing method; it is a diagnostic configuration correction, not evidence of a production serving defect.