Fix Phi-4 memory reproduction review findings
Reject package-local paths and non-finite logits. Surface sampler failures, handle missing CSV TTFT, and isolate opt-in arena logging.
Add regression coverage and refresh the related documentation and package checksums.
Validation: 82 host tests, Ruff and clang-format, offline integrity review, and Phase A builds against both historical runtime variants passed. No GPU inference was run.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
