[CompleteDotTech]

Imported evaluation rows could report accuracy above one or omit negative confidence from reliability bins. Validate finite, nonnegative measured fields and bound correctness/confidence before aggregation and bootstrap sampling. Preserve soft targets and optional ordinal error. Also run nine existing strict attention-oracle cases on CPU while retaining their CUDA cases and unchanged tolerances.

Related: #5 and #3. These changes improve reporting integrity and regression coverage; they do not establish representative quality, native-weight/CUDA acceptance or complete either issue.

Validation on this exact tree:

• Combined Python suite: 908 passed, 55 explicit skips; one upstream deprecation warning.
• Independent full-diff reviews for both contributions: no blocking findings; independent malformed-report probes passed.
• Compilation, critical Ruff, staged-source integrity, feature matrix, 24-configuration CPU FP32 numerical audit, local resume/calibration exercise, TypeScript tests, wheel/sdist and installed-wheel HTTP checks passed.
• CPU-compatible dependency installation, not a locked installation. Hosted CI is not claimed as passed.

Signed commit tree: 5527dcc3e757af9a10d5dcd1e2c293f97cdcd9b3.

Remain draft/unmerged pending required checks and policy verification. The integration denied the live protection read with HTTP403; that operation was not retried or bypassed. A pre-existing extreme-finite-value overflow limitation remains in standalone summarize(); the seed-report path rejects the nonfinite aggregate.