fix(reporting): preserve grading provenance in human exports

[rudycelekli]

Human exports drop caveats which the run already records: Markdown omits warnings, the interactive scorecard omits both warning collections, and evidence rows label every passed=False as a failed answer. An instrument error or operator diagnostic can consequently appear to be a measured failure without the self-judge or invalid-run warning.

Preserve the existing warning collections in both human exports. Carry extraction_error and is_diagnostic into interactive evidence, show “grading unavailable” with its cause and “diagnostic” for operator evidence, and keep actual failed answers labelled as failures. These are provenance labels; individual inspections' scoring and mandatory gate policies are unchanged.

Validation:

• Unchanged main fails three warning-preservation and five evidence-provenance regressions. The combined reporting, UTC and JSON audit controls pass: 26 tests.
• Execute the production artifact JavaScript with Node against a controlled scorecard: self-judge and invalid-run warnings are visible; a budget error renders “grading unavailable” with its cause; diagnostic evidence renders “diagnostic”; the failed-answer control still renders “fail”. JSON/script-boundary escaping remains covered.
• Required Ruff 0.16.9, Bandit, layout validation and all eleven shipped example fixtures pass.
• Advisory scoped Mypy retains seven identical current-main scorecard diagnostics; the artifact module has none. Hosted CI is reported separately.

Illustrative export control, built using the real reporting payload/render functions, not a model-quality measurement:

{"warnings":["self-judge bias: Standard-mode score needs independent verification"],"validation_warnings":["run_invalid: controlled measurement failure; ignore grade"],"evidence":{"extraction_error":"budget","is_diagnostic":false},"rendered_badge":"grading unavailable"}