[P0] Define eval runtime provenance and requested-vs-actual semantics

[spazyCZ]

Objective

Specify how eval tooling records requested versus actual runtime and model metadata.

Priority

P0 — Must Fix for v1.0

Details

The eval report format already distinguishes config values from observed runtime metadata, but the spec does not define the source-of-truth order between API responses, runtime JSON output, CLI metadata, and requested config/env values. This makes report provenance underspecified.

Acceptance Criteria

• The spec defines the precedence order for runtime/model provenance
• Requested and observed fields are defined consistently in eval report prose
• Fallback behavior is defined for runtimes without model introspection support

Notes

Source: work/eval-design-discussion.md