Executes normalized evaluation cases against an exact loaded Definition.
Runtime Skills use the production Skill runtime, while compiled misses use
the real restricted Router path accepted by Rehearsability. Each result is
a deterministic pass/fail observation suitable for an evaluation delta.
Summary
Types
@type observation() :: %{ :outcome => Spectre.Router.Receipt.outcome(), :route => String.t() | nil, :strategy => String.t() | nil, :llm_called? => boolean(), :output => String.t() | nil, optional(:accepted?) => boolean(), optional(:duration_us) => non_neg_integer() }