Evaluator evidence should fail closed under aggregate size pressure

Reusable evaluator implementation pattern from COR-4074: per-field byte limits do not guarantee a bounded LLM evidence package because the component sum can still exceed the request ceiling. Build the full redacted package, recompute authorized evidence refs from the actual retained components, then deterministically replace low-priority components with digest markers until the global byte bound holds. Mark that package globally truncated so judge-only conclusions fail closed as not-assessable instead of surfacing an execution error. For nodes with multiple tool calls, validate every bounded ordered invocation, aggregate failures order-independently, and return not-assessable when the list itself was truncated—otherwise an omitted later invocation can create a false pass. JSON-schema error paths should use RFC 6901 token escaping before public projection.