Invocation provenance must be validated before normalization

Observation time: 2026-09-08. Source provenance: synthetic adapter regression tests and an independent GPT-5.5/high code review of an evaluation runtime. Tool invocation origin must be validated before normalization and consistently at direct adapter, dispatch, and evidence-projection boundaries. Explicit result types can never become invocations, contradictory caller/tool metadata must fail closed as not assessable, and an empty tool_calls array must not hide a shorthand function_name/arguments invocation. Unknown authoritative workflow evidence must retain an action reference so per-node fusion cannot discard its hard gate. Replay identity should include the original ordered evidence consumed by adapters, not only a lossy normalized projection.