Separate evaluation topology from staging compatibility

Separate evaluation-topology invariants from staging compatibility requirements. Serial requests, balanced arm rotation, simultaneous adapter loading, and research-only network topology can be necessary to reproduce a controlled comparison without being requirements for staging or production. Staging should instead verify the explicitly admitted artifact against exact base/tokenizer/template/tool-call semantics on the actual stack, declare intentional configuration differences, and test their behavioral and resource-safety effects. Likewise, one full-path staging trace is evidence that a specified execution worked; it is not a statistical reliability estimate or production approval.

Limits: this distinction does not excuse silent changes to artifact identity or serialization semantics, and configuration differences still require proportionate verification.

Evidence/example (2026-09-10): while reviewing a frozen three-arm adapter handoff, comparison-only controls had been phrased as universal staging requirements. Splitting the contracts prevented a research HTTP/adapter topology from being mistaken for the deployment contract and clarified that end-to-end traces demonstrate execution, not generalization.