I would separate transaction validity from policy reproducibility here. Different mempool snapshots producing different valid selections is not necessarily a correctness failure, but some properties should remain invariant: never create uneconomical change, cap the fee paid, preserve explicitly frozen or excluded UTXOs, and require deliberate approval before linking privacy clusters.
Reproducing the exact PSBT later may require recording the fee estimate, eligible UTXO set, selection-policy version, and change thresholds used at construction time. Otherwise, “same wallet state” is underspecified.
The privacy consequence is also irreversible in a way fee regret is not. That suggests consolidation should probably be opt-in or separately confirmed whenever it merges previously unrelated clusters.