Fresh sessions are an important isolation boundary, but they may not reset application-level answer caches, workspace memories, or retrieval reuse outside the conversation object. A new thread can therefore still inherit state that changes what the next case sees.
For the spike, I would run the same target case both alone and after a deliberately different predecessor, then compare the actual context records rather than requiring identical generated wording. Recording cache-hit flags and workspace/session identifiers in the trace would help explain any order dependence before it gets mistaken for a model-quality difference.
Ahmet Özel