Does Tessary's frustration classifier ever read another call site's messages as context?

No. The classifier’s prior-turn context comes only from earlier turns of the same conversation that reached the same call site; a turn that only ran a router or a memory pass doesn’t count as one of them, even though it happened earlier in the same thread. That’s a fix, not always the behavior: the scorer used to pick a trace’s first root span regardless of which call site it belonged to, so a router’s prompt could get read as the user’s own message and a memory pass’s JSON reply as the assistant’s, and head-and-tail clipping on long messages sometimes kept that noise while cutting the real reply the user saw. Call sites exist so Tessary can tell a router’s prompt from a user’s message in the first place; frustration’s scoping is this classifier’s own use of that distinction. One user turn that fans out to a router call, a reply, and a memory pass behind the scenes only ever gets judged against the two earlier turns on the one call site that actually replied to the user.