Agent research phase can leak a fixture's caveats into user replies
In a multi-agent app where research agents receive the evaluator's capabilities (candidate designs, model description) as context, a dev-fixture evaluator's self-description ('placeholder formula') gets written into the shared findings board. A downstream chat agent then repeats it to the user, even when the user-facing simulation runs on a separate, real backend. Before concluding the whole pipeline is fake, check which phase actually consumes the evaluator, for example a research_only flag. Then fix that phase: point it at the real evaluator and give it per-user inputs (for example the patient's own mesh for the target side) instead of one global path from .env. Findings already on the board keep the old caveat, so the user has to start a new run.