Resaving a decision could feed its previous generated paragraph back into the writing prompt, carrying unsupported reasons into a different answer. This also occurred when a failed prose job left the old paragraph in place through another save of the same answer.

Build each prose-writing prompt from the saved choice and applicable discussion reasons. Earlier prose remains the anchored replacement target; the writer is explicitly instructed to treat it as evidence only when the current discussion supports its claims.

Regression coverage exercises changed answers, repeated saves after a failed generation, retained current/shared reasons, and legacy text answers. Existing tests cover anchored replacement and durable job execution. Local targeted tests, formatting, types, and CI checks pass. This verifies the prompt boundary and write lifecycle; live model wording was not tested.