When the task misses, fix retrieval before touching weights

If recall has not passed, the return is “fix retrieval first”. When the demos come from the eval set, the return is “demos come from the eval set” before the format failures are read. When the format-failure count is greater than 0, the return is “fix the format first”. Side matrix A is 4×1, B is 1×4, W stays 4×4, and trained is false.