Fix: chunk oversized examples (max 11k chars -> max 971) that caused NaN loss during training

Initial public corpus: redacted flock_dreams + memory, 9175 examples