Given that we've already chosen to use the fake `shape=...` keyword when there is a 0-length axis, what do you think we should do, consistency-wise?
np.empty([10, 0]) array([], shape=(10, 0), dtype=float64)
1. Follow the precedent and use the fake `shape=...` keyword in the summarized-array case. 2. Ignore the precedent and use a following `# shape=...` comment afterwards in the summarized-array case and leave the 0-length-axis case alone. 3. Fix the 0-length-axis case to use the following `# shape=...` comment too.
Consistency-wise, I guess option 3 (fix 0-length-axis case) is the best one; whether it's worth code churn in NumPy... this can go either way, so option 2 (ignore the precedent and keep 0-length-axis arrays alone) is fine, too, IMO. Evgeni