Fill the Gaps in Every Row

Sensor matrices arrive with holes, and the model downstream needs a dense array. Interpolating each row with a Python loop and np.interp is eight lines and works; doing it for a matrix with a hundred thousand rows in an inner loop is where it stops working.

Implement interpolate_rows(x) for a 2-D array of shape (n, d) containing nan. Along each row independently, left to right:

• an interior gap, with known values on both sides, is filled by linear interpolation between them,
• a leading gap, with no known value before it, holds the first known value in the row,
• a trailing gap holds the last known value,
• a row with no known values at all is returned unchanged, still all nan.

row      = [nan, 1.0, nan, nan, 4.0, nan]
result   = [1.0, 1.0, 2.0, 3.0, 4.0, 4.0]
             ^hold      ^interpolated      ^hold

Known values are never modified.

Do it without any Python loops, including no loop over rows. Every part of this is expressible with np.maximum.accumulate, np.minimum.accumulate and fancy indexing, and finding that is the exercise.