Collapse time, or preserve it

The two systems in the exercise want opposite things from the time axis.

• The recommender wants one embedding per video, so it must collapse time into a single vector.
• The surgical-phase model labels the current phase moment by moment, so it must preserve time all the way to the output.

Same input shape, same first three stages. The divergence is entirely about what the output has to be, and it decides every earlier design choice.