3) It is only afterwards that one can draw the sidewalks separately 
without sending the people into a dangerous situation.

In what I assumed up today to be the consensus, there are two cases when sidewalks should already now be drawn separately:

1. (as already mentioned) when separated from the street by space (i.e., not actually laterally connected to the street)
2. when separated from the street by a barrier (e.g. as seen here on the right-hand side)

In all other cases, for the time being, the presence (or absence) of sidewalks should indeed be represented by additional tags on the street's way rather than dedicated ways for the sidewalks, and I'm not even sure whether we should pursue

1) have a tag or relation with a meaning "This sidewalk is separated for 
  geographical precision, but for routing, it is permanently connected 
to the road".
2) Have at least one routing algorithm that can use it.

Maps (and map data) must abstract (and therefore, to some degree, simplify, aggregate and interpret) reality. It must do so, not (or at least not primarily) as a means to save data size and mapping effort, but to be adequately useful at all. Taking it to the extreme, to make that point obvious: The information "there is, here on this street, a rectangular yellow paint application on the ground, and next to it another one and another one and several more"* is much less useful than the information "here on this street, there's a marked pedestrian crossing".

For sidewalks this might be less obvious, but there's an important semantic difference between "there are a street and a sidewalk next to each other, with unknown lateral relation to each other" and "there's a street with a sidewalk at the right-hand side", because in the latter case, you can (even though you might not be allowed to) walk from the sidewalk onto the street or (even though you might not be allowed to) park a car on the sideway or let someone unboard the car right onto the sidewalk or let them board the car right from the sidewalk. Or you might inadvertently trip and step onto the street. When they are physically separated by space or a barrier, this isn't usually possible and thus this warrants mapping the lateral adjacent and the physically separated case so differently.

Given tools and maps that know about and use this convention to interpret the data (and given data that actually sticks to the convention) this will help blind or otherwise visually impaired people just as much (or even more) than fully seeing-capable ones.

Regards,
Raphael

*remember the cantonal GIS that had street ground markings mapped with the actual outlines of the colored area (i.e., dashed lines as series of rectangles, arrows as 9-or-more-vertice polygons, "STOP" markings as vectorization of the actual letters "S", "T", "O" and "P", etc.) instead of their semantics in traffic? While very accurate, that probably isn't useful, unless the road traffic department had planned to renew those markings with a street-wide inkjet printer vehicle or something.