It seems to me this thread is supporting the idea "physics is easy let's just enjoy writing a physics engine"!

The thread titles:

Collision detection with triangles as the basis of all computer graphics…?

It is not! As far as I can tell, triangles have never been used for collision detection extensively. Some people would point out this seemed to be the case in BSP driven engines. It is not. BSP works on hulls.

Leaving aside I don't see the necessity for CD to do graphics, , the idea you could use triangle soups for CD originated from the fact in some occasions you need per-triangle accurate effects. This is a special case, not the norm.

Reducing collision meshes to hulls is the norm. See for example kdops and other volumes, or scroll at the end of this page to see how different the representations can be.

You need the soup to create kdops and stuff. ;)

I don't think any of us has recommended doing CD on the polygon level as the primary means. But he did explicitly ask for 'small polygons'.

Depending on what kind of small polygons it is, then it would be better to use collision proxy shapes like spheres, oriented or axially aligned boxes, planes (even)..

Lots of physics libraries has the option of serializing the physics shapes so that they can be directly loaded (and not computed) at runtime.