Leaflet avoids all the complexity by relying on all the rendering done on the server, while all it has to do is load images.
Rendering vector data on the browser with the level of detalization necessary for modern cartography (hundreds of thousands of points in a single view) is not possible with the usual technologies like Canvas & SVG. The only option is WebGL, which is basically a very efficient way to draw triangles. You have to implement all the logic of turning all the data into triangles (tessellating lines, breaking down polygons, implementing anti-aliasing, figuring out collision detection between labels) from scratch, reimplementing difficult computer graphics algorithms from decades ago. This is a MASSIVE amount of work, and hard to do in a performant way.
Moreover, client-side vector rendering has even more complexity than server-side technologies like Mapnik, because you can't just render something once — you have to do it in real time, 60 frames per second, considering context, making fluid animations when zooming, panning and loading data, etc.
