Yes. Pipelining support was optional in HTTP/1.1 effectively.
Multiplexing in SPDY is so essential that if you mess it up, Google (and probably all other sites that use SPDY) won't work at all.
People are thus strongly motivated to get it right, which wasn't true of many of HTTP/1.1's effectively optional features.

As one of the creators of SPDY I can say: SPDY as specified today requires TLS and is only deployed using TLS.

yesh. :)

We definitely looked at SCTP.
It wasn't, and unfortunately isn't deployable.

I'm not sure I buy the argument that improving HTTP means we can't eventually improve the transport, btw. I think we can, but that it will take longer (e.g. ~10 years or more).

It all depends on the nature of the loss on the path the packets traverse.
Correlated (i.e. simultaneous) loss will be *worse* to the many-connection case. When a network is congested this is the likely loss-type.
Actual random loss (e.g. when using wifi and someone turns on the microwave) can cause a single connection to perform worse than many.
In most cases, the single connection can outperform multiple connections after a bit of startup time.

In all cases many connections adds to buffer bloat and decreases the ability of the TCP stack to react to real congestion.

As for server push... in that case, the server determines what the browser may need. The browser can cancel streams if it finds it already has the resource being pushed.
The server must announce what resources it will push before they're referenced in another resource (to avoid data races).
This assumes some smarts in the servers that doesn't exist in typical HTTP servers (unless they're doing inlining).

The browser signals the priority. The server does its best to respond to it.
The protocol allows this to occur.

Ask Patrick McManus, who implemented it :)
There was a spec, which has been public since day one (no reverse engineering required), along with open-source implementations of both client and server.
The SSL patch has been available as well and is hopefully going into the next OpenSSL release (but... those take a while).

Unfortunately, the spec was wrong in a couple of places, which Patrick pointed out as he went along (and so we fixed the spec). That is why it is good (and imho I wish it was still necessary) that the IETF wants to see multiple implementations by different implementors of specs.

SCTP is cool in a lot of ways. It just has some (substantial) deployment issues.
Sadly, this is true of any non-tcp protocol. I hope that we continue to play with such, but I think that replacing TCP as a transport for reliable communication is probably farther down the line.