Just to clarify one thing: buffer objects are frequently referred to as though they were a new GL feature but in fact they're not; they date all the way back to OpenGL 1.5 and the GL_ARB_vertex_buffer_object extension. This is well before the Longs Peak plans and what GL 3.x did that was new was make using them mandatory in core contexts.

The major missing feature from Longs Peak remains the object model; or at least part of it. Much of it is there but GL still suffers from atrocious type-safety which a well-specified object model should have fixed.

The biggest "feature" of longs peak is that it would've been a fresh API / broken backwards compatibility with existing GL code.

GL3 half-assed that by deprecating old API interfaces, but Windows drivers kept supporting them anyway (Mac actually killed them off, yay!).

And yes, Vulkan has actually achieved this goal by coming up with a new API from scratch (well, from Mantle), so there's not three decades of GL cruft hanging off of it.

...why did they scrap the new spec and wait to implement these features later?

There's never been an official statement about this.

If you were there at the time, the way it played out was that the ARB were making all the right noises about Longs Peak and everybody was excited, keen and supportive. Then they announced "some unresolved issues that we want addressed before we feel comfortable releasing a specification", went into a total media blackout for some time, and eventually emerged with the OpenGL 3.0 we know of today.

The closest to an answer you'll get is this post on the OpenGL forums:

What happened to Longs Peak?

In January 2008 the ARB decided to change directions. At that point it had become clear that doing Longs Peak, although a great effort, wasn't going to happen. We ran into details that we couldn't resolve cleanly in a timely manner. For example, state objects. The idea there is that of all state is immutable. But when we were deciding where to put some of the sample ops state, we ran into issues. If the alpha test is immutable, is the alpha ref value also? If we do so, what does this mean to a developer? How many (100s?) of objects does a developer need to manage? Should we split sample ops state into more than one object? Those kind of issues were taking a lot of time to decide.

Furthermore, the "opt in" method in Longs Peak to move an existing application forward has its pros and cons. The model of creating another context to write Longs Peak code in is very clean. It'll work great for anyone who doesn't have a large code base that they want to move forward incrementally. I suspect that that is most of the developers that are active in this forum. However, there are a class of developers for which this would have been a, potentially very large, burden. This clearly is a controversial topic, and has its share of proponents and opponents.

While we were discussing this, the clock didn't stop ticking. The OpenGL API *has to* provide access to the latest graphics hardware features. OpenGL wasn't doing that anymore in a timely manner. OpenGL was behind in features. All graphics hardware vendors have been shipping hardware with many more features available than OpenGL was exposing. Yes, vendor specific extensions were and are available to fill the gap, but that is not the same as having a core API including those new features. An API that does not expose hardware capabilities is a dead API.

Thus, prioritization was needed, and we made several decisons.

1) We set a goal of exposing hardware functionality of the latest generations of hardware by this Siggraph. Hence, the OpenGL 3.0 and GLSL 1.30 API you guys all seem to love

2) We decided on a formal mechanism to remove functionality from the API. We fully realize that the existing API has been around for a long time, has cruft and is inconsistent with its treatment of objects (how many object models are in the OpenGL 3.0 spec? You count). In its shortest form, removing functionality is a two-step process. First, functionality will be marked "deprecated" in the specification. A long list of functionality is already marked deprecated in the OpenGL 3.0 spec. Second, a future revision of the core spec will actually remove the deprecated functionality. After that, the ARB has options. It can decide to do a third step, and fold some of the removed functionality into a profile. Profiles are optional to implement (more below) and its functionality might still be very important to a sub-set of the OpenGL market. Note that we also decided that new functionality does not have to, and will likely not work with, deprecated functionality. That will make the spec easier to write, read and understand, and drivers easier to implement.

3) We decided to provide a way to create a forward-compatible context. That is an OpenGL 3.0 context with all deprecated features removed. Giving you, as a developer, a preview of what a next version of OpenGL might look like. Drivers can take advantage of this, and might be able to optimize certain code paths in the forward-compatible context only. This is described in the WGL_ARB_create_context extension spec.

4) We decided to have a formal way of defining profiles. During the Longs Peak design phase, we ran into disagreement over what features to remove from the API. Longs Peak removed quite a lot of features as you might remember. Not coincidentally, most of those features are marked deprecated in OpenGL 3.0. The disagreements happened because of different market needs. For some markets a feature is essential, and removing it will cause issues, whereas for another market it is not. We discovered we couldn't do one API to serve all. A profile encapsulates functionality needed to meet the needs of a particular market. Conformant OpenGL products may implement one or more profiles. A profile is by definition a subset of the whole core specification. The core OpenGL specification will contain all functionality, including what is in a profile, in a coherently designed whole. Profiles simply enable products for certain markets to not ship functionality that is not relevant to those markets in a well defined way. Only the ARB may define profiles, individual vendors may not (this in contrast to extensions).

5) We will keep working on object model issues. Yes, this work has been put on the back burner to get OpenGL 3.0 done, but we have picked that work up again. One of the early results of this is that we will work on folding object model improvements into the core in a more incremental manner.

6) We decided to provide functionality, where possible, as extensions to OpenGL 2.1. Any OpenGL 3.0 feature that does not require OpenGL 3.0 hardware is also available in extension form to OpenGL 2.1. The idea here is that new functionality on older hardware enables software vendors to provide upgrades to their existing users.

7) We decided that OpenGL is not going to evolve into a general GPU compute API. In the last two years or so compute using a GPU and a CPU has taken off, in fact is exploding. Khronos has recognized this and is on a fast track to define and release OpenCL, the open standard for compute programming. OpenGL and OpenCL will be able to share data, like buffer objects, in an efficient manner.

There are many good ideas in Longs Peak. They are not lost. We would be stupid to ignore it. We spent almost two years on it, and a lot of good stuff was designed. There is a desire to work on object model issues in the ARB, and we recently started doing that again. Did you know that you have no guarantee that if you change properties of a texture or render buffer attached to a framebuffer object that the framebuffer object will actually notice? It has to notice it, otherwise your next rendering command will not work. Each vendor's implementation deals with this case a bit differently. If you throw in multiple contexts in the mix, this becomes an even more interesting issue. The ARB wants to do object model improvements right the first time. We can't afford to do it wrong. At the same time, the ARB will work on exposing new hardware functionality in a timely manner.

What's clear about this is that the job of specifying Longs Peak was taking too long, design-by-committee wasn't working and OpenGL was once again at risk of being overtaken even further. So something had to be done, and what was done was a rush-release of OpenGL 3.0 with promises to fold in much of the Longs Peak features/improvements in future releases, which is what eventually happened.

At the time there were other rumours, one of which was that CAD interests on the ARB killed it, because their CAD programs were using the old/crufty behaviours that Longs peak threatened to remove, and they didn't want to upgrade them. The statement quoted that "there are a class of developers for which this would have been a, potentially very large, burden" suggests CAD vendors for sure, but IMO this rumour was never 100% credible, because of course CAD vendors would have always had the option to just continue using the older API (just as even today you can still write an OpenGL 1.1 program). A second rumour I recall is that one of the GPU vendors killed it for some non-specific reasons.

On balance of probability the situation I describe above (spec taking too long and competition getting further ahead again) seems most likely to me.

I recall hearing Blizzard mentioned as a possible villain of the piece too.

Amusingly, the one company we can be absolutely certain didn't kill it is Microsoft; they had long since ceased involvement with OpenGL by then.