Here's my source for the models thing...
From Brian Harris, of Id Software..."The only limitations I know of as far as texture size is set by the
video card. For example, my 5900 Ultra supports textures up to
4096x4096, of couse I'd quickly use up my video ram that way (a single
32 bit texture of that size is 64MB). As for polycounts, the only real
limitation has to do with the 8 bit stencil buffer and shadows. If
there are more than 256 overlapping shadow planes then the shadows will
get totally blown out (things that should be in shadow aren't etc..).
There's no real easy way to check for when this happens except for
through trial and error. At one time, we actually had the full poly
characters running around and it worked fine (just horribly slow)."
Also, yes Carmack mentioned the texture seam as a problem, but it's not an engine problem. Also, the plastic problem he mentioned wasn't anything that couldn't have been fixed (and it has been fixed with the glossmap mod).
Basically, someone took Humus' Doom 3 tweak which changed the specular lookup table into math operations, then the guy added a line or so of code which allowed the value "specExp" or Specular Exponent, which is what Carmack was talking about in his next engine doing that kind of lighting, rather than a bunch of biases and squares which could sort-of simulate a specular power, but due to the older cards he had to support, even the fragment program paths had to emulate the older GeForce 3 stuff, to be modified, through the diffuse alpha channel (it apparently doesn't work in the specular alpha channel).
So yeah, the specular problem wasn't at all a problem to fix, however it's fixed now, so you can't use that as a D3 limitation any longer, as you can specify one different gloss map per material, to simulate plastic, metal, skin, wood etc..
How many times have I repeatedly stated this problem has been fixed?
EDIT: He only mentioned you couldn't use a different interaction program per-surface. This isn't as necessary with gloss-maps (tighter/broader highlights).
Now what this prevents you from doing is stuff like sub-surface scatttering per-material, so light doesn't just bounce or transport the same way through everything. Also, partial translucency etc.. would best use multiple interaction programs.
Now, what more do I have to say?
EDIT2: Also regarding the HDR. It's not a low-dynamic range hack. It renders to a floating point buffer (16 bits per component) using floating-point blending (note, carmack has been advocating floating point render buffers since January 2003, so I don't think he'd HACK it). Basically the engine has the core HDR stuff, and you can turn it on, write custom fragment programs which use HDR etc.. but it's not integrated well enough for artists to have an easy time. So yes, it has HDR and it's rendering to a true floating point rendering buffer (of course there's that tone-mapping pass in the end which adds the post-processed effects to the 8bit per-component scene, but better than Valve's psedo-HDR)
Please get your facts straight. I've thoroughly researched the Doom 3 engine versus CryEngine etc... and I don't talk out of my ass.