Best practice for picking convenient IPv6 addresses for a few hosts on an isolated LAN

Here are things I've noticed after working with IPv6 addresses:

1. Don't use "full" addresses. The :: abbreviation goes a long way.
2. Inside your organization you'll probably deal with a single prefix. This means you can all but ignore the first part of the address.
3. Unless you're using glass teletypes copy and paste exists. Use it.
4. As @TomTom points out, DNS exists. Use it.
5. When you start working with v6 addresses patterns will quickly emerge. You start to recognize things like organizational prefixes (HE, Cisco, etc), Teredo, ISATAP, and stateless autoconfiguration.

IPv6 address don't have to be complex. For example, this is the address of a large company's public web server:

2001:420:80:1::5

I'm assuming the "80" is a subnet reserved for web servers. Unlike IPv4 you could also use the subnet "443" for HTTPS-only servers.

The "Unique Local Address" is exactly what you're looking for. fc00::/7 gives you enough bits that if you generate a random number instead of just picking one the chances of collision are small.

Does this mean I'll need extra protections so my router would not automatically start advertising these private IPv6 addresses to the world?

The RFC that covers these ULAs (RFC4193) specifically states that these numbers should not be routed on the internet, though two peers may mutually agree to pass certain prefixes. Unless Comcast decides to unilaterally route these (unlikely in the extreme) you should have no worries about route advertisement.

Assuming I will never, ever, tie that IPv6 address into the real internet (a router will NAT & firewall it), can I ignore the RFC to an extent and go with fc00::4:0/120?

Don't assume that. For instance, Comcast is currently doing IPv6 trials and they're passing out /64's to end-users (slide 5); not just the single address they're doing with IPv4. This means that their now-running IPv6 testers have the option of running with globally routeable addresses, but firewalled by their router, or do some kind of NAT with either link-local or unique-global-addresses.

However, running without any kind of address translation is not as insane as it sounds. Keep in mind a few points.

• Comcast is handing out a /64 subnet to you, so your attacker already knows what your IP space looks like.
• A /64 provides a mind bogglingly huge number of potential addresses. 2^64 worth! That's four billion IPv4 Internet's worth of IP addresses. (2^64 == 2^32 * 2^32. Four billion times four billion .) While the nature of IPv6 autoprovisioning reduces the actual number of addresses that need scanning, scanning it is still infeasible.
• Unless you set up your own domain to provide it, Comcast will not be providing forward or reverse DNS lookups to your /64-worth of IP addresses. This greatly reduces the ability of attackers to recon your network.
• Running without NAT makes certain network problems easier, and certainly makes undesirable but very popular peer-to-peer technologies (you know what I'm talking about) a lot easier to get up and running.

Running without a firewall is still just as insane as it sounds, though. Happily, you can do firewalling without having to NAT.

Second question, what's this link-local thing?

Think of it as able to reach anything in the current broadcast domain, and can not be routed. Like NetBEUI-of-old. In fact, if your home network is completely flat you can use these addresses instead of Unique Local Addresses.

Third question, what is scope id for?

It's used for two different things, which makes it annoying to describe:

Thing 1: Multicast. It defines how far the multicast packet is intended to reach.

Thing 2: (What I think you're referring to) This is used on a URI as a way of defining which interface to use. It's used primarily with link-local addresses. It should never be used in conjunction with CIDR notation, so the two syntaxes should never be combined.