The entire point of the HAL is to abstract hardware, any hardware, away from the OS. There's nothing that says it can't encompass more of the hardware than just the IO bus, CPU and MMU, like WinNT does. On an embedded device there's very little in terms of a standard IO bus that the OS can communicate through cleanly with peripherals, so might as well abstract the whole lot.

Perhaps by keeping the machine that hosts the seed secured? Like using a protocol between the publicly facing machine and the seed machine that doesn't allow for remote shell access? Really basic stuff, actually.

The unethical part, as far as I understand, is that smaller ISPs rent the "last mile" piece from Bell, which they're allowed to since the infrastructure is wholly, or partially, tax-payer funded. However, they don't buy big-pipe bandwidth from Bell, but instead peer with someone like Cogent. The cost of the bandwidth over the last mile is zero, since additional bytes don't degrade the infrastructure and therefore don't add to maintenance costs. However Bell wants to charge the ISP, for this zero-cost bandwidth, at the same scale as they charge their end-users, who, unlike the ISPs, *are* using their peering connection to talk to the rest of the internet.

I keep seeing this population density argument being thrown out as if it has some sort of a bearing on the issue.

You're not going to be rolling 100Mbs fiber to each farm house in Montana, now are you?

The fact that ISPs in North-America (Canada included) are unable to do this even in areas with high population density, simply indicates that they want to keep the status quo as long as customers keep paying. There is no technical or logistical issue preventing them from doing it. It's all about abject lack of competition and the precious dollar. Trotting out the density argument is specious at best.

Sure if you only implement the single-precision portion of IEEE-754 then you're still working with 32-bit quantities on an 8-bit computer. All that bit-jiggling really adds up quickly.

I once coded a fully IEEE-754 compliant single-precision floating point emulator for ARM which was about ~100 instructions on average per operation. And this is with an instruction set that handles 32-bit quantities natively.