This document is not a protocol design; it's a data store design. The protocol, Phoebe, is mentioned briefly in §5.2:
https://sentido-labs.com/en/library/201904240732/Xanadu%20Hy... and §7.4:
https://sentido-labs.com/en/library/201904240732/Xanadu%20Hy... but had not really been thought through yet. You can read the implementations of the client and server in the Udanax Green code I mentioned.
The primary concern of this architecture document from beginning to end is how to get adequate performance out of the computers of the day, given the features they considered to be fundamental.
BTW, as I understand it, the I-stream isn't the append-only data structure; the granfilade is. They contain mostly the same information, but in different orders, so that you can insert into the I-stream by appending to the granfilade.
I don't think it's true that forever storing all the text the system's users would write was infeasible in 01984; the revenues expected from readers would have paid orders of magnitude more than the cost of the space required, even in 01984. A book might be 1 megabyte; you might write another megabyte or two in the process of writing it. Someone writing a book in Xanadu — a level of success the system never in fact achieved — would have to pay to store those 2–3 megabytes forever, perhaps copying them onto a new computer every 6 years, at ⅛ the cost of the previous computer. How much might that cost?
I don't have good pricing data from 01984, but I do have here the Sun price list from December 01988; a Sun-3/100 at that time contained a 16.67-MHz 68020, 4–12 MiB of RAM, and if you got it with a disk, it would be a model 514D 141-MiB disk. The 3/140M-4-P5 configuration, with a 19" monitor, cost US$23700. 3 MB of text would have been about 2% of the machine's capacity, and so would have cost about US$237 in 01988. Four years earlier it probably would have cost about US$1000. Then it would have cost another US$125 for new storage in about 01990, US$16 in 01996, US$2 in 02002, etc., a total of a bit under US$1145. If you amortized the initial investment of the space for the book over the same 6 years, it would be about US$190 per year, so the system would be immediately profitable for any book that was expected to produce more than about US$190 of revenue per year across all its readers. A book with 100 readers per year, each paying US$10 to read the book, would earn five times the cost of its storage space.
Data transmission was plenty fast enough; people can only read at about 300 baud, or 150 baud with huffman coding. The Hayes Smartmodem 300 was introduced in 01981, and in 01984 1200-baud modems were common and 2400-baud modems were starting to appear.
Then the problem would merely be to build data centers fast enough, stuff them full of fileservers, and somehow keep the whole wild system from collapsing in chaos.
That wasn't the problem they had, though. The problem that they had was that their system design was so ambitious that they never shipped a working product.
As for video, well, even with modern codecs a two-hour-long movie can gobble up over a gigabyte of storage space (US$300 000 at the price I estimated above), and people were paying US$1 for a matinée movie ticket and US$0 to watch pan-scan-converted movies on late-night TV. There was actually a somewhat popular product a few years later to do data backup onto VCR tapes, called ArVid; it could get 4 gigabytes onto a 4-hour tape. (A few people were doing that starting in the early 01980s, but they were the people who had gigabytes of data, so there wasn't a mass-market product.) Although I was watching postage-stamp-sized MPEGs with mpeg_play_motif on SGI workstations in 01993, fullscreen digital video playback didn't become computationally feasible at all for mass-market hardware until the late 01990s; earlier systems like the Dragon's Lair arcade game overlaid the digital content on analog video backdrops.
So to deliver text to customers in 01984 and store it forever, you needed 300-baud modems, which were already a mass consumer product, a data center with fileservers, a packet-switching network like Tymnet, and textual data that averaged at least a few readers per month. Tymnet had been a commercial offering since 01972, and this whole package is something companies like The Source, CompuServe, and (as Animats pointed out) the vile Mead Data Central were already doing.
To deliver video without modern codecs, you would have needed four orders of magnitude more storage space in your data center, and you would have needed each video customer to either contract a T1 line to their premises or better, or also have the customers store the video data locally, perhaps on a VHS tape, so they could download it at slower than real time. With a 2400-baud modem you would be able to download a few seconds of video per day, though, so maybe you'd need to contract at least a 64-kbps frame relay link from the phone company, so you could download a minute of video per day. Then you could watch a single movie every other month.