Once you've written some very basic boostrap tools, the "second generation" of stuff that adds convenience and flexibility are a lot simpler.
A trivial example: 20 seconds after you wrote "directory listing", someone will say "I want a directory listing, but sorted by date, and it would be awesome if it didn't immediately scroll past the end of my screen."
With Unix Philosophy tools, you might already have have a "sort" and "paginate" command, so it's just piping stuff together. They can do it themselves, or it will take 20 seconds to explain.
Without it, you're going to have to add additional options to "directory listing" (or parallel commands) to handle the sorting and pagination features. The tools get bigger and buggier for the same functionality.
Early Unix machines weren't much bigger than mid-80s PCs-- 512K of memory or less-- but offered a very rich command line experience compared with DOS machines of similar sizes.
Programs like database or CAD packages probably go monolithic because they're more "state dependent" than your usual command-line tools. "sort" and "more" can take their inputs from stdin and feed them out to stdout, and when they're done, forget everything with no damage.
That wouldn't work well for other packages. You could probably make a database or CAD system that worked as composable units, like `echo db.sql | db-query "select username from accounts where credit < 0" | xargs delete-account` or `echo image.dxf | add-circle -x 200 -y 400 -r 60 > image.dxf` But you'd spend a lot of time reloading and reparsing the same files. A persistent monolith that keeps the data file open and in whatever internal representation is most efficient is going to perform better.
Some use cases also have limited composability, because the user can only plan a few moves ahead. Tools that encourage interactive/experimental usage, like drafting software, might involve the user stopping every step or two to make sure they're staying on plan, and queuing up a series of commands could wreak havoc. Some of these packages ended up simulating composable tools through internal macro/scripting languages which still avoided the penalty of having to rely on the OS to orchestrate every single action.
Sorting the output of textual tools like ls requires parsing which can be non-trivial. It's easier to do it by using a modern structural shell such as nushell.
(Offhand, I've never touched Plan 9 but...) Hypothetically /proc/SOMEPID/db/DATABASE/SCHEMA/TABLE/various views which provide expressions in some order. Or /proc/SOMEPID/containerofthings/ and the directory listing is serviced by the application, as an enumeration of keys (filenames) to values (datasets). For a database the API would behave similarly to how ORMs operate since filesystems are inherently similar to objects.
It kind of feels a bit of cargo cult, praising it all the time.
