Shell – What’s the POSIX specification on behavior of built-in commands with redirections and/or piping

Note that the problem is not with tail but with head here which reads from the pipe more than the first line it is meant to output (so there's nothing left for tail to read).

And yes, it's POSIX conformant.

head is required to leave the cursor within stdin just after the last line it has output when the input is seekable, but not otherwise.

http://pubs.opengroup.org/onlinepubs/9699919799/utilities/V3_chap01.html:

When a standard utility reads a seekable input file and terminates without an error before it reaches end-of-file, the utility shall ensure that the file offset in the open file description is properly positioned just past the last byte processed by the utility. For files that are not seekable, the state of the file offset in the open file description for that file is unspecified.

For head to be able to do that for a non-seekable file would mean it would have to read one byte at a time which would be terribly inefficient¹. That's what the read or line utility do or GNU sed with the -u option.

So you can replace head -n 20 with gsed -u 20q if you want that behaviour.

Though here, you'd rather want:

sed -e 1b -e '$b' -e d

instead. Here, only one tool invocation, so no problem with an internal buffer that can't be shared between two tool invocations. Note however that for large files, it's going to be less efficient as sed reads the whole file, while for seekable files tail would skip most of it by seeking near the end of the file.

See the related discussion about buffering at Why is using a shell loop to process text considered bad practice?.

Note that tail must output the tail of the stream on stdin. While, as an optimisation and for seekable files, implementations may seek to the end of the file to get the trailing data from there, it is not allowed to seek back to a point that would be before the initial position at the time tail was invoked (Busybox tail used to have that bug).

{ cat; tail -n 1; } < file

Even though tail could seek back to the last line of file, it does not. Its stdin is an empty stream as cat left the cursor at the end of the file; it's not allowed to reclaim data from that stream by seeking further backward in the file.

^{(Text above crossed out pending clarification by the Open Group and considering that it's not done correctly by several implementations)}

^{¹ The head builtin of ksh93 (enabled if you put /opt/ast/bin ahead of $PATH), for sockets (a type of non-seekable files) instead peeks at the input (using recvfrom(..., MSG_PEEK)) prior to actually reading it to see how much it needs to read to make sure it doesn't read too much. And falls back to reading one byte at a time for other types of files. That is slightly more efficient and I believe is the main reason why it implements its pipes with socketpair()s instead of pipe(). Note that it's not completely fool proof as there's a race condition that could be triggered if another process read from the socket in between the peek and the read.}

The grammar will only assign one word; successive words after that will be treated as a command.

Keep in mind that the rule you are citing refers to a TOKEN, and that cannot contain an unquoted space. See 2.3 Token Recognition, e.g.,

8. If the current character is an unquoted <blank>, any token containing the previous character is delimited and the current character shall be discarded.