While inspecting the tests you added, I realized with horror that
`(a ,@b) is NOT equivalent to `(a . ,b) despite the rewrite suggested
in CLHS 2.4.6,
because earlier in 2.4.6, it says that the former is like
(append 'a b nil) whereas the latter is like (append 'a b)
and that's different because the first one is invalid when b is not a
proper list.
I disabled in fare-quasiquote the optimizations that depended on this identity.
You might want to do the same in SBCL.

Nobody seems to take that identity seriously in the case of nil as the final atom.  Nil means that you just don't have a terminating atom to transform using the [x] rule, not that all preceding non-atoms are transformed and then followed by (QUOTE nil).
If they did it by the book, then sure, these tests are all invalid:

CLISP: (eval(read-from-string "`(,@1 . nil)")) => 1

ABCL: (eval(read-from-string "`(,@1 . nil)")) => 1

CCL: (eval(read-from-string "`(,@1 . nil)")) => 1

ECL: (eval(read-from-string "`(,@1 . nil)")) => 1