> ------------
> # -*- mode:my/test -*-
> This is test with keyword to be fontified.
>
> # Local Variables:
> # eval: (setq unsafe-variable t)
> # End:
> -------------
>
> Then, consider the following major mode:
>
> (define-derived-mode my/test-mode text-mode "Test"
>   ""
>   (add-hook 'hack-local-variables-hook
> 	    (lambda ()
> 	      (setq-local my/test-mode-keywords '(("keyword" . font-lock-keyword-face)))
> 	      (setq font-lock-defaults '(my/test-mode-keywords)))
>             nil 'local))
>
> 1. emacs -Q
> 2. Evaluate the major mode definition
> 3. Open the file
> 4. Answer "y" in the unsafe variable prompt
> 5. Observe "keyword" not being fontified.
> 6. Expected: "keyword" fontified using font-lock-keyword-face.

Hmm... AFAICT the problem here is that the implementation of
`global-font-lock-mode` ends up trying to enable `font-lock-mode` in
that file's buffer during execution of the
`after-major-mode-change-hook` of *another* buffer while querying
whether to obey those file-local settings, and then fails to try again
when the hook is run in the desired buffer.

I suspect the patch below might help (requires recompiling
`font-core.el` and re-dumping Emacs), but as the comment in there
explains it might not always be sufficient either.

Hmm...


        Stefan