1. isteacher is a deprecated functions, so in my patch, I use has_capability('moodle/site:doanything', $context) instead
2. I use return filter_phrases($text, $words, true); instead of a <dummytag>, cause your students might swear using <dummytag>swear<dummytag>

Please review

1) Look the get_context_instance(CONTEXT_SYSTEM, SITEID) as commented.
2) I like the approach of hiding the original to all but admins, but...
3) Require MD approval and fixfor versions (I think we could run it for some time in HEAD in our servers and then backport to 1.9 and 1.8).

If we do then it shouldn't rely on doanything, but some special capability like

filter/censor:seecensoredwords

One possible resolution to this issue is to cache twice - one cache with censored text and one for people with the capability to read uncensored swears.
When the page is being rendered the system would decide which cache to use according to users capabilities.

Another possible solution is to create a new filtering mechanism which is pre-write to the db.
Using this method, the text would be written to two tables if a swear was used.

I.e.

moodle_messages gets the censored text

a new table

moodle_prefilter gets the text prior to any filtering taking place

Obviously I've just scratched the surface with this pre-write filter idea but I can expand on it if you think it is a viable solution.

perhaps this is a matter, for 2.0 to do something like this (to bring this and similar issues fixed for ever):

1) OOP formats, making them pluggable (this is apart but could be good idea for 2.0)
2) OOP filters
3) Allow each filter to alter the md5key generated, via filter method. filter->gethash(), the results for this could be something like, in this example: "censor:seecensoredwords" for users having the capability, or: "censor:" for the rest. Note that adding this info to the hash sounds nice because it automatically invalidates records when filters are changed, and that isn't happening right now.
4) Make the rest of format_text() and format_string() to work as now, but using that "custom" $md5key that contains all the particularities of the text being formatted.

Note that right now, for example, we are adding manually the current_language() to the cache, because of the lang filter. This will fit in the schema commented above automatically.

Guy, I'm not really sure if the "intermediate" cache (before filtering) that you comment is really a good idea or no. Perhaps you can try to perform some basic implementation and measure differences? I guess we'll lost more time reading and witting it than the time needed to format the texts, but... who knows.

Dong, how do you see the proposal? IMO could be a nice thing to change (one more) for Moodle 2.0. Comment it with MD...ciao

http://docs.moodle.org/en/Development:Filter_enable/disable_by_context

Note that http://docs.moodle.org/en/Development:Filters was not updated, which is OK, because I may be about to change things a bit.