The offending section:
http://docs.python.org/devguide/faq.html#how-do-i-find-which-changeset-introduced-a-bug-or-regression

I think this could be improved a bit. The key point is that "hg bisect --bad/good" is a command relative to the checked out changeset. So the instructions tell me to run "hg bisect --bad" and then "hg bisect --good", but they could do with a more explicit instruction to run "hg update <ref-id>" in between.

- You can conveniently choose a faraway changeset (for example a former release), and check that it is indeed “good”

This could be construed as just peeking at the changeset using hg log or whatever, not actually checking it out.

- Mercurial will automatically bisect so as to narrow the range of possible culprits, until a single changeset is isolated.

Here too it could be made more explicit that mercurial will navigate to (ie. check out each changeset) as it's doing this, so that at every invocation of "hg bisect --bad/good" the changeset to be tested is checked out for you, not merely computed.