The idea of not using a pencil is to prevent changes later. It's really important not to make changes that cannot be tracked. The best is to cross out, write in new text, and sign+date in the margin. Leaving lots of space helps in this. Frankly, though, I've seldom found small changes. When an analysis turns out to be wrong, I simply make a new entry. With my forward- and backward-referencing notation, it won't matter if the new entry corrects an error later in the day, or a decade later.
I've never worried about sunlight because my books remain closed on a shelf, when not in use. As for water, choosing a good ink helps a lot (even my fountain pen scrawls can resist quickly-dabbed-up tea stains), but making copies/scans is really a great solution.
An addendum to my earlier post -- it helps to title each entry (which often corresponds to a day or two of work), and to type those entries (plus dates, book number and page number, and also some keywords) into a file that can be searched. I recommend one line per entry, and a format that is uniform enough that you can use grep (or similar) actions to produce specialized tables of contents for individual projects, years, etc.
During my PhD, when I was scared of losing things, I made photocopies of my books when I completed them, and kept the copies in a separate location ... now I can scan them.
I have a simple cross-reference scheme. To refer to a result on page N of book M, I write [M.N]. If it's a certain figure or equation on that page, I just write e.g. [M.N.f3] or [N.M.e5].
So far, I think this is all quite conventional. But I do one more thing that's simple but has proved to be very helpful. I write forward references in the top margin of pages. So, if book 10 page 3 refers to book 2 page 9, I write <10.3> in the top margin of page [2.9]. This simple scheme makes it really easy to take care of ripple effects of errors.
Oh, and you've got to use pen, and you've got to use clean cross-outs when you find errors. Otherwise you'll get hopelessly lost.