I based my number on this:
The Second Edition of the 20-volume Oxford English Dictionary contains full entries for 171,476 words in current use

So, I guess we should add "in current use" to the < 200k estimate.  Also, I can only guess that GLM's number includes every variant of every word (tense, subject, plurality, etc).  I expect it would be unwise to include all such variants for lists of words that must be precisely memorized.

In either event, I think we agree that the 1M or 200k discrepancy is largely irrelevant.  For brainwallets, there are two constraints on word selection: 1) They must be memorizable. 2) They must be randomly selectable.

Diceware uses five rolls of a six-sided die to do word selection.  This gives 7,776 possible "words", some of which aren't words, aren't well-known, and won't be easily memorized.  There are other lists out there, but they suffer the same constraints.  10,000 is a generous estimate of word pool size for this purpose.

Memorizing 12+ words, selected at random via dice roll, is a mathematically provable method to generate a sufficiently safe brainwallet.  Additional steps, shortcuts, obfuscations, etc are not necessary at best, and crippling to security at worst.