I think I finally found a solution, but it's ugly and hopefully someone knows a much cleaner way.
Let L be the positive solution to
L = \frac{1}{4}\left(1 + \frac{2}{L}\right)
or equivalently
4L^2 - L - 2 = 0
Thus, L \approx0.84307.
If (x_n) has a limit, then it must be L because it's easy to see that x_n is never negative.
The hard part is proving that (x_n) has a limit. I shall do this by proving that the sequence
y_n = x_n - L
has a limit of 0. To do this, I shall find a constant C < 1 such that
\left|\frac{y_{n+1}}{y_n}\right| < C
for all n.
First, I need a small lemma, namely:
|y_n| < 0.2 for all n
This will be proved inductively. Clearly it's true for n = 0, because
|y_0| = |x_0 - L| \approx |1 - 0.84307| \approx 0.15693
Now suppose that |y_n| < 0.2; we must show that this implies |y_{n+1}| < 0.2. Indeed, |y_n| < 0.2 implies that L - 0.2 < x_n < L + 0.2. Therefore,
\frac{2}{L+0.2} < \frac{2}{x_n} < \frac{2}{L-0.2}
and this means
\frac{1}{4}\left(1 + \frac{2}{L+0.2}\right) < \frac{1}{4}\left(1 + \frac{2}{x_n}\right) < \frac{1}{4}\left(1 + \frac{2}{L - 0.2}\right)
This is the same as
0.72935 < x_{n+1} < 1.0275
and we also have L - 0.2 < 0.72935 and 1.0275 < L + 0.2, so
L - 0.2 < x_{n+1} < L + 0.2
or equivalently
|y_{n+1}| < 0.2, concluding the proof of the lemma.
For the rest of the proof, we substitute x_n = y_n + L into the original recurrence relation, cross-multiply the denominator, and use the fact that 4L^2 - L - 2 = 0 to simplify the result to
4y_{n+1}y_n + 4L(y_n + y_{n+1}) = y_n
and solving for y_{n+1} we get
y_{n+1} = \frac{y_n(1 - 4L)}{4(y_n+L)}
Then, using the fact that
\frac{1}{|y_n+L|} = \frac{1}{|x_n|} = \frac{1}{x_n} < \frac{1}{L - 0.2}
we have
|y_{n+1}| = \frac{|y_n| \cdot |1 - 4L|}{4|y_n+L|} < \frac{|y_n| \cdot |1 - 4L|}{4|L - 0.2|} = |y_n| \cdot 0.92225
Thus
\frac{|y_{n+1}|}{|y_n|} < C
where C = 0.92225, concluding the proof.
It ain't pretty, but it gets the job done, unless I screwed up somewhere.
By the way, there is nothing magic about the number 0.2. At some point I concluded that I needed a number less than 0.25 but big enough to handle the variation of the sequence about its limit, and I chose 0.2 after "cheating" by plotting the sequence in Matlab.
