The formulation
If I were to ask you ____, would you answer affirmatively?
can get an honest answer for any yes-or-no question. So now we just have to find the best way to use the questions.
It is never optimal to choose a door until it is the only possibility left, because the chance of survival will always be better after asking a yes-no question (approximately halving the number of possibilities, so doubling the chance of survival with only a 5% penalty).
Now suppose that there is a system of questions such that it can always be narrowed down to one door (possibly choosing different questions depending on the response to previous ones). Each door can be uniquely identified by what sequence of yes/no answers leads to that door. The sequence of answers cannot start with the sequence of a different door, otherwise that door would already have been chosen. Therefore the sequences of answers form a prefix-free encoding of the doors.
Let the length of the code for door $i$ be $l_i$. If the safe door is $i$, it takes $l_i$ questions before choosing that door, so the probability of survival is $0.95^{l_i}$ This make the total probability of survival $\frac{1}{100}\sum_{i=1}^{100}0.95^{l_i}$. We want to maximize this value, under the constraint that $\sum_{i=1}^{100}0.5^{l_i}\le1$ (this is required for it to be a prefix-free encoding, since a code of length $l$ takes up $0.5^l$ of the code space).
If $\sum_{i=1}^{100}0.5^{l_i}<1$, there is some wasted sequence that does not match any door. This can only happen if some question is being asked in some situation where only one answer is possible. That question can be skipped to improve the chance of survival, so we can assume that $\sum_{i=1}^{100}0.5^{l_i}=1$.
Let the smallest $l_i$ be $m$ and the two largest ones be $n$ (there must be at least two tied for largest for the condition $\sum_{i=1}^{100}0.5^{l_i}=1$ to be true). Suppose $m+2\le n$. Then $$0.5^{m+1}+0.5^{m+1}+0.5^{n-1}=0.5^{m}+0.5^{n-1}=0.5^{m}+0.5^{n}+0.5^{n}$$ but $$0.95^{m+1}+0.95^{m+1}+0.95^{n-1}=1.9(0.95^{m})+0.95^{n-1}\\>0.95^{m}+0.95^3(0.95^{m})+0.95^{n-1}=0.95^{m}+0.95^{m+3}+0.95^{n-1}\\\ge0.95^{m}+0.95^{n+1}+0.95^{n-1}>0.95^{m}+0.95^{n}+0.95^{n}$$
So, while $\sum_{i=1}^{100}0.5^{l_i}$ remains equal to $1$, we can replace $m,n,n$ with $m+1,m+1,n-1$ to improve the chance of survival. This means that the best chance of survival is obtained when the maximum and minimum lengths differ by at most one. The only way to do this is to have 28 sequences of length 6 and 72 sequences of length 7 ($\frac{28}{64}+\frac{72}{128}=1$).
One easy way to implement this is to make the $x$th question asked (remembering to use the modification at the top of this answer):
When the number of the correct door is written in binary, is there a digit 1 in the place that represents $2^{x-1}$?
If the right door is between 37 and 64 inclusive, it will be the only one left after 6 questions. Otherwise, it will require a 7th question. The probability of survival is $0.28(0.95^6)+0.72(0.95^7)\approx70.863\%$, which is optimal as proven above.