Assume that $Δ=b^2−4ac≈0$ means that the actual value is $Δ=k^2·\mu$ where $μ=2^{-52}\sim 10^{-15}$ is the machine constant for the double type.
Then the floating point error for the computation of $fl(Δ)=fl(fl(b⋅b)−4⋅fl(a⋅c))$ and assuming "normal" values for $a,b,c$, for instance $a,b,c≈1,\pm2,1$, is of size $fl(Δ)-Δ=m·μ$ where $|m|<2$. Since all elementary operations are required to have an error smaller $μ/2$ one could argue that $|m|\le1$ is also true.
By Taylor or Newton, the transmitted error of the square root results from $$ \sqrt{k^2·μ+m·μ}=k·\sqrt{μ}·\left(1+\frac{m}{2k^2}+...\right) $$ so the actual error of the square root is of size $\frac{m}{2k}·\sqrt{μ}$, which can, depending on the actual size of the first ratio, contaminate about half the mantissa of the denominator.