Yes, running this gives a max lag at 0. That is indicating that y matches x the best in their current positions. Shifting y by any value will make it agree less. You're expecting to see a lag of 3/50 = 0.06 right? (This is equal to 6 samples).
clearvars
clc
close all
fs = 100;
t = 0:1/fs:2;
x = sin(t);
y = sin(t + 3/50);
[r1,lagsy] = xcorr(x,y);
[mr1,locmaxy] = max(r1);
lagsy(locmaxy)/fs
% ans = 0
y2 = circshift(y,6); % shift it by 6 samples
[r2,lagsy2] = xcorr(x,y2);
[mr2,locmaxy2] = max(r2);
lagsy2(locmaxy2)/fs
% ans = 0
% Plot the results
figure
plot(t,x,t,y,t,y2,'k--')
legend('x','y','y2','location','southeast')
Notice that mr1 > mr2. Look at the plot of y2. It's easy to see that any shift of y would lead to less correlation with x than the original signal.
HOWEVER, if you allow the oscillations to go through one full period, then yes, you get the correct value of 0.06.
clearvars
clc
close all
fs = 100;
t = 0:1/fs:2*pi;
x = sin(t);
y = sin(t + 3/50);
[r1,lagsy] = xcorr(x,y);
[mr1,locmaxy] = max(r1);
lagsy(locmaxy)/fs
% ans = 0.06
