If the proportional hazards (PH) assumption won't hold between your (+) RT and (-) RT groups for high-stage cases (e.g., RT was withheld because of prior cancer with radiation, or because of poor patient performance status, so a different shape of survival curve might be expected) then you can simply break your high-stage stratum in two: high-stage(+)RT, high-stage(-)RT. That just requires minimal recoding of the data.
If PH is expected to hold, then you could just add the +/- RT variable to all cases. Make sure that (-) RT is the reference level for that factor (whether the first or last level for a factor is the default for reference depends on software).
Then the RT value will be (-) for all low stages and the hazard ratios (HR) reported for the other predictors will be for (-) RT. The (+) RT HR will represent the (presumably lower) hazard for those at high stage receiving RT. You should avoid the temptation to use the (+) RT HR to estimate what might happen if RT were used in lower-stage cases.