In fact, the internal system uses Float32, which has a value of 23 bits. Using a floating point avoids most clipping problems, while ensuring good accuracy. This means that technically it makes no sense to try to create a wave curve that exceeds 8388608 (2 ^ 23); but in fact, the 16-bit curve is quite high resolution (signal-to-noise is ~ 96 dB). The big reason for 32-bit audio processing was to avoid clipping problems without improving the I / O SNR; using floating point helps this dramatically. By the way, WaveShaperNode is compressed to [-1, +1] (by the way, most nodes do not work).
In short, just use 16-bit (65535), but make sure your signal is in the range -1, + 1.