Actually, no, after the first three bytes (status byte and two data bytes), running status is employed, so that if the third byte received is not a status byte, it will be regarded as the first data byte of a subsequent pitch bend message.
e.g. for channel 1:
0xE0 0x00 0x40 is a complete pitch bend message, if two subsequent bytes 0x12 0x40 are received, they are interpreted as another pitch bend value, the 0xE0 being implicit since no new status byte has been given. If the bytes are sent in quick succession, the first bend value (0x00 0x40 in this case) might not have any audible effect, and the impression one gets as a listener is that it's the last two bytes which are the only ones received or acted on.
That's what I thought it was happening, and I don't know how to intercept the incoming byte data, but it looks to me that this does not involve the implicit status byte: what aseqdump actually prints (and I get from alsaseq) is a single pitchbend event using a value out of range, as I mentioned vkeybd too is able to send and get accepted a 8192 pitchbend. Correct me if I'm wrong, but looks like a single 3 bytes message (or a 4 byte one, maybe). If that's not the case, I think I should see two pitchbend event, being the first equal 0, and the second 127.
I don't know anything about pyalsa, but if it creates more bytes than specified in the protocol because higher values than allowed are requested I'd say it's buggy.
I think that it's "allowed" for performance reasons, provided that the programmer shouldn't send out of range events, still, the conversion can generate confusion, if there is no consistence with other 2 bytes limited events from alsa.
Maurizio