EDIT:
The issue was fixed with #3480. However, #3480 limits the maximum payload in a (transceiver-)packet, by keeping the theoretically maximum MAC-header length of 25 Bytes free. Ideally we want a dynamic solution that computes the actual MAC-header length (which may be shorter than 25 Bytes), to fill all packets with payload completely.

OLD:
In ng_sixlowpan the max. packet size NETCONF_OPT_MAX_PACKET_SIZE is requested from the hardware device which correlates with the max. fragment size of a 6LoWPAN packet. It is the basis for 6LoWPAN fragmentation, if payloads are too long. In the art86rfxx netdev-driver the device returns the value NG_AT86RF2XX_MAX_PKT_LENGTH (127). In the drivers _send function it is checked if the transceivers max. packet length plus FCS check sequence plus 802.15.4 header length is greater than the transceivers max. fifo size.
What I'm saying is: Shouldn't we return the max. packet size minus link layer header lengths to the 6LoWPAN fragmentation? Or do I get confused somewhere between fragmentation, header lengths and the transceivers fifo size?