It's a bit hard to remote-debug this. You absolutely need to look at the frames in wireshark. Use 'tcpdump -s 9999 -w dump.cap' to record a pcap file.

Classical bugs of home-grown MAC drivers:

 - Forgetting to pad frames to 60 bytes. Result: frame may be dropped by the receiving MAC because of undersize.
 - In the IRQ handler, reading only one frame from the MAC queue, instead of reading the whole queue. Result: frames may be delayed (by several seconds) until another frame arrives.
 - When copying data to/from the MAC word by word, pay attention to handling of an odd number of bytes. Otherwise the last byte may get mangled.
 - Endian confusion. Usually, everything is big endian (network order). It's easier to avoid mistakes if you assemble your frame byte by byte, rather than casting from integers or structs.

PS: have you chosen the MAC yourself? Are you sure it's not a multicast MAC? Try 00:00:00:AA:BB:CC instead.