Haskell, 76 characters

import Bits
z a=170-sum[a.&.p*p|p<-[1,2,4,8]]
y a=[z a,z$a`div`16]
m=(>>=y)

Test runs:

> testAll 
input      [10, 02]
encoded    [AA, A9, A6, AA]
  pass
input      [FF, 00, AA, 55]
encoded    [55, 55, AA, AA, 66, 66, 99, 99]
  pass
input      [12, 34, 56, 78, 90]
encoded    [A6, A9, 9A, A5, 96, 99, 6A, 95, AA, 69]
  pass
input      [01, 02, 03, F1, F2, F3]
encoded    [A9, AA, A6, AA, A5, AA, A9, 55, A6, 55, A5, 55]
  pass

Haskell, 76 characters

import Bits
z a=170-sum[a.&.p*p|p<-[1,2,4,8]]
y a=[z a,z$a`div`16]
m=(>>=y)

Test runs:

> testAll 
input      [10, 02]
encoded    [AA, A9, A6, AA]
  pass
input      [FF, 00, AA, 55]
encoded    [55, 55, AA, AA, 66, 66, 99, 99]
  pass
input      [12, 34, 56, 78, 90]
encoded    [A6, A9, 9A, A5, 96, 99, 6A, 95, AA, 69]
  pass
input      [01, 02, 03, F1, F2, F3]
encoded    [A9, AA, A6, AA, A5, AA, A9, 55, A6, 55, A5, 55]
  pass

Performance is well within spec. at 1MB in ~1.2s on my oldish laptop. It suffers because the input is convert to and from a list, rather then processed as a ByteArray.

> dd bs=1m count=1 if=/dev/urandom | time ./2040-Manchester > /dev/null
1+0 records in
1+0 records out
1048576 bytes transferred in 1.339130 secs (783028 bytes/sec)
        1.20 real         1.18 user         0.01 sys

The source, 2040-Manchester.hs, includes the code, tests, and main function for a command line filter.