+ move gfsmArcList back to a native type, ditching GSList
  - sorting, splicing, etc. is then tricky, but maybe we can use
    temporaries for this 
  - would save us sizeof(void*) per arc: all the space lost by adding source state!

+ (style): clearer API/internals demarcation

+ allow (arbitrary?) automaton indices
  - esp. (source_state,lower), (source_state,upper)
  - load & save with automaton
  - either:
    (1) maintain indices over destructive operations, or
        - generic hooks for destructive gfsmAutomaton API?
    (2) maintain a "dirty" marker to note when index need update, or
    (3) maintain a "smart list" of changed data and update on demand

+ get rid of "GArray fsm->states": it's just wasting space
  (30 of 64 bits are unused: track these at compile time with -Wpadded)
  - replace with (gfsmBitVector)s state_is_valid, state_is_final and
    a dedicated arc-list array (we can still call that "states" for
    backwards compatibility, esp. the fsm->states->len idiom)
    * tagh from 148MB to 129MB: -13%

+ allow gfsmArcList to be an array
  - better: allow *all* arcs for an automaton to be a single array

+ [DONE] source state tracking in gfsmArc
  - obsoletes gfsmReverseArcIndex
  - adds sizeof(gfsmStateId) bytes per arc (tagh from 148MB to 184MB: +24%)

+ Basic types
  - configure option for gfsmLabelId to fit a pointer
    * tricky to handle without a union or other explicit label API
    * without source tracking: tagh from 148MB to 184MB: +24%
    * with source tracking   : tagh from 184MB to 220MB: +20%
  - configure option for 64-bit gfsmStateId (also 64-bit gfsmLabelId?)
    * breaks binary compatibility: add header flag?

+ store alphabet with fsm
  - tricky with user alphabets: loadable modules?

+ loadable modules for user-defined alphabets & semirings
  - libgmodule ?
  - look at Micha's tomata2 code for ideas