Re: PEG Morphology Algorithm -- design
Jorge, this is just marvelous work — I'm in awe. (I'm also envious of the amount of free time you appear to have. 
However, I have a concern about the overall approach you're taking — the high-level design, as it were.
The grammar in its current state does four separable things:
1. It partitions the input stream into words.
2. It validates the words, rejecting invalid vowel and consonant patterns.
3. It determines the selma'o of a cmavo.
4. It categorizes brivla into gismu, lujvo and fu'ivla.
As a result, the grammar is fearsomely complex in spots. (OK, the part that recognizes selma'o isn't complex; it's just huge.) And it could be argued that categorizing brivla really belongs to semantic analysis, not parsing.
For the sake of modularity and reducing point-complexity, I think it would be worth considering splitting the job into its components, and writing separate grammars:
1. A partitioning grammar that considers an input string, and accepts a word (cmene, brivla, cmavo or non-Lojban) from its head.
2. A validating grammar that considers a Lojban word, and rejects it (re-categorizing it as non-Lojban?) if it has invalid vowel or consonant patterns.
3. Selma'o determination might be more easily described as a symbol table lookup than as a parsing problem.
4. A grammar that considers a valid Lojban brivla, and categorizes it.
Of course this scheme depends on being able to combine multiple PEG-generated parsers into a single program. But if the parser generator takes parameters which can be used to name the input and parser functions, that shouldn't be hard.
Or is there already a consensus that the requirement is for a single grand grammar covering every relevant aspect of the language?
Clark Nelson