build-number

Put the fada vowels in the speller alphabet The inclusion list had no á, é, í, ó or ú, so the ten a/á, e/é, i/í, o/ó and u/ú transition pairs were the only thing that could put them into the error model. giella-core 1.5.2 started dropping pair rules naming a symbol outside the alphabet, which removed the accented vowels from the model altogether: "bo" produced 205 corrections and not one of them was "bó". A missing or misplaced síneadh fada is the most common Irish spelling error, so the speller could not correct the error class it most needs to. giella-core 1.14.1 stopped dropping those rules, which restores the ten pairs. This fixes the underlying cause: the five vowels are letters of Irish and belong in the alphabet, so they now take part in insertion, deletion and substitution like any other letter. "bo" now yields 25 corrections carrying a fada, including "bó". Also drop č, î and ï. They are template leftovers: in this repository they occur only in downcasing filters, hyphenation, foreign character classes and stray North Sámi text, and in no Irish stem lexicon. Removing them roughly pays for the five vowels, so the compiled error model is 49572 bytes against 56616 for the pre-1.5.2 behaviour. Note that the weight column of the transition pairs has no effect: the @_UNKNOWN_SYMBOL_@ rule that editdist.py emits subsumes those arcs, so every substitution costs the same regardless of what the pairs declare. That is an upstream issue, identical in C++ and Rust hfst, and is not addressed here.