Define the task vocabulary once, and let the ABI ask about it

[christopherthompson81]

Define the task vocabulary once, and let the ABI ask about it

The fourteen task kinds are written out in four places that nothing compares, and they have drifted in every direction. Four separate-looking bugs turn out to be that one fact.

where	what it says
src/framework/model_spec/schema.cpp	the set a spec's tasks may contain
src/framework/model_spec/metadata.cpp	spec name → VoiceTaskKind
src/framework/runtime/session.cpp	ABI token → VoiceTaskKind, and back
include/audiocpp.h	prose examples of the spellings

The drift

• codec is in the schema's allowed set and no parser maps it to a task kind. model_specs/miocodec.json declares ["codec", "vc", "s2s"], and parse_tasks iterates in order, so it throws on the first element — the family never loads.
• audio_generation is the mirror image: parse_task_kind accepts it, the schema does not, so a spec using it fails validation instead.
• model_specs/moss_voicegen.json declares "vdes", the ABI token, where the spec vocabulary wants "design". Also unreachable.
• audiocpp.h documents "diarization" and "alignment" as task spellings. Neither has ever parsed. Because audiocpp_model_supports returns 0 for an unrecognised task, a caller following the header is told no for every model that does diarize, with nothing to indicate the question was malformed.
• parse_voice_task_kind's error message lists the tokens a fifth time.

The change

engine::runtime::task_vocabulary is now the single definition. The schema's allowed set, both parsers, to_string, and both error messages read it. They cannot disagree, and adding a task kind is one row.

The two specs that declared unmappable names are fixed: moss_voicegen uses design; miocodec drops codec, which names no task kind the runtime can serve. If codec should be a real task, that is now one row rather than four edits — flagging it as the one semantic decision here rather than deciding it quietly.

Three additive C entry points

None of this was askable. Every task-aware call takes an audiocpp_model, so a caller deciding what to install — reading model_specs/*.json to build a picker before anything is downloaded — has had to hardcode a copy of the mapping table:

size_t       audiocpp_task_count(void);
const char * audiocpp_task_name(size_t index);
const char * audiocpp_task_from_spec_name(const char * spec_task);

Nine of the fourteen spellings are identical between the two vocabularies, which is what makes comparing them directly appear to work. A binding that did so showed 0 packages for music generation when nine families serve it, and 1 of 33 for voice design.

And one more, for the same shape one layer down

audiocpp_status audiocpp_request_set_text_language(audiocpp_request *, const char *);

audiocpp_request_set_text writes options["language"] as well as the transcript language, mirroring the CLI's --language. The two cannot be separated through the ABI — and they have to be:

• parakeet_tdt validates request options against its contract and refuses a language it does not declare.
• qwen3_forced_aligner declares no language option and requires the transcript language anyway.

So "not declared" means must not send for one and must send for the other, and no static check decides it. A client can only discover which by failing a request.

The reference server needs exactly this separation and gets it by reaching around the ABI — drop_unsupported_language_option in app/server/runtime.cpp erases the option and keeps the transcript language. A C ABI client cannot express that function, so the ABI is strictly less expressive than the engine it wraps, in a way the server demonstrates is load-bearing.

Verification

Against parakeet_tdt on a real model, which is the exact pair from the report:

set_text(text, "en")                            -> unknown Parakeet TDT request option: language
set_text(text, NULL) + set_text_language("en")  -> accepted

Tests added to model_spec_system_test:

• the vocabulary is self-consistent in both directions — every token round-trips through parse_voice_task_kind/to_string, every spec alias maps to its token, and a name that maps to nothing (codec, diarization) reports as such rather than resolving;
• every spec this repository ships declares task names the engine knows. Reverting the two spec fixes fails it by name:

  unmappable task: moss_voicegen.json declares 'vdes'
  unmappable task: miocodec.json declares 'codec'

ctest: 35/35 pass. A C probe over the new entry points covers enumeration, out-of-range, NULL handling, and the four spec→token rows that differ.