chat: Muse Glimmer silently ignores JSON schema constraints #58

Upstream: ggml-org#29613

Muse Glimmer's chat preparation ignores a supplied JSON schema, producing no constrained-generation grammar.

Confirmed at the model-free chat layer on Spark Linux ARM64 CPU, upstream 1c4729414deba3bcacef6a5614e2ca04d069be3a. Full Glimmer model inference/HTTP was not tested.

Reproduction using common/chat.h and libllama-common: initialize templates with common_chat_templates_init(nullptr, template_text) using models/templates/muse-glimmer.jinja; call common_chat_templates_apply() with:

common_chat_templates_inputs input;
common_chat_msg msg;
msg.role = "user";
msg.content = "Tell me about Paris.";
input.messages = {msg};
input.reasoning_format = COMMON_REASONING_FORMAT_AUTO;
input.json_schema = R"({"type":"object","additionalProperties":false,"required":["city"],"properties":{"city":{"type":"string"}}})";
auto result = common_chat_templates_apply(templates.get(), input);

Observed: result.grammar.size() == 0; prompt, parser and grammar are identical with/without the schema. Control using google-gemma-4-31B-it.jinja yields a 1,599-byte grammar and changed parser.

Cause: common_chat_params_init_muse_glimmer() never consumes inputs.json_schema and only enables grammar when tools are present. Introducing commit not bisected.