[mudler-agent]

Description

Related to #12772.

Sound prompt expansion now requires exactly one strict, schema-backed
expand_sound tool call with tool_choice: required, instead of JSON in
assistant prose. The server parses and validates only the tool arguments.
Missing, wrong, multiple, or invalid calls stop before audio generation;
assistant prose and reasoning never provide fallback arguments.

Expansion requires llama-cpp, use_tokenizer_template: true, and a
model/template with native tool support. The request clears explicit grammar
locally without changing the installed configuration. Explicit controls,
authorization, cancellation, and requests without expansion keep their
existing behavior. Documentation and regression tests cover the new contract.

Notes for Reviewers

Verification passes for the backend, function helpers, and HTTP endpoint in
normal and auth builds. Both lint runs report zero issues. Reproduce with:

go test -p 2 ./core/backend ./pkg/functions ./core/http/endpoints/elevenlabs -count=1
go test -p 2 -tags auth ./core/backend ./pkg/functions ./core/http/endpoints/elevenlabs -count=1
golangci-lint run ./core/backend/... ./pkg/functions/... ./core/http/endpoints/elevenlabs/...
golangci-lint run --build-tags auth ./core/backend/... ./pkg/functions/... ./core/http/endpoints/elevenlabs/...

The native schema compilation check passes. Real-model inference remains
unverified because authentication blocked runtime validation. Mock-backed
HTTP tests do not establish compatibility with a particular model/template.
Enabling tokenizer templates alone does not guarantee native tool support.
Existing expansion configurations without native tools must select a
compatible model/template; unsupported configurations now fail closed.

Signed commits

• Yes, I signed my commits.
• Documentation updated (docs/content/) for user-facing changes, or not applicable