Clarify InferenceOptions usage in README and lock the API surface

[HaD0Yun]

Summary

Closes #15 by documenting the correct way to pass advanced generation controls through InferenceOptions in the README and by adding a small regression test that locks the README-facing API surface.

Problem statement

The README quickstart showed basic model.generate() usage, but it did not show how to configure advanced acoustic and flow-matching controls from Python. That made it easy for users to try passing values like noise_temperature, acoustic_cfg_scale, or num_flow_matching_steps directly to model.generate(), which is not the supported API shape.

Root cause

The runtime API already expects advanced generation controls to be bundled inside an InferenceOptions dataclass and passed via inference_options=..., but the main README did not surface that pattern where users look first.

What changed

• imported InferenceOptions in the main README inference snippet
• added a new Custom inference settings section under Run Inference
• documented that these controls belong in InferenceOptions, not as top-level model.generate() kwargs
• included a concrete example using inference_options=opts
• added tests/tada_api_surface_test.py to statically guard the InferenceOptions defaults and the TadaForCausalLM.generate(..., inference_options=...) signature

Why this design

This keeps the fix narrow and reviewable:

• it resolves the usability gap from Clarify usage of InferenceOptions for advanced audio parameters in model.generate() #15 without changing runtime behavior
• it documents the existing API instead of adding aliases or new argument plumbing
• it adds a cheap regression test so the README-facing API shape is less likely to drift in future changes

Overlap assessment

Classification: adjacent but distinct

• Issue Clarify usage of InferenceOptions for advanced audio parameters in model.generate() #15 / this PR: missing quickstart README guidance for using InferenceOptions with model.generate()
• Issue Mandarin output is much worse than VibeVoice #10: Mandarin quality / memory report; no README API-guidance overlap
• Closed PR Some optimizations for memory and speed #11: discussed inference-tuning concepts, but the actual change set did not add the missing README guidance for the supported InferenceOptions call pattern

So this PR is complementary rather than duplicative.

Validation

• git diff --check --cached
• uv run --python 3.12 pytest tests/tada_api_surface_test.py -q → 2 passed
• static README/API compatibility check against tada/modules/tada.py → passed

Risk / compatibility

Low risk.

• no runtime behavior changes
• no model default changes
• README examples now align more explicitly with the current API
• the added test is AST-based, so it is fast and avoids model downloads

Files changed

• README.md
• tests/tada_api_surface_test.py