[STT] Refactor request boundary and enforce fail-fast runner behavior (#157)

This PR is:
- To move STT request parsing and validation out of the runtime path.
- To make the v1 STT runner use one normalized internal request shape.
- To fail fast on bad STT requests instead of silently returning
fallback output.

### Refactor (why this is needed)
The STT runtime path was doing two jobs:
- parsing raw vLLM request shapes
- running STT encode/decode

That made the code harder to maintain and could hide invalid input.

This refactor keeps responsibilities clear:
- serve boundary: parse + validate
- runtime path: execute only

### Note
- Added STT request adapter and normalized DTO in
`vllm_metal/stt/serve.py`.
- Added fail-fast validation for missing/invalid `req_id`,
`prompt_token_ids`, `mm_features`, and `input_features`.
- Updated `_execute_stt` to normalize request first, then run
extract/decode.
- Removed silent malformed-input fallback behavior.

---------

Signed-off-by: Yuan Lik Xun <lxyuan0420@gmail.com>

Author: LxYuan0420

tests/test_stt_serve.py

@@ -0,0 +1,72 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Tests for STT serve-boundary request normalization."""
+
+from __future__ import annotations
+
+from collections import UserDict
+from types import SimpleNamespace
+
+import numpy as np
+import pytest
+
+from vllm_metal.stt.serve import STTRequestInput, VLLMSTTRequestAdapter
+
+
+class TestSTTServeRequestAdapter:
+    """Tests for STT serve-boundary normalization."""
+
+    @staticmethod
+    def _make_request(
+        *,
+        req_id: str = "req-1",
+        prompt_token_ids: list[int] | None = None,
+        mm_features=None,
+    ) -> SimpleNamespace:
+        return SimpleNamespace(
+            req_id=req_id,
+            prompt_token_ids=prompt_token_ids,
+            mm_features=mm_features or [],
+        )
+
+    def test_normalizes_multimodal_feature_spec_payload(self) -> None:
+        mel = np.zeros((80, 3000), dtype=np.float32)
+        field_elem = SimpleNamespace(data=mel)
+        feature_spec = SimpleNamespace(data=UserDict({"input_features": field_elem}))
+
+        normalized = VLLMSTTRequestAdapter.from_vllm_request(
+            self._make_request(prompt_token_ids=[1, 2], mm_features=[feature_spec])
+        )
+
+        assert isinstance(normalized, STTRequestInput)
+        assert normalized.req_id == "req-1"
+        assert normalized.prompt_token_ids == (1, 2)
+        assert normalized.input_features is mel
+
+    def test_rejects_empty_mm_features(self) -> None:
+        request = self._make_request(req_id="broken-req", mm_features=[])
+
+        with pytest.raises(ValueError, match="broken-req"):
+            VLLMSTTRequestAdapter.from_vllm_request(request)
+
+    def test_rejects_missing_input_features(self) -> None:
+        feature_spec = SimpleNamespace(data=UserDict({"other": "value"}))
+        request = self._make_request(req_id="broken-req", mm_features=[feature_spec])
+
+        with pytest.raises(ValueError, match="input_features"):
+            VLLMSTTRequestAdapter.from_vllm_request(request)
+
+    def test_rejects_missing_feature_payload(self) -> None:
+        feature_spec = SimpleNamespace(data=None)
+        request = self._make_request(req_id="broken-req", mm_features=[feature_spec])
+
+        with pytest.raises(ValueError, match="input_features"):
+            VLLMSTTRequestAdapter.from_vllm_request(request)
+
+    def test_rejects_wrapped_none_input_features(self) -> None:
+        field_elem = SimpleNamespace(data=None)
+        feature_spec = SimpleNamespace(data=UserDict({"input_features": field_elem}))
+
+        with pytest.raises(ValueError, match="input_features"):
+            VLLMSTTRequestAdapter.from_vllm_request(
+                self._make_request(req_id="broken-req", mm_features=[feature_spec])
+            )