Fix code review issues: security, race conditions, resource leaks

- Fix race condition: pass return_timestamps as parameter instead of mutating shared state
- Remove /local/transcriptions endpoint (path traversal vulnerability)
- Add timeout=30 and URL scheme validation to prevent SSRF
- Add temp file cleanup via try/finally in all route handlers
- AsyncTaskManager: add threading.Lock and TTL cleanup for completed tasks
- Change audio_rate to 16000 (matches Whisper's expected sample rate)
- Extract filename from URL via Content-Disposition/path
- Detect base64 audio format via python-magic
- Fix response_size_bytes to use json.dumps instead of str()
- Move scipy import to module level
- Clamp temperature to [0.0, 1.0]
- Deduplicate _load_model() code
- Fix docstrings, README structure, typo, log level

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

app/audio/utils.py

@@ -6,6 +6,7 @@ import os
 import subprocess
 import wave
 import numpy as np
+from scipy.signal import resample as scipy_resample
 import logging
 from typing import Tuple
 
@@ -33,9 +34,8 @@ def load_audio(file_path: str, sr: int = 16000) -> Tuple[np.ndarray, int]:
             sampling_rate = wav_file.getframerate()
 
             if sampling_rate != sr:
-                from scipy.signal import resample
                 num_samples = int(len(audio_array) * sr / sampling_rate)
-                audio_array = resample(audio_array, num_samples)
+                audio_array = scipy_resample(audio_array, num_samples)
                 sampling_rate = sr
 
             return audio_array, sampling_rate