fix(audio2x-common): close file handles and handle missing trtexec in…

[andrewwhitecdw]

… get_trt_cache_path

Summary

get_trt_cache_path() in audio2x-common/scripts/audio2x/data_utils.py (used to hash the ONNX model + trtexec binary + args into a TRT engine cache key) leaks two file handles, and crashes with a confusing TypeError: expected str, bytes or os.PathLike object, not NoneType when trtexec is not on PATH — a common situation since the TRT cache is an opt-in development speedup (A2X_SDK_USE_TRT_CACHE=true).

Root cause

digest.update(open(onnx_model_fpath, "rb").read())          # handle never closed
digest.update(open(shutil.which("trtexec"), "rb").read())   # which() may return None -> TypeError

shutil.which() returns None when the executable is not found, and open(None, "rb") raises TypeError, masking the actual problem (trtexec missing).

Fix

Read both files with with open(...) so handles are always closed, and raise a descriptive FileNotFoundError when shutil.which("trtexec") returns None.

Testing

Exercised get_trt_cache_path directly with uv run --with numpy (ephemeral env) against a fake network.onnx and a fake trtexec on PATH:

• hash is deterministic for identical inputs: OK
• changing args after the first 3 (trtexec/onnx/output paths) changes the hash: OK
• with trtexec removed from PATH:
  • after fix: raises FileNotFoundError("trtexec not found in PATH...")
  • before fix (verified with the change stashed): raises TypeError: expected str, bytes or os.PathLike object, not NoneType, confirming the bug is pre-existing

⚠️ The project's full C++/CUDA test suite cannot run locally (requires CUDA + TensorRT + generated test data); this change is confined to a pure-Python helper that was tested in isolation.

Why existing tests missed it

The TRT cache path is only exercised when A2X_SDK_USE_TRT_CACHE=true, which the header comment marks as an opt-in development speedup, and the missing-trtexec case only occurs on machines without TensorRT installed.

[andrewwhitecdw]

Closing this sweep-generated PR: PR has 2 commits; sweep requires exactly one commit per PR. It does not meet the sweep requirements (single signed-off commit).