Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@backends/advanced/init.py`:
- Around line 404-416: The embed model dict in the branch guarded by
embedding_model_name sets a hardcoded "embedding_dimensions": 1536 which can
break if the user supplies a different embedding model; update the construction
of that dict (the embed_model creation) so embedding_dimensions is not hardcoded
but derived or configurable: add a lookup or helper (e.g.,
get_embedding_dimensions(embedding_model_name)) that returns a sensible default
(1536) but first checks an env/config value (e.g., CUSTOM_EMBED_DIM) or prompts
the user for dimensions, and use that value for the "embedding_dimensions" key
instead of the literal 1536.

In `@config_manager.py`:
- Around line 328-345: add_or_update_model currently assumes model_def contains
a "name" key and will raise a raw KeyError; add an explicit guard at the start
of add_or_update_model to validate that model_def is a dict and contains a
non-empty string under "name" (e.g., if not isinstance(model_def, dict) or not
model_def.get("name") or not isinstance(model_def["name"], str): raise
ValueError("model_def must include a non-empty 'name' string")), then proceed to
call _load_config_yml, update or append the model, and _save_config_yml as
before so callers get a clear, consistent error message instead of a KeyError.

In `@extras/asr-services/tests/test_cuda_version_config.py`:
- Around line 162-167: Update the inaccurate inline comment in the arch_to_cuda
mapping: change the comment for the "sm_90" key (currently marked "RTX 4090,
H100") to reflect that sm_90 corresponds to Hopper (H100/H200), and update or
clarify the "sm_89" comment to indicate sm_89 is Ada Lovelace (RTX 4090); edit
the comments next to the keys "sm_90" and "sm_89" in the arch_to_cuda dict
accordingly.
- Around line 172-179: The test test_older_gpus_work_with_cu121 is tautological
because it asserts each arch from older_archs is in the same hard-coded list;
replace this with a real validation by comparing older_archs to the
authoritative CUDA 12.1 supported list (e.g., a constant or helper like
CU121_SUPPORTED_ARCHS or get_supported_archs_for_cuda("12.1")). Update the test
to either assert set(older_archs).issubset(set(CU121_SUPPORTED_ARCHS)) or
iterate and assert each arch is in CU121_SUPPORTED_ARCHS so the test actually
verifies that CUDA 12.1 supports those architectures rather than re-checking the
source list. Ensure you reference the test name test_older_gpus_work_with_cu121
and the older_archs variable when making the change.

In `@tests/configuration/test_transcription_url.robot`:
- Around line 16-23: The test shows a double-http prefix caused because
backends/advanced/init.py sets default prompt values for PARAKEET_ASR_URL and
VIBEVOICE_ASR_URL with the "http://" scheme, while the config template also
prepends "http://"; fix by removing the "http://" scheme from those default
values in backends/advanced/init.py (update the default strings for
PARAKEET_ASR_URL and VIBEVOICE_ASR_URL to e.g. host.docker.internal:8767) so the
existing Resolve Omega Config/template which prefixes http:// produces a single
correct URL (references: PARAKEET_ASR_URL, VIBEVOICE_ASR_URL, and the config
template key model_url).

In `@tests/libs/ConfigTestHelper.py`:
- Around line 49-59: Add a symmetric cleanup helper to remove the temp config
directory created by create_temp_config_structure: implement a method named
cleanup_temp_config_structure(self, base_path: str) that imports shutil,
computes config_dir = Path(base_path) / "config" and calls
shutil.rmtree(config_dir, ignore_errors=True) so stale config.yml files are
removed even if tests fail; ensure the new method is documented and intended to
be called from test teardowns.
- Around line 29-35: The current resolve_omega_config helper patches os.environ
(via patch.dict) which is process-global and not thread-safe; instead, modify
resolve_omega_config to avoid patching os.environ and register an explicit
OmegaConf resolver that reads from the provided env_vars dict (falling back to
os.environ) so resolution is isolated per call: convert env_vars values to
strings, register a resolver (OmegaConf.register_resolver) or use a custom
function passed into OmegaConf that returns env_vars.get(key) ||
os.environ.get(key), create the config with OmegaConf.create(config_template),
then call OmegaConf.to_container(conf, resolve=True) and return the result;
ensure you unregister or scope the resolver per-call if necessary to avoid
global side effects.

---

Nitpick comments:
In `@backends/advanced/init.py`:
- Around line 382-397: The code hardcodes "model_output": "json" in the
llm_model dict used by self.config_manager.add_or_update_model, which may be
unsupported by some OpenAI-compatible providers; change this to not force JSON
by either omitting the model_output key when unknown, defaulting to a safer
"text" value, or making it configurable (e.g., read from an env var like
CUSTOM_LLM_MODEL_OUTPUT) before building llm_model so add_or_update_model
receives a compatible model_output or none.

In `@extras/asr-services/docker-compose.yml`:
- Around line 117-119: The docker-compose.yml currently hardcodes Google DNS
under the dns directive for the vibevoice-asr service; change this to be
configurable via an environment variable so containers in restricted networks
can override it. Replace the literal dns list for the vibevoice-asr service with
a variable placeholder (e.g. use DNS_SERVERS with a default like
${DNS_SERVERS:-8.8.8.8,8.8.4.4}) and update README/compose usage to document the
expected comma-separated format (or how to pass a YAML list) and how to set
DNS_SERVERS in the environment or .env; ensure the docker-compose parsing
handles splitting or provide examples for both default and custom values so the
dns directive is not permanently hardcoded.

In `@extras/asr-services/tests/test_cuda_version_config.py`:
- Around line 89-112: The service-block regex is fragile to comment lines and
should match only the service header plus subsequent indented lines; in
test_docker_compose_cuda_arg_consistency replace the current service_match regex
(used to set service_match and service_text) with one that anchors the service
name at line start and captures following lines that begin with whitespace
(e.g., use a pattern like start-of-line service_name: followed by zero or more
lines that start with spaces/tabs) with re.MULTILINE (avoid DOTALL), so
service_text contains only the indented service block before the next top-level
key or EOF and the subsequent PYTORCH_CUDA_VERSION assertion remains valid.

In `@tests/libs/ConfigTestHelper.py`:
- Line 41: The local stdlib import "from urllib.parse import urlparse" should be
moved to the module's top-level imports in ConfigTestHelper to follow idiomatic
practice: add "from urllib.parse import urlparse" to the file's import block
alongside the other imports, remove the inline/local import statement (the one
currently at line containing "from urllib.parse import urlparse"), and keep
existing uses of urlparse in the function/method unchanged (e.g., any usage
inside methods of the ConfigTestHelper class).