feat: load bundled model providers by default

[lipikaramaswamy]

Summary

• Plain Anonymizer() now loads bundled providers.yaml and passes it to DataDesigner, so notebooks, CI, and benchmarks no longer depend on ~/.data-designer/model_providers.yaml.
• User-supplied model configs must declare explicit provider: fields; init validates provider names against the resolved provider list and rejects empty provider lists with InvalidConfigError.
• Docs, agent skill, and tests updated to reflect bundled defaults and hardened provider/model resolution.

Test plan

• make test (833 passed)
• make format-check
• Verified bundled models.yaml provider names match providers.yaml
• Verified Anonymizer() passes bundled providers to DataDesigner
• Verified custom model_providers= override and empty-list rejection
• Verified missing/unknown provider config raises InvalidConfigError

[greptile-apps]

Greptile Summary

This PR makes Anonymizer() load bundled providers.yaml by default instead of relying on DataDesigner's machine-local ~/.data-designer/model_providers.yaml, and adds validation that rejects empty provider lists and model configs that reference unknown provider names.

• Bundled provider defaults: _resolve_model_providers(None) now calls load_default_model_providers() to load the new src/anonymizer/config/default_model_configs/providers.yaml, which ships the NVIDIA build endpoint. Anonymizer.__init__ always passes a resolved list[ModelProvider] to DataDesigner — never None.
• Stricter config validation: User-supplied model configs must now declare an explicit provider: field; the new _validate_raw_model_configs_have_provider and validate_model_configs_reference_providers helpers enforce this at init time and wrap all ValueErrors as InvalidConfigError.
• detection_workflow.py bug fix: validation_single_chunk_full_text was silently dropped when detect_and_validate_entities called _build_detection_spec; the parameter is now forwarded correctly.

Confidence Score: 5/5

Safe to merge; the bundled defaults path is well-tested, existing callers that relied on None providers are handled by the new fallback, and all validation errors are correctly surfaced.

The change is focused and thoroughly tested (833 passing tests, dedicated new tests for all new code paths). The bundled providers.yaml matches the provider names in models.yaml, validated by a dedicated test. The only gaps are minor defensive-programming omissions in non-critical paths.

model_loader.py (load_default_model_providers lacks an empty-list guard) and detection_workflow.py (build_detection_config and build_detection_builder_for_seed still don't expose validation_single_chunk_full_text).

Important Files Changed

Filename	Overview
src/anonymizer/interface/anonymizer.py	Core init now loads bundled providers when None, wraps all ValueError from validation as InvalidConfigError, and removes the now-dead None-providers fallback in _resolve_model_hosts. Changes are correct and well-tested.
src/anonymizer/engine/ndd/model_loader.py	Adds load_default_model_providers, validate_model_configs_reference_providers, and _validate_raw_model_configs_have_provider. Logic is sound; load_default_model_providers lacks the empty-list guard present in the user-supplied path.
src/anonymizer/engine/detection/detection_workflow.py	Fixes validation_single_chunk_full_text not being forwarded in detect_and_validate_entities. build_detection_config and build_detection_builder_for_seed still don't expose the parameter.
src/anonymizer/config/default_model_configs/providers.yaml	New bundled provider file defining the NVIDIA build endpoint. Correctly structured with name, endpoint, provider_type, and api_key fields.
tests/engine/test_model_loader.py	Solid new tests covering missing provider, empty provider string, unknown provider, bundled coverage check, and empty list rejection. All existing tests updated to include provider field.
tests/interface/test_anonymizer_interface.py	New integration tests verify bundled providers are passed to DataDesigner, custom overrides work, and invalid configs raise InvalidConfigError. Good coverage of the new behavior.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["Anonymizer.__init__(model_configs, model_providers)"] --> B["parse_model_configs(model_configs)"]
    B --> C{model_configs is None?}
    C -->|Yes| D["Load bundled models.yaml\n→ ParsedModelConfigs"]
    C -->|No| E["Parse user YAML/Path\n_validate_raw_model_configs_have_provider\nload_model_configs()"]
    E --> F["ParsedModelConfigs"]
    D --> G["_resolve_model_providers(model_providers)"]
    F --> G
    G --> H{model_providers is None?}
    H -->|Yes| I["load_default_model_providers()\nbundled providers.yaml"]
    H -->|list| J{empty?}
    H -->|str/Path| K["load_config_file()\ncheck providers key\ncheck not empty"]
    J -->|Yes| L["raise ValueError"]
    J -->|No| M["list[ModelProvider]"]
    I --> N["list[ModelProvider]"]
    K --> O["list[ModelProvider]"]
    M --> P["validate_model_configs_reference_providers"]
    N --> P
    O --> P
    P --> Q{unknown providers?}
    Q -->|Yes| R["raise ValueError → InvalidConfigError"]
    Q -->|No| S["DataDesigner(model_providers=resolved_providers)"]

Loading

_{Reviews (2): Last reviewed commit: "fix: resolve CI failures after merging m..." | Re-trigger Greptile}

[greptile-apps]

Unreachable fallback with stale type hint

_resolve_model_providers now guarantees a non-empty list[ModelProvider] (loading bundled defaults when None, raising on empty list), so self._resolved_providers is always non-empty here. The if not providers: branch at line 874 is dead code and its return value was quietly changed from NVIDIA_BUILD to OTHER. The function signature also still advertises list[ModelProvider] | None, which contradicts the actual call-site type. If a future caller ever passes None or an empty list while relying on the OTHER fallback, the telemetry host would be misclassified compared to prior behaviour.