Extract config dataclasses from `openlrc.py` and defer heavy imports

[MaleicAcid]

Extract config dataclasses from openlrc.py and defer heavy imports

Problem

TranscriptionConfig and TranslationConfig are plain @dataclass classes that only depend on stdlib (dataclasses, pathlib.Path). But they live in openlrc/openlrc.py, so importing them triggers the entire module-level import chain:

# openlrc/openlrc.py — all executed on `from openlrc import TranscriptionConfig`
from faster_whisper.transcribe import Segment       # -> ctranslate2, onnxruntime
from openlrc.preprocess import Preprocessor          # -> torch, deepfilternet
from openlrc.transcribe import Transcriber           # -> faster_whisper, pysbd
from openlrc.translate import LLMTranslator          # -> openai, anthropic, google-genai
from openlrc.utils import Timer, ...                 # -> spacy, tiktoken, lingua, torch

Downstream projects that embed these dataclasses (e.g. via composition for their own config types) are forced to install and load ~3GB+ of dependencies just to use two stdlib-only classes. This also prevents packaging tools (Nuitka, PyInstaller) from producing lean binaries.

Proposed Changes

1. Move config dataclasses to a lightweight module

Move TranscriptionConfig and TranslationConfig to openlrc/config.py (or the existing openlrc/models.py which already only depends on stdlib). Re-export from __init__.py:

# openlrc/__init__.py
from openlrc.config import TranscriptionConfig, TranslationConfig  # stdlib only
from openlrc.openlrc import LRCer                                  # heavy

from openlrc import TranscriptionConfig becomes instant. Non-breaking — all existing import paths still work.

2. Defer heavy imports in openlrc.py to method level

Even from openlrc import LRCer (without calling any method) currently loads torch, spacy, faster-whisper, etc. These can be deferred to the methods that use them:

Import	Used in	Pulls in
openlrc.preprocess.Preprocessor	pre_process()	torch, deepfilternet
openlrc.transcribe.Transcriber	transcriber property (already lazy-init)	faster_whisper, ctranslate2
openlrc.translate.LLMTranslator	_translate()	openai, anthropic, google-genai
faster_whisper.transcribe.Segment	to_json() type hint	faster_whisper

Lightweight internal imports (openlrc.context, openlrc.defaults, openlrc.logger, openlrc.models, openlrc.opt, openlrc.subtitle) can stay at module level.

Test compatibility: Tests in test_preprocess.py patch module-level names like @patch("openlrc.preprocess.enhance"). After deferral these would need to target the source module (e.g. @patch("df.enhance.enhance")). I'm aware this was the concern in #87 — happy to include test updates in the same PR.

Benefits

• from openlrc import TranscriptionConfig loads stdlib only
• from openlrc import LRCer no longer triggers torch/spacy/whisper at import time
• No API changes, no new dependencies
• Complements Decouple torch/DeepFilterNet from core dependencies #88 — lazy imports make optional extras actually effective at runtime

Related

• chore: add cpu/cu124 extras for PyTorch index selection #87 — lazy import for preprocess.py (deferred due to test patch concerns)
• Decouple torch/DeepFilterNet from core dependencies #88 — decouple torch/DeepFilterNet into optional extras