Loading any mlx-community whisper repo (whisper-large-v3-mlx, whisper-base-mlx-4bit, whisper-base.en-mlx, etc.) crashes on first transcription with ValueError: Processor not found. These repos ship weights only, so WhisperProcessor.from_pretrained raises during load — leaving _processor = None after only a warning.

This PR adds a fallback: read the architecture signature from config.json and retry the processor load against the canonical openai/whisper-* repo that produced this architecture. Processor files are architecture-independent (~4 MB), so a one-time download recovers transcription with no user intervention.

Architecture-keyed lookup

The 5-tuple (n_audio_state, n_mels, n_audio_layer, n_text_layer, n_vocab) from config.json uniquely identifies each canonical openai/whisper variant. vocab_size = 51864 flags .en English-only models; n_mels = 128 flags large-v3 family; n_text_layer = 4 flags large-v3-turbo. large-v1 and large-v2 share dims (their processor files are interchangeable), so the lookup maps that signature to large-v2.

Identifying by dims rather than directory name handles the real mlx-community landscape uniformly — ~50+ repos with arbitrary suffixes like -4bit, -8bit, -q4, -fp32, -asr-*, plus user-renamed local directories. Both openai/mlx config keys (n_audio_state, n_mels, …) and HF Transformers keys (d_model, num_mel_bins, …) are read.

Error handling

Catches OSError specifically on the local load (transformers' signal for missing files) instead of bare Exception. Other failures — corrupt JSON, permission errors — propagate so a fine-tuned local checkpoint can't be silently masked by the canonical OpenAI processor; a vocab mismatch would generate garbage transcription with no error signal otherwise. The transformers import catches ImportError specifically.

Network behavior

When the fallback fires, the canonical processor is fetched from HF Hub (~4 MB). Set HF_HUB_OFFLINE=1 or TRANSFORMERS_OFFLINE=1 in air-gapped environments — transformers raises a clear offline-mode error rather than waiting on a network timeout. Documented in the load-helper docstring.

Behavior

Before, on mlx-community/whisper-base-mlx-4bit (or any non-canonical name):

warning: Could not load WhisperProcessor: Can't load feature extractor for '<path>'...
model._processor = None
# later, on first generate():
ValueError: Processor not found. Make sure the model was loaded with a HuggingFace processor.

After:

warning: Loaded WhisperProcessor from openai/whisper-base as fallback because <path> is missing processor files: ...
model._processor = <WhisperProcessor>
# transcription succeeds

Repos whose architecture isn't a recognized canonical variant, or that lack a readable config.json, preserve the existing "warn and set _processor = None" behavior.

Tests: 11 unittest cases covering dim-based resolution (tiny / quantized / large-v3 / large-v3-turbo / .en / HF Transformers config format), behavior preservation (missing config / unknown dims / local success), and error propagation (ValueError propagates, canonical fallback failure leaves _processor = None).

Fixes #645