Spoken Arabic dialect identification (DID) has advanced quickly, but Libyan Arabic remains under-served and, where it appears, is labeled only at the country level, hiding the variation between the Libyan coast and interior. We present a Libyan Arabic speech corpus annotated at the sub-regional level: Tripolitania (west), Cyrenaica (east), and Fezzan (south). It contains 133.7 hours of single-speaker segments from 645 speakers across 161 channels (YouTube and TikTok sources). An automated pipeline (diarization, speaker embedding, clustering, cross-episode linking) produces silver labels from channel provenance. One annotator then reviewed the resulting speaker clusters and confirmed the provenance region for 95.5% of the kept speakers (a within-annotator figure). We release speaker-disjoint splits. An off-the-shelf Arabic DID model recognizes our verified Libyan speech as Libyan only 74% of the time and cannot resolve the region. On the new three-way task, the encoder matters more than the classifier: a probe over speaker-recognition ECAPA embeddings reaches 0.525 macro-F1, over a language-identification ECAPA reaches 0.660, and fine-tuning that encoder reaches 0.702 (75.5% accuracy), far above the 0.20 floor. A channel-disjoint re-test reaches 0.609, versus 0.702 speaker-disjoint; Fezzan is hardest at 0.313. The task is learnable, but Fezzan performance remains low. We release the corpus, cached embeddings, and code.