morisien-embed: Text Embedding Models and Evaluation for Mauritian Creole (Kreol Morisien)

Second version of the preprint. It describes two released models, morisien-embed and morisien-embed-v1.5; the version number here is the paper's, not a model's. Revision of the first version. The first version reported that a dedicated Mauritian Creole embedding model beat the general multilingual baselines. On a benchmark built to have headroom it does not, and this version says so. What is new: an out-of-domain retrieval pool built from FLORES+ with the released xSIM++ augmentation, 996 queries over 45,029 passages, on which a purely lexical baseline scores 0.3273 accuracy@1 and LaBSE 0.6928; a controlled ablation showing hard-negative mining carries the result while Matryoshka loss contributes nothing to retrieval quality; and the finding that the base model matters more than any training choice. Fine-tuning LaBSE instead of multilingual-e5-base is the only configuration measured that beats untrained LaBSE out of domain, released as morisien-embed-v1.5. Scoring follows the margin-based similarity xSIM++ specifies rather than plain cosine. Under it morisien-embed-v1.5 reaches a 0.2932 error rate against 0.3343 for untrained LaBSE, McNemar exact p = 0.00083, paired bootstrap 95% interval [-0.0653, -0.0181]. Both models fail on the same distinction: distractors that reverse a causal relation are 4.35% of the pool and cause between 43% and 50% of the errors. Also corrected: nine claims from the first version, including that the corpus was 'effectively all publicly available parallel text' (it excludes google/smol), that leakage filters removed evaluation data (they remove 0 of 69,525 rows because the splits were already disjoint), and a test-visibility optimism bound that no committed script produced, which is withdrawn