quantized duplicated precisions.json, which is authoritative for the
quantized/non-quantized partition. default expressed a per-repository
recommendation nothing consumed: the library resolves its own per-platform
default. Both are gone from the schema.
MLX: quantize embeddings (4w/64) and emit get_model_schema
Adds qembedding_config=4w/group_size=64, matching the sibling embedding exporter: 284.9 MB -> 188.0 MB (-34%) with no measurable quality change (per-token cosine 0.99327 vs 0.99363, MaxSim retrieval identical) and +0.4 ms on an iPhone 16 at 128 tokens.
Also ships the get_model_schema companion method, so the runtime sees the {2..512} sequence range and the equality constraint tying input_ids/attention_mask/output seq dims instead of falling back to constant MethodMeta dims.
Also fixes the lowering config so the artifact loads at all: the previous build failed with Failed to load method forward, error 0x23 because it lacked both get_default_passes() and _skip_dim_order=True.
Add LFM2.5-ColBERT-350M (XNNPACK 8da4w + MLX int4) for react-native-executorch
