bug: PartitionTransformer::new unwraps a fallible index build

LABEL: bug

Description

PartitionTransformer::new in rust/lance-index/src/vector/ivf/transform.rs builds the centroid lookup index with .unwrap():

let index = SimpleIndex::may_train_index(
    centroids.values().clone(),
    centroids.value_length() as usize,
    distance_type,
)
.unwrap();

may_train_index is genuinely fallible — it calls FixedSizeListArray::try_new_from_values twice and SimpleIndex::try_new, which builds an HNSW over the centroids. rust/CLAUDE.md rules that out: "Never use .unwrap(), .expect(), panic!(), or assert!() in library code for fallible operations".

The fallible part is reached when the HNSW speedup for centroid assignment is in play: under SimpleIndexStatus::Auto that needs at least 1,000,000 centroid values, and LANCE_USE_HNSW_SPEEDUP_INDEXING can force it on regardless. So a large-scale index build is the case that turns a training error into a panic instead of a returned error.

Expected behavior

The error propagates. PartitionTransformer::new and the IvfTransformer constructors that call it return Result, the way new_ivf_transformer_with_quantizer already does.

Lance version

13.0.0-beta.4 (main)

Language binding

Rust

Start in rust/lance-index/src/vector/ivf/transform.rs at PartitionTransformer::new and trace the IvfTransformer constructors that call it. Follow may_train_index and its existing Result paths, including the HNSW speedup condition and environment-variable override. Done means index-training failures propagate as errors instead of causing a panic.