The OSQ vector scorer methods have grown cluttered with nested if/else branches testing NATIVE_SUPPORTED, SUPPORTS_HEAP_SEGMENTS, and VECTOR_BITSIZE , all static final constants. I am concerned about maintenance and JIT inlining, since the large bytecode size of these methods could prevent the JVM from inlining them on the hot path. So I split the native (FFI) and Panama (Vector API) implementations into separate class hierarchies.

A NativeMemorySegmentScorer abstract base provides template methods for quantize, bulk scoring, and corrections. Three small final subclasses (one per quant type) supply only the native dot-product function pointers and bit scales. The existing MS* scorers are now Panama-only, with all native branches removed. The right implementation is selected once at class-load time via a static factory map, keyed on query/index bit combination.

I also moved the dataLength >= 16 SIMD minimum check from every method call to construction time, so the native hot-path methods have no unnecessary runtime guards. A nice consequence of this is that the filtered KNN path (scoreBulkOffsets) now uses fully native scoring, both quantize and bulk corrections end-to-end. Previously it used native quantize but fell back to scalar corrections. I also replaced readInt() loops with bulk readInts() in the base ESNextOSQVectorsScorer.