Problem

Java-implemented TableProviders currently expose only a full-schema scan(BufferAllocator). The Rust adapter receives a projection and limit, but does not pass them to Java; Java sources also cannot report DataFusion's Exact/Inexact/Unsupported filter-pushdown capabilities. This prevents external Java sources from selecting a selective access path while preserving DataFusion's optimizer semantics.

Proposed scope

Add an opt-in scan-planning contract alongside the existing API:

• Immutable scan requests with projection indices (including an empty projection), structured DataFusion expression protos and an optional limit.
• Per-expression Exact/Inexact/Unsupported capability reporting.
• A scan plan that returns a fresh allocator-compatible Arrow reader only during physical execution. Planning must not read source data.
• Source projection with schema validation, while existing providers retain full-schema scans and the adapter's projection fallback.

The implementation would remain a thin, source-agnostic Java/JNI adapter. No storage SDK, source-specific predicate rules, query-language semantics or DataFusion-core changes are proposed. Tests would use an in-memory selective source to cover exact and inexact filtering, residual filters with LIMIT, projection, independent scans and resource/error handling.