SparseEncodingDecoder: silently ignores chunk_offsets metadata; does not reject 3rd child

[dfa1]

Root cause

The Sparse encoding spec (encoding-format/dict-runend-sparse.md) states:

Tags 4-6 (chunk_offsets) are in-memory-only. A conformant reader rejects a node that carries them and requires exactly two children (patch_indices, patch_values).

SparseEncodingDecoder.java:56-59 reads only tags 1-3 (len, offset, indices_ptype) and silently ignores any extra tags. The decoder also never validates that exactly two children are present.

Impact

Hardening gap only — no wrong output from production Rust-written files (no production writer emits chunk_offsets or a 3rd child). A malformed or future-format file carrying these fields would be silently accepted rather than rejected with a VortexException.

Fix

1. After decoding PatchesMetadata, assert ctx.childCount() == 2 — throw VortexException if not.
2. Consider validating that patch_indices are strictly sorted (another spec invariant currently unchecked); unsorted indices would silently corrupt output.

Reference

Spec: https://github.com/vortex-data/vortex/blob/mp/spec/docs/specification/encoding-format/dict-runend-sparse.md

[dfa1]

Fixed in 845ebb1