[Bug] C++ FLOAT/DOUBLE TS_2DIFF batch decoding skips segment prefix

[kkzi]

Description

The C++ FloatTS2DIFFDecoder::read_batch_float and DoubleTS2DIFFDecoder::read_batch_double implementations bypass the type-specific scalar decoders and delegate directly to the integer TS_2DIFF batch decoders.

FLOAT and DOUBLE TS_2DIFF segments use the Java-compatible layout and contain a maxPointNumber or overflow-bitmap prefix before the integer delta block. The integer batch decoders expect the stream to begin with the TS_2DIFF block header, so they interpret the prefix bytes as write_index and bit_width.

Symptoms

When a tree reader decodes a FLOAT or DOUBLE measurement using TS_2DIFF:

• the value stream becomes misaligned;
• the decoder reads invalid block metadata;
• end-of-input may not clear the decoder's internal remaining state;
• the tree reader can continue decoding indefinitely instead of completing the page.

Scalar read_float and read_double calls are not affected because they consume the segment prefix before decoding the integer delta block.

Root cause

The optimized batch overrides reuse TS2DIFFDecoder<int32_t>::read_batch_int32 and TS2DIFFDecoder<int64_t>::read_batch_int64, then bit-cast the decoded integers. This is only valid for the legacy raw C++ layout. It skips the scale/overflow handling required by the current FLOAT/DOUBLE segment format.

The affected methods are in cpp/src/encoding/ts2diff_decoder.h.

Expected behavior

FLOAT/DOUBLE batch decoding must consume each segment prefix and apply scale/overflow bitmap handling exactly as scalar decoding does. Batch decoding should return the page's declared number of values and terminate with no remaining decoder state.

Proposed fix

Route FLOAT/DOUBLE batch decoding through the segment-aware read_float/read_double path, and add regression tests that:

1. call read_batch_float and read_batch_double directly;
2. cross multiple TS_2DIFF segment boundaries;
3. cover both normal scale prefixes and overflow bitmap prefixes;
4. verify decoded values, decoded count, and end-of-input state.