[critical] Incorrect extra_infos indexing in build_by_batch_graph non-optimized path. The loop uses inner_id to index into the input extra_infos array, but when data_num != input_count (e.g., PIPNN deduplication path at line 622-628), input_indices is not a contiguous 0..N-1 sequence. The other data components in the same code block correctly use input_indices[inner_id] for indexing (see insert_codes lambda at line 672-675), but extra_info does not.

Current code:

for (InnerIdType inner_id = 0; inner_id < data_num; ++inner_id) {
    extra_infos_->InsertExtraInfo(extra_infos + inner_id * extra_info_size_, inner_id);
}

Suggested fix:

for (InnerIdType inner_id = 0; inner_id < data_num; ++inner_id) {
    extra_infos_->InsertExtraInfo(extra_infos + input_indices[inner_id] * extra_info_size_, inner_id);
}

Note: the optimized path through encode_add_batch already uses input_index correctly (line 1892).

[suggestion] The storage_preallocated condition in build_by_batch_graph (line 640-643) and prepare_add_batch (line 1817-1821) does not include extra_infos_. While resize() already grows extra_infos_ and encode_add_batch has a capacity guard, the storage_preallocated flag controls whether parallel encoding is used. If extra_infos_ capacity is insufficient but other components are fine, storage_preallocated could be false (or true in prepare_add_batch), leading to a potentially confusing error path instead of a clean preallocation.

Consider adding (extra_infos_ == nullptr or data_num > static_cast<int64_t>(extra_infos_->max_capacity_)) to both storage_preallocated conditions for consistency with the other data components.

[suggestion] Missing post-loop validation for loaded_extra_info in read_streaming_body. The EXTRA_INFO block is registered as critical in the streaming manifest (line 1432: StreamSerializationTagCritical(tag)), but after the read loop there is no check to ensure it was actually loaded when extra_info_size_ > 0 && extra_infos_ != nullptr. Other optional-but-critical blocks (precise_codes, raw_vector, paths) all have corresponding post-loop checks (lines 1678-1688), so a stream missing the EXTRA_INFO block would silently succeed without loading extra_info data.

Consider adding a check after the loop:

if (extra_info_size_ > 0 && extra_infos_ != nullptr && !loaded_extra_info) {
    throw VsagException(ErrorType::READ_ERROR,
                        "Pyramid streaming serialization extra_info block is missing");
}