fix: off-by-one error in RecursiveJsonSplitter.split_json

[Gamble Tan (gambletan)]

Summary

Fixes #29153

Two bugs in RecursiveJsonSplitter._json_split() that cause data loss at chunk boundaries:

1. Off-by-one boundary check: Changed size < remaining to size <= remaining so items that fit exactly at the chunk size boundary are added to the current chunk instead of being pushed to a new one.

2. Empty dict/leaf value loss: When a value doesn't fit in the current chunk and a new chunk is started, the code recursed into the value. For empty dicts {}, the for key, value in data.items() loop has zero iterations, silently dropping the key-value pair from all chunks. Added explicit handling to directly set leaf values and empty dicts instead of recursing.

Reproduction

from langchain_text_splitters import RecursiveJsonSplitter

data = {
    "projects": {
        "GTMS": {f"GTMS-{i}": {} for i in range(1, 23)},
        "ITSAMPLE": {f"ITSAMPLE-{i}": {} for i in range(1, 4)},
    }
}

splitter = RecursiveJsonSplitter(max_chunk_size=300)
chunks = splitter.split_json(data)

# Before fix: GTMS-10 and ITSAMPLE-2 are lost at chunk boundaries
# After fix: all items are present in chunks

Test plan

• Verify the reproduction case from Offset by 1 bug on RecursiveJsonSplitter::split_json() function #29153 no longer loses data
• Verify existing tests pass
• Verify chunk sizes still respect max_chunk_size

🤖 Generated with Claude Code

[era (avinashkamat48)]

This changes the splitter behavior for boundary-sized values and empty dict leaves, but I do not see a regression test in this PR. The previous bug sounds easy to reintroduce because the <= remaining condition and the non-empty-dict recursion are both subtle. Could you add tests for a value exactly equal to the remaining chunk size and for an empty dict / scalar leaf that previously got dropped?