Fix RecursionError in TokenTextSplitter & SentenceSplitter for units larger than chunk_size

[Incheonkirin]

What I hit

I was chunking Korean text (insurance policy clauses) with a small chunk_size and the splitters crashed instead of returning anything:

from llama_index.core.node_parser import TokenTextSplitter, SentenceSplitter

TokenTextSplitter(chunk_size=1, chunk_overlap=0).split_text("🚀")        # RecursionError
SentenceSplitter(chunk_size=2, chunk_overlap=0).split_text("보험" * 50)   # RecursionError

The trigger is a single indivisible unit whose token count is already larger than chunk_size. That's easy to hit with multi-token CJK characters and emoji — 🚀 is 3 tokens, and a single Korean syllable can be several.

Why

_split (in both token.py and sentence.py) recurses on the same text once the split functions can no longer break it down — there's no base case, so it loops until RecursionError. Once that's guarded, TokenTextSplitter._merge runs its overlap-trim loop on the oversized split and pops from an empty cur_chunk, raising IndexError.

SentenceSplitter._merge actually already handles this case — it raises ValueError("Single token exceeded chunk size") — but the recursion was crashing long before that line was ever reached.

Fix

I tried to keep each splitter's existing intent rather than invent new behavior:

• _split now keeps the indivisible unit as an oversized split instead of recursing. SentenceSplitter then reaches its existing ValueError, and TokenTextSplitter keeps it as a chunk (matching its existing warn-and-keep handling).
• TokenTextSplitter._merge stops trimming overlap once cur_chunk is empty.

Normal inputs are unaffected. Added regression tests in tests/text_splitter/test_splitter_oversized_unit.py covering the oversized cases and a check that ordinary text still splits the same way.

Note

There's also a separate, pre-existing malformed logging.warning call in TokenTextSplitter._merge (two positional args with no % placeholder) that raises TypeError when it actually fires. That one already has open PRs (#21796, #21363), so I left it untouched here and just raised that logger's level in the new TokenTextSplitter test to keep this change focused on the recursion. Happy to rebase around whichever of those lands first.