Findings

High: Table isolation is still violated when overlap_all=True, so adjacent text/table chunks continue to leak content across the boundary.

Description: The PR blocks table/text from sharing the same _elements list and blocks PreChunkCombiner merges, but it does not isolate the separate overlap state. PreChunkBuilder.flush() carries overlap_tail forward into _overlap_prefix, and tables still consume that prefix when the next pre-chunk is emitted. That means a "standalone" Table can still contain prior narrative text, and a text chunk after a table can still start with table content.

Why it is risky: The isolation invariant in the PR description is stronger than "not in the same CompositeElement". Here, hidden state crosses the boundary through _overlap_prefix, so callers still get mixed-content chunks under a supported option. That is silent data corruption, not just a presentation issue: embeddings, indexing, table reconstruction, and downstream classifiers will operate on polluted table/text payloads even though the element types look correct.

Concrete failure scenario: In production, a retrieval pipeline enables overlap_all=True for recall and chunks documents containing text/table boundaries. On this branch, chunk_elements([Text("Alpha beta gamma delta."), Table("H\nC")], overlap=5, overlap_all=True, new_after_n_chars=0) yields a table chunk with text "elta.\nH\nC". The inverse boundary leaks too: chunk_elements([Table("H\nC"), Text("Omega sigma tau upsilon.")], ...) yields the next CompositeElement as "H C\n\nOmega sigma tau upsilon.". So the PR still breaks its advertised isolation guarantee under a documented option.

The mechanism is visible here:

        # -- copy element list, don't use original or it may change contents as builder proceeds --
        pre_chunk = PreChunk(elements, self._overlap_prefix, self._opts)
        # -- clear builder before yield so we're not sensitive to the timing of how/when this
        # -- iterator is exhausted and can add elements for the next pre-chunk immediately.
        self._reset_state(pre_chunk.overlap_tail)
        yield pre_chunk
...
    def _reset_state(self, overlap_prefix: str) -> None:
        """Set working-state values back to "empty", ready to accumulate next pre-chunk."""
        self._overlap_prefix = overlap_prefix
        self._elements.clear()
        self._text_segments = [overlap_prefix] if overlap_prefix else []
        self._text_len = len(overlap_prefix)

        if len(self._elements) == 1 and isinstance(self._elements[0], Table):
            yield from _TableChunker.iter_chunks(
                self._elements[0], self._overlap_prefix, self._opts
            )
        else:
            yield from _Chunker.iter_chunks(self._elements, self._text, self._opts)
...
        if self._overlap_prefix:
            yield self._overlap_prefix
        for e in self._elements:
            if e.text and len(e.text):
                # -- preserve all whitespace for code snippets to maintain formatting --
                if isinstance(e, CodeSnippet):
                    yield e.text

    @cached_property
    def _text_with_overlap(self) -> str:
        """The text for this chunk, including the overlap-prefix when present."""
        overlap_prefix = self._overlap_prefix
        table_text = "" if not self._table.text else self._table.text.strip()
        # -- use row-separator between overlap and table-text --
        return overlap_prefix + "\n" + table_text if overlap_prefix else table_text

Missing High-Value Tests

• Add an end-to-end chunk_elements() test for text -> table with overlap_all=True asserting the Table text contains only table content.
• Add the symmetric table -> text test with overlap_all=True asserting the following CompositeElement does not start with table text.
• Add the same coverage for chunk_by_title(), since it shares the same PreChunkBuilder/PreChunk overlap path.
• The new test_table_isolation.py suite passes, but it never exercises overlap or overlap_all, so it would not catch this regression.

Verdict

Not Merge. I did not find other Critical/High issues with high confidence, but this one is a real correctness gap in the PR’s stated isolation guarantee.