fix: use deepcopy instead of shallow copy when stripping orig_elements

[chuenchen309]

Problem

_Chunker._orig_elements copies each element before clearing its metadata.orig_elements, with an explicit comment stating this is because "these elements don't belong to us (the user may have downstream purposes for them)." But copy.copy() is a shallow copy — the copy's .metadata is the same object as the original's .metadata — so orig_element.metadata.orig_elements = None mutates the caller's own element in place, exactly contradicting the stated intent.

Sibling _TableChunker._orig_elements already does this correctly with copy.deepcopy(self._table).

Fix

Use copy.deepcopy instead of copy.copy, matching the already-correct sibling implementation.

Testing

• Added a regression test confirming a caller's element still has its original .metadata.orig_elements value after _Chunker._orig_elements is computed.
• Confirmed red→green: reverting only base.py reproduces the caller mutation (assert None == [inner]); reapplying passes.
• Full test_unstructured/chunking/ suite: 334 passed.
• ruff check — clean.
• xref: no open PR touches this shallow/deep-copy site (checked fix(chunking): prevent infinite loop in token-based overlap splitting #4392, feat(pdf): flag invisible text in metadata #4383, feat: add max_page to chunk_by_title and remove multipage_sections #4382, feat: chunking by character and title now isolates tables #4197 — none overlap; feat: add max_page to chunk_by_title and remove multipage_sections #4382/feat: chunking by character and title now isolates tables #4197 have unrelated include_orig_elements matches, a different parameter name).

AI-Generated disclosure

I used an AI coding assistant (Claude) to help identify this bug and draft the fix. I independently reproduced the caller mutation with a throwaway script before writing the test, cross-checked against the already-correct sibling _TableChunker implementation, and ran the full local test suite before preparing this change.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com

[cubic-dev-ai]

No issues found across 4 files

<sub>Shadow auto-approve: would auto-approve. Fixes a shallow-copy bug that mutated the caller's element metadata. The change is bounded to one line, matches the already-correct sibling implementation, and adds a regression test. No operational, security, or product tradeoffs.

Re-trigger cubic</sub>

[zeweihan]

@/tmp/gh_comment_unstructured_4402.md

[chuenchen309]

Heads up — your last comment came through as just a literal @/tmp/gh_comment_unstructured_4402.md path, so the intended content didn't attach. Happy to take a look if you re-post it.