[Dodothereal]

Summary

Prevent an escaped physical newline inside a quoted string from being mistaken for a backslash continuation between tokens.

On current master, even a file without docstrings can be corrupted:

print("first\
second")

Formatting duplicates the opening print( and produces invalid Python. Assignments can become duplicated chained assignments, with more copies added on every pass.

Use the existing original-row comparison to distinguish a string starting on the previous token's row from a genuine inter-token continuation. Existing explicit-continuation behavior is retained.

Add 52 regression cases covering ordinary quote styles, raw/bytes/Unicode prefixes, calls and returns, repeated continuations, concatenation, missing final newlines, parenthesized strings, shifted rows after docstring formatting, LF/CRLF, and default/--black modes. They check exact output, AST equivalence, and second-pass idempotence.

This was found during source inspection and synthetic testing, not from a production incident. It is distinct from #406's docstring-classification guard and from the already merged #378/#387 fixes.

Validation

Against upstream 82069d7f1e105bd86c69fd9ae0587ca00da3c312, on Linux / CPython 3.12.14:

• Unchanged full suite: 686 passed, 3 skipped.
• New tests on unchanged production code: 48 failures, 4 passing controls.
• Fixed focused regressions: 52 passed.
• Fixed full suite, including CLI tests: 738 passed, 3 skipped.
• Pinned Black 26.3.1 and Ruff 0.15.11 pass both changed files. Black used explicit Python 3.10–3.12 targets under Python 3.12.
• Independent Codex review checked the exact diff and hashes and repeated the full suite and quality checks.
• Patch application and whitespace checks pass.

The complete tox/pre-commit aggregate and other interpreters/platforms' full test suites were not run locally. All seven hosted checks pass for b65883028ffa171b8f35a001993d177d16a04598.

AI assistance

Prepared with OpenAI Codex assistance for diagnosis, implementation, tests, and this description. A separate Codex agent independently reviewed and tested the exact final patch, with no actionable findings. The additional review was automated.