Summary

• Keep field brace-depth accounting stable when an opening brace is escaped.
• Preserve every field following the escaped-brace sequence.
• Add a public parse/write/reparse regression so a shallow successful parse
  cannot conceal later-field loss.

Follow-up to #452. The earlier report was closed after discussion of enclosing
middleware, but a stricter reproduction shows that current main silently
drops all fields following the reported abstract sequence.

Validation

• Complete enclosing regression matrix under warnings-as-errors: 917 passed, 10
  skipped.
• Complete upstream suite: 2,577 passed, 12 skipped, with the four existing
  deprecation warnings in tests/test_entrypoint.py.
• git diff --check: passed.
• The project pre-commit executable was not available in the isolated local
  environment; GitHub CI should independently run the configured hooks.

Review note

This change was developed and validated in a concentrated session rather than
exercised over a long period in production. Because the splitter previously
returned a normal entry while losing later fields, careful human review of the
brace-state change and regression fixture is requested.

AI assistance

This pull request was prepared with ChatGPT Codex using GPT-5.6 Sol with high
reasoning effort. Codex assisted with analysis, implementation, branch
isolation, and test execution. Automated validation is not a substitute for
maintainer review.