Add offline spellchecking for visible interview text

[nonprofittechy]

Summary

Adds a local spellcheck pass for the visible text interview users see: questions, subquestions, field labels, help, choice labels, and template content. It skips code, stored values, Mako, Markdown/HTML markup, and URLs.

This brings DAYamlChecker closer to parity with the original linter script, which previously ran spelling checks.

• Stable spelling rule codes: SP701 (spelling_possible_typo) and SP702 (spelling_common_legal_typo, such as judgement → judgment and HIPPA → HIPAA) under a new spelling finding class. The codes remain SP701/SP702 at every configured severity, so suppressions stay stable.
• On by default as warnings, using US English. Turn spellchecking off with --no-spellcheck or RuntimeOptions(spellcheck=None).
• Configurable severity: --spellcheck-severity info|warning|error changes the reported level and exit behavior without changing the rule code. This lets new lint rules begin as warnings and become enforceable when a project is ready.
• Languages: en, es, ru, and sv are built in. Custom Hunspell dictionaries can be added with --spellcheck-dictionary LANG=PATH. Blocks with a declared language outside the selected languages are skipped.
• Configuration: supports --spellcheck-language, --spellcheck-wordlist, --spellcheck-ignore-word, and --spellcheck-severity, plus SpellcheckOptions in Python. Findings can be suppressed with # no-dayc comments by code, message ID, or class.
• Designed for few false positives: skips acronyms, mixed-case words, short words, unfamiliar proper names, author names declared in metadata, names in credits and copyright sections, and geographic choice lists.
• New public API: exports find_spelling_findings_from_string and SpellcheckOptions. scripts/evaluate_spelling.py compares results on a corpus of interviews.

Licensing and dictionaries

Uses spylls (MIT; pure-Python Hunspell) as a dependency. Its English, Russian, and Swedish dictionaries require no network access.

The Spanish dictionary (RLA-ES, via LibreOffice) is not shipped, so the package remains plain MIT. The first run selecting es downloads es_US.aff and .dic (about 850 KB) from a pinned LibreOffice commit and verifies their SHA-256 hashes. Files are cached under $DAYAMLCHECKER_CACHE_DIR when configured; otherwise the platform user-cache directory is used ($XDG_CACHE_HOME/dayamlchecker, normally ~/.cache/dayamlchecker, or %LOCALAPPDATA%\dayamlchecker\cache on Windows). Download failures point users to --spellcheck-dictionary es=PATH for offline use.

Notes for reviewers

• Existing CI that uses --max-warnings 0 may begin failing on spelling warnings because spellchecking is enabled by default.
• Per-run cost is mostly loading the pure-Python dictionary: about 0.4 seconds for English, plus about 0.9 seconds when Spanish is enabled.

Testing

• uv run pytest -q tests: 748 passed, 56 subtests passed.
• Download, cache reuse, corrupted-cache recovery, hash mismatch, and offline failures are tested with a fake remote. Spanish tests use the real dictionary through pytest's cache and skip when offline except under CI.
• A clean wheel was checked to confirm that it contains no Spanish dictionary files and declares License-Expression: MIT.