Add HTML stripping for NodeBB integration and CI workflow

[anthony-tom1]

Context

The Ollama-backed translator works end-to-end when the model is running, but NodeBB stores post bodies as HTML (e.g. <p>Bonjour</p>). Sending that markup in the prompt adds noise and can hurt language detection and translation. Separately, small chat models often reply with slightly off-format lines (markdown like **LANGUAGE:**, numbered lists, or Language: instead of LANGUAGE:). The previous parser only recognized lines that started with exactly LANGUAGE: / TRANSLATION:, so a valid answer could be misread as “no language line” and collapsed to English with the original text. The repo also lacked a CI workflow to run tests on every push/PR.

Description

translate_content strips HTML tags before calling query_llm_robust, so the model sees plain text. Response parsing is more tolerant of common formatting variants, and the HTTP API now includes a language field (string or JSON null) parsed from the model when possible, so clients like NodeBB can show a real detected language instead of only English vs Unknown. A GitHub Actions workflow runs pytest on pushes and pull requests targeting f25 so regressions are caught automatically.

Changes in the codebase

• src/translator.py — Added _strip_html() (regex tag removal) and use it in translate_content(). Added _normalize_response_line() and updated _parse_model_content() to accept markdown/list prefixes and case-insensitive language: / translation: labels; return type now includes detected language where available. Error paths return (is_english, text, language=None) consistently.
• src/api.py — TranslateResponse includes optional language; GET / passes it through from translate_content.
• test/unit/test_translator.py — test_translate_content_strips_html (HTML stripped before the LLM call). New parser cases for **LANGUAGE:**, 1. LANGUAGE:, and Language: / Translation: casing. Assertions updated for the 3-tuple return and language on successful parses.
• .github/workflows/ci.yml (new) — On push/PR to f25: checkout, Python 3.12, pip install -r requirements.txt, pytest test/ -v.

How this was tested

• pytest test/ -v locally — all tests pass (currently 17), including HTML stripping and the new parser scenarios.
• Manual (optional): run uvicorn on port 5001, GET /?content=... with URL-encoded <p>Bonjour</p> and confirm behavior with Ollama running.

How to verify

• CI should go green on this PR once the workflow runs.
• JSON shape: GET /?content=<text> returns is_english, translated_content, and language. language is a string when a LANGUAGE: value is parsed from the model output, and null otherwise (e.g. parse failure, Ollama errors, or missing/invalid format). Clients that only use is_english and translated_content remain compatible.
• HTML-wrapped input from NodeBB should yield cleaner prompts and more reliable detection/translation than sending raw tags.
• Quick check (with Ollama up):

curl "http://127.0.0.1:5001/?content=%3Cp%3EBonjour%3C%2Fp%3E"