test_scripts.py

from __future__ import annotations

import pytest

from app.errors import LanguageUnsupportedError
from app.scripts import transcript_matches_language
from app.service import _require_matching_script

HINDI_LANGUAGE_CODE = "hi"
ENGLISH_LANGUAGE_CODE = "en"


def test_the_dolphin_cyrillic_bug_is_rejected() -> None:
    """The exact failure this guard exists for: dictating "नमस्ते" in Hindi to
    Dolphin returns "насте" — fluent, confident, and in the wrong alphabet."""
    assert transcript_matches_language("насте", HINDI_LANGUAGE_CODE) is False
    with pytest.raises(LanguageUnsupportedError, match="different language"):
        _require_matching_script("насте", HINDI_LANGUAGE_CODE)


def test_transliteration_into_another_scrip_aa() -> None:
    """A model under evaluation rendered "send me the report by Friday" in Arabic
    script. A presence test let it through on the single stray Latin "o" in
    "رoرت", which is why the check is proportional rather than binary."""
    transliterated = "ل سند مي  رoرت بي فريداي"
    assert "o" in transliterated  # the character that used to rescue it
    assert transcript_matches_language(transliterated, ENGLISH_LANGUAGE_CODE) is False
    with pytest.raises(LanguageUnsupportedError):
        _require_matching_script(transliterated, ENGLISH_LANGUAGE_CODE)


def test_code_switching_is_never_rejected() -> None:
    """Indian speakers routinely mix English into Hindi, so the threshold has to
    sit below the share of the base script even heavy Hinglish keeps."""
    # Including heavy Hinglish, which is where a proportional check could bite.
    for text in (
        "मैं office जा रहा हूँ",
        "मैं ठीक हूँ",
        "kal मैं आऊंगा",
        "मैं report Friday तक send करूंगा",
        "ठीक है OK",
    ):
        assert transcript_matches_language(text, HINDI_LANGUAGE_CODE) is True
        _require_matching_script(text, HINDI_LANGUAGE_CODE)


def test_uncertain_cases_pass_rather_than_fail() -> None:
    """Anything the check cannot judge must be let through: a guard that guesses
    would reject good transcripts."""
    assert transcript_matches_language("anything at all", "auto") is True
    assert transcript_matches_language("anything at all", "zz") is True  # unknown code
    assert transcript_matches_language("", HINDI_LANGUAGE_CODE) is True
    assert transcript_matches_language("1,200 $45 —", HINDI_LANGUAGE_CODE) is True  # no letters
    # Chinese and Japanese share Han characters, so CJK is deliberately lenient.
    assert transcript_matches_language("私は元気です", "zh") is True


def test_roman_hinglish_uses_the_latin_script_contract() -> None:
    assert transcript_matches_language("Aaj mujhe office jaana hai", "hinglish_roman") is True
    assert transcript_matches_language("आज मुझे जाना है", "hinglish_roman") is False
    assert transcript_matches_language("Aaj office में meeting hai", "hinglish_roman") is False


@pytest.mark.parametrize(
    ("text", "language"),
    [
        ("मैं ठीक हूँ", ENGLISH_LANGUAGE_CODE),
        ("hello there", HINDI_LANGUAGE_CODE),
        ("आज", "bn"),
        ("The quick brown fox", "ru"),
        ("مرحبا", HINDI_LANGUAGE_CODE),
    ],
)
def test_wrong_script_is_rejected(text: str, language: str) -> None:
    assert transcript_matches_language(text, language) is False


@pytest.mark.parametrize(
    ("text", "language"),
    [
        ("मैं कल बाजार जाऊंगा", HINDI_LANGUAGE_CODE),
        ("আমি ভালো আছি", "bn"),
        ("நான் நன்றாக இருக்கிறேன்", "ta"),
        ("私は元気です", "ja"),
        ("집에 갔어요", "ko"),
        ("ผมสบายดี", "th"),
        ("The quick brown fox", ENGLISH_LANGUAGE_CODE),
        ("Здравствуйте", "ru"),
        ("مان ٺيڪ آهيان", "sd"),
    ],
)
def test_matching_script_passes(text: str, language: str) -> None:
    assert transcript_matches_language(text, language) is True
    _require_matching_script(text, language)


def test_every_client_language_can_be_judged() -> None:
    """A language the clients offer but the table does not know would silently
    skip the guard, which is how this class of bug goes unnoticed."""
    from app.scripts import expected_scripts

    client_languages = [
        ENGLISH_LANGUAGE_CODE, "es", "ar", "ja", "ko", "zh", "uk", "ru", "vi", "fr", "de", "it",
        "pt", "nl", "pl", HINDI_LANGUAGE_CODE, "bn", "mr", "te", "ta", "gu", "ur", "kn", "ml",
        "pa", "as", "ne",
    ]  # fmt: skip
    unjudgeable = [code for code in client_languages if not expected_scripts(code)]
    assert not unjudgeable, f"no expected script for: {unjudgeable}"


@pytest.mark.parametrize(
    ("text", "language", "script"),
    [
        ("Здраво свете", "sr", "Serbian Cyrillic — official in Serbia"),
        ("Zdravo svete", "sr", "Serbian Latin — everyday use"),
        ("Салам дүнја", "az", "Azerbaijani Cyrillic"),
        ("Salam dünya", "az", "Azerbaijani Latin"),
        ("Салом дунё", "uz", "Uzbek Cyrillic"),
        ("Salom dunyo", "uz", "Uzbek Latin"),
        ("ਸਤ ਸ੍ਰੀ ਅਕਾਲ", "pa", "Punjabi Gurmukhi — India"),
        ("السلام علیکم", "pa", "Punjabi Shahmukhi — Pakistan"),
        ("Сәлеметсіз бе", "kk", "Kazakh Cyrillic"),
    ],
)
def test_languages_written_in_two_scripts_a_aaa(text: str, language: str, script: str) -> None:
    """A guard that rejects transcripts must not be wrong about a language's
    writing system. Holding Serbian to one alphabet would throw away half the
    country's writing, and a false rejection destroys a good dictation."""
    assert transcript_matches_language(text, language), script