[T3371] FIX: don't queue B2S letters when the language detector has no opinion

advanced_translation/models/langdetect.py

@@ -13,14 +13,17 @@ class OSD(models.AbstractModel):
 
     default_threshold = 0.95
     languages_langdetect = ["en", "de", "fr", "it", "es"]
+    # Below this many characters langdetect is unreliable: it returns unstable
+    # results across runs and confidently wrong ones ("Amen" is detected as Dutch).
+    min_length = 50
 
     def detect_language(self, text, threshold=None):
         if threshold is None:
             threshold = self.default_threshold
 
         language = self.env["res.lang.compassion"]
 
-        if not isinstance(text, str) or len(text) < 50:
+        if not isinstance(text, str) or len(text) < self.min_length:
             return language
 
         try:

partner_communication_compassion/models/correspondence.py

@@ -249,13 +249,5 @@ def _can_auto_send(self):
             and "Final Letter" not in types
             and "auto" in self.partner_id.letter_delivery_preference
         )
-        # Only auto-send if the letter's actual content language is one the
-        # sponsor reads (detected from content, not the field-office stamp).
-        detected_lang = self._detect_letter_language()
-        valid = (
-            valid
-            and bool(detected_lang)
-            and detected_lang in self.supporter_languages_ids
-        )
-
-        return valid
+        # Don't auto-send a letter the sponsor cannot read.
+        return valid and self._sponsor_can_read_letter()

sbc_compassion/models/correspondence.py

@@ -459,16 +459,7 @@ def _check_translation_language(self):
         for letter in self.with_context(skip_lang_detect=True):
             # Determine which text is analyzed
             is_translation = bool(letter.translated_text or letter.english_text)
-            letter_text = (
-                letter.translated_text or letter.english_text or letter.original_text
-            )
-            # Clean text for accurate detection
-            clean_text = (
-                letter_text.strip(" \t\n\r.")
-                .replace(BOX_SEPARATOR, "")
-                .replace(PAGE_SEPARATOR, "")
-                .strip()
-            )
+            clean_text = letter._clean_letter_text()
 
             if not clean_text:
                 # T2495 Default to English for empty B2S letters
@@ -492,17 +483,48 @@ def _check_translation_language(self):
                 ):
                     letter.original_language_id = detected_lang
 
-    def _detect_letter_language(self):
-        """Language the letter is actually written in. Detected from its text."""
+    def _clean_letter_text(self):
+        """Text of the letter, stripped of separators, ready for detection."""
         self.ensure_one()
         text = self.translated_text or self.english_text or self.original_text or ""
-        clean = (
+        return (
             text.strip(" \t\n\r.")
             .replace(BOX_SEPARATOR, "")
             .replace(PAGE_SEPARATOR, "")
             .strip()
         )
-        return self.env["langdetect"].detect_language(clean)
+
+    def _letter_language_verdict(self):
+        """Detected language of the letter, and whether we have an opinion at all.
+
+        has_opinion is False only when the text is shorter than
+        langdetect.min_length; the caller then falls back to the field-office
+        stamp (T3371). Anything longer counts as an opinion, so a letter we
+        cannot read is still queued for translation (T3339).
+        """
+        self.ensure_one()
+        clean = self._clean_letter_text()
+        lang_detector = self.env["langdetect"]
+        return (
+            lang_detector.detect_language(clean),
+            len(clean) >= lang_detector.min_length,
+        )
+
+    def _sponsor_can_read_letter(self):
+        """True if the sponsor reads the language the letter is actually in.
+
+        Judged from the content when the letter holds enough text, since the
+        field-office stamp is not always right (T3339). Below that the detector
+        has no opinion, so the stamp is all we have, and trusting it beats
+        queueing letters nobody needs translated (T3371).
+        """
+        self.ensure_one()
+        language, has_opinion = self._letter_language_verdict()
+        if has_opinion:
+            return language in self.supporter_languages_ids
+        return bool(self.beneficiary_language_ids & self.supporter_languages_ids) or (
+            self.translation_language_id in self.supporter_languages_ids
+        )
 
     @api.depends("uuid")
     def _compute_read_url(self):

sbc_compassion/tests/__init__.py

@@ -9,3 +9,4 @@
 ##############################################################################
 
 from . import test_sbc_compassion
+from . import test_letter_language_verdict

sbc_compassion/tests/test_letter_language_verdict.py

@@ -0,0 +1,72 @@
+##############################################################################
+#
+#    Copyright (C) 2026 Compassion CH (http://www.compassion.ch)
+#    Releasing children from poverty in Jesus' name
+#
+#    The licence is in the file __manifest__.py
+#
+##############################################################################
+
+from langdetect import DetectorFactory
+
+from odoo.tests import TransactionCase
+
+SHORT_CAPTION = "Salut, C'est Martial. Juste un bonjour!"
+LONG_ENGLISH = (
+    "Dear sponsor, thank you very much for your letter and for the gift you "
+    "sent me. I am doing well at school. God bless you."
+)
+LONG_SWAHILI = (
+    "Habari yako mlezi wangu, asante sana kwa barua yako na zawadi "
+    "uliyonitumia. Mimi ni mzima wa afya. Mungu akubariki sana."
+)
+
+
+class TestLetterLanguageVerdict(TransactionCase):
+    """Pins the verdict the B2S translation gate is built on (T3339 / T3371).
+
+    Covers `_letter_language_verdict` only. `_sponsor_can_read_letter` needs a
+    persisted letter, because `supporter_languages_ids` is related to
+    `partner_id.spoken_lang_ids` and comes back NewId-wrapped on an in-memory
+    record, and building one needs `BaseSponsorshipTest`, whose `setUpClass` is
+    currently broken for all 13 test classes that use it.
+    """
+
+    def setUp(self):
+        super().setUp()
+        # langdetect samples randomly and runs unseeded in production, so the
+        # same text occasionally misses the confidence threshold (~0.1% of runs
+        # for the English text below). Pin the seed here so these tests exercise
+        # verdict handling rather than detector luck.
+        previous_seed = DetectorFactory.seed
+        DetectorFactory.seed = 0
+        self.addCleanup(setattr, DetectorFactory, "seed", previous_seed)
+
+    def _letter(self, text, field="translated_text"):
+        return self.env["correspondence"].new(
+            {"page_ids": [(0, 0, {"paragraph_ids": [(0, 0, {field: text})]})]}
+        )
+
+    def test_short_text_gives_no_opinion(self):
+        """Photo captions are too short to judge: the gate must fall back."""
+        language, has_opinion = self._letter(SHORT_CAPTION)._letter_language_verdict()
+        self.assertFalse(language)
+        self.assertFalse(has_opinion)
+
+    def test_long_text_is_detected(self):
+        language, has_opinion = self._letter(LONG_ENGLISH)._letter_language_verdict()
+        self.assertTrue(has_opinion)
+        self.assertEqual(
+            language, self.env.ref("advanced_translation.lang_compassion_english")
+        )
+
+    def test_long_untranslatable_text_still_gives_an_opinion(self):
+        """A language we do not handle is an answer, not an absence of one.
+
+        Without this the letter falls back to the field-office stamp and can
+        reach the sponsor untranslated (T3339).
+        """
+        letter = self._letter(LONG_SWAHILI, field="original_text")
+        language, has_opinion = letter._letter_language_verdict()
+        self.assertFalse(language)
+        self.assertTrue(has_opinion)

sbc_translation/models/correspondence.py

@@ -225,13 +225,8 @@ def process_letter(self):
             if not force_publish:
                 letter._check_translation_language()
 
-            # Can sponser read the letter? Decide from the letter's ACTUAL
-            # content language, not the field-office TranslationLanguage stamp.
-            # If it can't be determined, fail safe to translation.
-            detected_lang = letter._detect_letter_language()
-            langs_match = (
-                bool(detected_lang) and detected_lang in letter.supporter_languages_ids
-            )
+            # Can sponser read the letter?
+            langs_match = letter._sponsor_can_read_letter()
 
             # Is the letter still in the translation process?
             translation_hold = (