From 4abb10df15d5298f3c07fa26c0cbc41df7cd0dfc Mon Sep 17 00:00:00 2001
From: DazedAnon <dazedanon@gmail.com>
Date: Thu, 9 Jul 2026 16:20:19 -0500
Subject: [PATCH] Pass 1 should TL Speakers

---
 modules/wolfdawn.py    | 35 +++++++++++++++++++++-----------
 tests/test_wolfdawn.py | 46 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 69 insertions(+), 12 deletions(-)

diff --git a/modules/wolfdawn.py b/modules/wolfdawn.py
index 6a449b3..efb2879 100644
--- a/modules/wolfdawn.py
+++ b/modules/wolfdawn.py
@@ -412,28 +412,37 @@ def _body_after_dropped_tag(text: str, original_speaker: str) -> str:
     return text
 
 
-def _maybe_fix_nameplate(entry, total_tokens: list) -> None:
-    """Swap a Japanese first-line nameplate for English on an otherwise-done line."""
-    if ESTIMATE or _batch_phase() == "collect":
-        return
+def _maybe_fix_nameplate(entry, total_tokens: list) -> bool:
+    """Resolve a Japanese first-line nameplate on an otherwise-done line.
+
+    Returns True when the nameplate was rewritten. During batch collect the
+    in-memory rewrite is discarded (translated/ is not written), but
+    ``getSpeaker`` still runs so the English gloss is cached before dialogue
+    batches are queued - otherwise those live speaker calls only show up in
+    Pass 2.
+    """
+    if ESTIMATE:
+        return False
     speaker_src = entry.get("speaker_src", "")
     if not wolf_speakers.is_firstline_enabled(speaker_src, SPEAKER_CONFIG):
-        return
+        return False
     txt = entry.get("text")
     if not isinstance(txt, str) or not txt.strip():
-        return
+        return False
     # Only touch finished dialogue (body already English / skip-translated).
     if _text_still_needs_translation(entry):
-        return
+        return False
     split = wolf_speakers.split_source(txt, speaker_src, SPEAKER_CONFIG)
     if split is None:
-        return
+        return False
     prefix, speaker, body = split
     if not re.search(LANGREGEX, speaker):
-        return
+        return False
     speaker_en = _resolve_nameplate(speaker, total_tokens)
     if speaker_en and speaker_en != speaker:
         entry["text"] = wolf_speakers.restore_source(prefix, speaker_en, body)
+        return True
+    return False
 
 
 def parseDocument(data, filename):
@@ -457,9 +466,11 @@ def parseDocument(data, filename):
             return False
         return True
 
-    # Fix Japanese nameplates on lines whose dialogue body is already done, so
-    # resume / re-wrap runs do not leave ``セルリア\\nPlease hold on...`` behind.
-    if IGNORETLTEXT and not ESTIMATE and _batch_phase() != "collect":
+    # Resolve Japanese nameplates on lines whose dialogue body is already done.
+    # Must run during batch collect too: those lines are skip-translated so they
+    # never hit the reshape/getSpeaker path, and deferring to consume is what
+    # made speaker Input/Output spam show up in Pass 2 after the batch returned.
+    if IGNORETLTEXT and not ESTIMATE:
         for entry in entries:
             if _translatable(entry):
                 continue
diff --git a/tests/test_wolfdawn.py b/tests/test_wolfdawn.py
index c02e5a0..22795cc 100644
--- a/tests/test_wolfdawn.py
+++ b/tests/test_wolfdawn.py
@@ -470,6 +470,52 @@ class TestTranslationWriteback(unittest.TestCase):
         # Short-string speaker resolve, then the remaining dialogue batch.
         self.assertEqual(captured, ["司祭", ["まだだ"]])
 
+    def test_collect_resolves_japanese_nameplate_before_queueing(self):
+        """Pass 1 must live-translate skip-translated JP nameplates (not defer to Pass 2)."""
+        doc = {
+            "kind": "map",
+            "scenes": [
+                {
+                    "event": 1,
+                    "name": "ev",
+                    "lines": [
+                        {
+                            "cmd": 0,
+                            "str": 0,
+                            "speaker": "客Ｆ",
+                            "speaker_src": "literal_line1_lowconf",
+                            "source": "客Ｆ\nまぁまぁ。",
+                            "text": "客Ｆ\nNow, now.",
+                        },
+                        {
+                            "cmd": 1,
+                            "str": 0,
+                            "speaker": "UI",
+                            "speaker_src": "ui",
+                            "source": "まだだ",
+                            "text": "まだだ",
+                        },
+                    ],
+                }
+            ],
+        }
+        orig_phase = os.environ.get("BATCH_PHASE")
+        os.environ["BATCH_PHASE"] = "collect"
+        try:
+            (data, _t, err), captured = _WolfTranslateHarness().run(doc, "collect_nameplate.mps.json")
+        finally:
+            if orig_phase is None:
+                os.environ.pop("BATCH_PHASE", None)
+            else:
+                os.environ["BATCH_PHASE"] = orig_phase
+
+        self.assertIsNone(err)
+        # Speaker resolved live during collect; dialogue list is also seen (queued).
+        self.assertEqual(captured[0], "客Ｆ")
+        self.assertEqual(captured[1], ["まだだ"])
+        # Collect does not persist translated/, but in-memory text is updated.
+        self.assertEqual(data["scenes"][0]["lines"][0]["text"], "En_客Ｆ\nNow, now.")
+
     def test_ignore_tl_text_false_retranslates(self):
         doc = {
             "kind": "map",