fix: repair split guillemet dialogue

CHANGELOG.md

@@ -59,6 +59,10 @@ Follows [Keep a Changelog](https://keepachangelog.com/en/1.0.0/) · versioned wi
   object URLs, reducing retained browser memory during repeated auditioning.
 - **Character Library Smoke Coverage.** Added deterministic browser coverage for opening Characters / Cast, profile
   sorting, column visibility, and assigned-voice rating persistence.
+- **Attribution Quote Repair.** Mixed German guillemet pairs are now handled by the deterministic fallback, and
+  normalized LLM segments are repaired when an epigraph is split before its closing quote.
+- **Attribution Regression Coverage.** Added focused tests for mixed guillemets and the Schwester Larissa epigraph
+  split regression.
 - **Security Regression Coverage.** Added a Python test gate for character metadata validation, private-download URL
   rejection and oversized JSON request rejection.
 - **Navigation Reliability.** Rehearser phase requests now survive the feature-module initialization boundary.

TODO.md

@@ -46,8 +46,8 @@ the first block before investing in architecture or polish.
   The remaining segmentation work is upstream splitting/fusion, not another merge pass.
 - [ ] **S — Finish the remaining Playwright smoke checks.** Characters/Cast, profile
   sorting, column visibility, voice ratings, and starting/cancelling a generation job.
-- [ ] **M — Fix the Schwester Larissa unclosed-guillemet split.** Add a regression test
+- [x] **M — Fix the Schwester Larissa unclosed-guillemet split.** Attribution normalization now repairs dialogue
-  for the epigraph and verify the parser does not split mid-sentence.
+  split before mixed German guillemet closes (`»…‹` / `«…›`), with an epigraph regression test.
 - [ ] **M — Make profile portraits and character sheets use one canonical identity.**
   Shared structured prompts now version consistent pairs as v3 and fingerprint canonical appearance fields; add
   cross-record validation so mismatched profile/reference files are detected even when both are current-version.
@@ -139,9 +139,9 @@ the first block before investing in architecture or polish.
 
 ## P0 — Data integrity and recovery
 
-- [ ] Validate character sheets and voice metadata with schemas.
+- [x] Validate character sheets and voice metadata with schemas.
-- [ ] Add database migrations instead of silently changing record shapes.
+- [x] Add database migrations instead of silently changing record shapes.
-- [ ] Use atomic writes for `.meta.json` files.
+- [x] Use atomic writes for `.meta.json` files.
 - [x] Create automatic backups before bulk deletes, voice moves, and recasts.
 - [x] Store prompt versions and generation history so bad regenerations can be audited and follow-up recovery can be built.
 
@@ -152,7 +152,7 @@ the first block before investing in architecture or polish.
 - [ ] Character parsing and phantom-speaker prevention.
 - [ ] Voice-matching constraints.
 - [ ] Profile-completeness counting.
-- [ ] Rating persistence.
+- [x] Rating persistence.
 - [ ] STT endpoint normalization.
 - [ ] ComfyUI workflow submission.
 - [ ] API failure and timeout behavior.
@@ -223,7 +223,8 @@ OCR/wording differences. Attribution work cannot fix these.
   separate is contributing.
 - The merge was predicted (on run 2 data) to cut unmatched 727 → 588. On run 3 it
   went to 750. **That prediction did not hold** and needs re-deriving on run 3.
-- Known unfixed: the Schwester Larissa epigraph still splits mid-sentence
+- Fixed: attribution normalization repairs the Schwester Larissa epigraph when the LLM splits it before a mixed
+  guillemet close; the regression is covered by `tests/test_conversation.py`.
   (unclosed-guillemet handling).
 
 ### Next steps, highest value first

routes/conversation.py

@@ -261,7 +261,7 @@ def _post_llm_chat_completion(llm_url: str, payload: dict, headers: dict, timeou
 def _fallback_attribute_response(text: str) -> dict:
     """Deterministic fallback for attribution when an upstream LLM fails."""
     quote_re = re.compile(
-        r"»([^«]+)«|«([^»]+)»|„([^“”]+)[“”]|“([^”]+)”|\"([^\"]+)\"|「([^」]+)」|『([^』]+)』",
+        r"»([^«‹]+)[«‹]|«([^»›]+)[»›]|„([^“”]+)[“”]|“([^”]+)”|\"([^\"]+)\"|「([^」]+)」|『([^』]+)』",
         re.DOTALL,
     )
     segments = []
@@ -288,6 +288,54 @@ def _fallback_attribute_response(text: str) -> dict:
     return {"segments": segments, "characters": chars, "fallback": True}
 
 
+def _repair_unclosed_dialogue_segments(segments: list) -> list:
+    """Join dialogue with a following narration fragment split before its quote close.
+
+    OCR and LLM output occasionally mix German guillemet pairs (``»…‹`` and
+    ``«…›``).  When that happens, a long epigraph can be split at an arbitrary
+    chunk boundary.  Only repair a dialogue segment whose opening quote is
+    still unmatched, and only consume the following narration through the
+    corresponding close marker; unrelated narration remains untouched.
+    """
+    pairs = (("»", ("«", "‹")), ("«", ("»", "›")), ("„", ("“", "”")), ("“", ("”",)), ("「", ("」",)), ("『", ("』",)))
+    repaired = [dict(seg) if isinstance(seg, dict) else seg for seg in segments]
+    index = 0
+    while index < len(repaired):
+        current = repaired[index]
+        if not isinstance(current, dict) or str(current.get("type") or "").lower() != "dialogue":
+            index += 1
+            continue
+        text = str(current.get("text") or "")
+        pair = next(((open_mark, closes) for open_mark, closes in pairs if open_mark in text), None)
+        if not pair:
+            index += 1
+            continue
+        open_mark, close_marks = pair
+        opening = text.find(open_mark)
+        if any(mark in text[opening + 1:] for mark in close_marks):
+            index += 1
+            continue
+        for following in range(index + 1, len(repaired)):
+            candidate = repaired[following]
+            if not isinstance(candidate, dict):
+                continue
+            if str(candidate.get("type") or "").lower() != "narration":
+                break
+            candidate_text = str(candidate.get("text") or "")
+            close_at = min((candidate_text.find(mark) for mark in close_marks if candidate_text.find(mark) >= 0), default=-1)
+            if close_at < 0:
+                continue
+            current["text"] = text + " " + candidate_text[:close_at + 1]
+            remainder = candidate_text[close_at + 1:].strip()
+            if remainder:
+                candidate["text"] = remainder
+            else:
+                repaired.pop(following)
+            break
+        index += 1
+    return repaired
+
+
 def _resolve_speak_voice(settings: dict, client_id: str, explicit_voice: str) -> str:
     if explicit_voice:
         return explicit_voice
@@ -3695,6 +3743,8 @@ def _attribution_parse(raw: str, text: str) -> dict:
         # Do NOT try to parse its reasoning as a script. Abort and preserve the book text.
         segments = [{"speaker": "Narrator", "type": "narration", "text": text, "emotion": ""}]
 
+    segments = _repair_unclosed_dialogue_segments(segments)
+
     # Normalise + collect speaker roster
     clean, chars = [], []
     for seg in segments:

tests/test_conversation.py

@@ -0,0 +1,34 @@
+"""Focused attribution parser regressions."""
+from __future__ import annotations
+
+import os
+import tempfile
+import unittest
+
+os.environ.setdefault("CONFIG_DIR", tempfile.mkdtemp(prefix="ttvc-conversation-test-config-"))
+
+from routes.conversation import _attribution_parse, _fallback_attribute_response
+
+
+class AttributionParserTests(unittest.TestCase):
+    def test_fallback_accepts_mixed_guillemets(self):
+        result = _fallback_attribute_response("Intro. »Ein langer Satz endet hier.‹ Danach.")
+        self.assertEqual([segment["type"] for segment in result["segments"]], ["narration", "dialogue", "narration"])
+        self.assertIn("Ein langer Satz endet hier.", result["segments"][1]["text"])
+
+    def test_repairs_epigraph_split_before_mixed_guillemet_close(self):
+        raw = '{"segments":[' \
+            '{"speaker":"Schwester Larissa","type":"dialogue","text":"»Von dem Windhager Felsgesteyn bis zur Kueste von Mendena ziehet es sich hin, von Ifirns Eisgebirg bis zu dem dampfenden Waldt der Mohas, und"},' \
+            '{"speaker":"Narrator","type":"narration","text":"fürwahr, es ist ein seltsam Land. Lieblich bisweilen, mild und voll der guten Menschen. ‹ Zitiert aus dem Buch."}' \
+            ']}'
+        result = _attribution_parse(raw, "unused")
+        self.assertEqual(len(result["segments"]), 2)
+        dialogue, narration = result["segments"]
+        self.assertEqual(dialogue["speaker"], "Schwester Larissa")
+        self.assertEqual(dialogue["type"], "dialogue")
+        self.assertTrue(dialogue["text"].endswith("Land. Lieblich bisweilen, mild und voll der guten Menschen. ‹"))
+        self.assertEqual(narration["text"], "Zitiert aus dem Buch.")
+
+
+if __name__ == "__main__":
+    unittest.main()