The dialogue line must quote the runtime it was measured against

First real scan put the feature in front of actual proposals and showed
the numbers disagreeing with each other: 'Sakura Wars — 69 words/min,
2 of 115 min without dialogue'. The rate was right (computed against one
episode) but the 115 minutes is the SERIES total from MediaTechProfile,
so an episode's silence was rendered against a season's length.

The measured runtime is now stored alongside the metrics and read back
from there, with a migration for installs that already have the table.
The track's language tag is also no longer printed unless it is a plain
code — this library has a track labelled Hindi holding no Devanagari and
one labelled German that was Thai, and a wrong label on an evidence line
is worse than no label.

Also: the warm-up log line now reports dialogue profiles alongside
articles and reception. The UI message already did; the logger did not,
which made a working feature look like it had never run.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: Randomname653

src/database/connection.py

@@ -78,6 +78,9 @@ def _migrate_columns() -> None:
     """Add columns introduced after initial schema creation (SQLite-safe)."""
     from sqlalchemy import text
     new_cols = [
+        # Subtitle metrics: the episode runtime they were measured against
+        # (MediaTechProfile holds the SERIES total, which is a different number)
+        ("media_subtitle_profiles", "duration_min", "FLOAT"),
         ("deletion_proposals",    "category",    "TEXT"),
         ("deletion_proposals",    "poster_url",  "TEXT"),
         ("deletion_proposals",    "synopsis",    "TEXT"),

src/database/models.py

@@ -687,6 +687,11 @@ class MediaSubtitleProfile(Base):
     mattr           = Column(Float, nullable=True)        # NOT raw TTR (length-biased)
     total_words     = Column(Integer, nullable=True)
     coverage        = Column(Float, nullable=True)        # cue span / runtime
+    # The runtime these figures were computed against. For a series that is ONE
+    # episode, never the season total — MediaTechProfile aggregates every
+    # episode, and rendering "2 of 115 min without dialogue" from an episode's
+    # silence and a season's length reads as nonsense.
+    duration_min    = Column(Float, nullable=True)
     checked         = Column(Boolean, default=False)
     metrics_v       = Column(String(16), nullable=True)
     updated_at      = Column(DateTime, default=datetime.utcnow, onupdate=datetime.utcnow)

src/services/recommendations_engine.py

@@ -1750,8 +1750,10 @@ def _cmp_vec(vec):
                      f"{_warmed_rec} reception, {_warmed_wd} on-record facts, "
                      f"{_warmed_sub} dialogue profiles")
                 logger.info("[deletions] %s: pre-judge warm-up added %d article(s), "
-                            "%d reception record(s), %d wikidata record(s)",
-                            category, _warmed, _warmed_rec, _warmed_wd)
+                            "%d reception record(s), %d wikidata record(s), "
+                            "%d dialogue profile(s)",
+                            category, _warmed, _warmed_rec, _warmed_wd,
+                            _warmed_sub)
         except Exception as _e:
             logger.debug("[deletions] pre-judge warm-up failed: %s", _e)
         _gate_label = f"deletion scan: {category}"

src/services/subtitle_signals.py

@@ -431,7 +431,11 @@ def format_dialogue_line(m: dict, lang: str = "") -> str:
     """
     if not m or not m.get("words_per_min"):
         return ""
-    lang_part = f"{lang} " if lang else ""
+    # The track's own language tag is demonstrably unreliable here (a track
+    # labelled Hindi contained no Devanagari; one labelled German was Thai),
+    # so it is only shown when the metrics were computed at all — and never
+    # presented as a fact about the work.
+    lang_part = f"{lang} " if lang and len(lang) <= 5 else ""
     sdh = ", SDH track" if m.get("is_sdh") else ""
     div = (f", lexical diversity {m['mattr']}" if m.get("mattr") is not None
            else "")
@@ -580,6 +584,7 @@ async def topup_subtitle_metrics(title: str, media_type: str, *,
                 row.mattr = m.get("mattr")
                 row.total_words = m.get("total_words")
                 row.coverage = m.get("coverage")
+                row.duration_min = m.get("duration_min")
             db.commit()
         return bool(m)
     except Exception as e:
@@ -610,7 +615,9 @@ def subtitle_facts(item: dict, media_type: str) -> str:
             return format_dialogue_line({
                 "words_per_min": row.words_per_min,
                 "silent_min": row.silent_min or 0.0,
-                "duration_min": dur or 0.0,
+                # The runtime we MEASURED against — for a series one episode,
+                # not the season total the tech profile carries.
+                "duration_min": row.duration_min or dur or 0.0,
                 "mattr": row.mattr,
                 "is_sdh": bool(row.is_sdh),
             }, lang=row.language or "")