update

backend/config.py

@@ -42,6 +42,30 @@ EMBEDDING_BLOCK_CAP = 25      # max. titles per block (keep the LLM list short/s
 # without exception true duplicates). Conservative 0.90 so different aspects (∈NP ≠ NP-hard) stay separate.
 EMBEDDING_SUB_DUP = 0.90
 
+# Umbrella grouping (block granularity level 2, step "Blocks-Gruppierung", AFTER the filter):
+# collapse sibling DEFINITIONS that are components of ONE umbrella concept (TM model:
+# Konfiguration/Start-/Folge-/Stopkonfiguration/Berechnung/Alphabet/δ; KNF: Literale/Klauseln/
+# Variablen) into ONE block whose description ENUMERATES the children — so the later subblock step
+# re-derives them from the source under the umbrella scope (demotion is re-derivation, not transfer).
+# If the flag is off or no embedding model → step is silently skipped (like Dedup).
+BLOCKS_GRUPPIERUNG_AKTIV = True
+# Lower floor than consolidation (0.5, paraphrase-tuned) for higher sibling recall; the LLM judge is
+# the precision gate. Smaller cap since a lower floor pulls in more nodes → keep the judge lists short.
+EMBEDDING_SIBLING_FLOOR = 0.35   # heterogeneous facets of one model co-cluster weakly → low floor (recall)
+EMBEDDING_SIBLING_CAP = 18       # a rich model (TM) can have many constituent parts
+# Reconcile pass: two independently-judged clusters can emit the SAME parent concept under different
+# titles (e.g. two „Turingmaschine"-umbrellas). Merge umbrella pairs whose title+description cosine is
+# ≥ this (conservative → only true same-parent duplicates, never two distinct umbrellas).
+GROUP_RECONCILE_FLOOR = 0.75
+# Over-merge backstop ONLY (no-structure floor). Research (meronymy ≠ similarity): parts of ONE model are
+# legitimately DISSIMILAR (TM: Alphabet/Konfiguration/δ ~0.22), while distinct same-type concepts (P/NP/…)
+# are SIMILAR (~0.85) — so member-vs-member cosine is the WRONG instrument for over-merge (empirically
+# inverted: TM 0.218 < the P/NP bundle 0.227). The real precision floor is the ATOMICITY type-guard
+# (_GROUP_STANDALONE: a member that is a named algorithm/problem/theorem/complexity-class dissolves the
+# umbrella). This floor is demoted to a near-zero backstop that only rejects a literally structureless
+# chain (random-pair baseline), set BELOW the legitimate heterogeneous minimum so it never kills a real model.
+GROUP_MIN_COS_FLOOR = 0.15
+
 # Cap for concurrent CLI agent processes (across all generations).
 # Own lane for interactive calls (chat, elements) so they don't hang behind
 # running writers in the queue.

backend/textkit.py

@@ -56,11 +56,14 @@ def _resolve_title(idx: dict[str, int], t: str) -> int | None:
 
 
 def _norm_dash(s: str) -> str:
-    """Space-surrounded dash variants (en/em/figure/bar/hyphen) → uniform separator ' — '.
+    """Dash variants (en/em/figure/bar) with whitespace on AT LEAST ONE side → uniform separator ' — '.
     Some models (especially non-western ones) use an en-dash "–" instead of the em-dash; without
-    normalization the ` — ` split fails entirely and the whole entry becomes the title.
-    The ASCII hyphen "-" is left untouched (otherwise it would split formulas like "n - 1")."""
-    return re.sub(r"\s+[‒–—―‐]\s+", " — ", s)
+    normalization the ` — ` split fails entirely and the whole entry becomes the title. A one-sided
+    space ("Titel —Beschreibung" / "Titel— Beschreibung") also breaks the split and leaks the source
+    filename into the description — so a dash with a space on either side is repaired too. The ASCII
+    hyphen "-" is deliberately NOT in the class (would split "n - 1"/"3-SAT"); requiring ≥1 surrounding
+    space keeps glued compounds like "Backtracking—Verfahren" and number ranges like "12–15" untouched."""
+    return re.sub(r"\s*[‒–—―]\s+|\s+[‒–—―]\s*", " — ", s)
 
 
 def _parse_selection(text: str) -> dict[int, str]: