update

backend/config.py

@@ -75,9 +75,6 @@ SEED_COVER_COS = 0.80
 # (Markdown: 50 pairs in the band, 4 above) — below every auto-merge threshold, so an LLM
 # judge decides. Candidates only; a merge still needs judge unanimity.
 SUB_DUP_KANDIDAT_COS = 0.75
-# Cross-block judge pairs per call: ONE call over all pairs scaled its timeout to 54 min
-# and a hung call blocked the barrier that long (aak: 196 pairs) — chunks cap it at ~15 min.
-CROSS_CHUNK_PAARE = 40
 
 # Umbrella grouping (block granularity level 2, step "Blocks-Gruppierung", AFTER the filter):
 # collapse sibling DEFINITIONS that are components of ONE umbrella concept (TM model:
@@ -94,16 +91,13 @@ EMBEDDING_SIBLING_CAP = 18       # a rich model (TM) can have many constituent p
 # 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
-# Fragment-demote backstop, same logic as GROUP_MIN_COS_FLOOR: fragment↔parent cosine is a BAD
-# fragment detector (measured, Markdown run: wrong demotes Blockzitate→Codeblöcke 0.353 and
+# Zielband der Themen-Blockzahl im Gruppierungs-Prompt: k = GROUP_THEMES_PER_SQRT·√n
+# (n = Items der Grouping-Welle), Band [0.7k, 1.3k] — weiche Vorgabe, kein Cap. √n hält
+# Blockzahl UND mittlere Blockgröße sublinear (n=30 → ~4–8, n=285 → ~12–22; das alte
+# statische „15–25" passte implizit nur zu n≈225–625).
+GROUP_THEMES_PER_SQRT = 1.0
+# Fragment-demote backstop (no-structure floor): fragment↔parent cosine is a BAD fragment
+# detector (measured, Markdown run: wrong demotes Blockzitate→Codeblöcke 0.353 and
 # Zeichenkodierung→Überschriften 0.640 sit ABOVE any usable floor, while true NP proof-gadget
 # demotes αu-Variablen→Cook/Levin 0.172 sit low). So this only vetoes judge/panel demotes with
 # NO containment match whose pair is literally structureless (Emoji→Tabelle 0.136).
@@ -135,10 +129,15 @@ CONSENSUS_GRACE = 300
 # check loops leave any remaining objections standing after that.
 CONSENSUS_MAX_ROUNDS = 3
 
-# QA gate: after the inventory phase an automatic QA run scores the blocks; below the
-# threshold the flow PAUSES before board 2 burns tokens (frontend offers force-continue).
-QA_GATE_NOTE = 9.5   # 0 = gate off; quota-based, so the tolerated finding count scales with topic size
+# QA gate: after the inventory phase an automatic QA run scores the blocks; below 100 %
+# the flow PAUSES before board 2 burns tokens (frontend offers force-continue).
+# QA_GATE_NOTE ist nur noch der An/Aus-Schalter (> 0 = Gate aktiv) — das Gate selbst
+# öffnet ausschließlich bei 100 % (auto_loop.VOLL); Repair/Recherche arbeiten dahin.
+QA_GATE_NOTE = 9.5   # 0 = gate off
 QA_GATE_LLM = True   # include the LLM samples (Echtheit/Dubletten) in the gate run
+# Lücken-Recherche im Auto-Loop: max. gezielte Nach-Recherche-Runden pro Lauf, bevor die
+# Rest-Lücken dem Stichentscheid vorgelegt werden (Repair-Budget, keine Detektor-Konstante).
+LUECKEN_RESEARCH_MAX = 2
 
 # Inline evidence for judge agents: corpus excerpts go INTO the prompt instead of letting
 # every judge re-search the source folder (measured: ~10 tool turns/judge, 82 % of the