update

backend/config.py

@@ -11,8 +11,10 @@ UNI_DIR = PROJECT_ROOT / "uni"
 
 
 def _load_env(path: Path) -> None:
-    """Mini .env loader (no dependency): KEY=VALUE lines; existing env always wins
-    (`make dev` already exports .env — this covers bare `uvicorn`/pytest starts)."""
+    """Mini .env loader (no dependency): KEY=VALUE lines. The FILE wins over inherited
+    env: a --reload master keeps its startup environment forever, so "existing env wins"
+    silently pinned stale values across .env edits (measured: file said 24, workers
+    inherited 15 for hours). Trade-off: ad-hoc shell overrides lose against the file."""
     try:
         text = path.read_text(encoding="utf-8")
     except OSError:
@@ -23,7 +25,7 @@ def _load_env(path: Path) -> None:
             continue
         key, _, value = line.partition("=")
         key, value = key.strip(), value.strip().strip('"').strip("'")
-        if key and key not in os.environ:
+        if key:
             os.environ[key] = value
 
 
@@ -59,6 +61,15 @@ EMBEDDING_BLOCK_CAP = 25      # max. titles per block (keep the LLM list short/s
 # block context — from this cosine on two are the same statement (checked on aak: ≥0.88 are
 # without exception true duplicates). Conservative 0.90 so different aspects (∈NP ≠ NP-hard) stay separate.
 EMBEDDING_SUB_DUP = 0.90
+# Variant folding BEFORE the subblock consensus count: finders rephrase the same concept each
+# round, so exact-norm counting starves real concepts (measured Markdown run: 623/965 mentions
+# discarded, „Zeichenkodierung" 73/74). 0.90 folds true paraphrases at ~0 false folds (0.85/0.88
+# fold distinct aspects like ** vs ***). Antonym pairs measure 0.91–0.95 → negation guard required.
+SUB_VARIANT_COS = 0.90
+# Seed coverage check is LEXICAL first (token containment) — seeds are short fragment NAMES,
+# subs are statements: true covers measure 0.27–0.38 while a wrong hit measured 0.76. The
+# embedding stage only backs up the lexical one (catches „Line Breaks (Soft)" 0.888).
+SEED_COVER_COS = 0.80
 
 # Umbrella grouping (block granularity level 2, step "Blocks-Gruppierung", AFTER the filter):
 # collapse sibling DEFINITIONS that are components of ONE umbrella concept (TM model:
@@ -83,6 +94,12 @@ GROUP_RECONCILE_FLOOR = 0.75
 # 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
+# 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).
+FRAGMENT_MIN_COS = 0.15
 
 # Caps for concurrent CLI agent processes (env-overridable). Two nested limits, both always active:
 # a per-topic cap and a global cap across all topics. Defaults 10/10 = previous behavior (global
@@ -125,15 +142,15 @@ QUELLE_RELEVANZ_SNIPPET = 800    # body characters per page in the prompt (URL i
 # Timeouts per agent step: (base seconds, seconds per block/section).
 # Applies equally to all providers — whoever is too slow gets restarted or overtaken.
 TIMEOUTS = {
-    "research":     (1800, 0),   # fixed 30 min
+    "research":     (900, 0),    # p95 measured 125 s (web mode); uni/link sections need headroom
     "research_mapping": (600, 3),  # n = pre-merged entries
     "selection_mapping":   (600, 2),  # n = remaining entries (block inventory)
-    "ergaenzung":    (900, 0),   # subject-field extension for projects (web research)
+    "ergaenzung":    (600, 0),   # subject-field extension for projects (web research)
     "plan":          (300, 5),
     "plan_judge":    (600, 5),    # judge reads up to 5 outlines, n = sections
-    "content":        (600, 90),   # identify content per block in the chunk (web search)
+    "content":        (450, 30),   # facts find/erg/fix — p95 measured 241 s (was 600+90n)
     "content_check":  (300, 10),   # content exam per block in the package
-    "subblock":       (900, 45),  # find subblocks per block in the chunk (web search)
+    "subblock":       (400, 15),  # finder round — p95 measured 124 s (was 900+45n)
     "subblock_check": (300, 15),  # judge decides contested subblocks in the chunk
     "level":             (300, 10),  # classify subblocks per chunk
     "level_check":       (300, 10),  # judge decides contested levels in the chunk
@@ -141,7 +158,7 @@ TIMEOUTS = {
     "relevance_check":    (300, 10),  # judge decides contested relevance in the chunk
     "question_pattern":      (300, 15),  # question patterns per block (subblocks × types)
     "question_pattern_check": (300, 10),  # critic cleans up the pattern table per block
-    "writer":        (600, 120),  # per section in the chunk
+    "writer":        (450, 60),   # per section — split keeps sections ≤30 subs
     "lese_check":    (300, 10),   # per section in the package
     # guide board (per card = one block)
     "lernziele":     (300, 5),    # backward-design objectives per block
@@ -194,15 +211,14 @@ PROVIDERS = {
     },
 }
 
-# Role routing ACROSS provider stacks: generation (quick/guide) and judging (judge)
-# may run on different providers within ONE run — judge model ≠ generator model
-# (research-backed: cross-model judging avoids self-preference bias).
-# Value: "" = provider of the run; "minimax" = that stack's role model;
-# "provider:model" = explicit model override.
+# Role routing: by DEFAULT the run's provider (the UI choice) handles ALL roles —
+# the role only picks the model WITHIN that stack (PROVIDERS[stack][role]).
+# Opt-in cross-provider mixing via env: ROLE_JUDGE=claude routes every judge call
+# to the claude stack regardless of the UI choice ("provider:model" pins a model).
 ROLE_ROUTING = {
-    "quick": os.getenv("ROLE_QUICK", "minimax"),
-    "judge": os.getenv("ROLE_JUDGE", "claude"),
-    "guide": os.getenv("ROLE_GUIDE", "minimax"),
+    "quick": os.getenv("ROLE_QUICK", ""),
+    "judge": os.getenv("ROLE_JUDGE", ""),
+    "guide": os.getenv("ROLE_GUIDE", ""),
     "fast":  os.getenv("ROLE_FAST", ""),
 }