fix(graph): harmonic damping for repeated same-edge_type evidence

First full-case run (runs/2026-05-20T20-15-04/) produced hypotheses
with log_odds +31 (8 direct_evidence + 15 supports). That's the
naive-Bayes independence assumption breaking down: 15 different
phenomena all "supporting" the same hypothesis from one source are
not 15 independent pieces of evidence, they're highly correlated.
DESIGN.md §4.5 last bullet flagged this as a "未实施旋钮" — this
commit implements it.

Rule: the k-th edge of a given (hyp_id, edge_type) contributes
log_lr_base / k instead of log_lr_base. Cumulative is harmonic
sum H_N, bounded by ~ ln N. Single-edge hypotheses unaffected
(k=1 → /1 → no change). Replaying the 2026-05-20 graph's 108
edges under the new rule pulls the top hypothesis from +31.0 →
+8.75; the smallest active hypothesis from +4.0 → +2.08.

Also adds rank + log_lr_base to confidence_log entries so the
math is auditable from the persisted graph.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

evidence_graph.py

@@ -1020,9 +1020,20 @@ class EvidenceGraph:
         **Idempotency**: if a ``(phenomenon, hypothesis, edge_type)`` edge
         already exists, this is a no-op — the same agent re-recording the
         same link (or two agents linking via the orchestrator's batch
-        judge and a manual override) does not double-count. Independent
-        evidence — *different* phenomena pointing the same way — still
-        accumulates fully.
+        judge and a manual override) does not double-count.
+
+        **Harmonic damping of repeated same-direction evidence** (added
+        post first full-case run, 2026-05-20): independent evidence —
+        different phenomena pointing the same way — still accumulates,
+        but with diminishing returns: the k-th edge of the same
+        ``(hyp_id, edge_type)`` contributes ``log_lr_base / k``. After N
+        same-direction edges the cumulative contribution is
+        ``log_lr_base · H_N`` (harmonic sum, grows as ln N). This
+        formalises the naive-Bayes-breakdown DESIGN.md §4.5 calls out:
+        "同一发现被多 agent 重复入图". Single-edge hypotheses are
+        unaffected (k=1, damping = 1.0). Cross-direction edges (supports
+        vs contradicts) keep their own independent counts so a strong
+        contradicting fact still bites against piled-on supports.
         """
         if edge_type not in self.edge_log_lr:
             raise ValueError(
@@ -1045,7 +1056,17 @@ class EvidenceGraph:
                 ):
                     return hyp.confidence
 
-            log_lr = self.edge_log_lr[edge_type]
+            # Harmonic damping rank: count existing edges of the SAME
+            # edge_type already incident on this hypothesis. The new edge
+            # becomes the (rank+1)-th of its kind. _adj_rev is keyed by
+            # target so this is O(in-degree(hyp)) without scanning all edges.
+            existing_same_type = sum(
+                1 for e in self._adj_rev.get(hyp_id, [])
+                if e.edge_type == edge_type
+            )
+            rank = existing_same_type + 1
+            log_lr_base = self.edge_log_lr[edge_type]
+            log_lr = log_lr_base / rank
             old_log_odds = hyp.log_odds
             old_conf = hyp.confidence
             new_log_odds = old_log_odds + log_lr
@@ -1065,7 +1086,9 @@ class EvidenceGraph:
                 "timestamp": datetime.now().isoformat(),
                 "phenomenon_id": phenomenon_id,
                 "edge_type": edge_type,
-                "log_lr": log_lr,
+                "log_lr_base": log_lr_base,
+                "rank": rank,
+                "log_lr": round(log_lr, 4),
                 "old_log_odds": round(old_log_odds, 4),
                 "new_log_odds": round(new_log_odds, 4),
                 "old_confidence": round(old_conf, 4),