feat(gnn): activate closed-loop scoring at agent → tactic-suggestion seam

Phase 3 (final piece) of the GNN feedback loop: connect the call site
where AgentCore.process_goal calls into provers/mod.rs::gnn_augment_tactics.
After this commit, the loop is fully closed end-to-end with no manual wiring:

  StatisticsTracker → /training/update → PROVER_DOMAIN_WEIGHTS
                                       ↓
  AgenticGoal.aspects → state.metadata["aspects"] → domain_hints in /gnn/rank
                                       ↓
                          weight-modulated premise scoring

- agent/mod.rs::process_goal: write goal.aspects into state.metadata
  under the "aspects" key alongside the existing coprocessor witness data
- provers/mod.rs::gnn_augment_tactics: read state.metadata["aspects"]
  and pass to gnn.rank_premises_with_aspects (falls back to empty for
  non-agent invocations like REPL — same behaviour as before)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Author: hyperpolymath

src/rust/agent/mod.rs

@@ -377,6 +377,13 @@ impl AgentCore {
                 metadata.insert(format!("coprocessor.{}", pre.aspect), v);
             }
         }
+        // Publish the goal's aspects so downstream neural ranking
+        // (gnn_augment_tactics) can pass them to Julia's /gnn/rank as
+        // domain_hints, activating the closed-loop weight-guided scoring
+        // from accumulated PROVER_DOMAIN_WEIGHTS.
+        if let Ok(v) = serde_json::to_value(&goal.aspects) {
+            metadata.insert("aspects".to_string(), v);
+        }
 
         // Step 5: Get tactic suggestions (neural guidance), using the
         // context-enriched initial state so the suggestion model sees the

src/rust/provers/mod.rs

@@ -1399,7 +1399,16 @@ pub(crate) async fn gnn_augment_tactics(
         use_attention: true,
     });
 
-    let result = gnn.rank_premises(&graph).await;
+    // Extract goal aspects from state metadata (written by AgentCore::process_goal)
+    // so Julia's /gnn/rank can apply per-domain weights from prior outcomes.
+    // When metadata has no "aspects" key (e.g. direct REPL invocation), aspects
+    // is empty and behaviour is identical to the no-hint path.
+    let aspects: Vec<String> = state
+        .metadata
+        .get("aspects")
+        .and_then(|v| serde_json::from_value(v.clone()).ok())
+        .unwrap_or_default();
+    let result = gnn.rank_premises_with_aspects(&graph, &aspects).await;
     // Prepend apply tactics for top premises (in score order, before heuristic hints)
     let mut gnn_tactics: Vec<Tactic> = result
         .ranked_premises