fix(502): don't demote merge-stage stories on mergemaster attach

start_agent unconditionally called move_story_to_current at the top of
its body, before the agent-stage check. When called for mergemaster (or
qa) on a story in 4_merge/ AND a stale 1_backlog/ shadow of the story
existed (post-491/492 split-brain artifact), the move would find the
shadow and yank it to 2_current/, find_active_story_stage would then
report 2_current/, the stage check would expect a Coder agent, and
mergemaster would be rejected — leaving the story in 2_current/ to be
re-promoted by the next auto-assign tick. Infinite loop.

Gate the move so it only fires for Coder-stage agents. QA and
Mergemaster now attach to the story at its existing stage.

Adds a regression test that reproduces the split-brain scenario by
seeding both 4_merge/ and 1_backlog/ copies of the same story and
asserting (1) the stage check does not reject mergemaster, and (2) the
4_merge/ copy is preserved (i.e. not demoted to 2_current/).

Observed live on 2026-04-09 while story 478 was looping. Filed as
bug 502.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

server/src/agents/pool/start.rs

@@ -47,10 +47,18 @@ impl AgentPool {
 
         // Move story from backlog/ to current/ before checking agent
         // availability so that auto_assign_available_work can pick it up even
-        // when all coders are currently busy (story 203).  This is idempotent:
-        // if the story is already in 2_current/ or a later stage, the call is
-        // a no-op.
+        // when all coders are currently busy (story 203). Only do this for
+        // Coder-stage agents — QA and Mergemaster must attach to the story
+        // at its existing stage (3_qa or 4_merge) and must NOT be demoted
+        // back to 2_current/ on attach (bug 502). When `agent_name` is None
+        // we are auto-selecting an idle coder, so still move.
+        let starting_a_coder = agent_name
+            .and_then(|n| config.find_agent(n).map(agent_config_stage))
+            .map(|s| s == PipelineStage::Coder)
+            .unwrap_or(true);
+        if starting_a_coder {
             crate::agents::lifecycle::move_story_to_current(project_root, story_id)?;
+        }
 
         // Validate that the agent's configured stage matches the story's
         // pipeline stage.  This prevents any caller (auto-assign, MCP tool,
@@ -1377,6 +1385,73 @@ stage = "coder"
         }
     }
 
+    /// Bug 502: when start_agent is called for a non-Coder agent (mergemaster
+    /// or qa) on a story that's in 4_merge/, the unconditional
+    /// move_story_to_current at the top of start_agent must NOT fire — even
+    /// when a stale split-brain shadow of the story exists in 1_backlog/.
+    ///
+    /// Pre-fix behaviour: move_story_to_current would find the 1_backlog
+    /// shadow and move it to 2_current/. find_active_story_stage would then
+    /// report 2_current/, the stage check would expect a Coder-stage agent,
+    /// and mergemaster would be rejected — leaving the story in 2_current/
+    /// to be picked up by the next auto-assign tick as a coder. Infinite loop.
+    /// Observed live on 2026-04-09 against story 478.
+    #[tokio::test]
+    async fn start_agent_does_not_demote_merge_stage_story_with_backlog_shadow() {
+        use std::fs;
+
+        let tmp = tempfile::tempdir().unwrap();
+        let root = tmp.path();
+
+        let sk_dir = root.join(".huskies");
+        fs::create_dir_all(sk_dir.join("work/1_backlog")).unwrap();
+        fs::create_dir_all(sk_dir.join("work/4_merge")).unwrap();
+        fs::write(
+            sk_dir.join("project.toml"),
+            "[[agent]]\nname = \"mergemaster\"\nstage = \"mergemaster\"\n",
+        )
+        .unwrap();
+        // Real copy in 4_merge/ (where the story actually is per the DB).
+        fs::write(
+            sk_dir.join("work/4_merge/502_story_split_brain.md"),
+            "---\nname: Split Brain\n---\n",
+        )
+        .unwrap();
+        // Stale split-brain shadow in 1_backlog/ (post-491/492 migration
+        // artifact — the filesystem shadow that bit us in production).
+        fs::write(
+            sk_dir.join("work/1_backlog/502_story_split_brain.md"),
+            "---\nname: Split Brain\n---\n",
+        )
+        .unwrap();
+
+        let pool = AgentPool::new_test(3098);
+        let result = pool
+            .start_agent(root, "502_story_split_brain", Some("mergemaster"), None)
+            .await;
+
+        // Stage check must not reject mergemaster.
+        if let Err(ref e) = result {
+            assert!(
+                !e.contains("cannot be assigned"),
+                "mergemaster on 4_merge/ story must not fail stage check even \
+                 when a 1_backlog shadow exists, got: '{e}'"
+            );
+        }
+
+        // Critical: the story must still be in 4_merge/ after the call.
+        // Before the fix, line 53 of start.rs would have demoted it to
+        // 2_current/ via move_story_to_current finding the 1_backlog shadow.
+        assert!(
+            sk_dir.join("work/4_merge/502_story_split_brain.md").exists(),
+            "story must still be in 4_merge/ after start_agent(mergemaster, ...)"
+        );
+        assert!(
+            !sk_dir.join("work/2_current/502_story_split_brain.md").exists(),
+            "story must NOT have been demoted to 2_current/ — that's bug 502"
+        );
+    }
+
     // ── front matter agent preference (bug 379) ──────────────────────────────
 
     #[tokio::test]