recon + sqlglot validator + drill_down package; guard ReAct dimension picks against candidate list

tests/test_events.py

@@ -2,6 +2,7 @@
 
 from api.analyses.types import Finding
 from api.runtime import events
+from api.runtime.context import current_analysis
 
 
 def test_run_start_shape():
@@ -22,10 +23,11 @@ def test_plan_ready_shape():
 
 
 def test_analysis_start_shape():
-    e = events.analysis_start("direct_answer", {"q": "?"}, "fits")
+    e = events.analysis_start("direct_answer", {"q": "?"}, "fits", step_id="direct_answer#0")
     assert e == {
         "type": "analysis_start",
         "analysis": "direct_answer",
+        "step_id": "direct_answer#0",
         "args": {"q": "?"},
         "why": "fits",
     }
@@ -33,19 +35,65 @@ def test_analysis_start_shape():
 
 def test_analysis_end_uses_finding():
     f = Finding(analysis="direct_answer", summary="ok", rows=[{"a": 1}], sql=["SELECT 1"])
-    e = events.analysis_end("direct_answer", f)
+    e = events.analysis_end("direct_answer", f, step_id="direct_answer#0")
     assert e["type"] == "analysis_end"
     assert e["analysis"] == "direct_answer"
+    assert e["step_id"] == "direct_answer#0"
     assert e["summary"] == "ok"
     assert e["error"] is None
     assert e["row_count"] == 1
 
 
-def test_tool_call_shapes():
+def test_analysis_fallback_shape():
+    e = events.analysis_fallback(
+        primary="drill_down", fallback="direct_answer",
+        reason="empty result", step_id="direct_answer#0.fallback",
+    )
+    assert e == {
+        "type": "analysis_fallback",
+        "primary": "drill_down",
+        "fallback": "direct_answer",
+        "reason": "empty result",
+        "step_id": "direct_answer#0.fallback",
+    }
+
+
+def test_tool_call_shapes_carry_analysis_field():
+    """Tool/LLM events must include the active Analysis name from the
+    contextvar — UI groups sub-events by it. None when outside an Analysis."""
+    # Outside any Analysis: analysis is None.
     s = events.tool_call_start("text_to_sql", input={"q": "?"})
-    assert s == {"type": "tool_call_start", "tool": "text_to_sql", "input": {"q": "?"}}
+    assert s["type"] == "tool_call_start"
+    assert s["tool"] == "text_to_sql"
+    assert s["analysis"] is None
+    assert s["input"] == {"q": "?"}
+
     e = events.tool_call_end("text_to_sql", output={"sql": "..."})
-    assert e == {"type": "tool_call_end", "tool": "text_to_sql", "output": {"sql": "..."}, "error": None}
+    assert e["analysis"] is None
+    assert e["output"] == {"sql": "..."}
+    assert e["error"] is None
+
+    # Inside an Analysis context: analysis field is populated.
+    token = current_analysis.set("direct_answer#0")
+    try:
+        s2 = events.tool_call_start("execute_sql", input={"sql": "SELECT 1"})
+        e2 = events.tool_call_end("execute_sql", output={"row_count": 1})
+        assert s2["analysis"] == "direct_answer#0"
+        assert e2["analysis"] == "direct_answer#0"
+    finally:
+        current_analysis.reset(token)
+
+
+def test_llm_call_carries_analysis_field():
+    token = current_analysis.set("compare_periods#0")
+    try:
+        e = events.llm_call("compare_periods.interpret", system_len=100, user_len=200)
+        assert e["type"] == "llm_call"
+        assert e["analysis"] == "compare_periods#0"
+        assert e["system_chars"] == 100
+        assert e["user_chars"] == 200
+    finally:
+        current_analysis.reset(token)
 
 
 def test_synth_start_shape():