fix: allow metrics/analyst sources in CategorizedRisk and GuidancePoint

CategorizedRisk.source and GuidancePoint.source hand-rolled a narrower
Literal["10-K", "10-Q", "transcript", "news"] instead of reusing the
shared EvidenceSource type (which also allows "metrics" and
"analyst"). Since get_financial_metrics legitimately tags its
evidence with source="metrics" (guidance figures and risk-relevant
metrics come straight from the ingested XBRL data), any brief where
the model correctly cited a metrics-sourced risk or guidance point
failed BriefOutput validation outright and fell back to an empty
PARTIAL brief — confirmed on a live AAPL run right after the
max_tokens fix (602548d) let synthesis complete for the first time.

ManagementCommentaryTopic.source is intentionally left unchanged
(deliberately narrower — no metrics/news commentary, per prompt and
_drop_news_commentary sanitizer).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

agent/schemas.py

@@ -478,7 +478,7 @@ class CategorizedRisk(BaseModel):
         description="Risk category."
     )
     text: str = Field(description="The risk, 1-2 sentences grounded in filing language.")
-    source: Literal["10-K", "10-Q", "transcript", "news"] = Field(
+    source: EvidenceSource = Field(
         description="Document type where this risk was cited."
     )
     reliability: Literal["HIGH", "MEDIUM", "LOW"] = Field(
@@ -550,7 +550,7 @@ class GuidancePoint(BaseModel):
 
     period: str = Field(description="The filing period when this guidance was given, e.g. 'Q1 2025'.")
     text: str = Field(description="The guidance statement, 1-2 sentences.")
-    source: Literal["10-K", "10-Q", "transcript", "news"] = Field(
+    source: EvidenceSource = Field(
         description="Document type where this guidance appeared."
     )
 

tests/test_schemas.py

@@ -99,6 +99,25 @@ def test_brief_output_valid():
     assert brief.what_to_watch == ["Q2 iPhone shipments", "AI feature adoption"]
 
 
+def test_brief_output_accepts_metrics_sources_for_risk_and_guidance():
+    brief = _minimal_brief(
+        risks_categorized=[dict(
+            category="Financial",
+            text="Margin compression risk.",
+            source="metrics",
+            reliability="HIGH",
+            is_new_this_filing=False,
+        )],
+        guidance_history=[dict(
+            period="Q3 2025",
+            text="Revenue guidance narrowed.",
+            source="metrics",
+        )],
+    )
+    assert brief.risks_categorized[0].source == "metrics"
+    assert brief.guidance_history[0].source == "metrics"
+
+
 def test_sourced_fact_ignores_extra_fields():
     # ConfigDict(extra="ignore") — extra fields are silently dropped, not rejected.
     # This is intentional: LLM output may include unexpected keys.
@@ -166,6 +185,17 @@ def test_management_commentary_topic_rejects_news_source():
         )
 
 
+def test_management_commentary_topic_rejects_metrics_source():
+    with pytest.raises(ValidationError):
+        ManagementCommentaryTopic(
+            topic="Revenue guidance",
+            summary="Revenue guidance was updated.",
+            source="metrics",
+            reliability="HIGH",
+            evidence_snippet="Revenue guidance was updated.",
+        )
+
+
 def test_management_commentary_topic_trims_long_snippet():
     snippet = " ".join([f"word{i}" for i in range(40)])
     t = ManagementCommentaryTopic(
@@ -228,6 +258,16 @@ def test_categorized_risk_accepts_demand():
     assert r.category == "Demand"
 
 
+def test_categorized_risk_accepts_metrics_source():
+    r = _risk(category="Financial", source="metrics")
+    assert r.source == "metrics"
+
+
+def test_categorized_risk_accepts_analyst_source():
+    r = _risk(category="Financial", source="analyst")
+    assert r.source == "analyst"
+
+
 def test_categorized_risk_case_folds_lowercase():
     r = _risk(category="regulatory")
     assert r.category == "Regulatory"
@@ -593,6 +633,16 @@ def _gp(**kw) -> GuidancePoint:
     return GuidancePoint(**(defaults | kw))
 
 
+def test_guidance_point_accepts_metrics_source():
+    g = _gp(source="metrics")
+    assert g.source == "metrics"
+
+
+def test_guidance_point_accepts_analyst_source():
+    g = _gp(source="analyst")
+    assert g.source == "analyst"
+
+
 def test_guidance_verdict_valid_passthrough():
     assert _gp(verdict="beat").verdict == "beat"
     assert _gp(verdict="in-line").verdict == "in-line"