fix: raise synthesis max_tokens to 64000 and surface truncation explicitly

The single-call synthesis (brief + company profile + per-fact
evidence_ref, merged in 8727e56) needs 25-40K output tokens; the
inherited 16384 cap (from the pre-merge two-call split) cut the JSON
mid-object every time (stop_reason=max_tokens confirmed on a live
AAPL run), failing parsing and producing an empty PARTIAL brief
(verified: 0/0) regardless of data quality or freshness. This was
the actual cause behind every "not enough verified evidence" report
throughout this incident, including the original stale NVDA brief.

64000 is claude-haiku-4-5's output ceiling (verified live), with
billing only for tokens actually generated. Also detect
stop_reason/finish_reason == max_tokens/length on the stream and
raise SynthesisTruncatedError, so a future truncation produces an
explicit "truncated by the token limit" reason instead of a generic
JSON-parse failure message.

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

agent/graph.py

@@ -362,6 +362,16 @@ def _format_signals_message(signals: list[dict]) -> str:
 MAX_EDGE_SIGNALS = 12
 MAX_FILING_SIGNALS = 8
 MAX_TRANSCRIPT_SIGNALS = 6
+# Output budget for the single-call synthesis (brief + company profile +
+# one full evidence_ref per fact). 16384 was inherited from the pre-merge
+# two-call split and truncated real AAPL briefs mid-JSON (stop_reason=
+# max_tokens at exactly 16384 output tokens). 64000 is the claude-haiku-4-5
+# output ceiling; billing only covers tokens actually generated.
+SYNTHESIS_MAX_TOKENS = 64000
+
+
+class SynthesisTruncatedError(RuntimeError):
+    """The synthesis stream was cut off by the max_tokens limit."""
 
 
 def _cap_signals(signals: list[dict], max_total: int = MAX_EDGE_SIGNALS) -> list[dict]:
@@ -495,7 +505,7 @@ def create_graph(config: Optional[RunConfig] = None):
 
     def synthesis_node(state: AgentState) -> dict:
         try:
-            llm_plain = make_chat_model(cfg, max_tokens=16384)
+            llm_plain = make_chat_model(cfg, max_tokens=SYNTHESIS_MAX_TOKENS)
             # Main prompt — cached (ephemeral) on Anthropic. Keep this block stable
             # so the cache hit rate is preserved regardless of the chosen language.
             lang = state.get("language") or "English"
@@ -525,10 +535,22 @@ def create_graph(config: Optional[RunConfig] = None):
                     HumanMessage(content="Now produce the structured research brief as a JSON object.")
                 ]
             chunks = []
+            stop_reason = None
             for chunk in llm_plain.stream(synthesis_messages):
+                metadata = getattr(chunk, "response_metadata", None) or {}
+                stop_reason = (
+                    metadata.get("stop_reason")
+                    or metadata.get("finish_reason")
+                    or stop_reason
+                )
                 text = chunk.content if isinstance(chunk.content, str) else ""
                 if text:
                     chunks.append(text)
+            if stop_reason in ("max_tokens", "length"):
+                raise SynthesisTruncatedError(
+                    "Synthesis output was truncated by the token limit "
+                    f"(max_tokens={SYNTHESIS_MAX_TOKENS}); the brief JSON was incomplete."
+                )
             raw = "".join(chunks)
             clean = _extract_json(raw)
             data = json.loads(clean)
@@ -596,7 +618,11 @@ def create_graph(config: Optional[RunConfig] = None):
         except Exception as exc:
             import sys
             print(f"[synthesis error] {exc}", file=sys.stderr)
-            partial = _partial_brief(state, f"Synthesis failed validation: {exc}")
+            if isinstance(exc, SynthesisTruncatedError):
+                reason = str(exc)
+            else:
+                reason = f"Synthesis failed validation: {exc}"
+            partial = _partial_brief(state, reason)
             return {
                 "brief": partial,
                 "brief_markdown": None,

tests/test_graph.py

@@ -243,6 +243,130 @@ def test_synthesis_zero_verified_keeps_filtered_brief_not_partial_skeleton(monke
     assert brief["coverage"]
 
 
+def test_synthesis_requests_expanded_max_tokens(monkeypatch):
+    fact = {
+        "text": "Revenue increased 12%.",
+        "source": "10-Q",
+        "reliability": "HIGH",
+        "evidence_snippet": "Revenue increased 12%.",
+    }
+    brief_json = json.dumps({
+        "ticker": "AAPL",
+        "company_name": "Apple Inc.",
+        "filing_date": "2026-04-30",
+        "what_matters_most": "Unsupported synthesis commentary.",
+        "standout_number": fact,
+        "what_changed": [],
+        "bull_points": [fact],
+        "bear_points": [],
+        "what_to_watch": ["Watch Q3 gross margin"],
+        "trends": [],
+        "mda_summary": {
+            "drivers": [],
+            "headwinds": [],
+            "language_shift": "No prior-period comparison was available.",
+            "key_quote": fact,
+        },
+        "risks_categorized": [],
+        "management_commentary": [],
+        "guidance_history": [],
+        "sentiment": None,
+        "market_expectations": None,
+    })
+
+    class FakeLLM:
+        def bind_tools(self, tools):
+            return self
+
+        def invoke(self, messages):
+            return AIMessage(content="done", tool_calls=[])
+
+        def stream(self, messages):
+            yield AIMessage(content=brief_json)
+
+    calls = []
+
+    def spy_make_chat_model(*args, **kwargs):
+        calls.append(kwargs)
+        return FakeLLM()
+
+    monkeypatch.setattr("analysis.textdiff.compute", lambda ticker: [])
+    monkeypatch.setattr("analysis.tone_drift.compute", lambda ticker: [])
+    monkeypatch.setattr(
+        "agent.company_profile.collect_profile_evidence",
+        lambda ticker, include_metrics=True: [],
+    )
+    monkeypatch.setattr("agent.graph.make_chat_model", spy_make_chat_model)
+
+    cfg = RunConfig(
+        provider="anthropic",
+        model="claude-haiku-4-5-20251001",
+        api_key="sk-ant-test",
+    )
+    initial_state = _state([
+        HumanMessage(content="Generate a research brief for AAPL."),
+        _tool_msg("get_financial_metrics"),
+        _tool_msg("search_filing"),
+        _tool_msg("search_filing", suffix="2"),
+        _tool_msg("search_transcript"),
+    ], 0)
+
+    create_graph(cfg).invoke(initial_state)
+
+    from agent.graph import SYNTHESIS_MAX_TOKENS
+
+    assert SYNTHESIS_MAX_TOKENS == 64000
+    assert any(
+        kwargs.get("max_tokens") == SYNTHESIS_MAX_TOKENS
+        for kwargs in calls
+    )
+
+
+def test_synthesis_truncation_produces_explicit_partial_reason(monkeypatch):
+    class FakeLLM:
+        def bind_tools(self, tools):
+            return self
+
+        def invoke(self, messages):
+            return AIMessage(content="done", tool_calls=[])
+
+        def stream(self, messages):
+            yield AIMessage(
+                content='{"ticker": "AAPL", "company_name": "Apple',
+                response_metadata={"stop_reason": "max_tokens"},
+            )
+
+    monkeypatch.setattr("analysis.textdiff.compute", lambda ticker: [])
+    monkeypatch.setattr("analysis.tone_drift.compute", lambda ticker: [])
+    monkeypatch.setattr(
+        "agent.company_profile.collect_profile_evidence",
+        lambda ticker, include_metrics=True: [],
+    )
+    monkeypatch.setattr("agent.graph.make_chat_model", lambda *args, **kwargs: FakeLLM())
+
+    cfg = RunConfig(
+        provider="anthropic",
+        model="claude-haiku-4-5-20251001",
+        api_key="sk-ant-test",
+    )
+    initial_state = _state([
+        HumanMessage(content="Generate a research brief for AAPL."),
+        _tool_msg("get_financial_metrics"),
+        _tool_msg("search_filing"),
+        _tool_msg("search_filing", suffix="2"),
+        _tool_msg("search_transcript"),
+    ], 0)
+
+    final = create_graph(cfg).invoke(initial_state)
+    brief = final["brief"]
+
+    assert brief["status"] == "PARTIAL"
+    assert "truncated" in brief["evidence_notes"][0]
+    assert "token limit" in brief["evidence_notes"][0]
+    assert not brief["evidence_notes"][0].startswith("Synthesis failed validation")
+    assert "truncated" in final["synthesis_error"]
+
+
 def test_partial_brief_falls_back_to_metrics_db(monkeypatch):
     monkeypatch.setattr(
         "storage.metrics_db.get_metrics",