fix: tolerate malformed LLM-copied EvidenceRef fields

EvidenceRef enforced its own format constraints (64-hex pattern on
content_hash, min_length on evidence_id/document_id, a closed source
Literal, extra="forbid") on fields the LLM copies verbatim from a
retrieved evidence.v1 record. A single mis-copied field (e.g. a
truncated SHA-256 hash) crashed BriefOutput.model_validate entirely,
wiping the whole brief -- even though agent/evidence.py::verify_fact
already re-checks every one of these fields deterministically against
the real record and fails just that one fact cleanly (FAILED /
content_hash_mismatch, reliability=LOW) without touching the rest.

Confirmed on a live AAPL run: the model cited the same real evidence
record twice with a hash truncated to 47 of 64 hex characters, which
crashed synthesis outright under the old constraints.

EvidenceRef fields are now plain strings (source normalized via the
existing _normalize_source instead of a closed Literal), extra
fields are ignored rather than rejected, and verify_fact remains the
single source of truth for whether a citation is trustworthy.

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

agent/schemas.py

@@ -8,18 +8,23 @@ VerificationStatus = Literal["VERIFIED", "UNVERIFIED", "FAILED"]
 
 
 class EvidenceRef(BaseModel):
-    """Immutable locator copied from a retrieved ``evidence.v1`` record."""
+    """LLM-copied locator whose canonical form is enforced only by ``verify_fact``."""
 
-    model_config = ConfigDict(extra="forbid")
+    model_config = ConfigDict(extra="ignore")
 
-    evidence_id: str = Field(min_length=4)
-    source: EvidenceSource
-    content_hash: str = Field(pattern=r"^[0-9a-f]{64}$")
-    document_id: str = Field(min_length=1)
+    evidence_id: str = ""
+    source: str = ""
+    content_hash: str = ""
+    document_id: str = ""
     chunk_id: Optional[str] = None
     source_url: Optional[str] = None
     as_of: Optional[str] = None
 
+    @field_validator("source", mode="before")
+    @classmethod
+    def _coerce_source(cls, v: object) -> object:
+        return _normalize_source(v)
+
 
 class EvidenceRecord(BaseModel):
     """Retrieved content plus its content-addressed reference."""

tests/test_evidence.py

@@ -1,8 +1,5 @@
 import json
 
-import pytest
-from pydantic import ValidationError
-
 from agent.evidence import (
     NO_VERIFIED_SYNTHESIS_MESSAGE,
     content_hash,
@@ -67,12 +64,25 @@ def test_ids_and_hashes_are_stable_across_whitespace_and_metadata_order():
     assert first.ref.content_hash == content_hash(second.content)
 
 
-def test_evidence_ref_rejects_noncanonical_source():
+def test_noncanonical_ref_source_fails_verification_not_validation():
     record = _filing_record()
     raw = record.ref.model_dump()
     raw["source"] = "10-Q, transcript"
-    with pytest.raises(ValidationError):
-        EvidenceRef.model_validate(raw)
+    supplied = EvidenceRef.model_validate(raw)
+    fact = {
+        "text": "Revenue grew 5%.",
+        "source": "10-Q",
+        "reliability": "HIGH",
+        "evidence_snippet": "Revenue grew 5% year over year to $12.0 billion.",
+        "evidence_ref": supplied.model_dump(mode="json"),
+    }
+
+    verify_fact(fact, [record])
+
+    assert supplied.source == "10-Q, transcript"
+    assert fact["verification_status"] == "FAILED"
+    assert fact["verification_reason"] == "source_mismatch"
+    assert fact["reliability"] == "LOW"
 
 
 def test_sourced_fact_supports_structured_sources_and_defaults_unverified():
@@ -118,6 +128,25 @@ def test_verified_fact_requires_matching_id_source_hash_snippet_and_numbers():
     assert fact["verification_status"] == "VERIFIED"
 
 
+def test_truncated_content_hash_fails_verification_per_fact():
+    record = _filing_record()
+    truncated_ref = record.ref.model_dump(mode="json")
+    truncated_ref["content_hash"] = truncated_ref["content_hash"][:47]
+    fact = {
+        "text": "Revenue increased 5% to $12 billion.",
+        "source": "10-Q",
+        "reliability": "HIGH",
+        "evidence_snippet": "Revenue grew 5% year over year to $12.0 billion.",
+        "evidence_ref": truncated_ref,
+    }
+
+    verify_fact(fact, [record])
+
+    assert fact["verification_status"] == "FAILED"
+    assert fact["verification_reason"] == "content_hash_mismatch"
+    assert fact["reliability"] == "LOW"
+
+
 def test_unsupported_number_fails_closed_to_low():
     record = _filing_record()
     fact = {

tests/test_schemas.py

@@ -99,6 +99,38 @@ def test_brief_output_valid():
     assert brief.what_to_watch == ["Q2 iPhone shipments", "AI feature adoption"]
 
 
+def test_brief_output_tolerates_truncated_content_hash_in_evidence_ref():
+    truncated_hash = "496fb91760e2df7fe4e59bb606127879006da861385c98"
+    evidence_ref = {
+        "evidence_id": "ev_abc123",
+        "source": "10-Q",
+        "content_hash": truncated_hash,
+        "document_id": "sec:AAPL:0001",
+    }
+    payload = _minimal_brief().model_dump(mode="json")
+    payload["bear_points"] = [{
+        "text": "China headwinds persist.",
+        "source": "10-Q",
+        "reliability": "HIGH",
+        "evidence_snippet": "China headwinds persist.",
+        "evidence_ref": evidence_ref,
+    }]
+    payload["risks_categorized"] = [{
+        "category": "Demand",
+        "text": "China demand may weaken.",
+        "source": "10-Q",
+        "reliability": "HIGH",
+        "is_new_this_filing": False,
+        "evidence_snippet": "China demand may weaken.",
+        "evidence_ref": evidence_ref,
+    }]
+
+    brief = BriefOutput.model_validate(payload)
+
+    assert brief.bear_points[0].evidence_ref.content_hash == truncated_hash
+    assert brief.risks_categorized[0].evidence_ref.content_hash == truncated_hash
+
+
 def test_brief_output_accepts_metrics_sources_for_risk_and_guidance():
     brief = _minimal_brief(
         risks_categorized=[dict(