fix: normalize signal_type case-insensitively + fix subtext typo

Lowercase before canonical check in _normalize_signal_type so mixed-case
LLM values like 'Accounting_Quality' resolve correctly instead of falling
through to the 'language_drift' fallback. Also fix 'subtexte' → 'subtext'
in SubtextRead docstring and BriefOutput.between_the_lines Field description.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

agent/schemas.py

@@ -374,14 +374,14 @@ def _normalize_signal_type(v: object) -> object:
     """
     if not isinstance(v, str):
         return v
-    s = v.strip()
+    s = v.strip().lower()           # lowercase first — canonicals are all lowercase
     if s in _CANONICAL_SIGNAL_TYPES:
         return s
-    mapped = _SIGNAL_TYPE_ALIASES.get(s.lower())
+    mapped = _SIGNAL_TYPE_ALIASES.get(s)   # s is already lowercased
     if mapped:
-        print(f"[signal-type] coercing '{s}' to '{mapped}'", file=sys.stderr)
+        print(f"[signal-type] coercing '{v}' to '{mapped}'", file=sys.stderr)
         return mapped
-    print(f"[signal-type] unknown signal_type '{s}', falling back to 'language_drift'", file=sys.stderr)
+    print(f"[signal-type] unknown signal_type '{v}', falling back to 'language_drift'", file=sys.stderr)
     return "language_drift"
 
 
@@ -662,7 +662,7 @@ class EarningsQualitySignal(BaseModel):
 
 
 class SubtextRead(BaseModel):
-    """A 'reading between the lines' item: surface observation → expert subtexte → implication."""
+    """A 'reading between the lines' item: surface observation → expert subtext → implication."""
     model_config = ConfigDict(extra="ignore")
 
     observation: str = Field(
@@ -717,7 +717,7 @@ class BriefOutput(BaseModel):
     between_the_lines: list[SubtextRead] = Field(
         default_factory=list,
         description=(
-            "0-3 expert 'between the lines' readings — surface observation + subtexte + implication. "
+            "0-3 expert 'between the lines' readings — surface observation + subtext + implication. "
             "EMPTY LIST IS VALID — never manufacture readings. Only emit items anchored to a "
             "precomputed edge signal or a verbatim cross-referenced evidence snippet."
         )