forge: tolerant JSON parsing — extract first object even if Sonnet

appends extra prose

Job 3 hit one recipe (healthy-chicken-stir-fry-with-vegetables) where
Sonnet returned a valid JSON object + appended prose afterward, in
violation of 'no prose' rule. json.loads choked with 'Extra data:
line 54'. _parse_json_blob now falls back to JSONDecoder.raw_decode
which extracts the first complete JSON value and ignores anything
after — any trailing notes / fence remnants / inline commentary get
silently dropped.

Plain json.loads is still tried first (fastest path on clean output).
The fallback only kicks in for malformed-but-recoverable cases.

cauldron/forge.py

@@ -1204,11 +1204,28 @@ def _extract_plan_payload(forge_result: dict) -> tuple[list, str]:
 
 
 def _parse_json_blob(s: str):
+    """Parse the FIRST balanced JSON value out of a string. Tolerates Sonnet
+    appending extra prose/notes after the JSON object (which violates the
+    'no prose' rule but happens occasionally). Also strips ```json fences."""
     s = s.strip()
     # Strip code fences if Sonnet wrapped its output
     s = re.sub(r"^```(?:json)?\s*", "", s)
     s = re.sub(r"\s*```$", "", s)
     try:
+        # Plain decode first — fastest path when output is clean
         return json.loads(s)
+    except Exception:
+        pass
+    # Fall back to raw_decode which extracts the first JSON value and
+    # tells us where it ended. Anything after gets ignored. Handles the
+    # "Extra data: line 54" failure mode where Sonnet appended notes.
+    try:
+        decoder = json.JSONDecoder()
+        # Skip any leading whitespace before scanning
+        idx = 0
+        while idx < len(s) and s[idx] in " \t\n\r":
+            idx += 1
+        obj, _end = decoder.raw_decode(s[idx:])
+        return obj
     except Exception as e:
         raise ForgeError(f"could not parse model JSON: {e}; head={s[:200]!r}") from e