fix: recognize currency-formatted numeric policy arguments

docs/SYSTEM_DESIGN.md

@@ -282,6 +282,7 @@ Context supplies deterministic facts (e.g. `customer.email`). No LLM calls.
 `ArgConstraint` supports deterministic matching over tool arguments:
 - **Equality / inequality:** `eq`, `ne`
 - **Numeric comparisons:** `gt`, `gte`, `lt`, `lte` (supports numeric string coercion per Issue #42)
+  also accept a leading `$` and comma-separated three-digit thousands groups, e.g. `$250` or `1,000.50`; malformed grouping and currency suffixes remain non-numeric. Equality and `$context.` references are unchanged.
 - **String patterns:** `contains`, `startswith`, `endswith` (Issue #23)
 
 Multiple operators on the same constraint are combined with logical AND: all present operators must evaluate to True. Values may reference context via `$context.path.to.value`. If an argument or expected pattern is not a string (or missing), string operator evaluation fails closed (`False`).

packages/mutiny_core/src/mutiny_core/policy/constraints.py

@@ -15,6 +15,7 @@
 
 from __future__ import annotations
 
+import re
 from typing import Any
 
 from pydantic import BaseModel, model_validator
@@ -84,6 +85,13 @@ def _as_number(value: Any) -> int | float | None:
         return value
     if isinstance(value, str):
         s = value.strip()
+        if s.startswith("$") or "," in s:
+            s = s.removeprefix("$")
+            # Validate grouping before removing separators: "1,00" is not 100.
+            # This also keeps $context references and currency suffixes non-numeric.
+            if not re.fullmatch(r"[+-]?(?:[0-9]{1,3}(?:,[0-9]{3})+|[0-9]+)(?:\.[0-9]+)?(?:[eE][+-]?[0-9]+)?", s):
+                return None
+            s = s.replace(",", "")
         if not s or s.lower() in ("nan", "inf", "-inf", "+inf", "infinity", "-infinity", "+infinity"):
             return None
         try:

tests/unit/test_policy_evaluator.py

@@ -747,6 +747,44 @@ def test_evidence_includes_arguments_and_rule(self):
 class TestNumericStringConstraintMatching:
     """Issue #42: coerce numeric-looking string args for inequality constraints."""
 
+    @pytest.mark.parametrize("actual,number", [
+        ("$250", 250), ("1,000.50", 1000.5), (" $1,000.50 ", 1000.5),
+        ("$-250", -250), ("-1,000", -1000), ("$0", 0),
+    ])
+    @pytest.mark.parametrize("operator", ["gt", "gte", "lt", "lte"])
+    def test_formatted_numbers_match_numeric_boundaries(self, actual, number, operator):
+        from mutiny_core.policy.constraints import matches_constraint
+
+        for boundary in (number - 1, number, number + 1):
+            constraint = ArgConstraint(**{operator: boundary})
+            assert matches_constraint(actual, constraint, context={}) == matches_constraint(
+                number, constraint, context={}
+            )
+
+    @pytest.mark.parametrize("actual", [
+        "$", "$$250", "$context.amount", "$nan", "$inf", "1,00", "1,,000",
+        ",250", "250,", "1,000.5,0", "250$", "$250usd",
+    ])
+    def test_malformed_formatted_numbers_do_not_match(self, actual):
+        from mutiny_core.policy.constraints import matches_constraint
+
+        assert not matches_constraint(actual, ArgConstraint(gt=-10000), context={"amount": 250})
+
+    @pytest.mark.parametrize("amount,violated", [("$250", True), ("1,000.50", True), ("$150", False)])
+    def test_formatted_amounts_reach_refund_policy(self, amount, violated):
+        hits = PolicyEvaluator().evaluate(
+            _refund_limit_policy(),
+            _trace(_tool("issue_refund", {"order_id": "o1", "amount": amount, "approved": False})),
+            {},
+        )
+        assert _hit_for(hits, "refund_limit").violated is violated
+
+    def test_context_references_and_equality_are_unchanged(self):
+        from mutiny_core.policy.constraints import matches_constraint
+
+        assert matches_constraint("$250", ArgConstraint(eq="$context.amount"), context={"amount": "$250"})
+        assert not matches_constraint("$250", ArgConstraint(eq=250), context={})
+
     @pytest.mark.parametrize(
         ("actual", "constraint", "expected"),
         [