fix(sweeper): key threshold cache by TTLPolicy value instead of id()

[35C4n0r]

• id(policy) returns the CPython memory address, which is reused as soon as the previous policy object is freed; within a single _plan() loop a policy_resolver that creates a fresh TTLPolicy per thread can produce two different policies with the same id(), causing the cache to return thresholds computed for the wrong policy
• TTLPolicy is a frozen dataclass so it has hash and eq based on its fields; using the policy object itself as the dict key is correct and safe regardless of object identity