Simulate context policies with and without caching

Everything rests on one rule, implemented in Context.send: compare this request's blocks with the previous request's, read the unchanged run at the start, write the rest. Policies only edit the block list before a request; they never touch prices. That separation is the decision that matters most: once the cache rule is right, clearing, compaction and selection each become a few lines, and every cost they cause or save, including the rewrite after an edit, falls out of the same comparison. No model is called.