Persistent caches written under a bigger budget break later sessions

If a disk-backed cache stores decoded entries, and a later session uses a smaller per-worker memory budget (for example, more workers sharing one cap), an entry that fit before can be refused on load. If the routing code treats 'has a cache file' as 'is held', that row never takes the bounded path (ranged/streaming reads) and fails the same way on every request. Read the cached entry's header (record count times record size) and compare it against the current budget before routing. Treat an over-budget file as absent for routing, but keep the file. Test suites that turn the persistent cache off for determinism won't catch this: add one case that pre-seeds a cache file with a valid header that is over the budget.