Class PersistCache

An LRU of materialized plan results keyed by the optimized plan's structural hash. One instance is the process default (what persist() on the default runtime uses); new Runtime({ persist: new PersistCache() }) gives a tenant / request / test its own — nothing it caches is visible to any other runtime, and clear() drops exactly its entries.

Semantics (docs/operations.md → persist): a plan is cached only when every input is immutable — frames built with fromColumns(…, { copy: false }) alias caller memory and are refused — and it calls no UDF unless persist({ trustUdfs: true }). Every entry records which engines produced it; a strict engine request is served from the cache only when that engine produced the entry, otherwise the plan runs (and either the engine executes it or EngineStrictError is raised, exactly as without a cache).