StreamReplay

How SemanticCache.getOrPutStreaming replays a cached answer on a hit.

A cache hit on a streaming path has to decide what a "stream" of an answer that is already written down even means, and the honest answers are few. This enum exists so the decision is made in the API, by the caller, rather than falling out of an implementation detail.

What is not here, and why

There is no option that reproduces the original timing. Doing it would mean recording how long each chunk took to arrive from the provider, figures that describe one model call, on one network, on one day, and then sleeping through them to serve an answer the cache already has in memory. That is a cache made deliberately slower in order to look like the thing it replaced, and the whole argument for this library is that it does not dress one measurement up as another. A hit is fast. Letting it look slow would be the pretence, not the honesty.

If a product genuinely needs a typewriter effect, that is a presentation decision and it belongs in the presentation layer, where it can be tuned, disabled and tested. It does not belong baked into a cache and paid for in stored bytes on every entry.