concurrency

type InFlightLoad struct {
	// Tokens is the in-flight token count this endpoint has committed to,
	// accumulated from past scheduling decisions. Updated by PreRequest (when
	// an endpoint is chosen) and OnEvicted (when its request stream ends);
	// snapshotted into this struct each cycle by InFlightLoadProducer.Produce.
	Tokens int64

	// Requests is the in-flight request count this endpoint has committed to,
	// maintained with the same lifecycle as Tokens.
	Requests int64
}

type UncachedRequestTokens struct {
	// Tokens is a speculative projection: the work the request currently being
	// scheduled would add to this endpoint if it landed here. Includes the
	// uncached input portion (accounting for prefix-cache hits) plus the
	// estimated output when the producer is configured with
	// AddEstimatedOutputTokens=true. Computed fresh by Produce on every cycle
	// from the request being scored and this endpoint's prefix-cache state;
	// never committed to endpoint state and not decremented on stream end.
	// Zero when no request is in scope (e.g., background snapshots).
	Tokens int64
}