The live in-flight cap. Always within [Min, Max].
Upper bound the cap can never rise above.
Lower bound the cap can never fall below.
Stop the cap from GROWING, without lowering it.
The cheapest possible response to memory pressure: everything already in flight keeps running at full speed and nothing new is admitted, so it costs no throughput at all. Without it the only options are "carry on" and "halve", and halving work that is already succeeding is a real cost paid for a problem that holding would have solved.
Additive increase: a healthy outcome nudges the cap up by 1, clamped to Max.
Multiplicative decrease: a throttle/error halves the cap, floored at Min.
Allow ramp-up again once pressure has cleared.
Holds a single adaptive concurrency cap and mutates it under the AIMD law. Stateless beyond the current cap and its bounds; safe to share across a single
RunAdaptiveinvocation (all calls run on one async thread, so no locking).