Hello Ien,
FluidSynth does some pre-calculations on sample data to determine when the sound becomes "imperceptible" ("reaches the noise floor"). I'm not really familiar with this code and have often wondered how well it works. For example, in the case of Swami, I don't believe I'm pre-calculating these values and I've questioned whether this is working right (in the case of Swami's use of FluidSynth at least). At any rate, if an instrument has a long release time, FluidSynth will try to optimize things to turn off the voice when it is no longer "perceptible". It will also prioritize voices to terminate if it runs out of the allocated polyphony. It seems you are saying that a sound continues past the note off though, which likely has to do with the instrument's release time. Are you certain that FluidSynth is maintaining the voice past the end of the designated release interval?
Best regards,
Element Green