The amplification bug, fixed

What happened. The replacement room tone was being level-matched to the audio immediately beside each region. Once the regions were widened to ten seconds either side, what sits beside them is usually speech — so the quiet room tone was being amplified up toward talking level. It was filling the gap with something louder than the rustle.

The target is now the quiet quarter of the room channel across the surrounding minute, with treated regions excluded, and the fill can never be raised above the donor's own level. Checked across all 30 spans in four episodes: nothing gets louder, where two of the first eight did before.

Where it stands honestly. The noise itself now falls by 3 to 24 dB, median 11 dB on episodes 1 and 6 and about 4 to 6 dB on 2 and 5. That is real but uneven, and I do not yet know why some spans move so little. The most likely explanation is that the separator puts part of the rustle into the voice channel, which is never touched — if so there is a ceiling here that no amount of work on the room channel will pass.

And your two sound-effect answers matter more than any of this. You marked episode 1 numbers 3 and 5 as you making a sound effect, not a defect. The detector cannot tell those apart, and treating your performance as noise is worse than leaving a rustle in. Keep marking them — that is the signal I need most.