Episodes 5 and 6 only, using the spans you drew and confirmed. No detector
is involved anywhere in this, so what you are judging is the fix alone.
The settings, as agreed. The voice stem untouched. The real room everywhere
outside the treated regions. Spans widened three seconds either side with 0.3 s
crossfades. Inside them, one ten second stretch of quiet room from the same episode —
the fix for the loop you heard, which came from a half-second tiled about eleven hundred
times.
Your level idea is what made it work. Each replacement is now matched to the room
measured in the three seconds either side of its own span, rather than one figure for the
whole episode. Room tone drifts as the phone moves, and a single figure fitted episode 6
but made three of episode 5's spans louder. There is also a hard clamp: the fill can
never exceed the room inside the span it replaces.
Measured across all eleven spans: every one is quieter, none louder.
Episode 6 down 29 to 46 dB. Episode 5 down 5 to 34 dB. Both masters pass every delivery
check, and the voice stem is bit-identical to the separator's output in both.