I think it's half #1, half #2.
When a singer (especially female) sings tenderly, some whispering (breathy) feature tends to be blended into the voice, resulting in incomplete voicing. It's a universal phenomenon. The whole phrase in your sound clip has underlying breathing, so in some ways you're true, these are not true voiced consonants.
What you hear at で or ぼ is what they call slack voice. Your vocal cords are set in the position where you pronounce voiced consonants, but the vibration is less than perfect, making a sound half voiced, half unvoiced. If you only focus on VOT (or, if there is vibration during plosion), you may think it's unvoiced, but you must also care about the sound quality, so that you'll notice it's different (much "thicker") than ordinary unvoiced consonant, namely か in your clip.
What's more, the singer properly tries to maintain glottal vibration till right before those consonants, which makes the consonants of で and ぼ effectively sound like //dt// and //bp//, unlike か's is just //kʰ//. But when it comes to this point, で's "voicedness" is certainly quite weaker than ぼ, as you observed.
cons. voice offset voice onset
で 80ms 50-70ms?
か 200ms 100ms
ぼ 10ms 10ms
(shoddily analyzed with Praat)