Government-imposed platform shutdowns fragment political discourse across languages, making it harder to detect the propagation of toxic threats. We analyse 172,059 tweets from the 2025 Nepal GenZ protests across five crisis phases using Detoxify multilingual scoring across seven toxicity dimensions, seven machine learning classifier families, and conversation-tree reconstruction across three generational ancestry levels (Parent, Grandparent, Great-Grandparent). Two findings emerge: cross-lingual reply chains selectively amplify threatening language ($d = +0.145$, $p < 0.001$), the only dimension escalating across language boundaries, while de-escalating interpersonal toxicity, partially disconfirming social identity theory and revealing a covert threat vector masked by aggregate de-escalation. Second, Immediate Parent features dominate toxicity prediction across active protest phases (F1 $= 97$-$98%$), shifting to Great-Grandparent dominance post-crisis (F1 $= 95%$), confirming post-shutdown discourse is anchored in deep ancestral frames rather than proximal provocation. Both findings advance the detection and characterisation of cyber social threats in multilingual crisis discourse.

PDF