fix(compaction): use provider-reported usage for triggers #243

i noticed one of my loger-term agents was compacting often, at or around 20% reported context. as it turns out, autocompaction can use total bytes / 4 as a rough metric of token usage, which is a useful approximation if the context in question is unencrypted. since new frontier models use encrypted thinking blocks, this metric is no longer accurate and it may not have ever been useful in the first place