I got this fixed, at least a very suspiciously related: same root cause as the multi-minute “Estimating” / idle-GPU gap before cached TRT: every new ffmpeg worker re-materializes the encrypted .tz3 into a process-local package map. TRT cache hits only after that.
On 1.6.1 / AIEngine 3.8.27 I measured Apollo-class packages at ~161–162 s cold per new worker, then ~250–280 ms on later workers after persisting that materialised map across process boundaries (same tile file). Iris-class control: ~39 s → ~61 ms.
Cache is per tile package. Product fix: durable package map and/or fast bulk decrypt. Concat same-settings jobs remains the zero-risk user workaround.
Full writeup + GPU screenshot:
