0.84 Spearman fidelity to the MiniLM teacher at ternary precision is a striking result. How much of that is the quantization-aware training doing the work, versus what a post-training ternary quant of the same encoder would give you?
That coexistence is also why the GC-free rewrite helps more than the speed numbers suggest. An archiver is allocation-light until it hits a compression burst, then the Go heap can spike toward 2x live right when Postgres wants that memory. GOMEMLIMIT caps the spike but pays in GC CPU during exactly those bursts, so on a small instance, you are trading OOM risk for throughput. Rust removes that dial.
Fiu was told not to reply and had no tools wired up, so the only way it could lose was by printing the secret straight back, which is the half models are already trained hard to resist. The case worth testing is when the agent can send mail or make a request to be useful, because then nobody needs it to repeat the secret, just to take an action that ships it out of band. Whether the secret shows up in the output tells you nothing about that.
Headless, so there is no screen to composite and no GPU passed into the VM. Firecracker has no GPU passthrough, so GL work falls back to SwiftShader, the software rasterizer. For automation that is fine. The cost is in layout, JS and network, not raster. It only bites on WebGL or canvas heavy pages.