EVA Vision v3.2.0 (Sep 16, 2026)

Expose end-to-end API latency: Add GET /system/latency, which reports per-route latency percentiles and the rate of responses that exceeded the camera-frame latency budget. This makes overload visible even when queued TorchServe requests ultimately return HTTP 200.
Add repeatable setup and test workflows: Add setup.sh for a fresh checkout and a ./run.sh test subcommand that checks prerequisites and runs the API and offline test suites. Add a serve harness and contract suites for detection, sync, health, and TorchServe-to-HTTP error mapping.

Harden service probes and TorchServe memory limits: Make both liveness and readiness probes verify TorchServe as well as FastAPI, make their targets configurable, and cap the TorchServe JVM heap and direct-memory pools through TORCHSERVE_JVM_XMX and TORCHSERVE_JVM_MAX_DIRECT_MEMORY.
Improve operating visibility and load-test support: Apply the request-path logging policy, add a camera-pattern load generator and an operational monitor, and record the measured latency baseline for the design load.
Reduce container image size: Remove unused model-weight formats and archive build context, reducing the runtime image from 33.8 GB to 28.8 GB.
Add development hygiene tooling: Add pre-commit hygiene hooks, Ruff formatting configuration, and a blame-ignore revision list.

Report actual pod memory usage: Use cgroup working-set memory rather than memory.current, preventing cache usage from being reported as a near-OOM condition.