SystemsLab 161.0.0

Added

• Release builds now publish .rpm packages for Rocky Linux 10 (el10).
• The context page's summary table can now be downloaded as a CSV file via a button next to the Summary heading. Each row is an experiment, with its varying parameters, one column per KPI (labelled name (unit), holding the raw numeric value), and its state.
• A new systemslab infra command group — status, log, pricing, requests, estimate, approve, and deny — lets an operator inspect the EC2 fleet and manage instance requests. approve/deny resolve a request parked under [infra] allocation = "manual", which previously failed at server startup since nothing could approve a parked request.
• systemslab infra estimate forecasts an experiment's on-demand/spot cost from run history and current prices, and systemslab infra recommend ranks instance types by expected cost per successful run. systemslab submit --budget <usd> (requires --host-count) runs the same estimate and refuses to submit when the on-demand-high figure exceeds the budget or the cost is unknown, unless --yes overrides.
• systemslab context new gains --max-instances, --pin-hosts, and --no-autoscale flags to set a context's infra scheduling policy at creation time; there is currently no command to change them afterward.
• A rezolus_addr option in agent.toml lets an operator point an agent's metrics recording at a different host's rezolus endpoint, for a workload that actually runs in a VM, container, or separate target the job only orchestrates. Leaving it unset keeps the existing behaviour of scraping the local machine.

Changed

• The metrics dashboard shown on experiment and context artifact pages now matches Rezolus's own viewer: a paired cgroup selector, a GPU selector, a PromQL Query Explorer tab, a regrouped nav, and "Charts with no data" instead of blank cards for a sampler the recording lacks. The old vllm-metrics, sgl-metrics, llm-perf, rpc-perf, and metrics-v2 URLs still work and redirect to the unified page. This also fixes opening an untagged workload recording sometimes rendering a different job's Rezolus data under it.

Fixed

• An agent whose host the server had dropped never came back. The agent registered once, at startup, and was passive after that, so nothing told it the server had forgotten it: the host stayed missing until someone restarted the agent, and experiments needing it queued indefinitely with no error surfaced anywhere. Agents now check that the server still has a host for them every registration_check_interval seconds (default 600, 0 to disable) and register again, under a new host id, when it does not. (#6079)
• Fixed /api/v1/artifact ignoring the experiment, context, and report filters for artifacts that are not attached to that kind of entity, so a query such as ?experiment=<id> also returned every context artifact in the installation. This made systemslab artifact list --experiment <id> and systemslab artifact download-all <id> report and download artifacts belonging to unrelated experiments. (#6082)
• Fixed a bug where a recording composed from multiple workloads (e.g. rezolus + cachecannon) could silently lose one workload's dashboard section, and sped up dashboard loads that were taking several seconds.
• Fixed several infra manager accounting bugs that could let fast, high-throughput jobs evade a context's max_instances cap, skew cost estimates after a server restart, or overcharge a drained machine.