Create a Murmur Grafana operations dashboard

Create an importable Grafana JSON file for Murmur operations. The user confirms observability is deployed and connected and explicitly clarified that no Grafana write is needed. Save the deliverable in ~/Downloads and let the import UI select the Prometheus datasource. Scope: - Use a standard Grafana datasource import input and dynamic job/instance/worker selectors derived from the known Murmur metric schema; do not guess datasource UIDs or the user's scrape job name. Live metadata access is unavailable because Grafana requires login and the configured credential helper returns no credential. - Show scrape health, fleet state and active work, report freshness, throughput/failures, processing time and per-worker CPU/RAM/GPU/VRAM/runtime details. - Treat a healthy zero-worker fleet as idle; do not mask scrape failures or missing GPU telemetry as healthy zero usage. Account for per-worker counter resets and ephemeral series. - Save readable, credential-free dashboard JSON only. Do not change Grafana, the repository, datasources, alerts or services. Verified deliverable: - Created /home/mathias/Downloads/murmur-grafana-dashboard.json (71,043 bytes), with a standard Prometheus import input, 33 panels including section headers, and dynamic job/instance/worker selectors. - Overview distinguishes healthy idle, active fleet, failed scrape and missing/stale reports. Includes selected-range success/failure/processing estimates, attempt rate/duration, GPU utilization/VRAM/telemetry availability, host CPU/RAM, model readiness/load time, process uptime and a current reporting table. - Offline validation passes: JSON/import datasource binding; unique panel IDs/non-overlapping layout; all 30 PromQL targets and 3 variable selectors parsed, including multi-select hostnames with escaped dots; all metric names match the exporter; credential/private-metadata safety checks. - All job/instance selection expands discovered Murmur options rather than matching unrelated Prometheus targets. GPU missing data is not replaced with zero; idle-only zero fallbacks require explicit fleet counts and a healthy scrape. - No Grafana write, remote service change or repository change was made. Live Grafana rendering/import and live Prometheus results remain untested; the user will import the file and select the datasource. Acceptance criteria: - Importable file has a Prometheus datasource picker, useful units/legends, time/worker selectors and clear scaled-to-zero/missing-data behavior. - Validate the JSON structure and parse every PromQL expression after variable substitution; worker-detail panels may be empty when no worker is active and must explain this. - Record that live Prometheus results and Grafana rendering were not validated; do not imply live installation. - No secrets enter the dashboard or repository and no GPU jobs are provisioned solely for validation.