Some shipped recipes and cloud stacks leave built-in metrics empty #524

What happened?

Some configurations Modelplane ships leave built-in modelplane_* metrics empty, with no error. The collector drops everything it didn't rename, so when a source metric isn't emitted, the modelplane_* name just never appears at the sink. In each case below Modelplane's own manifests or stacks produce the gap.

Engine recipes that don't emit the mapped names

• The llama-3.1-8b recipe pins vllm/vllm-openai:v0.7.3 (model-deployment.yaml#L47). That version runs the V0 engine by default (envs.py#L503-L504), which publishes vllm:gpu_cache_usage_perc and prefix-cache hit-rate gauges (engine/metrics.py#L85-L106) instead of the names in metrics.py#L54-L57. modelplane_kv_cache_utilization_ratio, modelplane_prefix_cache_hits_total and modelplane_prefix_cache_lookups_total are empty for it. Every other pinned vLLM version, v0.9.2 and later, publishes all of them.
• Neither SGLang recipe passes --enable-metrics: laguna/model-deployment-sglang.yaml#L46-L57 (v0.5.12.post1) and qwen3-coder/model-deployment-fp8.yaml#L46-L61 (v0.5.10.post1). SGLang only mounts /metrics with that flag (http_server.py#L302-L304), so every _SGLANG name is empty for both. The qwen3-coder recipe also lacks --collect-tokens-histogram, which v0.5.10 needs for the token histograms. telemetry.md#L240-L245 tells users SGLang needs both flags and that "vLLM needs nothing", which isn't true for v0.7.3.

Cloud stacks that install no GPU exporter

• Civo turns the exporter off: civo.py#L280 sets dcgmExporter.enabled: False because its ServiceMonitor needs the Prometheus CRDs (#L268-L272). Every modelplane_gpu_* series and modelplane_energy_joules_total is empty on Civo.

#519 covers the same GPU gap for source: Existing clusters, where Modelplane installs nothing. This is about clusters Modelplane builds.

How can we reproduce it?

1. Deploy the llama-3.1-8b recipe, or either SGLang recipe, with a TelemetryDestination configured.
2. Send it some requests.
3. Query the sink for modelplane_kv_cache_utilization_ratio (llama) or modelplane_requests_running (SGLang). No series exist for that deployment.

For GPUs, provision an inference cluster on Civo, deploy any model, and query modelplane_gpu_memory_used_bytes.

What environment did it happen in?

Modelplane version: main at 0ee6ab4 (also v0.5.0)
Engines: vLLM v0.7.3, SGLang v0.5.10.post1 and v0.5.12.post1, as pinned in the recipes
Clouds: Civo (from the stack code)
Found by reading the code and upstream source at the pinned versions, not on a live cluster