Follow a failure through the running system

A trace id is the one handle that joins the three places a failure leaves evidence:

• Tempo: one span per request, joined across services. The control plane's call to the worker is a child span of the browser's request, so one id covers the whole click.
• Loki: every log line either service wrote while serving that request. Each line carries traceId, and a run started by the request keeps writing lines with the same id after the response has gone.
• The container's own log on the host: the same JSONL lines as they reached stdout. Read it over SSH when Loki has nothing: export was off, the collector was down, or it was before tracing shipped.

The design and its limits are in docs/adr/0058-requests-are-traced-into-the-hosts-otel-lgtm.md.

The tools: task obs:*, and nothing else