Observability: evaluation hooks (LLM-as-judge, human annotation, custom scorers) #30

Split from #4 (LLM Observability Service). Depends on the base Phoenix tracing slice landing first.

Score traces to track quality over time, not just latency/cost:

• LLM-as-judge scorers (needs a design decision: which model judges — the same llama-server instance, a separate judge model/config, or admin-configurable per the existing LLM-provider-config pattern in app.services.system_settings_service)
• Human annotation (needs a UI surface — Phoenix has one natively; decide whether we expose it as-is or build an in-app equivalent)
• Custom scorers (deterministic, code-based checks — e.g. "did the answer cite a source when the tool fired")

Needs its own design pass: this is a genuinely separate feature with its own config surface (judge model selection), not just "more instrumentation." Do not bundle into the base tracing PR.

Out of scope for this issue: the base Phoenix service, OTel instrumentation of the chat agent, and trace search/clustering (see the separate issue for that).