Weaver registry for the agentevals attributes and the evaluation event #244

Why

agentevals sets agentevals.* attributes on SDK session resources and on gen_ai.evaluation.result records (#241). Nothing declares them, so a new key can ship without review and the docs are typed by hand.

Goal

An OpenTelemetry Weaver registry is the contract for every agentevals.* attribute and for what agentevals puts on gen_ai.evaluation.result. Code and registry cannot drift.

Scope

• docs/semconv/registry/: manifest, attributes (including the agentevals.metadata.* template), and the evaluation event with its upstream gen_ai.evaluation.* refs.
• A sidecar file for what Weaver cannot hold: the upstream GenAI semconv pin and the ingest aliases for old keys.
• A unit test that loads the registry and matches it against the emitter and SDK key sets, both ways.
• weaver registry check in CI from the pinned otel/weaver image. If the upstream dependency does not resolve transitively, redeclare the borrowed attributes locally.
• Optional: generate the attribute tables with weaver registry generate.

Out of scope

• live-check of incoming telemetry and the trace quality metric.
• Renaming the resource keys. That is decided in feat: OTel-native refactor #241.

Done when

• CI fails when an emitted key is missing from the registry or a registered key is never emitted.
• The registry passes weaver registry check.