feat(telemetry): adopt OTel GenAI semantic convention names for spans and metrics #1209

Mellea's OpenTelemetry instrumentation uses `gen_ai.*` span attributes, which is correct and ecosystem-aligned. However, the span name and all metric names deviate from the OTel GenAI semantic conventions, which means off-the-shelf GenAI observability tooling — Grafana GenAI dashboards, OTel Collector semantic processors, vendor-provided GenAI monitors — does not recognise mellea's signals out of the box. Users get data into their backend but none of the pre-built interpretability tooling lights up, and they cannot use community-maintained dashboards without rewriting every metric reference.

This issue tracks the full alignment checklist. Items are additive on the existing cpex hook/plugin architecture; none require structural changes to the backend or session APIs. M1/M2/M3 are breaking changes for existing dashboards and need a migration note.

Spans

• S1 Span name "chat" → "chat {gen_ai.request.model}" (SHOULD per semconv — enables model-level trace filtering without attribute filters)
• S2 Span kind should be CLIENT for remote-backend calls, not INTERNAL (SHOULD)
• S3 server.address / server.port not set on any span (Conditionally Required)
• S5 gen_ai.request.stream not set; mellea.streaming used instead (Conditionally Required)
• S6 gen_ai.provider.name enum values must match semconv (ibm.watsonx.ai, not "watsonx") (MUST where set)
• S8 gen_ai.output.type not set when structured output is requested; mellea.has_format used instead (Conditionally Required)
• S11 gen_ai.response.time_to_first_chunk missing as a span attribute; currently recorded only as a metric (Recommended)
• S9 gen_ai.usage.total_tokens is not in the spec; remove or document as a mellea extension
• S10 mellea.streaming / mellea.format_type duplicate information now in semconv attributes; clean up

Metrics

• M1 Token usage: two Counter instruments mellea.llm.tokens.input / mellea.llm.tokens.output → single Histogram gen_ai.client.token.usage with gen_ai.token.type dimension ("input" / "output") and semconv-specified bucket boundaries ⚠️ breaking change
• M2 Duration: mellea.llm.request.duration → gen_ai.client.operation.duration ⚠️ breaking change
• M3 TTFB: mellea.llm.ttfb → gen_ai.client.operation.time_to_first_chunk ⚠️ breaking change
• M4 gen_ai.operation.name dimension missing from all client metrics (Required dimension per spec)
• M5 server.address / server.port missing from all metrics (Recommended / Conditionally Required)
• M6 error.type missing on the duration metric for failed operations (Conditionally Required)
• M8 streaming dimension on duration is non-standard; superseded by gen_ai.operation.name + semconv approach
• M9 gen_ai.client.operation.time_per_output_chunk not implemented (Recommended)
• M10 error_type vs error.type key inconsistency on the error counter (cleanup)

Migration note (M1/M2/M3)

Renaming the three primary metrics is a breaking change for any existing dashboard or alert rule targeting the mellea.llm.* names. A migration plan (dual-emit period, deprecation notice, version bump) should accompany those three items.

Already tracked elsewhere (not duplicated here)

• gen_ai.request.* parameters (temperature, max_tokens, etc.) → feat(telemetry): emit gen_ai.request.* from tracing plugin (post-#1045) #1068
• Chat content events (gen_ai.content.prompt / gen_ai.content.completion) → feat(telemetry): emit gen_ai.input/output.messages content events from tracing plugin (post-#1045) #1073
• Prompt template attributes → feat(telemetry): emit llm.prompt_template.* from tracing plugin (post-#1045) #1067
• execute_tool span → feat: tool call spans via tool_post_invoke hook #1049
• Metrics plugins on generation_batch_* hooks → refactor: subscribe metrics plugins to generation_batch_* hooks #1182