[FEATURE]: propagate W3C traceparent on outbound LLM HTTP requests

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before. Related but different: #46856 asks for W3C trace context in MCP tools/call _meta (per SEP-414); this request is about the outbound HTTP requests to LLM providers (OpenAI, Anthropic, Azure, Copilot, etc.), which is a different surface. #18801 and #28295 were closed for inactivity and covered MCP HTTP/SSE, not LLM inference calls.

Describe the enhancement

OpenCode's outbound LLM HTTP requests do not carry a W3C traceparent header. When OpenCode is used behind a tracing-capable gateway or proxy, the gateway creates server-side spans for each LLM request, but these spans start a new trace because the inbound HTTP request has no trace context. Client-side and server-side spans end up as disjoint traces that can only be correlated by content and timing.

Suggested implementation

OpenCode already has full OTel infrastructure in packages/core/src/observability/otlp.ts: @opentelemetry/api, AsyncLocalStorageContextManager, and the OTLP HTTP trace exporter. The missing piece is injecting the active span context into outbound LLM requests.

Injection point: packages/llm/src/route/executor.ts, in the layer definition (~line 370). Before http.execute(request), inject the traceparent header from the active OTel context:

import { context, propagation } from "@opentelemetry/api"

// In executeOnce, before http.execute:
const carrier: Record<string, string> = {}
propagation.inject(context.active(), carrier)
const traced = Object.entries(carrier).reduce(
  (req, [k, v]) => HttpClientRequest.setHeader(req, k, v),
  request,
)
return yield* http.execute(traced)...

This requires adding @opentelemetry/api as a dependency to the llm package (it's already in the monorepo via packages/core).

Behavior when tracing is not configured: propagation.inject() is a no-op when no propagator is registered (no TracerProvider set up), so no headers are added. Zero behavioral change for users without tracing.

Verified end-to-end

Tested the equivalent fix (W3C traceparent injection via SDK middleware) against a live OTLP-capable gateway. With propagation, the gateway's server span becomes a child of the client's span in a single connected trace:

rag-pipeline (root)
|- retrieve (parent: rag-pipeline)
|- generate (parent: rag-pipeline)
|   \-- chat openai-gpt-5.4 [SERVER] (parent: generate)  <-- gateway span

Without it, the gateway span is in a separate trace.