CLAUDE_CODE_PROPAGATE_TRACEPARENT=1 does not inject traceparent header into outbound LLM requests

## Problem

When `CLAUDE_CODE_PROPAGATE_TRACEPARENT=1` is set, Claude Code emits client-side OTLP spans (`claude_code.interaction`, `claude_code.llm_request`) but does not inject a W3C `traceparent` header into the outbound HTTP requests to the LLM API. This means any OTLP-capable server (an API gateway, an observability proxy, etc.) creates its own trace rather than becoming a child span of Claude Code's client trace.

## Observed behavior

- `CLAUDE_CODE_ENABLE_TELEMETRY=1` + `CLAUDE_CODE_PROPAGATE_TRACEPARENT=1` + `OTEL_TRACES_EXPORTER=otlp`
- Client spans export correctly via OTLP
- Server spans from the upstream API have a different `trace_id` than the client spans
- No `traceparent` header appears on the outbound HTTP request (verified by checking `http.request.header.traceparent` on the server span — absent)

## Expected behavior

With `CLAUDE_CODE_PROPAGATE_TRACEPARENT=1`, the outbound HTTP request to the LLM API should carry a `traceparent` header derived from the active span context, per the [W3C Trace Context spec](https://www.w3.org/TR/trace-context/). The server then creates a child span linked to the client's trace, producing a single connected trace.

## Why this matters

Without traceparent propagation, client and server spans are disjoint traces. The only way to correlate them is by content + timing, which is fragile and breaks in concurrent workloads. This is the standard mechanism for distributed tracing across service boundaries, and the flag name suggests it's intended to work.

Start by locating the outbound LLM HTTP request path and the handling for CLAUDE_CODE_PROPAGATE_TRACEPARENT, then inspect how the active OTLP span context is created and exported. Verify behavior with CLAUDE_CODE_ENABLE_TELEMETRY=1, CLAUDE_CODE_PROPAGATE_TRACEPARENT=1, and OTEL_TRACES_EXPORTER=otlp; done means the request carries a W3C traceparent header and server spans share the client trace_id.

Written by the indexing model from the issue text.