[Feature] Expose developer (system) prompt as a separate observation in traces

Description:

Summary

The developer role message in the rollout (the system prompt / instructions) is currently silently dropped by the parser. This is the single largest token consumer in most Codex sessions — often 10k–20k+ tokens, yet it's invisible in Langfuse. I can only infer its size from the gap between input_tokens on each generation and the visible user prompt.

Motivation

When debugging cost or prompt quality, I need to see what is being sent as the system prompt, not just that it's expensive. Common use cases:

• Verifying that AGENTS.md / <environment_context> content is being injected as expected
• Comparing system prompt size across repos or configurations
• Auditing whether instructions are bloating token usage

Current rollout structure (per #47)

Proposal

Export the developer-role message as one of:

1. A dedicated span/observation (e.g. type span, name system_prompt) attached to the turn, with the full text as its input — or
2. Part of the generation's input (prefixed, clearly delimited) so it's visible on each LLM call.

Option 1 is preferred — it keeps the generation input clean and lets Langfuse show it as a collapsible sibling.

Implementation sketch

In parse.ts, when iterating rollout rows, capture payload.role === "developer" messages (currently skipped) and emit them as a span before the first generation of the turn:

// pseudo
if (p.role === "developer" && p.type === "message") {
  const sysSpan = startObservation({
    name: "system_prompt",
    type: "span",
    input: p.content,
    // attach to current turn
  });
  sysSpan.end();
}

No additional rollout data is needed, as the content is already parsed, just not exported.

Environment

• Plugin: tracing@codex-observability-plugin 0.3.0
• Codex: 26.904.11930
• Langfuse Cloud EU