Add OpenTelemetry-style spans

Apple does not ship an OpenTelemetry exporter for Foundation Models. You can still use the same span idea: a named interval, a parent, and a few attributes. On Apple platforms, OSSignposter is the recorder that Instruments already understands.

import os

enum AITrace {
    static let signposter = OSSignposter(
        subsystem: "dev.edgefde.app",
        category: "ai-pipeline"
    )

    static func span(_ name: StaticString, _ work: () async throws -> T) async rethrows -> T {
        let id = signposter.makeSignpostID()
        let state = signposter.beginInterval(name, id: id)
        defer { signposter.endInterval(name, state) }
        return try await work()
    }
}

// Sketch. Confirm OSSignposter names in the installed SDK.
let reply = try await AITrace.span("ai.request") {
    let packed = try await AITrace.span("ai.context_pack") { packContext() }
    let text = try await AITrace.span("ai.model") { try await session.respond(to: packed) }
    return try await AITrace.span("ai.guardrail") { try applyGuardrail(text) }
}

Use dotted names the way OpenTelemetry semantic conventions do: ai.request, ai.context_pack, ai.tool, ai.model, ai.guardrail. Put the tool name, model name, profile name, and prompt version on attributes or a short signpost message. Do not put the user text there.

If you later export to an OpenTelemetry collector from a debug Mac, map those same names. Production on-device builds should keep signposts cheap and empty of content. A collector is optional. The names are not.

Nest ai.tool under ai.request for each tool. If the session calls two tools, you get two child spans. That matches the Instruments tool-call tree. Pair the ai.model span with clocks from Stream tokens and measure speed. Turn a bad tool path into a sample in Test AI behavior before you ship.

Next, read a bad trace with those names.