feat: telemetry — tracing spans with token/cost fields (C5)

[yuanhao]

What

Phase C5 — the final Phase C item. The loop now emits structured tracing spans, not just log lines:

agent_loop                (model)
└─ llm_stream             (turn, model, tokens_in, tokens_out, tokens_cached, cost_usd)
└─ tool                   (tool, tool_call_id, is_error)

cargo run --example telemetry shows it live:

INFO agent_loop{model=mock}:llm_stream{turn=1 model=mock tokens_in=0 ...}: close time.busy=347µs
INFO agent_loop{model=mock}: close time.busy=1.32ms

Design

• OTel stays app-side (deviation from the roadmap's literal "feature flag", as discussed): the library emits plain tracing spans and gains no OTel dependency — apps install the tracing-opentelemetry layer and the same spans flow to Datadog/Tempo/Honeycomb/Jaeger. Zero overhead with no subscriber.
• cost_usd recorded from CostConfig when pricing is configured — cost attribution lands in dashboards for free.
• Futures instrumented with .instrument(span); no entered guards across .await.
• tracing-subscriber added as dev-dependency only (test + example).

Tests (1 new, 361 total green)

Capturing Layer + WithSubscriber-attached agent_loop run asserting exact span counts (1× agent_loop, 2× llm_stream, 1× tool).

Verification

clippy -Dwarnings --all-features clean · 361 passed / 0 failed · docs -Dwarnings clean · example smoke-run verified · new mdBook page concepts/telemetry.md with the OTLP bridge snippet.

🤖 Generated with Claude Code