[mheffner]

Both exporters built their request bodies by formatting serde_json::Value trees through Display (which goes through a fmt::Write adapter), after first copying already-formatted strings into new Values:

• X-Ray: each segment Value was formatted with to_string(), the strings were copied into json!({"TraceSegmentDocuments": [...]}) and that Value was formatted again.
• EMF: each EMF log line Value was formatted with to_string(), then every event (timestamp + message) was copied into a json! object for the PutLogEvents body, which was formatted again.

In profiles at saturation Value's Display impl was ~20% of CPU in both.

Use serde_json::to_string / to_vec, and serialize the request bodies from small borrowing structs. Field order matches what the Value-based code produced (serde_json objects are key-sorted), so the bodies are byte for byte the same; tests compare against the previous construction.

Benchmark (rotel-bench, Intel N95, mock endpoints, 5+5 runs ABBA, mean +- stddev, vs main):
EMF, 20k points/s: CPU/point -16.4% (10.8 -> 9.0 %/10k/s)
EMF, saturation (1 core): export requests/s 34.7 -> 45.3 (+30.7%),
CPU/point -24.6%
X-Ray, 50k spans/s: CPU/span -7.8% (11.0 -> 10.1 %/10k/s)
X-Ray, saturation: export requests/s 2,454 -> 2,840 (+15.7%),
client p99 363 -> 314 ms
(at saturation both exporters export slower than rotel ingests, so the
export rate is measured at the endpoint)