[bengl]

A trace that isn't sampled is dropped as a whole, but microtel still creates a span for every request in it, so the trace context reaches the services downstream and spans made while handling the request have a parent. Those spans were never exported, so the attributes, timestamps, and listeners they collected were wasted.

What changed. Each handler checks the span once, right after creating it, and an unsampled span skips everything only the export reads:

• the attribute work: URL parsing, header reads, and setAttribute() calls, for HTTP/1 servers and clients, HTTP/2 servers, and undici
• the attributes object, and the wall-clock and monotonic timestamps in the Span constructor and in end()
• the span stashed on the response, request, or stream, so the response listeners, status, exception events, and error handling that only record onto the span never run

What propagation and parenting need stays: trace and span IDs, trace flags, tracestate and baggage, removing a stale tracestate from outgoing requests, making the span current for the request handler, and the self-trace check. A sampled span exports the same fields as before, behind one more branch, though it now reads its start time just after its trace context is set up rather than just before. A Span.isSampled getter exposes the decision.

Performance. On a quiet machine, against a copy of main, on Node 26.12.1. microtel's own CPU time in profiles of 50k traced requests at sampleRatio: 0, over 7 interleaved runs:

Client	main	This PR	Change
HTTP	95.6 ms	63.8 ms	-33%
fetch	81.7 ms	55.1 ms	-33%

Sampled requests didn't change: 193.1 ms on main against 191.4 ms, over 8 interleaved runs at the default ratio.

Throughput of the new unsampled benchmarks, over 8 interleaved node:bench runs:

Benchmark	main	This PR	Change
http request (traced, unsampled)	26.45k ops/s	27.54k ops/s	+4.1%
fetch request (traced, unsampled)	19.77k ops/s	20.35k ops/s	+2.9%

Raw samples

• HTTP profile, main: 101.5, 111.5, 93.8, 95.6, 90.1, 88.6, 126.0 ms
• HTTP profile, this PR: 84.0, 57.6, 63.8, 61.4, 63.1, 74.5, 80.8 ms
• fetch profile, main: 81.7, 87.2, 81.9, 77.3, 75.2, 81.7, 77.8 ms
• fetch profile, this PR: 53.5, 49.2, 52.3, 60.3, 55.1, 64.3, 55.1 ms
• sampled HTTP profile, main: 191.6, 195.5, 191.0, 187.4, 201.5, 194.7, 202.2, 185.8 ms
• sampled HTTP profile, this PR: 190.8, 192.1, 185.7, 198.7, 184.5, 185.5, 201.0, 195.7 ms
• http unsampled throughput, main: 26.47k, 26.05k, 27.94k, 26.19k, 26.44k, 26.07k, 28.64k, 28.53k
• http unsampled throughput, this PR: 28.65k, 29.57k, 27.20k, 27.55k, 27.52k, 27.42k, 30.80k, 26.66k
• fetch unsampled throughput, main: 19.20k, 20.78k, 19.71k, 19.81k, 20.20k, 19.64k, 19.74k, 20.81k
• fetch unsampled throughput, this PR: 20.42k, 19.86k, 20.04k, 21.26k, 20.62k, 20.28k, 18.57k, 20.96k

benchmark/http-unsampled.bench.js and benchmark/fetch-unsampled.bench.js are new, so the CI benchmark comparison tracks this configuration from now on. The shared benchmark helper now passes sampleRatio through to init().

Not taken. Lazy attribute or timestamp storage. An unsampled span simply has none of the state the export reads, behind one undefined check in setAttribute() and addEvent(), and every span keeps the same shape.

Tests. A child process at sampleRatio: 0 drives sampled and unsampled requests through an HTTP frontend, fetch, and HTTP/2, and reports what each hop saw. An unsampled request leaves no microtel listener on its responses, the same counts plain Node has, and the HTTP/2 stream and the fetch hold no span for the export-only handlers. A direct test checks that an unsampled span gets no timestamps, attributes, or events, and that end() only marks it ended. Its traceparent, tracestate, and baggage still propagate, replacing a caller's stale ones the same way a sampled request does, and only the request that arrived sampled exports anything.