[mcollina]

Adds a self-contained, reproducible throughput benchmark harness under benchmarks/, plus the optional comparison libraries, npm scripts, and docs.

What's included

• benchmarks/bench.js — head-to-head encode/decode vs msgpackr, msgpack-lite, @msgpack/msgpack, and JSON, across the generated datasets. Each function runs under a fixed wall-clock budget (default 4000 ms, override BENCH_MS), reports ops/sec and ratios relative to msgpack5, and verifies every round-trip against the source JSON. Missing comparison libraries are skipped.
• benchmarks/stream.js — streaming .decoder() throughput vs msgpackr's UnpackrStream.
• benchmarks/ab.js — A/B harness comparing two local msgpack5 checkouts via AB_OLD/AB_NEW (interleaved rounds, median ops/sec, new/old ratio).
• benchmarks/gen-data.js — deterministic generator for the datasets (sample-large, example4, example5, mixed), written to the git-ignored benchmarks/data/.
• benchmarks/data/ is git-ignored (generated on demand by gen-data.js/bench.js).
• Comparison libraries added as devDependencies; npm run bench, bench:stream, bench:ab, bench:gen scripts; README "Benchmarks" section; AGENTS.md updated (decoder description and benchmark commands).

The existing benchmarks/encodedecode.js/parseshortmap.js are left as-is.

Notes

• benchmarks/ is linted by standard as part of npm test; the new scripts follow the existing project style. All 4391 tests still pass on Node 22/24/26.
• The A/B harness (bench:ab) just points AB_OLD/AB_NEW at two checkouts, so it works with any two local builds (no sandbox/symlink setup required).
• msgpackr's msgpackr-extract is an optional native dependency; if it can't be built, msgpackr still loads (pure-JS fallback) and the harness keeps working.