chore(benchmark): add Redis write benchmark and record the 0.18 to 0.19 measurements

[lan17]

Summary

Adds pnpm benchmark:redis-write, a maintainer benchmark for the Redis write path, and records the measured server-side cost of the 0.18.0 → 0.19.0 write-protocol change (#131). The 0.19.0 release notes carried an expectation extrapolated from #123's read measurements ("expect ~95% server-side command-time reduction at 1 MiB writes"); this PR replaces that expectation with numbers.

Measured: 0.18.0 (Lua write scripts) vs 0.19.0 (native SET + payload-free stamp)

Three full runs per version, sequential awaited writes via the node-redis adapter against a dedicated local Redis 6.2 container. Server-side cost is INFO commandstats µsec summed over the commands the client dispatches per logical write (0.18.0 tracked: EVALSHA; 0.19.0 tracked: SET + EVALSHA; the EVALSHA entry envelopes script-internal calls), divided by writes. Ranges are min–max across the three runs.

Path	Payload	0.18.0 server µs/write	0.19.0 server µs/write	Server-side reduction
tracked	100 B	6.8–8.8	6.7–8.2	parity (within noise)
tracked	10 KiB	12.7–15.7	10.5–15.7	parity (within noise)
tracked	100 KiB	24.0–25.7	7.3–10.3	60–69%
tracked	1 MiB	218–241	11.4–18.5	92–95%
untracked	100 B	6.0–7.8	0.5–0.7	~90%
untracked	10 KiB	12.2–13.4	0.8–1.6	87–94%
untracked	100 KiB	20.5–22.6	0.8–1.5	93–96%
untracked	1 MiB	210–237	2.4–3.0	~99%

Client-side latency (localhost, RTT-dominated): 1 MiB tracked p50 improved from 772–861 µs to 566–703 µs (~25%); small payloads are roughly flat with overlapping ranges.

Readings:

• The release-notes expectation is confirmed where it was made: ~95% server-side reduction at 1 MiB tracked writes (observed 92–95%).
• The new protocol's server cost is nearly payload-size-independent (tracked ~7–18 µs at every size), which is the direct signature of the payload no longer crossing the Redis↔Lua boundary; the old protocol's cost grows roughly linearly with payload.
• Small tracked payloads (≤10 KiB) are parity within this machine's noise floor: per-run deltas swung from −23% to +33% across runs, i.e. a few µs either way. The protocol's costs there are offset by its savings; neither wins measurably.
• Untracked writes win at every size, ~90%+ — one native SET versus a payload-carrying script.

Methodology caveats

Single machine (Apple Silicon laptop, Docker Desktop VM) with unrelated containers running; sequential awaited writes (per-op cost, not throughput under pipelined concurrency); Redis 6.2 only; three runs. Treat the small-payload cells as "no measurable difference" rather than precise deltas — the µs-scale numbers there sit at the environment's noise floor. The 100 KiB and 1 MiB signals are stable across runs and large relative to noise.

What ships in this PR

• scripts/benchmark-redis-write.mjs + pnpm benchmark:redis-write — the same methodology, pointed at the local build, for future regression checks (maintainer tool, not part of the published package, no timing thresholds asserted).
• A README section documenting how to run it.