Add Experiment 13: which payload sizes change ranking

[leo-gan]

Summary

• Asks which SPSC payload sizes change the ranking (1 B, 64 B, 256 B, 4 KiB, 64 KiB).
• Ran in Python, Rust, JavaScript, C#, and C (full, 99 trials after warmup).
• Combined page: experiments/13-payload-size-sweep/results.md.

Finding

• You were right that queue benches should split on size, not serializer-style data types. Runners already build opaque byte strings.
• Two published sizes are enough: 256 B (message) and 4 KiB (document).
• 1 B and 64 B match 256 B (same first place; mid-pack swaps are ties).
• 4 KiB is the first copy-cost knee in Rust (pack collapses, ~9× slower).
• 64 KiB inverts the Rust ranking (~280× slower). Keep it as this experiment, not a third default type.
• event / telemetry / strings (512 B / 1 KiB / 2 KiB) add no ranking question.
• Item count 1 / 100 / 10000 is a different axis. n=1 is wakeup. n=1000 is amortization. Do not use those as payload sizes.
• Default config/library/default.yaml is not collapsed in this PR.

How to read

• Times in two languages are not one contest.
• Read the similar / close sets. Do not crown a single winner.

Validation

• Timed runs completed for: python, rust, javascript, csharp, c. Failed: none.
• No language harness source changed; published site Results / dashboard language payloads were not regenerated.

Notes

• run.yaml is generated from the experiment settings (checked in).
• Experiment logs stay local. Do not commit them.

[leo-gan]

Superseded by #33, which includes Experiment 13 and the default-matrix collapse.