Publish a measured performance report with a regeneration script

[weiminglong]

Summary

The repo claimed performance characteristics without a single attributed measurement. Now:

• docs/PERFORMANCE.md — measured Criterion medians for every hot-path stage (wire deserialize, dispatcher normalize + shadow-book cross-check, WAL codec encode / append+flush / fdatasync-per-record, book delta/snapshot/top-of-book, read-model snapshot), stamped with CPU, RAM, OS, toolchain, commit, and date, plus the full 49-benchmark table. Headline medians from the committed run (Apple M3, 8 cores): book delta apply 9 ns, top-of-book read 1 ns, WAL codec encode 87 ns, WAL append+flush 509 ns/record, fdatasync-every-record 4.04 ms, dispatcher normalize 484 ns/entry (2.1 M entries/s), read-model snapshot 154 ns.
• scripts/bench-report.py — regenerates the report from target/criterion estimates (batch benches get a per-item column via encoded batch sizes); just bench-report runs the suite and rewrites the file.
• README's benchmarks section now leads with the measured numbers and links the report.

The report is explicit about methodology limits: single-machine wall-clock medians for order-of-magnitude reasoning; CI compiles benches but does not gate on timings.

Verification

• Generated from a full cargo bench --workspace run on current main (49 benchmarks collected)
• Script re-run is idempotent; hygiene gate clean; doc-only + script + justfile changes (no Rust code touched)