bench/: perf harness + regression-gated PERFORMANCE.md (#398)

[InauguralPhysicist]

What

Performance as an executable, regression-gated fact (Rust's credibility move) rather than a README claim.

Two metrics, on purpose

Wall-clock and instruction count answer different questions, and only one can gate CI honestly:

• bench/run_bench.sh — wall-clock n=5 medians (the house rule). Feeds docs/PERFORMANCE.md; environment-dependent, so published, not gated.
• bench/check_regression.sh — the gate uses cachegrind instruction counts (Ir), which are deterministic: same binary + input → identical Ir on any machine, every run. So it never flakes on runner load. In CI it compares this commit's Ir against origin/main, both built in the same runner (--vs) — a regression is a pure code-change signal, not machine noise. No cross-environment baseline drift.

Not vacuous: --selftest builds a 2×-work workload and asserts it's flagged (measured +105% Ir > 5% threshold).

The observer cost, now an executable document

observed_loop vs unobserved_loop is the same arithmetic; the difference is per-assignment observer bookkeeping — measured here at ~28% wall-clock / ~46% instructions. So unobserved: on a hot numeric loop is a real, now-gated win.

Acceptance

• bash bench/run_bench.sh reproduces medians locally ✓
• CI red on a deliberate 2× pessimization ✓ (the --selftest, deterministic + in-env)
• PERFORMANCE.md regenerates from the harness ✓
• New CI bench job; no runtime/C changes.

The MT spawn/channel cliff stays a documented contract in docs/CONCURRENCY.md; a dedicated MT workload is noted as future work.

Closes #398

🤖 Generated with Claude Code