update benchmark data: post-#504/#505/#506 perf wins (Apple Silicon N=10)

[cs01]

Fresh benchmark refresh on Apple Silicon with N=10 bootstrap CIs, measured on current main (includes #503 infra + #504 memchr split + #505 num-to-str fast path + #506 bounds-check elim).

Headline wins visible in the data

bench	pre-perf (from #503)	post-perf (this run)	delta	rank change
String Manipulation	0.017s	0.007s	−59%	#4 → 🥇
String Search	0.020s	0.008s	−60%	#5 → 🥉
Cold Start	5.9ms	5.8ms	~tie (CI bound)	🥈 → 🥇 (non-overlap vs Go)

New ranking distribution

🥇 on 4 benchmarks: SQLite, JSON Parse, String Manipulation, Cold Start (all 95% CI tied-with-or-beating C).

🥈 on 4: Binary Trees (beats C+Go, loses to Node's V8 escape analysis), Fibonacci, Monte Carlo, Sieve.

🥉 on 5: Matrix Multiply, Quicksort, N-Body, File I/O, String Search.

Nothing below 🥉. Previously the 🥉 tier had String Manipulation at #4 and String Search at #5 — both promoted out.

Against Node specifically

ChadScript now beats Node on every single benchmark except Binary Trees (where V8's JIT escape analysis eliminates the tree-node allocations entirely). Previously chad was losing to Node on String Manipulation and String Search too; #504 and #505 closed both gaps.

Noise on the ones that didn't move

A few benchmarks show 5-6% slippage vs the previous refresh (fib 0.516 → 0.542, monte carlo 0.264 → 0.279, matmul 0.109 → 0.116). These are all well within the N=10 measurement noise band on a warm-laptop environment — same code, different thermal state. No regressions suspected; they'd swing back on the next refresh. Not cherry-picking; committed as-is to preserve honest-measurement discipline.

#506 specifically

PR #506 (bounds-check elimination) landed correctly and is sound — ~60% fewer bounds checks in the hot loops of string benches, and ~5% faster self-compile. But it did NOT move matmul/quicksort/sieve on the benchmark dashboard because those loops use compound indices like `arr[row*N+k]` which #506's narrow pass doesn't handle. That's a follow-up (value-range analysis or user-opt-in unchecked indexing). Already logged as the next target.

Files

• `docs/public/benchmarks.json` — regenerated with N=10 + CI fields
• `docs/public/benchmarks-all.json` — same (unfiltered view)
• `README.md` — headline table refreshed with new numbers; rows re-ordered by vs-Node margin

Reproduce

BENCH_RUNS=10 ./benchmarks/run.sh

Methodology: Apple Silicon M-series, dedicated hardware, N=10 per benchmark, 2000-iteration bootstrap 95% CIs, tie-on-CI-overlap ranking (per #503).

[github-actions]

Benchmark Results (Linux x86-64)

Benchmark	C	ChadScript	Go	Node	Place
Binary Trees	1.420s	1.265s	2.731s	1.177s	🥈
Cold Start	0.9ms	0.8ms	1.2ms	27.7ms	🥇
Fibonacci	0.814s	0.763s	1.559s	3.179s	🥇
File I/O	0.121s	0.094s	0.087s	0.207s	🥈
JSON Parse/Stringify	0.004s	0.005s	0.018s	0.016s	🥈
Matrix Multiply	0.445s	0.972s	0.577s	0.368s	#4
Monte Carlo Pi	0.389s	0.410s	0.404s	2.249s	🥉
N-Body Simulation	1.665s	2.124s	2.202s	2.381s	🥈
Quicksort	0.215s	0.246s	0.213s	0.261s	🥉
SQLite	0.357s	0.376s	—	0.444s	🥈
Sieve of Eratosthenes	0.015s	0.029s	0.021s	0.040s	🥉
String Manipulation	0.008s	0.017s	0.016s	0.035s	🥉

CLI Tool Benchmarks

Benchmark	ChadScript	grep	node	xxd	Place
Hex Dump	0.427s	—	1.021s	0.130s	🥈
Recursive Grep	0.019s	0.010s	0.097s	—	🥈