chore(bench): add vm-linux-x86_64 benchmark run 2026-09-02

[chaliy]

What changed

Adds a fresh bashkit-bench run (bashkit in-process vs bash 5.2.21 subprocess) to crates/bashkit-bench/results/ and refreshes the generated /benches site timeline from it. No code changes.

Published reports:

• This run (bashkit vs bash): https://stend.sh/s/s10BFCg5pz2-DETCS3H-Gw/
• Companion: current HEAD vs bashkit from 3 months ago (v0.9.0, commit 87a5516): https://stend.sh/s/4F1nvasj53CGvu01pAkUjA/

Why

The last saved vm-linux-x86_64 run is from 2026-05-26. This refreshes the baseline on current main (v0.17.1) per the result-location contract in knowledge/operations/performance-results.md.

Before / After

Same machine class (vm, linux x86_64, 4 CPUs), 96 cases × 10 iterations, 0 errors, 100% output match in both runs.

	2026-05-26 (v0.9.0 era)	2026-09-02 (this PR)
bashkit avg / case	0.447 ms	0.539 ms
bash avg / case	9.363 ms	11.174 ms
bashkit vs bash (avg)	20.9×	20.7×
bashkit vs bash (geo-mean)	24.4×	29.8×

The companion 3-month comparison (3 alternating rounds × 30 iterations per version, same machine) shows a split rather than a uniform change: tool builtins are 20–55% faster (grep/sed, tool_jq_map 1.46×) and control is up 29%, while every case in the large category regressed 12–33% (category geo-mean 0.83×; e.g. large_function_calls_500 5.48 → 6.59 ms, large_string_append_100 0.286 → 0.372 ms). Overall geo-mean 1.02×, total time 40.9 → 47.3 ms. Worth a bisect on large_function_calls_500; not addressed in this PR.

Risk

• Low
• Results-only change. Site timeline regenerated with site/scripts/build-performance-data.mjs and verified deterministic on re-run.

Checklist

• Tests added or updated — N/A, no code changed
• Backward compatibility considered — N/A

---

Generated by Claude Code