Refresh engine comparison benchmark results

[lahma]

Re-runs the EngineComparison suite on the current main (14 commits past v4.10.1) and refreshes Jint.Benchmark/README.md with a fresh table, updated highlights, and a regression-focused summary.

Context

The competitor engine versions are unchanged since v4.10.1 (Jurassic 3.2.9, NiL.JS 2.6.1722, YantraJS.Core 1.2.406), so every movement in the table is Jint's own allocation/dispatch tuning landed since the tag: object-literal & constructor shapes (#2552, #2553), in-object properties (#2559), source-generated built-in shapes (#2554–#2557), the write-side (#2546) and prototype-method (#2558) inline caches, recursion env pooling (#2549), and zero-copy SlicedString search (#2547).

Regression check (the headline ask)

I ran a v4.10.1-vs-HEAD Jint-only A/B back-to-back in a single thermal window (both Jint re-parse and Jint_ParsedScript prepared paths), on top of the full all-engines run used for the table. Verdict: no alarming regressions.

• Real, consistent improvements (move the same way in both execution paths): dromaeo-object-array −12.7% (17.0 → 15.1 ms), dromaeo-object-array-modern −3%/−12%, array-stress −6.5%/−12% — all attributable to the prototype-method inline cache (Prototype-method inline cache for obj.method resolved on the direct prototype #2558).
• Allocation is flat (within ±2%) on every script except linq-js, which rises ~28 KB/op (175 → 203 KB prepared) — the one workload that builds many distinct object shapes per run, so the new shape metadata is rebuilt each operation. It is still 5× below YantraJS and 13× below NiL.JS.
• Apparent time slowdowns are measurement noise. stopwatch, dromaeo-object-string and dromaeo-string-base64 show deltas in the ±5–8% band, but they flip sign between the re-parse and prepared paths within the same run (e.g. stopwatch is +4.5% prepared / −6.7% re-parse), and intra-run the two paths disagree by more than the delta itself. These are call-dispatch-bound, high-variance scripts; the deltas are within their run-to-run noise floor.

Method

• Host: AMD Ryzen 9 5950X, .NET 10.0.9, BenchmarkDotNet v0.15.8 (DefaultJob), same machine as the previous refresh.
• Table = full all-engines run on main @ 9c0266e6c.
• Regression A/B = v4.10.1 worktree vs HEAD, Jint/Jint_ParsedScript only, adjacent runs.

🤖 Generated with Claude Code