bench(jit): measure mixed gpui-shell workload

[huacnlee]

Issue # (if available)

Follow-up to longbridge/gpui-kit#2903.

Description of changes

• add a real gpui-shell market workload that computes and aggregates 96 quote scores, sorts them, and builds a 12-row visible list
• extend the paired fresh-process runner and report format with native-required observational workloads
• support exact crates.io versions as well as Git revisions in the integration runner
• record clean x86_64 Linux evidence: mixed steady-state 2.32x..2.34x and P99 2.20x..2.27x, with snapshot parity and zero fallback/deopt
• document the next target: diagnose four invalid artifacts per mixed automatic process, then isolate array/object sorting and host-list stages

x86_64 Linux gpui-shell performance evidence

Measured from clean commits 034c9f04f80ef0464ee2589dbb202608ac090e4d (quickjs-jit) and c0b989691707d5c18d973cf9cf805bbcadde37b5 (gpui-kit).

Sampling: 5 discarded warmup processes, 30 interleaved paired fresh processes, 200 observations per process, and 10,000 paired bootstrap resamples. Every process was pinned to CPU 0.

Workload	Steady-state speed CI	P99 speed CI	Native entries	Fallback / deopt
Realistic 443-node host-heavy panel	0.99x..0.99x	0.99x..1.00x	0	0 / 0
Render-driven numeric layout checksum	15.34x..18.33x	10.73x..15.24x	7,028	0 / 0
Market compute, sort, aggregate, and visible list	2.32x..2.34x	2.20x..2.27x	732,467	0 / 0

Lifecycle speed CIs: first window 0.99x..1.00x; hot reload 0.99x..1.00x. Interpreter and automatic mode produced identical checksums, snapshot SHA-256 values, and script-render counts in all 30 pairs. The overall acceptance report passed.

The mixed workload computes 96 quote scores, aggregates and sorts them, then builds a 12-row visible list through the real GPUI builder API. This demonstrates a stable product-shaped gain above 2x while retaining the host-heavy panel regression guard.

Known follow-up: mixed automatic samples produced 120 optimizing-tier invalid-artifact compilation failures (four per process). They did not affect semantics or the installed native path, but tracing and eliminating or correctly classifying them is the next optimization target.

Verification:

• cargo test --workspace passed.
• gpui-shell release suite: 697 passed, 1 ignored.
• gpui-shell integration tests: 4 passed.
• gpui-shell doctests: 4 passed, 2 ignored.

Tracked evidence: benchmarks/results/gpui-shell-mixed-v1.json and .md.

Verification

• cargo test --workspace
• scripts/bench-gpui-shell.sh /home/jason/work/gpui-kit /tmp/gpui-shell-mixed-clean.json
• gpui-shell release suite: 697 passed, 1 ignored; integration tests: 4 passed; doctests: 4 passed, 2 ignored

Checklist

• Added change to the changelog (benchmark/documentation-only change)
• Created unit tests for my feature if needed

性能进展

场景	95% CI	结论
混合场景 steady-state	2.32x–2.34x	明确超过解释器 2 倍
混合场景 P99	2.20x–2.27x	尾延迟也稳定提升
纯计算 steady-state	15.34x–18.33x	显著加速
纯计算 P99	10.73x–15.24x	显著加速
Host-heavy panel	0.99x	慢约 1%，通过 5% guard
首窗口	0.99x–1.00x	基本持平
热重载	0.99x–1.00x	基本持平

[huacnlee]

x86_64 Linux gpui-shell performance evidence

Measured from clean commits 034c9f04f80ef0464ee2589dbb202608ac090e4d (quickjs-jit) and c0b989691707d5c18d973cf9cf805bbcadde37b5 (gpui-kit).

Sampling: 5 discarded warmup processes, 30 interleaved paired fresh processes, 200 observations per process, and 10,000 paired bootstrap resamples. Every process was pinned to CPU 0.

Workload	Steady-state speed CI	P99 speed CI	Native entries	Fallback / deopt
Realistic 443-node host-heavy panel	0.99x..0.99x	0.99x..1.00x	0	0 / 0
Render-driven numeric layout checksum	15.34x..18.33x	10.73x..15.24x	7,028	0 / 0
Market compute, sort, aggregate, and visible list	2.32x..2.34x	2.20x..2.27x	732,467	0 / 0

Lifecycle speed CIs: first window 0.99x..1.00x; hot reload 0.99x..1.00x. Interpreter and automatic mode produced identical checksums, snapshot SHA-256 values, and script-render counts in all 30 pairs. The overall acceptance report passed.

The mixed workload computes 96 quote scores, aggregates and sorts them, then builds a 12-row visible list through the real GPUI builder API. This demonstrates a stable product-shaped gain above 2x while retaining the host-heavy panel regression guard.

Known follow-up: mixed automatic samples produced 120 optimizing-tier invalid-artifact compilation failures (four per process). They did not affect semantics or the installed native path, but tracing and eliminating or correctly classifying them is the next optimization target.

Verification:

• cargo test --workspace passed.
• gpui-shell release suite: 697 passed, 1 ignored.
• gpui-shell integration tests: 4 passed.
• gpui-shell doctests: 4 passed, 2 ignored.

Tracked evidence: benchmarks/results/gpui-shell-mixed-v1.json and .md.