shell: switch to quickjs-jit 0.12.7 with upstream compatibility facade

[huacnlee]

Summary

Run GPUI Shell through quickjs-jit 0.12.7, preserving the existing QuickJS binding API and sharing one VM with LLRT. Pin the bindings, runtime, and upstream rquickjs compatibility facade to 9a83a7f9cdc178a0ab5d7b29c3c9227b0dcb8620; remove the compatibility crate previously maintained in this repository. The bundled facade is introduced by longbridge/quickjs-jit#22.

Production runtimes enable automatic JIT. Debug runtimes remain interpreter-only. Defer JIT profiling during prelude/module evaluation, construction, initialization, and the first render; warm eligible functions can compile afterward. Cached frames still replay Rust snapshots without rerunning JavaScript.

The benchmarks compare automatic JIT with a true interpreter runtime, check snapshot parity and native entry, and expose categorized compilation/fallback metrics. The current mixed-workload regression remains unresolved; this PR is not evidence that default JIT is ready to merge.

Performance evidence

The 0.12.7 release-tag retest on Apple M3 used release builds, 30 independent processes per version/mode/workload, interleaved serial execution, 64 warmup renders, and 200 measured renders per process. Values below are medians across processes; the P99 column is the median of per-process P99s. Timings include snapshot validation. The interpreter baseline is the same quickjs-jit release with JIT disabled, not a separate upstream rquickjs build.

workload	interpreter steady	automatic JIT steady	JIT speed relative to interpreter	interpreter / JIT P99
UI panel	0.8518 ms	0.8472 ms	1.01x, approximately equal	0.9575 / 0.9316 ms
numeric layout	0.1458 ms	0.0159 ms	9.19x	0.1767 / 0.0177 ms
mixed market	0.2156 ms	0.4548 ms	0.47x	0.3018 / 0.6050 ms

The panel entered no native code. The mixed workload fell back in 22/30 automatic samples; the other eight samples each recorded four rejected invalid artifacts. Snapshot checksums matched across all 360 samples. The zero-fallback mixed-workload assertion failed in 4/10 independent repeat runs. These findings supersede the earlier blanket acceptance conclusion; no cross-platform or memory/binary-size conclusions are drawn from this run.

Test Plan

• git diff --check.
• cargo tree --locked -p gpui-shell -i quickjs-jit-core: one shared binding core; all six QuickJS/facade packages use the pinned revision.
• Release build from the pinned Git source.
• cargo test -p gpui-shell --release tests::benchmark -- --include-ignored --nocapture --test-threads=1: 10 passed, 1 failed. The failure is the existing mixed_market_panel_matches_interpreter_and_enters_native_code zero-fallback assertion, also reproduced before the facade migration.
• Cached-frame assertions passed through 8,403 nodes with zero additional script renders.
• No fresh Windows/Linux validation was run for this migration.

Checklist

• Reviewed the AI-assisted dependency migration and benchmark findings.
• Keep all QuickJS packages on one upstream revision; no sibling-checkout paths are committed.
• Resolve mixed-workload fallback and nondeterministic performance before treating automatic JIT as merge-ready.

[huacnlee]

Apple Silicon Issue #3 diagnostics are now wired through quickjs-jit PR #15 (29d66ef).

Fresh release verification on Apple arm64 / macOS 27:

• 443-node panel: 66/66 compile failures were tier1_rejections; all unsupported/resource/cancelled/panic/invalid/install categories were zero, with zero installed/native entries and no pending worker jobs.
• Numeric layout: installed=1, native_entries=1, native_exits=1, osr_entries=1; zero fallback, deopt, invalid artifact, install failure, or pending worker jobs.
• Numeric interpreter and automatic runs used equal render counts and produced identical snapshot debug trees (layout:165580141).

Commands:

cargo fmt --all -- --check
git diff --check
cargo check -p gpui-shell --tests
cargo test -p gpui-shell --release tests::benchmark::describing_a_panel_stays_inside_the_frame_budget -- --exact --nocapture
cargo test -p gpui-shell --release tests::benchmark::numeric_layout_installs_and_enters_native_code -- --exact --nocapture

The remaining Issue #3 evidence is a paired panel parity run and the statistically sampled 0.12.5 performance comparison.

[huacnlee]

Follow-up 5fab7f4d adds paired panel parity: automatic and interpreter each execute the same 352 real renders, then assert identical complete snapshot debug trees and script-render counts. It also restores the rquickjs-jit dependency alias so cargo machete recognizes the runtime dependency; local cargo machete, debug test compilation, and both release acceptance tests pass. The untracked local assets/ directory was not staged or modified.