perf: optimize artifact dependencies, and fail the build when a consumer does not

[wan9chi]

Motivation

Cargo compiles artifact dependencies declared under [build-dependencies] with the build-override profile — opt-level 0 — even in release builds. It is the [build-dependencies] placement, not the artifact = ... key, that triggers this (artifact deps under [dependencies] get the normal profile). The output still lands under target/<triple>/release/, so nothing reports the omission. See rust-lang/cargo#16719; the per-package override used here is what maintainers recommend (rust-lang/cargo#11680).

Three workspace artifacts are build-dependencies and have been shipping unoptimized: the two preload libraries — injected into every traced process, running inside every intercepted libc call — and vt_client_napi.

What this does

Sets the profile. A [profile.release.package.<name>] block per artifact, with opt-level = 3, codegen-units = 1 (the host profile drops the workspace setting), and strip = "symbols" (a package override otherwise resets an inherited strip to "debuginfo", which would leave the preload .so — embedded into the fspy binary — larger than intended).

Then checks that it worked. Profiles only take effect from the root manifest of the workspace being built, so these entries do not reach workspaces that consume these crates. vite-plus builds the preload at opt-level 0 today and would keep doing so after this merges, silently. Rather than rely on every consumer knowing, each artifact crate now inspects its own OPT_LEVEL in its build script and fails a release build that is unoptimized:

`fspy_preload_unix` was built without optimizations in a release build.

It is an artifact dependency, which Cargo compiles with the
build-override profile (opt-level 0) even in release builds. Profiles
only take effect from the root manifest of the workspace being built,
so add this to the root Cargo.toml of *your* workspace:

    [profile.release.package.fspy_preload_unix]
    opt-level = 3
    codegen-units = 1
    strip = "symbols"

Keep all three: an override that omits `strip` resets it to
"debuginfo", and `codegen-units` is dropped by the build-override
profile. Without this block the crate ships unoptimized and nothing
else reports it.

Debug builds are unaffected, since opt-level 0 is correct there.

Only the optimization level is checked: Cargo passes OPT_LEVEL, DEBUG and PROFILE to build scripts, but not codegen-units or strip, so those cannot be verified — which is why the message prints the whole block rather than the one setting it can see. The check lives in each artifact crate rather than in materialized_artifact_build because that crate runs from the consuming build script, which reports its own OPT_LEVEL (3) while the artifact it just embedded was built at 0.

This is a breaking change for downstream workspaces. vite-plus will fail to build until it adds the three blocks to its root Cargo.toml — it already uses that pattern for vite_trampoline and vite_installer. That failure is the point: the alternative is shipping a preload that costs ~25% on every intercepted call, with no signal at all.

Measurements

x86_64 Linux, benchmark launcher built with vs. without the profile change, batch median per 8 tracked opens:

access             -24.8%
access-relative    -19.1%

Raising the dependency subtree's pre-LTO opt-level on top of this measured within noise — fat LTO re-optimizes the merged module at the cdylib's opt-level — so this keeps the smaller, cdylib-only override set.

🤖 Generated with Claude Code

[github-actions]

fspy benchmark

linux

dynamic/launch             change  -2.28%  [ -5.03% ..  +1.56%]  overhead   +58.68%
dynamic/access             change -29.92%  [-30.44% .. -29.39%]  overhead    +6.59%
static/launch              change  +0.84%  [ -3.32% ..  +5.65%]  overhead  +167.25%
static/access              change  +0.00%  [ -0.97% ..  +0.59%]  overhead  +827.91%

macos

dynamic/launch             change  -0.03%  [ -4.40% ..  +4.90%]  overhead  +219.41%
dynamic/access             change -14.88%  [-17.58% .. -11.32%]  overhead    +2.78%

windows

dynamic/launch             change  -2.47%  [ -5.60% ..  +0.43%]  overhead   +28.12%
dynamic/access             change  -5.86%  [ -6.69% ..  -4.86%]  overhead    +1.15%

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.