CI: skip framework-dependent bootstrapper in self-contained /perf Rust harness (fix #674 0xC0000135)

[azchohfi]

Summary

The on-demand /perf workflow's cross-framework Rust column still read n/a, with test_reactor_perf.exe exiting 0xC0000135 (DLL-not-found) at launch — even after the self-contained build (#662) and WebView2 Core staging (#689) fixes. This PR fixes the remaining root cause so the Rust column produces live numbers.

This is CI/build-script + docs only — no framework runtime code is touched.

Root cause (reproduced locally, hermetically)

I reproduced the runner failure on a local box with PATH scrubbed to System32 only (so a machine-wide Windows App SDK runtime cannot mask missing app-local DLLs), mirroring the GitHub windows-latest runner which has no machine-wide WinAppSDK runtime:

• The windows-rs reactor_perf harness main() unconditionally calls bootstrap() (MddBootstrapInitialize2) to locate a machine-wide Windows App SDK runtime.
• windows_reactor_setup::as_self_contained() (which the workflow already patches build.rs to call) never stages the framework-dependent bootstrapper — only as_framework_dependent calls copy_bootstrap_to.
• So the self-contained exe carries an unsatisfiable load-time PE import on Microsoft.WindowsAppRuntime.Bootstrap.dll (confirmed with dumpbin /imports). On the runtime-less runner the loader fails before main runs → 0xC0000135, instant exit, 0-byte stdout/stderr → the 100s watchdog reports n/a.
• Second layer: even if that DLL is staged, MddBootstrapInitialize2(..., OnNoMatch_ShowUI | ...) looks for a machine-wide framework package and fail-fasts headless (0xC0000602) when none is present. So staging the DLL is necessary-but-insufficient — the bootstrap call itself must be neutralized. (This is why CI: stage WebView2 Core DLL so the /perf Rust column runs (fix #674) #689's staging alone didn't fix it.)

MddBootstrapInitialize2's own documentation: "Self-contained apps do not need to call this." The harness's embedded reg-free app.manifest activates the runtime app-local.

The fix

In the workflow's "Prepare Rust harness" step, alongside the existing build.rs → as_self_contained() patch, comment out the single bootstrap()?; call in the cloned harness main.rs:

$patched = [regex]::Replace($src, '(?m)^(?<indent>[ \t]*)bootstrap\(\)\?;', '${indent}// bootstrap()?; // ...')

Why it's safe:

• Benchmark-neutral. main()'s bootstrap() is deployment-mode setup (runtime discovery), not the measured StocksGrid workload. Removing it is directly analogous to the existing build.rs as_self_contained() patch and does not affect the four target metrics.
• Best-effort. If upstream ever moves the call so the regex stops matching, the step logs a warning and the column reads n/a — it never fails the job, and the PR-vs-main C# comparison is entirely unaffected (the Rust leg is continue-on-error).
• Stage-RustRuntime is unchanged — the runtime MSIX payload + Microsoft.Web.WebView2.Core.dll (manifest-declared SxS) are still staged app-local; this fix removes the need for the bootstrapper, it doesn't change staging.

Local verification

Applying the exact workflow patches (build.rs + main.rs regex) to the pinned windows-rs checkout, rebuilding, and running hermetically:

• ✅ dumpbin confirms the Microsoft.WindowsAppRuntime.Bootstrap.dll import is gone from the rebuilt exe.
• ✅ Hermetic, PATH-scrubbed run → exit 0 with full metrics (Renders/sec, Reconcile, Diff, Memory).
• ✅ Loaded-module snapshot: all WinAppSDK DLLs resolve from the app-local staging dir (zero from C:\Program Files\WindowsApps) — proving the fix is genuinely runtime-independent, not masked by my box's installed runtime.
• ✅ Regex unit-validated: matches exactly one call site, preserves indentation, idempotent, leaves no bare bootstrap()?;.

I also confirmed the before state: the unpatched exit is 0xC0000135 (matching the runner), and the DLL-staged-but-call-present exit is 0xC0000602 (the fail-fast), exactly as the root-cause analysis predicts.

Validation run locally

• tests/stress_perf/ci/PerfLib.Tests.ps1 → 61/61 ✅
• tests/stress_perf/ci/RunPerfBenchmark.Tests.ps1 → 21/21 ✅
• Workflow YAML parses ✅; the step's embedded PowerShell AST parses with 0 errors ✅

Note: the local dotnet build Reactor.slnx -c Release surfaces a pre-existing, unrelated ARM64-dev-box XAML markup-compiler failure (issue #678, fixed by the separate open PR #684 — XamlCompiler.exe MSB3073 on long paths). It does not occur on x64 windows-latest CI runners and is independent of this YAML+docs change (neither file is an MSBuild compiler input).

Final acceptance test

This can only be fully validated by a live /perf re-run on a pre-gate PR (e.g. comment /perf on #656). A maintainer should run that as the final acceptance step; the local hermetic repro + module-path proof above are the strongest evidence achievable off-runner.

Refs

• Fixes the 0xC0000135 symptom tracked in Perf CI: Rust cross-framework column n/a — test_reactor_perf exits 0xC0000135 (DLL-not-found) at runtime #674 (auto-closed by CI: stage WebView2 Core DLL so the /perf Rust column runs (fix #674) #689, but the bug persisted on the live re-run).
• Builds on CI: on-demand perf comparison (/perf) vs main baseline #662 (self-contained harness), CI: make /perf work on pre-gate PRs (self-contained C# overlay) + fix Rust column (#674) #677 (csproj overlay), CI: stage WebView2 Core DLL so the /perf Rust column runs (fix #674) #689 (WebView2 Core staging).

Co-authored-by: Copilot 223556219+Copilot@users.noreply.github.com

[Copilot]

Pull request overview

This PR fixes the /perf workflow’s Rust cross-framework column by ensuring the pinned windows-rs perf harness can launch on runners without a machine-wide Windows App SDK runtime, by neutralizing the harness’s framework-dependent bootstrapper call while keeping the self-contained staging approach.

Changes:

• Patch the pinned windows-rs reactor_perf harness to comment out bootstrap()?; in main.rs, preventing load-time dependency on Microsoft.WindowsAppRuntime.Bootstrap.dll on runtime-less CI runners.
• Update the stress perf CI README to document why the bootstrap call must be removed for the self-contained Rust harness and how the failure surfaces (0xC0000135 / 0xC0000602).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
.github/workflows/perf-compare.yml	Adds a best-effort PowerShell patch that comments out the Rust harness bootstrap() call to avoid framework-dependent bootstrapper imports on CI.
tests/stress_perf/ci/README.md	Documents the Rust self-contained harness patching rationale and troubleshooting guidance for 0xC0000135 “n/a” failures.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.