Reduce repeated CI setup and fixture generation

[Sebastian Thiel (Byron)]

Tasks

• refackiew

---

Created by Codex on behalf of Byron. Byron will review before this is ready to merge.

Summary

Reduce repeated CI setup and fixture generation while preserving every distinct feature configuration, platform, and test environment.

• Generate the 410 shared URL baseline cases once and append their results to both platform baselines: 846 → 436 git fetch-pack --diag-url calls, with identical output. Start one test for each of three expensive Windows fixtures earlier so independent generation can overlap.
• Share Cargo's resolved root target directory across the 23 separate fuzz workspace checks, preserving their independent feature resolution. Reuse the Windows test-fast dependency cache in fixture jobs, retaining separate checkouts and parallel execution.
• Generate documentation once in the lint job, remove four literal duplicate test commands, and install the existing pinned cargo-machete version from a prebuilt binary. All 85 unique unit-test commands and the local ci-test recipe remain.
• Enforce expected-failure checks explicitly. Standalone ! commands suppress set -e, and the dependency guard also silently passed when rg was unavailable. Use the existing grep tool and report unexpected success as a failure.

Context

The audit compared #2847 and #3008: 798 changed files versus five, but both CI runs took approximately 22 minutes. Windows ARM fixture testing was the slowest job in both, with roughly 15 minutes of test execution alone. Each Linux fuzz-check loop also compiled libfuzzer-sys 23 times in separate target directories.

The requested priority is shorter completion time without risking missed feature combinations. Following the emphasis on repeated setup and slow tests in Linear's CI write-up, this keeps both Windows fixture modes parallel and preserves per-crate feature checks. It does not introduce change-based test selection, additional shards, or a new runner provider.

Windows scheduling gains and end-to-end CI savings still require measurement on the hosted runners. The changes do not increase test concurrency.

Validation

• etc/scripts/ci-check-local.sh --thorough passed on macOS with Rust 1.98.1, including the feature-check matrix, feature-specific tests, doctests, 4,265 workspace tests, and all four journey configurations. Tix recorded the checks-pass mark for commit 8932904.
• Targeted gix-url tests passed: 20 unit tests, 151 integration tests, and three doctests. An isolated before/after fixture comparison produced byte-identical Unix, Windows, and remote-helper baselines; one local measurement decreased from 6.48s to 3.89s.
• Confirmed every existing Cargo check, argument, and feature combination remains in order, and the unit-test recipe retains all 85 unique commands. Verified shared-target handling for environment and Cargo configuration overrides.
• Verified nextest configuration and unchanged test selection; tested the dependency guards' success/failure cases. Shell checks passed apart from existing intentional literal-backtick diagnostics. Workflow validation with actionlint 1.7.12 passed; its full ShellCheck integration reported the same diagnostics as the baseline.

User Prompts

Look the CI run of #2847 and #3008 for two very different PRs, and see where it duplicates work and where it could run more efficiently.

You can also run CI scripts locally to learn more.

I think that preventing to not test or compile something is preferable over risking to miss a feature flag combination, but I wonder if there are other places where it runs too much that are less risky to deduplicate.

Shortest completion time

Also look at https://linear.app/now/ci-bottleneck-reworked to learn what worked for others.

Implement the plan.