CI perf: coverage — cargo llvm-cov report runs twice over the same profiles (~2.5 min/run, ~530 job-min/day) #1174

Measurement

• In the coverage job of merge_group run 37828932401, the step "Run tests with coverage" takes 15.8 min.
  • The instrumented compile plus tests ends at about 13.3 min.
  • It then runs cargo llvm-cov report --lcov (78 s), followed by cargo llvm-cov report --summary-only (74 s).
• Across 7 days, coverage takes p50 12.5 min and p90 15.1 min (n=950 jobs).
• Since 2026-10-08 02:16Z, about 427 CI runs per day include coverage:
  • 287 pull_request
  • 105 merge_group
  • 35 push

Where the time goes

.github/workflows/ci.yml, at about lines 536–539:

cargo llvm-cov --workspace --no-fail-fast ...
cargo llvm-cov report --lcov --output-path coverage-host.lcov
cargo llvm-cov report --summary-only | tee coverage-summary.txt

Each report call merges the .profraw files with llvm-profdata and walks every instrumented object. The two calls do the same work.

Root cause

The summary is produced by a second full report pass. It is not derived from the lcov output that was just written.

Proposed fix

Pick one of these:

• (a) Run both report invocations concurrently, using & and wait. This cuts wall time by about 75 s but saves no CPU.
• (b) Preferred: emit lcov once and compute the per-file summary table from coverage-host.lcov. The LF/LH/BRF/BRH/FNF/FNH records hold the line, branch and function totals. A small Python helper in scripts/ can render the same fenced table the next step prints.

Keep coverage-summary.txt in the same format so that coverage-merge and the step summary don't change.

Expected saving

• With (b): about 1.25 min of job time per run, which is about 530 Linux job-min per day.
• coverage wall time drops by about 1.3 min.
• Critical path: about 0. coverage finishes before the Gradle e2e legs, and in the last 37 green merge_group runs it was never the job that gated ci-ok.

Coverage and risk

• Test execution is unchanged; only the reporting changes.
• Risk: summary numbers could drift from llvm-cov's own summary. Diff the two outputs on one run before switching.
• The required checks ci-ok and clippy are unchanged.

Effort

S

ROI

ROI = weighted saving × confidence / effort.

• Weighted saving: 0.53k Linux job-min per day.
• Confidence: 0.8. Effort: S, which counts as 1.
• ROI = 0.53 × 0.8 / 1 = 0.42