ci: run zebra-consensus benches on zebra-touching commits

[aphelionz]

Runs the zebra-consensus criterion benches on every zebra-touching commit (PRs and main pushes), posting the timings to the job's step summary. Informational, not a required check: absolute times on shared runners are noisy, but the ratios the benches watch (cache miss vs hit in #45, serial vs overlapped lookups in #46) are measured within one run on one box.

Follows the z3-regtest conventions: no paths: trigger filter, a Path gate job diffing the merge commit (manual dispatches bypass it), Swatinem rust-cache on the zebra workspace, saved from main only so per-PR saves cannot evict the smoke/regtest caches. Each bench step is guarded on its bench file existing, so this can merge before #45/#46 and activates as they land.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Adds an informational CI workflow to run Zebra consensus Criterion benchmarks on relevant PRs and main pushes.

Changes:

• Adds path-gated and manual workflow triggers.
• Runs available consensus benchmarks with Rust caching.
• Publishes benchmark results to the step summary.

Suppressed comments (5)

.github/workflows/zebra-bench.yml:67

• workflow_dispatch is filtered through the same path decision, so manually running this workflow on a commit whose immediate parent changed only documentation silently skips both benches. Since this trigger has no force-run input, special-case manual dispatches as relevant (as the other job-gated workflows do) or remove the trigger.

          if [ "$hits" -gt 0 ]; then relevant=true; else relevant=false; fi

.github/workflows/zebra-bench.yml:63

• On a push that contains multiple commits, HEAD^1..HEAD only examines the tip commit. For example, a zebra change followed by a docs-only commit makes hits=0 and skips the benches, so a zebra-touching commit is missed. Use the push event's before SHA (fetching it or using full history), while retaining the merge-parent range for pull requests.

          git diff --name-only HEAD^1 HEAD -- "${specs[@]}" > "$RUNNER_TEMP/matched"

.github/workflows/zebra-bench.yml:98

• This rust-cache setup does not provide a rolling Criterion baseline: for source-only commits with the same toolchain/job/lockfile, the cache key remains the same, and GitHub cache entries are immutable, so the post-run target/criterion data cannot replace the first cached entry. Later runs therefore compare against stale data rather than the immediately preceding run as the workflow comment claims. Store Criterion data with a per-run key plus a ref-scoped restore key, or otherwise provide an updateable baseline separately from the build cache.

        uses: Swatinem/rust-cache@v2
        with:
          prefix-key: v1-bench
          workspaces: |
            zebra -> target

.github/workflows/zebra-bench.yml:108

• These benchmark commands do not use --locked, so a manifest/lockfile mismatch can make Cargo resolve newer compatible dependencies and rewrite the lockfile during the run. The reported timing can then depend on registry state and represent a dependency graph different from the commit; fail loudly with --locked instead.

          cargo bench -p zebra-consensus --bench script 2>&1 | tee "$RUNNER_TEMP/script.out"

.github/workflows/zebra-bench.yml:124

• This benchmark command has the same reproducibility problem as the script bench: without --locked, a manifest/lockfile mismatch can resolve newer compatible dependencies during the run, making the timing depend on registry state. Pass --locked so the benchmark either uses the committed dependency graph or fails.

          cargo bench -p zebra-consensus --bench utxo_lookup 2>&1 | tee "$RUNNER_TEMP/utxo.out"

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated no new comments.