Measure bench baselines on the same runner

[owjs3901]

Problem

The Criterion regression gate compared measurements taken on two different machines.

The baseline was produced only on push to main, persisted through the actions cache, then restored in a later PR job and compared against numbers measured on whatever runner that PR happened to land on. GitHub's ubuntu-latest fleet is heterogeneous, so fleet variance registered as a code regression.

Run 33302880992 failed 15 benches at +25% to +148% on #141 — a changepacks release PR whose entire diff is version = "0.3.1" → "0.4.0" strings plus three deleted changepack logs. Zero bytes of crates/** changed. The same benchmark code had passed on main eight minutes earlier (run 33302590831).

The documented noise filter (mean > 10% AND CI lower > 5%) cannot catch this. Criterion's confidence interval measures within-run variance, not between-machine variance — when the two sides run on different hardware criterion becomes more confident, not less. Hence the observed signature of a huge mean shift inside a very tight band (+148.04% mean, +145.58% CI lower).

Secondarily, Cargo.toml is a path trigger, so every changepacks "Update Versions" PR drags the perf gate along despite touching no source.

Changes

• Same-runner comparison. Measure the PR merge base with --save-baseline main, then the PR head with --baseline main, back to back in one job. The cross-run baseline cache is removed entirely.
• No stale estimates. rm -rf target/criterion before the baseline run, so no change/estimates.json survives from an unrelated run.
• Version-only PRs skip the gate. Both revisions' Cargo.toml are parsed with tomllib and compared after removing only changepacks-managed release versions (workspace.package.version, plus version on path dependencies pointing at crates/<name>). Anything else — external dependency bumps, profile edits, new dependencies, an unparseable file — fails safe and runs the gate.
• Thresholds calibrated to the measured noise floor: +10%/+5% → +30%/+20%.

Why the thresholds moved

Same-runner comparison cut the noise by an order of magnitude, but did not remove it. Baseline and head are measured minutes apart, so runner drift and co-tenancy still shift the mean.

This PR is its own control: it touches only bench.yml, so its merge base and head hold byte-identical Rust. Every reported delta below is pure noise.

Run	Range	Max	Benches tripped at +10%/+5%
1	−12.33% … +12.20%	+12.20%	borrowed_pre_P2/1
2 (re-run)	−4.68% … +14.92%	+14.92%	borrowed_pre_P2/1024, owned_post_P2/1024
3	−2.07% … +12.33%	+12.33%	borrowed_pre_P2/64

All three identical-code runs would have failed the old gate, each on a different benchmark. A gate with a ~100% false-positive rate on unchanged code is not a gate. +30%/+20% sits at roughly 2x the highest observed noise, so the timing gate now catches only big regressions — which is what the header comment always claimed it was for. Precise, deterministic regressions remain covered noise-free by the alloc_budget integration test in the normal cargo test job.

Verification

Local, against the real revisions (merge base 31f7e652, #141 head):

Case	should_run	Expected
Real #141 (release version bump)	false	false
External dependency version change (garde)	true	true
Profile change (profile.bench.codegen-units)	true	true
New dependency added	true	true

• The Determine PR benchmark scope step was extracted from this YAML and executed under bash -eo pipefail (matching GitHub's shell: bash) against the real revisions — it emits merge_base=31f7e652… and should_run=false, confirming the heredoc-inside-block-scalar indentation is correct.
• Workflow parses cleanly; 8 steps with the expected conditionals.
• cargo bench -p vespera_inprocess --bench dispatch -- --list still resolves all 15 filtered benches.
• jni-bench.yml and CI.yml are untouched.
• CI on this PR is green, including the Criterion gate running its own new logic.

Cost

PR runs measure benches twice, so bench wall-clock roughly doubles (3m50s → ~6m40s). Cargo build outputs are reused across the two measurements within the job, so compile time is not doubled. Release PRs like #141 now skip the gate entirely and get faster.

Note

This must land before #141 can go green — #141 re-run under the current main workflow keeps comparing against the cached cross-machine baseline.

[github-actions]

Changepacks

vespera@0.3.1 → 0.4.0 - Cargo.toml

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.

vespera@0.3.1 → 0.4.0 - crates/vespera/Cargo.toml

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.

vespera_core@0.3.1 → 0.4.0 - crates/vespera_core/Cargo.toml

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.

vespera_inprocess@0.3.1 → 0.4.0 - crates/vespera_inprocess/Cargo.toml

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.

vespera_jni@0.3.1 → 0.4.0 - crates/vespera_jni/Cargo.toml

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.

vespera_macro@0.3.1 → 0.4.0 - crates/vespera_macro/Cargo.toml

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.
• Fail the build when merged apps define conflicting same-named OpenAPI schemas, turning a previously silent first-wins condition into an actionable compile error while preserving identical-schema deduplication.

Patch

• Cover every export prefix branch and require 100% Rust line coverage in CI.

vespera-bridge@0.3.1 → 0.4.0 - libs/vespera-bridge/build.gradle.kts

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.

vespera-bridge-gradle-plugin@0.3.1 → 0.4.0 - libs/vespera-bridge-gradle-plugin/build.gradle.kts

Minor

• Add an explicit export_app! prefix option that namespaces routes, OpenAPI paths, and generated component schemas.