ci: benchmark comparisons always use main branch baseline

[Copilot]

Benchmark comparisons were using PR-specific cache as the baseline, masking cumulative performance degradation across multiple commits. A PR with commit 1 degrading by 5% and commit 2 degrading by another 5% would show as two separate 5% regressions instead of cumulative 10% degradation from main.

Changes

Cache Strategy

• Before: Restore keys prioritized PR-specific cache, then fell back to main
• After: Only restore from main branch baseline
• All PR commits now compare against the same main baseline

Override Benchmark Threshold

• Added override-benchmark-threshold PR label support
• When present: benchmark runs and comments but doesn't fail PR
• Enables intentional performance trade-offs with full visibility
• Label check includes error handling with safe defaults (fails on error)
• Only applies to PRs; main branch always fails on regression

Configuration

restore-keys: |
  ${{ format('{0}-benchmark-main-', runner.os) }}

fail-on-alert: ${{ github.event_name != 'pull_request' || (steps.check-label.outputs.has_label || 'false') != 'true' }}

Documentation

• Updated benchmark README with comparison strategy and label usage
• Added micro-benchmark section to main README

Original prompt

The cache key for benchmarks should always compare to the last successful benchmark in the main branch. Otherwise, we might see failures when committing several commits to a branch that still succeed when compared to the merged code. We should also have a way to accept degraded performance in CI.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[LayZeeDK]

@copilot Change the PR label to override-benchmark-threshold.

[Copilot]

@copilot Change the PR label to override-benchmark-threshold.

Done in 739baef. Renamed the label from accept-performance-degradation to override-benchmark-threshold across all files (workflow, README, and benchmark documentation).