Release PRs trip the KNOWN_REGRESSIONS staleness guard when version jumps >1 minor ahead of the benchmark baseline

[carlos-alm]

Observed on PR #1701 (release v3.15.0)

The Pre-publish benchmark gate failed on the KNOWN_REGRESSIONS entries are not stale test (tests/benchmarks/regression-guard.test.ts:691): 11 entries keyed to 3.12.0/3.13.0 were flagged as >1 minor version behind the bumped package.json version (3.15.0).

Root cause

The committed benchmark baseline data (generated/benchmarks/*.md) was stuck at 3.13.0 — there was no docs: update performance benchmarks (3.14.0) PR (data jumps 3.13.0 → nothing, even though a v3.14.0 tag exists). Benchmark recording happens post-publish via benchmark.yml (triggered on the Publish workflow's workflow_run), so:

• The release PR bumps package.json to the new version immediately.
• The staleness guard measures entry age against package.json, not against the latest recorded baseline.
• When the package version is ≥2 minors ahead of the latest recorded baseline, every still-live exemption keyed to the previous baseline is flagged stale — even though it's still the active baseline for the dev-vs-baseline comparison.

This made the failure look like a routine 'prune stale entries' chore, but naively deleting the entries before the new baseline lands would have unmasked the dev-vs-old-baseline comparisons instead.

Why 3.14.0's recording is missing

Worth confirming whether 3.14.0 was published through publish.yml (which should have triggered benchmark.yml) and, if so, why no chore/bench-v3.14.0-* PR was created/merged.

Suggested fixes (pick one)

1. Measure staleness against the latest recorded benchmark version, not package.json — the guard should only consider an entry stale once a newer baseline that supersedes it has actually landed.
2. Allow a grace window (e.g. entries within MAX_VERSION_GAP minors of the latest recorded baseline are not stale), so a release PR doesn't fail before its post-publish benchmark PR merges.
3. Ensure the post-publish benchmark-recording PR is created/merged for every release (and investigate the missing 3.14.0 one) so the baseline never lags the package version by >1 minor.

Immediate unblock (already applied on #1701)

Merged the 3.15.0 benchmark data (PR #1702) into the release branch, then pruned the now-superseded 3.12.0/3.13.0 entries (commit 65f9ce7). Tracking the systemic fix here.