Overlay perf coverage must include annotation-backed widget churn

Problem

The decoration / annotation perf story looked more complete than it really was. We had browser timing for huge-document overlays and locality proof for overlay toggles plus hidden panes, but annotation-backed widget rebasing still had no perf-lane owner.

Symptoms

pnpm bench:react:rerender-breadth:local proved overlay-source toggle locality and hidden-pane Activity behavior, but nothing about annotation-driven widget churn.
Widget behavior was covered in correctness tests, which is not the same thing as a perf or locality benchmark.
It was too easy to say “overlay perf is covered” and miss the widget hole.

What Didn't Work

Treating correctness coverage as if it were perf coverage. The annotation-store-contract.tsx and widget-layer-contract.tsx tests proved semantics, not benchmark ownership.
Pretending the existing huge-document overlay lane covered everything. It measures browser wall time after overlay and sidebar churn, not mounted rerender breadth for annotation-backed widgets.

Solution

Extend the existing rerender-breadth lane with one annotation-backed widget row instead of inventing a new benchmark family.

The added row:

creates one annotation bookmark on the left block
derives one widget from that annotation
performs a local text insert that rebases the anchor
measures rerender breadth across:
- edited left text
- annotation projection slice
- annotation sidebar slice
- annotation-backed widget slice
- unrelated right text

Fresh result:

left text rerenders: 1
annotation projection rerenders: 1
annotation sidebar rerenders: 1
annotation widget rerenders: 1
unrelated right text rerenders: 0
mean edit cost: 0.29ms

Relevant files:

Why This Works

It closes the real gap instead of broadening the benchmark matrix for no reason.

Overlay toggle locality still proves projection-source churn.
Hidden-pane Activity still proves offscreen ownership and resume behavior.
The new row proves annotation store updates and annotation-backed widget visibility stay local too.

Together, those rows finally cover the overlay family as a family instead of as two isolated tricks.

Prevention

When auditing overlay perf, explicitly check three churn classes: projection-source toggles, hidden-pane resume, and annotation-backed widget rebasing.
Do not treat correctness tests as perf ownership.
Prefer extending an existing locality lane before creating a brand-new benchmark family.
If a rerender benchmark writes artifacts under a filtered package cwd, record the real artifact path in docs instead of guessing a repo-root .tmp/ path.