Bulk accept can send more than one smoke box per frame

Follow-up from #286 / #318, which added a server-side validator rejecting more than one smoke box per detection annotation.

buildQuickSubmitPlan (frontend/src/utils/annotation/quickSubmitUtils.ts:143) calls materializeReviewAnnotation with rejected: new Set() and humanRects: [], so items.length === winningBoxes.length — every box on the frame's winning model layer becomes a smoke item. If a frame's winning layer ever holds two or more boxes, the bulk-accept flows all fail with a 422 and surface the generic toast "Failed to accept boxes — try again":

• per-object quick accept (runLaneQuickAccept, LocalizeAlertPage.tsx:752)
• "Accept remaining" from the accept popover
• "Accept all & submit alert"

The LocalizeObjectEditor is unaffected — it enforces one box client-side, and it is the escape hatch if this ever fires (open the frame, reject the extra box).

Measured on the dev database, 2026-08-06: zero affected frames. Max boxes on the winning layer is 1, across 288 auto-layer and 9,759 engine-layer frames. This is latent, not live.

Why it is not structurally prevented: worker.py's keep_boxes_overlapping retains every sensitive-model prediction that overlaps the lane's anchor, so a model change could start emitting two boxes on one frame. algo_predictions could likewise carry two.

What this needs: a rule for which box wins when the winning layer has more than one — highest confidence, or reuse focusOnMainObject (already used for display in collectLaneBoxes, quickSubmitUtils.ts:106, but deliberately not in the submit plan). Then cap the submit payload at one smoke box per frame and cover it with a frontend test.

Note this is the same "which box wins" question that blocks #315 on the sequence-annotation side, though the two are otherwise independent.

Start in frontend/src/utils/annotation/quickSubmitUtils.ts at buildQuickSubmitPlan and compare its winning-box handling with collectLaneBoxes and focusOnMainObject. Trace the callers in LocalizeAlertPage.tsx, then add a frontend test covering multiple boxes on a winning layer; done means each bulk-accept payload contains at most one smoke box per frame and the existing flows remain covered.