Authored by Claude Fable 5 (Claude Code). Session 567191c3-16f6-4235-914c-b51fc94d1514.

Resolves #12880 Refs #12883 Refs #12886

The first fix produced WITH the #12886 eyes — every defect below was localized by trace evidence, not code-reading. Two root causes, both verified live on devindex (locked AND no-locked):

1. Duplicate drag:move delivery (container.SortZone): the same pointer position reaches the owning zone more than once (stacked delegated drag listeners along the DOM path — the listener census shows drag:move registered on every header button AND its toolbar). Re-processing an unchanged position re-qualifies delta against already-mutated itemRects → one spurious extra switch per duplicate. Fix: in-bound same-position guard (trace-visible as dup suppression events); the overdrag auto-scroll path is explicitly exempt (its recursion deliberately re-feeds the same position — the regression class that killed PR #12882 is structurally avoided).
2. Coordinate-space mismatch (the locked-only mechanism): grid/table header SortZones convert itemRects to owner-relative space via adjustProxyRectToParent, but the delta math only honored the adjustItemRectsToParent flag → viewport cursor compared against owner-relative rects → delta inflated by the toolbar's viewport origin (~388px for a locked grid's center region = every move switches; ~17px unlocked = imperceptible — exactly the operator's "only with locked columns" datum). Fix: the delta space-shift now fires when EITHER conversion mechanism is active.
3. Rides along: NeoArray import added to grid/Container.mjs (updateColCount's end-region-removal branch threw ReferenceError: NeoArray is not defined — caught live by the new eyes; it killed every end-region lock-change: gpt's deterministic end→center no-op and the C5 re-home race). The remaining cross-toolbar DOM-move gap (engine+vdom re-home correctly, the rendered DOM doesn't apply the move — verify_component_consistency's first real catch) stays on #12883 as a precisely-bounded follow-on leg.

Also upgrades the get_drag_trace move events with delta/offsetX/scrollLeft/slotLeft fields — the observability that convicted defect 2.

Evidence: L3 (live matrix on devindex via the #12886 eyes — get_drag_trace decision traces with dup-suppression visible, landing assertions, verify_component_consistency three-surface checks): Case 1 within-center: Total → index 2 EXACT intent (was 5); Case 2 within-start: [#, User, Rank] no ejection; no-locked control: +1 slot, geometry-equivalent parity with locked; end-column unlock: the exact op that threw ReferenceError now succeeds. → L3 required (#12880 AC2/AC3/AC4). Residual: none for #12880; the DOM-move leg is #12883's.

Test Evidence

• npx playwright test test/playwright/unit/draggable/container/SortZone.spec.mjs test/playwright/unit/draggable/grid/header/toolbar/SortZone.spec.mjs test/playwright/unit/ai/client/ComponentService.spec.mjs -c test/playwright/playwright.config.unit.mjs → 14 passed
• Decision traces (pre/post-fix) + the listener census + consistency reports: documented on #12880 / #12883

Deltas from ticket

• #12880's instrumentation-first AC1 is satisfied by the merged #12886 get_drag_trace (permanent substrate instead of throwaway logging) — the traces in this PR's evidence ARE that instrumentation.
• The fix landed in the parent for the delivery/space bugs (affects table headers too — same latent class at small offsets) with zero grid-subclass changes; gpt's #12882 resolver concept was ultimately unnecessary for landing accuracy once the REAL mechanisms were visible.

Post-Merge Validation

• Operator: locked devindex drag matches the no-locked feel (within-region landings at the drop band)
• #12883's DOM-move leg picks up from the consistency-probe evidence