Fix omnibox overlay hosting

docs/tables_frontend_overhaul_implementation_plan.md

@@ -60,6 +60,7 @@ It is intentionally implementation-focused:
 | 2026-03-21 | P2.8 | Completed | Added a shared `TableContextState` service on top of browser storage and the URL serializer, moved the `Tables` page off page-local table selection persistence, and switched diagnostics to the same restore/persist/build-URI flow so table context logic now lives in shared frontend state instead of being reinvented per page. |
 | 2026-03-21 | Post-P2 fix 1 | Completed | Fixed the shell omnibox drawer regression by adding explicit shell offset variables, constraining drawer/body scrolling, and giving the omnibox its own backdrop geometry so the flyout opens within the visible viewport instead of collapsing into invalid top/bottom positioning. |
 | 2026-03-21 | Post-P2 fix 2 | Completed | Rebuilt the shell omnibox as a dedicated command palette instead of a repurposed drawer, with shell-owned overlay markup, explicit viewport-safe geometry, autofocus, Escape and navigation close behavior, and a stable scrollable result body. |
+| 2026-03-21 | Post-P2 fix 3 | Completed | Moved omnibox overlay ownership from the header subtree into `AppShell` itself via a shared omnibox state service and a top-level palette host, which restored full-screen backdrop coverage and reliable outside-click close behavior. |
 
 ### Lessons Learned
 
@@ -89,6 +90,7 @@ It is intentionally implementation-focused:
 - The omnibox foundation does not need the full final interaction model to be useful. A drawer with real table search, real pinned/recent data, and a small slash-command set is enough to validate the shell surface before Phase 3 builds deeper index and inspector flows on top of it.
 - Shared overlay primitives should not depend on undeclared layout variables. If a drawer needs shell offsets, the shell must define them explicitly and overlay-specific backdrops should be adjustable instead of assuming full-screen dimming is always correct.
 - A command palette is not just a styled drawer. It needs shell-owned geometry, predictable focus behavior, and a bounded scroll region; treating it as a generic side panel led directly to the layout regressions found in Phase 2.
+- Backdrop and outside-click behavior depend on overlay ownership as much as CSS. If the trigger owns the overlay inside a sticky header subtree, fixed-position assumptions can break; shell-level overlays should be rendered by the shell, not by individual header controls.
 
 ## Target Outcomes