docs(ui): design global search over the board and chat corpora (RIG-1621)

docs/designs/DECISIONS.md

@@ -272,6 +272,7 @@ check enforces the mechanical half. Full rationale:
 | DL-208 | Both Solid-2 blocker deps are FORKED into the RigelBuild org and consumed as externalized pinned artifacts, never vendored into the compass tree (matching the org-wide fork-externalization ruling): `@tanstack/solid-virtual` is forked with Solid-2 support + an upstream PR to TanStack/virtual (a 136-line thin wrapper over the framework-agnostic `@tanstack/virtual-core`, likely to land), consumed via a `package.json` git ref (`github:RigelBuild/virtual`) until upstream merges then dropped; `solid-markdown` (effectively abandoned upstream) is forked-and-owned, published as `@rigelbuild/solid-markdown` with the `solid-js` 2.x peer bump + Solid-2 codemods (imports + effect lifecycle, no algorithmic rewrite, no in-repo AST renderer). Overrides this record's earlier adapter-over-core virtualizer recommendation | Active (Matt, 2026-08-18) / solid-markdown clause superseded by DL-218 (Matt, 2026-08-19) | [solid v2 migration §Resolved decisions](ui/compass-ui-solid-v2/design.md#resolved-decisions) |
 | DL-209 | The Wails packaged-shell seam under Solid 2 is low-risk-but-verified (no automation): the seam is Solid-free plain TS (`daemon-transport.ts` string-name `Call`/`Events`; `MarkdownText.tsx:1` `Browser.openExternal`), so effect-flush timing is the lesser risk; the packaged shell is the first place the babel-preset-solid 2 / vite-plugin-solid 3 / `@solidjs/web` production-build compilation and v2's render-root-owned delegated events run outside dev/happy-dom, so S7's manual packaged-shell smoke (RPC round-trip + openExternal) stays the mandatory gate — no automation justified for a one-shot migration | Active (Matt, 2026-08-18) | [solid v2 migration §Resolved decisions](ui/compass-ui-solid-v2/design.md#resolved-decisions) |
 | DL-218 | The owned `RigelBuild/solid-markdown` fork adopts upstream `andi23rosca/solid-markdown` #44 (the react-markdown-10 API rewrite — a third-party draft, upstream effectively inactive on it) + #45 (the bundle-deps jsdom-compat fix), re-ported to Solid 2 via a fork-local `src/jsx-runtime.ts` `jsx`/`jsxs`/`Fragment` triple over `createComponent`/`merge` + `@solidjs/web` `Dynamic` (solid-jsx dropped as Solid-1-bound), published as a clean `@rigelbuild/solid-markdown@3.0.0` (no `-rigel.N` suffix — the fork is ours and sets its own versions); this retires the fork's in-repo AST renderer in favor of the maintained `hast-util-to-jsx-runtime`. The apps/ui consumer injects @mention chips via a consumer-side rehype plugin over the existing `mentionRuns` splitter (the removed `text` component override), derives inline-vs-block code structurally via a `pre`/`code` override split (the removed `inline` prop), keeps `renderingStrategy="reconcile"` (un-deprecated in the owned fork), and adds a synchronous consumer-side `(lang, code)` highlight cache that `CodeBlock` reads at initial render (seeding highlighted HTML and bypassing the debounce + async `createResource` on a hit) to suppress the code-highlight flash caused by #44's per-tick component teardown (accepting marginal within-stream selection loss; the fine-grained-walker re-port is the escalation if it bites). Supersedes DL-208's `solid-markdown` clause ONLY (no-algorithmic-rewrite/no-in-repo-AST-renderer) — DL-208's `@tanstack/solid-virtual` clause and its org-wide fork-externalization stance stay Active | Active (Matt, 2026-08-19) | [markdown react10 §Approach](ui/compass-ui-markdown-react10/design.md#approach) |
+| DL-346 | Compass global search is a SERVER-side corpus reached over the wire, not a client-held index: a persistent center-of-top-bar input coexists with the D5-frozen Cmd+K palette (both feed one provider set), issues/PRs search the WHOLE board rather than the caller's assigned slice, and chat search reaches the already-shipped-but-UI-unreachable `rpc SearchMessages`. Issue search is new server work — an `issues.search_tsv` GENERATED ALWAYS … STORED column + GIN index + `SearchIssues` RPC, mirroring the messages pipeline. Two schema constraints are load-bearing and measured on PostgreSQL 18.4. (1) The labels term MUST route through a `compass_labels_text` SQL wrapper declared IMMUTABLE with a pinned `search_path`: bare `array_to_string` (and `labels::text`) is only STABLE and fails the migration with "generation expression is not immutable", while the IMMUTABLE `array_to_tsvector` alternative compiles but bypasses the parser so label search silently matches nothing. (2) The column labels its four fields with `setweight` (title A, body/summary B, labels C) and queries pass the matching `{0.1,0.2,0.4,1.0}` weights array to `ts_rank` — a flat concatenation ranks a short-title hit and a hit buried in a 10 KB body EXACTLY equal (0.06079 vs 0.06079 measured; weighted 0.60793 vs 0.24317), so relevance ordering is unachievable without it, and the column is STORED so this cannot be deferred without a table rewrite. PR results stay a client-side `prRows()` derivation over `Issue.prs` and are FIXTURE-ONLY at v1 — the field has zero Go writers — so no live-mode PR assertion may gate the release | Active (Matt, 2026-09-07) | [global search](ui/compass-global-search/design.md) |
 
 ## Threading & rendering