fix(search): consolidate dashboard, cross-mode, and composer fixes

docs/answer-page-redesign-handover.md

@@ -731,3 +731,85 @@ panel:
 card, the rail cards, the drawer and print are fixed together. Display only.
 
 ---
+
+### 12.9 What direction B cost, and how it was repaid (2026-09-07)
+
+Keeping one panel was right. Keeping the narrower one was a coverage loss nobody wrote down at
+the time, and it is invisible from a screenshot.
+
+`UniversalSearchAlsoMatches` is fed by `/api/search/universal` and reaches twelve domains.
+`CrossModeLinksSection` was fed by four catalogues loaded in the browser — medications,
+services, forms, differentials. So from 2026-08-26 a clinician who asked a question in Answer
+mode could never be shown a matching DSM diagnosis, dictionary term, formulation, specifier,
+therapy or tool, though all sixteen other modes could. The register in
+`tests/universal-also-matches-mode-coverage.test.ts` recorded the exemption honestly and
+recorded the wrong thing about it: that Answer answers cross-mode discovery elsewhere, without
+saying that "elsewhere" reached a third of the surface the tray reached.
+
+**Widened rather than replaced.** The line keeps its disclosure, its place inside the answer
+thread between the governed caution and the follow-ups, its walk back through earlier turns when
+a follow-up drops the entity name, and the per-record `Search in <mode>` control and
+`cross_mode_link_open` telemetry the mode-level tray has never had. What changed is only where
+the links come from: `buildCrossModeLinksFromUniversalSearch` maps universal-search groups into
+the same `CrossModeLink` shape, gated on a query term naming the record at a word boundary — the
+same weight and minimum term length the differentials path already used, because a content-only
+score cannot be trusted on a question full of filler.
+
+Three things hold it inside the 12.8 decision rather than undoing it:
+
+- The lookup is opt-in per surface (`universalMode`) and off by default. Every mode but Answer
+  mounts the tray, so a surface that ran this too would print one record twice — which is what
+  §12.8 removed, and what the prescribing page later shipped anyway.
+- The consumed domains and their complement come from one array. `documents` is excluded because
+  an answer already cites its documents in the evidence rail and the drawer, and the four
+  catalogue domains are excluded because the local half already resolves them.
+- It is withdrawn while a generation is in flight, matching `answer && !loading` on the tray, so
+  the lookup never races the answer stream and an open tray never holds matches for the question
+  being replaced. An errored or empty group contributes nothing.
+
+The strip's ceiling stays at four, the number the catalogue half alone allowed. Five was set
+while this was still a collapsed disclosure, where an extra row cost nothing until the reader
+opened it; with the strip open at every width every row is unconditional height on the answer
+surface, so the old ceiling holds and this change stays about reach rather than size. The cap
+only binds when the catalogue half filled the strip on its own, which is a question already well
+served; the common answer names none of those four record types, and that is where the widened
+reach earns its place.
+
+Calculators, Factsheets, Sources, On Call and Favourites are still unreachable from here, and
+not by omission — they contribute no cross-entity search domain at all
+(`universal-search-mode-context.ts`), so nothing can resolve a link to them from any surface.
+
+**The line keeps its single toggle, and now rests open (owner decision, 2026-09-07).** Both ends
+of this were built and looked at on the day the data was widened, and both were wrong.
+
+Closed by default, as direction B shipped it, hides the matches behind a tap nobody has a reason
+to take: a closed tray cannot show what it holds, so the reader has to spend the tap to find out
+whether it was worth spending. Removing the toggle entirely, which was the first correction, put
+four unconditional rows into the answer's evidence stack with no way to put them away — roughly
+90px of resting height on a phone, on the densest surface in the product, which is the trade
+§12.8 made in the other direction for a reason.
+
+The hybrid is one control that starts open. The matches are there to be read without a tap, and
+the space is one tap away for a reader who wants the answer and the follow-ups closer together.
+`variant="line"` and `CrossModeLinksLine` are the presentation again, so the `card` strip stays
+what the DSM diagnosis page and the mockups use.
+
+The header lost its preview of the first three record names in the same change. Those names
+earned their room while the line rested closed and had no other way to say what it held; open,
+they were naming the cards immediately below them. The header is label, rule and count now, at
+every width, which is the device the "Also matches" tray already uses.
+
+Worth recording because it nearly went the other way: `CrossModeLinksLine` was almost deleted
+while it briefly had no consumer. `check:dead-code-candidate` refused the removal — the symbol
+was introduced on 2026-08-26, inside the gate's 30-day window ("likely awaiting its consumer") —
+and the deletion was dropped rather than forced past the threshold. A day later it was the
+presentation again. The gate's heuristic looked wrong at the time and was right in substance.
+
+Three browser tests carry the disclosure as their contract. The 1280px case asserts the rail's
+own computed display in both states rather than the trigger's word for it: width is where this
+broke before, when a `hidden` beside a `md:flex` in one class list lost to the media-query rule
+from 768px up and left a rail painted open while its trigger reported it closed. Now that the
+line rests open, the same mechanic would hide a broken collapse instead, which is why the closed
+state is the one pinned on computed display.
+
+---

docs/branch-review-records/173c8dc623608fe60a08ef7070b302e3649b721fce26c5690363f025b23b6d14.record.md

@@ -0,0 +1 @@
+| 2026-09-07 | claude/side-navigation-reliability-0sg48b | ee0cf48b5ae370475de1526bd843ac9f62ef55a2 | prlanded | PR #2637 merged 2026-09-05T11:11:08Z by BigSimmo via true merge commit ee0cf48 (auto-merge armed, three update-branch merges of main by the owner beforehand). Tree of ee0cf48 is identical to branch tip a943909 (git diff --stat empty); fix commit 1a6adff is an ancestor of origin/main; no late or orphaned commits; remote branch already deleted by GitHub. Sidebar shortcuts now switch mode in place via onSelectMode and useHomeModeSeed consumes the UI-change flag one-shot. | git diff --stat ee0cf48 a943909 empty; git merge-base --is-ancestor 1a6adff origin/main true; PR read via GitHub MCP state=closed merged=true; CI on head a943909 green before merge |

docs/outstanding-issues-inbox/45cda6a2-725f-4fc7-bbd8-2350c7280d72.json

@@ -0,0 +1,14 @@
+{
+  "version": 2,
+  "id": "45cda6a2-725f-4fc7-bbd8-2350c7280d72",
+  "createdOn": "2026-09-07",
+  "action": "add",
+  "payload": {
+    "pri": "P2",
+    "type": "issue",
+    "summary": "PR required failed repeatedly on a UI-only branch with no cause ever identified",
+    "detail": "Session 2026-09-07, branch claude/jolly-keller-vy36h2 (the also-matches disclosure). The pr-required aggregate went red several times while every local gate was green: 15 static gates, whole-tree Prettier, 229 + 161 focused unit tests, 129 + 28 Chromium tests. A push-cancellation theory was offered and then disproved by a run with a quiet window. CI log access was blocked by the babysit marker for the whole session, so the failing job was never read and the branch merged with the cause unknown. Next action: on the next unexplained pr-required failure, unlock the marker first (CLAUDE_ALLOW_PR_FOLLOW=1) and read the failing job log before theorising. Two contributing factors worth ruling out: a second agent was pushing merge commits to the same branch concurrently, and main merged about twelve changes during the session.",
+    "source": "docs/branch-review-records + session 331c4d4d",
+    "issueUlid": "01M1XP1FE7K06J63XSYVBS40HA"
+  }
+}

docs/outstanding-issues-inbox/4d148590-e6c2-4dde-b07d-111f312c827e.json

@@ -0,0 +1,11 @@
+{
+  "version": 2,
+  "id": "4d148590-e6c2-4dde-b07d-111f312c827e",
+  "createdOn": "2026-09-07",
+  "action": "update",
+  "payload": {
+    "id": "#ZKR5YK",
+    "detail": "UPDATE 2026-09-06: this row is now partly stale and its Sources example is gone. PR #2672 consolidated /sources: its four-card mode home was deleted and the bare path now redirects to the shared home at /?mode=sources, so Sources no longer owns a home of its own and the mode pill landing there is correct rather than a miss. The row's remaining live cases are Favourites and Tools. Favourites is settled by an explicit owner decision (#V0EDR4, 2026-08-23: two purpose-specific surfaces are deliberate), which leaves Tools as the only open instance: changeMode still builds appModeSelectionHref, so picking Tools lands on /?mode=tools and the legacy client-only launcher (applications-launcher-page.tsx, data-testid tools-home) rather than canonical /tools (ToolsSearchResultsPage). Verified on main 0177bed by a browser probe: /?mode=tools renders tools-home. Documents and Medication also dropped off this row for the same reason as Sources - both bare paths redirect now. STILL AN OWNER RULING, not a defect to fix unilaterally: the pill deliberately always returns to the shared home, so special-casing Tools trades one inconsistency for another, and removing the second launcher means deleting a component that six assertions in tests/favourites-auth-gate.dom.test.tsx still render directly. Re-scope this row to Tools only.",
+    "baseRowFingerprint": "4a2c45298f781607394bb54bdde280518d7e1e6c1706ae00e2b05526613fa232"
+  }
+}

docs/outstanding-issues-inbox/aca18f9d-f748-40cd-918a-51ecb9b91f1c.json

@@ -0,0 +1,10 @@
+{
+  "version": 2,
+  "id": "aca18f9d-f748-40cd-918a-51ecb9b91f1c",
+  "createdOn": "2026-09-07",
+  "action": "cancel",
+  "payload": {
+    "requestId": "4d148590-e6c2-4dde-b07d-111f312c827e",
+    "reason": "Withdrawn in favour of the competing request 4036f064, a done on the same row backed by an actual implementation (outcome: mode pill now routes to dedicated mode homes). My request only re-scoped #ZKR5YK to Tools after verifying that its Sources, Documents and Medication examples are gone (all three bare paths redirect now) and that Favourites is settled by owner decision #V0EDR4. A re-scope is pointless against a close that supersedes the whole row, and two pending mutations on one row block reconciliation for everyone (#686WHW). The verified findings are recorded in the PR that carried this session's fixes instead, so they are not lost if that close is ever reopened."
+  }
+}