frontend: clear the 99 lint errors

`npm run lint` reported 99 errors and 26 warnings across 65 files and
had presumably been failing for a while — next.config.ts sets
eslint.ignoreDuringBuilds, so the build never surfaced it.

68 were in vendored code: components/charts and components/ai-elements,
pulled from upstream registries. Re-linting those reports upstream's
style back at us, and "fixing" them means diverging and eating conflicts
on every update — ai-elements already has exactly this carve-out on the
TypeScript side via ignoreBuildErrors. Ignore both, plus the two registry
files in components/ui (carousel publishes its api from an effect; the
sidebar skeleton picks a random width).

The other 31 were ours, nearly all react-hooks/set-state-in-effect on
the same shape: an effect that re-seeds form state when a dialog opens or
a selection changes. Moved to render-phase adjustment, which is both what
React recommends and a real fix — the effect version paints one frame of
the *previous* record's values before correcting itself. Two carried
sharper bugs: the employee dialog could keep a typed password across a
switch to another member, and use-wallet-sync could carry `linked` over
to a newly-selected patient, briefly offering to push a record to
someone else's wallet.

The rest: useIsMobile and speech-support detection become
useSyncExternalStore (correct on first paint, no mount flash); refs
mirroring state are written in effects rather than during render; the
care-team fetch moves into its effect behind a reload key, dropping an
exhaustive-deps suppression.

Two effects in chat-panel keep the rule disabled with a reason. Both are
what effects are for: draining the queued-message buffer when the
transport goes idle, and resolving ?thread from the URL — the latter
mints an id with nanoid(), so moving it into render would just trade this
error for a purity one.

Also removes a dead /explore-era import and unused directives found on
the way. lint now exits 0, with the ai-elements tsc carve-out unchanged.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

frontend/components/chat/chat-input.tsx

@@ -9,6 +9,7 @@ import {
   useEffect,
   useRef,
   useState,
+  useSyncExternalStore,
 } from "react";
 import { useTranslation } from "react-i18next";
 
@@ -78,17 +79,23 @@ export function ChatInput({
   const [addKey, setAddKey] = useState(0);
   const fileInputRef = useRef<HTMLInputElement>(null);
 
-  // Voice dictation (Web Speech API). Detected client-side so SSR markup and the
-  // first client render agree (button starts disabled, enabled by the effect).
-  const [speechSupported, setSpeechSupported] = useState(false);
+  // Voice dictation (Web Speech API). Read through useSyncExternalStore so SSR
+  // markup and the first client render agree: the server snapshot is `false`
+  // (no Speech API to detect), and the client re-reads on hydration. Support
+  // never changes for the life of the page, so the subscription is a no-op.
+  const speechSupported = useSyncExternalStore(
+    () => () => {},
+    () => getSpeechRecognition() !== null,
+    () => false,
+  );
   const [isListening, setIsListening] = useState(false);
   const recognitionRef = useRef<SpeechRecognitionLike | null>(null);
   // The textarea contents when dictation started; transcript is appended to it.
   const dictationBaseRef = useRef("");
 
   useEffect(() => {
-    setSpeechSupported(getSpeechRecognition() !== null);
-    return () => recognitionRef.current?.stop();
+    const recognition = recognitionRef;
+    return () => recognition.current?.stop();
   }, []);
 
   const toggleDictation = useCallback(() => {