The user-label cache is re-parsed from localStorage on every render

Start with readCache and readCachedUserLabels in desktop/src/features/profile/lib/userLabelStorage.ts, then inspect resolveUserLabelPlaceholderData at desktop/src/features/profile/hooks.ts:378 and the existing memoization at hooks.ts:98. Add the regression coverage described in the issue and verify unchanged cached bytes are parsed once while returned values remain correct.

Describe the bug

readCache() in desktop/src/features/profile/lib/userLabelStorage.ts runs once
per React render, per query observer. Each run does a localStorage.getItem, a
JSON.parse of up to MAX_CACHED_LABELS (1,000) profile entries, and a full
rebuild of the lowercased profile map. The caller then throws the result away.

Mechanism

readCache is reached through resolveUserLabelPlaceholderData, which
desktop/src/features/profile/hooks.ts:378 hands to React Query as
placeholderData:

placeholderData: (previousData) =>
  resolveUserLabelPlaceholderData(
    previousData,
    relayUrl,
    normalizedPubkeys,
  ),

React Query invokes placeholderData on every result computation — once per
render, per observer. useUsersBatchQuery has roughly 35 call sites across 31
files, and UserProfilePopover calls it twice and mounts per message row, per
avatar, per member-list entry. Every rendered username is an observer.

The worst case is the common one. readCachedUserLabels calls readCache
before it inspects pubkeys:

export function readCachedUserLabels(relayUrl, pubkeys) {
  const cache = readCache(relayUrl);   // full 1,000-entry parse happens here
  if (!cache) return undefined;
  ...                                  // pubkeys not consulted until here
}

UserProfilePopover gates its queries on enabled: isOpen and passes
open ? [pubkey] : []. So while a popover is closed there is no
previousData, the ?? falls through, and a full parse runs to produce a
guaranteed undefined. Two observers per closed popover.

Steps to reproduce

1. Use Buzz until the label cache is populated — check
   localStorage["buzz-user-labels.v1:<relay>"].
2. Open a channel with long scrollback so many UserProfilePopovers mount.
3. Minimize the window. Do not interact with the app.
4. Attach a CPU profiler to the WebView renderer, or just watch the process.

Quicker check, in the devtools console with the app idle:

const orig = JSON.parse; let n = 0;
JSON.parse = (...a) => { n++; return orig(...a); };
setTimeout(() => { console.log("parses in 10s:", n); JSON.parse = orig; }, 10_000);

Expected behavior

An idle, minimized window should be near 0% CPU. A cache read that returns the
same bytes should not re-parse them.

Measurement

45-second CPU profile of the minimized app over the WebView2 debug port:

[profile] window 46446 ms, 17245 samples

React rendering, total          23.0% of one core
readCache                       largest JS leaf by self time
(garbage collector)              6.7%
main thread idle                66.5%

No hot loop — thousands of small parses. The garbage collector sitting second is
the signature: allocate a 1,000-entry object graph, read one field, drop it,
repeat.

Profiled call chain from the running app, minified names:

readCache
  <- createResult
  <- getOptimisticResult
  <- UserProfilePopover
  <- React scheduler

Version and platform

• Built from main. The file is unchanged since d0d4acd4f (#3317), which
  introduced it.
• Windows 11, WebView2.

Additional context

This looks like an oversight rather than a design decision, because the same fix
is already applied one file over. hooks.ts:98 wraps readSelfProfileCache in a
React.useMemo with the comment "Parse localStorage once per relayUrl/pubkey
pair — not on every render", and selfProfileStorage.ts carries a module-level
memo. userLabelStorage.ts simply never got the same treatment.

No existing gate catches this: the returned value is correct, only the call
count is wrong. Unit tests, typecheck and lint are all green today.

I have a fix — memoise the parse keyed on the raw string, about six lines, plus a
regression test that asserts the JSON.parse call count rather than the returned
value, because asserting the value passes identically with and without the fix.
Verified against the unfixed file: 25 parses across 25 lookups before, 1 after.
Happy to open a PR if the approach sounds right.

Related but deliberately not bundled: normalizedPubkeys in
useUsersBatchQuery (hooks.ts:325) is rebuilt every render — map + Set +
filter + sort, no useMemo — and it also feeds the query key. Smaller cost,
separate change, happy to file separately.