fix(desktop): make Check all match checking each harness on its own

Problem

In Settings > Harnesses, Check all marks every harness as failing. Pressing Check on each card one at a time mostly reports Ready.

runAllHarnessChecks (apps/desktop/src/state/harnessChecks.ts:73) starts every check at once with Promise.all. Each check then:

calls checkHarnessAccounts(true, [id], false) → _rustcode/harnesses/account/check with redetect: [id], and each of those also fires refreshPlanCache(true) (a Keychain read for Claude);
then calls fetchHarnessModels(id, true), which starts the harness process to list its models (60 s timeout), followed by refreshHarnesses().

So N harness processes, N redetects, N login-shell/Keychain reads and N full snapshot refreshes all run concurrently. Likely failure modes:

Each answer replaces the whole harnesses snapshot. A stale snapshot that arrives late (from an earlier request, holding another harness's cached detection) overwrites a fresh one, and currentResult then marks a Ready card as Not installed or Needs sign-in.
Concurrent harness starts and probes hit timeouts (the shell probe has a 3 s limit, see the Homebrew PATH issue) that a single check never hits.
N parallel Keychain reads raise prompts or get denied.

Run Check all with limited concurrency (2 at a time) or one after another. Show progress ("Checking 2 of 5…").
Do one account/check with redetect set to all ids and one plan read up front, then list models per harness.
Apply a snapshot only if it is newer than the one held (a request sequence number in harnessCall).
Show each failure's real error text on the card, not a generic failed state.

In the mock and in a unit test with delayed, out-of-order responses, Check all ends in the same per-card results as checking each card alone.
Only one Keychain/plan read per Check all.
bun run check and harnessChecks.test.ts cover the ordering.