[laurentiu021]

Closes #2782.

What changes for a user

• Drive names come back after a failed read. When a read of the drives failed, a WMI service still starting at logon being enough, the Dashboard's temperature card kept the failed read as if it were the drives. Run as administrator, a drive the sensors could not name stayed "Drive 1" until SysManager was restarted. Without administrator, the drive temperatures were gone for 30 seconds. Only a read Windows answers in full is kept now, and a failed one is tried again at the next refresh, two seconds later.
• No more refused query at every refresh. Run as administrator, the card also asked for the drive models with SELECT Model FROM Win32_DiskDrive ORDER BY Index. WQL has no ORDER BY, so WMI refused the query as invalid at every refresh, and SysManager logged each refusal. It never named a drive in any release since 1.17.0. It is removed.
• System Health says when it could not read the drives. Run SMART check said "Collected 0 disk report(s)." when Windows did not answer. It now says the check failed, and why. A read that stopped partway says so after the drives it did read.

How

• DiskHealthService.ReadAsync returns a DiskHealthRead: the drives, and the fault that stopped the read, or null when it finished. CollectAsync still returns the drives alone for its callers (health score, Tune-up, the report), which judge each drive and already count an empty list as not checked. A run of failed reads is logged once, the way NvidiaGpuReader logs, and after a read succeeds the next failure is logged again. An internal constructor takes the read and the log, so a test can make a read fail.
• TemperatureService.EnrichStorageNamesAsync caches names only from a complete read and names nothing on a failed one. ReadDiskTemperaturesAsync shows a failed read's drives for that refresh but caches only a complete one. Its three catches, for exceptions DiskHealthService never lets out, are gone. Both methods are internal for the tests.
• GetDiskNamesFromWmi and its cache are gone, and ReadHardware takes no name list. A drive LibreHardwareMonitor cannot name is "Drive N", and the enricher names it, as it already did. The prefetch paired names by position anyway, and a drive with no temperature sensor adds no reading, so if the query had worked, each drive after one like that would have shown the previous drive's model.
• SystemHealthViewModel.DescribeDiskRead writes the status line. The catch for ManagementException or COMException could never run and is gone.
• ArchitectureTests.NoWmiQuery_SortsWithOrderBy scans the app's code, comments left out, for ORDER BY. It also checks that at least 20 WMI queries were read, so an empty scan cannot pass.

Tests

• Names: a failed read leaves "Drive 1" for one refresh and the next read names it. A read cut short names nothing, even when its drive count matches. A complete read is read once over three refreshes. The last one replaces the integration test TemperatureServiceMemoizeTests, which was written on the opposite premise, that an empty result from a failed read was "still a valid, cached resolution".
• Temperatures: a failed read is read again at the next refresh, a read cut short shows its drives but is not kept, and a complete read is kept for exactly DiskTemperatureTtl and read again one second later. Before this the TTL could not be reached from a test.
• DiskHealthService: three failures, a success and a failure log twice; a read that works logs nothing; CollectAsync still returns the drives of a read cut short. System Health: a failed check's status, and both wordings of DescribeDiskRead.
• ReadHardware: a drive LibreHardwareMonitor cannot name is a numbered placeholder.

Red before the fix, by mutation, each restored byte for byte afterwards:

• Caching names from any read: the failed-read and cut-short name tests fail, and the read-once test passes.
• Caching temperatures from any read: the failed-read and cut-short temperature tests fail, and the TTL test passes.
• Logging every failure: the log-once test fails.
• The old "Collected N" status: the failed-check test fails, and so does the cut-short row of DescribeDiskRead.
• An ORDER BY in a real query: the new guard fails.

Checks

• All four projects build with 0 warnings, and dotnet format --verify-no-changes is clean.
• The unit suite passed twice on the final tree: 10,706 tests, 0 failed. One skip was there before this change.
• The affected integration classes pass: 63 tests, with DiskHealthServiceTests against real WMI.
• A console harness over the built app read the real drives three times, elevated and with a standard user's token: every read complete, 2 drives. Elevated, both had temperatures; with the standard token, neither did.
• CHANGELOG (1.136.104), the csproj bump, README (Dashboard Temperatures, System Health SMART) and ARCHITECTURE (DiskHealthService, TemperatureService) are in this pull request.