[A13xB0]

For #447. I checked the updater on Linux against the real feed, using the published v0.2.0 compact tarball and its PyPI client. Each run had its own config and cache directories.

Check	Result
0.2.0 notices 0.3.0 without being told where	yes: latest 0.3.0, available, the compact tarball chosen to match its own variant
Nothing new	0.3.0 says "this is the newest release: 0.3.0"
Network down	error names the failure and available is false; the status line did not say it (fixed below)
Download verified	the staged file's SHA-256 matches the published SHA256SUMS; corrupt and truncated downloads are covered by download_test.go
Version afterwards	the unpacked build reports v0.3.0
Old client against the new workbench	refused, naming both versions and the fix; the Setup row says so before the update
Off stays off	update_checks: false survives a restart, and no check runs
Swap while running	unpacking over a running 0.2.0 leaves it running; the next start is 0.3.0

Fixed

• Tarball detection. update.Detect recognised a Linux tarball install by a radioserver beside the binary, which the tarball no longer ships. Every tarball install reported unpackaged and got the generic instruction. It now looks for VARIANT, and the real 0.2.0 tree is detected as tarball.
• Headless echo. Headless echoed the latest Status every 100 ms. It dropped every line said between two looks, and the last ones before exit, so a failed check left nothing in the log. It now prints each new log line once, and flushes before exiting.

Not done: macOS and Windows. Each needs a real install of the previous release on that platform.