Both genesis downloads reached for a package-level client, so a test had
no way to point them at a mock server. Take an api.HttpPkgInterface like
the release checks, the catchpoint lookup and the short links already do.
GetGenesis has no callers; handleDataDirMac gets the client from
algod.Install, which the install and bootstrap commands now supply.

Drop main_test.go with it. A test binary is not linked as package main,
so the linker never stamps main.version there and both sides of the
comparison were the same "dev" default: the test passed whether or not
init handed the version over. api/useragent_test.go covers SetVersion.

Every outbound request went out as Go's default "Go-http-client/1.1",
which tells a node operator reading their algod access log, or the
GitHub release API, nothing about what is calling them. Send
"Nodekit v<version>" instead, built from the version the linker stamps
into main so that releases can be told apart.

The shared HttpPkg wrapper covers the release checks, the catchpoint
lookup, the short links and the upgrade download; the generated algod
client gets the header through a request editor, since api/lf.go is
overwritten by `make generate`.

--follow substituted a backlog of 10 when --lines was not given, so -n
meant one thing alone and another beside -f: `-n 0 -f` replayed the
whole history, where `tail -n 0 -f` shows none of it. --follow now
replays the same set the command shows without it, --lines N shortens
that backlog, and a stream that starts at now is --since 0s.

The cost is that a bare -f reads the whole history before the first
new entry arrives, as the command without -f already does.