chore: clear golangci-lint findings and handle the errors they flagged - #5306

The walk used a loop that always ended in its first pass to get
"break to the fallback" control flow. A function returning whether it
resolved the application says the same without the loop. A test pins
that each failure to resolve still falls back to the HelmRelease's own
label and that a resolved one ends the walk.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Drop a field-manager constant nothing uses in the MariaDB driver, switch
on the MongoDB operator's state as a tagged switch, and reword the error
strings of the MongoDB driver and the workload monitor so they do not
start with a capital. The errors keep their meaning and nothing matches
their text.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

For the commands this change touches (cozypkg list, the validation
report of validate, push and tap, tap's summary, and check-readiness),
a write error the process observes is a failure: the command exits
non-zero and says whether the action it reports on already happened.
Progress lines printed before the decisive step stay best-effort. A
closed pipe ends the process with SIGPIPE, as for other Unix tools.

`cozypkg list` and the validation report dropped write errors, and the
list deferred its tabwriter flush, so a closed or full stdout still
reported success with nothing printed. The list now writes to the
command's writer and returns the flush error, and the report returns
its write error. When validation refuses a push or a tap, or fails a
validate run, the refusal stays the reported reason and the report's
write error is joined to it. A push whose report cannot be written
stops before flux runs and says "nothing was pushed"; a tap whose
summary cannot be written fails with "tap applied, but writing its
summary failed".

check-readiness fails --help, --version and its readiness result the
same way, in one-shot, --wait and --watch mode, with exit 1 as for a
result that is not ready, which its help now lists among the exit
statuses along with SIGPIPE for a closed pipe; the watch loop stops at
the first frame it cannot write, and a timed-out wait still ends with
its verdict.

push writes its "Pushing" line before it runs flux and stops with
"nothing was pushed" when that line cannot be written, with validation
skipped or after the report.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Every report, check and verify result goes to stdout, and every write
error there was dropped, so a run whose output never reached its
reader still exited 0. The terminal writer now keeps the first stdout
error, and the run fails with it unless it already failed. Only stdout
is tracked; a write error on stderr is still discarded.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Shutdown closed every open client's channel, and the client's handler
closed the same channel again when its request ended, so any stream
still open at shutdown panicked with a close of a closed channel.
Stopping a client is now idempotent.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Write errors on the event stream were ignored, so a client whose
connection broke kept its handler, pump and registration until the
request context ended, and every event was written into the dead
connection meanwhile. A failed write now ends the stream and
unregisters the client, and the handler also returns once its pump
ends, so Shutdown no longer leaves streams open.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

The status stream was started in a bare goroutine whose error was
dropped, and the runnable's registration error was discarded too, so a
port that could not be bound left the plunger running with no status
endpoint and nothing in its logs. The server now runs inside its
runnable: a listen error fails the manager, as a metrics server bind
failure does, and stopping the manager is not reported as an error.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

The drain tests now assert that RunThenDrain returns the run's nil
error, the idle-tail test fails on any error other than the stream
ending or being cut, and a test fixture file fails the test when it
cannot be closed.

Removing a temporary directory and closing a response body that has
been read are best-effort: a failure leaves nothing for the caller to
act on and does not change any result, so those errors are discarded
explicitly. The same holds for the stub servers' response writes, whose
failure surfaces on the client side of the test.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

The Prometheus response body is closed after it has been read, and the
test stubs' writes fail on the client side of the test, so neither
error changes a result.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

A temporary file or directory that cannot be removed, or closed after
its write already failed, leaves nothing to act on: the run's result is
the same, and the primary error is the one returned. The test that
unsets an environment variable now fails if it cannot.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Direct access to FieldsV1.Raw is deprecated; GetRawBytes returns the
same bytes and is nil-safe, and NewFieldsV1 builds the test fixtures.

The write-options test still covers a raw-only orphanDependents delete
option, now decoded from its wire form, and asserts that the converted
options keep it.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

The pointer package is deprecated; its Bool and BoolPtr are aliases of
ptr.To, so the values written are unchanged.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Go names sentinel error variables Err...; the package is internal and
the webhook is their only user.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

When a client disconnected, the handler stopped its pump and returned at
once, while the pump could still be writing an event or a heartbeat to
the response. net/http does not allow a write to a response after
ServeHTTP has returned. The handler now waits for the pump to exit; a
pump blocked in a write returns once the connection is gone.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

backport-batch writes its body, summary, findings and warnings before
it pushes, opens or edits a pull request, closes a batch or posts a
comment. A run that kept going after that output was lost acted on
what nobody read, and a closed pipe was worse: with Go's default
SIGPIPE handling the process died silently mid-run. Lines that confirm
an action after it happened, such as the URL of an opened pull request,
go to stderr, so a lost stderr matters as much as a lost stdout.

A failed write to either stream is kept, and from then on every push and
every GitHub write is refused at the point it would be made: the GitHub
client is wrapped so that each write checks first, and each push checks
before it runs. Once output is lost, the run changes nothing more: no
push, no GitHub write, no local branch move. No new unit starts after
output is lost; a unit already in flight finishes its rollback. A reset
whose own hint, the only way back to the commits it drops, cannot be
written does not run either. The run exits 2 saying the output was lost;
a confirmation lost after an action fails the run too. Such a
confirmation is lost only when stderr itself fails, so a lost
confirmation can only be reported by the exit code. A test holds every
method of the GitHub interface to being a known read or a guarded write.

The tool catches SIGPIPE with signal.Notify, so a closed pipe is a
write error like any other. It does not use signal.Ignore, whose ignored
disposition the git and gh children would inherit across exec; a test
checks that a child still dies of SIGPIPE.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

go-unit-tests ran only pkg/registry, pkg/config and pkg/cmd/server, so
the tests in pkg/lineage, pkg/ovnstatus, pkg/reservation, cmd/cozypkg
and cmd/check-readiness ran on no CI job. None of them needs a cluster
or the network.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>

Use tagged switches where staticcheck suggests them and drop two test
assignments whose values are never read. Behaviour is unchanged.

Assisted-by: LLM
Signed-off-by: Aleksei Sviridkin <[email protected]>