[AdaAibaby]

Problem

When a Firecracker VM exits, in-flight NBD commands and the reply socket fail in predictable, non-fault ways — yet all of them were logged at ERROR, so every normal teardown with in-flight I/O produced a burst of ERROR lines that drown out the real backend failures ERROR is meant to surface (#3274):

• cmdRead / cmdWrite / cmdWriteZeroes select on the dispatch context, so an in-flight backend call observes context.Canceled (or DeadlineExceeded) the instant the sandbox is torn down.
• completeAsync's reply-socket write fails with use of closed network connection (net.ErrClosed) or broken pipe (EPIPE) once the client side is gone.

Fix — dispatch.go

• Add isExpectedTeardownError(err) — true for context.Canceled, context.DeadlineExceeded, net.ErrClosed, syscall.EPIPE.
• Add logCmdFailure(ctx, msg, cmd, err) and route all four failure logs through it: expected teardown errors go to Warn, everything else stays at Error.

Genuine backend errors (corrupt block, ECONNRESET, unexpected EOF) still surface at Error.

Before (every teardown with in-flight I/O):

level=error msg="nbd backend read failed"  error="context canceled"
level=error msg="nbd backend write failed" error="context canceled"
level=error msg="nbd command failed while the dispatcher was already failing" error="write unix @->@: use of closed network connection"

After: the same lines at level=warn; real backend faults unchanged at level=error.

Tests — dispatch_teardown_test.go

Table-driven test over the classifier (the single predicate both log levels hang off): wrapped context.Canceled, net.ErrClosed / EPIPE inside *net.OpError (how a closed-socket write actually surfaces), plus the negatives that must stay Error (ECONNRESET, unexpected EOF, a plain error, nil). go build ./... and go vet clean; full nbd package green.

Scope

Revives only the NBD-dispatch half of the closed #3277. Its other half — the FC metrics flusher select-race guard (#3275) — is already fixed on main by 50c64bee8 ("stop flushing Firecracker metrics after the VM exits"), so it is intentionally dropped. #3275 is CLOSED COMPLETED; #3274 is still open.

Adapted to main's refactor since the original PR (single completeAsync error log, command.fields(), injected d.logger). The fix is also broader-but-precise vs the original: the original only special-cased context.Canceled on the backend path and blanket-downgraded the fatal path; here one classifier covers both backend-cancel and closed-socket errors and leaves real failures at Error.

Closes #3274. Supersedes #3277 (closed for inactivity / merge drift; maintainer invited a fresh PR against current main).