fix(libstore): improve legacy ssh addToStore diagnostics

[znaniye]

Motivation

nix copy --to ssh://... can fail with a misleading low-level deserialisation error when the remote side returns a plain-text rejection instead of the expected binary status reply.

In this situation, users see a transport/protocol-looking failure even though the real issue is a policy rejection on the remote side.

Context

In a real-world flow, copying an unsigned store path to a remote machine as a non-trusted user produced:

error (ignored): error: interrupted by the user
error: serialised integer 7142773272180060773 is too large for type 'j'

This is hard to act on because it hides the underlying remote reason.

The actual remote-side cause was a trust/signature policy failure (path lacks a signature by a trusted key), but the client attempted to parse that text as a binary integer.

User-Visible Result

With this change, the same class of failure now reports a directly actionable error, for example:

error (ignored): failed to add path '/nix/store/<hash>-<name>' to remote host '<remote-host>': remote returned text where a binary reply was expected: 'error: cannot add path '/nix/store/<hash>-<name>' because it lacks a signature by a trusted key'
error: failed to add path '/nix/store/<hash>-<name>' to remote host '<remote-host>': remote returned text where a binary reply was expected: 'error: cannot add path '/nix/store/<hash>-<name>' because it lacks a signature by a trusted key'

This keeps protocol handling behavior intact while making failures diagnosable at the right abstraction level.

[znaniye]

legacy-ssh-store-no-write.sh is an intentional test fixture.

It strips --write from nix-store --serve --write ..., forcing the remote to return a real textual rejection (importing paths is not allowed) instead of a binary reply. This deterministically exercises the legacy-ssh-store.cc fix for “remote returned text where a binary reply was expected” and prevents regression to low-level serialization errors. Using plain remote-program=nix-store does not trigger this path, and a broken/missing program fails for unrelated reasons.

If you guys know a better way to trigger this binary-vs-text reply path in tests, I’m happy to switch. Suggestions welcome, and thanks in advance.