require user presence on security key signatures

[plotnick]

A security key signature carries a short trailer of a flags byte and a signature counter. The authenticator folds both into the signed data, so a verifier can trust them. Bit 0 of the flags records user presence, the authenticator's assertion that somebody touched the key.

#8 verified sk signatures without looking at the flags. A signature produced without a touch, whether by a key enrolled with no-touch-required or by a purely software authenticator, was therefore accepted. Refuse those by default, as sshd does. Verifiers that deliberately accept no-touch enrollments can relax the check per verification with require_user_presence(false).

The new tests emulate an authenticator in software, which is the only way to produce a signature without user presence set. They check that a touched signature verifies end to end, and that an untouched one is refused by default but accepted under the opt-out. That acceptance shows the refusal comes from the flags rather than from a bad signature. An ignored test runs the same verification against a real key in the SSH agent:

    SSHAUTH_TEST_SK_KEY=<comment or SHA256 fingerprint> \
        cargo test -- --ignored sk_hardware

The signature counter remains unused. Authenticators without a counter legitimately report zero, so the only meaningful check would be per-key monotonicity, which needs state the verifier does not have.

🤖 Disclaimer: LLM assistance was used to find this issue and prepare this patch, but human 👀 and 🧠 reviewed it.