landlock: fix disable() of FS protections; reject disable(FsRefer)

[congwang-mk]

Follow-up to the protection opt-out work merged in #71, addressing the blocking review comment about disable() of FS protections (#71 (comment)).

1. Blocking bug: disable() of an FS protection fails with EINVAL when the sandbox has a writable path

compute_fs_mask removes the disabled bit (REFER / TRUNCATE / IOCTL_DEV) from handled_access_fs, but the per-path write mask was derived from write_access(abi) alone. So with disable(FsTruncate) (or FsIoctlDev, or the now-rejected FsRefer) plus any writable path, the writable-path rule still requested the dropped bit, it was no longer a subset of the ruleset's handled accesses, and landlock_add_rule(2) rejected it with EINVAL. That broke nearly every real sandbox (any with a writable path). The net path was already gated this way via net_tcp_active; the FS path did not get the matching treatment.

Fix (landlock.rs): intersect the per-path write mask with the resolved handled set, so every installed rule is a subset by construction.

let fs_write_mask = write_access(abi) & handled_access_fs;

This covers both the directory rule (access) and the file rule (access & ACCESS_FILE).

Regression test: a forked confine_filesystem against the real host kernel with disable(FsTruncate) / disable(FsIoctlDev) plus fs_write("/tmp"), asserting clean confinement. It fails (EINVAL) before the fix and passes after. The existing mask_contract_tests could not catch this class because they exercise compute_fs_mask in isolation and never install a path rule.

2. disable(FsRefer) is rejected at build

Per landlock(7), REFER is denied by default by every ruleset even when it is not handled. Controlled cross-directory rename within writable areas works precisely because REFER is handled and granted on writable paths. Disabling REFER un-handles it, which can only make the sandbox stricter, never looser, so it cannot do what disable() promises and is a footgun.

build() and build_unchecked() now return SandboxError::Invalid for disable(Protection::FsRefer), surfaced uniformly through the Rust, C ABI, and Python bindings. allow_degraded(FsRefer) stays meaningful and is unchanged. The disable() rustdoc and the --disable CLI help carry the caveat.

Validation

Full workspace green on a v7 host: 334 core lib + 242 core integration + 70 FFI/CLI, plus 12 Python protection tests. New tests: disable_fs_protection_with_writable_path_does_not_einval and disable_fsrefer_is_rejected_at_build.

[congwang-mk]

@dzerik Please take a look