capsule_runtime/

sealed.rs

//! The `sealed` containment mechanism: a private mount namespace pivoted onto a composed root.
//!
//! `scoped` (Landlock + seccomp, see [`crate::sandbox`]) can deliver "nothing outside the workdir
//! is readable or writable". It can never deliver "nothing outside the workdir *exists*", because
//! Landlock is an access filter — it does not change what a process can see, name or traverse. A
//! mount namespace can, and that is the whole of what this module adds.
//!
//! ## What the mechanism is
//!
//! Inside the forked child's `pre_exec` window — before fd hygiene, before the capability drop,
//! before the seccomp filter, and before the Landlock ruleset — the child:
//!
//!   1. `unshare(CLONE_NEWUSER | CLONE_NEWNS)`, then maps its own uid/gid one-to-one into the new
//!      user namespace (`/proc/self/uid_map`), which is what gives it `CAP_SYS_ADMIN` *inside that
//!      namespace* and nowhere else. Nothing here needs the host's root.
//!
//!      On every kernel tier this now runs *after* [`crate::network_namespace`] has already
//!      created a user+network namespace, so the one built here is **nested** inside it. That is
//!      load-bearing rather than incidental: a task in a descendant user namespace holds no
//!      capability in the ancestor, so a `sealed` capsule cannot reconfigure the network namespace
//!      it is confined to.
//!   2. Makes mount propagation private (`MS_REC | MS_PRIVATE` on `/`), so nothing it does leaks
//!      back into the host's mount table.
//!   3. Mounts a fresh `tmpfs` over a base directory ([`choose_root_base`]) and composes the new
//!      root inside it: the host runtime directories ([`SEALED_RUNTIME_PATHS`]) and a fixed,
//!      narrow `/etc` set ([`SEALED_ETC_PATHS`]) bind-mounted **read-only**, the session workdir
//!      bind-mounted **read-write at its own absolute path**, a private `/dev` tmpfs carrying the
//!      OCI default device set ([`SEALED_DEVICE_NODES`]), and a `/proc` — masked with `hidepid`
//!      where the kernel permits it, bound from the host where it does not
//!      ([`PROC_HIDEPID_OPTIONS`] documents which is which, and why).
//!   4. `pivot_root`s onto it, detaches the old root, remounts the new root read-only, and
//!      `chdir`s back into the workdir.
//!
//! Everything outside that root is then not merely access-denied — it is *absent*. There is no
//! pathname for `/etc/shadow`, for a Docker socket, or for `/dev/sda`, so the escape classes that
//! `scoped` has to enumerate and refuse (pathname unix sockets, device nodes, metadata visibility,
//! stdlib directory enumeration) stop being reachable rather than being individually denied.
//!
//! `/proc` is the one exception to that sentence: mounting a private `procfs` unprivileged needs a
//! PID namespace this module deliberately does not create, so on a bare host the composed root
//! carries a bind of the host's `/proc` and
//! host process metadata stays *visible* there (opens through it are still Landlock's to refuse).
//! See [`PROC_HIDEPID_OPTIONS`].
//!
//! ## What it does *not* replace
//!
//! Landlock and seccomp still install afterwards, now scoped to the private root. `sealed` is
//! defence in depth over `scoped`'s mechanism, not a substitute for it: the syscall allowlist is
//! still the only thing standing between a capsule and `io_uring`, `bpf` or `ptrace`, none of
//! which a mount namespace has an opinion about.
//!
//! ## Why the plan is precomputed
//!
//! Everything the child does is driven by [`SealedRootSpec`], a fully-resolved list of
//! NUL-terminated paths built in the **parent**, before `fork()`. The `pre_exec` window permits
//! only async-signal-safe work: allocating there (a `format!`, a `PathBuf::join`) can deadlock on
//! the allocator lock a *different* thread of the parent held at fork time. Splitting the decision
//! ([`plan_composed_root`], pure and unit-testable against a fake host layout) from the execution
//! ([`construct_composed_root`], raw syscalls over already-built `CString`s) is what keeps the
//! child side allocation-free on its success path — and it is what makes the composed-root layout
//! testable at all on a host that cannot create a namespace.

// The composed-root planner and the synthetic `/etc` content are pure so they stay unit-tested on
// every OS, but outside tests only the Linux `linux` submodule and `sandbox`'s Linux enforcement
// call them, so off Linux they are dead by construction.
#![cfg_attr(not(target_os = "linux"), allow(dead_code))]

use std::path::{Path, PathBuf};

// ---------------------------------------------------------------- AppArmor

/// Name of the AppArmor profile this runtime ships, as it appears in
/// `/sys/kernel/security/apparmor/profiles` and in `/proc/self/attr/current`.
///
/// On a host where AppArmor's `restrict_unprivileged_userns` knob is on (Ubuntu 23.10+ and
/// derivatives), an unconfined process's `unshare(CLONE_NEWUSER)` is transitioned into the
/// restricted `unprivileged_userns` profile, which denies `CAP_SYS_ADMIN` and therefore every
/// `mount(2)` that follows. Loading this profile — which attaches to the `mur` binary and carries
/// `userns,` — is what lifts that, exactly as Ubuntu's own shipped profiles do for Chrome and
/// Firefox.
pub const SEALED_APPARMOR_PROFILE_NAME: &str = "mur-sealed";

/// Where [`scripts/install.sh`](../../../scripts/install.sh) installs the shipped profile, and the
/// path the refusal message tells an operator to load.
pub const SEALED_APPARMOR_PROFILE_PATH: &str = "/etc/apparmor.d/mur-sealed";

/// SHA-256 (lowercase hex) of `packaging/apparmor/mur-sealed` as this build ships it.
///
/// A literal rather than an `include_str!` digest: `capsule-runtime` is published to crates.io,
/// `cargo package` copies only files under the crate's own directory, and a path escaping the
/// crate root compiles in the workspace and then fails the packaging verification build. The unit
/// test `shipped_profile_digest_constant_matches_the_file` reads the real file through
/// `CARGO_MANIFEST_DIR` and fails with the digest to paste in here — that test does not run during
/// `cargo package`, so publishing stays green while a profile edit that forgot this constant is
/// caught by `cargo test`.
///
/// Compares *file bytes*. It says nothing about what `apparmor_parser` has actually loaded — see
/// [`classify_installed_profile`].
pub const SEALED_APPARMOR_PROFILE_SHA256: &str =
    "1669f6c0038dddea393cfacd95b078ac99ec718a538c8ebb52701f6ba686a892";

/// Where this host's permission to create an unprivileged user namespace comes from.
///
/// Three of the four variants permit the namespace, and they are not interchangeable: AppArmor
/// being absent is a distribution fact nobody chose, the shipped profile confining `mur` is the
/// configuration murmur ships, and the restriction being switched off host-wide removes the
/// hardening for *every* binary on the machine. The three reach the same achieved class, so
/// `mur doctor`, `--explain-scope` and the session trace each report the variant: without it a
/// `sealed` result obtained on a weakened host cannot be told apart from one obtained through the
/// profile.
///
/// The grant never decides anything: whether a namespace, a private `/tmp` or `sealed` is
/// available is decided by [`crate::namespace_probe`] creating them. The grant only words a
/// refusal once an attempt has failed.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Default, serde::Serialize)]
#[serde(rename_all = "snake_case")]
pub enum UsernsGrant {
    /// AppArmor is not enabled on this host (Fedora, Arch, a kernel built without it, …). There is
    /// no restriction to lift and no profile to install.
    ApparmorAbsent,
    /// AppArmor is enabled but `kernel.apparmor_restrict_unprivileged_userns` is `0`, so no binary
    /// on this host is transitioned into the restricted `unprivileged_userns` profile. Legitimate,
    /// and on some hosts the only option — but it is not the configuration murmur ships, and it is
    /// what `W-SEC-013` reports.
    RestrictionDisabledHostWide,
    /// The restriction is on and this process is confined by a profile whose name begins with
    /// [`SEALED_APPARMOR_PROFILE_NAME`] — the shipped profile, or the checkout profile
    /// `scripts/install-dev-apparmor.sh` generates. The grant is narrow: it applies to this binary
    /// and nothing else.
    ProfileConfining,
    /// The restriction is on and nothing grants this binary anything. A container whose own
    /// profile permits user namespaces reads this too, and is not refused for it. The default, so
    /// an unprobed host claims nothing.
    #[default]
    Withheld,
}

impl UsernsGrant {
    /// Every variant, so a caller reasoning about "which grant did this host give" cannot silently
    /// miss one. Same convention as [`SealedBlocker::ALL`].
    pub const ALL: &'static [UsernsGrant] = &[
        UsernsGrant::ApparmorAbsent,
        UsernsGrant::RestrictionDisabledHostWide,
        UsernsGrant::ProfileConfining,
        UsernsGrant::Withheld,
    ];

    /// Stable wire name, as it appears in `--explain-scope --json`, in `session_start` and in
    /// `mur doctor`. Distinct per variant, and not a `Debug` rendering, which is free to change.
    #[must_use]
    pub fn wire_name(self) -> &'static str {
        match self {
            UsernsGrant::ApparmorAbsent => "apparmor_absent",
            UsernsGrant::RestrictionDisabledHostWide => "restriction_disabled_host_wide",
            UsernsGrant::ProfileConfining => "profile_confining",
            UsernsGrant::Withheld => "withheld",
        }
    }

    /// One line naming the mechanism in words, for the `mur doctor` block. States what the grant
    /// covers — one binary or the whole host — because that is the difference the wire name alone
    /// does not spell out.
    #[must_use]
    pub fn summary(self) -> &'static str {
        match self {
            UsernsGrant::ApparmorAbsent => {
                "AppArmor is not enabled on this host, so nothing restricts unprivileged user \
                 namespaces and no profile is needed"
            }
            UsernsGrant::RestrictionDisabledHostWide => {
                "kernel.apparmor_restrict_unprivileged_userns is off, so unprivileged user \
                 namespaces are unrestricted for every binary on this host, not just for mur"
            }
            UsernsGrant::ProfileConfining => {
                "the restriction is on and the mur-sealed AppArmor profile is confining this \
                 binary, so the grant covers mur alone — the configuration murmur ships"
            }
            UsernsGrant::Withheld => {
                "the restriction is on and no mur-sealed AppArmor profile is confining this \
                 binary, so nothing murmur ships grants it a user namespace"
            }
        }
    }

    /// The whole decision, over the three host readings that produce it: whether AppArmor is
    /// enabled, whether its `restrict_unprivileged_userns` knob is on, and what profile confines
    /// this process. Each argument is the file's trimmed contents, or `None` when it is missing or
    /// empty.
    ///
    /// Pure, and separate from the reading for the same reason [`classify_installed_profile`] is:
    /// every combination is then testable on any OS, with no `/sys` and no `/proc`.
    /// `linux::userns_grant` is the reader and holds no logic of its own.
    ///
    /// A missing or unreadable restriction knob counts as *not restricted*, not as withheld — a
    /// kernel that does not implement the knob does not restrict anything.
    #[must_use]
    pub fn from_readings(
        enabled: Option<&str>,
        restricted: Option<&str>,
        current_profile: Option<&str>,
    ) -> UsernsGrant {
        if !matches!(enabled, Some("Y") | Some("1")) {
            return UsernsGrant::ApparmorAbsent;
        }
        if !matches!(restricted, Some("Y") | Some("1")) {
            return UsernsGrant::RestrictionDisabledHostWide;
        }
        match current_profile {
            Some(profile) if profile.starts_with(SEALED_APPARMOR_PROFILE_NAME) => {
                UsernsGrant::ProfileConfining
            }
            _ => UsernsGrant::Withheld,
        }
    }
}

/// How `/etc/apparmor.d/mur-sealed` on this host compares to the profile this build ships.
///
/// A *file-level* finding and deliberately nothing more. AppArmor loads profiles from the kernel's
/// own policy cache, not from this path at call time, so a file can be edited without
/// `apparmor_parser -r` ever running and a profile can be loaded from a file that has since been
/// deleted. [`UsernsGrant`] is the behavioural answer and stays the source of truth; this only
/// tells an operator whether the bytes on disk are the bytes murmur ships.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum InstalledProfileState {
    /// The installed file hashes to [`SEALED_APPARMOR_PROFILE_SHA256`].
    Matches,
    /// The file is present and hashes to something else — an older revision, or an edited copy.
    /// Operator customisation belongs in `/etc/apparmor.d/local/mur-sealed`, which both shipped
    /// profiles `include if exists` and which is not hashed here.
    Drifted {
        /// Lowercase hex SHA-256 of the bytes actually on disk.
        installed_sha256: String,
    },
    /// No file at [`SEALED_APPARMOR_PROFILE_PATH`]. Expected on a host without AppArmor, and on a
    /// checkout build using `scripts/install-dev-apparmor.sh`, which writes its own separate file.
    Absent,
    /// The path exists and could not be read — a permission or I/O error, kept apart from
    /// [`Self::Absent`] so "I could not look" never reads as "it is not there".
    Unreadable {
        /// The `std::io::Error`, rendered.
        error: String,
    },
}

/// Classifies a read of [`SEALED_APPARMOR_PROFILE_PATH`] against the shipped digest.
///
/// Takes the already-performed read rather than the path, so all four outcomes are unit-testable
/// with no filesystem mutation and no privilege.
#[must_use]
pub fn classify_installed_profile(read: Result<Vec<u8>, std::io::Error>) -> InstalledProfileState {
    match read {
        Ok(bytes) => {
            let installed_sha256 = murmur_artifact::sha256_hex(&bytes);
            if installed_sha256 == SEALED_APPARMOR_PROFILE_SHA256 {
                InstalledProfileState::Matches
            } else {
                InstalledProfileState::Drifted { installed_sha256 }
            }
        }
        Err(error) if error.kind() == std::io::ErrorKind::NotFound => InstalledProfileState::Absent,
        Err(error) => InstalledProfileState::Unreadable {
            error: error.to_string(),
        },
    }
}

/// [`classify_installed_profile`] over the real [`SEALED_APPARMOR_PROFILE_PATH`]. Reads one file
/// and needs no privilege.
#[must_use]
pub fn inspect_installed_profile() -> InstalledProfileState {
    classify_installed_profile(std::fs::read(SEALED_APPARMOR_PROFILE_PATH))
}

/// The paths the two shipped profiles in `packaging/apparmor/mur-sealed` attach to, as their
/// AppArmor path globs.
///
/// **Remediation text only.** Nothing decides anything by matching the running binary's path
/// against this list: matching a path is inference, while reading `/proc/self/attr/current` is
/// what the kernel actually did, and [`classify_profile_attachment`] uses only the latter. This
/// exists so `mur doctor` can tell an operator running a binary from an unusual location where the
/// shipped profile *does* attach, without pretending to have measured anything.
pub const SEALED_APPARMOR_ATTACHMENT_PATHS: &[&str] = &[
    "/{usr/local/,usr/,opt/mur/}bin/mur",
    "@{HOME}/{.local,.cargo}/bin/mur",
];

/// Which AppArmor profile, if any, the kernel reports as confining *this* process.
///
/// AppArmor attaches profiles by executable path, so a `mur` at a bind mount, a build output or a
/// test fixture gets no profile however correctly the shipped one is installed. That is invisible
/// in [`UsernsGrant`] alone — it reports [`UsernsGrant::Withheld`] identically for "the profile is
/// not installed" and "the profile is installed and does not attach here" — which is why this is
/// reported beside it rather than folded into it.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum ProfileAttachment {
    /// A profile whose name begins with [`SEALED_APPARMOR_PROFILE_NAME`] confines this process —
    /// the shipped profile, or the checkout profile `scripts/install-dev-apparmor.sh` generates.
    /// The same condition [`UsernsGrant::ProfileConfining`] is decided on.
    Sealed {
        /// The profile name as the kernel reports it, without the trailing enforcement mode.
        profile: String,
    },
    /// Some other profile confines this process. It grants nothing murmur asked for, and it may
    /// be *narrower* than unconfined — a snap or a container runtime profile, for instance.
    Other {
        /// The profile name as the kernel reports it, without the trailing enforcement mode.
        profile: String,
    },
    /// The kernel reports `unconfined`: AppArmor is present and no profile attaches to this
    /// binary's path.
    Unconfined,
    /// `/proc/self/attr/current` is absent or unreadable — no AppArmor on this host, or not
    /// Linux. Never collapsed into [`Self::Unconfined`], on the same terms
    /// [`InstalledProfileState::Unreadable`] is kept apart from
    /// [`InstalledProfileState::Absent`]: "nothing attaches" and "I could not look" are different
    /// findings.
    NotReported,
}

/// Classifies the contents of `/proc/self/attr/current` into a [`ProfileAttachment`].
///
/// Takes the already-performed read rather than the path, so all four outcomes are unit-testable
/// on any OS — the same split [`classify_installed_profile`] uses.
///
/// The kernel writes the profile name followed by its enforcement mode in parentheses
/// (`mur-sealed (unconfined)`, `unconfined`), so the mode is stripped before the name is compared.
#[must_use]
pub fn classify_profile_attachment(current_profile: Option<&str>) -> ProfileAttachment {
    let Some(raw) = current_profile
        .map(str::trim)
        .filter(|value| !value.is_empty())
    else {
        return ProfileAttachment::NotReported;
    };
    let name = raw
        .split_once(" (")
        .map_or(raw, |(name, _mode)| name)
        .trim();
    if name.is_empty() || name == "unconfined" {
        return ProfileAttachment::Unconfined;
    }
    if name.starts_with(SEALED_APPARMOR_PROFILE_NAME) {
        return ProfileAttachment::Sealed {
            profile: name.to_string(),
        };
    }
    ProfileAttachment::Other {
        profile: name.to_string(),
    }
}

/// [`classify_profile_attachment`] over the real `/proc/self/attr/current`. Holds no logic: the
/// file is missing on every non-Linux host and on a Linux host without AppArmor, and both read as
/// [`ProfileAttachment::NotReported`].
#[must_use]
pub fn inspect_profile_attachment() -> ProfileAttachment {
    classify_profile_attachment(
        std::fs::read_to_string("/proc/self/attr/current")
            .ok()
            .as_deref(),
    )
}

// ---------------------------------------------------------------- blockers

/// The one mechanism that stands between this host and `sealed`, named specifically enough that
/// the refusal carries a command the operator can run.
///
/// Deliberately a small closed enum rather than a free-form string: `containment.rs` renders it,
/// `sandbox.rs` derives it from the probe, and the manual-verification document asserts the exact
/// text of each rendering.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, serde::Serialize)]
#[serde(rename_all = "snake_case")]
pub enum SealedBlocker {
    /// Not Linux. Mount namespaces do not exist here and never will — the same permanence
    /// `EnforcementTier::EnvironmentOnly` carries.
    NotLinux,
    /// AppArmor's unprivileged-user-namespace restriction is active and this binary is not
    /// confined by the shipped [`SEALED_APPARMOR_PROFILE_NAME`] profile.
    AppArmorProfileMissing,
    /// `unshare(CLONE_NEWUSER | CLONE_NEWNS)` was refused outright — the container case: no
    /// `CAP_SYS_ADMIN`, or the container's own seccomp filter blocking the syscall.
    NamespaceCreationDenied,
    /// The namespace was created, but writing the identity `uid_map`/`gid_map` was refused, so the
    /// process cannot own it. Nothing to do with `CAP_SYS_ADMIN` or containers.
    IdMapDenied,
    /// The namespace was created, but `mount(2)` inside it was refused. The signature of a
    /// confinement that permits `userns_create` and then denies `CAP_SYS_ADMIN`.
    MountDenied,
    /// The namespace was created and `mount(2)` inside it worked, but the `pivot_root(2)`
    /// sequence the composed root's step 6 performs did not. A separate finding from
    /// [`Self::MountDenied`] because the two have separate remediations: `mount` is what an
    /// AppArmor profile grants, while `pivot_root` is a distinct rule in that profile and a
    /// distinct entry in a container runtime's seccomp allowlist, so a host can grant one and
    /// refuse the other.
    PivotRootDenied,
    /// The kernel has no unprivileged user namespace support at all (`CONFIG_USER_NS=n`, or
    /// `user.max_user_namespaces=0`).
    KernelUnsupported,
    /// Namespaces are available but Landlock is not, so the defence-in-depth layer `sealed`
    /// is required to keep would be missing. `sealed` is strictly stronger than `scoped`; a host
    /// that cannot back `scoped` cannot back `sealed` either.
    LandlockUnavailable,
    /// The composed root is built inside the capsule's network namespace, and this host created
    /// the user namespace but could not finish the network namespace or enter it.
    NetworkNamespaceUnavailable,
    /// The probe itself could not run — a child could not be forked, waited for, or reported
    /// nothing — so whether this host can back `sealed` is unknown.
    ProbeUnavailable,
}

impl SealedBlocker {
    /// Every variant, so a caller that has to reason about "which refusal did this host give"
    /// cannot silently miss one. Added because a hand-maintained list of the expected refusal
    /// texts in `murmur-cli`'s CLI test drifted the moment a variant was added, and the drift
    /// showed up as a failing assertion about the *host* rather than about the list.
    pub const ALL: &'static [SealedBlocker] = &[
        SealedBlocker::NotLinux,
        SealedBlocker::AppArmorProfileMissing,
        SealedBlocker::NamespaceCreationDenied,
        SealedBlocker::IdMapDenied,
        SealedBlocker::MountDenied,
        SealedBlocker::PivotRootDenied,
        SealedBlocker::KernelUnsupported,
        SealedBlocker::LandlockUnavailable,
        SealedBlocker::NetworkNamespaceUnavailable,
        SealedBlocker::ProbeUnavailable,
    ];

    /// One sentence naming the missing mechanism, plus the exact remediation. Rendered into
    /// `RuntimeError::ContainmentFloorUnmet`'s `reason`, so this is the text an operator sees
    /// under `E-CAP-003`.
    #[must_use]
    pub fn reason(self) -> String {
        match self {
            SealedBlocker::NotLinux => {
                "sealed requires a Linux mount namespace + pivot_root; this \
                 platform has no such primitive and never will (macOS and every non-Linux target \
                 stay at advisory permanently)"
                    .to_string()
            }
            SealedBlocker::AppArmorProfileMissing => format!(
                "sealed requires an unprivileged user+mount namespace, and AppArmor's \
                 unprivileged-userns restriction is active on this host while the '{name}' profile \
                 is not confining this binary. Install and load the profile shipped with mur: \
                 `sudo install -m 644 packaging/apparmor/{name} {path} && sudo apparmor_parser -r \
                 {path}` (or re-run the mur installer as root), then re-run. Building out of a \
                 checkout, where the binary sits at ./target/{{debug,release}}/mur and no shipped \
                 profile attaches to it: run `scripts/install-dev-apparmor.sh`, which generates \
                 and loads the same grant for those two paths. LAST RESORT, only where a profile \
                 genuinely cannot be loaded: `sudo sysctl -w \
                 kernel.apparmor_restrict_unprivileged_userns=0` — this removes \
                 unprivileged-userns hardening from every program on the machine, not just from \
                 mur, and is not the configuration murmur ships.",
                name = SEALED_APPARMOR_PROFILE_NAME,
                path = SEALED_APPARMOR_PROFILE_PATH,
            ),
            SealedBlocker::NamespaceCreationDenied => {
                "sealed requires an unprivileged user namespace, which this host refused to create. \
                 This is the usual answer inside a container: the container runtime's seccomp \
                 profile refuses unshare(CLONE_NEWUSER), and refuses setns, mount, umount2 and \
                 pivot_root, which sealed needs next. Run the container with a seccomp profile \
                 that permits those calls (Docker: `--security-opt seccomp=<file>`; Kubernetes: \
                 `seccompProfile: Localhost`), or create the namespaces outside the container and \
                 run mur inside them. Adding capabilities does not lift a seccomp refusal. The \
                 runtime will not fall back to a weaker class."
                    .to_string()
            }
            SealedBlocker::IdMapDenied => {
                "sealed created a user namespace, but writing the identity uid_map/gid_map into \
                 it was refused, so the process cannot own the namespace it just made. This \
                 is an id-mapping problem, not a missing capability: check that \
                 /proc/sys/user/max_user_namespaces is non-zero, that no LSM policy blocks \
                 uid_map writes for this binary, and that mur is not already running inside an \
                 unmapped user namespace. The runtime will not fall back to a weaker class."
                    .to_string()
            }
            SealedBlocker::MountDenied => format!(
                "sealed created a user+mount namespace but mount(2) inside it was refused, which \
                 is what a confinement that permits userns_create and then denies CAP_SYS_ADMIN \
                 looks like. On an AppArmor host, load the shipped profile: `sudo apparmor_parser \
                 -r {path}`. Inside a container, the container runtime's seccomp profile refuses \
                 mount, umount2 and pivot_root whatever capabilities the container holds: run it \
                 with a seccomp profile that permits them (Docker: `--security-opt \
                 seccomp=<file>`; Kubernetes: `seccompProfile: Localhost`), or establish the \
                 mount namespace outside the container.",
                path = SEALED_APPARMOR_PROFILE_PATH,
            ),
            SealedBlocker::PivotRootDenied => format!(
                "sealed created a user+mount namespace and mounted inside it, but pivot_root(2) \
                 was refused, so the composed root cannot become this capsule's root. mount(2) \
                 working while pivot_root(2) does not is what a policy that grants one and not \
                 the other looks like: on an AppArmor host, check that the loaded profile carries \
                 a `pivot_root,` rule and reload it with `sudo apparmor_parser -r {path}`; inside \
                 a container, the container runtime's seccomp profile refuses pivot_root whatever \
                 capabilities the container holds, so run it with a seccomp profile that permits \
                 pivot_root (Docker: `--security-opt seccomp=<file>`; Kubernetes: \
                 `seccompProfile: Localhost`), or establish the mount namespace outside the \
                 container instead. The runtime will not fall back to a weaker class.",
                path = SEALED_APPARMOR_PROFILE_PATH,
            ),
            SealedBlocker::KernelUnsupported => {
                "sealed requires unprivileged user namespaces, which this kernel does not provide \
                 (CONFIG_USER_NS=n, or user.max_user_namespaces=0). Raise \
                 `sudo sysctl -w user.max_user_namespaces=10000` if the sysctl is merely zeroed, \
                 otherwise run on a kernel built with CONFIG_USER_NS=y."
                    .to_string()
            }
            SealedBlocker::LandlockUnavailable => {
                "sealed keeps Landlock and seccomp inside the composed root as defence in depth, \
                 and this host provides no usable Landlock ABI (Linux 5.13+ required). A host that \
                 cannot back scoped cannot back sealed either."
                    .to_string()
            }
            SealedBlocker::NetworkNamespaceUnavailable => {
                "sealed is built inside the capsule's network namespace and this host could not \
                 create or enter it, so there is no namespace to build the composed root in. \
                 Every capsule that can spawn a subprocess needs that namespace: see E-CAP-005 \
                 for the step that failed and its remedy, then re-run. The runtime will not fall \
                 back to a weaker class."
                    .to_string()
            }
            SealedBlocker::ProbeUnavailable => {
                "sealed could not be probed: the probe could not run, so whether this host can \
                 back sealed is unknown; refused rather than assumed. A process or descriptor \
                 limit, or memory, ran out, or a policy killed the probe's child: re-run once \
                 the limit is relieved, or allow the probe's syscalls."
                    .to_string()
            }
        }
    }
}

/// Why this host cannot back `sealed`: the blocker, and, when a probe attempt failed, the call it
/// saw fail and the readings the refusal is worded with.
///
/// [`Self::reason`] is the text under `E-CAP-003` for `sealed`, so the floor refusal,
/// `--explain-scope` and `mur doctor` cannot explain one host two ways.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SealedShortfall {
    /// Decided by the attempts and Landlock alone; the readings below never change whether there
    /// is a shortfall.
    pub blocker: SealedBlocker,
    /// The probe's call that failed, in its step's words (`unshare(CLONE_NEWUSER | CLONE_NEWNET)`,
    /// `pivot_root(2)`, …), or, for [`SealedBlocker::ProbeUnavailable`], the runtime's account of
    /// why the probe could not run. `None` when no attempt failed: [`SealedBlocker::NotLinux`] and
    /// [`SealedBlocker::LandlockUnavailable`].
    pub failed_call: Option<String>,
    /// The `errno` the failed call left. `None` when no call failed, and for
    /// [`SealedBlocker::ProbeUnavailable`].
    pub errno: Option<i32>,
    /// AppArmor's unprivileged-userns restriction as read with the attempts.
    pub userns_grant: UsernsGrant,
    /// The profile the kernel reports as confining this process, read with the attempts.
    pub attachment: ProfileAttachment,
}

impl SealedShortfall {
    /// [`SealedBlocker::reason`]; then, when an attempt failed, `The probe's <call> failed with
    /// <ERRNO> (<words>).` (or `The probe could not run: <account>.`); then, last, the same `Host
    /// readings: …` line `E-CAP-005` ends with.
    #[must_use]
    pub fn reason(&self) -> String {
        let mut reason = self.blocker.reason();
        let Some(call) = &self.failed_call else {
            return reason;
        };
        let what = match self.errno {
            Some(errno) => crate::namespace_probe::probe_failure_sentence(call, errno),
            None => format!("The probe could not run: {call}."),
        };
        reason.push(' ');
        reason.push_str(&what);
        reason.push(' ');
        reason.push_str(&crate::network_namespace::host_readings(
            self.userns_grant,
            &self.attachment,
        ));
        reason
    }
}

/// Which single mechanism keeps this host below `sealed`, decided from the probe's attempts and
/// Landlock. Pure; `None` means the composed root was built and Landlock is fully enforced.
///
/// The readings decide nothing but which words an attempt that failed with `EPERM`/`EACCES` is
/// named in: a network namespace refused under the restriction with `mur` unconfined is
/// [`SealedBlocker::AppArmorProfileMissing`], the same refusal under a container's profile is the
/// container case. The `/tmp` attempt never affects `sealed`: the composed root binds `/tmp`
/// itself.
pub(crate) fn sealed_shortfall(
    is_linux: bool,
    landlock_fully_enforced: bool,
    reading: &crate::namespace_probe::NamespaceReading,
) -> Option<SealedShortfall> {
    use crate::namespace_probe::{MountAttempt, MountAttemptError, MountStep};
    use crate::network_namespace::{CreationStep, NamespaceAttemptError};

    let shortfall = |blocker, failed_call: Option<String>, errno| SealedShortfall {
        blocker,
        failed_call,
        errno,
        userns_grant: reading.userns_grant,
        attachment: reading.attachment.clone(),
    };
    if !is_linux {
        return Some(shortfall(SealedBlocker::NotLinux, None, None));
    }
    const KERNEL_ERRNOS: [i32; 4] = [libc::ENOSPC, libc::EINVAL, libc::ENOSYS, libc::EUSERS];
    match &reading.network {
        Ok(()) => {}
        Err(NamespaceAttemptError::Creator(detail)) => {
            return Some(shortfall(
                SealedBlocker::ProbeUnavailable,
                Some(detail.clone()),
                None,
            ))
        }
        Err(NamespaceAttemptError::Step { step, errno }) => {
            let errno = *errno;
            let blocker = match step {
                CreationStep::Unshare if KERNEL_ERRNOS.contains(&errno) => {
                    SealedBlocker::KernelUnsupported
                }
                CreationStep::Unshare | CreationStep::IdentityMaps
                    if matches!(errno, libc::EPERM | libc::EACCES)
                        && reading.userns_grant == UsernsGrant::Withheld
                        && reading.attachment == ProfileAttachment::Unconfined =>
                {
                    SealedBlocker::AppArmorProfileMissing
                }
                CreationStep::Unshare => SealedBlocker::NamespaceCreationDenied,
                CreationStep::IdentityMaps => SealedBlocker::IdMapDenied,
                CreationStep::Loopback
                | CreationStep::LocalRoute
                | CreationStep::TcpListener
                | CreationStep::Resolver
                | CreationStep::NamespaceHandles
                | CreationStep::HandOff => SealedBlocker::NetworkNamespaceUnavailable,
            };
            return Some(shortfall(
                blocker,
                Some(step.describe().to_string()),
                Some(errno),
            ));
        }
    }
    match &reading.sealed_root {
        MountAttempt::Created => {}
        // Not attempted with the network namespace created is no state the probe produces; it is
        // refused as unknown rather than read as built.
        MountAttempt::NotAttempted => {
            return Some(shortfall(
                SealedBlocker::ProbeUnavailable,
                Some("the sealed-root rehearsal did not run".to_string()),
                None,
            ))
        }
        MountAttempt::Failed(MountAttemptError::Rehearsal(detail)) => {
            return Some(shortfall(
                SealedBlocker::ProbeUnavailable,
                Some(detail.clone()),
                None,
            ))
        }
        MountAttempt::Failed(MountAttemptError::Step { step, errno }) => {
            let errno = *errno;
            let blocker = match step {
                MountStep::JoinSession => SealedBlocker::NetworkNamespaceUnavailable,
                MountStep::UnshareUserMount if KERNEL_ERRNOS.contains(&errno) => {
                    SealedBlocker::KernelUnsupported
                }
                MountStep::UnshareUserMount => SealedBlocker::NamespaceCreationDenied,
                MountStep::IdentityMaps => SealedBlocker::IdMapDenied,
                MountStep::MakePrivateNested | MountStep::Tmpfs => SealedBlocker::MountDenied,
                MountStep::PivotPrepare | MountStep::PivotRoot | MountStep::DetachOldRoot => {
                    SealedBlocker::PivotRootDenied
                }
                // The `/tmp` rehearsal's steps; the sealed-root rehearsal never reports them.
                MountStep::UnshareMount | MountStep::MakePrivate | MountStep::BindTmp => {
                    SealedBlocker::MountDenied
                }
            };
            return Some(shortfall(
                blocker,
                Some(step.describe().to_string()),
                Some(errno),
            ));
        }
    }
    if landlock_fully_enforced {
        None
    } else {
        Some(shortfall(SealedBlocker::LandlockUnavailable, None, None))
    }
}

// ---------------------------------------------------------------- the composed root, declared

/// Host directories bind-mounted read-only into every composed root, in order.
///
/// Fixed, not derived: a whole-tree bind of `/usr` plus the non-usrmerge spellings of the same
/// tree. On a usrmerge distro `/bin`, `/sbin`, `/lib`,
/// `/lib64`… are *symlinks* into `/usr`; [`plan_composed_root`] recreates them as symlinks rather
/// than bind-mounting through them, so the composed root has the same shape as the host and a
/// hashbang like `#!/bin/sh` resolves.
///
/// Deliberately a whole-directory bind rather than a curated per-file set: the point of `sealed`
/// is that it *deletes* the ELF-closure derivation `scoped` needs, along with its lock-time
/// pinning problem. `capabilities.shell.staged_runtime` is **additive** to this list, not a
/// replacement for it: a declared runtime tree arrives through
/// [`plan_composed_root`]'s own `staged_runtime_read_only` parameter, as a *required* bind, while
/// this fixed list stays exactly as it is. Anything a `shell.allow` binary needs at runtime must
/// therefore be under one of these, under a staged runtime tree, or under a directory derived
/// from the manifest — see `extra_read_only` in [`plan_composed_root`].
pub const SEALED_RUNTIME_PATHS: &[&str] = &[
    "/usr", "/bin", "/sbin", "/lib", "/lib32", "/lib64", "/libx32",
];

/// One entry of the fixed `/etc` allowlist: the host path to bind, plus whether it is a directory.
///
/// The `directory` bit exists because the same list drives two mechanisms with different needs.
/// [`plan_composed_root`] does not need it — it asks the host what each path is via
/// [`HostLayout::kind`] and mirrors accordingly. `sandbox::resolve_sealed_etc_landlock_grants` does:
/// it must decide `ReadDir` per entry, and it is required to be pure and syscall-free (it runs on
/// any platform, for paths that need not exist). Carrying the classification *in the list* rather
/// than in a second parallel constant is what makes the bound set and the granted set structurally
/// incapable of drifting apart — the same reason [`SealedDevice`] carries its own `writable` bit
/// rather than deriving it from a separate table.
///
/// It is a static claim about mainstream Linux, not a fact read off this host: `/etc/ssl` and
/// friends are directories on Debian, Ubuntu, Fedora, RHEL, Arch and Alpine alike. A host that
/// disagrees costs the entry its enumerability and nothing else — `open_landlock_fds` narrows
/// `list_dir` back to `false` for anything that does not `fstat` as a directory, because the kernel
/// rejects (`EINVAL`) a `ReadDir` rule on a non-directory and that would otherwise refuse the whole
/// launch.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct SealedEtcPath {
    /// Absolute path inside the composed root. Also the host path bind-mounted there, except for
    /// an entry carrying [`Self::synthetic`], whose content the parent writes itself.
    pub path: &'static str,
    /// Whether this path is a directory on a mainstream Linux host, and therefore whether its
    /// Landlock rule carries `ReadDir`. `false` for regular files and for symlinks (the symlink is
    /// followed when the rule's fd is opened, so `/etc/localtime`'s rule lands on the zoneinfo file
    /// it points at).
    pub directory: bool,
    /// `Some` when [`Self::path`] is backed by a file the *parent* synthesises rather than by the
    /// host's own file at that path — the account databases, and nothing else.
    ///
    /// Carried on the entry for the same reason as `directory`: the alternative is a second
    /// parallel list of "the ones we do not mirror", which is free to drift out of step with this
    /// one. Typed rather than a `bool` so the content each synthetic entry gets is decided by an
    /// exhaustive `match` — adding a third one is then a compile error until it is given content,
    /// instead of a launch-time failure to bind a file nobody wrote.
    pub synthetic: Option<SyntheticEtcFile>,
}

/// An `/etc` file a composed root gets in synthesised form instead of the host's.
///
/// Both are account databases, and both are world-readable on every distribution — so binding the
/// host's copy hands the capsule the machine's full account list for no benefit. What a capsule
/// actually needs from them is that `getpwuid(3)`/`getgrgid(3)` resolve *its own* id: a shell's `~`
/// expansion, `os.path.expanduser`, `whoami`, and anything reading `$HOME` through the password
/// database. Two entries — `root` and the id the capsule runs as — satisfy all of that.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum SyntheticEtcFile {
    /// `/etc/passwd`, in the seven colon-separated fields of `passwd(5)`.
    Passwd,
    /// `/etc/group`, in the four colon-separated fields of `group(5)`.
    Group,
}

/// The uid/gid a capsule's subprocesses run as, and the `HOME` they are given.
///
/// Read off the session workdir the parent created rather than from `getuid(2)`: this crate is
/// `#![deny(unsafe_code)]` and the workdir's owner *is* the spawning process's identity, so
/// `std::fs::metadata(workdir).uid()` answers the question in safe std — the same pattern
/// `crate::resources` already uses.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub(crate) struct SealedAccountIdentity<'a> {
    pub(crate) uid: u32,
    pub(crate) gid: u32,
    /// The synthetic `HOME`, exactly as `crate::shell::build_shell_env` sets it in the subprocess
    /// environment. Byte-identical or the whole exercise is pointless: a `pw_dir` that disagrees
    /// with `$HOME` is worse than no entry at all, because tools silently pick one of the two.
    pub(crate) home: &'a str,
}

/// Login shell recorded for both synthetic accounts. Present in the composed root
/// (`/bin` is on [`SEALED_RUNTIME_PATHS`]) and *not* thereby made runnable — the exec allowlist is
/// Landlock's, and a shell field is a string in a database, not a grant.
const SYNTHETIC_LOGIN_SHELL: &str = "/bin/sh";

/// Account and group name for the capsule's own entry. Deliberately not the host's login name:
/// that name is itself part of the account list this file exists to stop leaking.
const SYNTHETIC_ACCOUNT_NAME: &str = "capsule";

impl SyntheticEtcFile {
    /// File name under the session's `sealed-etc/`. The basename of the target path, so that
    /// directory reads like the `/etc` it stands in for.
    pub(crate) fn file_name(self) -> &'static str {
        match self {
            Self::Passwd => "passwd",
            Self::Group => "group",
        }
    }

    /// The exact bytes written to the staging file, for one capsule identity.
    ///
    /// At most two lines: `root`, plus the capsule's own id when that is not already 0. A capsule
    /// running as root gets **one** line — the `root` line, with `$HOME` as its home directory —
    /// rather than a duplicate entry for uid 0, which `getpwuid(3)` would resolve to whichever came
    /// first.
    ///
    /// Fails rather than emits a corrupt database when the home path contains a field separator or
    /// a newline: `passwd(5)` has no escaping, so a workdir named with a `:` would silently shift
    /// every field after it. Parent-side and synchronous, so it surfaces as a named launch failure.
    pub(crate) fn render(self, identity: &SealedAccountIdentity) -> Result<String, String> {
        let SealedAccountIdentity { uid, gid, home } = *identity;
        match self {
            Self::Passwd => {
                if home.contains(':') || home.contains('\n') {
                    return Err(format!(
                        "sealed: the synthetic HOME {home} contains a character /etc/passwd cannot \
                         quote (':' or a newline), so no valid passwd entry can name it"
                    ));
                }
                // name:password:uid:gid:gecos:home:shell — `x` in the password field is the
                // universal "look in shadow", and there is no shadow file in a composed root.
                if uid == 0 {
                    return Ok(format!(
                        "root:x:0:{gid}:root:{home}:{SYNTHETIC_LOGIN_SHELL}\n"
                    ));
                }
                Ok(format!(
                    "root:x:0:0:root:/root:{SYNTHETIC_LOGIN_SHELL}\n\
                     {SYNTHETIC_ACCOUNT_NAME}:x:{uid}:{gid}:Murmur capsule:{home}:\
                     {SYNTHETIC_LOGIN_SHELL}\n"
                ))
            }
            // name:password:gid:members — the member list is empty because membership of the
            // capsule's own group comes from its passwd entry's gid field.
            Self::Group => {
                if gid == 0 {
                    return Ok("root:x:0:\n".to_string());
                }
                Ok(format!("root:x:0:\n{SYNTHETIC_ACCOUNT_NAME}:x:{gid}:\n"))
            }
        }
    }
}

/// The only `/etc` entries that exist inside a composed root, each bind-mounted read-only and each
/// silently skipped when the host does not have it.
///
/// An allowlist, never `/etc` wholesale, and never a denylist of sensitive names: `/etc/shadow`,
/// `/etc/sudoers`, `/etc/ssh`, cloud-init credentials and every future addition to `/etc` are
/// absent by construction rather than by enumeration. The entries here are the ones without which
/// ordinary tooling misreports rather than fails cleanly — the dynamic loader's cache, the TLS
/// trust store, DNS configuration, the timezone, and the passwd/group databases `getpwuid(3)`
/// needs.
///
/// `/etc/passwd` and `/etc/group` are the two entries the host does *not* supply. They are
/// world-readable on every distribution, so binding the host's copies would hand the capsule the
/// machine's full account list; instead the parent writes a two-line database naming only `root`
/// and the capsule's own id, and the composed root binds *that* at the same path, read-only. See
/// [`SyntheticEtcFile`]. It stays on this list because the list is also the grant list — an entry
/// removed from here loses its Landlock rule and becomes unreadable, which is the opposite of what
/// this narrowing is for.
///
/// Binding these is only half the job: Landlock still mediates *inside* the composed root, so an
/// entry bound here but absent from the ruleset exists and is unopenable — a trust store present
/// at `/etc/ssl/certs/ca-certificates.crt` but `EACCES` on open fails TLS verification.
/// `sandbox::resolve_sealed_etc_landlock_grants` turns this list into the matching read grants.
/// Keep the two sets aligned. A Landlock rule names the *inode* an fd resolved to, not the path
/// string it was opened by, so an entry whose bind source is not the host path of the same name
/// needs its rule taken on the file actually bound — see `sandbox::LandlockChildFds`'s
/// `sealed_identity_fds`, which does that for the two synthetic entries.
pub const SEALED_ETC_PATHS: &[SealedEtcPath] = &[
    SealedEtcPath {
        path: "/etc/ld.so.cache",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/ld.so.conf",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/ld.so.conf.d",
        directory: true,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/alternatives",
        directory: true,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/ssl",
        directory: true,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/pki",
        directory: true,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/ca-certificates",
        directory: true,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/ca-certificates.conf",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/resolv.conf",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/hosts",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/nsswitch.conf",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/localtime",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/timezone",
        directory: false,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/terminfo",
        directory: true,
        synthetic: None,
    },
    SealedEtcPath {
        path: "/etc/passwd",
        directory: false,
        synthetic: Some(SyntheticEtcFile::Passwd),
    },
    SealedEtcPath {
        path: "/etc/group",
        directory: false,
        synthetic: Some(SyntheticEtcFile::Group),
    },
];

/// One entry of the private `/dev` tmpfs.
///
/// `major`/`minor`/`mode` are recorded for the record and for the manual-verification document to
/// assert against — they are **not** used to create the node. Device nodes cannot be `mknod`ed
/// from inside a non-initial user namespace at all (the kernel refuses regardless of
/// `CAP_MKNOD`), so each entry is created as an empty file in the tmpfs and the host's node is
/// bind-mounted over it. This is what bubblewrap and rootless podman do, and it means the numbers
/// a capsule actually sees are the host's — which on every mainstream distribution are exactly the
/// ones below.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct SealedDevice {
    /// Name under `/dev`, without the directory.
    pub name: &'static str,
    /// The host path bind-mounted over the tmpfs entry — `/dev/` + [`Self::name`], spelled out
    /// because it is also what the composed root's Landlock rule set is built from.
    pub path: &'static str,
    pub major: u32,
    pub minor: u32,
    pub mode: u32,
    /// Whether the composed root's Landlock rule for this device carries `WriteFile` on top of
    /// `ReadFile`.
    ///
    /// Landlock still mediates inside a composed root — that is the whole point of keeping it as
    /// defence in depth — so a device present in `/dev` but absent from the rule set would exist
    /// and be unopenable, which is worse than not being there. The read-only pair is the two
    /// entropy sources, where nothing legitimate writes; the rest carry write because their write
    /// side is defined to discard (`null`, `zero`), to report `ENOSPC` (`full`), or to be a
    /// terminal (`tty`).
    pub writable: bool,
}

/// The OCI runtime-spec default device set, which is what a `sealed` capsule's `/dev` contains and
/// all it contains.
///
/// This is deliberately *independent* of `sandbox::CAPSULE_DEVICE_GRANTS`, the three-device
/// Landlock grant set that services `scoped`. That set is a rule list layered over the host's real
/// `/dev`; this one is the entire contents of a private tmpfs. They answer different questions and
/// neither should be derived from the other.
///
/// `/dev/console` is omitted (the OCI spec creates it only for a terminal-attached container, and
/// a capsule subprocess has pipes, not a controlling terminal). `/dev/ptmx` and `/dev/pts` come
/// from a `devpts` mount rather than from this list, and `/dev/shm` is omitted deliberately: it is
/// writable, and the session workdir, `/tmp` and the home folder are the only writable paths in a
/// composed root.
pub const SEALED_DEVICE_NODES: &[SealedDevice] = &[
    SealedDevice {
        name: "null",
        path: "/dev/null",
        major: 1,
        minor: 3,
        mode: 0o666,
        writable: true,
    },
    SealedDevice {
        name: "zero",
        path: "/dev/zero",
        major: 1,
        minor: 5,
        mode: 0o666,
        writable: true,
    },
    SealedDevice {
        name: "full",
        path: "/dev/full",
        major: 1,
        minor: 7,
        mode: 0o666,
        writable: true,
    },
    SealedDevice {
        name: "random",
        path: "/dev/random",
        major: 1,
        minor: 8,
        mode: 0o666,
        writable: false,
    },
    SealedDevice {
        name: "urandom",
        path: "/dev/urandom",
        major: 1,
        minor: 9,
        mode: 0o666,
        writable: false,
    },
    SealedDevice {
        name: "tty",
        path: "/dev/tty",
        major: 5,
        minor: 0,
        mode: 0o666,
        writable: true,
    },
];

/// The OCI default `/dev` symlinks, `(link name under /dev, target)`.
pub const SEALED_DEVICE_SYMLINKS: &[(&str, &str)] = &[
    ("fd", "/proc/self/fd"),
    ("stdin", "/proc/self/fd/0"),
    ("stdout", "/proc/self/fd/1"),
    ("stderr", "/proc/self/fd/2"),
    ("ptmx", "pts/ptmx"),
];

/// Host path of the file backing one synthetic `/etc` entry, in the session's `sealed-etc/`
/// directory beside its runtime directory, which no capsule command can write.
///
/// The single place this path is composed. Three callers need to agree on it byte for byte — the
/// planner that binds it, the parent that writes it, and the Landlock fd that grants it — and two
/// of them run in different phases of the launch.
pub(crate) fn synthetic_etc_source(sealed_etc_dir: &Path, file: SyntheticEtcFile) -> PathBuf {
    sealed_etc_dir.join(file.file_name())
}

/// The host-side sources a composed root binds that the runtime itself provides, each resolved
/// by the parent before the fork.
#[derive(Debug, Clone, Copy)]
pub(crate) struct SealedBindSources<'a> {
    /// The session's temporary folder ([`crate::temp_dir`]), bound at `/tmp`, with the
    /// `(st_dev, st_ino)` of the descriptor the session's records root holds on it — the identity
    /// the bind verifies, so a link swapped onto its path moves nothing ([`RootOp::BindVerified`]).
    pub(crate) tmp_store: (&'a Path, (u64, u64)),
    /// The session's `sealed-etc/` directory, holding the synthetic `/etc/passwd` and
    /// `/etc/group`.
    pub(crate) etc_dir: &'a Path,
    /// The session's runtime directory and the identity of the runtime root's own descriptor,
    /// bound read-only and required at its own absolute path.
    pub(crate) runtime_dir: Option<(&'a Path, (u64, u64))>,
    /// The session's records directory and the identity of the records root's own descriptor,
    /// bound read-only and required at its own absolute path, on [`Self::runtime_dir`]'s terms.
    pub(crate) records_dir: Option<(&'a Path, (u64, u64))>,
    /// The session's home folder, `home/` in its records directory, and the identity of the
    /// descriptor its home root holds on it, bound read-write and required at its own absolute
    /// path, after [`Self::records_dir`], so `$HOME` names it inside the root.
    pub(crate) home_dir: Option<(&'a Path, (u64, u64))>,
    /// A native tool's durable store, `~/.murmur/state/<store>/`, and the identity of the
    /// descriptor its spawn took on it, bound read-write and required at its own absolute path,
    /// after [`Self::home_dir`], so `$MURMUR_STATE_DIR` names it inside the root. `None` for every
    /// spawn but the declaring tool's.
    pub(crate) state_dir: Option<(&'a Path, (u64, u64))>,
    /// The host paths `~/.murmur/config.yaml` lends the session, in config order, each bound
    /// verified and required at its own absolute path before everything else.
    pub(crate) host_paths: &'a [HostPathBind],
    /// Every execute root inside a `read-write` host path, bound read-only and required on top of
    /// it so the capsule cannot rewrite what it may run.
    pub(crate) host_path_read_only: &'a [PathBuf],
}

/// One host path's bind: the directory the entry resolved to, the `(st_dev, st_ino)` of the
/// descriptor the parent took on it, and whether it is bound read-only (`access: read`).
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) struct HostPathBind {
    pub(crate) path: PathBuf,
    pub(crate) identity: (u64, u64),
    pub(crate) read_only: bool,
}

/// Directory the old root is parked in between `pivot_root(2)` and `umount2(MNT_DETACH)`.
/// Removed immediately afterwards, so it does not exist for any process the capsule can run.
const OLD_ROOT_NAME: &str = ".mur-oldroot";

/// [`OLD_ROOT_NAME`] relative, and rooted at `/` after the pivot, as NUL-terminated literals.
///
/// The probe's sealed-root rehearsal pivots after `fork()`, where composing a `CString` is not
/// allowed. `old_root_names_agree` keeps the three spellings in step.
pub(crate) const OLD_ROOT_NAME_C: &std::ffi::CStr = c".mur-oldroot";

/// [`OLD_ROOT_NAME`] as it is named from inside the new root, once `pivot_root(2)` has parked the
/// old one there.
pub(crate) const OLD_ROOT_PIVOTED_C: &std::ffi::CStr = c"/.mur-oldroot";

/// Base directories the new-root `tmpfs` may be mounted over, most preferred first.
///
/// The base is overmounted inside the private namespace, so whatever the host has there becomes
/// unreachable for the rest of the setup — which is why the chosen base must not be an ancestor of
/// the session workdir (a workdir under `/tmp` is entirely normal) and must not be one of the
/// directories the composed root goes on to bind-mount from.
pub const SEALED_ROOT_BASE_CANDIDATES: &[&str] = &["/tmp", "/run", "/var/tmp", "/mnt", "/media"];

/// `hidepid` spellings tried, in order, when mounting the composed root's `/proc` with
/// `mount -t proc`. The numeric form is the legacy parser's; `invisible` is the Linux 5.8+
/// spelling; the empty string is a plain private `procfs` with no masking.
///
/// **On a bare host none of the three succeeds, and that is a kernel rule.**
/// `proc_fill_super` requires `CAP_SYS_ADMIN` over the user namespace *owning the PID namespace*.
/// `unshare(CLONE_NEWUSER | CLONE_NEWNS)` leaves the process in the host's initial PID namespace,
/// whose owner is the initial user namespace — where an unprivileged process has no capabilities
/// at all — so every `mount -t proc` returns `EPERM`. Adding `CLONE_NEWPID` does not help either:
/// `unshare` moves only *future children* into the new PID namespace, so the mounting process is
/// still judged against the old one. Mounting a private `procfs` unprivileged genuinely requires
/// forking so the child becomes PID 1, which changes reaping and signal semantics for the whole
/// capsule subprocess tree and is out of scope here (see the PID-namespace note in
/// `docs/content/reference/sealed-containment-manual-verification.md`).
///
/// They are still tried first, and in this order, because they *do* succeed where the process has
/// that privilege — as root, and inside a container granted `CAP_SYS_ADMIN` — and because
/// a later PID-namespace slice makes them succeed everywhere without touching this list.
///
/// When they all fail the executor binds the host's `/proc` instead. The capsule can then
/// enumerate host PIDs and read the `/proc/<pid>/root` and `/proc/<pid>/cwd` symlinks, so `/proc`
/// is the one part of the composed root where "outside does not exist" degrades to `scoped`'s
/// "outside is denied" — Landlock's ruleset covers no path under `/proc`, so opens through it are
/// refused, and `ptrace_may_access` gates the rest. Every other axis of the root stays absolute. The alternative is a capsule with no `/proc` at all, which
/// breaks `/dev/fd`, process substitution and every runtime that reads `/proc/self/*`.
const PROC_HIDEPID_OPTIONS: &[&str] = &["hidepid=2", "hidepid=invisible", ""];

// ---------------------------------------------------------------- host layout (test seam)

/// What a path is on the host, as far as composing a root cares.
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) enum PathKind {
    Dir,
    /// Any non-directory: a regular file, a device node, a socket. All are bind-mount targets that
    /// need an empty *file* created for them rather than a directory.
    File,
    /// A symlink, with its literal (un-resolved) target.
    Symlink(PathBuf),
}

/// The host filesystem facts [`plan_composed_root`] reads. A trait purely so the planner can be
/// unit-tested against a synthetic usrmerge / non-usrmerge layout on any machine — production has
/// exactly one implementation, [`RealHostLayout`].
pub(crate) trait HostLayout {
    fn kind(&self, path: &Path) -> Option<PathKind>;
}

/// The real filesystem, read with `symlink_metadata` so a usrmerge `/bin -> usr/bin` reports as a
/// symlink rather than as the directory it points at.
pub(crate) struct RealHostLayout;

impl HostLayout for RealHostLayout {
    fn kind(&self, path: &Path) -> Option<PathKind> {
        let meta = std::fs::symlink_metadata(path).ok()?;
        if meta.file_type().is_symlink() {
            return Some(PathKind::Symlink(std::fs::read_link(path).ok()?));
        }
        if meta.is_dir() {
            return Some(PathKind::Dir);
        }
        Some(PathKind::File)
    }
}

// ---------------------------------------------------------------- the plan

/// One filesystem operation in the composed-root plan.
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) enum RootOp {
    /// `mkdir(path, 0755)`. `EEXIST` is success.
    MkDir(PathBuf),
    /// Create an empty regular file, to be bind-mounted over.
    MkFile(PathBuf),
    /// `symlink(target, link)` — used to reproduce the host's usrmerge shape and the OCI `/dev`
    /// symlinks.
    Symlink { target: PathBuf, link: PathBuf },
    /// `mount(source, target, MS_BIND | MS_REC)`, followed by a second
    /// `MS_REMOUNT | MS_BIND | MS_RDONLY` call when `read_only`. The second call is required: a
    /// single `MS_BIND | MS_RDONLY` mount does *not* produce a read-only bind.
    Bind {
        source: PathBuf,
        target: PathBuf,
        read_only: bool,
    },
    /// [`Self::Bind`] of a directory the parent holds a descriptor on: the child opens `source`
    /// after it has entered its own mount namespace, refuses unless the directory it reached has
    /// `identity` — the `(st_dev, st_ino)` of the parent's descriptor — and binds through the
    /// descriptor it opened. A link swapped onto `source` after the parent took its descriptor
    /// fails the step rather than redirecting the bind.
    ///
    /// The parent's descriptor itself cannot be the source: a bind's source must be a mount in
    /// the caller's own namespace, and one opened before `unshare(CLONE_NEWNS)` is not (`EINVAL`).
    BindVerified {
        source: PathBuf,
        target: PathBuf,
        read_only: bool,
        identity: (u64, u64),
    },
    /// A fresh `tmpfs`.
    Tmpfs {
        target: PathBuf,
        options: &'static str,
    },
    /// A `/proc`: a fresh `procfs` masked with `hidepid` where the kernel permits one, a bind of
    /// the host's where it does not. Which of the two a host gets is decided at execution time, not
    /// here — see [`PROC_HIDEPID_OPTIONS`].
    Proc { target: PathBuf },
    /// A fresh `devpts` instance, so the capsule can allocate its own ptys without seeing the
    /// host's.
    DevPts { target: PathBuf },
    /// `mount(NULL, target, NULL, MS_REMOUNT | MS_BIND | MS_RDONLY)` — used to seal `/dev` after
    /// its device nodes have been bound into it.
    RemountReadOnly(PathBuf),
}

/// One planned operation plus whether failing it aborts the launch.
///
/// `required: false` is the "shrink, do not fail" convention this crate already uses for optional
/// device grants in `sandbox::open_landlock_fds`: losing an optional bind makes the composed root
/// *narrower*, never wider, so refusing the launch over it would trade a working capsule for no
/// security gain.
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) struct RootStep {
    pub(crate) op: RootOp,
    pub(crate) required: bool,
}

/// The full, host-resolved recipe for one composed root.
#[derive(Debug, Clone, PartialEq, Eq)]
pub(crate) struct ComposedRootPlan {
    /// Directory the new-root `tmpfs` is mounted over, and the directory `pivot_root` is handed.
    pub(crate) base: PathBuf,
    pub(crate) steps: Vec<RootStep>,
    /// The workdir's path *inside* the composed root. Identical to its host path by design: env
    /// vars, `PWD`, log paths and anything a previous turn recorded keep working, and the capsule
    /// never has to be told it moved.
    pub(crate) workdir_in_root: PathBuf,
}

/// Picks the directory the new-root `tmpfs` is mounted over.
///
/// Rejects any candidate that is the workdir or an ancestor of it — overmounting an ancestor of
/// the workdir would hide the very directory the plan then has to bind-mount — on the same terms
/// any candidate that is a host path or an ancestor of one, and any candidate the host does not
/// have. Pure apart from the injected `exists` predicate.
pub(crate) fn choose_root_base(
    workdir: &Path,
    host_paths: &[PathBuf],
    candidates: &[&str],
    exists: impl Fn(&Path) -> bool,
) -> Option<PathBuf> {
    candidates
        .iter()
        .map(Path::new)
        .find(|candidate| {
            !workdir.starts_with(candidate)
                && !host_paths
                    .iter()
                    .any(|host_path| host_path.starts_with(candidate))
                && exists(candidate)
        })
        .map(Path::to_path_buf)
}

/// Builds the ordered operation list for one composed root. Pure with respect to `host`.
///
/// `extra_read_only` carries the host directories this *particular* capsule needs beyond
/// [`SEALED_RUNTIME_PATHS`] — the containing directory of each resolved `shell.allow` binary and
/// each `capabilities.shell.interpreter_runtime` directory. They are whole-directory binds, not a
/// per-file grant set: a bind carries a directory's whole contents, so this must not grow into a
/// second ELF-closure derivation. Entries already covered by a fixed path are dropped.
///
/// `staged_runtime_read_only` carries the `source_path` of each declared
/// `capabilities.shell.staged_runtime` grant. It is a *separate* parameter from `extra_read_only`
/// rather than more entries in it, because the two have opposite failure semantics and the
/// difference is the whole point of the capability. See [`PlanBuilder::require_bind`].
pub(crate) fn plan_composed_root(
    accessible_workdir: &Path,
    sources: &SealedBindSources<'_>,
    base: &Path,
    extra_read_only: &[PathBuf],
    staged_runtime_read_only: &[PathBuf],
    host: &dyn HostLayout,
) -> ComposedRootPlan {
    let mut builder = PlanBuilder::new(base);

    // -1. Host paths from `~/.murmur/config.yaml`, verified and REQUIRED, shallowest first so a
    //     nested entry inside another lands on top of it; then every execute root inside a
    //     read-write one, read-only and REQUIRED on top of that, so the capsule cannot rewrite
    //     what it may run. First because a host path is disjoint from every fixed path the root
    //     binds: every runtime folder beneath one (the workdir, the runtime, records and home
    //     folders, a native tool's store) is bound later, on top, and keeps its own access.
    let mut host_paths: Vec<&HostPathBind> = sources.host_paths.iter().collect();
    host_paths.sort_by_key(|host_path| host_path.path.components().count());
    for host_path in &host_paths {
        builder.host_path_bind(host_path);
    }
    for path in sources.host_path_read_only {
        builder.require_bind(path);
    }

    // 0. Declared `staged_runtime` trees, read-only and REQUIRED — before everything else.
    //
    //    Ordering is load-bearing, not tidiness. `mirror` below deduplicates by target path via
    //    the `made` set, so whichever loop registers a target first wins it and any later loop
    //    silently no-ops on the same target. Running the required binds first means a staged path
    //    that happens to collide with an optional one can only ever be *upgraded* to required,
    //    never silently downgraded to optional.
    //
    //    These are also planned ahead of /dev, /proc, /tmp and the workdir bind — and therefore
    //    ahead of `pivot_root`, which `construct_composed_root` performs only after every planned
    //    step has run. A capsule never reaches a shell tool call inside a root that is missing a
    //    runtime tree it declared.
    for path in staged_runtime_read_only {
        builder.require_bind(path);
    }

    // 1. The host runtime tree, read-only. A usrmerge symlink is reproduced as a symlink so the
    //    composed root has the host's shape; a real directory is bind-mounted.
    for path in SEALED_RUNTIME_PATHS
        .iter()
        .map(Path::new)
        .chain(extra_read_only.iter().map(PathBuf::as_path))
    {
        builder.mirror(path, host, /* required */ false);
    }

    //    A step-0 or step-1 bind that landed over a directory holding a host path (a host path
    //    inside an `interpreter_runtime` directory or a staged tree) hid it, so each such host path
    //    and its read-only execute roots are bound again, on top.
    builder.rebind_covered_host_paths(&host_paths, sources.host_path_read_only);

    // 2. The narrow /etc allowlist, read-only. Everything else under /etc simply does not exist.
    //
    //    An entry marked synthetic is not mirrored from the host at all: its target is bound from
    //    the file the parent wrote in the session's `sealed-etc/` directory.
    //    `required: true` there, unlike the host-mirrored entries' `required: false` — those are
    //    optional because a distribution may genuinely not have them, whereas the parent wrote
    //    these two itself, so a bind that fails is a broken runtime rather than a narrower one, and
    //    the empty `MkFile` left in its place would be a passwd database with no accounts in it.
    for entry in SEALED_ETC_PATHS {
        let path = Path::new(entry.path);
        match entry.synthetic {
            None => builder.mirror(path, host, /* required */ false),
            Some(file) => builder.bind_file(
                synthetic_etc_source(sources.etc_dir, file),
                rebase(base, path),
                /* required */ true,
            ),
        }
    }

    // 3. A private /dev tmpfs carrying the OCI default device set, sealed read-only once it is
    //    populated. Device nodes are bind-mounted from the host because `mknod` of a device is
    //    refused inside a user namespace — see `SealedDevice`.
    let dev = base.join("dev");
    builder.mkdir_p(&dev);
    builder.push(
        RootOp::Tmpfs {
            target: dev.clone(),
            options: "mode=0755,size=1m",
        },
        true,
    );
    for device in SEALED_DEVICE_NODES {
        let source = PathBuf::from(device.path);
        if host.kind(&source).is_none() {
            continue;
        }
        let target = dev.join(device.name);
        builder.push(RootOp::MkFile(target.clone()), false);
        builder.push(
            RootOp::Bind {
                source,
                target,
                read_only: false,
            },
            false,
        );
    }
    let pts = dev.join("pts");
    builder.push(RootOp::MkDir(pts.clone()), false);
    builder.push(RootOp::DevPts { target: pts }, false);
    for (link, target) in SEALED_DEVICE_SYMLINKS {
        builder.push(
            RootOp::Symlink {
                target: PathBuf::from(*target),
                link: dev.join(link),
            },
            false,
        );
    }
    builder.push(RootOp::RemountReadOnly(dev), false);

    // 4. /proc — masked with hidepid where the kernel allows it, bound from the host where it does
    //    not. See `PROC_HIDEPID_OPTIONS`.
    let proc = base.join("proc");
    builder.push(RootOp::MkDir(proc.clone()), true);
    builder.push(RootOp::Proc { target: proc }, true);

    // 5. /tmp, the session's temporary folder in its records directory, counted by the same
    //    size guard as the workdir.
    //
    //    Before the workdir, not after, and the ordering is load-bearing rather than tidy. A
    //    session workdir under `/tmp` is not an edge case — it is `mur run`'s default, since
    //    `--workdir` falls back to a temporary directory. With the workdir reproduced first, this
    //    mount landed *on top of* the path leading to it, the workdir bind vanished under it, and
    //    the construction died at `chdir into the workdir inside the root: ENOENT` — after the
    //    pivot, with the host root already detached. Mounting `/tmp` first means the workdir's path
    //    components are created inside it, so the deeper mount is the one that survives.
    let tmp = base.join("tmp");
    builder.push(RootOp::MkDir(tmp.clone()), true);
    builder.push(
        RootOp::BindVerified {
            source: sources.tmp_store.0.to_path_buf(),
            target: tmp,
            read_only: false,
            identity: sources.tmp_store.1,
        },
        true,
    );

    // 6. The capsule's workdir — the accessible one, which is also the subprocess cwd — at its
    //    own absolute path, read-write. With /tmp above and the home folder below, the only
    //    writable paths in the composed root.
    let workdir_target = rebase(base, accessible_workdir);
    builder.mkdir_p(&workdir_target);
    builder.push(
        RootOp::Bind {
            source: accessible_workdir.to_path_buf(),
            target: workdir_target,
            read_only: false,
        },
        true,
    );

    // 7. The session's runtime directory, read-only and REQUIRED, at its own absolute path, so
    //    `$MURMUR_RUNTIME_DIR` names the same directory inside the root as on the host. After the
    //    workdir for the reason /tmp precedes it: a runtime directory under either is created
    //    inside it, and the deeper mount is the one that survives. A runtime directory that
    //    cannot be bound ends the session rather than leaving its commands without their task
    //    and tools.
    if let Some((path, identity)) = sources.runtime_dir {
        let target = rebase(base, path);
        builder.mkdir_p(&target);
        builder.push(
            RootOp::BindVerified {
                source: path.to_path_buf(),
                target,
                read_only: true,
                identity,
            },
            true,
        );
    }

    // 8. The session's records directory, read-only and REQUIRED, at its own absolute path, so
    //    `$MURMUR_RECORDS_DIR` names it inside the root: a command reads the trace and logs and
    //    changes none of them. Bound after the runtime directory, so the deeper bind wins when one
    //    lies beneath the other. A records directory that cannot be bound ends the session rather than leaving
    //    it writable, or absent from a root whose commands are told where it is.
    if let Some((path, identity)) = sources.records_dir {
        let target = rebase(base, path);
        builder.mkdir_p(&target);
        builder.push(
            RootOp::BindVerified {
                source: path.to_path_buf(),
                target,
                read_only: true,
                identity,
            },
            true,
        );
    }

    // 9. The session's home folder, read-write and REQUIRED, at its own absolute path, so `$HOME`
    //    and the synthetic passwd's home directory name it inside the root. After the records
    //    directory it lies in, so this deeper, writable bind is the one that survives on top of
    //    that read-only one. Verified like `/tmp`: a link swapped onto its path moves nothing.
    if let Some((path, identity)) = sources.home_dir {
        let target = rebase(base, path);
        builder.mkdir_p(&target);
        builder.push(
            RootOp::BindVerified {
                source: path.to_path_buf(),
                target,
                read_only: false,
                identity,
            },
            true,
        );
    }

    // 10. A native tool's durable store, read-write and REQUIRED, at its own absolute path, on
    //     the home folder's terms: a tool that declared a store never runs in a root without it.
    //     Last, so a store beneath any directory bound above it survives on top.
    if let Some((path, identity)) = sources.state_dir {
        let target = rebase(base, path);
        builder.mkdir_p(&target);
        builder.push(
            RootOp::BindVerified {
                source: path.to_path_buf(),
                target,
                read_only: false,
                identity,
            },
            true,
        );
    }

    ComposedRootPlan {
        base: base.to_path_buf(),
        steps: builder.steps,
        workdir_in_root: accessible_workdir.to_path_buf(),
    }
}

/// `base` + `path`, where `path` is absolute: `/tmp` + `/usr/lib` → `/tmp/usr/lib`.
///
/// `pub(crate)` so [`crate::staged_runtime::target_under_root`] can reuse this instead of
/// reimplementing the same re-basing rule.
pub(crate) fn rebase(base: &Path, path: &Path) -> PathBuf {
    let relative = path.strip_prefix("/").unwrap_or(path);
    base.join(relative)
}

struct PlanBuilder {
    base: PathBuf,
    steps: Vec<RootStep>,
    /// Directories already scheduled for creation, so a shared prefix (`/usr`, `/etc`) is not
    /// `mkdir`ed once per entry under it.
    made: std::collections::HashSet<PathBuf>,
    /// Directories in [`Self::made`] only because a host path lies beneath them: a later bind of
    /// one of these is still planned, rather than skipped as already made.
    host_path_ancestors: std::collections::HashSet<PathBuf>,
    /// Targets bound over a host path ancestor, which [`Self::rebind_covered_host_paths`] binds
    /// the hidden host paths again on top of.
    covering: Vec<PathBuf>,
}

impl PlanBuilder {
    fn new(base: &Path) -> Self {
        let mut made = std::collections::HashSet::new();
        made.insert(base.to_path_buf());
        Self {
            base: base.to_path_buf(),
            steps: Vec::new(),
            made,
            host_path_ancestors: std::collections::HashSet::new(),
            covering: Vec::new(),
        }
    }

    /// Schedules one host path's verified, required bind at its own absolute path.
    fn host_path_bind(&mut self, host_path: &HostPathBind) {
        let target = rebase(&self.base, &host_path.path);
        let mut current = self.base.clone();
        if let Ok(relative) = target.strip_prefix(&self.base) {
            let components: Vec<_> = relative.components().collect();
            for component in &components[..components.len().saturating_sub(1)] {
                current = current.join(component);
                if !self.made.contains(&current) {
                    self.host_path_ancestors.insert(current.clone());
                }
            }
        }
        self.mkdir_p(&target);
        self.host_path_ancestors.remove(&target);
        self.push(
            RootOp::BindVerified {
                source: host_path.path.clone(),
                target,
                read_only: host_path.read_only,
                identity: host_path.identity,
            },
            /* required */ true,
        );
    }

    /// Whether a bind at `target` is skipped: something already owns it. A directory made only as
    /// a host path's ancestor does not count, and binding over one is recorded in
    /// [`Self::covering`].
    fn already_owned(&mut self, target: &Path) -> bool {
        if !self.made.contains(target) {
            return false;
        }
        if self.host_path_ancestors.remove(target) {
            self.covering.push(target.to_path_buf());
            return false;
        }
        true
    }

    /// Binds again, on top, every host path (and every read-only execute root) a bind in
    /// [`Self::covering`] hid.
    fn rebind_covered_host_paths(&mut self, host_paths: &[&HostPathBind], read_only: &[PathBuf]) {
        if self.covering.is_empty() {
            return;
        }
        let covering = std::mem::take(&mut self.covering);
        let hidden = |target: &Path| {
            covering
                .iter()
                .any(|cover| target != cover && target.starts_with(cover))
        };
        for host_path in host_paths {
            let target = rebase(&self.base, &host_path.path);
            if hidden(&target) {
                self.push(
                    RootOp::BindVerified {
                        source: host_path.path.clone(),
                        target,
                        read_only: host_path.read_only,
                        identity: host_path.identity,
                    },
                    true,
                );
            }
        }
        for path in read_only {
            let target = rebase(&self.base, path);
            if hidden(&target) {
                self.push(
                    RootOp::Bind {
                        source: path.clone(),
                        target,
                        read_only: true,
                    },
                    true,
                );
            }
        }
    }

    fn push(&mut self, op: RootOp, required: bool) {
        self.steps.push(RootStep { op, required });
    }

    /// Schedules `mkdir` for every not-yet-scheduled component of `target` under the base.
    fn mkdir_p(&mut self, target: &Path) {
        let mut current = self.base.clone();
        let Ok(relative) = target.strip_prefix(&self.base) else {
            return;
        };
        for component in relative.components() {
            current = current.join(component);
            if self.made.insert(current.clone()) {
                self.steps.push(RootStep {
                    op: RootOp::MkDir(current.clone()),
                    required: true,
                });
            }
        }
    }

    /// Schedules an unconditional, **required** read-only directory bind of `source`.
    ///
    /// The deliberate difference from [`Self::mirror`] is that this never consults [`HostLayout`].
    /// `mirror` asks the host whether a path exists and, when it does not, returns having
    /// scheduled *nothing at all* — not even a step that would fail. That is right for
    /// `extra_read_only`, where a missing path should make the composed root narrower rather than
    /// refuse the launch. It is exactly wrong for `staged_runtime`, where a missing source must
    /// fail the launch: routed through `mirror`, a capsule declaring a runtime tree the host does
    /// not have would launch successfully into a root silently lacking it.
    ///
    /// So no existence pre-check happens here, or anywhere else in this module. The plan always
    /// carries the step, `required: true`, and the real `mount(2)` in `execute_step`'s
    /// [`CStepKind::Bind`] arm is the single source of truth for "does this exist" — a missing
    /// source fails with `ENOENT` at construction time, in the child, which
    /// `construct_composed_root`'s `step.required` check turns into a `SealedRootFailure`. That
    /// path already reaches the operator as `RuntimeError::SealedRootConstructionFailed`
    /// (`E-RUN-014`), already names the offending path via [`SealedRootSpec::describe`], and is
    /// already session-fatal, so no error variant of its own is needed.
    ///
    /// Always a directory bind: a staged runtime tree is a tree. `mkdir_p` registers the target in
    /// `made`, so a later `mirror` of the same path is the one that no-ops, never this.
    fn require_bind(&mut self, source: &Path) {
        let target = rebase(&self.base, source);
        if self.already_owned(&target) {
            return;
        }
        self.mkdir_p(&target);
        self.push(
            RootOp::Bind {
                source: source.to_path_buf(),
                target,
                read_only: true,
            },
            /* required */ true,
        );
    }

    /// Reproduces one host path inside the composed root: a symlink as a symlink, a directory or
    /// file as a read-only bind mount. A path the host does not have contributes nothing.
    fn mirror(&mut self, source: &Path, host: &dyn HostLayout, required: bool) {
        let Some(kind) = host.kind(source) else {
            return;
        };
        let target = rebase(&self.base, source);
        if self.already_owned(&target) {
            return;
        }
        match kind {
            PathKind::Symlink(link_target) => {
                if let Some(parent) = target.parent() {
                    self.mkdir_p(parent);
                }
                self.made.insert(target.clone());
                self.push(
                    RootOp::Symlink {
                        target: link_target,
                        link: target,
                    },
                    required,
                );
            }
            PathKind::Dir => {
                self.mkdir_p(&target);
                self.push(
                    RootOp::Bind {
                        source: source.to_path_buf(),
                        target,
                        read_only: true,
                    },
                    required,
                );
            }
            PathKind::File => self.bind_file(source.to_path_buf(), target, required),
        }
    }

    /// Schedules the mountpoint-then-bind pair one file bind needs: `mkdir -p` of its parent, an
    /// empty file to mount over (a bind needs an existing target of the right kind), then the
    /// read-only bind itself.
    ///
    /// Takes `source` and `target` separately rather than deriving one from the other, which is the
    /// whole reason it is a method: [`Self::mirror`] binds a host path at its own path, while a
    /// synthetic `/etc` entry binds a file from inside the workdir at a path that names something
    /// else entirely.
    fn bind_file(&mut self, source: PathBuf, target: PathBuf, required: bool) {
        if self.made.contains(&target) {
            return;
        }
        if let Some(parent) = target.parent() {
            self.mkdir_p(parent);
        }
        self.made.insert(target.clone());
        self.push(RootOp::MkFile(target.clone()), required);
        self.push(
            RootOp::Bind {
                source,
                target,
                read_only: true,
            },
            required,
        );
    }
}

// ---------------------------------------------------------------- Linux: probe + execution

/// Marker every composed-root failure message carries, so `shell.rs` can tell a `pre_exec`
/// sealed-root failure apart from a Landlock or seccomp one after reading it back off the
/// diagnostic pipe — the pipe carries one flat string, so the classification has to live in the
/// text. Declared here rather than inside the Linux submodule because the reader is
/// cross-platform code.
pub(crate) const SEALED_ROOT_FAILURE_PREFIX: &str = "sealed-root:";

#[cfg(target_os = "linux")]
pub(crate) use linux::{build_sealed_root_spec, construct_composed_root, SealedRootSpec};

/// The post-`fork()` namespace primitives [`crate::network_namespace`] reuses rather than
/// reimplementing.
///
/// The `dumpable` pair is shared because a non-dumpable task's `/proc/self/uid_map` is root-owned
/// and therefore unopenable by the task itself, and every namespace `mur` creates hits that trap
/// identically — see [`linux::make_dumpable_for_map_writes`].
///
/// `userns_grant` is shared because it answers where the permission for `unshare(CLONE_NEWUSER)`
/// comes from, which `sealed` and the capsule network namespace both need. Two implementations of
/// one host question could return two answers, and the operator would be told to fix two different
/// things.
#[cfg(target_os = "linux")]
pub(crate) use linux::{
    make_dumpable_for_map_writes, open_verified_dir, proc_fd_path, remount_read_only,
    restore_dumpable, userns_grant, write_decimal_map, PROC_FD_PATH_LEN,
};

#[cfg(target_os = "linux")]
#[allow(unsafe_code)]
mod linux {
    use std::ffi::CString;
    use std::os::unix::ffi::OsStrExt;
    use std::path::Path;

    use super::{
        ComposedRootPlan, RootOp, RootStep, UsernsGrant, OLD_ROOT_NAME, PROC_HIDEPID_OPTIONS,
        SEALED_ROOT_FAILURE_PREFIX,
    };

    /// Reads the three host files [`UsernsGrant::from_readings`] decides on, which holds the
    /// whole decision.
    ///
    /// The knob is read from `/sys/module/apparmor/parameters` first and from `/proc/sys/kernel`
    /// only if that is missing: the two are the same value, and the `/sys` path is the one that
    /// exists on kernels where the `/proc` alias does not. Reading `/proc/self/attr/current`
    /// rather than the loaded-profile list is deliberate: a profile that is loaded but does not
    /// *attach* to the path `mur` was installed at helps nobody, and this asks the question that
    /// decides the outcome.
    pub(crate) fn userns_grant() -> UsernsGrant {
        let enabled = read_trimmed("/sys/module/apparmor/parameters/enabled");
        let restricted =
            read_trimmed("/sys/module/apparmor/parameters/restrict_unprivileged_userns")
                .or_else(|| read_trimmed("/proc/sys/kernel/apparmor_restrict_unprivileged_userns"));
        let current = read_trimmed("/proc/self/attr/current");
        UsernsGrant::from_readings(
            enabled.as_deref(),
            restricted.as_deref(),
            current.as_deref(),
        )
    }

    fn read_trimmed(path: &str) -> Option<String> {
        std::fs::read_to_string(path)
            .ok()
            .map(|value| value.trim().to_string())
            .filter(|value| !value.is_empty())
    }

    /// Flips `PR_SET_DUMPABLE` for the duration of the identity-map writes, and back again.
    ///
    /// Not optional. `mur` marks itself
    /// non-dumpable at startup (`security::harden_process_dumpable`, `prctl(PR_SET_DUMPABLE, 0)`)
    /// so that no same-uid process can read its `/proc/<pid>/environ`. That flag is inherited
    /// across `fork()`, and the kernel's `task_dump_owner()` reassigns *every* `/proc/<pid>/*`
    /// entry of a non-dumpable task to root — including `setgroups`, `uid_map` and `gid_map`. An
    /// unprivileged process therefore cannot open its own `uid_map` for writing, and the namespace
    /// it just created is one it can never own.
    ///
    /// The symptom is misleading: `unshare` succeeds, the map write fails with `EACCES`, and the
    /// refusal blames the host's id-mapping policy — on a host whose id-mapping policy is fine,
    /// and where the identical syscall sequence run from any other program succeeds. Without this
    /// flip, `sealed` is unreachable on every host for that reason alone.
    ///
    /// The window is reopened for exactly the three `open`/`write`/`close` pairs that need it and
    /// closed again before the first mount, so the child spends no longer readable than the map
    /// writes take.
    ///
    /// [`restore_dumpable`] puts back whatever the flag *was*, read here and returned, rather than
    /// assuming `0`. Hardcoding `0` looked equivalent — `mur` always runs non-dumpable — and was
    /// not: it also cleared the flag for every process that had *not* hardened itself, which
    /// includes this crate's own integration tests, and a non-dumpable child is one whose
    /// `/proc/<pid>/mem` the seccomp-notify supervisor could not read — every allowlisted `execve`
    /// then failed closed to `EACCES`.
    ///
    /// # Safety
    /// Post-`fork()`, pre-exec child context only — `prctl(2)` is async-signal-safe, but the
    /// dumpable window this reopens is only sound while this process is the single-threaded child
    /// that is about to `execve` a sandboxed binary.
    pub(crate) unsafe fn make_dumpable_for_map_writes() -> libc::c_int {
        // SAFETY: both `prctl` calls take four integer arguments; no pointers cross the boundary.
        // `PR_GET_DUMPABLE` returning `-1` means the query failed, in which case the restore below
        // is skipped rather than guessing.
        unsafe {
            let previous = libc::prctl(libc::PR_GET_DUMPABLE, 0, 0, 0, 0);
            libc::prctl(libc::PR_SET_DUMPABLE, 1 as libc::c_int, 0, 0, 0);
            previous
        }
    }

    /// Undoes [`make_dumpable_for_map_writes`], restoring the exact value it found.
    ///
    /// # Safety
    /// Same window and same constraints as [`make_dumpable_for_map_writes`]; `previous` must be
    /// the value that call returned.
    pub(crate) unsafe fn restore_dumpable(previous: libc::c_int) {
        if previous < 0 {
            return;
        }
        // SAFETY: integer-only `prctl`, in the post-fork child.
        unsafe {
            libc::prctl(libc::PR_SET_DUMPABLE, previous, 0, 0, 0);
        }
    }

    /// Writes either the literal `deny` (when `id` is `None`) or an identity map line
    /// `"<id> <id> 1\n"`, formatting the number into a stack buffer.
    ///
    /// Hand-rolled rather than `format!` because this runs in a forked child: the parent may have
    /// been mid-allocation on another thread when `fork()` returned, so touching the allocator
    /// here can deadlock. `id` never exceeds `u32::MAX`, so ten digits is the whole range.
    ///
    /// # Safety
    /// Must only be called from the post-`fork()`, pre-exec window in the child, before any other
    /// thread could exist in this process — the same async-signal-safety constraint documented on
    /// [`construct_composed_root`]'s `unsafe` block, which this function's callers all sit inside.
    pub(crate) unsafe fn write_decimal_map(
        path: &std::ffi::CStr,
        id: Option<libc::uid_t>,
        value: libc::uid_t,
    ) -> Result<(), ()> {
        let Some(_) = id else {
            return write_file(path, b"deny");
        };

        let mut buf = [0u8; 32];
        let mut len = 0usize;
        push_decimal(&mut buf, &mut len, value);
        buf[len] = b' ';
        len += 1;
        push_decimal(&mut buf, &mut len, value);
        buf[len] = b' ';
        len += 1;
        buf[len] = b'1';
        len += 1;
        buf[len] = b'\n';
        len += 1;

        write_file(path, &buf[..len])
    }

    // ------------------------------------------------------------ spec (parent side)

    /// One planned operation, lowered to NUL-terminated C strings plus a pre-rendered label.
    ///
    /// The label is built here, in the parent, precisely so the child never has to: a failure
    /// message that names its step costs one `String` per step at spawn time and zero allocation
    /// inside `pre_exec`.
    struct CStep {
        kind: CStepKind,
        required: bool,
        label: String,
    }

    enum CStepKind {
        MkDir(CString),
        MkFile(CString),
        Symlink {
            target: CString,
            link: CString,
        },
        Bind {
            source: CString,
            target: CString,
            read_only: bool,
        },
        BindVerified {
            source: CString,
            target: CString,
            read_only: bool,
            identity: (u64, u64),
        },
        Tmpfs {
            target: CString,
            options: CString,
        },
        Proc {
            target: CString,
        },
        DevPts {
            target: CString,
        },
        RemountReadOnly(CString),
    }

    /// Everything the child needs to build a composed root, fully resolved before `fork()`.
    pub(crate) struct SealedRootSpec {
        uid_map: Vec<u8>,
        gid_map: Vec<u8>,
        base: CString,
        /// `<base>/.mur-oldroot` — created before the pivot.
        old_root_full: CString,
        /// `.mur-oldroot`, relative, as `pivot_root(2)` wants it with `cwd == base`.
        old_root_relative: CString,
        /// `/.mur-oldroot`, as it is named after the pivot, for the detach and the rmdir.
        old_root_pivoted: CString,
        steps: Vec<CStep>,
        workdir_in_root: CString,
    }

    /// Where a composed root's construction failed. `Copy` and allocation-free by construction —
    /// it is produced inside `pre_exec`; the message is rendered by the caller, afterwards.
    #[derive(Debug, Clone, Copy, PartialEq, Eq)]
    pub(crate) struct SealedRootFailure {
        /// Index into the spec's step list, or `usize::MAX` for one of the fixed
        /// unshare/pivot/chdir stages.
        pub(crate) step: usize,
        /// A fixed stage name when `step` is `usize::MAX`; otherwise unused.
        pub(crate) stage: &'static str,
        pub(crate) errno: i32,
    }

    impl SealedRootFailure {
        fn stage(stage: &'static str) -> Self {
            Self {
                step: usize::MAX,
                stage,
                errno: std::io::Error::last_os_error().raw_os_error().unwrap_or(0),
            }
        }

        fn step(index: usize) -> Self {
            Self {
                step: index,
                stage: "",
                errno: std::io::Error::last_os_error().raw_os_error().unwrap_or(0),
            }
        }
    }

    impl SealedRootSpec {
        /// Renders a failure into the message that reaches the operator, naming both the step and
        /// the errno. Runs in the *parent's* error path (via the diagnostic pipe) or in the
        /// child's, where an allocation is already unavoidable because the launch is over.
        pub(crate) fn describe(&self, failure: SealedRootFailure) -> String {
            let what = if failure.step == usize::MAX {
                failure.stage.to_string()
            } else {
                self.steps
                    .get(failure.step)
                    .map(|step| step.label.clone())
                    .unwrap_or_else(|| format!("step {}", failure.step))
            };
            format!(
                "{SEALED_ROOT_FAILURE_PREFIX} {what} failed: {}",
                std::io::Error::from_raw_os_error(failure.errno)
            )
        }
    }

    /// Remounts the bind at `target` read-only and `nosuid`, keeping the flags its source mount
    /// locks. `target` may be a `/proc/self/fd/<n>` path from [`proc_fd_path`].
    ///
    /// # Safety
    /// Allocation-free; callable post-`fork()`, pre-exec, with the capabilities to remount.
    pub(crate) unsafe fn remount_read_only(target: &std::ffi::CStr) -> Result<(), ()> {
        // Inside a user namespace the bind inherits the source mount's flags locked: a remount
        // that does not repeat them is refused with `EPERM`. `statvfs` on the new bind reports
        // them from the kernel's `statfs` `f_flags`, with no allocation.
        let mut stat: libc::statvfs = std::mem::zeroed();
        let kept = if libc::statvfs(target.as_ptr(), &mut stat) == 0 {
            remount_flags_kept(stat.f_flag)
        } else {
            0
        };
        if libc::mount(
            std::ptr::null(),
            target.as_ptr(),
            std::ptr::null(),
            libc::MS_REMOUNT | libc::MS_BIND | libc::MS_RDONLY | libc::MS_NOSUID | kept,
            std::ptr::null(),
        ) != 0
        {
            return Err(());
        }
        Ok(())
    }

    /// `"/proc/self/fd/"` and the longest decimal `c_int` and a NUL.
    pub(crate) const PROC_FD_PATH_LEN: usize = 32;

    /// Writes `/proc/self/fd/<fd>` and a NUL into `out`, with no allocation.
    pub(crate) fn proc_fd_path(
        fd: libc::c_int,
        out: &mut [u8; PROC_FD_PATH_LEN],
    ) -> Result<(), ()> {
        const PREFIX: &[u8] = b"/proc/self/fd/";
        if fd < 0 {
            return Err(());
        }
        out[..PREFIX.len()].copy_from_slice(PREFIX);
        let mut len = PREFIX.len();
        push_decimal(out, &mut len, fd as u32);
        out[len] = 0;
        Ok(())
    }

    /// Appends `n` in decimal to `buf` at `*len`, advancing `*len`. Allocation-free, for the
    /// post-`fork()` window; `buf` must have room for ten more bytes.
    fn push_decimal(buf: &mut [u8], len: &mut usize, mut n: u32) {
        let mut digits = [0u8; 10];
        let mut count = 0;
        loop {
            digits[count] = b'0' + (n % 10) as u8;
            count += 1;
            n /= 10;
            if n == 0 {
                break;
            }
        }
        while count > 0 {
            count -= 1;
            buf[*len] = digits[count];
            *len += 1;
        }
    }

    /// Opens the directory at `source` with `O_PATH`, and returns its descriptor only when its
    /// `(st_dev, st_ino)` is `identity`. A mismatch closes it and fails with `ESTALE`.
    ///
    /// # Safety
    /// Allocation-free; callable post-`fork()`, pre-exec.
    pub(crate) unsafe fn open_verified_dir(
        source: &CString,
        identity: (u64, u64),
    ) -> Result<libc::c_int, ()> {
        let fd = libc::open(
            source.as_ptr(),
            libc::O_PATH | libc::O_DIRECTORY | libc::O_CLOEXEC,
        );
        if fd < 0 {
            return Err(());
        }
        let mut stat: libc::stat = std::mem::zeroed();
        if libc::fstat(fd, &mut stat) != 0 {
            let errno = *libc::__errno_location();
            libc::close(fd);
            *libc::__errno_location() = errno;
            return Err(());
        }
        if (stat.st_dev as u64, stat.st_ino as u64) != identity {
            libc::close(fd);
            *libc::__errno_location() = libc::ESTALE;
            return Err(());
        }
        Ok(fd)
    }

    /// The mount flags a read-only remount of a bind has to repeat from the mount it was bound
    /// from, given that mount's `statvfs` `f_flag`: `nodev`, `noexec`, `nosuid` and the atime
    /// mode. A user namespace locks each of them on a mount it did not create, and a remount that
    /// drops one is refused with `EPERM`. No atime flag at all is `strictatime`, which is stated
    /// explicitly because a remount naming none defaults to `relatime`.
    pub(crate) fn remount_flags_kept(f_flags: libc::c_ulong) -> libc::c_ulong {
        let has = |flag: libc::c_ulong| f_flags & flag != 0;
        let mut kept: libc::c_ulong = 0;
        for (st, ms) in [
            (libc::ST_NODEV, libc::MS_NODEV),
            (libc::ST_NOEXEC, libc::MS_NOEXEC),
            (libc::ST_NOSUID, libc::MS_NOSUID),
            (libc::ST_NOATIME, libc::MS_NOATIME),
            (libc::ST_NODIRATIME, libc::MS_NODIRATIME),
            (libc::ST_RELATIME, libc::MS_RELATIME),
        ] {
            if has(st) {
                kept |= ms;
            }
        }
        if !has(libc::ST_NOATIME) && !has(libc::ST_RELATIME) {
            kept |= libc::MS_STRICTATIME;
        }
        kept
    }

    /// Lowers a [`ComposedRootPlan`] into the C-string form the child executes. Fails only on a
    /// path containing an interior NUL, which no real path does.
    pub(crate) fn build_sealed_root_spec(
        plan: &ComposedRootPlan,
    ) -> Result<SealedRootSpec, String> {
        // SAFETY: `getuid`/`getgid` take no arguments, dereference nothing and cannot fail.
        let (uid, gid) = unsafe { (libc::getuid(), libc::getgid()) };

        let mut steps = Vec::with_capacity(plan.steps.len());
        for RootStep { op, required } in &plan.steps {
            let (kind, label) = match op {
                RootOp::MkDir(path) => (
                    CStepKind::MkDir(cstr(path)?),
                    format!("mkdir {}", path.display()),
                ),
                RootOp::MkFile(path) => (
                    CStepKind::MkFile(cstr(path)?),
                    format!("create {}", path.display()),
                ),
                RootOp::Symlink { target, link } => (
                    CStepKind::Symlink {
                        target: cstr(target)?,
                        link: cstr(link)?,
                    },
                    format!("symlink {} -> {}", link.display(), target.display()),
                ),
                RootOp::Bind {
                    source,
                    target,
                    read_only,
                } => (
                    CStepKind::Bind {
                        source: cstr(source)?,
                        target: cstr(target)?,
                        read_only: *read_only,
                    },
                    format!(
                        "bind{} {} -> {}",
                        if *read_only { " (ro)" } else { "" },
                        source.display(),
                        target.display()
                    ),
                ),
                RootOp::BindVerified {
                    source,
                    target,
                    read_only,
                    identity,
                } => (
                    CStepKind::BindVerified {
                        source: cstr(source)?,
                        target: cstr(target)?,
                        read_only: *read_only,
                        identity: *identity,
                    },
                    format!(
                        "bind{} {} (dev:ino {}:{}) -> {}",
                        if *read_only { " (ro)" } else { "" },
                        source.display(),
                        identity.0,
                        identity.1,
                        target.display()
                    ),
                ),
                RootOp::Tmpfs { target, options } => (
                    CStepKind::Tmpfs {
                        target: cstr(target)?,
                        options: CString::new(*options).map_err(|error| error.to_string())?,
                    },
                    format!("tmpfs on {}", target.display()),
                ),
                RootOp::Proc { target } => (
                    CStepKind::Proc {
                        target: cstr(target)?,
                    },
                    format!("proc on {}", target.display()),
                ),
                RootOp::DevPts { target } => (
                    CStepKind::DevPts {
                        target: cstr(target)?,
                    },
                    format!("devpts on {}", target.display()),
                ),
                RootOp::RemountReadOnly(path) => (
                    CStepKind::RemountReadOnly(cstr(path)?),
                    format!("remount read-only {}", path.display()),
                ),
            };
            steps.push(CStep {
                kind,
                required: *required,
                label,
            });
        }

        Ok(SealedRootSpec {
            // Identity maps: the capsule keeps the uid it already had, so files it writes into the
            // bind-mounted workdir are owned by the real user rather than by `nobody`.
            uid_map: format!("{uid} {uid} 1\n").into_bytes(),
            gid_map: format!("{gid} {gid} 1\n").into_bytes(),
            base: cstr(&plan.base)?,
            old_root_full: cstr(&plan.base.join(OLD_ROOT_NAME))?,
            old_root_relative: CString::new(OLD_ROOT_NAME).map_err(|e| e.to_string())?,
            old_root_pivoted: CString::new(format!("/{OLD_ROOT_NAME}"))
                .map_err(|e| e.to_string())?,
            steps,
            workdir_in_root: cstr(&plan.workdir_in_root)?,
        })
    }

    fn cstr(path: &Path) -> Result<CString, String> {
        CString::new(path.as_os_str().as_bytes())
            .map_err(|_| format!("path contains an interior NUL: {}", path.display()))
    }

    // ------------------------------------------------------------ execution (child side)

    /// Builds the composed root and pivots onto it. Runs inside the forked child's `pre_exec`
    /// window, before fd hygiene, the capability drop, seccomp and Landlock — which is the only
    /// window in which `unshare`/`mount`/`pivot_root` are available at all, since
    /// `sandbox::SECCOMP_MUST_STAY_DENIED` denies all three to every process the filter covers.
    ///
    /// Allocation-free on the success path: every path was turned into a `CString` by
    /// [`build_sealed_root_spec`] before `fork()`.
    pub(crate) fn construct_composed_root(spec: &SealedRootSpec) -> Result<(), SealedRootFailure> {
        // SAFETY: every call below is a bare syscall over pointers into `spec`, which outlives
        // this function. No allocation, no locks, no reentrancy — the constraints of the
        // post-fork/pre-exec window.
        unsafe {
            // 1. The namespaces. `CLONE_NEWUSER` is what makes `CLONE_NEWNS` available without
            //    host root; asking for both in one call means the mount namespace is created with
            //    the new user namespace's credentials already in force.
            if libc::unshare(libc::CLONE_NEWUSER | libc::CLONE_NEWNS) != 0 {
                return Err(SealedRootFailure::stage(
                    "unshare(CLONE_NEWUSER|CLONE_NEWNS)",
                ));
            }

            // 2. Identity uid/gid maps. `setgroups=deny` first, which the kernel requires before
            //    an unprivileged process may write `gid_map`. A host without `setgroups` (pre-3.19)
            //    is tolerated; a failing `uid_map` is not, because running as the overflow uid
            //    would leave the capsule unable to write its own workdir.
            //
            //    Dumpable is flipped on for exactly these three writes and put back before the
            //    first mount: `mur` runs non-dumpable, and a non-dumpable task's `/proc/self/*`
            //    entries are owned by root, which makes its own `uid_map` unopenable. See
            //    [`make_dumpable_for_map_writes`].
            let previous_dumpable = make_dumpable_for_map_writes();
            let _ = write_file(c"/proc/self/setgroups", b"deny");
            let uid_mapped = write_file(c"/proc/self/uid_map", &spec.uid_map).is_ok();
            let gid_mapped = uid_mapped && write_file(c"/proc/self/gid_map", &spec.gid_map).is_ok();
            restore_dumpable(previous_dumpable);
            if !uid_mapped {
                return Err(SealedRootFailure::stage("write /proc/self/uid_map"));
            }
            if !gid_mapped {
                return Err(SealedRootFailure::stage("write /proc/self/gid_map"));
            }

            // 3. Private propagation, so none of what follows escapes back to the host's mounts.
            if libc::mount(
                std::ptr::null(),
                c"/".as_ptr(),
                std::ptr::null(),
                libc::MS_REC | libc::MS_PRIVATE,
                std::ptr::null(),
            ) != 0
            {
                return Err(SealedRootFailure::stage("mount --make-rprivate /"));
            }

            // 4. The new root: an empty tmpfs over the chosen base. It is a mount point, which is
            //    what `pivot_root(2)` requires of its new root.
            if libc::mount(
                c"tmpfs".as_ptr(),
                spec.base.as_ptr(),
                c"tmpfs".as_ptr(),
                libc::MS_NOSUID | libc::MS_NODEV,
                c"mode=0755".as_ptr() as *const libc::c_void,
            ) != 0
            {
                return Err(SealedRootFailure::stage(
                    "mount tmpfs for the composed root",
                ));
            }

            // 5. The plan.
            for (index, step) in spec.steps.iter().enumerate() {
                if execute_step(step).is_err() && step.required {
                    return Err(SealedRootFailure::step(index));
                }
            }

            // 6. Pivot. `chdir` into the new root first so `pivot_root` can be handed the pair of
            //    relative paths the syscall is happiest with.
            if libc::mkdir(spec.old_root_full.as_ptr(), 0o700) != 0 {
                return Err(SealedRootFailure::stage(
                    "mkdir the old-root parking directory",
                ));
            }
            if libc::chdir(spec.base.as_ptr()) != 0 {
                return Err(SealedRootFailure::stage("chdir into the composed root"));
            }
            if libc::syscall(
                libc::SYS_pivot_root,
                c".".as_ptr(),
                spec.old_root_relative.as_ptr(),
            ) != 0
            {
                return Err(SealedRootFailure::stage("pivot_root"));
            }
            if libc::chdir(c"/".as_ptr()) != 0 {
                return Err(SealedRootFailure::stage("chdir / after pivot_root"));
            }

            // 7. Detach and remove the old root. After this there is no pathname anywhere in this
            //    namespace that reaches the host filesystem.
            if libc::umount2(spec.old_root_pivoted.as_ptr(), libc::MNT_DETACH) != 0 {
                return Err(SealedRootFailure::stage("umount2 the old root"));
            }
            // Best-effort, deliberately: `MNT_DETACH` already removed the old root from this
            // namespace's tree, so a leftover empty directory here is cosmetic. Failing a launch
            // over it would trade a working capsule for nothing.
            libc::rmdir(spec.old_root_pivoted.as_ptr());

            // 8. Seal the root tmpfs itself. `MS_REMOUNT | MS_BIND` changes per-mount flags only,
            //    which is the well-defined way to make an existing mount read-only; the workdir
            //    bind underneath it keeps its own (writable) flags.
            if libc::mount(
                std::ptr::null(),
                c"/".as_ptr(),
                std::ptr::null(),
                libc::MS_REMOUNT
                    | libc::MS_BIND
                    | libc::MS_RDONLY
                    | libc::MS_NOSUID
                    | libc::MS_NODEV,
                std::ptr::null(),
            ) != 0
            {
                return Err(SealedRootFailure::stage(
                    "remount the composed root read-only",
                ));
            }

            // 9. Back into the workdir, at the same absolute path it had on the host. `Command`
            //    already `chdir`ed here before this closure ran, but that was in the old root.
            if libc::chdir(spec.workdir_in_root.as_ptr()) != 0 {
                return Err(SealedRootFailure::stage(
                    "chdir into the workdir inside the root",
                ));
            }
        }

        Ok(())
    }

    /// Executes one planned step. `Err(())` carries no detail — the caller pairs the index with
    /// the spec's pre-rendered label and the live errno.
    ///
    /// # Safety
    /// Must only be called from the post-`fork()`, pre-exec window: every branch is a bare
    /// syscall over `CStr` pointers borrowed from `step`, which the caller guarantees outlives
    /// the call, and none may allocate or take a lock that a sibling thread could hold frozen at
    /// `fork()` time.
    unsafe fn execute_step(step: &CStep) -> Result<(), ()> {
        match &step.kind {
            CStepKind::MkDir(path) => {
                if libc::mkdir(path.as_ptr(), 0o755) != 0
                    && *libc::__errno_location() != libc::EEXIST
                {
                    return Err(());
                }
            }
            CStepKind::MkFile(path) => {
                let fd = libc::open(
                    path.as_ptr(),
                    libc::O_CREAT | libc::O_WRONLY | libc::O_CLOEXEC,
                    libc::mode_t::from(0o644u16),
                );
                if fd < 0 {
                    return Err(());
                }
                libc::close(fd);
            }
            CStepKind::Symlink { target, link } => {
                if libc::symlink(target.as_ptr(), link.as_ptr()) != 0 {
                    return Err(());
                }
            }
            CStepKind::Bind {
                source,
                target,
                read_only,
            } => {
                if libc::mount(
                    source.as_ptr(),
                    target.as_ptr(),
                    std::ptr::null(),
                    libc::MS_BIND | libc::MS_REC,
                    std::ptr::null(),
                ) != 0
                {
                    return Err(());
                }
                if *read_only {
                    remount_read_only(target)?;
                }
            }
            CStepKind::BindVerified {
                source,
                target,
                read_only,
                identity,
            } => {
                let fd = open_verified_dir(source, *identity)?;
                let mut path = [0u8; PROC_FD_PATH_LEN];
                let bound = proc_fd_path(fd, &mut path).is_ok()
                    && libc::mount(
                        path.as_ptr() as *const libc::c_char,
                        target.as_ptr(),
                        std::ptr::null(),
                        libc::MS_BIND | libc::MS_REC,
                        std::ptr::null(),
                    ) == 0;
                // `close` may overwrite `errno`; the mount's is the one the failure reports.
                let errno = *libc::__errno_location();
                libc::close(fd);
                if !bound {
                    *libc::__errno_location() = errno;
                    return Err(());
                }
                if *read_only {
                    remount_read_only(target)?;
                }
            }
            CStepKind::Tmpfs { target, options } => {
                if libc::mount(
                    c"tmpfs".as_ptr(),
                    target.as_ptr(),
                    c"tmpfs".as_ptr(),
                    libc::MS_NOSUID,
                    options.as_ptr() as *const libc::c_void,
                ) != 0
                {
                    return Err(());
                }
            }
            CStepKind::Proc { target } => {
                // Try each `hidepid` spelling in turn, then a recursive bind of the host's
                // `/proc` — see `PROC_HIDEPID_OPTIONS` for the bind's rationale and exposure.
                for option in PROC_HIDEPID_OPTIONS {
                    let mut buffer = [0u8; 32];
                    let bytes = option.as_bytes();
                    if bytes.len() >= buffer.len() {
                        continue;
                    }
                    buffer[..bytes.len()].copy_from_slice(bytes);
                    let data = if bytes.is_empty() {
                        std::ptr::null()
                    } else {
                        buffer.as_ptr() as *const libc::c_void
                    };
                    if libc::mount(
                        c"proc".as_ptr(),
                        target.as_ptr(),
                        c"proc".as_ptr(),
                        libc::MS_NOSUID | libc::MS_NODEV | libc::MS_NOEXEC,
                        data,
                    ) == 0
                    {
                        return Ok(());
                    }
                }

                // Last resort: bind the host's `/proc` in. `mount -t proc` needs `CAP_SYS_ADMIN`
                // over the *PID* namespace's user namespace, which an unprivileged user namespace
                // never has, so on a bare host every spelling above fails with `EPERM` and this is
                // the branch that actually runs. A bind of an existing mount needs no such
                // privilege. Recursive so the `/proc` submounts a modern host carries come with it
                // rather than leaving empty stubs.
                if libc::mount(
                    c"/proc".as_ptr(),
                    target.as_ptr(),
                    std::ptr::null(),
                    libc::MS_BIND | libc::MS_REC,
                    std::ptr::null(),
                ) != 0
                {
                    return Err(());
                }
                // Best-effort hardening of the bind. `MS_REMOUNT | MS_BIND` changes per-mount
                // flags only — unlike `hidepid`, which is a superblock option and stays refused —
                // so this narrows what the bind can do without touching the host's own `/proc`.
                // Not fatal: a `/proc` that is present and merely unhardened beats no `/proc`.
                libc::mount(
                    std::ptr::null(),
                    target.as_ptr(),
                    std::ptr::null(),
                    libc::MS_REMOUNT
                        | libc::MS_BIND
                        | libc::MS_NOSUID
                        | libc::MS_NODEV
                        | libc::MS_NOEXEC,
                    std::ptr::null(),
                );
                return Ok(());
            }
            CStepKind::DevPts { target } => {
                if libc::mount(
                    c"devpts".as_ptr(),
                    target.as_ptr(),
                    c"devpts".as_ptr(),
                    libc::MS_NOSUID | libc::MS_NOEXEC,
                    c"newinstance,ptmxmode=0666,mode=0620".as_ptr() as *const libc::c_void,
                ) != 0
                {
                    return Err(());
                }
            }
            CStepKind::RemountReadOnly(target) => {
                if libc::mount(
                    std::ptr::null(),
                    target.as_ptr(),
                    std::ptr::null(),
                    libc::MS_REMOUNT | libc::MS_BIND | libc::MS_RDONLY | libc::MS_NOSUID,
                    std::ptr::null(),
                ) != 0
                {
                    return Err(());
                }
            }
        }
        Ok(())
    }

    /// `open`/`write`/`close` over a static path, with no allocation — the `/proc/self/*_map`
    /// writes are the only file I/O the composed-root construction performs.
    ///
    /// # Safety
    /// Must only be called from the post-`fork()`, pre-exec window, for the same reason as
    /// [`execute_step`]: `path` and `data` are borrowed for the duration of the call only, and the
    /// raw `open`/`write`/`close` sequence must stay allocation- and lock-free in that window.
    unsafe fn write_file(path: &std::ffi::CStr, data: &[u8]) -> Result<(), ()> {
        let fd = libc::open(path.as_ptr(), libc::O_WRONLY | libc::O_CLOEXEC);
        if fd < 0 {
            return Err(());
        }
        let mut written = 0usize;
        while written < data.len() {
            let n = libc::write(
                fd,
                data[written..].as_ptr() as *const libc::c_void,
                data.len() - written,
            );
            if n < 0 {
                if *libc::__errno_location() == libc::EINTR {
                    continue;
                }
                libc::close(fd);
                return Err(());
            }
            if n == 0 {
                break;
            }
            written += n as usize;
        }
        libc::close(fd);
        if written == data.len() {
            Ok(())
        } else {
            Err(())
        }
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::HashMap;

    struct FakeHost(HashMap<PathBuf, PathKind>);

    impl FakeHost {
        fn usrmerge() -> Self {
            let mut map = HashMap::new();
            map.insert(PathBuf::from("/usr"), PathKind::Dir);
            map.insert(
                PathBuf::from("/bin"),
                PathKind::Symlink(PathBuf::from("usr/bin")),
            );
            map.insert(
                PathBuf::from("/sbin"),
                PathKind::Symlink(PathBuf::from("usr/sbin")),
            );
            map.insert(
                PathBuf::from("/lib"),
                PathKind::Symlink(PathBuf::from("usr/lib")),
            );
            map.insert(
                PathBuf::from("/lib64"),
                PathKind::Symlink(PathBuf::from("usr/lib64")),
            );
            map.insert(PathBuf::from("/etc/ld.so.cache"), PathKind::File);
            map.insert(PathBuf::from("/etc/ssl"), PathKind::Dir);
            // Present on the fake host precisely so the synthesis tests below can show that having
            // them is not what decides whether they are mirrored.
            map.insert(PathBuf::from("/etc/passwd"), PathKind::File);
            map.insert(PathBuf::from("/etc/group"), PathKind::File);
            for device in SEALED_DEVICE_NODES {
                map.insert(PathBuf::from(device.path), PathKind::File);
            }
            Self(map)
        }
    }

    impl HostLayout for FakeHost {
        fn kind(&self, path: &Path) -> Option<PathKind> {
            self.0.get(path).cloned()
        }
    }

    /// The session's temporary folder in its records directory, and the identity of the
    /// descriptor its records root holds on it.
    const TEST_TMP_SOURCE: &str = "/home/u/.murmur/records/home/u/w/ses_plan_tests/tmp";
    const TEST_TMP_IDENTITY: (u64, u64) = (64, 7);

    /// The session's `sealed-etc/`, beside its runtime directory.
    const TEST_ETC_DIR: &str = "/home/u/.murmur/sessions/ses_plan_tests/sealed-etc";

    /// The runtime directory and the identity of its root's descriptor.
    const TEST_RUNTIME_DIR: &str = "/home/u/.murmur/sessions/ses_plan_tests/runtime";
    const TEST_RUNTIME_IDENTITY: (u64, u64) = (64, 8);

    /// The runtime-provided bind sources, with no runtime directory: the planner's steps 1–6.
    fn test_sources() -> SealedBindSources<'static> {
        SealedBindSources {
            tmp_store: (Path::new(TEST_TMP_SOURCE), TEST_TMP_IDENTITY),
            etc_dir: Path::new(TEST_ETC_DIR),
            runtime_dir: None,
            records_dir: None,
            home_dir: None,
            state_dir: None,
            host_paths: &[],
            host_path_read_only: &[],
        }
    }

    fn plan_for(workdir: &str) -> ComposedRootPlan {
        plan_composed_root(
            Path::new(workdir),
            &test_sources(),
            Path::new("/tmp"),
            &[],
            &[],
            &FakeHost::usrmerge(),
        )
    }

    fn ops(plan: &ComposedRootPlan) -> Vec<RootOp> {
        plan.steps.iter().map(|step| step.op.clone()).collect()
    }

    #[test]
    fn base_selection_skips_any_ancestor_of_the_workdir() {
        // The default case: a workdir under the user's home, so `/tmp` is free.
        assert_eq!(
            choose_root_base(
                Path::new("/home/u/w"),
                &[],
                SEALED_ROOT_BASE_CANDIDATES,
                |_| { true }
            ),
            Some(PathBuf::from("/tmp"))
        );
        // A workdir under /tmp — overmounting /tmp would hide the workdir before it can be bound.
        assert_eq!(
            choose_root_base(
                Path::new("/tmp/session/w"),
                &[],
                SEALED_ROOT_BASE_CANDIDATES,
                |_| true
            ),
            Some(PathBuf::from("/run"))
        );
        // A host missing the first two candidates falls through to the third.
        assert_eq!(
            choose_root_base(
                Path::new("/home/u/w"),
                &[],
                SEALED_ROOT_BASE_CANDIDATES,
                |path| { path != Path::new("/tmp") && path != Path::new("/run") }
            ),
            Some(PathBuf::from("/var/tmp"))
        );
        assert_eq!(
            choose_root_base(
                Path::new("/home/u/w"),
                &[],
                SEALED_ROOT_BASE_CANDIDATES,
                |_| { false }
            ),
            None
        );
    }

    #[test]
    fn base_selection_skips_any_ancestor_of_a_host_path() {
        assert_eq!(
            choose_root_base(
                Path::new("/home/u/w"),
                &[PathBuf::from("/run/media/data")],
                &["/run", "/var/tmp"],
                |_| true
            ),
            Some(PathBuf::from("/var/tmp"))
        );
        assert_eq!(
            choose_root_base(
                Path::new("/home/u/w"),
                &[PathBuf::from("/var/tmp")],
                &["/var/tmp", "/mnt"],
                |_| true
            ),
            Some(PathBuf::from("/mnt"))
        );
        // No host path, no restriction.
        assert_eq!(
            choose_root_base(Path::new("/home/u/w"), &[], &["/run"], |_| true),
            Some(PathBuf::from("/run"))
        );
    }

    fn host_bind(path: &str, read_only: bool, ino: u64) -> HostPathBind {
        HostPathBind {
            path: PathBuf::from(path),
            identity: (70, ino),
            read_only,
        }
    }

    fn host_path_plan(
        host_paths: &[HostPathBind],
        read_only: &[PathBuf],
        extra_read_only: &[PathBuf],
    ) -> ComposedRootPlan {
        let mut host = FakeHost::usrmerge();
        for path in extra_read_only {
            host.0.insert(path.clone(), PathKind::Dir);
        }
        plan_composed_root(
            Path::new("/home/u/w"),
            &SealedBindSources {
                runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
                home_dir: Some((Path::new("/home/u/.murmur/records/x/home"), (64, 9))),
                state_dir: Some((Path::new("/home/u/.murmur/state/s"), (64, 10))),
                host_paths,
                host_path_read_only: read_only,
                ..test_sources()
            },
            Path::new("/tmp"),
            extra_read_only,
            &[PathBuf::from("/opt/staged")],
            &host,
        )
    }

    fn position(plan: &ComposedRootPlan, matches: impl Fn(&RootStep) -> bool) -> usize {
        plan.steps
            .iter()
            .position(matches)
            .expect("the plan holds the step")
    }

    #[test]
    fn host_paths_are_planned_first_shallowest_first_verified_and_required() {
        let host_paths = [
            host_bind("/home/u/data/inner", false, 2),
            host_bind("/home/u/data", true, 1),
        ];
        let plan = host_path_plan(&host_paths, &[], &[]);
        let outer = position(&plan, |step| {
            step.op
                == RootOp::BindVerified {
                    source: PathBuf::from("/home/u/data"),
                    target: PathBuf::from("/tmp/home/u/data"),
                    read_only: true,
                    identity: (70, 1),
                }
                && step.required
        });
        let inner = position(&plan, |step| {
            step.op
                == RootOp::BindVerified {
                    source: PathBuf::from("/home/u/data/inner"),
                    target: PathBuf::from("/tmp/home/u/data/inner"),
                    read_only: false,
                    identity: (70, 2),
                }
                && step.required
        });
        assert!(outer < inner, "the shallower entry is bound first");
        // Before every other step that binds anything: staged runtime, runtime tree, /tmp, the
        // workdir, runtime directory, records, home and store.
        let first_other_bind = position(&plan, |step| {
            matches!(
                &step.op,
                RootOp::Bind { .. }
                    | RootOp::BindVerified { .. }
                    | RootOp::Symlink { .. }
                    | RootOp::Tmpfs { .. }
            ) && !matches!(&step.op, RootOp::BindVerified { source, .. } if source.starts_with("/home/u/data"))
        });
        assert!(inner < first_other_bind);
        for source in [
            "/opt/staged",
            "/home/u/w",
            TEST_RUNTIME_DIR,
            "/home/u/.murmur/records/x/home",
            "/home/u/.murmur/state/s",
        ] {
            let at = position(&plan, |step| match &step.op {
                RootOp::Bind { source: s, .. } | RootOp::BindVerified { source: s, .. } => {
                    s == Path::new(source)
                }
                _ => false,
            });
            assert!(inner < at, "{source} is bound after every host path");
        }
    }

    #[test]
    fn an_execute_root_inside_a_read_write_host_path_is_bound_read_only_and_required_on_top() {
        let host_paths = [host_bind("/home/u/tools", false, 1)];
        let plan = host_path_plan(
            &host_paths,
            &[PathBuf::from("/home/u/tools/bin")],
            &[PathBuf::from("/home/u/tools/bin")],
        );
        let host = position(
            &plan,
            |step| matches!(&step.op, RootOp::BindVerified { source, read_only: false, .. } if source == Path::new("/home/u/tools")),
        );
        let protect = position(&plan, |step| {
            step.op
                == RootOp::Bind {
                    source: PathBuf::from("/home/u/tools/bin"),
                    target: PathBuf::from("/tmp/home/u/tools/bin"),
                    read_only: true,
                }
                && step.required
        });
        assert!(host < protect);
        // The optional mirror of the same directory adds nothing: the required bind owns it.
        let binds_of_bin = plan
            .steps
            .iter()
            .filter(|step| matches!(&step.op, RootOp::Bind { source, .. } if source == Path::new("/home/u/tools/bin")))
            .count();
        assert_eq!(binds_of_bin, 1);
    }

    #[test]
    fn a_host_path_inside_a_later_bound_directory_is_bound_again_on_top() {
        let host_paths = [host_bind("/home/u/py/site/data", true, 1)];
        let plan = host_path_plan(&host_paths, &[], &[PathBuf::from("/home/u/py")]);
        let cover = position(
            &plan,
            |step| matches!(&step.op, RootOp::Bind { source, .. } if source == Path::new("/home/u/py")),
        );
        let binds: Vec<usize> = plan
            .steps
            .iter()
            .enumerate()
            .filter(|(_, step)| matches!(&step.op, RootOp::BindVerified { source, .. } if source == Path::new("/home/u/py/site/data")))
            .map(|(at, _)| at)
            .collect();
        assert_eq!(binds.len(), 2, "{:#?}", plan.steps);
        assert!(binds[0] < cover && cover < binds[1]);
    }

    #[test]
    fn no_host_paths_add_no_step() {
        let without = host_path_plan(&[], &[], &[]);
        let baseline = plan_composed_root(
            Path::new("/home/u/w"),
            &SealedBindSources {
                runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
                home_dir: Some((Path::new("/home/u/.murmur/records/x/home"), (64, 9))),
                state_dir: Some((Path::new("/home/u/.murmur/state/s"), (64, 10))),
                ..test_sources()
            },
            Path::new("/tmp"),
            &[],
            &[PathBuf::from("/opt/staged")],
            &FakeHost::usrmerge(),
        );
        assert_eq!(without, baseline);
    }

    #[test]
    fn usrmerge_symlinks_are_reproduced_rather_than_bound_through() {
        let plan = plan_for("/home/u/w");
        assert!(ops(&plan).contains(&RootOp::Symlink {
            target: PathBuf::from("usr/bin"),
            link: PathBuf::from("/tmp/bin"),
        }));
        assert!(ops(&plan).contains(&RootOp::Bind {
            source: PathBuf::from("/usr"),
            target: PathBuf::from("/tmp/usr"),
            read_only: true,
        }));
        // `/lib32` and `/libx32` are absent from this fake host, so they contribute nothing.
        assert!(!ops(&plan)
            .iter()
            .any(|op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/lib32"))));
    }

    #[test]
    fn every_runtime_and_etc_bind_is_read_only_and_only_the_workdir_is_not() {
        let plan = plan_for("/home/u/w");
        let writable: Vec<&PathBuf> = plan
            .steps
            .iter()
            .filter_map(|step| match &step.op {
                RootOp::Bind {
                    source,
                    read_only: false,
                    ..
                } => Some(source),
                _ => None,
            })
            .collect();

        // The workdir and the /dev nodes (whose write side is the device driver's, not the
        // filesystem's) — and nothing else. /tmp is a verified bind, not among these.
        for source in &writable {
            assert!(
                source.starts_with("/home/u/w") || source.starts_with("/dev/"),
                "unexpected writable bind of {}",
                source.display()
            );
        }
        assert!(writable.contains(&&PathBuf::from("/home/u/w")));
    }

    #[test]
    fn etc_is_an_allowlist_so_shadow_is_absent_by_construction() {
        let plan = plan_for("/home/u/w");
        let bound: Vec<String> = plan
            .steps
            .iter()
            .filter_map(|step| match &step.op {
                RootOp::Bind { source, .. } => Some(source.display().to_string()),
                _ => None,
            })
            .collect();

        assert!(bound.iter().any(|path| path == "/etc/ld.so.cache"));
        assert!(!bound.iter().any(|path| path == "/etc"));
        assert!(!bound.iter().any(|path| path.starts_with("/etc/shadow")));
        assert!(!bound.iter().any(|path| path.starts_with("/root")));
        assert!(!bound.iter().any(|path| path.starts_with("/home/u/.ssh")));
    }

    /// The host has both files (see `FakeHost::usrmerge`) and the plan mirrors neither: their
    /// targets are bound from the files the parent writes into `sealed-etc/` instead.
    #[test]
    fn passwd_and_group_are_bound_from_the_staging_files_never_from_the_host() {
        let plan = plan_for("/home/u/w");

        for host_path in ["/etc/passwd", "/etc/group"] {
            assert!(
                !plan.steps.iter().any(|step| matches!(
                    &step.op,
                    RootOp::Bind { source, .. } if source == Path::new(host_path)
                )),
                "the host's {host_path} carries every account on the machine and must not be \
                 bound into a capsule",
            );
        }

        for (file, target) in [
            (SyntheticEtcFile::Passwd, "/tmp/etc/passwd"),
            (SyntheticEtcFile::Group, "/tmp/etc/group"),
        ] {
            let expected = RootStep {
                op: RootOp::Bind {
                    source: synthetic_etc_source(Path::new(TEST_ETC_DIR), file),
                    target: PathBuf::from(target),
                    read_only: true,
                },
                // Not the host-mirrored entries' `required: false`: the parent wrote this file, so
                // a failed bind is a broken runtime, not a distribution that lacks the path.
                required: true,
            };
            assert!(
                plan.steps.contains(&expected),
                "expected {expected:?} in {:?}",
                plan.steps,
            );
            // And the mountpoint the bind lands on, since a bind needs an existing target.
            assert!(ops(&plan).contains(&RootOp::MkFile(PathBuf::from(target))));
        }

        // Every other /etc entry is still the host's own file, untouched by this.
        assert!(ops(&plan).contains(&RootOp::Bind {
            source: PathBuf::from("/etc/ld.so.cache"),
            target: PathBuf::from("/tmp/etc/ld.so.cache"),
            read_only: true,
        }));
    }

    /// Two lines, and the `pw_dir` field is the synthetic `HOME` verbatim — the property a capsule
    /// observes as `pwd.getpwuid(os.getuid()).pw_dir == os.environ["HOME"]`.
    #[test]
    fn the_synthetic_databases_name_root_and_the_capsules_own_id_and_nothing_else() {
        let identity = SealedAccountIdentity {
            uid: 1000,
            gid: 1000,
            home: "/r/ses/home",
        };

        let passwd = SyntheticEtcFile::Passwd.render(&identity).unwrap();
        assert_eq!(
            passwd,
            "root:x:0:0:root:/root:/bin/sh\n\
             capsule:x:1000:1000:Murmur capsule:/r/ses/home:/bin/sh\n"
        );
        let group = SyntheticEtcFile::Group.render(&identity).unwrap();
        assert_eq!(group, "root:x:0:\ncapsule:x:1000:\n");

        for (file, contents) in [
            (SyntheticEtcFile::Passwd, &passwd),
            (SyntheticEtcFile::Group, &group),
        ] {
            assert_eq!(
                contents.lines().count(),
                2,
                "{file:?} must stay a two-line database"
            );
            assert!(
                contents.ends_with('\n'),
                "{file:?} must end its last record with a newline"
            );
        }
    }

    /// A capsule already running as uid 0 gets *one* passwd line, not a second `root` — two entries
    /// for uid 0 would leave `getpwuid(0)` resolving to whichever the parser saw first.
    #[test]
    fn a_root_capsule_gets_a_single_passwd_line_carrying_its_own_home() {
        let identity = SealedAccountIdentity {
            uid: 0,
            gid: 0,
            home: "/r/ses/home",
        };

        assert_eq!(
            SyntheticEtcFile::Passwd.render(&identity).unwrap(),
            "root:x:0:0:root:/r/ses/home:/bin/sh\n"
        );
        assert_eq!(
            SyntheticEtcFile::Group.render(&identity).unwrap(),
            "root:x:0:\n"
        );
    }

    /// `passwd(5)` has no escaping, so a home path containing the field separator would silently
    /// shift every field after it — a `pw_shell` of half a path, or worse.
    #[test]
    fn a_home_path_that_cannot_be_spelled_in_a_passwd_field_is_refused() {
        for home in ["/w/od:d", "/w/two\nlines"] {
            let identity = SealedAccountIdentity {
                uid: 1000,
                gid: 1000,
                home,
            };
            let error = SyntheticEtcFile::Passwd.render(&identity).unwrap_err();
            assert!(
                error.contains(home),
                "the message must name the offending path: {error}"
            );
        }
    }

    /// Step 7: the runtime directory, read-only and required, at its own absolute path, from the
    /// descriptor the parent took, after the workdir's own bind.
    #[test]
    fn the_runtime_directory_is_bound_read_only_and_required_after_the_workdir() {
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &SealedBindSources {
                runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
                ..test_sources()
            },
            Path::new("/tmp"),
            &[],
            &[],
            &FakeHost::usrmerge(),
        );
        let bind = RootOp::BindVerified {
            source: PathBuf::from(TEST_RUNTIME_DIR),
            target: PathBuf::from("/tmp").join(TEST_RUNTIME_DIR.trim_start_matches('/')),
            read_only: true,
            identity: TEST_RUNTIME_IDENTITY,
        };
        let at = plan
            .steps
            .iter()
            .position(|step| step.op == bind)
            .expect("the runtime directory is bound");
        assert!(
            plan.steps[at].required,
            "a missing runtime directory ends the session"
        );
        let workdir = plan
            .steps
            .iter()
            .position(|step| {
                matches!(&step.op, RootOp::Bind { source, .. } if source == Path::new("/home/u/w"))
            })
            .unwrap();
        assert!(
            workdir < at,
            "the runtime directory is planned after the workdir"
        );
        assert_eq!(at, plan.steps.len() - 1);
        assert!(plan_for("/home/u/w").steps.iter().all(|step| !matches!(
            &step.op,
            RootOp::BindVerified { source, .. } if source == Path::new(TEST_RUNTIME_DIR)
        )));
    }

    /// The records directory is bound read-only and required at its own path, after the runtime
    /// directory, so the deeper bind survives a records directory beneath either.
    #[test]
    fn the_records_directory_is_bound_read_only_and_required_after_the_runtime_directory() {
        const TEST_RECORDS_DIR: &str = "/home/u/.murmur/records/home/u/w/ses_x";
        const TEST_RECORDS_IDENTITY: (u64, u64) = (7, 77);
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &SealedBindSources {
                runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
                records_dir: Some((Path::new(TEST_RECORDS_DIR), TEST_RECORDS_IDENTITY)),
                ..test_sources()
            },
            Path::new("/tmp"),
            &[],
            &[],
            &FakeHost::usrmerge(),
        );
        let position = |source: &str| {
            plan.steps
                .iter()
                .position(|step| {
                    matches!(&step.op, RootOp::BindVerified { source: s, read_only: true, .. } if s == Path::new(source))
                })
                .unwrap_or_else(|| panic!("{source} is bound read-only"))
        };
        let records = position(TEST_RECORDS_DIR);
        assert!(
            plan.steps[records].required,
            "a missing records directory ends the session"
        );
        assert!(position(TEST_RUNTIME_DIR) < records);
        assert_eq!(
            plan.steps[records].op,
            RootOp::BindVerified {
                source: PathBuf::from(TEST_RECORDS_DIR),
                target: PathBuf::from("/tmp").join(TEST_RECORDS_DIR.trim_start_matches('/')),
                read_only: true,
                identity: TEST_RECORDS_IDENTITY,
            }
        );
    }

    /// The home folder is bound read-write, verified and required at its own path, after the
    /// records directory's read-only bind it lies beneath, so the deeper, writable bind survives.
    #[test]
    fn home_folder_is_bound_read_write_and_verified_after_the_records_directory() {
        const TEST_RECORDS_DIR: &str = "/home/u/.murmur/records/home/u/w/ses_x";
        const TEST_RECORDS_IDENTITY: (u64, u64) = (7, 77);
        const TEST_HOME_DIR: &str = "/home/u/.murmur/records/home/u/w/ses_x/home";
        const TEST_HOME_IDENTITY: (u64, u64) = (7, 78);
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &SealedBindSources {
                runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
                records_dir: Some((Path::new(TEST_RECORDS_DIR), TEST_RECORDS_IDENTITY)),
                home_dir: Some((Path::new(TEST_HOME_DIR), TEST_HOME_IDENTITY)),
                ..test_sources()
            },
            Path::new("/tmp"),
            &[],
            &[],
            &FakeHost::usrmerge(),
        );
        let position = |source: &str| {
            plan.steps
                .iter()
                .position(|step| {
                    matches!(&step.op, RootOp::BindVerified { source: s, .. } if s == Path::new(source))
                })
                .unwrap_or_else(|| panic!("{source} is bound"))
        };
        let records = position(TEST_RECORDS_DIR);
        let home = position(TEST_HOME_DIR);
        assert!(records < home, "the home bind comes after the records bind");
        assert!(
            plan.steps[home].required,
            "a missing home folder ends the session"
        );
        assert_eq!(
            plan.steps[home].op,
            RootOp::BindVerified {
                source: PathBuf::from(TEST_HOME_DIR),
                target: PathBuf::from("/tmp").join(TEST_HOME_DIR.trim_start_matches('/')),
                read_only: false,
                identity: TEST_HOME_IDENTITY,
            }
        );
        assert!(matches!(
            plan.steps[records].op,
            RootOp::BindVerified {
                read_only: true,
                ..
            }
        ));
    }

    /// A native tool's store is bound read-write, verified and required at its own absolute
    /// path, after the home folder; a spawn with no store binds nothing there.
    #[test]
    fn a_native_tools_store_is_bound_read_write_verified_and_required_at_its_own_path() {
        const TEST_HOME_DIR: &str = "/home/u/.murmur/records/home/u/w/ses_x/home";
        const TEST_STATE_DIR: &str = "/home/u/.murmur/state/config-capsule";
        const TEST_STATE_IDENTITY: (u64, u64) = (7, 79);
        let sources = SealedBindSources {
            runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
            home_dir: Some((Path::new(TEST_HOME_DIR), (7, 78))),
            ..test_sources()
        };
        let plan = |sources: &SealedBindSources<'_>| {
            plan_composed_root(
                Path::new("/home/u/w"),
                sources,
                Path::new("/tmp"),
                &[],
                &[],
                &FakeHost::usrmerge(),
            )
        };
        let with_store = plan(&SealedBindSources {
            state_dir: Some((Path::new(TEST_STATE_DIR), TEST_STATE_IDENTITY)),
            ..sources
        });
        let position = |plan: &ComposedRootPlan, source: &str| {
            plan.steps.iter().position(|step| {
                matches!(&step.op, RootOp::BindVerified { source: s, .. } if s == Path::new(source))
            })
        };
        let home = position(&with_store, TEST_HOME_DIR).expect("the home folder is bound");
        let store = position(&with_store, TEST_STATE_DIR).expect("the store is bound");
        assert!(home < store, "the store bind comes after the home bind");
        assert!(
            with_store.steps[store].required,
            "a store that cannot be bound ends the session"
        );
        let target = PathBuf::from("/tmp").join(TEST_STATE_DIR.trim_start_matches('/'));
        assert_eq!(
            with_store.steps[store].op,
            RootOp::BindVerified {
                source: PathBuf::from(TEST_STATE_DIR),
                target: target.clone(),
                read_only: false,
                identity: TEST_STATE_IDENTITY,
            }
        );
        assert!(
            with_store.steps[..store]
                .iter()
                .any(|step| step.op == RootOp::MkDir(target.clone())),
            "the store's mount point is made before its bind"
        );
        assert_eq!(position(&plan(&sources), TEST_STATE_DIR), None);
    }

    /// A delegating session's root binds nothing beneath the directory its delegated children keep
    /// their records in, `records_group(<records dir>/children)`: the parent's runtime reads a
    /// child's answer and hands it over, so no command of the parent's reaches a child's records.
    #[test]
    fn nothing_is_bound_beneath_the_childrens_records() {
        const TEST_RECORDS_DIR: &str = "/home/u/.murmur/records/home/u/w/ses_x";
        const TEST_HOME_DIR: &str = "/home/u/.murmur/records/home/u/w/ses_x/home";
        const TEST_CHILDREN_RECORDS: &str =
            "/home/u/.murmur/records/home/u/.murmur/records/home/u/w/ses_x/children";
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &SealedBindSources {
                runtime_dir: Some((Path::new(TEST_RUNTIME_DIR), TEST_RUNTIME_IDENTITY)),
                records_dir: Some((Path::new(TEST_RECORDS_DIR), (7, 77))),
                home_dir: Some((Path::new(TEST_HOME_DIR), (7, 78))),
                ..test_sources()
            },
            Path::new("/tmp"),
            &[],
            &[],
            &FakeHost::usrmerge(),
        );
        let bound: Vec<&Path> = plan
            .steps
            .iter()
            .filter_map(|step| match &step.op {
                RootOp::BindVerified { source, .. } | RootOp::Bind { source, .. } => {
                    Some(source.as_path())
                }
                _ => None,
            })
            .collect();
        assert!(
            bound.contains(&Path::new(TEST_RECORDS_DIR)),
            "the plan binds the session's own records: {bound:?}"
        );
        for source in &bound {
            assert!(
                !source.starts_with(TEST_CHILDREN_RECORDS)
                    && !Path::new(TEST_CHILDREN_RECORDS).starts_with(source),
                "{} is bound, which reaches the children's records",
                source.display()
            );
        }
    }

    /// The remount keeps every flag a user namespace locks on the source mount.
    #[test]
    #[cfg(target_os = "linux")]
    fn a_read_only_remount_keeps_the_source_mounts_locked_flags() {
        let kept = linux::remount_flags_kept(
            libc::ST_NODEV | libc::ST_NOEXEC | libc::ST_NOSUID | libc::ST_RELATIME,
        );
        for flag in [
            libc::MS_NODEV,
            libc::MS_NOEXEC,
            libc::MS_NOSUID,
            libc::MS_RELATIME,
        ] {
            assert_ne!(kept & flag, 0, "{flag:#x}");
        }
        assert_eq!(kept & libc::MS_STRICTATIME, 0);
        let strict = linux::remount_flags_kept(0);
        assert_ne!(strict & libc::MS_STRICTATIME, 0);
        assert_eq!(strict & (libc::MS_NODEV | libc::MS_NOEXEC), 0);
    }

    #[test]
    fn the_workdir_keeps_its_absolute_path_inside_the_root() {
        let plan = plan_for("/home/u/.murmur/sessions/abc");
        assert_eq!(
            plan.workdir_in_root,
            PathBuf::from("/home/u/.murmur/sessions/abc")
        );
        assert!(ops(&plan).contains(&RootOp::Bind {
            source: PathBuf::from("/home/u/.murmur/sessions/abc"),
            target: PathBuf::from("/tmp/home/u/.murmur/sessions/abc"),
            read_only: false,
        }));
        // Every parent component is created before the bind, exactly once.
        let mkdirs: Vec<&RootOp> = plan
            .steps
            .iter()
            .map(|step| &step.op)
            .filter(|op| matches!(op, RootOp::MkDir(path) if path.starts_with("/tmp/home")))
            .collect();
        assert_eq!(mkdirs.len(), 5, "got {mkdirs:?}");
    }

    #[test]
    fn tmp_is_bound_only_when_it_is_the_directory_the_parent_holds() {
        let plan = plan_for("/home/u/w");
        assert!(ops(&plan).contains(&RootOp::BindVerified {
            source: PathBuf::from(TEST_TMP_SOURCE),
            target: PathBuf::from("/tmp/tmp"),
            read_only: false,
            identity: TEST_TMP_IDENTITY,
        }));
        // Never an unverified bind of a path a capsule's command could have replaced with a link.
        assert!(!ops(&plan).iter().any(|op| matches!(
            op,
            RootOp::Bind { source, .. } if source == Path::new(TEST_TMP_SOURCE)
        )));
    }

    /// The child's check: the directory reached is the one whose identity the parent recorded,
    /// or nothing is bound. A link swapped in after the parent's descriptor was taken reaches
    /// another directory, and is refused with `ESTALE`.
    #[test]
    #[cfg(target_os = "linux")]
    #[allow(unsafe_code)]
    fn a_verified_bind_source_refuses_a_directory_swapped_in_after_the_descriptor() {
        use std::os::unix::fs::MetadataExt;

        let dir = tempfile::tempdir().unwrap();
        let store = dir.path().join("tmp");
        let outside = dir.path().join("outside");
        std::fs::create_dir_all(&store).unwrap();
        std::fs::create_dir_all(&outside).unwrap();
        let meta = std::fs::metadata(&store).unwrap();
        let identity = (meta.dev(), meta.ino());
        let source = std::ffi::CString::new(store.to_str().unwrap()).unwrap();

        // SAFETY: open, fstat and close on a path this test owns; nothing else holds the fd.
        let fd = unsafe { linux::open_verified_dir(&source, identity) }.expect("same directory");
        let mut path = [0u8; 32];
        linux::proc_fd_path(fd, &mut path).unwrap();
        let resolved = std::fs::canonicalize(
            std::ffi::CStr::from_bytes_until_nul(&path)
                .unwrap()
                .to_str()
                .unwrap(),
        )
        .unwrap();
        assert_eq!(resolved, store.canonicalize().unwrap());
        // SAFETY: `fd` was returned open by `open_verified_dir` and is closed once.
        unsafe { libc::close(fd) };

        std::fs::rename(&store, dir.path().join("moved")).unwrap();
        std::os::unix::fs::symlink(&outside, &store).unwrap();
        // SAFETY: `source` is a NUL-terminated path this test owns; on the identity mismatch it
        // reports, `open_verified_dir` has already closed the descriptor it opened.
        assert!(unsafe { linux::open_verified_dir(&source, identity) }.is_err());
        assert_eq!(
            std::io::Error::last_os_error().raw_os_error(),
            Some(libc::ESTALE)
        );
    }

    /// The composed root's own `/tmp` must not land on top of the path leading to the workdir.
    ///
    /// This is `mur run`'s default layout, not a curiosity: with no `--workdir`, the session
    /// workdir is a temporary directory, which is under `/tmp`. Planned in the other order, the
    /// `/tmp` bind hid the workdir bind underneath it and the construction died at `chdir` —
    /// *after* `pivot_root`, with the host root already detached, so the failure arrived as a bare
    /// `ENOENT` from inside a root that no longer had a way back.
    #[test]
    fn a_workdir_under_tmp_survives_the_composed_tmp_mounted_over_its_path() {
        let workdir = "/tmp/session-1234/w";
        let plan = plan_for(workdir);
        let operations = ops(&plan);

        // A workdir under /tmp forces a base other than /tmp, so both mounts are still expressible.
        assert_eq!(
            choose_root_base(Path::new(workdir), &[], SEALED_ROOT_BASE_CANDIDATES, |_| {
                true
            }),
            Some(PathBuf::from("/run")),
        );

        let tmp_index = operations
            .iter()
            .position(
                |op| matches!(op, RootOp::BindVerified { target, .. } if target == Path::new("/tmp/tmp")),
            )
            .expect("the composed /tmp is bound");
        let workdir_index = operations
            .iter()
            .position(
                |op| matches!(op, RootOp::Bind { source, .. } if source == Path::new(workdir)),
            )
            .expect("the workdir is bound");
        assert!(
            tmp_index < workdir_index,
            "/tmp must be mounted before the workdir path is built inside it, or the workdir bind \
             is hidden by it",
        );
    }

    #[test]
    fn dev_carries_the_oci_default_set_and_is_sealed_afterwards() {
        let plan = plan_for("/home/u/w");
        let operations = ops(&plan);

        for device in SEALED_DEVICE_NODES {
            assert!(
                operations.contains(&RootOp::Bind {
                    source: PathBuf::from("/dev").join(device.name),
                    target: PathBuf::from("/tmp/dev").join(device.name),
                    read_only: false,
                }),
                "missing /dev/{}",
                device.name
            );
        }
        // No block device, and nothing outside the OCI default set.
        assert!(!operations.iter().any(
            |op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/dev/sda"))
        ));

        let tmpfs_index = operations
            .iter()
            .position(
                |op| matches!(op, RootOp::Tmpfs { target, .. } if target == Path::new("/tmp/dev")),
            )
            .expect("/dev tmpfs");
        let sealed_index = operations
            .iter()
            .position(
                |op| matches!(op, RootOp::RemountReadOnly(path) if path == Path::new("/tmp/dev")),
            )
            .expect("/dev sealed read-only");
        assert!(
            tmpfs_index < sealed_index,
            "the device nodes must be bound before /dev is sealed"
        );
    }

    /// The plan always asks for a masked `/proc`; whether the kernel grants one is a runtime
    /// question the executor answers, and on a bare host the answer is no — see
    /// [`PROC_HIDEPID_OPTIONS`] for the `CAP_SYS_ADMIN`-over-the-PID-namespace rule that forces
    /// the bind fallback. What this pins is the *preference order*: masking is tried first, and
    /// the unmasked private `procfs` is the last of the three before the fallback, never the
    /// first.
    #[test]
    fn proc_is_planned_and_masking_is_preferred_over_an_unmasked_mount() {
        let plan = plan_for("/home/u/w");
        assert!(ops(&plan).contains(&RootOp::Proc {
            target: PathBuf::from("/tmp/proc")
        }));
        assert_eq!(PROC_HIDEPID_OPTIONS[0], "hidepid=2");
        assert_eq!(
            PROC_HIDEPID_OPTIONS.last(),
            Some(&""),
            "the unmasked spelling must be tried last, after every masked one",
        );
    }

    #[test]
    fn manifest_derived_directories_are_whole_directory_binds_and_deduped() {
        let mut host = FakeHost::usrmerge();
        host.0
            .insert(PathBuf::from("/opt/python3.12"), PathKind::Dir);
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &test_sources(),
            Path::new("/tmp"),
            &[PathBuf::from("/opt/python3.12"), PathBuf::from("/usr")],
            &[],
            &host,
        );
        let operations = ops(&plan);

        assert!(operations.contains(&RootOp::Bind {
            source: PathBuf::from("/opt/python3.12"),
            target: PathBuf::from("/tmp/opt/python3.12"),
            read_only: true,
        }));
        // `/usr` was already bound by the fixed list; naming it again adds nothing.
        assert_eq!(
            operations
                .iter()
                .filter(
                    |op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/usr"))
                )
                .count(),
            1
        );
    }

    /// The property the whole capability rests on: `extra_read_only` shrinks when the host is
    /// missing a path, `staged_runtime_read_only` fails. Both directions are asserted from the
    /// same plan so the contrast cannot drift apart.
    #[test]
    fn a_staged_runtime_path_absent_from_the_host_still_plans_a_required_bind() {
        // Neither path is in `FakeHost::usrmerge()` — the host has neither.
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &test_sources(),
            Path::new("/tmp"),
            &[PathBuf::from("/opt/optional-tree")],
            &[PathBuf::from("/opt/staged-tree")],
            &FakeHost::usrmerge(),
        );

        // The optional entry contributes nothing at all — not even a step that would fail. This
        // is `mirror`'s silent-skip-on-absence, and it is why `extra_read_only` could not be the
        // destination for a staged grant.
        assert!(
            !plan.steps.iter().any(|step| matches!(
                &step.op,
                RootOp::Bind { source, .. } if source == Path::new("/opt/optional-tree")
            )),
            "an absent extra_read_only path must plan no step",
        );

        // The staged entry is planned regardless, and is required — so the real `mount(2)`
        // returning ENOENT aborts the construction before `pivot_root`.
        let staged = plan
            .steps
            .iter()
            .find(|step| {
                matches!(
                    &step.op,
                    RootOp::Bind { source, .. } if source == Path::new("/opt/staged-tree")
                )
            })
            .expect("an absent staged_runtime path must still plan a bind");
        assert_eq!(
            staged.op,
            RootOp::Bind {
                source: PathBuf::from("/opt/staged-tree"),
                target: PathBuf::from("/tmp/opt/staged-tree"),
                read_only: true,
            },
            "staged trees are re-based read-only binds at their own absolute path",
        );
        assert!(
            staged.required,
            "a staged_runtime bind must be required, so a missing source is session-fatal \
             rather than a silently narrower root",
        );
    }

    /// Staged binds land before `/etc`, `/dev`, `/proc`, `/tmp` and the workdir — and therefore
    /// before the `pivot_root` that `construct_composed_root` performs after every planned step.
    #[test]
    fn staged_runtime_binds_are_planned_before_every_other_step() {
        let mut host = FakeHost::usrmerge();
        host.0
            .insert(PathBuf::from("/opt/staged-tree"), PathKind::Dir);
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &test_sources(),
            Path::new("/tmp"),
            &[],
            &[PathBuf::from("/opt/staged-tree")],
            &host,
        );

        let position = |predicate: &dyn Fn(&RootOp) -> bool| {
            plan.steps
                .iter()
                .position(|step| predicate(&step.op))
                .expect("step must be planned")
        };

        let staged = position(
            &|op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/opt/staged-tree")),
        );
        let fixed_runtime_tree = position(
            &|op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/usr")),
        );
        let etc = position(
            &|op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/etc/ssl")),
        );
        let dev = position(
            &|op| matches!(op, RootOp::Tmpfs { target, .. } if target == Path::new("/tmp/dev")),
        );
        let proc = position(&|op| matches!(op, RootOp::Proc { .. }));
        let tmp = position(
            &|op| matches!(op, RootOp::BindVerified { target, .. } if target == Path::new("/tmp/tmp")),
        );
        let workdir = position(
            &|op| matches!(op, RootOp::Bind { source, .. } if source == Path::new("/home/u/w")),
        );

        // Ahead of the fixed runtime tree / `extra_read_only` loop, so a coincidental target
        // collision is won by the required registration rather than the optional one.
        assert!(
            staged < fixed_runtime_tree,
            "staged binds precede the fixed runtime tree"
        );
        assert!(staged < etc, "staged binds precede /etc");
        assert!(staged < dev, "staged binds precede /dev");
        assert!(staged < proc, "staged binds precede /proc");
        assert!(staged < tmp, "staged binds precede /tmp");
        assert!(staged < workdir, "staged binds precede the workdir bind");
    }

    /// The collision case the ordering above exists to protect: a path named by *both* parameters
    /// is registered once, by the required loop.
    #[test]
    fn a_path_named_both_staged_and_optional_stays_required() {
        let mut host = FakeHost::usrmerge();
        host.0.insert(PathBuf::from("/opt/shared"), PathKind::Dir);
        let plan = plan_composed_root(
            Path::new("/home/u/w"),
            &test_sources(),
            Path::new("/tmp"),
            &[PathBuf::from("/opt/shared")],
            &[PathBuf::from("/opt/shared")],
            &host,
        );

        let binds: Vec<&RootStep> = plan
            .steps
            .iter()
            .filter(|step| {
                matches!(
                    &step.op,
                    RootOp::Bind { source, .. } if source == Path::new("/opt/shared")
                )
            })
            .collect();
        assert_eq!(
            binds.len(),
            1,
            "the shared target is registered exactly once"
        );
        assert!(
            binds[0].required,
            "and the required registration is the one that won"
        );
    }

    // ---- `UsernsGrant` ---------------------------------------------------------------------

    /// Every variant, produced from the combination of readings that must produce it. The three
    /// permitting cases are the point: they reach the same achieved class, so only the variant
    /// separates a `sealed` result on a host whose hardening is switched off from one obtained
    /// through the shipped profile.
    #[test]
    fn every_userns_grant_comes_from_its_own_combination_of_readings() {
        let cases = [
            // AppArmor off, or the knob file missing entirely: nothing restricts anything.
            ((None, None, None), UsernsGrant::ApparmorAbsent),
            (
                (Some("N"), Some("1"), Some("unconfined")),
                UsernsGrant::ApparmorAbsent,
            ),
            // Enabled, restriction explicitly off host-wide — the `/etc/sysctl.d` drop-in case.
            (
                (Some("Y"), Some("0"), Some("unconfined")),
                UsernsGrant::RestrictionDisabledHostWide,
            ),
            // Enabled and the knob unreadable: not restricted, so not withheld.
            (
                (Some("1"), None, Some("unconfined")),
                UsernsGrant::RestrictionDisabledHostWide,
            ),
            // Restriction on, and a profile whose name begins with `mur-sealed` is attached —
            // both the shipped profile and the checkout profile the dev script generates.
            (
                (Some("Y"), Some("Y"), Some("mur-sealed (unconfined)")),
                UsernsGrant::ProfileConfining,
            ),
            (
                (Some("Y"), Some("1"), Some("mur-sealed-home (unconfined)")),
                UsernsGrant::ProfileConfining,
            ),
            (
                (Some("Y"), Some("1"), Some("mur-sealed-dev (unconfined)")),
                UsernsGrant::ProfileConfining,
            ),
            // Restriction on and nothing attached, or something else attached.
            (
                (Some("Y"), Some("Y"), Some("unconfined")),
                UsernsGrant::Withheld,
            ),
            (
                (Some("Y"), Some("1"), Some("firefox (enforce)")),
                UsernsGrant::Withheld,
            ),
            ((Some("Y"), Some("1"), None), UsernsGrant::Withheld),
        ];

        for ((enabled, restricted, current), expected) in cases {
            assert_eq!(
                UsernsGrant::from_readings(enabled, restricted, current),
                expected,
                "enabled={enabled:?} restricted={restricted:?} current={current:?}"
            );
        }
    }

    /// The wire names are what `--explain-scope --json`, `session_start` and `mur doctor` all
    /// print, so two variants sharing one would re-collapse exactly the distinction this enum
    /// exists to make.
    #[test]
    fn every_grant_has_a_distinct_wire_name_and_summary() {
        let names: Vec<&str> = UsernsGrant::ALL
            .iter()
            .map(|grant| grant.wire_name())
            .collect();
        let mut unique = names.clone();
        unique.sort_unstable();
        unique.dedup();
        assert_eq!(unique.len(), names.len(), "two grants share a wire name");
        assert_eq!(
            names,
            [
                "apparmor_absent",
                "restriction_disabled_host_wide",
                "profile_confining",
                "withheld",
            ]
        );

        for grant in UsernsGrant::ALL {
            assert_eq!(
                serde_json::to_value(grant).unwrap(),
                serde_json::Value::String(grant.wire_name().to_string()),
                "the serialized form and the wire name must be one string"
            );
        }

        // The `mur doctor` line under each wire name has to say something different for each, or
        // the block repeats the wire name in longer words and tells an operator nothing.
        let mut summaries: Vec<&str> = UsernsGrant::ALL.iter().map(|g| g.summary()).collect();
        summaries.sort_unstable();
        summaries.dedup();
        assert_eq!(
            summaries.len(),
            UsernsGrant::ALL.len(),
            "two grants share a summary"
        );

        // An unprobed host claims nothing, exactly as the `bool` field's `false` default did.
        assert_eq!(UsernsGrant::default(), UsernsGrant::Withheld);
    }

    // ---- the installed profile ------------------------------------------------------------

    /// All four outcomes, driven by a `Result` rather than by the filesystem, so a host with no
    /// AppArmor and a host with an unreadable `/etc/apparmor.d` are both covered without root and
    /// without touching a byte of `/etc`.
    #[test]
    fn the_installed_profile_classifier_separates_all_four_outcomes() {
        let shipped = std::fs::read(concat!(
            env!("CARGO_MANIFEST_DIR"),
            "/../../packaging/apparmor/mur-sealed"
        ))
        .expect("the shipped profile is in the workspace");
        assert_eq!(
            classify_installed_profile(Ok(shipped)),
            InstalledProfileState::Matches
        );

        let drifted = classify_installed_profile(Ok(b"# an older revision\n".to_vec()));
        match drifted {
            InstalledProfileState::Drifted { installed_sha256 } => {
                assert_ne!(installed_sha256, SEALED_APPARMOR_PROFILE_SHA256);
                assert_eq!(installed_sha256.len(), 64);
            }
            other => panic!("edited bytes must read as drift, got {other:?}"),
        }

        assert_eq!(
            classify_installed_profile(Err(std::io::Error::from(std::io::ErrorKind::NotFound))),
            InstalledProfileState::Absent
        );

        // "I could not look" must never read as "it is not there".
        match classify_installed_profile(Err(std::io::Error::from(
            std::io::ErrorKind::PermissionDenied,
        ))) {
            InstalledProfileState::Unreadable { error } => assert!(!error.is_empty()),
            other => panic!("a read error must stay distinct from absence, got {other:?}"),
        }
    }

    /// The probe child cannot compose a `CString`, so it carries NUL-terminated copies of the
    /// old-root parking directory's name: three spellings of one name.
    #[test]
    fn old_root_names_agree() {
        assert_eq!(OLD_ROOT_NAME_C.to_str().unwrap(), OLD_ROOT_NAME);
        assert_eq!(
            OLD_ROOT_PIVOTED_C.to_str().unwrap(),
            format!("/{OLD_ROOT_NAME}")
        );
    }

    /// Four outcomes, none of them inferred from a path. `unconfined` and "I could not look" stay
    /// apart for the same reason [`InstalledProfileState`]'s two absences do.
    #[test]
    fn profile_attachment_reports_only_what_the_kernel_said() {
        assert_eq!(
            classify_profile_attachment(Some("mur-sealed (unconfined)")),
            ProfileAttachment::Sealed {
                profile: "mur-sealed".to_string()
            }
        );
        assert_eq!(
            classify_profile_attachment(Some("mur-sealed-home")),
            ProfileAttachment::Sealed {
                profile: "mur-sealed-home".to_string()
            }
        );
        assert_eq!(
            classify_profile_attachment(Some("snap.chromium.chromium (enforce)")),
            ProfileAttachment::Other {
                profile: "snap.chromium.chromium".to_string()
            }
        );
        assert_eq!(
            classify_profile_attachment(Some("unconfined")),
            ProfileAttachment::Unconfined
        );
        assert_eq!(
            classify_profile_attachment(Some("unconfined\n")),
            ProfileAttachment::Unconfined
        );
        assert_eq!(
            classify_profile_attachment(None),
            ProfileAttachment::NotReported
        );
        assert_eq!(
            classify_profile_attachment(Some("   ")),
            ProfileAttachment::NotReported
        );
    }

    /// The attachment list is documentation of the shipped profile, never a verdict — nothing
    /// decides anything by matching a path against it. This keeps it honest about what the file
    /// actually declares.
    #[test]
    fn attachment_paths_are_the_ones_the_shipped_profile_declares() {
        let path = concat!(
            env!("CARGO_MANIFEST_DIR"),
            "/../../packaging/apparmor/mur-sealed"
        );
        let profile =
            std::fs::read_to_string(path).expect("the shipped profile is in the workspace");
        for attachment in SEALED_APPARMOR_ATTACHMENT_PATHS {
            assert!(
                profile.contains(attachment),
                "{attachment} is not an attachment in packaging/apparmor/mur-sealed"
            );
        }
    }

    /// The digest constant is a literal because `capsule-runtime` is published to crates.io and
    /// `cargo package` would not carry a file from outside the crate directory. This test is what
    /// keeps the literal honest, and it does not run during `cargo package`.
    #[test]
    fn shipped_profile_digest_constant_matches_the_file() {
        let path = concat!(
            env!("CARGO_MANIFEST_DIR"),
            "/../../packaging/apparmor/mur-sealed"
        );
        let bytes = std::fs::read(path).expect("the shipped profile is in the workspace");
        let actual = murmur_artifact::sha256_hex(&bytes);
        assert_eq!(
            actual, SEALED_APPARMOR_PROFILE_SHA256,
            "packaging/apparmor/mur-sealed changed. Update SEALED_APPARMOR_PROFILE_SHA256 in \
             crates/capsule-runtime/src/sealed.rs to:\n    {actual}"
        );
    }

    /// The profile is the fix and the sysctl is a fallback that costs the whole host its
    /// unprivileged-userns hardening. Both must be in the text — on a host where no profile can be
    /// loaded the sysctl is the right answer — but their order and the stated cost are what stop a
    /// reader taking them for peers.
    #[test]
    fn the_apparmor_reason_offers_the_profile_before_the_sysctl() {
        let reason = SealedBlocker::AppArmorProfileMissing.reason();
        let profile = reason
            .find("apparmor_parser -r")
            .expect("the profile install command must be named");
        let sysctl = reason
            .find("kernel.apparmor_restrict_unprivileged_userns=0")
            .expect("the fallback must still be named for hosts that cannot load a profile");
        assert!(
            profile < sysctl,
            "the profile install must come first: {reason}"
        );
        assert!(
            reason.contains("LAST RESORT"),
            "the sysctl must be labelled a fallback, not offered as a peer: {reason}"
        );
        assert!(
            reason.contains("every program on the machine"),
            "the sysctl's host-wide cost must be stated: {reason}"
        );
        assert!(
            reason.contains("scripts/install-dev-apparmor.sh"),
            "a checkout build must be told how to get the narrow grant: {reason}"
        );
    }

    /// `SealedBlocker::ALL` is what `murmur-cli`'s containment test matches a refusal against, so a
    /// variant missing from it makes that test assert something false about the *host*. Adding a
    /// variant must break this test — loudly, here — rather than surface three crates away.
    #[test]
    fn every_blocker_is_listed_in_all_with_its_own_actionable_reason() {
        // The exhaustive match is the mechanism: a new variant fails to compile until it is added
        // here, and the `contains` below then forces it into `ALL` too.
        for blocker in SealedBlocker::ALL {
            let named = match blocker {
                SealedBlocker::NotLinux
                | SealedBlocker::AppArmorProfileMissing
                | SealedBlocker::NamespaceCreationDenied
                | SealedBlocker::IdMapDenied
                | SealedBlocker::MountDenied
                | SealedBlocker::PivotRootDenied
                | SealedBlocker::KernelUnsupported
                | SealedBlocker::LandlockUnavailable
                | SealedBlocker::NetworkNamespaceUnavailable
                | SealedBlocker::ProbeUnavailable => *blocker,
            };
            assert!(
                SealedBlocker::ALL.contains(&named),
                "{named:?} is missing from SealedBlocker::ALL",
            );
        }

        let reasons: Vec<String> = SealedBlocker::ALL
            .iter()
            .map(|blocker| blocker.reason())
            .collect();
        for reason in &reasons {
            assert!(reason.starts_with("sealed "), "got: {reason}");
            assert!(!reason.contains('\n'), "got: {reason}");
            assert!(
                reason.len() > 80,
                "a blocker reason must name the mechanism and its remediation, not just fail: \
                 {reason}",
            );
        }
        let mut unique = reasons.clone();
        unique.sort();
        unique.dedup();
        assert_eq!(
            unique.len(),
            reasons.len(),
            "two blockers share a reason, so the refusal cannot say which one applies",
        );
    }

    #[test]
    fn the_device_set_is_exactly_the_oci_default_and_excludes_shm() {
        let names: Vec<&str> = SEALED_DEVICE_NODES.iter().map(|d| d.name).collect();
        assert_eq!(names, ["null", "zero", "full", "random", "urandom", "tty"]);
        assert!(!names.contains(&"shm"));
        assert!(!names.contains(&"console"));
        // Every entry records the canonical major/minor the OCI spec names.
        assert_eq!(
            SEALED_DEVICE_NODES
                .iter()
                .map(|d| (d.path, d.major, d.minor, d.mode, d.writable))
                .collect::<Vec<_>>(),
            [
                ("/dev/null", 1, 3, 0o666, true),
                ("/dev/zero", 1, 5, 0o666, true),
                ("/dev/full", 1, 7, 0o666, true),
                ("/dev/random", 1, 8, 0o666, false),
                ("/dev/urandom", 1, 9, 0o666, false),
                ("/dev/tty", 5, 0, 0o666, true),
            ]
        );
    }

    // ---- `sealed_shortfall` -----------------------------------------------------------------

    mod shortfall {
        use super::super::*;
        use crate::namespace_probe::{
            MountAttempt, MountAttemptError, MountStep, NamespaceReading,
        };
        use crate::network_namespace::{CreationStep, NamespaceAttemptError};

        fn docker() -> ProfileAttachment {
            ProfileAttachment::Other {
                profile: "docker-default".to_string(),
            }
        }

        fn created(grant: UsernsGrant, attachment: ProfileAttachment) -> NamespaceReading {
            NamespaceReading {
                network: Ok(()),
                mount: MountAttempt::Created,
                sealed_root: MountAttempt::Created,
                userns_grant: grant,
                attachment,
            }
        }

        fn network_failed(step: CreationStep, errno: i32) -> NamespaceReading {
            NamespaceReading {
                network: Err(NamespaceAttemptError::Step { step, errno }),
                mount: MountAttempt::NotAttempted,
                sealed_root: MountAttempt::NotAttempted,
                ..created(UsernsGrant::Withheld, ProfileAttachment::Unconfined)
            }
        }

        fn root_failed(step: MountStep, errno: i32) -> NamespaceReading {
            NamespaceReading {
                sealed_root: MountAttempt::Failed(MountAttemptError::Step { step, errno }),
                ..created(UsernsGrant::Withheld, ProfileAttachment::Unconfined)
            }
        }

        /// Every mapping of the design, with the blocker it decides and the rendering it carries:
        /// the class sentence, the failed call, the `errno` symbol and words, and the readings
        /// line last.
        #[test]
        fn every_failed_attempt_names_its_blocker_call_errno_and_readings() {
            use CreationStep as C;
            use MountStep as M;
            let docker_eperm = |step| NamespaceReading {
                attachment: docker(),
                ..network_failed(step, libc::EPERM)
            };
            let cases: Vec<(NamespaceReading, SealedBlocker, &str, i32)> = vec![
                (
                    network_failed(C::Unshare, libc::ENOSPC),
                    SealedBlocker::KernelUnsupported,
                    C::Unshare.describe(),
                    libc::ENOSPC,
                ),
                (
                    network_failed(C::Unshare, libc::EINVAL),
                    SealedBlocker::KernelUnsupported,
                    C::Unshare.describe(),
                    libc::EINVAL,
                ),
                (
                    network_failed(C::Unshare, libc::ENOSYS),
                    SealedBlocker::KernelUnsupported,
                    C::Unshare.describe(),
                    libc::ENOSYS,
                ),
                (
                    network_failed(C::Unshare, libc::EUSERS),
                    SealedBlocker::KernelUnsupported,
                    C::Unshare.describe(),
                    libc::EUSERS,
                ),
                (
                    network_failed(C::Unshare, libc::EPERM),
                    SealedBlocker::AppArmorProfileMissing,
                    C::Unshare.describe(),
                    libc::EPERM,
                ),
                (
                    network_failed(C::IdentityMaps, libc::EPERM),
                    SealedBlocker::AppArmorProfileMissing,
                    C::IdentityMaps.describe(),
                    libc::EPERM,
                ),
                (
                    network_failed(C::IdentityMaps, libc::EACCES),
                    SealedBlocker::AppArmorProfileMissing,
                    C::IdentityMaps.describe(),
                    libc::EACCES,
                ),
                (
                    docker_eperm(C::Unshare),
                    SealedBlocker::NamespaceCreationDenied,
                    C::Unshare.describe(),
                    libc::EPERM,
                ),
                (
                    docker_eperm(C::IdentityMaps),
                    SealedBlocker::IdMapDenied,
                    C::IdentityMaps.describe(),
                    libc::EPERM,
                ),
                (
                    network_failed(C::Loopback, libc::EPERM),
                    SealedBlocker::NetworkNamespaceUnavailable,
                    C::Loopback.describe(),
                    libc::EPERM,
                ),
                (
                    network_failed(C::LocalRoute, libc::EPERM),
                    SealedBlocker::NetworkNamespaceUnavailable,
                    C::LocalRoute.describe(),
                    libc::EPERM,
                ),
                (
                    network_failed(C::Resolver, libc::EADDRINUSE),
                    SealedBlocker::NetworkNamespaceUnavailable,
                    C::Resolver.describe(),
                    libc::EADDRINUSE,
                ),
                (
                    network_failed(C::HandOff, libc::EMFILE),
                    SealedBlocker::NetworkNamespaceUnavailable,
                    C::HandOff.describe(),
                    libc::EMFILE,
                ),
                (
                    root_failed(M::JoinSession, libc::EINVAL),
                    SealedBlocker::NetworkNamespaceUnavailable,
                    M::JoinSession.describe(),
                    libc::EINVAL,
                ),
                (
                    root_failed(M::UnshareUserMount, libc::EUSERS),
                    SealedBlocker::KernelUnsupported,
                    M::UnshareUserMount.describe(),
                    libc::EUSERS,
                ),
                (
                    root_failed(M::UnshareUserMount, libc::ENOSPC),
                    SealedBlocker::KernelUnsupported,
                    M::UnshareUserMount.describe(),
                    libc::ENOSPC,
                ),
                (
                    root_failed(M::UnshareUserMount, libc::EPERM),
                    SealedBlocker::NamespaceCreationDenied,
                    M::UnshareUserMount.describe(),
                    libc::EPERM,
                ),
                (
                    root_failed(M::IdentityMaps, libc::EPERM),
                    SealedBlocker::IdMapDenied,
                    M::IdentityMaps.describe(),
                    libc::EPERM,
                ),
                (
                    root_failed(M::MakePrivateNested, libc::EACCES),
                    SealedBlocker::MountDenied,
                    M::MakePrivateNested.describe(),
                    libc::EACCES,
                ),
                (
                    root_failed(M::Tmpfs, libc::EPERM),
                    SealedBlocker::MountDenied,
                    M::Tmpfs.describe(),
                    libc::EPERM,
                ),
                (
                    root_failed(M::PivotPrepare, libc::EACCES),
                    SealedBlocker::PivotRootDenied,
                    M::PivotPrepare.describe(),
                    libc::EACCES,
                ),
                (
                    root_failed(M::PivotRoot, libc::EPERM),
                    SealedBlocker::PivotRootDenied,
                    M::PivotRoot.describe(),
                    libc::EPERM,
                ),
                (
                    root_failed(M::DetachOldRoot, libc::EINVAL),
                    SealedBlocker::PivotRootDenied,
                    M::DetachOldRoot.describe(),
                    libc::EINVAL,
                ),
            ];
            for (reading, blocker, call, errno) in cases {
                for landlock in [true, false] {
                    let shortfall = sealed_shortfall(true, landlock, &reading)
                        .unwrap_or_else(|| panic!("{reading:?} must be refused"));
                    assert_eq!(shortfall.blocker, blocker, "{reading:?}");
                    assert_eq!(shortfall.failed_call.as_deref(), Some(call));
                    assert_eq!(shortfall.errno, Some(errno));
                    let reason = shortfall.reason();
                    let name = crate::network_namespace::errno_name(errno).unwrap();
                    let words = crate::network_namespace::errno_words(errno);
                    assert!(reason.starts_with(&blocker.reason()), "{reason}");
                    assert!(
                        reason
                            .contains(&format!("The probe's {call} failed with {name} ({words}).")),
                        "{reason}"
                    );
                    let readings = reason.find("Host readings:").expect("readings line");
                    assert!(reason[readings..].ends_with('.'), "{reason}");
                    assert!(
                        !reason[readings..].contains("The probe's"),
                        "the readings come last: {reason}"
                    );
                }
            }
        }

        /// A probe that could not run is refused as unknown, with the runtime's own account and
        /// no errno.
        #[test]
        fn a_probe_that_could_not_run_is_refused_rather_than_assumed() {
            for reading in [
                NamespaceReading {
                    network: Err(NamespaceAttemptError::Creator(
                        "fork() for the namespace creator failed: EAGAIN".to_string(),
                    )),
                    mount: MountAttempt::NotAttempted,
                    sealed_root: MountAttempt::NotAttempted,
                    ..created(UsernsGrant::Withheld, ProfileAttachment::Unconfined)
                },
                NamespaceReading {
                    sealed_root: MountAttempt::Failed(MountAttemptError::Rehearsal(
                        "fork() for the sealed-root rehearsal failed: EAGAIN".to_string(),
                    )),
                    ..created(UsernsGrant::Withheld, ProfileAttachment::Unconfined)
                },
            ] {
                let shortfall = sealed_shortfall(true, true, &reading).unwrap();
                assert_eq!(shortfall.blocker, SealedBlocker::ProbeUnavailable);
                assert_eq!(shortfall.errno, None);
                let reason = shortfall.reason();
                assert!(
                    reason.contains("The probe could not run: fork() for the"),
                    "{reason}"
                );
                assert!(reason.contains("Host readings:"), "{reason}");
            }
        }

        /// `AppArmorProfileMissing` comes only from a failed attempt: every grant and attachment
        /// over a reading whose attempts all created their namespaces is no shortfall at all.
        #[test]
        fn the_readings_never_refuse_a_host_whose_attempts_all_succeeded() {
            for grant in UsernsGrant::ALL {
                for attachment in [
                    ProfileAttachment::Unconfined,
                    docker(),
                    ProfileAttachment::NotReported,
                    ProfileAttachment::Sealed {
                        profile: "mur-sealed".to_string(),
                    },
                ] {
                    let reading = created(*grant, attachment);
                    assert_eq!(sealed_shortfall(true, true, &reading), None, "{reading:?}");
                    assert_eq!(
                        sealed_shortfall(true, false, &reading).map(|s| s.blocker),
                        Some(SealedBlocker::LandlockUnavailable)
                    );
                    assert_eq!(
                        sealed_shortfall(false, true, &reading).map(|s| s.blocker),
                        Some(SealedBlocker::NotLinux)
                    );
                }
            }
        }

        /// An `EPERM` the readings cannot explain — a grant other than `withheld` — is the
        /// namespace or its map refused, never the missing `mur-sealed` profile.
        #[test]
        fn a_refusal_under_any_other_grant_is_not_blamed_on_the_profile() {
            for grant in [
                UsernsGrant::ApparmorAbsent,
                UsernsGrant::ProfileConfining,
                UsernsGrant::RestrictionDisabledHostWide,
            ] {
                for (step, blocker) in [
                    (
                        CreationStep::Unshare,
                        SealedBlocker::NamespaceCreationDenied,
                    ),
                    (CreationStep::IdentityMaps, SealedBlocker::IdMapDenied),
                ] {
                    let reading = NamespaceReading {
                        userns_grant: grant,
                        ..network_failed(step, libc::EPERM)
                    };
                    assert_eq!(
                        sealed_shortfall(true, true, &reading).map(|s| s.blocker),
                        Some(blocker),
                        "{reading:?}"
                    );
                }
            }
        }

        /// The `/tmp` attempt is the mount namespace a `scoped` command uses; the composed root
        /// binds `/tmp` itself, so a refused `/tmp` mapping says nothing about `sealed`.
        #[test]
        fn the_tmp_attempt_never_affects_sealed() {
            let reading = NamespaceReading {
                mount: MountAttempt::Failed(MountAttemptError::Step {
                    step: MountStep::MakePrivate,
                    errno: libc::EACCES,
                }),
                ..created(UsernsGrant::Withheld, ProfileAttachment::Unconfined)
            };
            assert_eq!(sealed_shortfall(true, true, &reading), None);
        }

        /// No attempt failed for `NotLinux` and `LandlockUnavailable`, so their reason is the
        /// class sentence alone, with no readings.
        #[test]
        fn a_shortfall_without_a_failed_attempt_carries_no_readings() {
            let reading = created(UsernsGrant::Withheld, ProfileAttachment::Unconfined);
            for shortfall in [
                sealed_shortfall(false, true, &reading).unwrap(),
                sealed_shortfall(true, false, &reading).unwrap(),
            ] {
                assert_eq!(shortfall.reason(), shortfall.blocker.reason());
            }
        }

        /// The container case and the id-mapping case keep their own advice: a host that creates
        /// the namespace and refuses its map is not told to add a capability it is not missing,
        /// and a refused `pivot_root(2)` is not given the `mount(2)` remedy.
        #[test]
        fn each_blocker_keeps_its_own_remedy() {
            let container = SealedBlocker::NamespaceCreationDenied.reason();
            assert!(!container.contains("--cap-add"), "{container}");
            assert!(container.contains("seccomp profile"), "{container}");
            for call in [
                "unshare(CLONE_NEWUSER)",
                "setns",
                "mount",
                "umount2",
                "pivot_root",
            ] {
                assert!(container.contains(call), "{call}: {container}");
            }
            assert!(
                container.contains("--security-opt seccomp=<file>")
                    && container.contains("seccompProfile: Localhost"),
                "{container}"
            );
            assert!(container.contains("outside the container"), "{container}");
            let mount = SealedBlocker::MountDenied.reason();
            assert!(!mount.contains("--cap-add"), "{mount}");
            assert!(
                mount.contains("seccomp profile")
                    && mount.contains("mount, umount2 and pivot_root")
                    && mount.contains("--security-opt seccomp=<file>")
                    && mount.contains("apparmor_parser -r"),
                "{mount}"
            );
            let map = SealedBlocker::IdMapDenied.reason();
            assert!(map.contains("uid_map"), "{map}");
            assert!(!map.contains("--cap-add"), "{map}");
            let pivot = SealedBlocker::PivotRootDenied.reason();
            assert!(
                pivot.contains("pivot_root(2)") && pivot.contains("seccomp profile"),
                "{pivot}"
            );
            assert!(!pivot.contains("CAP_SYS_ADMIN"), "{pivot}");
            assert!(pivot.contains("seccompProfile: Localhost"), "{pivot}");
            assert_ne!(pivot, SealedBlocker::MountDenied.reason());
            let network = SealedBlocker::NetworkNamespaceUnavailable.reason();
            assert!(network.contains("E-CAP-005"), "{network}");
            let apparmor = SealedBlocker::AppArmorProfileMissing.reason();
            assert!(apparmor.contains("mur-sealed") && apparmor.contains("apparmor_parser -r"));
        }
    }
}