dracon-system 0.112.44

use super::*;

#[cfg(unix)]
use std::os::unix::fs::PermissionsExt;

/// Serializes the "create an executable fixture, then exec it" window
/// across the test binary's threads. ADDED 2026-09-30.
///
/// WHY THIS EXISTS (root cause of the long-standing ETXTBSY flakes in
/// `renice_process_with_bin_reports_success_and_failure`,
/// `nix_cleanup_apply_preserves_configured_profile_generations`,
/// `restore_runtime_adjustments_composes_overlapping_nice_limiters`,
/// `restore_runtime_adjustments_preserves_current_pid_incarnation`,
/// `restore_runtime_adjustments_restores_renice_and_oom` and
/// `restore_runtime_adjustments_retains_failed_cpu_cap`):
///
/// `execve()` of a file that this process created milliseconds earlier
/// can fail with `ETXTBSY` ("Text file busy", os error 26). The kernel
/// refuses to execute an inode it still considers write-busy, and the
/// window in which a just-written ext4 inode stays in that state is
/// long enough for a sibling test thread to collide with it. `libtest`
/// runs tests in parallel threads inside one process, so several
/// fixtures are being written and exec'd concurrently by construction.
///
/// The contended resource is the fixture inode of THIS process, and
/// that is what fixes the scope of the remedy. Discriminators measured
/// with a standalone 8-thread reproducer containing none of this crate's
/// code (write -> chmod -> fork -> execv, unique paths, 4800-6400 execs
/// per configuration):
///
///   | configuration                          | ETXTBSY |
///   |----------------------------------------|---------|
///   | 1 thread, no contention                |       0 |
///   | 4 threads, nothing serialized          |  27-155 |
///   | 8 threads, nothing serialized          |      25 |
///   | 8 threads, whole create->exec serialized|       0 |
///   | 8 threads, ONLY the write serialized   |       5 |
///   | 8 threads, ONLY the exec serialized    |       0 |
///   | 8 threads, per-thread interpreter copy  |      53 |
///   | 8 threads, path reuse (O_TRUNC)        |     155 |
///   | 8 threads, fsync(fd) before close      |    3033 |
///   | 8 threads + 4 external /bin/sh hammer
///     processes, whole window serialized    |       0 |
///
/// Read together: concurrent *execs* are required, serializing the
/// writes alone does not help, and a private interpreter copy does not
/// help — so it is neither a shared `/bin/sh` inode nor inode reuse.
/// The last row is why this is a process-local lock: 24,000 external
/// `/bin/sh` execs running against a serialized fixture test produced
/// zero failures, so a test binary running twice at once, or any other
/// process on the host, cannot reintroduce the flake.
///
/// THE LOCK MUST SPAN THE WRITE *AND* THE EXEC. In the C sketch
/// serializing only the exec half looks sufficient, but the Rust harness
/// disagrees: the tripwire below was first written that way and still
/// failed (`left: 1`) twice in ~89 suite runs. The broader window is
/// therefore the invariant: with the guard held across
/// `write_test_script` and the exec, the tripwire gave 140 consecutive
/// clean runs (112,000 create->exec cycles), while removing the guard
/// from the same test was detected in 20/20 runs. That statistical
/// tripwire has since been replaced by the deterministic guards named
/// below, for the reason given under "SERIALIZATION IS NECESSARY BUT NOT
/// SUFFICIENT".
///
/// SERIALIZATION IS NECESSARY BUT NOT SUFFICIENT UNDER LOAD. With the
/// lock in place the flake still appeared in 3 of 48 full-suite runs
/// (all inside the test that measures the raw hazard, 1-2 ETXTBSY per 800
/// cycles) while this host
/// was running at load average 87: with the process descheduled between
/// the write and the exec, a serialized create->exec can still land
/// inside the busy window. So the lock removes the in-process collision
/// — the reported root cause — and `settle_fixture` closes what is left.
/// Settling is not a retry of an assertion: it establishes a real
/// invariant, namely that the fixture has been exec'd at least once, and
/// a fixture that has been exec'd once is provably no longer write-busy
/// (re-exec'ing a pre-written script never failed, even from many
/// threads at once). `write_test_script` therefore settles every fixture
/// it creates, so the tests' own execs always run against a settled
/// inode no matter how busy the machine is.
///
/// Re-verify any of the above without this crate, in seconds:
/// `cc -O2 -pthread -o repro scripts/etxtbsy-repro.c && ./repro 600 1`
/// (0 failures) versus `./repro 600 8` (dozens of ETXTBSY). The shipped
/// copy of that program documents every variant measured here and sets
/// `O_CLOEXEC` on its fixtures because `std::fs` always does; its numbers
/// are a mechanism sketch, and the authoritative measurements are the
/// Rust ones in the tripwire below.
///
/// INVARIANT: any test that writes an executable fixture and then execs
/// it must hold `fixture_exec_guard()` from before the first
/// `write_test_script` until after the last exec — the write must be
/// inside the critical section, not just the exec — and must obtain the
/// fixture through `write_test_script`, which settles it.
/// `fixture_writing_tests_hold_the_fixture_exec_guard` enforces the first
/// half on every fixture-writing test in this module;
/// `settle_fixture_probes_without_side_effects_and_fails_loudly` and
/// `settled_fixture_survives_parallel_reexec` cover the second. All three
/// are deterministic, because an assertion of the form "zero ETXTBSY" on
/// the raw hazard is statistical and host load alone breaks it. `libtest`
/// prints captured
/// stdout only for failing tests, which is why the earlier
/// instrumentation also reported `Is a directory (os error 21)` and
/// `process identity unavailable` — those are by-design fixtures in
/// `restore_runtime_adjustments_restores_renice_and_oom` (a missing
/// proc root, a `renice` that exits 1, and a directory standing in for
/// an unwritable `oom_score_adj`), not additional failure modes.
///
/// THE PRODUCTION CODE IS NOT INVOLVED and needs no change: it only ever
/// execs resolved system binaries (`resolve_bin_strict("renice")`,
/// `nix-env`, `systemctl`, `ps`), which the process never writes. That
/// is also why "400 concurrent /bin/sh spawns succeed" never disproved
/// this: spawning a pre-existing binary concurrently is safe; it is a
/// write immediately before the exec of that same file that opens the
/// window.
#[cfg(unix)]
static FIXTURE_EXEC_LOCK: std::sync::Mutex<()> = std::sync::Mutex::new(());

/// Hold the fixture-exec lock for the whole create->exec window.
///
/// Poisoning is recovered rather than propagated: one panicking fixture
/// test must not cascade into every other test that needs the lock.
#[cfg(unix)]
fn fixture_exec_guard() -> std::sync::MutexGuard<'static, ()> {
    FIXTURE_EXEC_LOCK
        .lock()
        .unwrap_or_else(|poisoned| poisoned.into_inner())
}

#[test]
fn defaults_are_expected() {
    assert_eq!(default_min_size_mb(), 512);
    assert_eq!(default_kinds(), "rust-build,node-deps,build-output,cache");
}

#[test]
fn nix_delete_generations_arg_means_keep_last_n() {
    assert_eq!(nix_delete_generations_arg(5), Some("+5".to_string()));
    assert_eq!(nix_delete_generations_arg(1), Some("+1".to_string()));
    assert_eq!(nix_delete_generations_arg(0), None);
}

#[cfg(unix)]
#[tokio::test]
async fn nix_cleanup_apply_preserves_configured_profile_generations() {
    let _fixture_exec = fixture_exec_guard();
    fn shell_quote(value: &str) -> String {
        format!("'{}'", value.replace('\'', "'\\''"))
    }

    let root = std::env::temp_dir().join(format!(
        "dracon_system_nix_cleanup_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&root).expect("create temp dir");
    let log = root.join("calls.log");
    let quoted_log = shell_quote(log.to_str().expect("log path"));
    let nix_env = root.join("nix-env");
    let nix_gc = root.join("nix-collect-garbage");
    write_test_script(
        &nix_env,
        &format!("echo \"nix-env:$*\" >> {quoted_log}\nexit 0"),
    );
    write_test_script(
        &nix_gc,
        &format!(
            r#"echo "nix-gc:$*" >> {quoted_log}
if [ "$1" = "-d" ]; then
    echo 'would delete retained profile generations' >&2
    exit 99
fi
exit 0"#
        ),
    );

    let result = clean_nix_garbage_with_bins(5, true, &nix_env, &nix_gc)
        .await
        .expect("Nix cleanup should succeed without destructive garbage mode");
    assert_eq!(result, (0, Vec::new()));

    let calls = fs::read_to_string(&log).expect("read command log");
    let lines: Vec<&str> = calls.lines().collect();
    assert_eq!(
        lines.len(),
        3,
        "two profile prunes and one store GC expected"
    );
    assert_eq!(lines[0], "nix-env:--delete-generations +5");
    assert_eq!(
        lines[1],
        "nix-env:--delete-generations +5 -p /nix/var/nix/profiles/default"
    );
    assert_eq!(lines[2], "nix-gc:");
    assert!(
        !lines[2].contains("-d"),
        "apply must not use nix-collect-garbage -d, which discards retained generations"
    );
    let _ = fs::remove_dir_all(&root);
}

#[test]
fn guard_clean_bare_command_selects_all_targets() {
    let cli = Cli::try_parse_from(["dracon-system", "guard", "clean"])
        .expect("bare guard clean should parse");
    let Commands::Guard {
        cmd:
            GuardCommands::Clean {
                all,
                rust,
                trash,
                nix,
                caches,
                node_modules,
                docker,
                ..
            },
    } = cli.cmd
    else {
        panic!("expected guard clean command");
    };
    assert!(
        !all,
        "the bare command should not need an explicit --all flag"
    );

    let targets = resolve_clean_targets(
        all,
        &CleanTargets {
            rust,
            trash,
            nix,
            caches,
            node_modules,
            docker,
        },
    );
    assert!(
        targets.rust
            && targets.trash
            && targets.nix
            && targets.caches
            && targets.node_modules
            && targets.docker,
        "bare guard clean must resolve every cleanup target"
    );
}

#[test]
fn guard_clean_target_flags_remain_selective_and_all_overrides_them() {
    let rust_only = CleanTargets {
        rust: true,
        ..CleanTargets::default()
    };
    let resolved = resolve_clean_targets(false, &rust_only);
    assert!(resolved.rust);
    assert!(!resolved.trash);

    let resolved_all = resolve_clean_targets(true, &rust_only);
    assert!(resolved_all.rust && resolved_all.trash && resolved_all.docker);
}

#[cfg(unix)]
#[tokio::test]
async fn renice_process_with_bin_reports_success_and_failure() {
    let _fixture_exec = fixture_exec_guard();
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_renice_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("temp dir");
    let success = tmp.join("renice-success");
    write_test_script(&success, "echo 'ok' >&2\nexit 0");

    let failure = tmp.join("renice-failure");
    write_test_script(&failure, "echo 'permission denied' >&2\nexit 1");

    renice_process_with_bin(&success, 123, 5)
        .await
        .expect("success renice");
    let err = renice_process_with_bin(&failure, 123, 5)
        .await
        .expect_err("failure renice");
    assert!(err.to_string().contains("permission denied"));
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
fn write_process_fixture(
    root: &std::path::Path,
    pid: i32,
    comm: &str,
    starttime: u64,
    oom_score_adj: Option<i32>,
    cgroup: Option<&str>,
) {
    // Mirror the `/proc/self` availability marker used by the production
    // identity check so this fixture can distinguish a gone PID from a
    // missing/unavailable proc tree.
    fs::create_dir_all(root.join("self")).expect("create proc availability fixture");
    let dir = root.join(pid.to_string());
    fs::create_dir_all(&dir).expect("create process fixture");
    fs::write(dir.join("comm"), format!("{comm}\n")).expect("write comm fixture");
    let mut stat_fields = vec!["S".to_string()];
    stat_fields.extend(std::iter::repeat_n("0".to_string(), 18));
    stat_fields.push(starttime.to_string());
    fs::write(
        dir.join("stat"),
        format!("{pid} ({comm}) {}\n", stat_fields.join(" ")),
    )
    .expect("write stat fixture");
    if let Some(adj) = oom_score_adj {
        fs::write(dir.join("oom_score_adj"), format!("{adj}\n")).expect("write oom fixture");
    }
    if let Some(cgroup) = cgroup {
        fs::write(dir.join("cgroup"), cgroup).expect("write cgroup fixture");
    }
}

/// Marker argument used only by `settle_fixture`. No production caller
/// passes it, so a fixture that short-circuits on it is a no-op.
const FIXTURE_SETTLE_ARG: &str = "--dracon-fixture-settle";

/// Prove that `path` is executable *now*.
///
/// `execve` of a file this process just wrote can fail with `ETXTBSY`
/// because the kernel still considers the inode write-busy, and the
/// window survives a millisecond or two of quiet. That transient is
/// indistinguishable from a real test failure, so rather than retrying
/// an assertion this establishes the invariant the tests actually need:
/// *a fixture that has been exec'd once is no longer write-busy.*
/// Measured: re-exec'ing an already-exec'd script is always safe even
/// from many threads at once, so once this probe succeeds every later
/// exec of the same inode is safe too.
///
/// Bounded and loud by design — it is fixture construction, not a
/// swallow: a fixture that cannot be exec'd within the budget is a real
/// bug (missing interpreter, no execute bit) and must fail loudly.
#[cfg(unix)]
fn settle_fixture(path: &std::path::Path) {
    const ETXTBSY: i32 = 26;
    const ATTEMPTS: usize = 40;
    for attempt in 0..ATTEMPTS {
        match std::process::Command::new(path)
            .arg(FIXTURE_SETTLE_ARG)
            .output()
        {
            Ok(_) => return,
            Err(e) if e.raw_os_error() == Some(ETXTBSY) => {
                std::thread::sleep(std::time::Duration::from_millis(1 + (attempt as u64) / 8));
            }
            Err(e) => panic!("fixture {} is not executable: {e}", path.display()),
        }
    }
    panic!(
        "fixture {} stayed exec-busy (ETXTBSY) across {ATTEMPTS} settle attempts",
        path.display()
    );
}

/// Write an executable `#!/bin/sh` fixture without settling it.
///
/// The settle-marker guard is prepended here rather than in each body so
/// that no test can forget it: a fixture invoked with the marker exits
/// immediately, so probing it can never append to the log file (or emit
/// the stderr) that the real invocation is asserted against. Fixtures
/// that must observe a zero-argument call — `nix-collect-garbage` runs
/// without arguments — are covered by the marker too, because the probe
/// always passes exactly one argument.
#[cfg(unix)]
fn write_executable_script(path: &std::path::Path, body: &str) {
    fs::write(
        path,
        format!("#!/bin/sh\ncase \"$1\" in\n{FIXTURE_SETTLE_ARG}) exit 0;;\nesac\n{body}\n"),
    )
    .expect("write test script");
    fs::set_permissions(path, fs::Permissions::from_mode(0o755)).expect("chmod test script");
}

/// Write an executable fixture and settle it before any test can depend
/// on it. Every fixture a test actually asserts on goes through here.
#[cfg(unix)]
fn write_test_script(path: &std::path::Path, body: &str) {
    write_executable_script(path, body);
    settle_fixture(path);
}

/// Guards the settling half of the fix: once `settle_fixture` has
/// exec'd a fixture, re-exec'ing it must be safe no matter how busy the
/// host is.
///
/// This is the pattern every fixture test actually uses, and it is
/// deterministic rather than statistical: settling retries until the
/// inode has been exec'd once, and nothing writes the fixture afterwards,
/// so the busy window cannot reopen. Run on a host at load average 100+,
/// where an *unsettled* serialized create->exec still produced ETXTBSY in
/// 3 of 48 full-suite runs.
#[cfg(unix)]
#[test]
fn settled_fixture_survives_parallel_reexec() {
    const THREADS: usize = 8;
    const REEXECS: usize = 100;
    const ETXTBSY: i32 = 26;

    let etxtbsy = std::sync::Arc::new(std::sync::atomic::AtomicU64::new(0));
    let workers: Vec<_> = (0..THREADS)
        .map(|thread| {
            let etxtbsy = std::sync::Arc::clone(&etxtbsy);
            std::thread::spawn(move || {
                let dir = unique_test_home(&format!("settled_{thread}"));
                fs::create_dir_all(&dir).expect("create fixture dir");
                let script = dir.join("fixture");
                {
                    // Settling happens inside the guard; the re-execs below
                    // need no lock because settling already proved the
                    // inode has been exec'd.
                    let _fixture_exec = fixture_exec_guard();
                    write_test_script(&script, "exit 0");
                }
                for _ in 0..REEXECS {
                    if let Err(e) = std::process::Command::new(&script).output() {
                        if e.raw_os_error() == Some(ETXTBSY) {
                            etxtbsy.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
                        } else {
                            panic!("settled fixture exec failed unexpectedly: {e}");
                        }
                    }
                }
                let _ = fs::remove_dir_all(&dir);
            })
        })
        .collect();
    for worker in workers {
        worker.join().expect("worker thread must not panic");
    }

    assert_eq!(
        etxtbsy.load(std::sync::atomic::Ordering::SeqCst),
        0,
        "a settled fixture must never come back exec-busy ({} execs across {} threads)",
        THREADS * REEXECS,
        THREADS
    );
}

/// Guards the settling half of the fix, deterministically.
///
/// `settle_fixture` is what makes the fixture execs safe, so what needs
/// guarding is its contract rather than the raw kernel hazard it absorbs:
/// that the probe it runs has no side effects, and that a fixture it
/// cannot exec fails loudly instead of being swallowed. Asserting "zero
/// ETXTBSY" here instead would be a 1-in-20 statistical check that host
/// load alone can break.
#[cfg(unix)]
#[test]
fn settle_fixture_probes_without_side_effects_and_fails_loudly() {
    let dir = unique_test_home("settle_contract");
    fs::create_dir_all(&dir).expect("create fixture dir");

    // (1) The probe must not touch what the test asserts on. A fixture
    // that appends its argv to a log must leave that log untouched by
    // settling, and must still record exactly the real invocation later.
    let log = dir.join("calls.log");
    let escaped = log.to_string_lossy().replace('\'', "'\\''");
    let script = dir.join("recording");
    write_test_script(&script, &format!("printf '%s\\n' \"$*\" >> '{escaped}'"));
    assert!(
        !log.exists(),
        "settling must not run the fixture body: the probe created the log"
    );
    std::process::Command::new(&script)
        .args(["-n", "5", "-p", "123"])
        .output()
        .expect("a settled fixture must exec");
    assert_eq!(
        fs::read_to_string(&log).expect("log written by the real exec"),
        "-n 5 -p 123\n",
        "the real invocation must record exactly one line, probe or no probe"
    );

    // (2) A fixture that cannot be exec'd must fail loudly, naming the
    // path, rather than hanging or silently succeeding. The panic is
    // caught so the expectation is explicit instead of fatal.
    let broken = dir.join("not-executable");
    fs::write(&broken, "#!/bin/sh\nexit 0\n").expect("write broken fixture");
    fs::set_permissions(&broken, fs::Permissions::from_mode(0o644)).expect("clear the execute bit");
    let outcome = std::panic::catch_unwind(|| settle_fixture(&broken));
    let payload = outcome.expect_err("settling a non-executable fixture must panic, not swallow");
    let message = payload
        .downcast_ref::<String>()
        .cloned()
        .unwrap_or_else(|| "<non-string panic payload>".to_string());
    assert!(
        message.contains("is not executable") && message.contains("not-executable"),
        "the panic must name the fixture and the reason, got: {message}"
    );

    let _ = fs::remove_dir_all(&dir);
}

/// Guards the locking half of the fix, deterministically and without
/// running any fixture.
///
/// Rather than assert a property of the raw kernel hazard — which is
/// statistical, and which host load alone can still produce even with the
/// lock correctly held — this checks the invariant directly: every test in
/// this module that writes an executable fixture obtains
/// `fixture_exec_guard()` in the same body. A fixture test that forgets
/// the guard is the actual regression, and it is caught here instantly
/// instead of resurfacing as a 1-in-125 flake. Same shape as the coverage
/// tripwires elsewhere in this fleet: a developer adding a fixture test
/// cannot make the suite flaky by omission, they make this test fail.
#[cfg(unix)]
#[test]
fn fixture_writing_tests_hold_the_fixture_exec_guard() {
    const SOURCE: &str = include_str!("tests.rs");
    // The two writer helpers themselves: their bodies name them in the
    // signature and they create no fixtures of their own.
    const HELPERS: [&str; 2] = ["write_test_script", "write_executable_script"];

    // Slice the module into top-level function bodies. Every test and
    // helper here is a free function, so a body spans from one
    // column-zero `fn`/`async fn` line to the next.
    let mut bodies: Vec<(String, String)> = Vec::new();
    for line in SOURCE.lines() {
        if line.starts_with("fn ") || line.starts_with("async fn ") {
            let name = line
                .split_once('(')
                .map(|(head, _)| head)
                .unwrap_or(line)
                .trim_start_matches("async fn ")
                .trim_start_matches("fn ")
                .trim()
                .to_string();
            bodies.push((name, String::new()));
            continue;
        }
        if let Some((_, body)) = bodies.last_mut() {
            body.push_str(line);
            body.push('\n');
        }
    }

    let mut guarded = 0usize;
    let mut offenders: Vec<&str> = Vec::new();
    for (name, body) in &bodies {
        if HELPERS.contains(&name.as_str()) {
            continue;
        }
        let writes_fixture = ["write_test_script(", "write_executable_script("]
            .iter()
            .any(|call| body.contains(call));
        if !writes_fixture {
            continue;
        }
        guarded += 1;
        if !body.contains("fixture_exec_guard()") {
            offenders.push(name.as_str());
        }
    }

    assert!(
        guarded >= 10,
        "the source scan should find every fixture-writing test; found {guarded}, \
         which means the top-level body parser drifted from this module's layout"
    );
    assert!(
        offenders.is_empty(),
        "these tests write an executable fixture and then exec it but never hold \
         fixture_exec_guard() for the create->exec window (see FIXTURE_EXEC_LOCK): \
         {offenders:?}"
    );
}

#[cfg(unix)]
#[tokio::test]
async fn restore_runtime_adjustments_restores_renice_and_oom() {
    let _fixture_exec = fixture_exec_guard();
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_restore_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    let renice = tmp.join("renice");
    let systemctl = tmp.join("systemctl");
    write_test_script(&renice, "exit 0");
    write_test_script(&systemctl, "exit 0");
    write_process_fixture(&tmp, 101, "zsh", 77, Some(250), None);

    let identity = ProcessIdentity {
        comm: "zsh".to_string(),
        starttime: 77,
    };
    assert_eq!(
        process_identity_status(&tmp, 101, &identity),
        ProcessIdentityStatus::Match
    );
    assert_eq!(
        process_identity_status(
            &tmp,
            101,
            &ProcessIdentity {
                comm: "zsh".to_string(),
                starttime: 78,
            },
        ),
        ProcessIdentityStatus::Mismatch
    );
    assert_eq!(
        process_identity_status(&tmp, 102, &identity),
        ProcessIdentityStatus::Gone
    );
    let missing_proc_root = tmp.join("missing-proc-root");
    assert_eq!(
        process_identity_status(&missing_proc_root, 101, &identity),
        ProcessIdentityStatus::Unavailable,
        "a missing proc root must not be treated as a gone PID"
    );
    let mut unavailable_state = GuardRuntimeState::default();
    unavailable_state.reniced_pids.insert(
        101,
        LegacyReniceState {
            original_nice: 0,
            applied_nice: 5,
            identity: identity.clone(),
        },
    );
    unavailable_state.capped_pids.insert(
        101,
        (
            "dracon-cap.service".to_string(),
            "user.slice".to_string(),
            identity.clone(),
        ),
    );
    assert!(
        !restore_runtime_adjustments_with(
            &mut unavailable_state,
            &renice,
            &systemctl,
            &missing_proc_root,
        )
        .await,
        "missing proc root must retain every indeterminate adjustment"
    );
    assert!(unavailable_state.reniced_pids.contains_key(&101));
    assert!(unavailable_state.capped_pids.contains_key(&101));

    // Failed renice and OOM writes must retain their entries for the next
    // release retry rather than reporting success and dropping tracking.
    let failing_renice = tmp.join("renice-failure");
    write_test_script(&failing_renice, "exit 1");
    write_process_fixture(&tmp, 103, "worker", 79, None, None);
    write_process_fixture(&tmp, 104, "worker", 80, None, None);
    fs::create_dir(tmp.join("104/oom_score_adj")).expect("create failing oom fixture");
    let mut failure_state = GuardRuntimeState::default();
    failure_state.reniced_pids.insert(
        103,
        LegacyReniceState {
            original_nice: 0,
            applied_nice: 5,
            identity: ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 79,
            },
        },
    );
    failure_state.oom_biased_pids.insert(
        104,
        (
            -100,
            ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 80,
            },
        ),
    );
    assert!(
        !restore_runtime_adjustments_with(&mut failure_state, &failing_renice, &systemctl, &tmp)
            .await,
        "failed release operations must prevent runtime reset"
    );
    assert!(failure_state.reniced_pids.contains_key(&103));
    assert!(failure_state.oom_biased_pids.contains_key(&104));

    // A kernel comm name can be truncated or differ from argv[0]. The
    // release identity check must use the stable starttime, not cmdline.
    write_process_fixture(&tmp, 105, "very-long-process", 81, None, None);
    let long_comm_identity = ProcessIdentity {
        comm: "very-long".to_string(),
        starttime: 81,
    };
    assert_eq!(
        process_identity_status(&tmp, 105, &long_comm_identity),
        ProcessIdentityStatus::Match
    );

    write_process_fixture(&tmp, 106, "memory-worker", 82, None, None);
    let renice_log = tmp.join("renice.log");
    let recording_renice = tmp.join("renice-recording");
    let escaped_renice_log = renice_log.to_string_lossy().replace('\'', "'\\''");
    write_test_script(
        &recording_renice,
        &format!("printf '%s\\n' \"$*\" >> '{escaped_renice_log}'\nexit 0"),
    );

    let mut state = GuardRuntimeState::default();
    state.reniced_pids.insert(
        101,
        LegacyReniceState {
            original_nice: 0,
            applied_nice: 5,
            identity: identity.clone(),
        },
    );
    state.reniced_pids.insert(
        105,
        LegacyReniceState {
            original_nice: 0,
            applied_nice: 5,
            identity: long_comm_identity,
        },
    );
    state.oom_biased_pids.insert(101, (-100, identity));
    state.memory_reniced_pids.insert(
        106,
        MemoryReniceState {
            original_nice: 10,
            applied_nice: 5,
            identity: ProcessIdentity {
                comm: "memory-worker".to_string(),
                starttime: 82,
            },
        },
    );

    assert!(
        restore_runtime_adjustments_with(&mut state, &recording_renice, &systemctl, &tmp).await,
        "all successful restorations should permit runtime reset"
    );
    assert!(state.reniced_pids.is_empty());
    assert!(state.memory_reniced_pids.is_empty());
    assert!(state.oom_biased_pids.is_empty());
    assert_eq!(
        fs::read_to_string(tmp.join("101/oom_score_adj")).unwrap(),
        "-100\n"
    );
    let renice_calls = fs::read_to_string(&renice_log).unwrap();
    assert!(
        renice_calls.lines().any(|call| call == "-n 10 -p 106"),
        "memory release must restore the captured original nice value"
    );
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
#[tokio::test]
async fn restore_runtime_adjustments_composes_overlapping_nice_limiters() {
    let _fixture_exec = fixture_exec_guard();
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_overlap_restore_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    let renice_log = tmp.join("renice.log");
    let renice = tmp.join("renice");
    let systemctl = tmp.join("systemctl");
    let escaped_log = renice_log.to_string_lossy().replace('\'', "'\\''");
    write_test_script(
        &renice,
        &format!("printf '%s\\n' \"$*\" >> '{escaped_log}'\nexit 0"),
    );
    write_test_script(&systemctl, "exit 0");
    write_process_fixture(&tmp, 107, "overlap-worker", 83, None, None);

    let identity = ProcessIdentity {
        comm: "overlap-worker".to_string(),
        starttime: 83,
    };
    let mut state = GuardRuntimeState::default();
    state.reniced_pids.insert(
        107,
        LegacyReniceState {
            original_nice: 3,
            applied_nice: 8,
            identity: identity.clone(),
        },
    );
    state.memory_reniced_pids.insert(
        107,
        MemoryReniceState {
            original_nice: 3,
            applied_nice: 12,
            identity,
        },
    );

    assert!(
        restore_runtime_adjustments_with(&mut state, &renice, &systemctl, &tmp).await,
        "overlapping nice limiters should restore successfully"
    );
    assert!(state.reniced_pids.is_empty());
    assert!(state.memory_reniced_pids.is_empty());
    assert_eq!(
        fs::read_to_string(&renice_log)
            .expect("read renice log")
            .lines()
            .collect::<Vec<_>>(),
        vec!["-n 3 -p 107"],
        "one compositional restore must target the pre-limiter nice value"
    );
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
#[tokio::test]
async fn restore_runtime_adjustments_retains_overlapping_nice_limiters_on_failure() {
    let _fixture_exec = fixture_exec_guard();
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_overlap_failure_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    let renice = tmp.join("renice-failure");
    let systemctl = tmp.join("systemctl");
    write_test_script(&renice, "exit 1");
    write_test_script(&systemctl, "exit 0");
    write_process_fixture(&tmp, 108, "overlap-worker", 84, None, None);

    let identity = ProcessIdentity {
        comm: "overlap-worker".to_string(),
        starttime: 84,
    };
    let mut state = GuardRuntimeState::default();
    state.reniced_pids.insert(
        108,
        LegacyReniceState {
            original_nice: 4,
            applied_nice: 9,
            identity: identity.clone(),
        },
    );
    state.memory_reniced_pids.insert(
        108,
        MemoryReniceState {
            original_nice: 4,
            applied_nice: 13,
            identity,
        },
    );

    assert!(
        !restore_runtime_adjustments_with(&mut state, &renice, &systemctl, &tmp).await,
        "a failed compositional restore must remain retryable"
    );
    assert!(state.reniced_pids.contains_key(&108));
    assert!(state.memory_reniced_pids.contains_key(&108));
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
#[tokio::test]
async fn restore_runtime_adjustments_preserves_current_pid_incarnation() {
    let _fixture_exec = fixture_exec_guard();
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_pid_reuse_restore_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    let renice_log = tmp.join("renice.log");
    let renice = tmp.join("renice");
    let systemctl = tmp.join("systemctl");
    let escaped_log = renice_log.to_string_lossy().replace('\'', "'\\''");
    write_test_script(
        &renice,
        &format!("printf '%s\\n' \"$*\" >> '{escaped_log}'\nexit 0"),
    );
    write_test_script(&systemctl, "exit 0");
    // The legacy entry belongs to an exited incarnation; the memory entry is
    // for the replacement process now occupying the same PID.
    write_process_fixture(&tmp, 109, "reused-worker", 86, None, None);
    let mut state = GuardRuntimeState::default();
    state.reniced_pids.insert(
        109,
        LegacyReniceState {
            original_nice: 3,
            applied_nice: 8,
            identity: ProcessIdentity {
                comm: "reused-worker".to_string(),
                starttime: 85,
            },
        },
    );
    state.memory_reniced_pids.insert(
        109,
        MemoryReniceState {
            original_nice: 7,
            applied_nice: 12,
            identity: ProcessIdentity {
                comm: "reused-worker".to_string(),
                starttime: 86,
            },
        },
    );

    assert!(
        restore_runtime_adjustments_with(&mut state, &renice, &systemctl, &tmp).await,
        "stale PID state must not prevent current-incarnation restoration"
    );
    assert!(state.reniced_pids.is_empty());
    assert!(state.memory_reniced_pids.is_empty());
    assert_eq!(
        fs::read_to_string(&renice_log)
            .expect("read renice log")
            .lines()
            .collect::<Vec<_>>(),
        vec!["-n 7 -p 109"],
        "the replacement process must restore its own captured nice value"
    );
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
#[test]
fn stale_nice_state_does_not_override_current_pid_incarnation() {
    let mut state = GuardRuntimeState::default();
    state.reniced_pids.insert(
        110,
        LegacyReniceState {
            original_nice: 2,
            applied_nice: 8,
            identity: ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 90,
            },
        },
    );
    state.memory_reniced_pids.insert(
        110,
        MemoryReniceState {
            original_nice: 6,
            applied_nice: 12,
            identity: ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 91,
            },
        },
    );
    let current_identity = ProcessIdentity {
        comm: "worker".to_string(),
        starttime: 91,
    };

    drop_stale_nice_adjustments(&mut state, 110, &current_identity);

    assert!(!state.reniced_pids.contains_key(&110));
    assert_eq!(
        state
            .memory_reniced_pids
            .get(&110)
            .map(|entry| entry.original_nice),
        Some(6)
    );
}

#[cfg(unix)]
#[test]
fn sweep_stranded_oom_descendants_restores_only_post_bias_children() {
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_oom_descendants_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    write_process_fixture(&tmp, 101, "oom-parent", 77, Some(250), None);
    write_process_fixture(&tmp, 102, "existing-child", 78, Some(250), None);
    write_process_fixture(&tmp, 103, "forked-child", 79, Some(250), None);
    write_process_fixture(&tmp, 104, "unrelated", 80, Some(250), None);
    write_process_fixture(&tmp, 105, "failed-child", 81, None, None);
    fs::create_dir(tmp.join("105/oom_score_adj")).expect("create failing child oom fixture");

    let sample = |pid: i32, ppid: i32, starttime: u64, command: &str| ProcSample {
        pid,
        ppid,
        cpu_percent: 0.0,
        rss_mb: 1,
        nice: 0,
        command: command.to_string(),
        args: String::new(),
        starttime,
    };
    let samples = vec![
        sample(101, 1, 77, "oom-parent"),
        sample(102, 101, 78, "existing-child"),
        sample(103, 101, 79, "forked-child"),
        sample(104, 1, 80, "unrelated"),
        sample(105, 101, 81, "failed-child"),
    ];
    let identity = ProcessIdentity {
        comm: "oom-parent".to_string(),
        starttime: 77,
    };
    let mut state = GuardRuntimeState::default();
    state.oom_biased_pids.insert(101, (-100, identity));
    state
        .oom_known_descendants
        .insert(101, std::collections::HashSet::from([(102, 78)]));

    let actions = sweep_stranded_oom_descendants(
        &tmp,
        &samples,
        &mut state,
        &std::collections::HashSet::new(),
    );
    assert_eq!(
        actions.restored,
        vec!["oom-restore-descendant forked-child=-100"]
    );
    assert_eq!(actions.deferred, 1);
    assert!(state.oom_pending_descendants.contains_key(&(105, 81)));
    remove_oom_bias(&mut state, 101);
    assert!(
        state.oom_biased_pids.contains_key(&101),
        "root bias must remain tracked while child restoration is pending"
    );
    assert_eq!(
        fs::read_to_string(tmp.join("102/oom_score_adj")).unwrap(),
        "250\n"
    );
    assert_eq!(
        fs::read_to_string(tmp.join("103/oom_score_adj")).unwrap(),
        "-100\n"
    );
    assert_eq!(
        fs::read_to_string(tmp.join("104/oom_score_adj")).unwrap(),
        "250\n"
    );
    assert!(state
        .oom_known_descendants
        .get(&101)
        .unwrap()
        .contains(&(103, 79)));

    fs::remove_dir(tmp.join("105/oom_score_adj")).expect("remove failing child fixture");
    fs::write(tmp.join("105/oom_score_adj"), "250\n").expect("restore child fixture");
    let retry = sweep_stranded_oom_descendants(
        &tmp,
        &samples,
        &mut state,
        &std::collections::HashSet::new(),
    );
    assert_eq!(
        retry.restored,
        vec!["oom-restore-descendant failed-child=-100"]
    );
    assert_eq!(retry.deferred, 0);
    assert!(!state.oom_pending_descendants.contains_key(&(105, 81)));
    remove_oom_bias(&mut state, 101);
    assert!(!state.oom_biased_pids.contains_key(&101));
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
#[test]
fn sweep_stranded_oom_descendants_uses_nearest_biased_ancestor() {
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_nested_oom_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    write_process_fixture(&tmp, 201, "outer", 91, Some(250), None);
    write_process_fixture(&tmp, 202, "inner", 92, Some(250), None);
    write_process_fixture(&tmp, 203, "nested-child", 93, Some(250), None);
    let sample = |pid: i32, ppid: i32, starttime: u64, command: &str| ProcSample {
        pid,
        ppid,
        cpu_percent: 0.0,
        rss_mb: 1,
        nice: 0,
        command: command.to_string(),
        args: String::new(),
        starttime,
    };
    let samples = vec![
        sample(201, 1, 91, "outer"),
        sample(202, 201, 92, "inner"),
        sample(203, 202, 93, "nested-child"),
    ];
    let mut state = GuardRuntimeState::default();
    state.oom_biased_pids.insert(
        201,
        (
            -100,
            ProcessIdentity {
                comm: "outer".to_string(),
                starttime: 91,
            },
        ),
    );
    state.oom_biased_pids.insert(
        202,
        (
            25,
            ProcessIdentity {
                comm: "inner".to_string(),
                starttime: 92,
            },
        ),
    );
    state
        .oom_known_descendants
        .insert(201, std::collections::HashSet::from([(202, 92)]));
    state
        .oom_known_descendants
        .insert(202, std::collections::HashSet::new());

    let result = sweep_stranded_oom_descendants(
        &tmp,
        &samples,
        &mut state,
        &std::collections::HashSet::new(),
    );
    assert_eq!(
        result.restored,
        vec!["oom-restore-descendant nested-child=25"]
    );
    assert_eq!(result.deferred, 0);
    assert_eq!(
        fs::read_to_string(tmp.join("203/oom_score_adj")).unwrap(),
        "25\n"
    );
    let _ = fs::remove_dir_all(&tmp);
}

#[cfg(unix)]
#[tokio::test]
async fn restore_runtime_adjustments_retains_failed_cpu_cap() {
    let _fixture_exec = fixture_exec_guard();
    let tmp = std::env::temp_dir().join(format!(
        "dracon_system_cap_restore_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&tmp).expect("create temp dir");
    fs::create_dir_all(tmp.join("self")).expect("create proc availability marker");
    let renice = tmp.join("renice");
    let systemctl = tmp.join("systemctl");
    write_test_script(&renice, "exit 0");
    write_test_script(&systemctl, "exit 1");
    let mut state = GuardRuntimeState::default();
    state.capped_pids.insert(
        999_999,
        (
            "dracon-cap.service".to_string(),
            "user.slice".to_string(),
            ProcessIdentity {
                comm: "gone".to_string(),
                starttime: 1,
            },
        ),
    );

    assert!(
        !restore_runtime_adjustments_with(&mut state, &renice, &systemctl, &tmp).await,
        "failed systemd cleanup must prevent runtime reset"
    );
    assert!(state.capped_pids.contains_key(&999_999));

    // Once systemd succeeds, the previously retained cap may be removed.
    write_test_script(&systemctl, "exit 0");
    assert!(restore_runtime_adjustments_with(&mut state, &renice, &systemctl, &tmp).await);
    assert!(state.capped_pids.is_empty());

    // A live process whose cgroup cannot be read is indeterminate and must
    // also remain tracked even when systemd itself reports success.
    write_process_fixture(&tmp, 1000, "worker", 88, None, None);
    fs::create_dir(tmp.join("1000/cgroup")).expect("create unreadable cgroup fixture");
    state.capped_pids.insert(
        1000,
        (
            "dracon-cap.service".to_string(),
            "user.slice".to_string(),
            ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 88,
            },
        ),
    );
    assert!(
        !restore_runtime_adjustments_with(&mut state, &renice, &systemctl, &tmp).await,
        "cgroup read failure must prevent runtime reset"
    );
    assert!(state.capped_pids.contains_key(&1000));

    let systemctl_probe = tmp.join("systemctl-probe");
    let probe_log = tmp.join("systemctl-probe.log");
    let escaped_probe_log = probe_log.to_string_lossy().replace('\'', "'\\''");
    write_test_script(
        &systemctl_probe,
        &format!("printf 'called\\n' >> '{escaped_probe_log}'\nexit 0"),
    );

    // A missing cgroup file with a live PID directory is also indeterminate;
    // it may mean the cgroup source is unavailable rather than that the
    // process exited.
    write_process_fixture(&tmp, 1001, "worker", 89, None, None);
    state.capped_pids.insert(
        1001,
        (
            "dracon-cap.service".to_string(),
            "user.slice".to_string(),
            ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 89,
            },
        ),
    );
    assert!(
        !restore_runtime_adjustments_with(&mut state, &renice, &systemctl_probe, &tmp).await,
        "a missing live-PID cgroup file must prevent runtime reset"
    );
    assert!(state.capped_pids.contains_key(&1001));

    // Malformed cgroup data is equally indeterminate and must not trigger
    // systemd cleanup or allow the cap entry to be discarded.
    write_process_fixture(&tmp, 1002, "worker", 90, None, Some("malformed\n"));
    state.capped_pids.insert(
        1002,
        (
            "dracon-cap.service".to_string(),
            "user.slice".to_string(),
            ProcessIdentity {
                comm: "worker".to_string(),
                starttime: 90,
            },
        ),
    );
    assert!(
        !restore_runtime_adjustments_with(&mut state, &renice, &systemctl_probe, &tmp).await,
        "malformed cgroup data must prevent runtime reset"
    );
    assert!(state.capped_pids.contains_key(&1002));
    assert!(
        !probe_log.exists(),
        "indeterminate cgroup state must not stop the transient service"
    );
    let _ = fs::remove_dir_all(&tmp);
}

#[test]
fn human_bytes_formats_units() {
    assert_eq!(human_bytes(1), "1.0 B");
    assert_eq!(human_bytes(1024), "1.0 KiB");
    assert_eq!(human_bytes(1024 * 1024), "1.0 MiB");
}

#[test]
fn parse_kinds_trims_and_dedupes() {
    let kinds = parse_kinds(" rust-build, node-deps ,rust-build,,cache ");
    assert_eq!(kinds.len(), 3);
    assert!(kinds.contains("rust-build"));
    assert!(kinds.contains("node-deps"));
    assert!(kinds.contains("cache"));
}

#[test]
fn expand_tilde_resolves_to_home_dir() {
    // dirs::home_dir() uses getpwuid() on Linux, not $HOME.
    // Just verify ~ expands to whatever dirs reports.
    let home = dirs::home_dir().expect("home dir should exist");
    assert_eq!(expand_tilde("~"), home);
    assert_eq!(expand_tilde("~/foo/bar"), home.join("foo/bar"));
    assert_eq!(expand_tilde("/x/y"), PathBuf::from("/x/y"));
}

#[test]
fn expand_tilde_with_home_unset_falls_back_to_dot() {
    // We can't actually unset home for dirs::home_dir() (it uses getpwuid),
    // but we can verify the fallback path is wired correctly by testing
    // the helper directly if we could mock it. Instead, just verify
    // non-tilde paths pass through unchanged.
    assert_eq!(
        expand_tilde("/absolute/path"),
        PathBuf::from("/absolute/path")
    );
    assert_eq!(
        expand_tilde("relative/path"),
        PathBuf::from("relative/path")
    );
}

#[test]
fn guard_log_path_expands_the_example_configuration() {
    let policy: SystemPolicy = toml::from_str(include_str!("../dracon-system.example.toml"))
        .expect("shipped example policy should parse");
    assert_eq!(
        policy.guard.guard_log_file,
        "~/.local/state/dracon/dracon-system-guard.log"
    );

    let home = dirs::home_dir().expect("home directory");
    let resolved = resolve_guard_log_path(&policy.guard.guard_log_file)
        .expect("example guard log path should be enabled");
    assert_eq!(
        resolved,
        home.join(".local/state/dracon/dracon-system-guard.log")
    );
    assert!(!resolved.to_string_lossy().contains('~'));
    assert!(resolve_guard_log_path("  ").is_none());
    assert_eq!(
        resolve_guard_log_path("logs/guard.log"),
        Some(PathBuf::from("logs/guard.log"))
    );
}

#[test]
fn guard_log_example_path_writes_and_startup_rotates_in_isolated_home() {
    let policy: SystemPolicy = toml::from_str(include_str!("../dracon-system.example.toml"))
        .expect("shipped example policy should parse");
    assert_eq!(
        policy.guard.guard_log_file,
        "~/.local/state/dracon/dracon-system-guard.log"
    );

    let home = std::env::temp_dir().join(format!(
        "dracon_system_guard_log_home_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("clock after Unix epoch")
            .as_nanos()
    ));
    let mut guard = policy.guard;
    guard.guard_log_max_mb = 1;
    let path = resolve_guard_log_path_with_home(&guard.guard_log_file, Some(&home))
        .expect("example guard log path should be enabled");
    assert_eq!(
        path,
        home.join(".local/state/dracon/dracon-system-guard.log")
    );

    fs::create_dir_all(path.parent().expect("guard log parent")).expect("create guard log parent");
    fs::write(&path, vec![b'x'; 1024 * 1024 + 1]).expect("seed oversized guard log");
    rotate_guard_log_for_policy_with_home(&guard, true, Some(&home));
    assert!(
        !path.exists(),
        "startup rotation must use the expanded advertised log path"
    );

    log_guard_event_with_home(&guard, "f66-test", "expanded path", Some(&home));
    let contents = fs::read_to_string(&path).expect("expanded guard log should be written");
    let record: serde_json::Value =
        serde_json::from_str(contents.trim()).expect("guard log should contain JSONL");
    assert_eq!(record["event"], "f66-test");
    assert_eq!(record["details"], "expanded path");
    assert!(
        fs::metadata(&path).expect("guard log metadata").len() < 1024 * 1024,
        "guard log should remain below the configured limit after rotation"
    );

    fs::remove_dir_all(&home).expect("remove isolated guard log home");
}

#[test]
fn build_link_report_counts_states() {
    let policy = SystemPolicy {
        storage: StoragePolicy::default(),
        links: LinkPolicy {
            entries: vec![LinkEntry {
                link: "/tmp/does-not-exist-link".into(),
                target: "/tmp/does-not-exist-target".into(),
            }],
        },
        guard: GuardPolicy::default(),
    };
    let report = build_link_report(&policy);
    assert_eq!(report.total, 1);
    assert_eq!(report.healthy, 0);
    assert_eq!(report.drifted, 1);
    assert_eq!(report.missing_target, 1);
}

#[cfg(unix)]
#[test]
fn evaluate_link_handles_missing_and_sync_cases() {
    let base = std::env::temp_dir().join(format!(
        "dracon_system_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&base).expect("base dir");

    let target = base.join("target.txt");
    fs::write(&target, "x").expect("target");

    let missing_link = LinkEntry {
        link: base.join("missing-link").display().to_string(),
        target: target.display().to_string(),
    };
    let s1 = evaluate_link(&missing_link);
    assert_eq!(s1.issue, "link_missing");

    let normal_file_link = base.join("normal-file");
    fs::write(&normal_file_link, "x").expect("file");
    let not_symlink = LinkEntry {
        link: normal_file_link.display().to_string(),
        target: target.display().to_string(),
    };
    let s2 = evaluate_link(&not_symlink);
    assert_eq!(s2.issue, "path_not_symlink");

    let good_link = base.join("good-link");
    symlink(&target, &good_link).expect("symlink");
    let synced = LinkEntry {
        link: good_link.display().to_string(),
        target: target.display().to_string(),
    };
    let s3 = evaluate_link(&synced);
    assert_eq!(s3.issue, "ok");
    assert!(s3.in_sync);

    let wrong_target = base.join("other.txt");
    fs::write(&wrong_target, "y").expect("other");
    let mismatch_link = base.join("mismatch-link");
    symlink(&wrong_target, &mismatch_link).expect("symlink mismatch");
    let mismatch = LinkEntry {
        link: mismatch_link.display().to_string(),
        target: target.display().to_string(),
    };
    let s4 = evaluate_link(&mismatch);
    assert_eq!(s4.issue, "link_target_mismatch");

    let _ = fs::remove_dir_all(&base);
}

#[test]
fn parse_and_format_repeated_scenarios() {
    for i in 0..220usize {
        let csv = if i % 2 == 0 {
            "rust-build,node-deps,cache"
        } else {
            " rust-build , build-output , cache , rust-build "
        };
        let kinds = parse_kinds(csv);
        assert!(kinds.contains("rust-build"));
        assert!(kinds.contains("cache"));

        let bytes = (i as u64 + 1) * 2048;
        let out = human_bytes(bytes);
        assert!(!out.is_empty());
        assert!(out.contains(' '));
    }
}

#[test]
fn parse_df_use_percent_works() {
    let sample =
        "Filesystem 1024-blocks Used Available Capacity Mounted on\n/dev/root 100 91 9 91% /\n";
    assert_eq!(parse_df_use_percent(sample), Some(91));
}

#[test]
fn parse_ps_output_works() {
    let sample = "123 1 250.5 4194304 5 git\n456 2 12.0 2048 0 zsh\n";
    let rows = parse_ps_output(sample);
    assert_eq!(rows.len(), 2);
    assert_eq!(rows[0].pid, 123);
    assert_eq!(rows[0].ppid, 1);
    assert_eq!(rows[0].command, "git");
    assert_eq!(rows[0].rss_mb, 4096);
    assert_eq!(rows[0].nice, 5);
    assert_eq!(rows[0].args, "");
}

#[test]
fn is_protected_ancestor_exact_match() {
    assert!(is_protected_ancestor("/home", "/home"));
    assert!(is_protected_ancestor("/etc", "/etc"));
    assert!(is_protected_ancestor("/", "/"));
}

#[test]
fn is_protected_ancestor_descendant_match() {
    assert!(is_protected_ancestor("/home/dracon", "/home"));
    assert!(is_protected_ancestor("/home/dracon/Dev", "/home"));
    assert!(is_protected_ancestor("/etc/nginx/nginx.conf", "/etc"));
}

#[test]
fn is_protected_ancestor_rejects_partial_prefix() {
    assert!(!is_protected_ancestor("/homefoo", "/home"));
    assert!(!is_protected_ancestor("/homefoo/bar", "/home"));
    assert!(!is_protected_ancestor("/etcnginx", "/etc"));
}

#[test]
fn is_protected_ancestor_root_matches_exact_only() {
    assert!(is_protected_ancestor("/", "/"));
    assert!(!is_protected_ancestor("/anything", "/")); // root only matches exact to allow cleanup
    assert!(!is_protected_ancestor("/home", "/"));
}

#[test]
fn check_path_str_blocks_descendants() {
    assert!(!check_path_str("/home/dracon", &[]));
    assert!(!check_path_str("/home/dracon/Dev", &[]));
    assert!(!check_path_str("/etc/nginx", &[]));
    assert!(check_path_str("/safe/path", &[]));
    assert!(check_path_str("/homefoo", &[])); // partial prefix should be safe
}

#[test]
fn parse_ps_output_extracts_all_fields() {
    let sample = "9999 1 75.0 8192000 0 git-fetch origin main\n";
    let rows = parse_ps_output(sample);
    assert_eq!(rows.len(), 1);
    assert_eq!(rows[0].pid, 9999);
    assert_eq!(rows[0].ppid, 1);
    assert_eq!(rows[0].cpu_percent, 75.0);
    assert_eq!(rows[0].rss_mb, 8192000 / 1024);
    assert_eq!(rows[0].command, "git-fetch");
    assert_eq!(rows[0].args, "origin main");
}

#[test]
fn process_cmdline_is_sanitized_for_logs() {
    let raw = b"python3\0worker.py\0line\nwith\tcontrol\0";
    assert_eq!(
        sanitize_process_cmdline(raw, false),
        "python3 worker.py line\\nwith\\tcontrol"
    );
}

#[test]
fn process_cmdline_is_bounded() {
    let raw = vec![b'x'; MAX_PROCESS_CMDLINE_BYTES];
    let output = sanitize_process_cmdline(&raw, true);
    assert!(output.ends_with(" [truncated]"));
    assert!(output.len() <= MAX_PROCESS_CMDLINE_BYTES + " [truncated]".len());
}

#[test]
fn disk_state_transitions_at_thresholds() {
    let guard = GuardPolicy {
        disk_warn_percent: 70,
        disk_action_percent: 85,
        disk_critical_percent: 95,
        ..GuardPolicy::default()
    };
    assert_eq!(disk_state(50, &guard), "ok");
    assert_eq!(disk_state(70, &guard), "warn");
    assert_eq!(disk_state(84, &guard), "warn");
    assert_eq!(disk_state(85, &guard), "action");
    assert_eq!(disk_state(94, &guard), "action");
    assert_eq!(disk_state(95, &guard), "critical");
    assert_eq!(disk_state(100, &guard), "critical");
}

#[test]
fn should_notify_respects_cooldown() {
    let mut state = GuardRuntimeState::default();
    let key = "test-key";
    assert!(
        should_notify(&mut state, key, 60),
        "first notify should succeed"
    );
    assert!(
        !should_notify(&mut state, key, 60),
        "immediate second notify should be blocked"
    );
    assert!(
        should_notify(&mut state, "other-key", 60),
        "different key should succeed"
    );
}

#[test]
fn predict_fill_time_requires_minimum_samples() {
    let history: Vec<(Instant, u8)> = vec![(Instant::now(), 50), (Instant::now(), 51)];
    assert!(
        predict_fill_time(&history).is_none(),
        "needs at least 3 samples"
    );
}

#[test]
fn predict_fill_time_returns_none_for_stable_disk() {
    let base = Instant::now();
    let history: Vec<(Instant, u8)> = vec![
        (base, 50),
        (base + Duration::from_secs(10), 50),
        (base + Duration::from_secs(20), 50),
    ];
    assert!(
        predict_fill_time(&history).is_none(),
        "stable disk should not predict fill"
    );
}

#[test]
fn predict_fill_time_estimates_for_filling_disk() {
    let base = Instant::now();
    let history: Vec<(Instant, u8)> = vec![
        (base, 50),
        (base + Duration::from_secs(3600), 60),
        (base + Duration::from_secs(7200), 70),
    ];
    let hours = predict_fill_time(&history);
    assert!(hours.is_some(), "should predict fill time for rising disk");
    let h = hours.unwrap();
    assert!(h > 0.0, "predicted hours should be positive");
    assert!(
        h < 100.0,
        "predicted hours should be reasonable for 10%/hr rate"
    );
}

#[tokio::test]
async fn docker_prune_returns_zero_on_dry_run() {
    // When apply=false, docker_prune should return immediately without
    // invoking docker, yielding 0 bytes reclaimed.
    let result = docker_prune(false, true, true).await;
    assert!(result.is_ok(), "dry-run docker_prune should not error");
    assert_eq!(result.unwrap(), 0, "dry-run should reclaim 0 bytes");
}

#[tokio::test]
async fn empty_trash_credential_guard_blocks_dry_run_estimate() {
    let home = std::env::temp_dir().join(format!(
        "dracon_system_trash_guard_test_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    let trash_files = home.join(".local/share/Trash/files");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    let credential_fixture = trash_files.join("CREDENTIALS.md");
    fs::write(&credential_fixture, b"fixture contents").expect("write credential fixture");

    let (reclaimed, cleaned) = empty_trash_at(&home, false, &[], true, 0)
        .await
        .expect("dry-run trash scan");
    assert_eq!(
        reclaimed, 0,
        "blocked dry-run must not report reclaimable bytes"
    );
    assert!(
        cleaned.is_empty(),
        "blocked dry-run must report no cleanup action"
    );
    assert!(
        credential_fixture.exists(),
        "dry-run must not delete the credential-like fixture"
    );

    let _ = fs::remove_dir_all(&home);
}

// --- 2026-08-25 (v0.112.39): age-based trash + tmp cleanup regression tests ---

fn write_file_with_mtime(path: &std::path::Path, contents: &[u8], age_secs: u64) {
    fs::write(path, contents).expect("write fixture");
    let f = fs::OpenOptions::new()
        .append(true)
        .open(path)
        .expect("open fixture for mtime set");
    let mtime = SystemTime::now() - Duration::from_secs(age_secs);
    f.set_modified(mtime).expect("set_modified");
}

fn unique_test_home(tag: &str) -> std::path::PathBuf {
    std::env::temp_dir().join(format!(
        "dracon_system_{}_{}_{}",
        tag,
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ))
}

#[tokio::test]
async fn empty_trash_min_age_keeps_recent_entries() {
    let home = unique_test_home("trash_age");
    let trash_files = home.join(".local/share/Trash/files");
    let trash_info = home.join(".local/share/Trash/info");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    fs::create_dir_all(&trash_info).expect("create info fixture");

    // Old entry (30 days) must go; recent entry (1 hour) must stay.
    let old = trash_files.join("old-stuff.txt");
    let recent = trash_files.join("recent-stuff.txt");
    write_file_with_mtime(&old, b"old", 30 * 86_400);
    write_file_with_mtime(&recent, b"new", 3_600);
    fs::write(trash_info.join("old-stuff.txt.trashinfo"), b"[Trash Info]").expect("info old");
    fs::write(
        trash_info.join("recent-stuff.txt.trashinfo"),
        b"[Trash Info]",
    )
    .expect("info recent");

    let (reclaimed, cleaned) = empty_trash_at(&home, true, &[], false, 7)
        .await
        .expect("aged purge");
    assert!(reclaimed > 0, "aged purge should reclaim the old entry");
    assert_eq!(cleaned.len(), 1, "exactly one cleanup line expected");
    assert!(!old.exists(), "old entry must be removed");
    assert!(recent.exists(), "recent entry must be kept");
    assert!(
        !trash_info.join("old-stuff.txt.trashinfo").exists(),
        "paired .trashinfo for the old entry must be removed"
    );
    assert!(
        trash_info.join("recent-stuff.txt.trashinfo").exists(),
        "paired .trashinfo for the kept entry must survive"
    );

    let _ = fs::remove_dir_all(&home);
}

#[tokio::test]
async fn empty_trash_zero_age_empties_everything() {
    let home = unique_test_home("trash_zero");
    let trash_files = home.join(".local/share/Trash/files");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    write_file_with_mtime(&trash_files.join("a.txt"), b"a", 0);
    write_file_with_mtime(&trash_files.join("b.txt"), b"b", 0);

    let (_reclaimed, _cleaned) = empty_trash_at(&home, true, &[], false, 0)
        .await
        .expect("whole-empty purge");
    assert!(!trash_files.join("a.txt").exists());
    assert!(!trash_files.join("b.txt").exists());

    let _ = fs::remove_dir_all(&home);
}

fn set_mtime_old(path: &std::path::Path, age_secs: u64) {
    let f = fs::File::open(path).expect("open fixture for mtime set");
    let mtime = SystemTime::now() - Duration::from_secs(age_secs);
    f.set_modified(mtime).expect("set_modified");
}

#[tokio::test]
async fn empty_trash_credential_guard_skips_flagged_purges_rest() {
    // 2026-09-20 (space audit): a single flagged entry must not abort the
    // whole purge. The old all-or-nothing return let one trashed target/
    // dir (cargo fingerprints named *mint_dev_token*) block 46 GiB of
    // unrelated trash every cycle at 100% disk.
    let home = unique_test_home("trash_skip");
    let trash_files = home.join(".local/share/Trash/files");
    let trash_info = home.join(".local/share/Trash/info");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    fs::create_dir_all(&trash_info).expect("create info fixture");

    // Flagged OLD entry: nested credential-signal filename inside a dir
    // whose own mtime is past the age cutoff (so only the skip set keeps
    // it, not the age gate).
    let flagged = trash_files.join("target");
    fs::create_dir_all(flagged.join("debug/.fingerprint/x")).expect("flagged nest");
    write_file_with_mtime(
        &flagged.join("debug/.fingerprint/x/example-mint_dev_token"),
        b"t",
        30 * 86_400,
    );
    set_mtime_old(&flagged, 30 * 86_400);
    let benign = trash_files.join("old-notes.txt");
    write_file_with_mtime(&benign, b"notes", 30 * 86_400);
    fs::write(trash_info.join("target.trashinfo"), b"[Trash Info]").expect("info flagged");
    fs::write(trash_info.join("old-notes.txt.trashinfo"), b"[Trash Info]").expect("info benign");

    let (reclaimed, _cleaned) = empty_trash_at(&home, true, &[], true, 7)
        .await
        .expect("purge with flagged entry");
    assert!(
        reclaimed > 0,
        "benign entry must be purged despite the flag"
    );
    assert!(!benign.exists(), "benign old entry must be removed");
    assert!(flagged.exists(), "flagged entry must be kept");
    assert!(
        trash_info.join("target.trashinfo").exists(),
        "kept entry's .trashinfo must survive"
    );
    assert!(
        !trash_info.join("old-notes.txt.trashinfo").exists(),
        "purged entry's .trashinfo must be removed"
    );

    let _ = fs::remove_dir_all(&home);
}

#[tokio::test]
async fn empty_trash_zero_age_with_flagged_keeps_flagged() {
    // Age-0 whole-empty mode must degrade to per-entry purge (not a wipe
    // and not an abort) when the guard flags entries.
    let home = unique_test_home("trash_zero_skip");
    let trash_files = home.join(".local/share/Trash/files");
    let trash_info = home.join(".local/share/Trash/info");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    fs::create_dir_all(&trash_info).expect("create info fixture");
    write_file_with_mtime(&trash_files.join("id_ed25519.key"), b"k", 0);
    write_file_with_mtime(&trash_files.join("a.txt"), b"a", 0);
    fs::write(trash_info.join("id_ed25519.key.trashinfo"), b"[Trash Info]").expect("info flagged");
    fs::write(trash_info.join("a.txt.trashinfo"), b"[Trash Info]").expect("info benign");

    let (reclaimed, _cleaned) = empty_trash_at(&home, true, &[], true, 0)
        .await
        .expect("zero-age purge with flagged entry");
    assert!(reclaimed > 0, "benign entry must be purged");
    assert!(!trash_files.join("a.txt").exists());
    assert!(
        trash_files.join("id_ed25519.key").exists(),
        "flagged entry must be kept"
    );
    assert!(
        trash_info.join("id_ed25519.key.trashinfo").exists(),
        "kept entry's .trashinfo must survive the zero-age path"
    );

    let _ = fs::remove_dir_all(&home);
}

#[tokio::test]
async fn empty_trash_credential_guard_keeps_depth_capped_subtree() {
    // 2026-10-02 (audit L6): the scan is capped at max_depth(8) — a
    // directory AT the cap has unobservable children, so its whole
    // top-level entry must be kept (fail closed per-entry). The nested
    // content here is BENIGN: the keep must happen with no credential
    // signal at all, purely because the depth-9 file cannot be seen.
    let home = unique_test_home("trash_capped");
    let trash_files = home.join(".local/share/Trash/files");
    let trash_info = home.join(".local/share/Trash/info");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    fs::create_dir_all(&trash_info).expect("create info fixture");

    // deep(1)/d1(2)/d2(3)/d3(4)/d4(5)/d5(6)/d6(7)/d7(8)/notes.txt(9):
    // d7 sits AT the cap, notes.txt is never yielded.
    let mut deep = trash_files.join("deep");
    for d in ["d1", "d2", "d3", "d4", "d5", "d6", "d7"] {
        deep = deep.join(d);
    }
    fs::create_dir_all(&deep).expect("deep nest");
    write_file_with_mtime(&deep.join("notes.txt"), b"benign", 0);
    write_file_with_mtime(&trash_files.join("a.txt"), b"a", 0);
    fs::write(trash_info.join("deep.trashinfo"), b"[Trash Info]").expect("info deep");
    fs::write(trash_info.join("a.txt.trashinfo"), b"[Trash Info]").expect("info benign");

    let (reclaimed, _cleaned) = empty_trash_at(&home, true, &[], true, 0)
        .await
        .expect("zero-age purge with capped subtree");
    assert!(reclaimed > 0, "benign entry must be purged");
    assert!(
        !trash_files.join("a.txt").exists(),
        "benign entry must be removed"
    );
    assert!(
        trash_files.join("deep").exists(),
        "depth-capped entry must be kept (children unobservable)"
    );

    let _ = fs::remove_dir_all(&home);
}

#[cfg(unix)]
#[tokio::test]
async fn empty_trash_credential_guard_keeps_unreadable_subtree() {
    // 2026-10-02 (audit L6): a walk error (EPERM here) must fail closed
    // per-entry — the old `_ => {}` swallowed it and purged a subtree
    // the guard never saw.
    use std::os::unix::fs::PermissionsExt;

    let home = unique_test_home("trash_unreadable");
    let trash_files = home.join(".local/share/Trash/files");
    let trash_info = home.join(".local/share/Trash/info");
    fs::create_dir_all(&trash_files).expect("create trash fixture");
    fs::create_dir_all(&trash_info).expect("create info fixture");

    let locked = trash_files.join("top").join("locked");
    fs::create_dir_all(&locked).expect("locked nest");
    fs::write(locked.join("inner.txt"), b"x").expect("inner file");
    write_file_with_mtime(&trash_files.join("a.txt"), b"a", 0);
    fs::write(trash_info.join("top.trashinfo"), b"[Trash Info]").expect("info top");
    fs::write(trash_info.join("a.txt.trashinfo"), b"[Trash Info]").expect("info benign");
    fs::set_permissions(&locked, fs::Permissions::from_mode(0o000)).expect("lock dir");

    let (reclaimed, _cleaned) = empty_trash_at(&home, true, &[], true, 0)
        .await
        .expect("zero-age purge with unreadable subtree");
    fs::set_permissions(&locked, fs::Permissions::from_mode(0o755)).expect("unlock dir");
    assert!(reclaimed > 0, "benign entry must be purged");
    assert!(
        !trash_files.join("a.txt").exists(),
        "benign entry must be removed"
    );
    assert!(
        trash_files.join("top").exists(),
        "entry with an unreadable subtree must be kept"
    );

    let _ = fs::remove_dir_all(&home);
}

#[cfg(unix)]
#[tokio::test]
async fn nix_cleanup_gen_prune_failure_still_runs_gc() {
    let _fixture_exec = fixture_exec_guard();
    // 2026-09-20 (space audit): nix-env prune failures (a user-session
    // guard can never prune the root-owned system profile) must not fail
    // the pass or discard store-GC results.
    let root = unique_test_home("nix_warn");
    fs::create_dir_all(&root).expect("create temp dir");
    let nix_env = root.join("nix-env");
    let nix_gc = root.join("nix-collect-garbage");
    write_test_script(&nix_env, "echo 'cannot lock profile' >&2\nexit 1");
    write_test_script(&nix_gc, "echo 'deleting /nix/store/abc'\nexit 0");

    let (_bytes, cleaned) = clean_nix_garbage_with_bins(5, true, &nix_env, &nix_gc)
        .await
        .expect("GC success must win over prune failure");
    assert_eq!(cleaned.len(), 1, "GC results must survive prune failure");

    let _ = fs::remove_dir_all(&root);
}

#[cfg(unix)]
#[tokio::test]
async fn critical_tier_bypass_cleans_fresh_target() {
    // 2026-09-20 (space audit): the lingering age gate must lift at the
    // critical tier — a fresh 79 GiB workspace target held a 100% disk
    // hostage. Gated pass protects it; bypass pass (dry-run) candidates it.
    let root = unique_test_home("critical_rust");
    let proj = root.join("proj");
    let target = proj.join("target");
    fs::create_dir_all(target.join("debug")).expect("target fixture");
    fs::write(proj.join("Cargo.toml"), b"[package]\nname = \"x\"\n").expect("manifest");
    fs::write(target.join("debug/blob"), vec![0u8; 1024]).expect("blob");
    // Fresh (1h: past the 60s active-build backstop, inside the 7d gate).
    set_mtime_old(&target, 3_600);

    let policy = GuardPolicy {
        rust_search_roots: root.to_str().expect("utf8 root").to_string(),
        cleanup_min_size_mb: 0,
        rust_target_action_min_age_days: 7,
        ..Default::default()
    };
    let mut gated_state = GuardRuntimeState::default();
    let gated = auto_cleanup_rust_targets(&policy, &mut gated_state, false, false)
        .await
        .expect("gated scan");
    assert_eq!(
        gated.cleaned_count, 0,
        "age gate must protect the fresh target on the normal path"
    );
    let mut crit_state = GuardRuntimeState::default();
    let crit = auto_cleanup_rust_targets(&policy, &mut crit_state, false, true)
        .await
        .expect("bypass scan");
    assert_eq!(
        crit.cleaned_count, 1,
        "critical bypass must candidate the fresh target"
    );

    let _ = fs::remove_dir_all(&root);
}

#[tokio::test]
async fn clean_tmp_paths_respects_age_dry_run_and_open_fds() {
    let root = unique_test_home("tmp_clean");
    fs::create_dir_all(&root).expect("create tmp root");
    let old_file = root.join("stale-pi-bash.log");
    let new_file = root.join("active-session.log");
    write_file_with_mtime(&old_file, b"old log data", 2 * 86_400);
    write_file_with_mtime(&new_file, b"live", 60);

    // Hold an old-aged file open via this process so /proc/self/fd sees it.
    let held = root.join("held-open.bin");
    write_file_with_mtime(&held, b"held", 2 * 86_400);
    let handle = fs::File::open(&held).expect("open held file");

    let roots = vec![root.display().to_string()];

    // Dry-run measures but deletes nothing.
    let (dry_bytes, dry_lines) = clean_tmp_paths(false, &roots, 24, &[])
        .await
        .expect("dry run");
    assert!(dry_bytes > 0, "dry run should report reclaimable bytes");
    assert!(old_file.exists(), "dry run must not delete");
    assert!(
        !dry_lines.iter().any(|l| l.contains("active-session")),
        "young entries are not candidates"
    );
    assert!(
        !dry_lines.iter().any(|l| l.contains("held-open")),
        "open-held entries are skipped even when old"
    );

    // Apply removes only aged, unheld entries.
    let (bytes, lines) = clean_tmp_paths(true, &roots, 24, &[]).await.expect("apply");
    assert_eq!(
        bytes, dry_bytes,
        "apply should reclaim exactly what dry-run measured"
    );
    assert!(!old_file.exists(), "aged unheld entry must be removed");
    assert!(new_file.exists(), "young entry must be kept");
    assert!(
        held.exists(),
        "entry held open by a live process must be kept"
    );
    assert_eq!(lines.len(), 1);

    drop(handle);
    let _ = fs::remove_dir_all(&root);
}

#[tokio::test]
async fn clean_tmp_keeps_tree_with_fresh_nested_content() {
    // Audit MEDIUM: directory mtime only bumps on direct-child changes,
    // so a stale top dir can hide a freshly written nested file. The
    // cleaner must judge the tree by its newest entry, not the top dir.
    use std::time::Duration;
    let root = unique_test_home("tmp_nested");
    fs::create_dir_all(&root).expect("create tmp root");
    let stale_dir = root.join("old-run");
    let nested = stale_dir.join("logs");
    fs::create_dir_all(&nested).expect("create nested dir");
    write_file_with_mtime(&nested.join("fresh.log"), b"just written", 60);
    // Backdate everything except the fresh log, top dir included.
    let old = SystemTime::now() - Duration::from_secs(2 * 86_400);
    for p in [&stale_dir, &nested] {
        fs::OpenOptions::new()
            .read(true)
            .open(p)
            .expect("open dir for mtime set")
            .set_modified(old)
            .expect("backdate dir");
    }
    // A fully stale tree next to it must still be cleaned.
    let dead_dir = root.join("dead-run");
    fs::create_dir_all(&dead_dir).expect("create dead dir");
    write_file_with_mtime(&dead_dir.join("old.log"), b"stale", 2 * 86_400);
    fs::OpenOptions::new()
        .read(true)
        .open(&dead_dir)
        .expect("open dir for mtime set")
        .set_modified(old)
        .expect("backdate dir");

    let roots = vec![root.display().to_string()];
    let (bytes, lines) = clean_tmp_paths(true, &roots, 24, &[]).await.expect("apply");
    assert!(bytes > 0, "the dead tree must be reclaimed");
    assert!(!dead_dir.exists(), "fully stale tree must be removed");
    assert!(
        stale_dir.join("logs").join("fresh.log").exists(),
        "tree with fresh nested content must be kept: {lines:?}"
    );
    assert!(
        !lines.iter().any(|l| l.contains("old-run")),
        "kept tree must not be reported cleaned: {lines:?}"
    );
    let _ = fs::remove_dir_all(&root);
}

#[tokio::test]
async fn clean_tmp_paths_rejects_home_search_root_before_apply() {
    let home = dirs::home_dir().expect("home directory");
    let sentinel = home.join(format!(
        ".dracon_system_f65_home_sentinel_{}_{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    write_file_with_mtime(&sentinel, b"must survive", 2 * 86_400);

    let result = clean_tmp_paths(true, &["~".to_string()], 24, &[]).await;
    let sentinel_survived = sentinel.exists();
    let _ = fs::remove_file(&sentinel);

    let error = result.expect_err("home must not be accepted as a tmp root");
    assert!(
        error.to_string().contains("invalid tmp_search_paths entry"),
        "invalid root should be reported, got: {error:#}"
    );
    assert!(
        sentinel_survived,
        "an apply config targeting ~ must not delete home data"
    );
}

#[cfg(unix)]
#[tokio::test]
async fn clean_tmp_paths_keeps_old_process_cwd_directory() {
    let root = unique_test_home("tmp_cwd");
    let proc_root = unique_test_home("proc_cwd");
    let cwd_dir = root.join("stale-working-directory");
    let unheld_file = root.join("stale-unheld-file");
    fs::create_dir_all(&cwd_dir).expect("create cwd fixture");
    write_file_with_mtime(&unheld_file, b"unheld", 2 * 86_400);

    // A directory's mtime determines whether it is a tmp cleanup candidate.
    // Use an open directory handle so the fixture models an old process cwd,
    // rather than relying on the current time when the directory is created.
    let cwd_handle = File::open(&cwd_dir).expect("open cwd fixture for mtime");
    cwd_handle
        .set_modified(SystemTime::now() - Duration::from_secs(2 * 86_400))
        .expect("set cwd fixture mtime");
    drop(cwd_handle);

    // Model /proc/4242/cwd without changing this test process's real cwd.
    let process_dir = proc_root.join("4242");
    fs::create_dir_all(process_dir.join("fd")).expect("create proc fixture");
    symlink(&cwd_dir, process_dir.join("cwd")).expect("create cwd proc link");

    let roots = vec![root.display().to_string()];
    let unheld_bytes = unheld_file.metadata().expect("unheld metadata").len();
    let open_paths = collect_open_paths_under_from(&proc_root, std::slice::from_ref(&root))
        .await
        .expect("readable proc fixture must scan");
    assert!(
        open_paths.contains(&cwd_dir),
        "process cwd must be included in open-path protection"
    );

    let (reclaimed, cleaned) = clean_tmp_paths_with_proc(true, &roots, 24, &[], &proc_root)
        .await
        .expect("tmp cleanup");
    assert_eq!(reclaimed, unheld_bytes);
    assert_eq!(cleaned.len(), 1, "only the unheld old entry is removed");
    assert!(cwd_dir.exists(), "an old process cwd must remain protected");
    assert!(
        !unheld_file.exists(),
        "an old unheld entry remains removable"
    );

    let _ = fs::remove_dir_all(&root);
    let _ = fs::remove_dir_all(&proc_root);
}

#[cfg(unix)]
#[tokio::test]
async fn clean_tmp_paths_refuses_blind_cleanup_when_proc_unreadable() {
    // R3-L31: an unreadable /proc must abort the pass LOUD (no
    // silent empty protection set) and delete nothing — in BOTH
    // apply and dry-run (dry-run would otherwise over-report).
    let root = unique_test_home("tmp_blind");
    let stale = root.join("stale-file");
    fs::create_dir_all(&root).expect("create fixture");
    write_file_with_mtime(&stale, b"stale", 2 * 86_400);
    let roots = vec![root.display().to_string()];
    let missing_proc = root.join("no-such-proc");

    for apply in [false, true] {
        let err = clean_tmp_paths_with_proc(apply, &roots, 24, &[], &missing_proc)
            .await
            .expect_err("unreadable proc must fail the pass");
        assert!(
            err.to_string().contains("refusing tmp cleanup blind"),
            "failure must name the blind-cleanup refusal, got: {err:#}"
        );
        assert!(
            stale.exists(),
            "nothing may be deleted blind (apply={apply})"
        );
    }

    let _ = fs::remove_dir_all(&root);
}

#[test]
fn guard_policy_defaults_cover_tmp_and_trash_age_fields() {
    let guard = GuardPolicy::default();
    assert!(guard.clean_tmp, "clean_tmp should default on");
    assert_eq!(guard.tmp_search_paths, "/tmp");
    assert_eq!(guard.tmp_min_age_hours, 24);
    assert_eq!(guard.trash_min_age_days, 7);
}

#[test]
fn guard_startup_config_errors_use_non_restarting_status() {
    let wrapped = anyhow::Error::new(PolicyLoadError(anyhow::anyhow!("invalid policy")));
    assert_eq!(exit_status_for_error(&wrapped), CONFIG_ERROR_EXIT_STATUS);
    assert_eq!(
        exit_status_for_error(&anyhow::anyhow!("runtime failure")),
        1
    );
}

#[cfg(unix)]
#[test]
fn non_unicode_policy_override_is_authoritative() {
    use std::os::unix::ffi::OsStringExt;

    let custom = std::ffi::OsString::from_vec(b"/tmp/policy-\xff.toml".to_vec());
    let resolved = resolve_system_policy_path_with(Some(custom.clone()), PathBuf::from("/nowhere"));

    assert_eq!(resolved.unwrap(), Some(PathBuf::from(custom)));
}

#[test]
fn shipped_guard_service_exposes_host_tmp_for_clean_tmp() {
    let service = include_str!("../dracon-system-guard.service");
    assert!(
        service
            .lines()
            .any(|line| line.trim() == "PrivateTmp=false"),
        "the shipped guard must share the host temporary namespace"
    );
    assert!(
        !service.lines().any(|line| line.trim() == "PrivateTmp=true"),
        "the shipped guard must not use a private temporary namespace"
    );
    assert!(
        service
            .lines()
            .any(|line| line.trim() == "WorkingDirectory=%h"),
        "relative policy paths need a stable user-service working directory"
    );

    let writable_paths = service
        .lines()
        .find_map(|line| line.strip_prefix("ReadWritePaths="))
        .expect("the guard service must declare writable paths");
    assert!(
        writable_paths.split_whitespace().any(|path| path == "/tmp"),
        "host /tmp must be writable despite ProtectSystem=strict"
    );

    let guard = GuardPolicy::default();
    assert!(
        guard.clean_tmp,
        "the production default must enable clean_tmp"
    );
    assert_eq!(
        guard.tmp_search_paths, "/tmp",
        "the policy and service must target the same host namespace"
    );
}

#[test]
fn shipped_guard_service_restart_policy_handles_disabled_and_bad_config() {
    let service = include_str!("../dracon-system-guard.service");
    assert!(
        service
            .lines()
            .any(|line| line.trim() == "Restart=on-failure"),
        "clean enabled=false exits must not restart the guard"
    );
    assert!(
        !service.lines().any(|line| line.trim() == "Restart=always"),
        "the guard must not restart after a clean policy-disabled exit"
    );
    assert!(
        service
            .lines()
            .any(|line| line.trim() == "RestartPreventExitStatus=2 78"),
        "usage errors and EX_CONFIG policy errors must remain stopped"
    );
}

#[tokio::test]
async fn guard_report_completes_for_ok_disk() {
    let fixture = unique_test_home("guard_report");
    fs::create_dir(&fixture).expect("create report fixture");
    let mut state = GuardRuntimeState::default();
    // This is a report-assembly smoke test, not a cleanup integration test.
    // The host disk may be full. Force the OK branch and explicitly disable
    // every mutating/expensive subsystem instead of inheriting live defaults.
    // 101 is a test-only unreachable threshold for a percentage (0..=100).
    let guard = GuardPolicy {
        disk_early_warn_percent: 101,
        disk_warn_percent: 101,
        disk_action_percent: 101,
        disk_critical_percent: 101,
        proactive_cleanup_percent: 101,
        disk_mount_path: fixture.display().to_string(),
        sync_freeze_marker: fixture.join("freeze").display().to_string(),
        guard_log_file: fixture.join("guard.log").display().to_string(),
        freeze_sync_at_action: false,
        track_trends: false,
        notify: false,
        auto_cleanup_apply: false,
        auto_cleanup_rust: false,
        docker_prune: false,
        clean_package_caches: false,
        clean_trash: false,
        clean_nix_garbage: false,
        clean_node_modules: false,
        clean_tmp: false,
        monitor_inodes: false,
        monitor_zombies: false,
        monitor_memory: false,
        monitor_logs: false,
        auto_renice: false,
        auto_renice_on_memory: false,
        bias_oom_on_pressure: false,
        ..GuardPolicy::default()
    };
    let report = tokio::time::timeout(Duration::from_secs(10), run_guard_once(&guard, &mut state))
        .await
        .expect("report assembly must be bounded")
        .expect("report succeeds");
    assert_eq!(report.disk_state, "ok");
    assert!(!report.sync_frozen);
    assert!(!fixture.join("freeze").exists());
    fs::remove_dir_all(fixture).expect("remove report fixture");
}

#[test]
fn guard_never_removes_or_overwrites_operator_freeze_marker() {
    let marker = std::env::temp_dir().join(format!(
        "dracon-system-freeze-operator-{}-{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    let operator_contents = "dracon-sync pause\n";
    fs::write(&marker, operator_contents).expect("write operator marker");

    let guard = GuardPolicy {
        freeze_sync_at_action: false,
        sync_freeze_marker: marker.display().to_string(),
        unfreeze_below_percent: 70,
        ..GuardPolicy::default()
    };
    let mut sync_frozen = true;
    manage_sync_freeze(&guard, 68, "ok", &mut sync_frozen);
    assert!(marker.exists(), "guard-disabled policy must not remove it");
    assert_eq!(
        fs::read_to_string(&marker).expect("read marker"),
        operator_contents
    );
    assert!(sync_frozen);

    // Even when automatic freezing is enabled, an existing marker written by
    // dracon-sync remains operator-owned and must not be replaced or removed.
    let enabled_guard = GuardPolicy {
        freeze_sync_at_action: true,
        ..guard
    };
    manage_sync_freeze(&enabled_guard, 95, "action", &mut sync_frozen);
    assert!(marker.exists(), "guard must not replace an operator marker");
    assert_eq!(
        fs::read_to_string(&marker).expect("read marker"),
        operator_contents
    );

    let _ = fs::remove_file(marker);
}

#[test]
fn guard_removes_only_its_own_freeze_marker() {
    let marker = std::env::temp_dir().join(format!(
        "dracon-system-freeze-owned-{}-{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    let guard = GuardPolicy {
        freeze_sync_at_action: true,
        sync_freeze_marker: marker.display().to_string(),
        unfreeze_below_percent: 70,
        ..GuardPolicy::default()
    };

    let mut sync_frozen = false;
    manage_sync_freeze(&guard, 95, "action", &mut sync_frozen);
    assert!(sync_frozen);
    assert!(marker.exists());
    assert!(fs::read_to_string(&marker)
        .expect("read guard marker")
        .starts_with(GUARD_FREEZE_MARKER_PREFIX));

    manage_sync_freeze(&guard, 68, "ok", &mut sync_frozen);
    assert!(!marker.exists(), "guard-owned marker should be released");
    assert!(!sync_frozen);
}

#[test]
fn test_graduated_nice_value_cpu_tiers() {
    assert_eq!(graduated_nice_value(100.0, 0, 5, 0), 5);
    assert_eq!(graduated_nice_value(180.0, 0, 5, 0), 5);
    assert_eq!(graduated_nice_value(250.0, 0, 5, 0), 5);
    assert_eq!(graduated_nice_value(300.0, 0, 5, 0), 10);
    assert_eq!(graduated_nice_value(450.0, 0, 5, 0), 10);
    assert_eq!(graduated_nice_value(500.0, 0, 5, 0), 15);
    assert_eq!(graduated_nice_value(900.0, 0, 5, 0), 15);
}

#[test]
fn test_graduated_nice_value_memory_tiers() {
    assert_eq!(graduated_nice_value(0.0, 2000, 5, 0), 5);
    assert_eq!(graduated_nice_value(0.0, 4096, 5, 0), 5);
    assert_eq!(graduated_nice_value(0.0, 5000, 5, 0), 5);
    assert_eq!(graduated_nice_value(0.0, 8192, 5, 0), 10);
    assert_eq!(graduated_nice_value(0.0, 16000, 5, 0), 10);
}

#[test]
fn test_graduated_nice_value_cpu_plus_memory() {
    assert_eq!(graduated_nice_value(300.0, 8192, 5, 0), 10);
    assert_eq!(graduated_nice_value(500.0, 4096, 5, 0), 15);
    assert_eq!(graduated_nice_value(180.0, 8192, 5, 0), 10);
}

#[test]
fn test_graduated_nice_value_clamped() {
    assert_eq!(graduated_nice_value(0.0, 0, 5, 0), 5);
    assert_eq!(graduated_nice_value(0.0, 0, 0, 0), 0);
}

#[test]
fn test_graduated_nice_value_negative_base_clamped() {
    assert_eq!(graduated_nice_value(0.0, 0, -5, 0), 0);
}

#[test]
fn test_graduated_nice_value_high_base_clamped() {
    assert_eq!(graduated_nice_value(0.0, 0, 20, 0), 19);
}

#[test]
fn test_graduated_nice_value_memory_boundary() {
    assert_eq!(graduated_nice_value(0.0, 4095, 0, 0), 0);
    assert_eq!(graduated_nice_value(0.0, 4096, 0, 0), 5);
    assert_eq!(graduated_nice_value(0.0, 8191, 0, 0), 5);
    assert_eq!(graduated_nice_value(0.0, 8192, 0, 0), 10);
}

fn guard_test_tmp(name: &str) -> std::path::PathBuf {
    std::env::temp_dir().join(format!(
        "dracon_test_{}_{}_{}",
        name,
        std::process::id(),
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap()
            .as_nanos()
    ))
}

#[test]
fn safe_tmp_root_policy_allows_tmp_descendants_and_rejects_home() {
    let tmp = guard_test_tmp("safe_tmp_root");
    std::fs::create_dir_all(&tmp).expect("tmp root");

    let resolved = check_safe_tmp_root(&tmp).expect("tmp descendant should be allowed");
    assert!(resolved.starts_with("/tmp"));

    let home = dirs::home_dir().expect("home directory");
    let error = check_safe_tmp_root(&home).expect_err("home must not be a tmp root");
    assert!(error.to_string().contains("must be /tmp (or a descendant)"));

    let _ = std::fs::remove_dir_all(&tmp);
}

/// Set a file or dir mtime `hours` in the past (`File::set_modified`
/// is stable; opening a dir read-only suffices for futimens on Linux).
fn set_mtime_hours_ago(path: &std::path::Path, hours: u64) {
    let f = std::fs::OpenOptions::new()
        .read(true)
        .open(path)
        .unwrap_or_else(|e| panic!("open {} to set mtime: {e}", path.display()));
    f.set_modified(std::time::SystemTime::now() - std::time::Duration::from_secs(hours * 3600))
        .unwrap_or_else(|e| panic!("set mtime on {}: {e}", path.display()));
}

#[tokio::test]
async fn tree_has_fresh_content_fails_closed_on_walk_errors() {
    // R3-L22: unreadable subtrees (or any walk error) must read as
    // FRESH (keep), never skip toward a delete verdict. Controls:
    // old content reads stale, fresh content reads fresh.
    use std::os::unix::fs::PermissionsExt;
    let now = std::time::SystemTime::now();
    let cutoff = now - std::time::Duration::from_secs(60);

    let stale = guard_test_tmp("fresh_closed_stale");
    std::fs::create_dir_all(&stale).unwrap();
    let old_file = stale.join("old");
    std::fs::write(&old_file, b"old").unwrap();
    set_mtime_hours_ago(&old_file, 1);
    set_mtime_hours_ago(&stale, 1);
    assert!(
        !tree_has_fresh_content(&stale, cutoff).await,
        "all-old tree must read stale"
    );

    let fresh = guard_test_tmp("fresh_closed_fresh");
    std::fs::create_dir_all(&fresh).unwrap();
    std::fs::write(fresh.join("new"), b"new").unwrap();
    assert!(
        tree_has_fresh_content(&fresh, cutoff).await,
        "tree with a fresh file must read fresh"
    );

    // chmod-000 subdir: readdir fails EACCES → fail closed (fresh).
    let locked = guard_test_tmp("fresh_closed_locked");
    let sub = locked.join("sub");
    std::fs::create_dir_all(&sub).unwrap();
    std::fs::write(sub.join("x"), b"x").unwrap();
    set_mtime_hours_ago(&sub.join("x"), 1);
    set_mtime_hours_ago(&sub, 1);
    set_mtime_hours_ago(&locked, 1);
    std::fs::set_permissions(&sub, std::fs::Permissions::from_mode(0o000)).unwrap();
    let verdict = tree_has_fresh_content(&locked, cutoff).await;
    std::fs::set_permissions(&sub, std::fs::Permissions::from_mode(0o755)).unwrap();
    assert!(verdict, "unreadable subtree must fail closed (fresh/keep)");

    let _ = std::fs::remove_dir_all(&stale);
    let _ = std::fs::remove_dir_all(&fresh);
    let _ = std::fs::remove_dir_all(&locked);
}

#[test]
fn tmp_entry_must_remain_under_validated_root() {
    let root = guard_test_tmp("tmp_root_containment");
    let outside = guard_test_tmp("tmp_outside_containment");
    std::fs::create_dir_all(&root).expect("tmp root");
    std::fs::create_dir_all(&outside).expect("outside root");

    let validated = check_safe_tmp_root(&root).expect("tmp root should be valid");
    let result = check_safe_to_delete_tmp_entry(&outside, &validated, &[]);
    assert!(
        result.is_err(),
        "a candidate outside the validated tmp root must be rejected"
    );

    let _ = std::fs::remove_dir_all(&root);
    let _ = std::fs::remove_dir_all(&outside);
}

#[test]
fn guard_safe_delete_allows_paths_under_system_protected() {
    let tmp = guard_test_tmp("guard_safe_1");
    let target = tmp.join("target");
    std::fs::create_dir_all(&target).unwrap();
    let result = check_safe_to_delete_guard(&target, &[]);
    assert!(
        result.is_ok(),
        "guard safe delete should allow paths under /home (system-protected skipped)"
    );
    let _ = std::fs::remove_dir_all(&tmp);
}

#[test]
fn guard_safe_delete_blocks_user_protected() {
    let tmp = guard_test_tmp("guard_safe_2");
    let target = tmp.join("target");
    std::fs::create_dir_all(&target).unwrap();
    let user_protected = vec![tmp.display().to_string()];
    let result = check_safe_to_delete_guard(&target, &user_protected);
    assert!(
        result.is_err(),
        "guard safe delete should block user-protected paths"
    );
    let _ = std::fs::remove_dir_all(&tmp);
}

#[test]
fn guard_safe_delete_rejects_symlink() {
    let tmp = guard_test_tmp("guard_safe_3");
    let real = tmp.join("real_target");
    std::fs::create_dir_all(&real).unwrap();
    let link = tmp.join("link_target");
    let _ = std::fs::remove_file(&link);
    std::os::unix::fs::symlink(&real, &link).unwrap();
    let result = check_safe_to_delete_guard(&link, &[]);
    assert!(result.is_err(), "guard safe delete should reject symlinks");
    let _ = std::fs::remove_dir_all(&tmp);
}

#[test]
fn guard_safe_delete_rejects_exact_system_roots() {
    for prot in SYSTEM_PROTECTED {
        // Some build sandboxes (including Nix) intentionally omit host roots
        // such as /home.  A missing path is already safe to delete, while
        // every existing protected root must still be rejected here.
        if !Path::new(prot).exists() {
            continue;
        }
        let result = check_safe_to_delete_guard(Path::new(prot), &[]);
        assert!(
            result.is_err(),
            "guard safe delete should reject exact protected system root {prot}"
        );
    }
}

#[test]
fn check_safe_to_delete_rejects_log_symlink_before_truncate() {
    let tmp = guard_test_tmp("log_symlink");
    std::fs::create_dir_all(&tmp).unwrap();
    let real = tmp.join("real.log");
    let link = tmp.join("link.log");
    std::fs::write(&real, "line1\nline2\nline3\n").unwrap();
    std::os::unix::fs::symlink(&real, &link).unwrap();

    let result = check_safe_to_delete(&link, &[]);
    assert!(
        result.is_err(),
        "symlink log should be rejected before truncate"
    );
    assert_eq!(
        std::fs::read_to_string(&real).unwrap(),
        "line1\nline2\nline3\n"
    );

    let _ = std::fs::remove_dir_all(&tmp);
}

#[test]
fn proactive_cleanup_defaults() {
    assert_eq!(default_proactive_cleanup_percent(), 80);
    assert_eq!(default_auto_cleanup_interval_secs(), 1800);
    assert_eq!(default_rust_target_max_age_days(), 14);
    assert_eq!(default_rust_target_action_min_age_days(), 7);
    assert_eq!(default_proactive_cleanup_interval_cycles(), 120);
}

#[test]
fn action_cleanup_age_gate_lingering_definition() {
    // 2026-09-14: action-level cleanup must not touch active projects.
    const DAY: u64 = 86400;
    // Freshly built (hours old) is never a candidate.
    assert!(!rust_target_old_enough_for_action_cleanup(3600, 7));
    // Touched yesterday is still active.
    assert!(!rust_target_old_enough_for_action_cleanup(DAY, 7));
    // Exactly at the boundary counts as lingering.
    assert!(rust_target_old_enough_for_action_cleanup(7 * DAY, 7));
    // Untouched for weeks is a candidate.
    assert!(rust_target_old_enough_for_action_cleanup(30 * DAY, 7));
}

#[test]
fn action_cleanup_age_gate_zero_disables() {
    // 0 restores the pre-fix delete-anything posture for operators who
    // want it; normalize must not clamp it away.
    assert!(rust_target_old_enough_for_action_cleanup(0, 0));
    let policy = GuardPolicy {
        rust_target_action_min_age_days: 0,
        ..Default::default()
    };
    let mut policy = policy;
    normalize_guard_policy(&mut policy);
    assert_eq!(policy.rust_target_action_min_age_days, 0);
}

#[test]
fn normalize_proactive_cleanup_percent_bounded_by_action() {
    let policy = GuardPolicy {
        disk_action_percent: 85,
        proactive_cleanup_percent: 90,
        ..Default::default()
    };
    let mut policy = policy;
    normalize_guard_policy(&mut policy);
    assert!(
        policy.proactive_cleanup_percent < policy.disk_action_percent,
        "proactive_cleanup_percent must be below disk_action_percent"
    );
}

#[test]
fn normalize_rust_target_max_age_days_min_1() {
    let policy = GuardPolicy {
        rust_target_max_age_days: 0,
        ..Default::default()
    };
    let mut policy = policy;
    normalize_guard_policy(&mut policy);
    assert!(policy.rust_target_max_age_days >= 1);
}

#[test]
fn normalize_proactive_interval_min_1() {
    let policy = GuardPolicy {
        proactive_cleanup_interval_cycles: 0,
        ..Default::default()
    };
    let mut policy = policy;
    normalize_guard_policy(&mut policy);
    assert!(policy.proactive_cleanup_interval_cycles >= 1);
}

#[test]
fn normalize_auto_cleanup_interval_min_60() {
    let policy = GuardPolicy {
        auto_cleanup_interval_secs: 0,
        ..Default::default()
    };
    let mut policy = policy;
    normalize_guard_policy(&mut policy);
    assert!(policy.auto_cleanup_interval_secs >= 60);
}

#[test]
fn auto_cleanup_cadence_is_stateful() {
    let now = Instant::now();
    let mut state = GuardRuntimeState::default();
    assert!(auto_cleanup_due_at(&state, 1800, now));

    state.last_auto_cleanup = Some(now - Duration::from_secs(1799));
    assert!(!auto_cleanup_due_at(&state, 1800, now));

    state.last_auto_cleanup = Some(now - Duration::from_secs(1800));
    assert!(auto_cleanup_due_at(&state, 1800, now));
}

#[test]
fn guard_runtime_state_default_cycle_zero() {
    let state = GuardRuntimeState::default();
    assert_eq!(state.guard_cycle, 0);
    assert!(state.last_proactive_cleanup.is_none());
    assert!(state.last_auto_cleanup.is_none());
}

#[test]
fn target_dir_info_has_mtime() {
    let info = TargetDirInfo {
        path: PathBuf::from("/tmp/test/target"),
        bytes: 1024,
        mtime_secs_ago: 86400 * 15,
    };
    assert_eq!(info.mtime_secs_ago, 86400 * 15);
    assert_eq!(info.bytes, 1024);
}

#[test]
fn truncate_log_preserves_open_writer_inode() {
    let td =
        std::env::temp_dir().join(format!("dracon-system-truncate-log-{}", std::process::id()));
    let _ = std::fs::remove_dir_all(&td);
    std::fs::create_dir_all(&td).expect("temp dir");
    let path = td.join("guard.log");
    std::fs::write(&path, "header\nbody-one\nbody-two\n").expect("log");

    // Keep an append handle open as a long-running logger would. A rename-
    // based truncation leaves this handle on an unlinked inode; in-place
    // truncation keeps subsequent lines visible at the original path.
    let mut writer = std::fs::OpenOptions::new()
        .append(true)
        .open(&path)
        .expect("open writer");
    let reclaimed = truncate_log_file(&path, 12, 1).expect("truncate");
    assert!(reclaimed > 0);
    use std::io::Write;
    writer.write_all(b"tail\n").expect("append after truncate");

    let content = std::fs::read_to_string(&path).expect("read log");
    assert!(content.starts_with("header\n"));
    assert!(
        content.contains("tail\n"),
        "writer output must remain visible"
    );
    assert!(!content.contains("body-two"));
    let _ = std::fs::remove_dir_all(&td);
}

#[test]
fn truncate_log_preserves_non_utf8_lines_verbatim() {
    // 2026-10-02 (audit L8): `BufRead::lines()` yields Err on invalid
    // UTF-8 and both loops STOPPED there, silently dropping the bad
    // line AND everything after it. Byte-split lines preserve them.
    let td = std::env::temp_dir().join(format!(
        "dracon-system-truncate-utf8-{}",
        std::process::id()
    ));
    let _ = std::fs::remove_dir_all(&td);
    std::fs::create_dir_all(&td).expect("temp dir");
    let path = td.join("app.log");
    let mut raw = b"header\nok-line\n".to_vec();
    raw.extend_from_slice(b"bad-\xff\xfe-bytes\n");
    raw.extend_from_slice(b"after-bad-line\n");
    raw.extend_from_slice(&[b'x'; 200]);
    std::fs::write(&path, &raw).expect("log");

    let reclaimed = truncate_log_file(&path, 64, 1).expect("truncate");
    assert!(reclaimed > 0);
    let out = std::fs::read(&path).expect("read log");
    assert!(out.starts_with(b"header\n"), "header must survive: {out:?}");
    assert!(
        out.windows(8).any(|w| w == b"bad-\xff\xfe-b"),
        "non-UTF8 line must be preserved verbatim: {out:?}"
    );
    assert!(
        out.windows(14).any(|w| w == b"after-bad-line"),
        "lines AFTER the bad line must survive (old code stopped): {out:?}"
    );
    assert!(
        out.len() as u64 <= 64,
        "budget must hold: {} bytes",
        out.len()
    );
    let _ = std::fs::remove_dir_all(&td);
}

#[test]
fn truncate_log_handles_missing_trailing_newline_and_blank_lines() {
    // Byte-split must not invent or drop lines at the edges: a missing
    // trailing newline gains one terminator (as before), blank lines
    // survive, and the terminator artifact is never kept as a line.
    let td = std::env::temp_dir().join(format!(
        "dracon-system-truncate-edge-{}",
        std::process::id()
    ));
    let _ = std::fs::remove_dir_all(&td);
    std::fs::create_dir_all(&td).expect("temp dir");
    let path = td.join("app.log");
    std::fs::write(&path, b"h1\n\nbody-no-newline").expect("log");

    let reclaimed = truncate_log_file(&path, 10, 1).expect("truncate");
    assert!(reclaimed > 0);
    let out = std::fs::read(&path).expect("read log");
    assert_eq!(
        out, b"h1\n\n",
        "header + blank line, terminator artifact dropped"
    );
    let _ = std::fs::remove_dir_all(&td);
}

#[tokio::test]
async fn clean_old_node_modules_counts_nested_tree_once() {
    let td = std::env::temp_dir().join(format!(
        "dracon-system-node-modules-{}-{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    let outer = td.join("project/node_modules");
    let inner = outer.join("nested/node_modules");
    std::fs::create_dir_all(&inner).expect("create nested node_modules");
    std::fs::write(outer.join("outer.bin"), vec![b'o'; 17]).expect("write outer file");
    std::fs::write(inner.join("inner.bin"), vec![b'i'; 29]).expect("write inner file");

    let expected = get_dir_size(&outer).await.expect("measure outer tree");
    let (reclaimed, cleaned) = clean_old_node_modules(
        std::slice::from_ref(&td),
        0,
        false,
        &[],
        false,
        std::path::Path::new(""),
    )
    .await
    .expect("dry-run node_modules cleanup");

    assert_eq!(
        cleaned.len(),
        1,
        "nested node_modules must not be listed twice"
    );
    assert_eq!(reclaimed, expected, "outer tree must be counted once");
    let _ = std::fs::remove_dir_all(&td);
}

#[test]
fn rust_build_process_detection_covers_long_lived_tooling() {
    // Pre-existing classes stay detected (substring + exact).
    assert!(is_rust_build_process("cargo"));
    assert!(is_rust_build_process("cargo-build"));
    assert!(is_rust_build_process("rustc"));
    assert!(is_rust_build_process("rustc-1.94.1"));
    assert!(is_rust_build_process("clippy-driver"));
    // Long-lived Rust tooling added 2026-08-21 (detection gap): a running
    // analyzer or watch session holds the target dir just like a build.
    assert!(is_rust_build_process("rust-analyzer"));
    assert!(is_rust_build_process("cargo-watch"));
    // Unrelated processes must not match.
    assert!(!is_rust_build_process(""));
    assert!(!is_rust_build_process("firefox"));
    assert!(!is_rust_build_process("dracon-sync"));
    assert!(!is_rust_build_process("node"));
}

#[test]
fn package_cache_process_detection_covers_direct_and_wrapped_tools() {
    let proc_root = unique_test_home("package_proc");
    fs::create_dir_all(proc_root.join("self")).expect("create proc fixture");

    for (pid, cmdline) in [
        (101, b"/usr/bin/python3\0-m\0pip\0install\0".as_slice()),
        (
            102,
            b"/usr/bin/node\0/usr/lib/npm/npm-cli.js\0install\0".as_slice(),
        ),
        (103, b"/bin/sh\0-c\0go\0build\0./...\0".as_slice()),
    ] {
        let process_dir = proc_root.join(pid.to_string());
        fs::create_dir_all(&process_dir).expect("create process fixture");
        fs::write(process_dir.join("cmdline"), cmdline).expect("write cmdline fixture");
    }

    let active = detect_active_package_manager_operations_from(
        "100 cargo\n101 python3\n102 node\n103 sh\n104 firefox\n",
        &proc_root,
    )
    .expect("process metadata fixture should be readable");
    assert_eq!(active.len(), 4);
    assert!(active.contains(&PackageCacheKind::Cargo));
    assert!(active.contains(&PackageCacheKind::Npm));
    assert!(active.contains(&PackageCacheKind::Pip));
    assert!(active.contains(&PackageCacheKind::Go));

    let _ = fs::remove_dir_all(proc_root);
}

#[cfg(unix)]
#[test]
fn package_cache_process_detection_fails_closed_on_unreadable_cmdline() {
    let proc_root = unique_test_home("package_proc_unreadable");
    let process_dir = proc_root.join("201");
    fs::create_dir_all(proc_root.join("self")).expect("create proc fixture");
    fs::create_dir_all(&process_dir).expect("create process fixture");
    let cmdline = process_dir.join("cmdline");
    fs::write(
        &cmdline,
        b"/usr/bin/node\0/usr/lib/npm/npm-cli.js\0install\0",
    )
    .expect("write cmdline fixture");
    fs::set_permissions(&cmdline, fs::Permissions::from_mode(0o000))
        .expect("make cmdline unreadable");

    let result = detect_active_package_manager_operations_from("201 python3\n", &proc_root);
    assert!(
        result.is_err(),
        "an unreadable wrapper command line must abort cache protection"
    );

    let _ = fs::remove_dir_all(proc_root);
}

#[test]
fn package_cache_process_detection_fails_closed_on_missing_cmdline() {
    let proc_root = unique_test_home("package_proc_missing_cmdline");
    fs::create_dir_all(proc_root.join("self")).expect("create proc fixture");
    fs::create_dir_all(proc_root.join("202")).expect("create process fixture");

    let result = detect_active_package_manager_operations_from("202 python3\n", &proc_root);
    assert!(
        result.is_err(),
        "an existing process without cmdline metadata must abort cache protection"
    );

    let _ = fs::remove_dir_all(proc_root);
}

#[test]
fn package_cache_process_detection_rejects_malformed_ps_output() {
    let proc_root = unique_test_home("package_proc_malformed");
    fs::create_dir_all(proc_root.join("self")).expect("create proc fixture");

    let result = detect_active_package_manager_operations_from("not-a-process\n", &proc_root);
    assert!(
        result.is_err(),
        "partial process metadata must abort cache protection"
    );

    let _ = fs::remove_dir_all(proc_root);
}

#[cfg(unix)]
#[tokio::test]
async fn package_cache_process_listing_failure_is_an_error() {
    let _fixture_exec = fixture_exec_guard();
    let proc_root = unique_test_home("package_proc_failed_ps");
    let bin_root = unique_test_home("package_ps_failed");
    fs::create_dir_all(proc_root.join("self")).expect("create proc fixture");
    fs::create_dir_all(&bin_root).expect("create bin fixture");
    let ps = bin_root.join("ps");
    write_test_script(&ps, "exit 7");

    let result = detect_active_package_manager_operations_with(&ps, &proc_root).await;
    assert!(result.is_err(), "a failed ps command must fail closed");

    let _ = fs::remove_dir_all(proc_root);
    let _ = fs::remove_dir_all(bin_root);
}

#[tokio::test]
async fn package_cache_apply_refuses_all_cache_deletion() {
    let home = unique_test_home("package_cache_protection");
    let targets = [
        (PackageCacheKind::Cargo, ".cargo/registry/cache"),
        (PackageCacheKind::Npm, ".npm"),
        (PackageCacheKind::Pip, ".cache/pip"),
        (PackageCacheKind::Go, ".cache/go-build"),
    ];
    let mut active = HashSet::new();
    for (kind, relative) in targets {
        let cache = home.join(relative);
        fs::create_dir_all(&cache).expect("create cache fixture");
        fs::write(cache.join("in-use.bin"), b"must survive").expect("write cache fixture");
        active.insert(kind);
    }

    let (reclaimed, cleaned) =
        clean_package_caches_at(&home, true, true, true, true, true, &active)
            .await
            .expect("protected cache cleanup");
    assert_eq!(
        reclaimed, 0,
        "apply cache deletion is disabled, so caches must not be counted as reclaimed"
    );
    assert!(
        cleaned.is_empty(),
        "apply cache deletion is disabled, so caches must not be reported cleaned"
    );
    for (_, relative) in targets {
        assert!(
            home.join(relative).exists(),
            "package cache {relative} must survive apply"
        );
    }

    let _ = fs::remove_dir_all(home);
}

#[tokio::test]
async fn package_cache_apply_refuses_uncoordinated_start_after_snapshot() {
    let home = unique_test_home("package_cache_adversarial_start");
    let proc_root = unique_test_home("package_proc_adversarial_start");
    fs::create_dir_all(proc_root.join("self")).expect("create proc fixture");
    let initial_process = proc_root.join("400");
    fs::create_dir_all(&initial_process).expect("create initial process fixture");
    fs::write(initial_process.join("cmdline"), b"/usr/bin/sleep\0").expect("write initial cmdline");

    // The first snapshot sees no package operation.
    let initial = detect_active_package_manager_operations_from("400 sleep\n", &proc_root)
        .expect("initial process snapshot");
    assert!(initial.is_empty());

    // An npm operation starts after that snapshot, exactly in the window the
    // old detector-then-remove implementation could not coordinate.
    let new_process = proc_root.join("401");
    fs::create_dir_all(&new_process).expect("create adversarial process fixture");
    fs::write(
        new_process.join("cmdline"),
        b"/usr/bin/node\0/usr/lib/npm/npm-cli.js\0install\0",
    )
    .expect("write adversarial cmdline");

    let npm_cache = home.join(".npm");
    fs::create_dir_all(&npm_cache).expect("create npm cache fixture");
    fs::write(npm_cache.join("in-use.bin"), b"must survive").expect("write cache fixture");

    // Apply refuses all package-cache deletion because no lock shared with
    // external package managers can make this lifecycle-safe.
    let result = clean_package_caches_at(&home, false, true, false, false, true, &initial)
        .await
        .expect("uncoordinated apply must be a safe no-op");
    assert_eq!(result, (0, Vec::new()));
    assert!(
        npm_cache.exists(),
        "an operation starting after detection must not race deletion"
    );

    let _ = fs::remove_dir_all(home);
    let _ = fs::remove_dir_all(proc_root);
}

#[tokio::test]
async fn active_package_operations_skip_matching_cache_estimates_in_dry_run() {
    let home = unique_test_home("package_cache_dry_run");
    let cargo_cache = home.join(".cargo/registry/cache");
    let npm_cache = home.join(".npm");
    fs::create_dir_all(&cargo_cache).expect("create cargo cache fixture");
    fs::create_dir_all(&npm_cache).expect("create npm cache fixture");
    fs::write(cargo_cache.join("cargo.bin"), b"cargo").expect("write cargo fixture");
    fs::write(npm_cache.join("npm.bin"), b"npm").expect("write npm fixture");
    let active = HashSet::from([PackageCacheKind::Npm]);

    let result = clean_package_caches_at(&home, true, true, false, false, false, &active)
        .await
        .expect("dry-run cache estimate");
    assert_eq!(result.0, 5);
    assert!(result.1.iter().any(|entry| entry.contains("cargo")));
    assert!(!result.1.iter().any(|entry| entry.contains("npm")));
    assert!(cargo_cache.exists() && npm_cache.exists());

    let _ = fs::remove_dir_all(home);
}

#[test]
fn storage_cleanup_apply_accepts_home_artifact_dirs_and_refuses_system_roots() {
    // Regression (2026-08-21): `storage --cleanup --apply` used the strict
    // classifier, refusing EVERY path under /home ("under system root
    // /home") — i.e. every real candidate on a workstation — while the
    // guard's auto-cleanup deleted the same class of paths. The apply path
    // must accept an artifact dir under /home …
    let home = dirs::home_dir().expect("home dir available");
    let fixture = home.join(format!(
        ".dracon_storage_apply_test_{}_{}",
        std::process::id(),
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap()
            .as_nanos()
    ));
    std::fs::create_dir_all(&fixture).unwrap();
    let validated = validate_storage_cleanup_path(&fixture, &[]);
    assert!(
        validated.is_ok(),
        "apply must accept artifact dirs under /home: {:?}",
        validated.err()
    );
    let _ = std::fs::remove_dir_all(&fixture);

    // … while exact system roots stay refused.
    for prot in SYSTEM_PROTECTED {
        if !Path::new(prot).exists() {
            continue;
        }
        assert!(
            validate_storage_cleanup_path(Path::new(prot), &[]).is_err(),
            "apply must refuse exact system root {prot}"
        );
    }

    // The strict classifier remains strict: the two rule sets are
    // intentionally different, this pins WHY the apply path must not use it.
    if Path::new("/home").exists() {
        assert!(check_safe_to_delete(Path::new("/home"), &[]).is_err());
    }
}

#[test]
fn storage_cache_hotspots_share_package_apply_safety_gate() {
    assert!(storage_hotspot_apply_is_blocked("cache", true));
    assert!(!storage_hotspot_apply_is_blocked("cache", false));
    assert!(!storage_hotspot_apply_is_blocked("node-deps", true));
}

#[test]
fn filter_selectable_cleanup_kinds_drops_git_db_and_keeps_artifact_kinds() {
    // Audit M2 (2026-08-21): git-db is report-only — it must be filtered
    // out of every cleanup selection source.
    let requested: HashSet<String> = ["git-db", "rust-build", "node-deps"]
        .iter()
        .map(|s| s.to_string())
        .collect();
    let (kept, excluded) = filter_selectable_cleanup_kinds(requested);
    assert!(kept.contains("rust-build") && kept.contains("node-deps"));
    assert!(
        !kept.contains("git-db"),
        "git-db must never survive filtering"
    );
    assert_eq!(excluded, vec!["git-db".to_string()]);

    // Absent kind → no exclusions reported; empty request stays empty.
    let (kept, excluded) =
        filter_selectable_cleanup_kinds(["cache".to_string()].into_iter().collect());
    assert!(excluded.is_empty() && kept == ["cache".to_string()].into_iter().collect());
    let (kept, excluded) = filter_selectable_cleanup_kinds(HashSet::new());
    assert!(kept.is_empty() && excluded.is_empty());

    // Every entry in NON_CLEANUP_KINDS is actually enforced by the filter.
    for kind in NON_CLEANUP_KINDS {
        let (kept, excluded) =
            filter_selectable_cleanup_kinds([(kind.to_string())].into_iter().collect());
        assert!(!kept.contains(*kind));
        assert_eq!(excluded, vec![kind.to_string()]);
    }
}

#[test]
fn storage_cleanup_apply_refuses_git_database_dirs() {
    // Audit M2 backstop: even if a .git path ever reaches
    // validate_storage_cleanup_path directly, it must be refused — git
    // never tracks its own database, so the allow_tracked gate cannot
    // protect project history from remove_dir_all.
    let home = dirs::home_dir().expect("home dir available");
    let fixture = home.join(format!(
        ".dracon_storage_gitdb_test_{}_{}",
        std::process::id(),
        std::time::SystemTime::now()
            .duration_since(std::time::UNIX_EPOCH)
            .unwrap()
            .as_nanos()
    ));
    let git_db = fixture.join("project").join(".git");
    std::fs::create_dir_all(&git_db).unwrap();
    let validated = validate_storage_cleanup_path(&git_db, &[]);
    let _ = std::fs::remove_dir_all(&fixture);
    let err = validated.expect_err("apply must refuse a .git directory");
    let msg = format!("{err:#}");
    assert!(
        msg.contains("git database"),
        "refusal must name the reason, got: {msg}"
    );
}

#[test]
fn storage_cleanup_apply_refuses_git_database_subpaths() {
    // ADDED 2026-09-09 (audit F38): the M2 backstop matched only
    // `file_name() == ".git"`, so `/repo/.git/objects` passed. Any
    // path with a `.git` component must be refused.
    let err = validate_storage_cleanup_path(Path::new("/home/user/Dev/project/.git/objects"), &[])
        .expect_err("apply must refuse paths under a .git directory");
    assert!(format!("{err:#}").contains("git database"));
}

#[test]
fn normalize_early_warn_clamped_to_warn() {
    // ADDED 2026-09-09 (audit F43): early > warn made the early band
    // (`used >= early && used < warn`) permanently empty.
    let policy = GuardPolicy {
        disk_early_warn_percent: 95,
        disk_warn_percent: 75,
        ..Default::default()
    };
    let mut policy = policy;
    normalize_guard_policy(&mut policy);
    assert!(
        policy.disk_early_warn_percent <= policy.disk_warn_percent,
        "early warn must not exceed warn, got early={} warn={}",
        policy.disk_early_warn_percent,
        policy.disk_warn_percent
    );
}

/// The 2026-10-01 audit: when the freeze marker disappears out from under the
/// daemon (the freeze watchdog's 30m auto-clear, or an operator running
/// `dracon-sync resume`), the in-memory flag used to stay true forever, so
/// every later pass re-entered the unfreeze branch and logged a remove failure
/// for a file that no longer existed. The flag must resync instead.
#[test]
fn sync_frozen_resyncs_when_the_marker_disappears_externally() {
    let fixture = std::env::temp_dir().join(format!(
        "dracon-system-freeze-resync-{}-{}",
        std::process::id(),
        SystemTime::now()
            .duration_since(UNIX_EPOCH)
            .expect("time")
            .as_nanos()
    ));
    fs::create_dir_all(&fixture).expect("create fixture");
    let marker = fixture.join("freeze");

    let guard = GuardPolicy {
        freeze_sync_at_action: true,
        sync_freeze_marker: marker.display().to_string(),
        unfreeze_below_percent: 70,
        ..GuardPolicy::default()
    };

    // Action level with no marker yet: the guard writes it and claims the
    // freeze.
    let mut sync_frozen = false;
    manage_sync_freeze(&guard, 95, "action", &mut sync_frozen);
    assert!(sync_frozen, "guard must claim a freeze it just performed");
    assert!(marker.exists(), "guard must write its own marker");

    // Pressure recovers but the marker is still there: normal unfreeze.
    manage_sync_freeze(&guard, 68, "ok", &mut sync_frozen);
    assert!(!sync_frozen);
    assert!(!marker.exists(), "guard must remove its own marker");

    // Freeze again, then let something ELSE clear the marker, then recover.
    manage_sync_freeze(&guard, 95, "action", &mut sync_frozen);
    assert!(sync_frozen);
    fs::remove_file(&marker).expect("simulate an external clear");

    manage_sync_freeze(&guard, 68, "ok", &mut sync_frozen);
    assert!(
        !sync_frozen,
        "an externally cleared marker must resync the in-memory flag"
    );

    fs::remove_dir_all(fixture).expect("remove fixture");
}

/// The 2026-10-01 audit: `cleanup_stale_cooldowns` used
/// `Instant::now() - Duration::from_secs(cooldown * 2)`, and std PANICS on that
/// subtraction once the duration underflows the clock's representation. A
/// policy cooldown at 2^63 seconds or more killed the daemon on its first
/// pass. The subtraction must not be able to panic for any u64.
#[test]
fn cleanup_stale_cooldowns_survives_an_extreme_cooldown() {
    let mut state = GuardRuntimeState::default();
    state.notify_cooldowns.insert(
        "pid-1".to_string(),
        Instant::now() - Duration::from_secs(60),
    );
    // u64::MAX saturates at 2x itself and underflows Instant by construction.
    cleanup_stale_cooldowns(&mut state, u64::MAX);
    // u64::MAX/2 * 2 == u64::MAX - 1, the largest value that still subtracts.
    cleanup_stale_cooldowns(&mut state, u64::MAX / 2);
    // A normal cooldown still prunes an entry older than 2x the window.
    state.notify_cooldowns.insert(
        "pid-2".to_string(),
        Instant::now() - Duration::from_secs(600),
    );
    cleanup_stale_cooldowns(&mut state, 5);
    assert!(
        !state.notify_cooldowns.contains_key("pid-2"),
        "a 600s-old entry must be pruned at a 5s cooldown"
    );
}

#[test]
fn guard_runtime_state_default_no_expiry_yet() {
    let state = GuardRuntimeState::default();
    assert!(state.last_quarantine_expire.is_none());
}

#[test]
fn quarantine_expire_interval_zero_is_never_clamped() {
    // 0 means "never expire automatically". A floor!() here would silently
    // re-enable unattended deletion for an operator who explicitly opted out,
    // which is the opposite failure from the one a floor guards against.
    let mut policy = GuardPolicy {
        quarantine_expire_interval_secs: 0,
        ..Default::default()
    };
    normalize_guard_policy(&mut policy);
    assert_eq!(policy.quarantine_expire_interval_secs, 0);
}

#[test]
fn quarantine_expire_interval_defaults_to_a_day() {
    let policy = GuardPolicy::default();
    assert_eq!(policy.quarantine_expire_interval_secs, 86_400);
}

/// The two opt-outs that stop `maybe_expire_quarantine` from deleting anything.
/// Both are about not deleting what the operator did not ask to be deleted:
/// interval 0 disables the unattended path, and an unarmed
/// `clean_quarantine_first` means any entries present came from a human running
/// `quarantine move` by hand.
#[test]
fn maybe_expire_quarantine_respects_both_opt_outs() {
    let root = std::env::temp_dir().join(format!(
        "dracon-expire-gate-{}-{}",
        std::process::id(),
        "gate"
    ));
    let _ = std::fs::remove_dir_all(&root);
    let src = root.join("work").join("proj");
    std::fs::create_dir_all(&src).unwrap();
    std::fs::write(src.join("a.txt"), b"hello").unwrap();
    let qdir = root.join("q");
    // Move it in, then age the manifest far past any TTL so expiry WOULD act.
    let manifest = quarantine_move(&src, &qdir, &[]).unwrap();
    let aged = QuarantineManifest {
        name: manifest.name.clone(),
        origin: manifest.origin.clone(),
        moved_at_unix: now_unix().saturating_sub(400 * 86_400),
        bytes: manifest.bytes,
        files: manifest.files,
    };
    std::fs::write(
        qdir.join(&manifest.name).join(".quarantine.json"),
        serde_json::to_string_pretty(&aged).unwrap(),
    )
    .unwrap();
    let entry_dir = qdir.join(&manifest.name);

    let armed = GuardPolicy {
        quarantine_dir: qdir.display().to_string(),
        clean_quarantine_first: true,
        quarantine_ttl_days: 30,
        quarantine_expire_interval_secs: 0,
        ..Default::default()
    };
    let mut state = GuardRuntimeState::default();

    // Opt-out 1: interval 0.
    maybe_expire_quarantine(&armed, &mut state, Instant::now());
    assert!(
        entry_dir.exists(),
        "interval 0 must disable the unattended expiry entirely"
    );

    // Opt-out 2: quarantine not armed, interval positive.
    let unarmed = GuardPolicy {
        quarantine_expire_interval_secs: 86_400,
        clean_quarantine_first: false,
        ..armed.clone()
    };
    let mut state2 = GuardRuntimeState::default();
    maybe_expire_quarantine(&unarmed, &mut state2, Instant::now());
    assert!(
        entry_dir.exists(),
        "entries from a manual `quarantine move` must not become an unattended delete"
    );

    // Neither opt-out: the entry is past TTL and must actually be removed, and
    // the cooldown must be stamped so the next pass is a no-op.
    // (auto_cleanup_apply must be armed: audit M5 made expiry honor it.)
    let active = GuardPolicy {
        quarantine_expire_interval_secs: 86_400,
        auto_cleanup_apply: true,
        ..armed.clone()
    };
    let mut state3 = GuardRuntimeState::default();
    maybe_expire_quarantine(&active, &mut state3, Instant::now());
    assert!(!entry_dir.exists(), "a past-TTL entry must be expired");
    assert!(state3.last_quarantine_expire.is_some());

    // And the cooldown holds: a second immediate pass must not re-scan. Proven
    // by re-creating an expired entry and confirming it survives the next pass.
    let src2 = root.join("work2").join("proj");
    std::fs::create_dir_all(&src2).unwrap();
    std::fs::write(src2.join("b.txt"), b"world!").unwrap();
    let m2 = quarantine_move(&src2, &qdir, &[]).unwrap();
    let aged2 = QuarantineManifest {
        name: m2.name.clone(),
        origin: m2.origin.clone(),
        moved_at_unix: now_unix().saturating_sub(400 * 86_400),
        bytes: m2.bytes,
        files: m2.files,
    };
    std::fs::write(
        qdir.join(&m2.name).join(".quarantine.json"),
        serde_json::to_string_pretty(&aged2).unwrap(),
    )
    .unwrap();
    maybe_expire_quarantine(&active, &mut state3, Instant::now());
    assert!(
        qdir.join(&m2.name).exists(),
        "the cooldown must stop a per-pass rescan"
    );

    let _ = std::fs::remove_dir_all(&root);
}

/// Audit M5 (2026-10-02): `maybe_expire_quarantine` must honor
/// `auto_cleanup_apply` — previously it hardcoded apply=true, arming
/// unattended deletes for operators who believed `false` meant dry-run.
/// A dry-run pass scans and reports but removes nothing.
#[test]
fn maybe_expire_quarantine_dry_run_deletes_nothing() {
    let root = std::env::temp_dir().join(format!(
        "dracon-expire-dryrun-{}-{}",
        std::process::id(),
        "dryrun"
    ));
    let _ = std::fs::remove_dir_all(&root);
    let src = root.join("work").join("proj");
    std::fs::create_dir_all(&src).unwrap();
    std::fs::write(src.join("a.txt"), b"hello").unwrap();
    let qdir = root.join("q");
    let manifest = quarantine_move(&src, &qdir, &[]).unwrap();
    let aged = QuarantineManifest {
        name: manifest.name.clone(),
        origin: manifest.origin.clone(),
        moved_at_unix: now_unix().saturating_sub(400 * 86_400),
        bytes: manifest.bytes,
        files: manifest.files,
    };
    std::fs::write(
        qdir.join(&manifest.name).join(".quarantine.json"),
        serde_json::to_string_pretty(&aged).unwrap(),
    )
    .unwrap();
    let entry_dir = qdir.join(&manifest.name);

    // Fully armed EXCEPT auto_cleanup_apply: the past-TTL entry must
    // survive, and the cooldown must still stamp (no rescan spam).
    let dry = GuardPolicy {
        quarantine_dir: qdir.display().to_string(),
        clean_quarantine_first: true,
        quarantine_ttl_days: 30,
        quarantine_expire_interval_secs: 86_400,
        auto_cleanup_apply: false,
        ..Default::default()
    };
    let mut state = GuardRuntimeState::default();
    maybe_expire_quarantine(&dry, &mut state, Instant::now());
    assert!(
        entry_dir.exists(),
        "dry-run expiry must not delete a past-TTL entry"
    );
    assert!(state.last_quarantine_expire.is_some());

    let _ = std::fs::remove_dir_all(&root);
}

/// The objective requires expiry to "tolerate absurd interval values without
/// panicking", and until now that held only by inspection. `cooldown_due` is the
/// only place the interval is used, and the classic failure here is
/// `Instant - Duration`, which PANICS on underflow rather than saturating —
/// `cleanup_stale_cooldowns` already guards that with `checked_sub` and a
/// comment saying a large policy must not kill the daemon. The same reasoning
/// has to hold for the expiry cadence, so it is pinned rather than reasoned.
///
/// Two halves, because the cadence has two distinct behaviours and conflating
/// them hides both:
///
/// - With a cooldown already recorded, an absurd interval must read as NOT
///   due, so a fat-fingered policy cannot turn into a per-pass delete storm.
/// - With no prior run, the very first pass is due for ANY interval — that is
///   the documented "nothing has run yet" case — so an absurd value must still
///   complete normally instead of panicking.
///
/// Values include `u64::MAX`: any `*2` overflows, any subtraction underflows,
/// and `Duration::from_secs` on it sits on the boundary `Instant` uses.
const ABSURD_INTERVALS: &[u64] = &[
    0,
    1,
    60,
    86_400,
    u32::MAX as u64,
    u64::MAX / 2,
    u64::MAX - 1,
    u64::MAX,
];

/// Build an expired entry in a fresh quarantine root, returning
/// `(root, quarantine_dir, entry_name)`.
fn absurd_interval_fixture(tag: &str) -> (PathBuf, PathBuf, String) {
    let root = std::env::temp_dir().join(format!(
        "dracon-expire-absurd-{}-{}",
        std::process::id(),
        tag
    ));
    let _ = std::fs::remove_dir_all(&root);
    let qdir = root.join("q");
    let src = root.join("work").join("proj");
    std::fs::create_dir_all(&src).unwrap();
    std::fs::write(src.join("a.txt"), b"hello").unwrap();
    let manifest = quarantine_move(&src, &qdir, &[]).unwrap();
    let aged = QuarantineManifest {
        name: manifest.name.clone(),
        origin: manifest.origin.clone(),
        moved_at_unix: now_unix().saturating_sub(400 * 86_400),
        bytes: manifest.bytes,
        files: manifest.files,
    };
    std::fs::write(
        qdir.join(&manifest.name).join(".quarantine.json"),
        serde_json::to_string_pretty(&aged).unwrap(),
    )
    .unwrap();
    (root, qdir, manifest.name)
}

fn expiry_policy(qdir: &Path, interval: u64) -> GuardPolicy {
    GuardPolicy {
        quarantine_dir: qdir.display().to_string(),
        clean_quarantine_first: true,
        quarantine_ttl_days: 30,
        quarantine_expire_interval_secs: interval,
        // FIXED 2026-10-03 (audit M5 follow-up): expiry honors
        // `auto_cleanup_apply` — the default false is dry-run (pinned
        // by `maybe_expire_quarantine_dry_run_deletes_nothing`), so
        // these interval tests must arm apply to exercise deletion.
        // This also strengthens the not-due test: the cooldown gates
        // even a fully armed policy.
        auto_cleanup_apply: true,
        ..Default::default()
    }
}

#[test]
fn absurd_expiry_intervals_are_not_due_once_the_cooldown_exists() {
    for &interval in ABSURD_INTERVALS {
        let (root, qdir, name) = absurd_interval_fixture("not-due");
        // Stamped 2s ago, not `now`: that is what pins the `.max(60)` floor.
        // With a zero-aged stamp even a dropped floor still reads "not due", so
        // the floor would go unpinned; 2 seconds is past a typo'd interval of 1
        // but far short of 60, so the floor is the only thing keeping this
        // entry alive. Without it a policy of 1 becomes a per-pass delete.
        let mut state = GuardRuntimeState {
            last_quarantine_expire: Some(Instant::now() - Duration::from_secs(2)),
            ..Default::default()
        };
        maybe_expire_quarantine(&expiry_policy(&qdir, interval), &mut state, Instant::now());
        if interval != 0 {
            assert!(
                qdir.join(&name).exists(),
                "a non-zero interval must not be due immediately (interval={interval})"
            );
        }
        let _ = std::fs::remove_dir_all(&root);
    }
}

#[test]
fn absurd_expiry_intervals_complete_normally_on_the_first_pass() {
    let (root, qdir, name) = absurd_interval_fixture("first-pass");
    // No prior run: the first pass is due for every interval, so this exercises
    // the actual expiry path with the knob at its extremes. It must not panic,
    // and it must expire the entry rather than silently doing nothing.
    for &interval in ABSURD_INTERVALS {
        let mut state = GuardRuntimeState::default();
        maybe_expire_quarantine(&expiry_policy(&qdir, interval), &mut state, Instant::now());
        if interval == 0 {
            // Opting out returns before the cooldown is touched, so the
            // runtime state is left exactly as an operator who never configured
            // the knob would leave it.
            assert!(
                state.last_quarantine_expire.is_none(),
                "interval 0 must leave the runtime state untouched"
            );
        } else {
            assert!(
                state.last_quarantine_expire.is_some(),
                "the cooldown must be stamped (interval={interval})"
            );
        }
    }
    assert!(
        !qdir.join(&name).exists(),
        "the first pass must still expire a past-TTL entry"
    );
    let _ = std::fs::remove_dir_all(&root);
}

/// The dry-run preview listed every user-protected tree as reclaimable and only
/// refused it at apply time, because the refusal lived inside the `if apply`
/// block. On this host that overstated reclaimable space by ~66 GiB and
/// misled the operator twice in one session.
///
/// This drives `auto_cleanup_rust_targets` in dry-run mode over a real tempdir
/// tree, because the preview is the thing that was wrong — asserting on the
/// apply path would pass even with the bug in place.
#[tokio::test]
async fn dry_run_preview_omits_user_protected_targets() {
    let root = std::env::temp_dir().join(format!(
        "dracon-preview-protected-{}-{}",
        std::process::id(),
        "t"
    ));
    let _ = std::fs::remove_dir_all(&root);

    // One target under a protected project, one under an unprotected sibling.
    // Both large enough to clear the min-size gate so the scan really sees them.
    let protected_proj = root.join("protected-proj");
    let free_proj = root.join("free-proj");
    for proj in [&protected_proj, &free_proj] {
        // The scanner only treats a `target/` as a Rust build cache when the
        // project parent has a Cargo.toml, so the fixture needs one or both
        // dirs are invisible and the test passes vacuously.
        std::fs::create_dir_all(proj).unwrap();
        std::fs::write(
            proj.join("Cargo.toml"),
            b"[package]\nname = \"probe\"\nversion = \"0.0.0\"\nedition = \"2021\"\n",
        )
        .unwrap();
        let target = proj.join("target");
        std::fs::create_dir_all(target.join("debug")).unwrap();
        std::fs::write(
            target.join("debug").join("blob.bin"),
            vec![7u8; 4 * 1024 * 1024],
        )
        .unwrap();
        // The scan has a "modified <60s ago means an active build" backstop, so
        // a fixture written a moment ago is protected on that ground alone and
        // the candidate list comes back empty. Age the tree instead of trying
        // to relax the guard: `touch -d` needs no new dependency, and the
        // alternative (setting mtime from Rust) needs a crate this repo does
        // not carry.
        let aged = std::process::Command::new("touch")
            .arg("-d")
            .arg("20 days ago")
            .arg(&target)
            .output()
            .expect("touch must be available to age the fixture");
        assert!(
            aged.status.success(),
            "touch failed: {}",
            String::from_utf8_lossy(&aged.stderr)
        );
    }

    let guard = GuardPolicy {
        rust_search_roots: root.display().to_string(),
        cleanup_min_size_mb: 1,
        protected_paths: vec![protected_proj.display().to_string()],
        // Nothing here may touch the real quarantine dir.
        quarantine_dir: root.join("q").display().to_string(),
        clean_quarantine_first: true,
        ..Default::default()
    };

    let mut runtime = GuardRuntimeState::default();
    // apply = false: this is the preview path that was broken.
    let result = auto_cleanup_rust_targets(&guard, &mut runtime, false, true)
        .await
        .expect("dry-run scan must succeed");

    let cleaned: Vec<&String> = result.cleaned_paths.iter().collect();
    assert!(
        !cleaned.iter().any(|p| p.contains("protected-proj")),
        "the protected tree must NOT appear as reclaimable, got: {:?}",
        result.cleaned_paths
    );
    assert!(
        cleaned.iter().any(|p| p.contains("free-proj")),
        "an unprotected sibling must still be offered, got: {:?}",
        result.cleaned_paths
    );
    assert!(
        result
            .protected_paths
            .iter()
            .any(|p| p.contains("protected-proj")),
        "the protected tree must be reported as protected, got: {:?}",
        result.protected_paths
    );
    // The reason must survive, or the operator sees a refusal with no cause.
    // Asserted on a phrase that cannot come from the PATH — an earlier version
    // matched on "protected", which the fixture directory name itself contains,
    // so it passed with the reason deleted entirely.
    assert!(
        result
            .protected_paths
            .iter()
            .any(|p| p.contains("refusing to delete protected path")),
        "the protected verdict must carry its reason, got: {:?}",
        result.protected_paths
    );
    // Nothing was deleted. `reclaimed_bytes` is non-zero in a dry run BY
    // DESIGN — it is the "would reclaim" figure the preview's
    // "Total reclaimable" line is built from — so the proof that this is a
    // preview and not a delete is that the trees are still on disk.
    assert!(protected_proj.join("target").exists());
    assert!(free_proj.join("target").exists());
    assert!(
        free_proj
            .join("target")
            .join("debug")
            .join("blob.bin")
            .exists(),
        "the offered candidate must still be intact after a dry run"
    );

    let _ = std::fs::remove_dir_all(&root);
}

/// A multi-line error must not be able to break the preview's table alignment
/// by embedding a newline in a cell.
#[test]
fn one_line_collapses_multiline_messages() {
    assert_eq!(
        one_line("refusing to delete protected path\n  under user-protected path X"),
        "refusing to delete protected path under user-protected path X"
    );
    assert_eq!(one_line("already one line"), "already one line");
    assert_eq!(one_line(""), "");
}

#[tokio::test]
async fn storage_cleanup_activity_refuses_recent_artifacts() {
    let project = tempfile::tempdir().unwrap();
    let target = project.path().join("target");
    fs::create_dir(&target).unwrap();
    let error = validate_storage_cleanup_activity("rust-build", &target)
        .await
        .unwrap_err();
    assert!(error.to_string().contains("modified <60s"), "{error}");
    assert!(target.exists());
}

#[cfg(target_os = "linux")]
#[tokio::test]
async fn storage_cleanup_activity_refuses_live_rust_and_node_processes() {
    // Run this same test in a child harness with an explicit Linux process
    // name. Multicall `sleep` binaries can reset comm after exec, so a link
    // named rustc is not a reliable live-process fixture on NixOS.
    if let Ok(comm) = std::env::var("DRACON_TEST_BUILD_COMM") {
        fs::write("/proc/self/comm", comm).unwrap();
        fs::write(".build-fixture-ready", b"ready").unwrap();
        loop {
            std::thread::sleep(Duration::from_secs(300));
        }
    }
    struct ChildGuard(std::process::Child);
    impl Drop for ChildGuard {
        fn drop(&mut self) {
            let _ = self.0.kill();
            let _ = self.0.wait();
        }
    }
    for (comm, kind, artifact) in [
        ("rustc", "rust-build", "target"),
        ("cargo", "build-output", "build"),
        ("node", "node-deps", "node_modules"),
    ] {
        let project = tempfile::tempdir().unwrap();
        let path = project.path().join(artifact);
        fs::create_dir(&path).unwrap();
        fs::write(path.join("keep"), b"synthetic artifact").unwrap();
        let old = SystemTime::now() - Duration::from_secs(120);
        File::open(&path)
            .unwrap()
            .set_times(std::fs::FileTimes::new().set_modified(old))
            .unwrap();
        let mut child = ChildGuard(
            std::process::Command::new(std::env::current_exe().unwrap())
                .args([
                    "--exact",
                    "tests::storage_cleanup_activity_refuses_live_rust_and_node_processes",
                ])
                .env("DRACON_TEST_BUILD_COMM", comm)
                .current_dir(project.path())
                .stdout(std::process::Stdio::null())
                .spawn()
                .unwrap(),
        );
        let ready = project.path().join(".build-fixture-ready");
        let deadline = Instant::now() + Duration::from_secs(5);
        while !ready.exists() && Instant::now() < deadline {
            assert!(
                child.0.try_wait().unwrap().is_none(),
                "fixture exited before ready"
            );
            std::thread::sleep(Duration::from_millis(10));
        }
        assert!(ready.exists(), "fixture did not become ready");
        let error = validate_storage_cleanup_activity(kind, &path)
            .await
            .unwrap_err();
        assert!(error.to_string().contains("active"), "{kind}: {error}");
        assert!(child.0.try_wait().unwrap().is_none());
        assert_eq!(fs::read(path.join("keep")).unwrap(), b"synthetic artifact");
    }
}

#[cfg(unix)]
#[tokio::test]
async fn rust_build_listing_failure_or_malformed_output_is_an_error() {
    let _fixture_exec = fixture_exec_guard();
    let bin = tempfile::tempdir().unwrap();
    for body in ["exit 7", "printf 'broken\\n'", "exit 0"] {
        let ps = bin.path().join("ps");
        write_test_script(&ps, body);
        assert!(detect_active_rust_builds_with(&ps).await.is_err());
    }
}