Artifact [80bb6fef67]

Artifact 80bb6fef67c88a29a2b6bdc372996c9f2c4ad65db289cddebc7a427b42ee5d39:

• File src/sim/checkout_sweep.zig — part of check-in [0cea021cf9] at 2026-07-20 04:21:55 on branch h2-checkout-dst — sim: working-tree crash-recovery DST sweep for update

  Add sim/faulty_dir.zig, a decorator std.Io that cuts a run at a seeded mutating-op boundary the way a power loss does: the crashing write persists at most a torn prefix and every later mutation is a frozen no-op, so the checkout layer's own recovery moves no byte. Crash-point enumeration follows TigerBeetle's storage-fault-injection lineage, ported to the directory surface update.zig and checkout.zig write through.

  Add sim/checkout_sweep.zig: it materializes a check-in, replays an A-to-B update.applyCheckin under a crash at every reachable cut, reopens, and asserts unmanaged files are preserved, the checkout record names only a complete target, and the tree recovers to the target via the undo journal. Joins the dst gate per seed at bounded cuts; the soak walks every boundary.

  Finding: update.writePayload writes working files in place (createFile truncate, write, sync), not through the atomic write-temp+fsync+rename that checkout.write, the managed set, and the undo journal use. A crash mid-write can leave a file matching neither its pre- nor its post-image. The sweep pins this; a working-tree write-journal is the fix. (user: dmestas, size: 23601) [annotate] [blame] [check-ins using] [more...]

//! The working-tree crash-recovery DST sweep (H2, native harness design — no RFC
//! clause touches the working tree; contracts/undo.md and update.zig's own
//! transactional discipline are the oracle). It drives the real
//! `update.applyCheckin` public API over a real directory whose every write goes
//! through `sim/faulty_dir.zig`, cuts the run at every reachable mutating-op
//! boundary, reopens, and asserts the crash-recovery invariants.
//!
//! Crash-point enumeration — count the mutating operations one `update` performs,
//! then replay it once per cut `k ∈ [1, total]` and assert recovery from each —
//! is TigerBeetle's storage-fault-injection lineage (the VOPR walks every
//! disk-write boundary); this ports it from block storage to the checkout layer's
//! POSIX directory surface. The seed is the whole universe: the same seed picks
//! the same content, the same op count, and the same torn-prefix lengths, so any
//! failure replays verbatim from `reproduce with seed=N`.
//!
//! Invariants (the model tracks each managed path's pre-op A-bytes and post-op
//! B-bytes, plus the untouched user witness):
//!   (i)   no silent hybrid — a managed file present after a cut is byte-identical
//!         to its A-image or its B-image, never a splice. **This does not hold
//!         today** and is checked only by the finding test below, never the gate:
//!         `update.writePayload` materializes working files in place
//!         (`createFile(truncate)` + `writePositionalAll` + `sync`), NOT through
//!         the atomic write-temp+fsync+rename that `checkout.write`/`managed`/the
//!         undo journal use — so a crash between the truncate and the durable
//!         write leaves a file matching neither image. See the finding test.
//!   (ii)  user-edit preservation — an unmanaged witness file is byte-identical to
//!         the user's bytes under every cut (update never touches out-of-set paths).
//!   (iii) recoverability — after any cut, a healthy `update` to the target either
//!         completes directly or, once `zit undo` replays the journal, reaches the
//!         declared target; the two-phase liveness mirrors storage_sweep.zig's
//!         healed phase.
//!   (iv)  checkout-record consistency — the record names only A, the target B, or
//!         unborn; and when it already names B, B is fully materialized (the record
//!         is the atomic commit point, written last).

const std = @import("std");
const zit = @import("zit");

const artifact = zit.artifact;
const checkout = zit.checkout;
const logstore = zit.logstore;
const managed = zit.managed;
const storage = zit.storage;
const update = zit.update;

const faulty_dir = @import("faulty_dir.zig");

const Dir = std.Io.Dir;
const Io = std.Io;

pub const SweepError = error{
    OutOfMemory,
    /// A backend/setup filesystem op failed unexpectedly (never a fault the
    /// sweep injected — those are handled).
    Io,
    /// Invariant (iv): the record named a check-in other than A/B/unborn, or it
    /// named B while B was not fully on disk.
    RecordNamesIncompleteTarget,
    /// Invariant (ii): a crashed update destroyed an unmanaged user file.
    UnmanagedFileDestroyed,
    /// Invariant (iii): after a cut, no healthy path (direct or via `zit undo`)
    /// returns the tree to the declared target.
    Unrecoverable,
    /// Invariant (i): a managed file survived a cut matching neither its pre- nor
    /// its post-image — a torn, spliced write. THE FINDING (checked only by the
    /// finding test).
    HybridFile,
};

const witness_path = "witness.txt";
const witness_bytes = "the user's own uncommitted scratch\n";

/// One materialized check-in's file set: path -> uncompressed bytes, owned.
const Tree = struct {
    paths: std.ArrayList([]u8) = .empty,
    blobs: std.ArrayList([]u8) = .empty,

    fn deinit(self: *Tree, gpa: std.mem.Allocator) void {
        for (self.paths.items) |p| gpa.free(p);
        for (self.blobs.items) |b| gpa.free(b);
        self.paths.deinit(gpa);
        self.blobs.deinit(gpa);
        self.* = undefined;
    }

    fn add(self: *Tree, gpa: std.mem.Allocator, path: []const u8, blob: []const u8) SweepError!void {
        const p = gpa.dupe(u8, path) catch return error.OutOfMemory;
        errdefer gpa.free(p);
        const b = gpa.dupe(u8, blob) catch return error.OutOfMemory;
        errdefer gpa.free(b);
        self.paths.append(gpa, p) catch return error.OutOfMemory;
        self.blobs.append(gpa, b) catch return error.OutOfMemory;
    }

    /// The image bytes for `path`, or null when this tree does not track it.
    fn get(self: *const Tree, path: []const u8) ?[]const u8 {
        for (self.paths.items, self.blobs.items) |p, b| {
            if (std.mem.eql(u8, p, path)) return b;
        }
        return null;
    }
};

/// The full managed-path universe across A and B (used by the hybrid and record
/// oracles). A is `{f.txt}`, B is `{f.txt, g.txt}`.
const managed_paths = [_][]const u8{ "f.txt", "g.txt" };

/// The seeded corpus for one run: the two check-in ids and the A/B expected
/// trees. Pure data — the log is (re)built from it per workspace. A's tracked
/// file is rewritten and a second file added at B, so an A->B update deletes,
/// rewrites, and creates: a spread of mutating ops for the cut enumeration.
const Corpus = struct {
    a_id: [64]u8,
    a_len: usize,
    b_id: [64]u8,
    b_len: usize,
    a_tree: Tree,
    b_tree: Tree,

    fn aId(self: *const Corpus) []const u8 {
        return self.a_id[0..self.a_len];
    }
    fn bId(self: *const Corpus) []const u8 {
        return self.b_id[0..self.b_len];
    }
    fn deinit(self: *Corpus, gpa: std.mem.Allocator) void {
        self.a_tree.deinit(gpa);
        self.b_tree.deinit(gpa);
        self.* = undefined;
    }
};

fn completeArtifact(gpa: std.mem.Allocator, body: []const u8) SweepError![]u8 {
    var checksum_buf: [32]u8 = undefined;
    const checksum = artifact.selfChecksum(body, &checksum_buf);
    return std.fmt.allocPrint(gpa, "{s}Z {s}\n", .{ body, checksum }) catch return error.OutOfMemory;
}

fn putRaw(log: *logstore.LogStore, name: []const u8, bytes: []const u8) SweepError!void {
    var digest: [logstore.max_hash_len]u8 = undefined;
    const raw = digest[0 .. name.len / 2];
    _ = std.fmt.hexToBytes(raw, name) catch return error.Io;
    _ = log.put(raw, bytes) catch return error.Io;
}

fn seedBlob(gpa: std.mem.Allocator, random: std.Random, tag: []const u8) SweepError![]u8 {
    // A few hundred bytes of tagged, seed-varied content: distinct per blob so
    // every artifact hashes distinctly, and non-empty so a torn/empty survivor is
    // unambiguously neither image.
    const rows = 6 + random.uintLessThan(usize, 10);
    var buf: std.ArrayList(u8) = .empty;
    errdefer buf.deinit(gpa);
    var i: usize = 0;
    while (i < rows) : (i += 1) {
        buf.print(gpa, "{s} line {d} salt {x}\n", .{ tag, i, random.int(u32) }) catch return error.OutOfMemory;
    }
    return buf.toOwnedSlice(gpa) catch return error.OutOfMemory;
}

fn manifestA(gpa: std.mem.Allocator, fa: []const u8) SweepError![]u8 {
    var fname: [64]u8 = undefined;
    const fa_id = artifact.artifactName(.sha1, fa, &fname);
    const body = std.fmt.allocPrint(gpa, "D 2020-01-01T00:00:00\nF f.txt {s}\n", .{fa_id}) catch return error.OutOfMemory;
    defer gpa.free(body);
    return completeArtifact(gpa, body);
}

fn manifestB(gpa: std.mem.Allocator, fb: []const u8, gb: []const u8, a_name: []const u8) SweepError![]u8 {
    var fname: [64]u8 = undefined;
    var gname: [64]u8 = undefined;
    const fb_id = artifact.artifactName(.sha1, fb, &fname);
    const gb_id = artifact.artifactName(.sha1, gb, &gname);
    // F cards filename-sorted (f before g); card order D, F, F, P is canonical.
    const body = std.fmt.allocPrint(gpa, "D 2020-01-02T00:00:00\nF f.txt {s}\nF g.txt {s}\nP {s}\n", .{ fb_id, gb_id, a_name }) catch return error.OutOfMemory;
    defer gpa.free(body);
    return completeArtifact(gpa, body);
}

/// Compute the corpus ids and A/B trees from a seeded content stream. Pure: no
/// log, no filesystem — the log is seeded from this per workspace.
fn buildCorpus(gpa: std.mem.Allocator, random: std.Random) SweepError!Corpus {
    const fa = try seedBlob(gpa, random, "f-at-a");
    defer gpa.free(fa);
    const fb = try seedBlob(gpa, random, "f-at-b");
    defer gpa.free(fb);
    const gb = try seedBlob(gpa, random, "g-at-b");
    defer gpa.free(gb);

    const a_bytes = try manifestA(gpa, fa);
    defer gpa.free(a_bytes);
    var a_id: [64]u8 = undefined;
    const a_name = artifact.artifactName(.sha1, a_bytes, &a_id);

    const b_bytes = try manifestB(gpa, fb, gb, a_name);
    defer gpa.free(b_bytes);
    var b_id: [64]u8 = undefined;
    const b_name = artifact.artifactName(.sha1, b_bytes, &b_id);

    var a_tree = Tree{};
    errdefer a_tree.deinit(gpa);
    try a_tree.add(gpa, "f.txt", fa);
    var b_tree = Tree{};
    errdefer b_tree.deinit(gpa);
    try b_tree.add(gpa, "f.txt", fb);
    try b_tree.add(gpa, "g.txt", gb);

    return .{ .a_id = a_id, .a_len = a_name.len, .b_id = b_id, .b_len = b_name.len, .a_tree = a_tree, .b_tree = b_tree };
}

/// Seed `log` with the corpus's blobs and manifests (byte-identical every call,
/// since content is a pure function of the corpus).
fn seedLog(gpa: std.mem.Allocator, corpus: *const Corpus, log: *logstore.LogStore) SweepError!void {
    const fa = corpus.a_tree.get("f.txt").?;
    const fb = corpus.b_tree.get("f.txt").?;
    const gb = corpus.b_tree.get("g.txt").?;
    for ([_][]const u8{ fa, fb, gb }) |blob| {
        var name: [64]u8 = undefined;
        const id = artifact.artifactName(.sha1, blob, &name);
        try putRaw(log, id, blob);
    }
    const a_bytes = try manifestA(gpa, fa);
    defer gpa.free(a_bytes);
    try putRaw(log, corpus.aId(), a_bytes);
    const b_bytes = try manifestB(gpa, fb, gb, corpus.aId());
    defer gpa.free(b_bytes);
    try putRaw(log, corpus.bId(), b_bytes);
    log.commit() catch return error.Io;
}

/// A live workspace: a fresh subdir under `root`, an open repo.zit seeded with
/// the corpus, and its LogStore.
const Workspace = struct {
    name: []u8,
    dir: Dir,
    repo_file: std.Io.File,
    /// Heap-allocated so `StorageFile.context = &fs` stays valid after this
    /// Workspace is returned by value (a moved local would dangle it).
    fs: *storage.FileStorage,
    log: logstore.LogStore,
    root: Dir,
    io: Io,

    fn open(gpa: std.mem.Allocator, io: Io, root: Dir, corpus: *const Corpus, nonce: u64) SweepError!Workspace {
        const name = std.fmt.allocPrint(gpa, "ws-{x}", .{nonce}) catch return error.OutOfMemory;
        errdefer gpa.free(name);
        root.deleteTree(io, name) catch {}; // a prior aborted run may have left it
        root.createDir(io, name, .default_dir) catch return error.Io;
        var dir = root.openDir(io, name, .{ .access_sub_paths = true, .follow_symlinks = false }) catch return error.Io;
        errdefer dir.close(io);

        var repo_file = dir.createFile(io, "repo.zit", .{ .read = true, .truncate = true, .resolve_beneath = true }) catch return error.Io;
        errdefer repo_file.close(io);

        const fs = gpa.create(storage.FileStorage) catch return error.OutOfMemory;
        errdefer gpa.destroy(fs);
        fs.* = storage.FileStorage.init(io, repo_file, 0);
        var log = logstore.LogStore.open(gpa, fs.storageFile(), 0) catch return error.Io;
        errdefer log.deinit(gpa);
        try seedLog(gpa, corpus, &log);

        return .{
            .name = name,
            .dir = dir,
            .repo_file = repo_file,
            .fs = fs,
            .log = log,
            .root = root,
            .io = io,
        };
    }

    fn close(self: *Workspace, gpa: std.mem.Allocator) void {
        self.log.deinit(gpa);
        gpa.destroy(self.fs);
        self.repo_file.close(self.io);
        self.dir.close(self.io);
        self.root.deleteTree(self.io, self.name) catch {};
        gpa.free(self.name);
        self.* = undefined;
    }
};

/// Seed the working tree to a clean check-out of A, plus the user's witness file.
fn seedAtA(gpa: std.mem.Allocator, io: Io, ws: *Workspace, corpus: *const Corpus) SweepError!void {
    // Unborn record + empty managed set, exactly what `init` leaves.
    checkout.write(io, ws.dir, .{ .version = "", .branch = "", .merge_parent = "", .special_op = "", .manifest_flags = "" }) catch return error.Io;
    var empty = managed.ManagedSet.init(gpa);
    defer empty.deinit();
    empty.write(io, ws.dir) catch return error.Io;

    writeUserFile(io, ws.dir, witness_path, witness_bytes) catch return error.Io;
    update.applyCheckin(gpa, io, ws.dir, &ws.log, corpus.aId(), .inherit) catch return error.Io;
}

fn writeUserFile(io: Io, dir: Dir, path: []const u8, bytes: []const u8) !void {
    var file = try dir.createFile(io, path, .{ .read = false, .truncate = true, .resolve_beneath = true });
    defer file.close(io);
    try file.writePositionalAll(io, bytes, 0);
    try file.sync(io);
}

/// Count the mutating ops one healthy A->B update performs on a throwaway
/// workspace — the crash-cut range `[1, total]`.
fn countOps(gpa: std.mem.Allocator, io: Io, root: Dir, corpus: *const Corpus) SweepError!u64 {
    var ws = try Workspace.open(gpa, io, root, corpus, 0xC0);
    defer ws.close(gpa);
    try seedAtA(gpa, io, &ws, corpus);

    var fprng = std.Random.DefaultPrng.init(0xC0);
    var faulty = faulty_dir.FaultyDir.init(io, fprng.random(), .{}); // no crash: count only
    faulty.install();
    defer faulty_dir.FaultyDir.uninstall();
    update.applyCheckin(gpa, faulty.io(), ws.dir, &ws.log, corpus.bId(), .inherit) catch return error.Io;
    return faulty.op_index;
}

const FileState = union(enum) { missing, bytes: []u8 };

fn readFile(gpa: std.mem.Allocator, io: Io, dir: Dir, path: []const u8) SweepError!FileState {
    const bytes = dir.readFileAlloc(io, path, gpa, .limited(1 << 20)) catch |e| switch (e) {
        error.FileNotFound => return .missing,
        error.OutOfMemory => return error.OutOfMemory,
        else => return error.Io,
    };
    return .{ .bytes = bytes };
}

/// Invariant (ii): the witness is byte-identical to the user's bytes.
fn checkWitness(gpa: std.mem.Allocator, io: Io, dir: Dir) SweepError!void {
    var st = try readFile(gpa, io, dir, witness_path);
    switch (st) {
        .missing => return error.UnmanagedFileDestroyed,
        .bytes => |*b| {
            defer gpa.free(b.*);
            if (!std.mem.eql(u8, b.*, witness_bytes)) return error.UnmanagedFileDestroyed;
        },
    }
}

/// Invariant (i): every managed path present on disk matches its A-image or its
/// B-image. Returns error.HybridFile on the first splice. Only the finding test
/// calls this — the property does not hold today.
fn checkNoHybrid(gpa: std.mem.Allocator, io: Io, dir: Dir, corpus: *const Corpus) SweepError!void {
    for (managed_paths) |path| {
        var st = try readFile(gpa, io, dir, path);
        switch (st) {
            .missing => {}, // absence is an (iii) concern, not a splice
            .bytes => |*b| {
                defer gpa.free(b.*);
                const a_img = corpus.a_tree.get(path);
                const b_img = corpus.b_tree.get(path);
                const matches_a = a_img != null and std.mem.eql(u8, b.*, a_img.?);
                const matches_b = b_img != null and std.mem.eql(u8, b.*, b_img.?);
                if (!matches_a and !matches_b) return error.HybridFile;
            },
        }
    }
}

/// Invariant (iv): the record names only A/B/unborn, and if it names B then B is
/// fully materialized on disk.
fn checkRecord(gpa: std.mem.Allocator, io: Io, dir: Dir, corpus: *const Corpus) SweepError!void {
    var buf: [checkout.max_ck_payload + 1]u8 = undefined;
    const state = checkout.read(io, dir, &buf) catch return error.Io;
    const ver = if (state) |s| s.version else "";
    if (ver.len == 0) return; // unborn/mid — permitted
    if (std.mem.eql(u8, ver, corpus.aId())) return; // still A — permitted
    if (!std.mem.eql(u8, ver, corpus.bId())) return error.RecordNamesIncompleteTarget;
    // Record already names B: B must be fully on disk (the record is written last).
    try expectTree(gpa, io, dir, &corpus.b_tree);
}

/// Assert every path in `tree` is present with its bytes.
fn expectTree(gpa: std.mem.Allocator, io: Io, dir: Dir, tree: *const Tree) SweepError!void {
    for (tree.paths.items, tree.blobs.items) |path, want| {
        var st = try readFile(gpa, io, dir, path);
        switch (st) {
            .missing => return error.RecordNamesIncompleteTarget,
            .bytes => |*b| {
                defer gpa.free(b.*);
                if (!std.mem.eql(u8, b.*, want)) return error.RecordNamesIncompleteTarget;
            },
        }
    }
}

/// Invariant (iii): a healthy `update` to B either completes directly, or — once
/// `zit undo` replays the journal — reaches B. The final tree must equal B.
fn checkRecoverable(gpa: std.mem.Allocator, io: Io, ws: *Workspace, corpus: *const Corpus) SweepError!void {
    if (update.applyCheckin(gpa, io, ws.dir, &ws.log, corpus.bId(), .inherit)) |_| {
        // Direct completion.
    } else |err| switch (err) {
        error.OutOfMemory => return error.OutOfMemory,
        // Any clean refusal (dirty tree, collision, invalid current state) is the
        // documented interrupted-state report. Recovery is the undo journal.
        else => {
            update.journal.undo(gpa, io, ws.dir) catch return error.Unrecoverable;
            update.applyCheckin(gpa, io, ws.dir, &ws.log, corpus.bId(), .inherit) catch return error.Unrecoverable;
        },
    }
    try expectTree(gpa, io, ws.dir, &corpus.b_tree);
}

/// Options controlling one enumeration pass.
const CutOpts = struct {
    torn_write: bool = false,
    corrupt_torn: bool = false,
    check_hybrid: bool = false,
    /// Cap on cuts actually replayed. 0 = every cut. When `total` exceeds it, the
    /// cuts are sampled evenly across `[1, total]` (endpoints always included), so
    /// the gate stays small while the soak (larger cap) walks every boundary — the
    /// ci.yml gate-vs-sweep division, applied here (TEST-HARDENING §H2/§6).
    max_cuts: usize = 0,
};

/// Result of an enumeration: whether any cut tripped the hybrid detector (the
/// finding), reported so the finding test can assert on it without treating it as
/// a run failure.
const CutOutcome = struct { hybrid_seen: bool = false };

/// Build the corpus once, then for every cut `k ∈ [1, total]` replay A->B under a
/// crash at op k and check the invariants. Deterministic: `seed` fixes content,
/// op count, and every torn length.
fn sweepSeed(gpa: std.mem.Allocator, io: Io, root: Dir, seed: u64, opts: CutOpts) SweepError!CutOutcome {
    var prng = std.Random.DefaultPrng.init(seed);
    const random = prng.random();

    var corpus = try buildCorpus(gpa, random);
    defer corpus.deinit(gpa);

    const total = try countOps(gpa, io, root, &corpus);
    if (total == 0) return .{};

    var outcome = CutOutcome{};
    const budget = if (opts.max_cuts == 0) total else @min(opts.max_cuts, total);
    var j: u64 = 0;
    while (j < budget) : (j += 1) {
        // Evenly spaced across [1, total], endpoints included: cut = 1 + round(j *
        // (total-1) / (budget-1)); with budget == total this is every k.
        const k: u64 = if (budget == 1) 1 else 1 + (j * (total - 1)) / (budget - 1);
        if (try sweepCut(gpa, io, root, &corpus, seed, k, opts)) outcome.hybrid_seen = true;
    }
    return outcome;
}

/// One cut: fresh workspace at A, faulted A->B at op `k`, then the oracle.
/// Returns whether the hybrid detector tripped (only meaningful when
/// `opts.check_hybrid`).
fn sweepCut(gpa: std.mem.Allocator, io: Io, root: Dir, corpus: *const Corpus, seed: u64, k: u64, opts: CutOpts) SweepError!bool {
    const nonce = std.hash.Wyhash.hash(seed, std.mem.asBytes(&k));
    var ws = try Workspace.open(gpa, io, root, corpus, nonce);
    defer ws.close(gpa);
    try seedAtA(gpa, io, &ws, corpus);

    var fprng = std.Random.DefaultPrng.init(seed *% 0x9E37_79B9 +% k);
    var faulty = faulty_dir.FaultyDir.init(io, fprng.random(), .{
        .crash_at_op = k,
        .torn_write = opts.torn_write,
        .corrupt_torn = opts.corrupt_torn,
    });
    faulty.install();
    // The faulted update is EXPECTED to fail (that is the crash); ignore its
    // return and inspect the disk with the healthy io.
    update.applyCheckin(gpa, faulty.io(), ws.dir, &ws.log, corpus.bId(), .inherit) catch {};
    faulty_dir.FaultyDir.uninstall();

    // (ii) and (iv) hold at every cut and are always checked.
    try checkWitness(gpa, io, ws.dir);
    try checkRecord(gpa, io, ws.dir, corpus);

    var hybrid = false;
    if (opts.check_hybrid) {
        checkNoHybrid(gpa, io, ws.dir, corpus) catch |e| switch (e) {
            error.HybridFile => hybrid = true, // the finding; recoverability still proven below
            else => return e,
        };
    }
    // (iii) recoverability — proven even when a hybrid is present: the tree is
    // torn but never lost.
    try checkRecoverable(gpa, io, &ws, corpus);
    return hybrid;
}

/// The gate entry: the bounded crash-recovery sweep. Checks invariants (ii),
/// (iii), (iv) at every reachable cut of an A->B update. Invariant (i) is the
/// documented open finding and is NOT gated here (see the module header and the
/// finding test). `root` is a caller-owned directory the sweep makes throwaway
/// workspaces under.
pub fn run(gpa: std.mem.Allocator, io: Io, root: Dir, seed: u64) SweepError!void {
    // 16 evenly-spaced cuts keep the gate to seconds; H1's soak runs `sweepSeed`
    // with `max_cuts = 0` (every boundary) across many seeds for the depth.
    _ = try sweepSeed(gpa, io, root, seed, .{ .torn_write = false, .check_hybrid = false, .max_cuts = 16 });
}

// ---------------------------------------------------------------------------
// Inline tests (join `zig build test` via dst.zig's aggregation). They use
// std.testing.io + a tmpDir; the gate `run` is driven from dst.zig's main over a
// process-created workspace.
// ---------------------------------------------------------------------------

test "checkout_sweep: A->B recovers at every crash cut (invariants ii/iii/iv)" {
    const t = std.testing;
    var tmp = t.tmpDir(.{});
    defer tmp.cleanup();
    // Two seeds so content and op counts vary; both must stay green.
    try run(t.allocator, t.io, tmp.dir, 1);
    try run(t.allocator, t.io, tmp.dir, 2);
}

test "checkout_sweep FINDING (H2): a torn in-place payload write leaves a hybrid file" {
    const t = std.testing;
    var tmp = t.tmpDir(.{});
    defer tmp.cleanup();

    // Enumerate every cut with torn writes and the (i) detector armed. Because
    // update.writePayload is a non-atomic createFile(truncate)+write+sync, at
    // least one cut tears a managed file into a splice matching neither image —
    // WHILE recoverability (iii) and witness preservation (ii) still hold (proven
    // inside sweepCut). This PINS the current gap: when a write-journal /
    // staged-materialization fix lands, no cut will produce a hybrid and this
    // assertion flips — convert it then to `try t.expect(!outcome.hybrid_seen)`.
    const outcome = try sweepSeed(t.allocator, t.io, tmp.dir, 1, .{ .torn_write = true, .corrupt_torn = true, .check_hybrid = true });
    try t.expect(outcome.hybrid_seen);
}