coordinode-lsm-tree 7.4.2

use super::{BlobDropReason, DropReason, salvage_blob_file, salvage_sst};
// The options-bearing entry is exercised on every feature set: the
// prefix-extractor round-trip below is ungated (extractors are core
// configuration), so the import must not hide behind the encrypted /
// dictionary / delete-resurrection gates its other consumers carry.
use super::{SalvageOptions, salvage_sst_with_options};
use crate::comparator::default_comparator;
use crate::fs::{Fs, StdFs};
use crate::table::{Table, Writer};
use crate::{InternalValue, ValueType};
use alloc::sync::Arc;
use tempfile::tempdir;
use test_log::test;

/// Runs the whole reconcile family and returns the error, asserting it came
/// from the gate under test.
///
/// The gates share one decode of each block, so they are driven together; this
/// keeps each test pinning BOTH its own failure message and the fact that the
/// combined pass still routes that forgery to the right check.
#[cfg(feature = "std")]
fn reconcile_error(
    table: &Table,
    expected: crate::table::ReconcileGate,
    prefix_extractor: Option<&Arc<dyn crate::prefix::PrefixExtractor>>,
) -> crate::Error {
    match table.verify_reconcile_gates(prefix_extractor, false) {
        Ok(()) => panic!("the forgery must be rejected, the reconcile pass accepted it"),
        Err((gate, e)) => {
            assert_eq!(gate, expected, "the wrong gate rejected the table: {e}");
            e
        }
    }
}

/// Asserts the whole reconcile family accepts an honest table.
#[cfg(feature = "std")]
fn reconcile_clean(
    table: &Table,
    prefix_extractor: Option<&Arc<dyn crate::prefix::PrefixExtractor>>,
) {
    if let Err((gate, e)) = table.verify_reconcile_gates(prefix_extractor, false) {
        panic!("an honest table must pass every gate, {gate:?} refused it: {e}");
    }
}

/// Rot in a frame's length field is caught by the header CRC, after which the
/// consumed lengths cannot locate the next frame: the salvage walk resyncs at
/// the next magic instead of stopping. But the resync magic may be an original
/// boundary OR a `BLO4` frame nested in the rotted frame's user bytes, and every
/// frame CHAINED past it inherits that unproven anchor: a fabricated chain can
/// plant one valid frame after another. There is no independent anchor
/// mid-stream, so the taint is STICKY: the whole tail after the first resync is
/// dropped (fail closed), not just the frame reached immediately by the scan.
#[test]
fn salvage_blob_file_drops_the_whole_tail_after_a_resync() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_rot");
    let dest = dir.path().join("blob_rot_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    build_blob(
        &source,
        &fs,
        &[
            (b"aaaa", b"AAAAAAAA"),
            (b"bbbb", b"BBBBBBBB"),
            (b"cccc", b"CCCCCCCC"),
            (b"dddd", b"DDDDDDDD"),
        ],
    )?;

    // Rot the SECOND frame's key_len (4 -> 6). Frame layout: magic 4 |
    // checksum 16 | seqno 8 | key_len 2 | real_val_len 4 | on_disk_val_len 4
    // | header_crc 4 | key | value = 42 + 4 + 8 bytes; the data section
    // starts at file offset 0.
    let frame_len = 42 + 4 + 8;
    let kl_off = frame_len + 4 + 16 + 8;
    let mut bytes = std::fs::read(&source)?;
    let Some(slot) = bytes.get_mut(kl_off..kl_off + 2) else {
        panic!("second frame's key_len within the file");
    };
    slot.copy_from_slice(&6u16.to_le_bytes());
    std::fs::write(&source, &bytes)?;

    let report = salvage_blob_file(
        &source,
        dest,
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    // aaaa recovered; bbbb drops (header CRC) and arms the resync taint; cccc is
    // the first frame reached by the byte scan (unproven boundary). The taint is
    // sticky, so the walk STOPS there and reports the whole surrendered tail
    // (cccc and everything chained past it) as ONE drop. Only the frame BEFORE the
    // rot survives.
    assert_eq!(
        report.records_salvaged, 1,
        "only the frame before the rot is provable; the whole tail after the resync drops: {report:?}",
    );
    assert!(
        report
            .dropped
            .iter()
            .any(|d| matches!(&d.reason, BlobDropReason::Corrupt(msg) if msg.contains("HeaderCrcMismatch"))),
        "the rotted frame drops as a header CRC mismatch: {report:?}",
    );
    assert_eq!(
        report
            .dropped
            .iter()
            .filter(
                |d| matches!(&d.reason, BlobDropReason::Corrupt(msg) if msg.contains("surrendered"))
            )
            .count(),
        1,
        "the surrendered tail is recorded ONCE, not per tainted frame: {report:?}",
    );

    // The recovered copy carries only the single provable record.
    let Some(salvaged) = report.salvaged_path else {
        panic!("one record was salvaged");
    };
    let keys: Vec<Vec<u8>> = BlobScanner::new(&salvaged, &*fs, 0)?
        .map(|r| r.map(|e| e.key.to_vec()))
        .collect::<crate::Result<_>>()?;
    assert_eq!(keys, vec![b"aaaa".to_vec()]);
    Ok(())
}

/// A forged `tli_tail` that DROPS a handle passes every byte-level check
/// (fresh checksum, valid framing, Index role): a walk following the tail
/// would silently skip the hidden block and "recover" a copy missing its
/// keys — no drop, no error, just vanished rows. Pins that the salvage open
/// walks the HEAD `tli` copy, so a forged tail cannot steer which blocks
/// are recovered.
#[test]
fn salvage_recovers_all_blocks_under_a_forged_tli_tail() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Tiny block budget so the SST spills several data blocks (several TLI
    // handles — the forge needs at least two).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    crate::test_forge::forge_tli_tail_truncated(&source, 0, None)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(
        report.dropped.is_empty(),
        "every block is intact, nothing may drop: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 64,
        "the block the forged tail hides must still be recovered: {report:?}",
    );
    Ok(())
}

/// BOTH TLI mirrors forged to the SAME truncated handle list pass every
/// byte-level check AND the mirror comparison (two forged copies prove
/// nothing), so the salvage open's block index simply omits the hidden
/// block. A walk trusting that index neither recovers the block nor
/// reports it dropped — repair then installs an apparently complete copy
/// with the block's keys silently missing. The physical data-section
/// tiling is the only ground truth: salvage must frame and recover the
/// bytes the index does not cover.
#[test]
fn salvage_recovers_a_block_hidden_by_forged_tli_mirrors() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Tiny block budget so the SST spills several data blocks (several TLI
    // handles — the forge needs at least two).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    crate::test_forge::forge_tli_mirrors_truncated(&source, 0, None)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(
        report.dropped.is_empty(),
        "every block is intact, nothing may drop: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 64,
        "the block both forged mirrors hide must still be recovered: {report:?}",
    );
    Ok(())
}

/// BOTH TLI mirrors re-encoded with the first two handles SWAPPED keep
/// every block present and the section fully covered — but the stored
/// order is no longer the offset order the physical tiling assumes. A
/// tiling pass trusting that order frames the out-of-place block through
/// the gap probe AND pushes its handle again: the duplicate emit is
/// rejected by the writer's ordering validation, so an intact block is
/// reported dropped and the block totals are inflated. The tiling must
/// re-sort by offset and skip spans it already covered.
#[test]
fn salvage_recovers_all_blocks_under_a_reordered_tli() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    crate::test_forge::forge_tli_mirrors_swap_first_two(&source, 0, None)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(
        report.dropped.is_empty(),
        "every block is intact and covered once, nothing may drop: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 64,
        "a reordered index must not double-walk or lose blocks: {report:?}",
    );
    Ok(())
}

/// A partitioned index whose MIDDLE leaf partition is corrupt yields only the
/// earlier partitions' handles before erroring, setting `index_broken`. The
/// physical data section is still intact, writer-ordered, and self-framing, so
/// salvage must walk it independently and recover EVERY data block instead of
/// re-emitting only the enumerated prefix and silently dropping the failed and
/// later partitions' blocks.
#[test]
fn salvage_recovers_physical_blocks_past_a_broken_index_partition() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Tiny data blocks + a partitioned index spread the 256 blocks across
    // several leaf partitions (no range tombstone, so salvage re-emits rather
    // than failing closed).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(128)
        .use_partitioned_index();
    for i in 0u64..256 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Corrupt the middle of the `index` section (a leaf partition), desyncing
    // the handle enumeration partway while leaving the data section intact.
    let (index_pos, index_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = crate::sfa::Reader::from_reader(&mut f)?;
        let Some((pos, len)) = reader
            .toc()
            .iter()
            .find(|e| e.name() == b"index")
            .map(|e| (e.pos(), e.len()))
        else {
            panic!("a partitioned-index SST must carry an index section");
        };
        (pos, len)
    };
    let Ok(flip) = usize::try_from(index_pos + index_len / 2) else {
        panic!("the index-section offset fits usize");
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert_eq!(
        report.entries_salvaged, 256,
        "every data block must be recovered via the physical walk when the \
         index enumeration breaks: {report:?}",
    );
    // The index enumeration broke, but the physical walk recovered every data
    // block: that structural index damage is NOT data loss, so the report must
    // grade complete rather than counting the index error as a dropped block.
    assert!(
        report.is_complete(),
        "a broken index the physical walk fully recovers around is not data loss: {report:?}",
    );
    Ok(())
}

/// A range tombstone hidden by RENAMING its section to a recognized name whose
/// block decodes cleanly (here an empty `filter`, since filtering is disabled)
/// must fail salvage closed. The rename keeps the catalogue uniquely named and
/// tiled and the block loads without degrading, so neither the TOC-concealment
/// check nor the rebuildable-section degradation flag fires — but the persisted
/// `range_tombstone_count` still records the tombstone, so salvage cross-checks
/// it and refuses rather than re-emitting the covered keys as live.
#[test]
fn salvage_refuses_a_range_tombstone_hidden_as_a_recognized_section() -> crate::Result<()> {
    use crate::UserKey;
    use crate::config::BloomConstructionPolicy;
    use crate::range_tombstone::RangeTombstone;
    use crate::table::block::BlockType;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Filtering disabled, so renaming range_tombstones to `filter` yields a
    // UNIQUE recognized name (no existing filter to duplicate).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_bloom_policy(BloomConstructionPolicy::BitsPerKey(0.0));
    for i in 0u64..8 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    writer.write_range_tombstone(RangeTombstone::new(
        UserKey::from(b"key-002".as_slice()),
        UserKey::from(b"key-005".as_slice()),
        9,
    ));
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Rename range_tombstones -> filter and re-role its block to Filter: the
    // block loads as a (garbage) filter WITHOUT degrading, so the
    // rebuildable-section flag stays clear and the parsed table reports no range
    // tombstones — but the persisted count still records them.
    crate::test_forge::forge_duplicate_section_name(
        &source,
        b"range_tombstones",
        b"filter",
        BlockType::Filter,
    )?;

    let Err(err) = salvage_sst(&source, dest, &fs) else {
        panic!("a hidden range tombstone must fail salvage");
    };
    let crate::Error::FeatureUnsupported(reason) = &err else {
        panic!("the refusal must be FeatureUnsupported, got {err:?}");
    };
    assert!(
        reason.contains("range tombstones"),
        "the refusal must name range tombstones specifically, got {reason:?}",
    );
    Ok(())
}

/// When the index is UNTRUSTED (its mirror comparison, binary-index
/// authentication, or section tiling fails), an indexed offset is no more
/// provable than a byte-scanned one: a checksum-restamped TLI could point a
/// handle at a frame nested inside a corrupt block's value bytes. So once the
/// physical chain breaks, later indexed handles must be dropped too — the walk
/// recovers only the anchored prefix. (With a TRUSTED index the same corruption
/// stays block-granular; that path is covered by the recovery tests.)
#[test]
fn salvage_drops_indexed_blocks_after_a_break_when_the_index_is_untrusted() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(128)
        .use_partitioned_index();
    for i in 0u64..256 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Resolve a middle block offset BEFORE forging (data blocks precede the TLI,
    // so the TLI forge below leaves this offset valid).
    let smash_offset = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|h| *h.as_ref().offset())
            .collect();
        let Some(&off) = offsets.get(offsets.len() / 2) else {
            panic!("need several data blocks, got {}", offsets.len());
        };
        off
    };

    // Corrupt the tli_tail mirror so it disagrees with the head: the index
    // structure no longer authenticates (the mirrors diverge — exactly a
    // checksum-restamped-TLI signal), though head enumeration still yields
    // every handle.
    crate::test_forge::forge_flip_section_last_payload_byte(&source, b"tli_tail", None)?;

    // Smash the header of the middle data block: the physical chain breaks there.
    {
        let mut bytes = std::fs::read(&source)?;
        let Ok(at) = usize::try_from(smash_offset) else {
            panic!("data block offset {smash_offset} fits usize");
        };
        let Some(b) = bytes.get_mut(at) else {
            panic!("the block header at {at} lies within the file");
        };
        *b ^= 0xFF;
        std::fs::write(&source, &bytes)?;
    }

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    // The anchored prefix before the smash recovers.
    assert!(
        reopen_get(dest.clone(), &fs, b"key-000")?.is_some(),
        "the contiguous prefix before the broken chain must recover: {report:?}",
    );
    // A key in a block AFTER the smash is reachable only through its UNTRUSTED
    // index entry, past a broken boundary — it must NOT be re-emitted.
    assert!(
        reopen_get(dest, &fs, b"key-250")?.is_none(),
        "an indexed block after a broken chain, under an untrusted index, must \
         be dropped, not emitted: {report:?}",
    );
    assert!(
        report.dropped.iter().any(|d| matches!(
            &d.reason,
            DropReason::HeaderCorrupted(msg) if msg.contains("broken")
        )),
        "the surrendered tail must be reported as a broken-chain drop: {report:?}",
    );
    Ok(())
}

/// A PERSISTENT read failure on one block (a bad-sector `Other` / EIO, or a
/// truncated final block) must drop just that block, not abort the whole salvage
/// — otherwise one unreadable block sinks every intact sibling. Only a TRANSIENT
/// read aborts for a retry. The victim block's positioned read faults once with a
/// persistent kind; the rest of the file is intact.
#[test]
fn salvage_drops_a_persistently_unreadable_block_and_keeps_the_rest() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let clean: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer =
        Writer::new(source.clone(), 0, 0, Arc::clone(&clean))?.use_data_block_size(128);
    for i in 0u64..256 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Resolve a MIDDLE data block's offset.
    let victim_offset = {
        let table = open(source.clone(), &clean)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|h| *h.as_ref().offset())
            .collect();
        let Some(&off) = offsets.get(offsets.len() / 2) else {
            panic!("need several data blocks, got {}", offsets.len());
        };
        off
    };

    // Salvage through a fs whose positioned read of that ONE block ALWAYS fails
    // (a persistent bad sector); every read at a different offset succeeds.
    let fault = FaultFs::new(StdFs);
    fault.injector().arm(
        FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Other)).at_offset(victim_offset),
    );
    let faulting: Arc<dyn Fs> = Arc::new(fault);

    // Before the fix this aborted (`Err`); now it recovers, dropping just the
    // unreadable block. (The trusted index resolves the probe failure to a
    // broken-boundary drop rather than a ReadError, but the point is that ONE
    // unreadable block no longer sinks the salvage.)
    let report = salvage_sst(&source, dest.clone(), &faulting)?;
    assert!(
        !report.dropped.is_empty(),
        "a persistently unreadable block must be dropped, not abort the salvage: {report:?}",
    );
    assert!(
        report.blocks_salvaged >= report.blocks_total - report.dropped.len(),
        "every block except the dropped one is salvaged: {report:?}",
    );
    assert!(
        reopen_get(dest, &clean, b"key-000")?.is_some(),
        "an intact block must still be recovered past the dropped one: {report:?}",
    );
    Ok(())
}

/// When the broken-index physical walk hits an UNFRAMEABLE block header, the
/// bytes after it are reachable only by byte-scan resync, so their block
/// boundaries are unprovable: an uncompressed block can carry a complete
/// checksum-valid SST block inside a user value, and a resync would frame that
/// NESTED forge and re-emit its interior entries as genuine data. The walk must
/// fail closed like the blob resync path — drop the whole gap tail after the
/// broken boundary rather than emit unanchored candidates. Only the contiguous,
/// fully-framed prefix (anchored to the trusted section start) is recovered.
#[test]
fn salvage_drops_the_gap_tail_after_an_unframeable_block() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(128)
        .use_partitioned_index();
    for i in 0u64..256 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Smash the header of a block in the FIRST THIRD of the data section (its
    // first byte, so `probe_block_handle_in` cannot frame it). The blocks after
    // it are reachable only by resync, so they must be dropped, not emitted.
    let smash_offset = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|h| *h.as_ref().offset())
            .collect();
        let Some(&off) = offsets.get(offsets.len() / 3) else {
            panic!("need several data blocks, got {}", offsets.len());
        };
        off
    };
    {
        let mut bytes = std::fs::read(&source)?;
        let Ok(at) = usize::try_from(smash_offset) else {
            panic!("data block offset {smash_offset} fits usize");
        };
        let Some(b) = bytes.get_mut(at) else {
            panic!("the block header at {at} lies within the file");
        };
        *b ^= 0xFF;
        std::fs::write(&source, &bytes)?;
    }

    // Corrupt the `index` section so enumeration breaks and the physical walk
    // tiles the whole data section from the start.
    let (index_pos, index_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = crate::sfa::Reader::from_reader(&mut f)?;
        let Some((pos, len)) = reader
            .toc()
            .iter()
            .find(|e| e.name() == b"index")
            .map(|e| (e.pos(), e.len()))
        else {
            panic!("a partitioned-index SST must carry an index section");
        };
        (pos, len)
    };
    let Ok(flip) = usize::try_from(index_pos + index_len / 2) else {
        panic!("the index-section offset fits usize");
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    // The anchored prefix (contiguous from the trusted section start) recovers.
    assert!(
        reopen_get(dest.clone(), &fs, b"key-000")?.is_some(),
        "the contiguous prefix before the broken boundary must recover: {report:?}",
    );
    // A late key past the smashed block is reachable only by resync, so it must
    // NOT be re-emitted — its boundary cannot be proven original.
    assert!(
        reopen_get(dest, &fs, b"key-250")?.is_none(),
        "a block reached only by resync must be dropped, not emitted: {report:?}",
    );
    // The dropped tail is reported once with the unanchored reason.
    assert!(
        report.dropped.iter().any(|d| matches!(
            &d.reason,
            DropReason::HeaderCorrupted(msg) if msg.contains("unanchored")
        )),
        "the resync tail must be reported as an unanchored drop: {report:?}",
    );
    Ok(())
}

/// A header-checksum-valid FAKE header can declare an oversized `data_length`
/// that spans the real blocks after it. The walk LOADS each candidate so it
/// never advances by the unvalidated framed size, but the load failure it hits
/// at the fake header breaks the contiguous chain: the swallowed span is then
/// reachable only by byte-scan resync, whose boundaries are unprovable (a
/// nested checksum-valid frame in a user value is indistinguishable from a real
/// block). The walk must fail closed — drop the tail past the fake header
/// rather than resync into it. The anchored prefix before the forge recovers.
#[test]
fn salvage_drops_the_tail_past_a_fake_oversized_block_header() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(128)
        .use_partitioned_index();
    for i in 0u64..256 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Re-stamp a block a quarter of the way in so its header FRAMES (valid
    // header checksum) but claims a size reaching the three-quarter mark — a
    // fake oversized header whose enlarged payload cannot load. The blocks it
    // swallows hold real, mid-table keys.
    let (forge_off, forged_end) = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|h| *h.as_ref().offset())
            .collect();
        let n = offsets.len();
        assert!(n >= 8, "need several data blocks, got {n}");
        let (Some(&off), Some(&end)) = (offsets.get(n / 4), offsets.get(3 * n / 4)) else {
            panic!("the quarter and three-quarter block boundaries exist");
        };
        (off, end)
    };
    crate::test_forge::forge_data_block_oversized_header(&source, forge_off, forged_end)?;

    // Corrupt the `index` section so enumeration breaks and the physical walk
    // tiles the whole data section, reaching the fake header.
    let (index_pos, index_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = crate::sfa::Reader::from_reader(&mut f)?;
        let Some((pos, len)) = reader
            .toc()
            .iter()
            .find(|e| e.name() == b"index")
            .map(|e| (e.pos(), e.len()))
        else {
            panic!("a partitioned-index SST must carry an index section");
        };
        (pos, len)
    };
    let Ok(flip) = usize::try_from(index_pos + index_len / 2) else {
        panic!("the index-section offset fits usize");
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    // The anchored prefix before the forge (a quarter of the way in) recovers.
    assert!(
        reopen_get(dest.clone(), &fs, b"key-010")?.is_some(),
        "the contiguous prefix before the fake header must recover: {report:?}",
    );
    // A key inside the swallowed span is reachable only by resync past the fake
    // header, so it must NOT be re-emitted.
    assert!(
        reopen_get(dest, &fs, b"key-128")?.is_none(),
        "a block reached only by resync past the fake header must drop: {report:?}",
    );
    // Only the anchored prefix survives, so the count reflects the first
    // quarter, not the near-full 256 an unbounded resync would emit.
    assert!(
        report.entries_salvaged < 200,
        "the resync tail must not be recovered, got {} of 256: {report:?}",
        report.entries_salvaged,
    );
    assert!(
        report.dropped.iter().any(|d| matches!(
            &d.reason,
            DropReason::HeaderCorrupted(msg) if msg.contains("unanchored")
        )),
        "the swallowed tail must be reported as an unanchored drop: {report:?}",
    );
    Ok(())
}

/// An INTERIOR handle omitted from both forged mirrors leaves the hidden
/// block between two indexed neighbours, exercising the mid-list gap
/// probe (the truncated-mirror sibling only covers the section-tail
/// probe). The block must be framed from the physical tiling and
/// recovered like any indexed block.
#[test]
fn salvage_recovers_an_interior_block_hidden_by_forged_tli_mirrors() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    crate::test_forge::forge_tli_mirrors_drop_interior(&source, 0, None)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(
        report.dropped.is_empty(),
        "every block is intact, nothing may drop: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 64,
        "the interior block both forged mirrors hide must still be recovered: {report:?}",
    );
    Ok(())
}

/// BOTH TLI mirrors re-encoded as a SINGLE handle spanning the whole data
/// section pass the cumulative tiling (one span covers the section), the
/// mirror comparison, and the separator cross-check (only the FIRST
/// payload decodes; the tail reads as an unrecognized trailer on a
/// non-ECC block) — yet every later physical block is unreachable through
/// the index, so reads silently miss its keys after the table is
/// accepted. Each handle must frame EXACTLY one physical block.
#[test]
fn verify_tli_mirrors_rejects_a_section_spanning_handle() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Sanity: the intact table passes the mirror gate.
    open(source.clone(), &fs)?.verify_tli_mirrors()?;

    crate::test_forge::forge_tli_mirrors_span_single_handle(&source, 0)?;

    let table = open(source, &fs)?;
    // Match the frame-check reason specifically: the gate returns several
    // distinct InvalidHeader reasons, and an unrelated one would keep this
    // green without proving the per-handle physical-frame check ran.
    let Err(err) = table.verify_tli_mirrors() else {
        panic!("a handle spanning several physical blocks must fail the mirror gate");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "an index handle's size disagrees with its block's physical frame"
            )
        ),
        "the spanning handle must be rejected by the physical-frame check, got {err:?}",
    );
    Ok(())
}

/// BOTH TLI mirrors re-encoded as a SINGLE handle spanning the whole data
/// section must not blind the SALVAGE walk: advancing the physical cursor
/// by the untrusted indexed size skips every later block, and the oversized
/// non-ECC handle still decodes its first payload — one block "salvaged",
/// zero drops, and repair installs a copy with the rest silently lost. The
/// tiler must trust each handle's span only after the block's own header
/// confirms it, falling back to the physically framed span otherwise.
#[test]
fn salvage_recovers_all_blocks_under_a_section_spanning_handle() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    crate::test_forge::forge_tli_mirrors_span_single_handle(&source, 0)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(
        report.dropped.is_empty(),
        "every block is intact, nothing may drop: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 64,
        "the blocks a spanning handle hides must still be recovered: {report:?}",
    );
    Ok(())
}

/// An index handle whose offset sits BEYOND the data section (a checksum-
/// repatched / forged entry) must not set the gap-probe's upper bound: probing
/// the range from the cursor up to that out-of-section offset would scan past
/// the section, potentially to an attacker-controlled `u64` (an unbounded hang,
/// and later SST sections read as candidate data frames). The tiler must skip
/// the out-of-section handle and bound the walk by the section end.
#[test]
fn salvage_ignores_an_index_handle_beyond_the_data_section() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    let n = 64u32;
    for i in 0..u64::from(n) {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Append an index handle at u64::MAX/2 (far past the data section). The real
    // handles stay intact, so a walk that bounds the gap probe to the section
    // recovers every block and finishes; the pre-fix walk scans up to the bogus
    // offset instead. The nextest slow-timeout terminates a hang, so completing
    // at all — with the full key range recovered — is the proof.
    crate::test_forge::forge_tli_mirrors_offset_beyond_section(&source, 0)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every real block is recovered once the out-of-section handle is skipped: {report:?}",
    );
    assert!(
        reopen_get(dest, &fs, b"key-060")?.is_some(),
        "a late key must survive the bounded walk: {report:?}",
    );
    Ok(())
}

/// A cleanly enumerated index handle whose block header does NOT frame and
/// whose stored span is oversized (a spanning forge with a smashed header)
/// must not advance the physical cursor by that UNVERIFIED size: trusting it
/// covers the whole remaining data section, hiding later blocks from the gap
/// walk. The tiler leaves the cursor where it is and lets the physical walk
/// tile the gap, but the smashed header breaks the contiguous chain at the
/// section start, so the tail is reachable only by unprovable byte-scan resync.
/// The walk must fail closed and drop that tail rather than emit it.
#[test]
fn salvage_drops_the_tail_after_an_unframeable_oversized_handle() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Resolve the FIRST data block's offset (the block the spanning handle
    // keeps) BEFORE the forge — the data section is first, so the tli rewrite
    // that follows only shifts later sections and leaves this offset valid.
    let first_off = {
        let table = open(source.clone(), &fs)?;
        let Some(off) = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|h| *h.as_ref().offset())
            .next()
        else {
            panic!("the source carries data blocks");
        };
        off
    };

    // Collapse both TLI mirrors into ONE handle spanning the whole data section
    // (size = sum of every block size), then smash that block's header first
    // byte so it cannot frame: the handle enumerates cleanly through the
    // mirror gate, but its oversized span is now UNVERIFIABLE.
    crate::test_forge::forge_tli_mirrors_span_single_handle(&source, 0)?;
    {
        let mut bytes = std::fs::read(&source)?;
        let Ok(at) = usize::try_from(first_off) else {
            panic!("data block offset {first_off} fits usize");
        };
        let Some(b) = bytes.get_mut(at) else {
            panic!("the block header lies within the file");
        };
        *b ^= 0xFF;
        std::fs::write(&source, &bytes)?;
    }

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    // The smashed header is the FIRST data block, so the contiguous chain never
    // starts: the whole section is reachable only by unprovable resync and must
    // drop. Nothing is recovered, so no salvaged table is written.
    assert_eq!(
        report.entries_salvaged, 0,
        "nothing is anchored, so no entry is recovered: {report:?}",
    );
    assert!(
        !dest.exists(),
        "an all-dropped section produces no salvaged table: {report:?}",
    );
    assert!(
        report.dropped.iter().any(|d| matches!(
            &d.reason,
            DropReason::HeaderCorrupted(msg) if msg.contains("broken")
        )),
        "the dropped section must be reported as a broken-chain drop: {report:?}",
    );
    Ok(())
}

/// A forged `data` SFA-section length whose `pos + len` OVERFLOWS `u64` breaks
/// the TOC tiling: the catalogue can no longer prove that no deletion section is
/// concealed, so standalone salvage fails CLOSED, the same decision repair
/// reaches by dropping such a table. It must reject PROMPTLY (the coverage
/// check reads the TOC, never scans to the overflowed bound), so the hang the
/// naive byte-at-a-time resync would suffer is avoided by refusing before the
/// walk rather than by tiling to a rejected bound.
#[test]
fn salvage_refuses_an_overflowing_data_section() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    let n = 64u32;
    for i in 0..u64::from(n) {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Forge the `data` section's advertised length to `u64::MAX` so its end
    // overflows: the tiling can no longer be proven, so a deletion section could
    // be hidden behind the oversized span.
    crate::test_forge::forge_section_len(&source, b"data", u64::MAX)?;

    // The nextest slow-timeout terminates a hang; a prompt error (before any
    // scan to the overflowed bound) is the proof the coverage check rejected
    // the catalogue instead of tiling to it.
    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("an overflowing data section must fail salvage closed");
    };
    assert!(
        matches!(err, crate::Error::FeatureUnsupported(msg) if msg.contains("TOC may hide a deletion section")),
        "the refusal must name the TOC-coverage gate, not any other refusal, got {err:?}",
    );
    assert!(
        !std::path::Path::new(&dest).exists(),
        "no salvaged copy is produced when the catalogue may hide a deletion",
    );
    Ok(())
}

/// A source whose `seqno_bounds` section is PRESENT but does not decode must
/// fail SALVAGE closed on a table that exposes NO deletion metadata: a
/// re-stamped TOC can rename a `range_tombstones` / `delete_bitmap` section to
/// `seqno_bounds` and re-role its block, leaving a uniquely named, tiled
/// catalogue whose parsed table reports no deletion. Salvage re-derives the
/// seqno bounds from the recovered entries, so it would discard that section
/// and re-emit the suppressed rows as live. A genuinely rotted seqno-bounds
/// section is indistinguishable from the relabel, so both fail closed; those
/// rows come back from a replica, a checkpoint plus journal replay, or a backup.
#[test]
fn salvage_refuses_a_corrupt_seqno_bounds_that_may_hide_a_deletion() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_seqno_in_index(true);
    for i in 0u64..50 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Rot one payload byte of the seqno_bounds block WITHOUT re-stamping its
    // checksum: the recover-time load fails and degrades the map to empty.
    {
        let pos = {
            let mut f = std::fs::File::open(&source)?;
            let reader = match crate::sfa::Reader::from_reader(&mut f) {
                Ok(r) => r,
                Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
            };
            let Some(entry) = reader.toc().iter().find(|e| e.name() == b"seqno_bounds") else {
                panic!("the source carries a seqno_bounds section");
            };
            let Ok(pos) = usize::try_from(entry.pos()) else {
                panic!("pos fits usize");
            };
            pos
        };
        let mut bytes = std::fs::read(&source)?;
        // Past the block header, inside the payload.
        let at = pos + 40;
        let Some(slot) = bytes.get_mut(at) else {
            panic!("payload byte within the file");
        };
        *slot ^= 0xFF;
        std::fs::write(&source, bytes)?;
    }

    // Salvage fails closed: the seqno-bounds section did not decode and the
    // table exposes no deletion, so it may be a relabeled deletion salvage
    // would discard — refuse instead of resurrecting rows.
    let Err(err) = salvage_sst(&source, dest, &fs) else {
        panic!("a corrupt seqno-bounds section with no visible deletion must fail salvage");
    };
    assert!(
        matches!(err, crate::Error::FeatureUnsupported(_)),
        "the refusal names the unsupported salvage, got {err:?}",
    );
    Ok(())
}

/// When BOTH meta mirrors decode under the expected id but DIVERGE (a forged,
/// internally-consistent tail: `compression#data` re-stamped None -> Lz4,
/// `meta_mid` untouched), the tail-first open decodes every data block under
/// the wrong codec, drops them all, and repair would discard a table whose
/// intact MID mirror recovers everything. Salvage must arbitrate the mirrors
/// and keep the attempt that recovers more.
#[cfg(feature = "lz4")]
#[test]
fn salvage_arbitrates_divergent_meta_mirrors() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Forge the TAIL meta's data-block compression from the written None
    // (tag 0) to Lz4 (tag 1) — fresh block checksum, `meta_mid` untouched.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(report.blocks_total > 0, "the walk saw the data blocks");
    assert_eq!(
        report.blocks_salvaged, report.blocks_total,
        "every block is recoverable through the intact MID mirror: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 100,
        "all rows recovered: {report:?}"
    );
    assert!(
        report.dropped.is_empty(),
        "nothing should drop under the intact mirror: {report:?}",
    );
    assert!(report.salvaged_path.is_some(), "a copy was written");

    // The MID attempt's sibling temp path must not survive the arbitration:
    // the winner is renamed over dest, the loser is discarded.
    for entry in std::fs::read_dir(dir.path())? {
        let name = entry?.file_name().to_string_lossy().into_owned();
        assert!(
            !name.contains(".healtmp-"),
            "the arbitration temp file must be renamed or removed, found {name:?}",
        );
    }
    Ok(())
}

/// A MID-arbitration publish whose post-rename directory sync FAILS must not
/// leave the owned destination behind: the rename already populated `dest`,
/// so returning the durability error without cleanup would leave a partial
/// SST a standalone retry's `create_new` open trips over (and the repair
/// caller's rebuilt manifest omits, leaving an orphan). Every other salvage
/// failure path removes its owned destination; this one must too.
///
/// `lz4`-gated: the divergent-mirror arbitration is driven by a
/// `compression#data` forge to Lz4, so the tail attempt only mis-decodes
/// (and the MID attempt wins, taking the publish path) when lz4 is built.
#[cfg(feature = "lz4")]
#[test]
fn salvage_removes_the_mid_copy_when_the_publish_dir_sync_fails() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Divergent mirrors so the MID attempt wins and publishes via rename.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // Fail the publish's post-rename directory sync. The tail attempt
    // mis-decodes under the forged codec and drops its only block, so it
    // never finishes; a table writer syncs only its file, so the publish
    // after the MID rename is the first directory sync. Fire on it.
    injector.arm(FaultRule::new(FaultOp::SyncDirectory, Fault::Error(ErrorKind::Other)).once());

    // The error must be the INJECTED directory-sync fault: only the publish
    // path both consumes the fault and propagates it. If a future change
    // adds an earlier directory sync, the fault fires there instead and this
    // assertion flags that the test no longer covers the publish-sync
    // cleanup path.
    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("a failed publish directory sync must fail the salvage");
    };
    assert!(
        err.to_string().contains("injected fault on SyncDirectory"),
        "the salvage error must be the injected publish-sync fault, got {err:?}",
    );
    assert!(
        !std::path::Path::new(&dest).exists(),
        "the owned MID copy must be removed when the publish sync fails",
    );
    for entry in std::fs::read_dir(dir.path())? {
        let name = entry?.file_name().to_string_lossy().into_owned();
        assert!(
            !name.contains(".healtmp-"),
            "the MID temp copy must not survive either, found {name:?}",
        );
    }
    Ok(())
}

/// The MID publish must fall back to a best-effort rename when the backend
/// leaves `Fs::hard_link` unsupported. Such a backend can still create and
/// rename ordinary files, so a divergent-meta salvage must still publish the
/// recovered copy rather than drop a recoverable table.
#[cfg(feature = "lz4")]
#[test]
fn salvage_publishes_the_mid_copy_when_hard_link_is_unsupported() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Divergent mirrors so the MID attempt wins and reaches the publish path.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // The backend cannot hard-link: the publish must fall back to a rename.
    injector.arm(FaultRule::new(
        FaultOp::HardLink,
        Fault::Error(ErrorKind::Unsupported),
    ));

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.salvaged_path.as_deref(),
        Some(dest.as_path()),
        "the MID copy must be published via the rename fallback: {report:?}",
    );
    assert!(
        report.entries_salvaged > 0,
        "the recovered table must carry entries: {report:?}",
    );
    assert!(
        std::path::Path::new(&dest).exists(),
        "the recovered table must land at the destination",
    );
    Ok(())
}

/// A MID publish that cannot claim `dest` (the link fails, or on a backend
/// without links the claim or the rename after it does) is a failure of the
/// OUTPUT: the salvage reports it as one, so repair stops instead of grading
/// the source as damage, and it leaves neither `dest` nor its temp copy behind.
///
/// `lz4`-gated for the same reason as the tests around it: the MID attempt
/// only runs when the meta mirrors diverge.
#[cfg(feature = "lz4")]
#[test]
fn a_mid_publish_that_cannot_claim_the_destination_fails_as_an_output_failure() -> crate::Result<()>
{
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    // Without links: the claim of `dest` fails, or the rename onto it does.
    for (unsupported_link, failing) in [
        (false, None),
        (true, Some(FaultOp::Rename)),
        (true, Some(FaultOp::Open)),
    ] {
        let dir = tempdir()?;
        let source = dir.path().join("source");
        // The temp copies are named `salvaged.healtmp-{n}`, so a fault on
        // this name strikes `dest` alone.
        let dest = dir.path().join("salvaged.sst");
        let fault = FaultFs::new(StdFs);
        let injector = fault.injector();
        let fs: Arc<dyn Fs> = Arc::new(fault);

        let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
        for i in 0u64..100 {
            writer.write(InternalValue::from_components(
                format!("key-{i:03}").into_bytes(),
                format!("val-{i:03}").into_bytes(),
                i + 1,
                ValueType::Value,
            ))?;
        }
        assert!(writer.finish()?.is_some(), "source SST is non-empty");
        crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

        if let Some(op) = failing {
            injector.arm(FaultRule::new(
                FaultOp::HardLink,
                Fault::Error(ErrorKind::Unsupported),
            ));
            injector.arm(
                FaultRule::new(op, Fault::Error(ErrorKind::Other)).on_path(dest.to_string_lossy()),
            );
        } else {
            injector.arm(FaultRule::new(
                FaultOp::HardLink,
                Fault::Error(ErrorKind::Other),
            ));
        }

        let result = salvage_sst(&source, dest.clone(), &fs);
        assert!(
            matches!(result, Err(crate::Error::SalvageOutput(_))),
            "unsupported link: {unsupported_link}, failing: {failing:?}: {result:?}",
        );
        assert!(!dest.exists(), "no claim is left at the destination");
        for entry in std::fs::read_dir(dir.path())? {
            let name = entry?.file_name().to_string_lossy().into_owned();
            assert!(
                !name.contains(".healtmp-"),
                "the temp copy does not survive, found {name:?}"
            );
        }
    }
    Ok(())
}

/// The MID publish must claim `dest` with an atomic no-replace operation, never
/// a blind rename that clobbers an unowned file. When `dest` is already
/// occupied, the tail attempt's `create_new` open fails against it (so the MID
/// attempt wins and reaches the publish path), and the publish must surface an
/// error while leaving the occupant's bytes untouched and leaking no temp copy.
///
/// The narrower check-then-rename TOCTOU — `dest` appearing only AFTER an
/// existence probe reports it free — is not deterministically reachable through
/// the current fault surface (it cannot force `Fs::exists` to report a present
/// file absent); the no-replace `hard_link` publish closes that window
/// structurally rather than by a racy probe.
///
/// `lz4`-gated: the MID attempt only runs when the meta mirrors DIVERGE, driven
/// by a `compression#data` forge to Lz4 (as in the sibling MID-arbitration
/// tests). Without divergence `salvage_with_context` returns the tail attempt
/// and never reaches the publish path this test targets.
#[cfg(feature = "lz4")]
#[test]
fn salvage_does_not_clobber_an_occupied_destination_when_the_mid_attempt_wins() -> crate::Result<()>
{
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Diverge the meta mirrors so the MID attempt actually runs (the tail
    // attempt mis-decodes under the forged Lz4 codec and drops its block).
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // A racing worker already owns `dest`. The tail attempt's `create_new` open
    // fails against it, so the tail errors and the MID attempt (writing to a
    // temp) wins arbitration and must PUBLISH into the occupied path.
    std::fs::write(&dest, b"racing worker's file")?;

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "publishing over an occupied destination must fail, got {result:?}",
    );
    assert_eq!(
        std::fs::read(&dest)?,
        b"racing worker's file",
        "the occupant's bytes must survive the failed publish",
    );
    for entry in std::fs::read_dir(dir.path())? {
        let name = entry?.file_name().to_string_lossy().into_owned();
        assert!(
            !name.contains(".healtmp-"),
            "the MID temp copy must not leak, found {name:?}",
        );
    }
    Ok(())
}

/// A STALE arbitration temp file left by a crashed predecessor must not
/// block the MID attempt: the temp-name sequence is process-local, so a
/// fresh process would otherwise pick the same `.healtmp-0` name, fail its
/// `create_new` open, and return the tail attempt's inferior (or failing)
/// result without ever trying the recoverable MID mirror — and tree
/// recovery only sweeps this namespace inside table folders, so a
/// standalone salvage destination stays blocked indefinitely. The
/// arbitration must probe forward to a free name; the foreign artifact is
/// never reclaimed (it may belong to a concurrently running salvage).
///
/// `lz4`-gated: the divergent-mirror arbitration is driven by a
/// `compression#data` forge to Lz4 (see the sibling MID-cleanup test).
#[cfg(feature = "lz4")]
#[test]
fn salvage_arbitration_skips_a_stale_crash_artifact() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Divergent mirrors so the MID attempt wins and publishes via rename.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // A crashed predecessor's artifact occupies the FIRST temp name this
    // process would pick (the sequence is a process-local counter and
    // nextest runs each test in its own process, so it starts at zero).
    let stale = dest.with_extension("healtmp-0");
    std::fs::write(&stale, b"crashed predecessor artifact")?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert_eq!(
        report.entries_salvaged, 100,
        "the MID attempt must pick a fresh temp name past the stale artifact: {report:?}",
    );
    assert!(
        std::path::Path::new(&stale).exists(),
        "a foreign artifact is never reclaimed",
    );
    assert_eq!(
        std::fs::read(&stale)?,
        b"crashed predecessor artifact".to_vec(),
        "the stale artifact's bytes stay untouched",
    );
    Ok(())
}

/// When two processes salvage the SAME divergent-mirror SST concurrently, the
/// process-local temp counter can hand both the same `.healtmp-0` name if one's
/// existence probe RACES the other's creation. The loser's `create_new` then
/// fails `AlreadyExists`, and it must NOT discard that path: the file there is
/// the winner's in-progress output. A stale existence probe (a Metadata fault
/// that reports the occupied name absent) plus a pre-existing foreign temp
/// reproduces the race: the tail attempt loses, mid wins, and the foreign
/// `.healtmp-0` must survive the arbitration cleanup.
///
/// `lz4`-gated: the divergent-mirror arbitration is driven by a
/// `compression#data` forge to Lz4.
#[cfg(feature = "lz4")]
#[test]
fn salvage_arbitration_keeps_a_foreign_temp_it_did_not_create() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Divergent mirrors so the arbitration runs both attempts.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // A concurrent process's in-progress temp already occupies the FIRST name
    // this process would pick.
    let foreign = dest.with_extension("healtmp-0");
    let sentinel = b"concurrent salvage in-progress output".to_vec();
    std::fs::write(&foreign, &sentinel)?;

    // The tail attempt's existence probe RACES: a Metadata fault makes it see
    // `.healtmp-0` as absent, so it returns that name and the tail's create_new
    // then fails AlreadyExists against the foreign file. The mid attempt picks
    // `.healtmp-1` and wins.
    injector.arm(
        FaultRule::new(FaultOp::Metadata, Fault::Error(ErrorKind::Other))
            .on_path("healtmp-0")
            .once(),
    );

    let report = salvage_sst(&source, dest, &fs)?;
    assert_eq!(
        report.entries_salvaged, 100,
        "the mid attempt recovers every entry: {report:?}",
    );
    assert!(
        std::path::Path::new(&foreign).exists(),
        "the foreign concurrent temp must not be discarded by this salvage",
    );
    assert_eq!(
        std::fs::read(&foreign)?,
        sentinel,
        "the foreign temp's bytes stay untouched",
    );
    Ok(())
}

/// A TRANSIENT failure to open the source for the mirror-divergence probe must
/// not be laundered into `Ok(false)`: that would skip the dual-mirror
/// arbitration and salvage from the tail view alone, so a divergent source (a
/// forged tail layout) could omit healthy blocks the intact MID mirror keeps.
/// The source is being salvaged, so it EXISTS — an open failure here is
/// retryable I/O and must propagate, aborting the salvage so repair retries.
///
/// A single `Open` fault on the source hits exactly the probe's open (the first
/// source open in the flow); the later `salvage_attempt` opens then succeed, so
/// WITHOUT the fix salvage completes on the tail alone (`Ok`). WITH the fix it
/// returns the propagated I/O error.
///
/// `lz4`-gated: the divergent mirror is forged via `compression#data` → Lz4.
#[cfg(feature = "lz4")]
#[test]
fn salvage_propagates_a_transient_open_failure_from_the_mirror_probe() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Divergent mirrors: without the fault, the probe would return `Ok(true)`
    // and run the arbitration. The transient open fault must abort instead.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    injector.arm(
        FaultRule::new(FaultOp::Open, Fault::Error(ErrorKind::Other))
            .on_path("source")
            .once(),
    );

    let result = salvage_sst(&source, dest, &fs);
    assert!(
        matches!(result, Err(crate::Error::Io(_))),
        "a transient open failure in the mirror-divergence probe must propagate, not \
         skip arbitration and salvage tail-only: {result:?}",
    );
    Ok(())
}

/// The ENVIRONMENTAL counterpart: a `PermissionDenied` (or any access failure
/// that does not implicate the bytes) while reading ONE metadata mirror must
/// propagate too, not fall through to "mirrors agree". Falling through skips
/// arbitration and salvages from the remaining mirror alone — and if that one
/// only supports a partial recovery, the repair publishes the lossy result
/// even though a fixed environment would have recovered more from the other.
///
/// `lz4`-gated: the divergent mirror is forged via `compression#data` → Lz4.
#[cfg(feature = "lz4")]
#[test]
fn salvage_propagates_an_environmental_failure_from_the_mirror_probe() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // Fault the positioned reads the probe makes on the mirrors themselves —
    // one mirror unreadable, the other fine — sweeping the skip count so the
    // fault lands inside the probe wherever its reads fall.
    let mut reached = false;
    for skip in 0..24u64 {
        injector.clear();
        injector.arm(
            FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::PermissionDenied))
                .on_path("source")
                .skip(skip)
                .times(1),
        );
        let options = SalvageOptions {
            encryption: None,
            #[cfg(zstd_any)]
            zstd_dictionary: None,
            #[cfg(zstd_any)]
            zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
            table_id: 0,
            expected_stored_id: None,
            output_id: None,
            allow_delete_resurrection: false,
            sync_mode: crate::fs::SyncMode::Normal,
            prefix_extractor: None,
            blob_rewrite: None,
            owner_promotions: None,
            progress: None,
            clock: None,
        };
        match super::meta_mirrors_diverge(&source, &fs, &options) {
            Ok(_) => {}
            Err(crate::Error::Io(io)) if io.kind() == ErrorKind::PermissionDenied => {
                reached = true;
            }
            Err(e) => panic!("unexpected probe failure at skip {skip}: {e:?}"),
        }
    }
    assert!(
        reached,
        "no skip count made the probe read a mirror; the sweep proves nothing",
    );
    let _ = dest;
    Ok(())
}

/// A PRE-EXISTING destination must survive a divergent-mirror salvage: the
/// tail attempt correctly fails its `create_new` open, but the MID attempt
/// writes to a sibling temp path and wins the arbitration — publishing it
/// by renaming over `dest` would silently destroy an unrelated file that an
/// API call refusing "destination occupied" is supposed to leave untouched.
///
/// `lz4`-gated: the divergent-mirror arbitration is driven by a
/// `compression#data` forge to Lz4 (see the sibling MID-cleanup test).
#[cfg(feature = "lz4")]
#[test]
fn salvage_refuses_to_overwrite_a_preexisting_destination() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Divergent mirrors, so the arbitration runs a MID attempt after the
    // tail attempt fails on the occupied destination.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    // An unrelated file already occupies the destination.
    let sentinel = b"unrelated pre-existing file".to_vec();
    std::fs::write(&dest, &sentinel)?;

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "an occupied destination must fail the salvage: {result:?}",
    );
    assert_eq!(
        std::fs::read(&dest)?,
        sentinel,
        "the pre-existing destination must survive byte-for-byte",
    );
    // The losing / refused MID copy must not leak its temp file either.
    for entry in std::fs::read_dir(dir.path())? {
        let name = entry?.file_name().to_string_lossy().into_owned();
        assert!(
            !name.contains(".healtmp-"),
            "the refused MID attempt must clean up its temp copy, found {name:?}",
        );
    }
    Ok(())
}

/// DIVERGENT meta mirrors disable the verbatim copy-through: a divergence
/// confined to a DECODE-TRANSPARENT layout field (`restart_interval#data` —
/// full block decoding is trailer-driven, so every block still reads clean
/// and nothing drops) lets the tail-first attempt finish perfectly, yet a
/// byte-copied block keeps the ORIGINAL encoding while the copy's meta is
/// stamped with the forged interval. The partial-decode read path
/// reconstructs prefix-compressed entries FROM that metadata interval, so
/// the inconsistent pair silently truncates synthesized blocks. Since
/// neither mirror is provably genuine, every block must be RE-ENCODED under
/// the chosen meta — self-consistent whichever mirror wins.
#[test]
fn salvage_reencodes_all_blocks_when_meta_mirrors_diverge() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Forge the TAIL meta's data restart interval from the written 16 to 4 —
    // fresh block checksum, `meta_mid` untouched. Blocks keep decoding
    // (trailer-driven), so the tail attempt sees zero drops.
    crate::test_forge::forge_tail_meta_value(&source, b"restart_interval#data", &[4])?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert_eq!(
        report.entries_salvaged, 100,
        "all rows recovered: {report:?}"
    );
    assert!(report.salvaged_path.is_some(), "a copy was written");
    assert_eq!(
        report.blocks_copied_verbatim, 0,
        "divergent mirrors must force the re-encode path: a byte-copied \
         block would keep the original encoding under the chosen meta's \
         forged layout: {report:?}",
    );
    Ok(())
}

/// When the meta mirrors diverge ONLY in `created_at`, both salvage attempts
/// recover identical blocks and entries, so the completeness tie-break picks the
/// tail. That is SAFE because the recovered copy re-stamps `created_at` fresh at
/// finish and re-derives every other authoritative field (key range, seqnos,
/// counts) from the actual re-emitted entries; the layout is mirrored but
/// re-encoded, so a backdated tail cannot be laundered into the copy. This
/// regression guards that property: if a future change ever carried the source
/// `created_at` forward, the forged backdated value would surface here.
#[test]
fn salvage_does_not_carry_a_forged_created_at_into_the_copy() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..100 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let backdated: u128 = 1;
    crate::test_forge::forge_tail_meta_value(&source, b"created_at", &backdated.to_le_bytes())?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.entries_salvaged, 100, "{report:?}");
    let recovered = open(dest, &fs)?;
    assert_ne!(
        *recovered.metadata.created_at, backdated,
        "the recovered copy must not carry the forged backdated created_at",
    );
    Ok(())
}

/// The point-read reachability walk must reject a cross-block internal-key
/// order inversion. A checksum-restamped later block that raises the seqno of a
/// key ending the previous block leaves both blocks decoding and probing
/// cleanly on their own, yet the global (user key ASC, seqno DESC) order is
/// broken across the boundary: after reopen the index seeks the first block and
/// a later compaction could persist the stale (lower-seqno) version.
#[test]
fn verify_point_read_reachability_rejects_a_cross_block_seqno_inversion() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Two versions of the SAME key, one per block (a 1-byte block budget forces
    // each entry into its own block), so the key's versions span the boundary:
    // block 0 = kkk@2, block 1 = kkk@1 (equal key, descending seqno: valid).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(1);
    writer.write(InternalValue::from_components(
        b"kkk",
        b"v2",
        2,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"kkk",
        b"v1",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Raise the SECOND block's first entry seqno from 1 to 3: block 0 now ends
    // kkk@2 and block 1 starts kkk@3 (equal key, seqno RAISED across the
    // boundary: the invalid inversion).
    let block1_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: alloc::vec::Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        let Some(&second) = offsets.get(1) else {
            panic!("the key's versions must span two blocks, got {offsets:?}");
        };
        let Ok(off) = usize::try_from(second) else {
            panic!("the block offset fits usize");
        };
        off
    };
    crate::test_forge::forge_raise_data_block_first_seqno(&source, block1_off, 3)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(
        &table,
        crate::table::ReconcileGate::PointReadReachability,
        None,
    );
    assert!(
        matches!(&err, crate::Error::InvalidHeader(msg) if msg.contains("out of order")),
        "a cross-block seqno inversion must be rejected, got {err:?}",
    );
    Ok(())
}

/// A checksum-clean row block that ITERATES to fewer entries than its
/// trailer declares must be dropped, not marked recovered: the entry decoder
/// turns a mid-stream parse failure into an ordinary end of iteration, so a
/// block with a valid prefix and a malformed tail would otherwise be
/// counted salvaged while silently losing the remaining keys (or byte-copied
/// verbatim, still malformed).
#[test]
fn salvage_drops_a_row_block_that_decodes_fewer_entries_than_declared() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..3 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Inflate the trailer's item count by one behind a re-stamped block
    // checksum: iteration now yields fewer entries than the block declares.
    crate::test_forge::forge_inflated_item_count(&source)?;

    let report = salvage_sst(&source, dest, &fs)?;
    assert_eq!(
        report.entries_salvaged, 0,
        "an under-decoding block must not contribute recovered entries: {report:?}",
    );
    assert!(
        report
            .dropped
            .iter()
            .any(|d| matches!(d.reason, DropReason::DecodeError(_))),
        "the count mismatch is a dropped DecodeError, not a clean recovery: {report:?}",
    );
    Ok(())
}

/// Regression: a data block can hold several MVCC versions of one user key
/// (same key, descending seqno). The verbatim copy-through path must accept
/// equal user keys — only columnar *ingest* requires strictly-unique keys — so
/// salvaging such a block recovers every version instead of erroring.
#[test]
fn salvage_recovers_a_block_with_multiple_versions_of_one_key() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // One block holding the same user key at several seqnos, surrounded by unique
    // keys. Valid SST order: user key ascending, seqno descending within a key.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    writer.write(InternalValue::from_components(
        b"a".to_vec(),
        b"a".to_vec(),
        1,
        ValueType::Value,
    ))?;
    for seqno in [3u64, 2, 1] {
        writer.write(InternalValue::from_components(
            b"dup".to_vec(),
            format!("v{seqno}").into_bytes(),
            seqno,
            ValueType::Value,
        ))?;
    }
    writer.write(InternalValue::from_components(
        b"z".to_vec(),
        b"z".to_vec(),
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "a healthy SST with MVCC duplicates salvages cleanly: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged, 5,
        "every version is recovered, including all 3 of `dup`: {report:?}",
    );

    let recovered = open(dest, &fs)?;
    assert_eq!(
        recovered.metadata.item_count, 5,
        "all 5 entries (3 versions of `dup`) are recovered",
    );
    Ok(())
}

/// A block where a weak tombstone is immediately followed by a value for the
/// same key (a reclaimable pair) salvages verbatim and recovers both entries —
/// exercising the reclaimable-weak-tombstone accounting on the copy-through path.
#[test]
fn salvage_recovers_a_reclaimable_weak_tombstone_pair() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // SST order is user key ascending, seqno descending: the weak tombstone
    // (higher seqno) precedes the value it reclaims (lower seqno) for `dup`.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    writer.write(InternalValue::from_components(
        b"a".to_vec(),
        b"a".to_vec(),
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"dup".to_vec(),
        b"".to_vec(),
        3,
        ValueType::WeakTombstone,
    ))?;
    writer.write(InternalValue::from_components(
        b"dup".to_vec(),
        b"v1".to_vec(),
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let report = salvage_sst(&source, dest, &fs)?;
    assert!(
        report.is_complete(),
        "healthy SST salvages cleanly: {report:?}"
    );
    assert_eq!(
        report.entries_salvaged, 3,
        "the weak tombstone and both values are recovered: {report:?}",
    );
    assert!(
        report.blocks_copied_verbatim >= 1,
        "the clean block is copied verbatim: {report:?}",
    );
    Ok(())
}

/// `verify_locator` must validate the SLOT hint, not only the block id: a
/// `Restart`-precision locator can keep the correct block id yet redirect
/// a key's slot to a later restart interval holding an older version, so
/// `point_read_at_slot` returns the stale value without falling back to
/// the sorted index. The `Writer` is driven directly (two versions of one
/// key in one block, `restart_interval = 1`, locator enabled) and the
/// locator section is rebuilt with that key's slot pushed from the newest
/// head (0) to the older head (1).
#[test]
fn verify_locator_rejects_a_slot_hint_pointing_at_an_older_version() -> crate::Result<()> {
    use crate::config::{LocatorPolicyEntry, LocatorPrecision};
    use crate::runtime_config::{ChecksumAlgorithm, KvChecksumPolicy};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_restart_interval(1)
        .use_kv_checksums(KvChecksumPolicy::AllLevels, ChecksumAlgorithm::Xxh3_64)
        .use_locator(LocatorPolicyEntry::Enabled {
            precision: LocatorPrecision::Restart,
            block_id_bits: None,
            slot_bits: None,
        });
    // Restart heads (restart_interval = 1) in one block: key-a@2 (head 0),
    // key-a@1 (head 1), key-b (head 2), key-c (head 3).
    writer.write(InternalValue::from_components(
        b"key-a".to_vec(),
        b"new".to_vec(),
        2,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"key-a".to_vec(),
        b"old".to_vec(),
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"key-b".to_vec(),
        b"vb".to_vec(),
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"key-c".to_vec(),
        b"vc".to_vec(),
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Sanity: the honest locator passes the gates.
    reconcile_clean(&open(source.clone(), &fs)?, None);

    // Rebuild the locator with key-a's slot pushed to head 1 (the older
    // version); key-b / key-c keep their honest heads (2 / 3). block_id 0
    // for the single block.
    let h = |k: &[u8]| crate::hash::hash64(k);
    crate::test_forge::forge_locator_slots(
        &source,
        0,
        &[
            (h(b"key-a"), 0, 1),
            (h(b"key-b"), 0, 2),
            (h(b"key-c"), 0, 3),
        ],
    )?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::Locator, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "locator slot hint does not resolve a key's newest version"
            )
        ),
        "the redirect must be rejected by the slot-hint check, got {err:?}",
    );
    Ok(())
}

/// `verify_locator` must reject a present locator that gives NO answer for a
/// decoded key. The writer omits the section when it cannot build one and
/// otherwise encodes every unique key, so a no-answer locator is a forgery (a
/// `delete_bitmap` relabeled to a locator that resolves nothing). The read path
/// falls back to the sorted index on a miss, but the verifier must treat the
/// unanswered key as corrupt so the relabel cannot pass as deletion-free.
#[test]
fn verify_locator_rejects_a_no_answer_locator() -> crate::Result<()> {
    use crate::config::{LocatorPolicyEntry, LocatorPrecision};
    use crate::runtime_config::{ChecksumAlgorithm, KvChecksumPolicy};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_restart_interval(1)
        .use_kv_checksums(KvChecksumPolicy::AllLevels, ChecksumAlgorithm::Xxh3_64)
        .use_locator(LocatorPolicyEntry::Enabled {
            precision: LocatorPrecision::Restart,
            block_id_bits: None,
            slot_bits: None,
        });
    for k in [b"key-a".as_slice(), b"key-b", b"key-c"] {
        writer.write(InternalValue::from_components(
            k.to_vec(),
            b"v".to_vec(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Sanity: the honest locator passes the gates.
    reconcile_clean(&open(source.clone(), &fs)?, None);

    // Rebuild the locator so key-c resolves to an OUT-OF-RANGE block id: the
    // handle lookup fails and `locate_block` yields no answer for that decoded
    // key (the read path would fall back to the sorted index; the verifier must
    // not). key-a / key-b keep their honest block 0.
    let h = |k: &[u8]| crate::hash::hash64(k);
    crate::test_forge::forge_locator_slots(
        &source,
        0,
        &[
            (h(b"key-a"), 0, 0),
            (h(b"key-b"), 0, 1),
            (h(b"key-c"), 99, 2),
        ],
    )?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::Locator, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "locator gives no answer for a decoded key it should resolve"
            )
        ),
        "the miss must be rejected by the no-answer check, got {err:?}",
    );
    Ok(())
}

/// First four key bytes (keys are `keyNNNNN`, so the prefix is `key0`).
struct FixedLengthPrefix;

impl crate::prefix::PrefixExtractor for FixedLengthPrefix {
    fn name(&self) -> &'static str {
        "fixed-4"
    }

    fn prefixes<'a>(&self, key: &'a [u8]) -> Box<dyn Iterator<Item = &'a [u8]> + 'a> {
        key.get(..4).map_or_else(
            || Box::new(std::iter::empty()) as Box<dyn Iterator<Item = &'a [u8]>>,
            |p| Box::new(std::iter::once(p)),
        )
    }

    fn is_valid_scan_boundary(&self, prefix: &[u8]) -> bool {
        prefix.len() == 4
    }
}

/// `verify_filter` must probe the extractor's PREFIX hashes, not only
/// complete-key hashes: a full filter that claims an extractor's prefixes
/// yet holds none of them still answers every complete-key probe, while a
/// prefix scan under that extractor trusts it to rule the prefixes out and
/// silently omits the table's rows. The writer is given an extractor under
/// the real one's name that emits no prefix, which builds exactly that
/// filter.
#[test]
fn verify_filter_rejects_missing_prefix_hashes() -> crate::Result<()> {
    /// Claims the real extractor's name, emits nothing.
    struct Mute;
    impl crate::prefix::PrefixExtractor for Mute {
        fn name(&self) -> &'static str {
            "fixed-4"
        }

        fn prefixes<'a>(&self, _: &'a [u8]) -> Box<dyn Iterator<Item = &'a [u8]> + 'a> {
            Box::new(std::iter::empty())
        }
    }

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let extractor: Arc<dyn crate::prefix::PrefixExtractor> = Arc::new(FixedLengthPrefix);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_prefix_extractor(Some(Arc::new(Mute)));
    for i in 0u32..100 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let table = open(source, &fs)?;
    // Without an extractor the gate only probes complete keys and passes.
    reconcile_clean(&table, None);
    // WITH the extractor it claims, it probes each key's prefix hash, which
    // the filter never indexed — a false negative on an existing key's
    // prefix.
    let err = reconcile_error(
        &table,
        crate::table::ReconcileGate::Filter,
        Some(&extractor),
    );
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "filter reports an existing key's prefix as definitely absent"
            )
        ),
        "the missing prefix hash must be the rejection reason, got {err:?}",
    );
    Ok(())
}

/// A filter written without an extractor indexes no prefix, and says so: the
/// gate does not probe it for the prefixes of an extractor configured later,
/// which it was never asked to hold, so a healthy table is not judged corrupt
/// when a tree gains an extractor.
#[test]
fn verify_filter_accepts_a_table_written_before_the_extractor() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let extractor: Arc<dyn crate::prefix::PrefixExtractor> = Arc::new(FixedLengthPrefix);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u32..100 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let table = open(source, &fs)?;
    reconcile_clean(&table, Some(&extractor));
    Ok(())
}

/// The point-read reachability gate must require the probe to return the
/// NEWEST version of a key, not merely SOME version. A key spanning two
/// restart intervals has a CONFLICT-marked hash bucket; redirecting that
/// bucket to the later interval (an OLDER version) still makes
/// `point_read(MAX_SEQNO)` return `Some`, so an `is_none` check would pass
/// — yet reads after reopen return the stale value. The `Writer` is driven
/// directly (two versions of one key in one block, `restart_interval = 1`,
/// hashed, footered) because a memtable flush deduplicates shadowed
/// versions and cannot produce this layout.
#[test]
fn verify_point_read_reachability_rejects_a_bucket_redirected_to_an_older_version()
-> crate::Result<()> {
    use crate::runtime_config::{ChecksumAlgorithm, KvChecksumPolicy};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_hash_ratio(2.0)
        .use_data_block_restart_interval(1)
        .use_kv_checksums(KvChecksumPolicy::AllLevels, ChecksumAlgorithm::Xxh3_64);
    // "key-a" twice: newest (seqno 2) then older (seqno 1), so both are
    // restart heads in the first block and the key's bucket conflicts.
    writer.write(InternalValue::from_components(
        b"key-a".to_vec(),
        b"new".to_vec(),
        2,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"key-a".to_vec(),
        b"old".to_vec(),
        1,
        ValueType::Value,
    ))?;
    for k in [b"key-b", b"key-c"] {
        writer.write(InternalValue::from_components(
            k.to_vec(),
            b"v".to_vec(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Sanity: the intact table passes the reachability gate.
    reconcile_clean(&open(source.clone(), &fs)?, None);

    // Redirect the conflicting bucket to binary-index pos 1 — the second
    // restart head, holding the OLDER (seqno 1) version.
    crate::test_forge::forge_hash_index_bucket(&source, b"key-a", 1, None)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(
        &table,
        crate::table::ReconcileGate::PointReadReachability,
        None,
    );
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "a decoded key's point_read does not return its newest version \
                 (an in-block index disagrees with the entries)"
            )
        ),
        "the redirect must be rejected by the newest-version check, got {err:?}",
    );
    Ok(())
}

/// A `filter_tli` block that does not decode as a filter index must fail
/// SALVAGE closed on a table that exposes NO deletion metadata: a re-stamped
/// TOC can rename a `range_tombstones` / `delete_bitmap` section to
/// `filter_tli` and re-role its block, leaving a uniquely named, tiled
/// catalogue whose parsed table reports no deletion. Salvage re-derives the
/// filter from the recovered keys, so it would discard that section and
/// re-emit the suppressed rows as live. A genuinely rotted filter index is
/// indistinguishable from the relabel, so both fail closed; those rows come back
/// from a replica, a checkpoint plus journal replay, or a backup.
#[test]
fn salvage_refuses_a_corrupt_filter_index_that_may_hide_a_deletion() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_partitioned_filter()
        // A tiny partition budget so several filter partitions spill and
        // the writer emits the `filter_tli` top-level index over them.
        .use_meta_partition_size(3);
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Rot one payload byte of the filter_tli block WITHOUT re-stamping its
    // checksum, so loading it fails like any bit-rotted block.
    {
        let pos = {
            let mut f = std::fs::File::open(&source)?;
            let reader = match crate::sfa::Reader::from_reader(&mut f) {
                Ok(r) => r,
                Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
            };
            let Some(entry) = reader.toc().iter().find(|e| e.name() == b"filter_tli") else {
                panic!("the source carries a filter_tli section");
            };
            let Ok(pos) = usize::try_from(entry.pos()) else {
                panic!("pos fits usize");
            };
            pos
        };
        let mut bytes = std::fs::read(&source)?;
        // First payload byte, derived from the block header length (filter_tli
        // is an Index block) so a header-layout change cannot slide the flip
        // into the header and leave the payload untouched.
        let at =
            pos + crate::table::block::Header::header_len(crate::table::block::BlockType::Index);
        let Some(slot) = bytes.get_mut(at) else {
            panic!("payload byte within the file");
        };
        *slot ^= 0xFF;
        std::fs::write(&source, bytes)?;
    }

    // Provably corrupt: a LIVE open fails closed on the rotted filter index.
    assert!(
        open(source.clone(), &fs).is_err(),
        "the rotted filter index must fail a live open",
    );

    // Salvage fails closed: the filter index did not decode and the table
    // exposes no deletion, so it may be a relabeled deletion salvage would
    // discard — refuse instead of resurrecting rows.
    let Err(err) = salvage_sst(&source, dest, &fs) else {
        panic!("a corrupt filter index with no visible deletion must fail salvage");
    };
    assert!(
        matches!(err, crate::Error::FeatureUnsupported(_)),
        "the refusal names the unsupported salvage, got {err:?}",
    );
    Ok(())
}

/// The rebuilt filter must carry the source's PREFIX hashes: the SST names
/// its extractor but cannot run it, so the salvage writer can only index
/// prefixes when the caller threads the tree's extractor through
/// [`SalvageOptions::prefix_extractor`]. With it, the salvaged copy indexes
/// and names the same prefixes, so a prefix scan keeps skipping by it.
#[test]
fn salvage_preserves_prefix_filter_hashes() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let extractor: Arc<dyn crate::prefix::PrefixExtractor> = Arc::new(FixedLengthPrefix);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_prefix_extractor(Some(Arc::clone(&extractor)));
    for i in 0u32..100 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Sanity: the SOURCE's filter answers the prefix probes (keys are
    // `key000NN`, so the indexed 4-byte prefixes are `key0`).
    let prefix_hash = crate::hash::hash64(b"key0");
    {
        let table = open_under(source.clone(), &fs, &extractor)?;
        assert!(table.has_prefix_filter(), "the source filter is asked");
        assert!(
            table.maybe_contains_prefix(prefix_hash)?,
            "the source filter indexes the prefix",
        );
    }

    let options = SalvageOptions {
        prefix_extractor: Some(Arc::clone(&extractor)),
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.entries_salvaged, 100,
        "all rows recovered: {report:?}"
    );

    // A missing prefix hash is a DEFINITE-absent answer from the rebuilt
    // filter — the salvaged copy would silently vanish from prefix scans.
    let table = open_under(dest, &fs, &extractor)?;
    assert!(
        table.has_prefix_filter(),
        "the salvaged copy's filter indexes the extractor's prefixes, and says so"
    );
    assert!(
        table.maybe_contains_prefix(prefix_hash)?,
        "the salvaged copy's filter must keep the source's prefix hashes",
    );
    Ok(())
}

/// Without a prefix extractor, salvage rebuilds a filter of the keys alone
/// and names no extractor: point reads keep their filter, and a prefix scan
/// under the tree's extractor reads the copy instead of trusting a filter
/// that holds none of its prefixes, so no recovered row drops out of it.
#[test]
fn salvage_without_a_prefix_extractor_keeps_a_key_filter_prefix_scans_read_past()
-> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let extractor: Arc<dyn crate::prefix::PrefixExtractor> = Arc::new(FixedLengthPrefix);

    // Source built WITH a prefix extractor: its filter indexes the 4-byte
    // prefixes (`key0` for keys `key000NN`).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_prefix_extractor(Some(Arc::clone(&extractor)));
    for i in 0u32..100 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Salvage via the default path with NO extractor threaded (the CLI / API
    // default), so the source's prefix indexing intent is unknown here.
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.entries_salvaged, 100,
        "all rows recovered: {report:?}"
    );

    let prefix_hash = crate::hash::hash64(b"key0");
    let table = open_under(dest, &fs, &extractor)?;
    assert_eq!(
        table.metadata.filter_hashes,
        Some(100),
        "the copy keeps a filter of its keys"
    );
    assert_eq!(table.metadata.prefix_extractor, None, "naming no extractor");
    assert!(
        !table.has_prefix_filter(),
        "so a prefix scan does not ask it"
    );
    // Its filter holds no prefix: asked, it would rule `key0` out and hide
    // every recovered row from a prefix scan.
    assert!(
        table.maybe_contains_prefix(prefix_hash)?,
        "without the extractor the salvaged copy must not report a prefix as definitely absent",
    );
    Ok(())
}

fn iv(i: u32) -> InternalValue {
    InternalValue::from_components(
        format!("key{i:05}").into_bytes(),
        format!("val{i:05}").into_bytes(),
        1,
        ValueType::Value,
    )
}

/// File offset of the named SFA section, for a fault rule that targets one
/// section's positional read.
fn section_pos(path: &std::path::Path, name: &[u8]) -> u64 {
    let mut f = match std::fs::File::open(path) {
        Ok(f) => f,
        Err(e) => panic!("opening the source failed: {e:?}"),
    };
    let reader = match crate::sfa::Reader::from_reader(&mut f) {
        Ok(r) => r,
        Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
    };
    let Some(entry) = reader.toc().iter().find(|e| e.name() == name) else {
        panic!(
            "source must carry a {} section",
            String::from_utf8_lossy(name)
        );
    };
    entry.pos()
}

/// Re-stamps one block of an unencrypted, uncompressed SST in place: replaces
/// its payload with `new_payload` (which must keep the payload's length, so
/// every later block stays where the index says it is) and recomputes the
/// header's data checksum, so the forgery stays checksum-consistent.
#[cfg(feature = "columnar")]
fn restamp_block(bytes: &mut [u8], at: usize, new_payload: &[u8]) -> crate::Result<()> {
    use crate::coding::{Decode, Encode};
    use crate::table::block::Header;

    let Some(frame) = bytes.get(at..) else {
        panic!("block start within the file");
    };
    let header = Header::decode_from(&mut &frame[..])?;
    assert_eq!(
        new_payload.len(),
        header.data_length as usize,
        "a restamp must keep the payload length",
    );
    let start = Header::header_len(header.block_type);
    let Some(old_payload) = frame.get(start..start + new_payload.len()) else {
        panic!("payload within the file");
    };
    let old = crate::hash::hash128(old_payload);
    let new_header = crate::test_forge::restamp(header, old, new_payload);
    let mut new_block =
        Vec::with_capacity(Header::header_len(header.block_type) + new_payload.len());
    new_header.encode_into(&mut new_block)?;
    new_block.extend_from_slice(new_payload);
    let Some(target) = bytes.get_mut(at..at + new_block.len()) else {
        panic!("block range within the file");
    };
    target.copy_from_slice(&new_block);
    Ok(())
}

/// The payload of the block at `at` in an unencrypted, uncompressed SST.
#[cfg(feature = "columnar")]
fn block_payload(bytes: &[u8], at: usize) -> crate::Result<Vec<u8>> {
    use crate::coding::Decode;
    use crate::table::block::Header;

    let Some(frame) = bytes.get(at..) else {
        panic!("block start within the file");
    };
    let header = Header::decode_from(&mut &frame[..])?;
    let start = Header::header_len(header.block_type);
    let Some(payload) = frame.get(start..start + header.data_length as usize) else {
        panic!("payload within the file");
    };
    Ok(payload.to_vec())
}

/// Where the `group`-th row group of a columnar SST sits, and its directory.
#[cfg(feature = "columnar")]
fn row_group(
    source: &std::path::Path,
    fs: &Arc<dyn Fs>,
    group: usize,
) -> crate::Result<(usize, crate::table::column_page::PageDirectory)> {
    let table = open(source.to_path_buf(), fs)?;
    let Some(kh) = table.data_block_handles().filter_map(Result::ok).nth(group) else {
        panic!("source must have at least {} row groups", group + 1);
    };
    let at = usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX);
    let bytes = std::fs::read(source)?;
    let directory = crate::table::column_page::PageDirectory::decode(&block_payload(&bytes, at)?)?;
    Ok((at, directory))
}

/// Forges the page of `column_id` in the `group`-th row group of a columnar
/// SST: decodes the column, lets `mutate` alter it in place, re-encodes it at
/// the SAME length and re-stamps the page's checksum. The group stays
/// checksum-consistent, so whatever `mutate` broke surfaces as a CONTENT
/// failure, not as an ordinary checksum drop. Returns the group's row count.
#[cfg(feature = "columnar")]
fn forge_row_group_column(
    source: &std::path::Path,
    fs: &Arc<dyn Fs>,
    group: usize,
    column_id: u16,
    mutate: impl FnOnce(&mut crate::table::columnar::Column, u32),
) -> crate::Result<u32> {
    use crate::table::columnar::Column;

    let (group_at, directory) = row_group(source, fs, group)?;
    let row_count = directory.row_count();
    let Some(page) = directory
        .entries()
        .iter()
        .find(|e| e.id.column_id == column_id)
    else {
        panic!("row group holds a page for column {column_id}");
    };
    let directory_len = {
        let bytes = std::fs::read(source)?;
        let header = {
            use crate::coding::Decode;
            let Some(frame) = bytes.get(group_at..) else {
                panic!("row group within the file");
            };
            crate::table::block::Header::decode_from(&mut &frame[..])?
        };
        header.on_disk_size_with(None) as usize
    };
    let page_at =
        group_at + directory_len + directory.pages_start() as usize + page.offset as usize;

    // The page keeps its stamp: the forgery targets its content, not which
    // group it claims to belong to.
    let stamp = directory.stamp_for(page);
    let mut bytes = std::fs::read(source)?;
    let mut column = Column::decode_page(
        &block_payload(&bytes, page_at)?.into(),
        row_count,
        stamp,
        &mut 0,
        &mut crate::table::columnar::DecodeBudget::default(),
    )?;
    mutate(&mut column, row_count);
    // Encoded as the tree wrote it, plain by default, so the forged page
    // keeps the length its frame was restamped around.
    let new_payload = column.encode_page(row_count, stamp, crate::config::ColumnEncoding::Plain)?;
    restamp_block(&mut bytes, page_at, &new_payload)?;
    std::fs::write(source, &bytes)?;
    Ok(row_count)
}

/// A columnar scan that fails on a later page of a group still counts what it
/// copied building the pages before it: those copies were made whatever the
/// scan returns. Forged: the key page of the group's last row page keeps its
/// length and still parses, but one flipped byte leaves its encoding unable to
/// build the column, so it fails only once the earlier row pages were built.
#[cfg(all(feature = "columnar", feature = "metrics"))]
#[test]
fn a_columnar_scan_failing_on_a_later_page_still_counts_its_copies() -> crate::Result<()> {
    use crate::coding::Decode;
    use crate::table::columnar::{COL_USER_KEY, Column, DecodeBudget};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_column_encoding(crate::config::ColumnEncoding::Auto)
        .use_row_group_size(4_096)
        .use_columnar_page_size(256);
    for i in 0..40 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    let (group_at, directory) = row_group(&source, &fs, 0)?;
    let Some(last) = directory
        .entries()
        .iter()
        .rfind(|e| e.id.column_id == COL_USER_KEY)
    else {
        panic!("the group holds key pages");
    };
    assert!(last.row_page > 0, "the group holds several row pages");
    let rows = directory.row_page_rows(last.row_page).unwrap_or_default();
    let stamp = directory.stamp_for(last);
    let mut bytes = std::fs::read(&source)?;
    let directory_len = {
        let Some(frame) = bytes.get(group_at..) else {
            panic!("row group within the file");
        };
        crate::table::block::Header::decode_from(&mut &frame[..])?.on_disk_size_with(None) as usize
    };
    let page_at =
        group_at + directory_len + directory.pages_start() as usize + last.offset as usize;
    let page = block_payload(&bytes, page_at)?;
    let Some(forged) = (0..page.len()).find_map(|at| {
        let mut flipped = page.clone();
        *flipped.get_mut(at)? ^= 1;
        let slice = crate::Slice::from(flipped.clone());
        let builds = Column::decode_page(&slice, rows, stamp, &mut 0, &mut DecodeBudget::default());
        (Column::parse_page(&slice, rows, stamp).is_ok() && builds.is_err()).then_some(flipped)
    }) else {
        panic!("some byte of the key page parses but no longer builds once flipped");
    };
    restamp_block(&mut bytes, page_at, &forged)?;
    std::fs::write(&source, &bytes)?;

    let table = open(source, &fs)?;
    // The page fails when the scan reads its group, so the scan is drained.
    assert!(
        table
            .columnar_scan(&[COL_USER_KEY], None)
            .and_then(Iterator::collect::<crate::Result<Vec<_>>>)
            .is_err(),
        "the forged page fails the scan",
    );
    assert!(
        table.metrics.bytes_copied() > 0,
        "the key pages built before the forged one are counted",
    );
    Ok(())
}

/// Listing a table's page encodings refuses a page whose own column header
/// names another column than the directory filed it under, as every read of
/// the group does, rather than reporting its encoding as the directory's
/// column. Forged: the value page's header renamed to column 9, its length
/// and stamp kept and its checksum re-stamped.
#[cfg(feature = "columnar")]
#[test]
fn listing_page_encodings_refuses_a_page_filed_under_another_column() -> crate::Result<()> {
    use crate::coding::Decode;
    use crate::table::column_page::PageStamp;
    use crate::table::columnar::COL_VALUE;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true);
    for i in 0..20 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );
    assert!(
        crate::inspect::read_column_encodings(&source).is_ok(),
        "the untouched table lists"
    );

    let (group_at, directory) = row_group(&source, &fs, 0)?;
    let Some(value) = directory
        .entries()
        .iter()
        .find(|e| e.id.column_id == COL_VALUE)
    else {
        panic!("the group holds a value page");
    };
    let mut bytes = std::fs::read(&source)?;
    let directory_len = {
        let Some(frame) = bytes.get(group_at..) else {
            panic!("row group within the file");
        };
        crate::table::block::Header::decode_from(&mut &frame[..])?.on_disk_size_with(None) as usize
    };
    let page_at =
        group_at + directory_len + directory.pages_start() as usize + value.offset as usize;
    let mut page = block_payload(&bytes, page_at)?;
    // The column header opens with its id, right after the stamp.
    let Some(id) = page.get_mut(PageStamp::LEN..PageStamp::LEN + 2) else {
        panic!("the page holds a column header");
    };
    id.copy_from_slice(&9u16.to_le_bytes());
    restamp_block(&mut bytes, page_at, &page)?;
    std::fs::write(&source, &bytes)?;

    assert!(
        crate::inspect::read_column_encodings(&source).is_err(),
        "a page under another column's id is refused, not listed as the directory's column",
    );
    Ok(())
}

/// Swaps the first two rows' user keys inside a key column, in place.
/// Equal-length keys keep the `Bytes` framing intact, so the column still
/// decodes and its rows still materialize — only the ordering invariant is
/// broken.
#[cfg(feature = "columnar")]
fn swap_first_two_keys(key_col: &mut crate::table::columnar::Column, row_count: u32) {
    assert!(row_count >= 2, "group holds at least two rows");
    // Key column framing: (row_count + 1) LE u32 offsets, then payload.
    let table_len = (row_count as usize + 1) * 4;
    let off = |data: &[u8], idx: usize| -> usize {
        let Some(b) = data.get(idx * 4..idx * 4 + 4) else {
            panic!("offset {idx} within the frame table");
        };
        u32::from_le_bytes(b.try_into().unwrap_or([0; 4])) as usize
    };
    let (o0, o1, o2) = (
        off(&key_col.data, 0),
        off(&key_col.data, 1),
        off(&key_col.data, 2),
    );
    assert_eq!(o1 - o0, o2 - o1, "adjacent keys are equal-length");
    let len = o1 - o0;
    let Some(first) = key_col.data.get(table_len + o0..table_len + o0 + len) else {
        panic!("first key within the column");
    };
    let first = first.to_vec();
    let Some(second) = key_col.data.get(table_len + o1..table_len + o1 + len) else {
        panic!("second key within the column");
    };
    let second = second.to_vec();
    // Column bytes are an immutable view — rebuild the swapped column.
    let mut swapped = key_col.data.to_vec();
    let Some(dst0) = swapped.get_mut(table_len + o0..table_len + o0 + len) else {
        panic!("first key range within the column");
    };
    dst0.copy_from_slice(&second);
    let Some(dst1) = swapped.get_mut(table_len + o1..table_len + o1 + len) else {
        panic!("second key range within the column");
    };
    dst1.copy_from_slice(&first);
    key_col.data = swapped.into();
}

/// Forges the `group`-th row group's directory to declare `row_count` rows,
/// leaving its row pages and pages alone, and re-stamps the directory's
/// checksum. The directory keeps its length; only its row count field (after
/// the version byte and the page count) changes, so it no longer agrees with
/// its own row pages.
#[cfg(feature = "columnar")]
fn forge_row_group_row_count(
    source: &std::path::Path,
    fs: &Arc<dyn Fs>,
    group: usize,
    row_count: u32,
) -> crate::Result<()> {
    let (group_at, directory) = row_group(source, fs, group)?;
    let mut payload = Vec::new();
    directory.encode_into(&mut payload);
    let Some(field) = payload.get_mut(3..7) else {
        panic!("a directory holds its row count");
    };
    field.copy_from_slice(&row_count.to_le_bytes());
    let mut bytes = std::fs::read(source)?;
    restamp_block(&mut bytes, group_at, &payload)?;
    std::fs::write(source, &bytes)?;
    Ok(())
}

/// Opens an SST as a `Table`, stamping the open with the file's current digest
/// (the source may be corrupt; per-block checksums catch the actual damage).
fn open(path: std::path::PathBuf, fs: &Arc<dyn Fs>) -> crate::Result<Table> {
    open_with_id(path, fs, 0)
}

/// As [`open`] but under an explicit expected table id (the recover
/// cross-checks it against the SST's stored `table_id`).
fn open_with_id(
    path: std::path::PathBuf,
    fs: &Arc<dyn Fs>,
    table_id: crate::TableId,
) -> crate::Result<Table> {
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&**fs, &path)?);
    let mut params = crate::table::RecoverParams::new(
        path,
        checksum,
        table_id,
        Arc::clone(fs),
        default_comparator(),
        Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
    );
    params.descriptor_table = Some(Arc::new(crate::descriptor_table::DescriptorTable::new(8)));
    Table::recover(params)
}

/// As [`open`] but under a tree's prefix extractor, so a prefix probe asks a
/// filter that indexes its prefixes.
fn open_under(
    path: std::path::PathBuf,
    fs: &Arc<dyn Fs>,
    extractor: &Arc<dyn crate::prefix::PrefixExtractor>,
) -> crate::Result<Table> {
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&**fs, &path)?);
    let mut params = crate::table::RecoverParams::new(
        path,
        checksum,
        0,
        Arc::clone(fs),
        default_comparator(),
        Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
    );
    params.prefix_extractor = Some(Arc::clone(extractor));
    Table::recover(params)
}

/// As [`open`] but threads an encryption provider so a keyed SST recovers.
#[cfg(feature = "encryption")]
fn open_encrypted(
    path: std::path::PathBuf,
    fs: &Arc<dyn Fs>,
    encryption: Arc<dyn crate::encryption::EncryptionProvider>,
) -> crate::Result<Table> {
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&**fs, &path)?);
    let mut params = crate::table::RecoverParams::new(
        path,
        checksum,
        0,
        Arc::clone(fs),
        default_comparator(),
        Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
    );
    params.descriptor_table = Some(Arc::new(crate::descriptor_table::DescriptorTable::new(8)));
    params.encryption = Some(encryption);
    Table::recover(params)
}

/// A reopen of a salvaged SST: recover it and return its live item count.
fn reopen_item_count(path: std::path::PathBuf, fs: &Arc<dyn Fs>) -> crate::Result<u64> {
    Ok(open(path, fs)?.metadata.item_count)
}

/// Point-reads `key` from the SST at `path` at the latest snapshot — the
/// LOGICAL visibility check behind the physical row counts (a delete either
/// masks the key or, under the resurrection opt-in, leaves it readable).
fn reopen_get(
    path: std::path::PathBuf,
    fs: &Arc<dyn Fs>,
    key: &[u8],
) -> crate::Result<Option<crate::InternalValue>> {
    open(path, fs)?.get(key, crate::MAX_SEQNO, crate::hash::hash64(key))
}

/// An SST from a KV-separated tree carries a `linked_blob_files` section
/// naming every blob file its `ValueHandle`s point into; blob GC / relocation
/// consults it (via `list_blob_file_references`) to decide whether a blob is
/// still referenced. The salvaged copy must carry the SOURCE's links —
/// omitting the section would let GC rewrite or delete a blob that only this
/// table still references, silently breaking its indirections.
#[test]
fn salvage_preserves_the_source_linked_blob_files() -> crate::Result<()> {
    use crate::AbstractTree;

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A KV-separated tree: large values go to a blob file, the SST holds
    // indirections plus a linked_blob_files section.
    let crate::AnyTree::Blob(tree) = crate::Config::new(
        dir.path(),
        crate::SequenceNumberCounter::default(),
        crate::SequenceNumberCounter::default(),
    )
    .with_kv_separation(Some(crate::KvSeparationOptions::default()))
    .open()?
    else {
        unreachable!("kv separation configured");
    };
    let big = |i: u32| format!("{i:08}").repeat(512);
    for i in 0u32..10 {
        tree.insert(format!("key{i:05}"), big(i), u64::from(i) + 1);
    }
    tree.flush_active_memtable(10)?;
    let source = {
        let binding = tree.index.version_history.read().latest_version();
        let Some(table) = binding.version.iter_tables().next() else {
            panic!("flush produced one table");
        };
        (*table.path).clone()
    };
    drop(tree);

    let dest = dir.path().join("salvaged");
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(report.is_complete(), "healthy SST: {report:?}");

    // The source's blob links survive into the recovered copy.
    let Some(source_links) = open(source, &fs)?.list_blob_file_references()? else {
        panic!("the source carries a linked_blob_files section");
    };
    assert!(!source_links.is_empty(), "the source references blob files");
    let Some(recovered_links) = open(dest, &fs)?.list_blob_file_references()? else {
        panic!("the salvaged copy carries a linked_blob_files section");
    };
    assert_eq!(
        recovered_links, source_links,
        "the salvaged copy references the same blob files as the source",
    );
    Ok(())
}

/// `verify_blob_links` must reject a table whose cell rows own blob objects
/// but whose `owned_blob_objects` section is gone: a drop of the table would
/// then release nothing, and a reference to an object it owned could be
/// written back after it. The honest table passes.
#[test]
fn verify_blob_links_rejects_a_missing_owned_objects_section() -> crate::Result<()> {
    use crate::AbstractTree;
    use crate::blob_tree::field_row::{FIRST_FIELD_COLUMN, Field};

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let crate::AnyTree::Blob(tree) = crate::Config::new(
        dir.path(),
        crate::SequenceNumberCounter::default(),
        crate::SequenceNumberCounter::default(),
    )
    .with_kv_separation(Some(
        crate::KvSeparationOptions::default().separation_threshold(64),
    ))
    .open()?
    else {
        unreachable!("kv separation configured");
    };
    let body = vec![b'b'; 4_096];
    for i in 0u32..4 {
        tree.insert_cells(
            format!("key{i:05}"),
            &[
                Field::bytes(FIRST_FIELD_COLUMN, b"draft"),
                Field::bytes(FIRST_FIELD_COLUMN + 1, &body),
            ],
            u64::from(i) + 1,
        )?;
    }
    tree.flush_active_memtable(10)?;
    let source = {
        let binding = tree.index.version_history.read().latest_version();
        let Some(table) = binding.version.iter_tables().next() else {
            panic!("flush produced one table");
        };
        assert_eq!(table.owned_blob_objects()?.len(), 4, "one owned body a row");
        (*table.path).clone()
    };
    drop(tree);

    open(source.clone(), &fs)?.verify_blob_links()?;

    crate::test_forge::forge_section_omitted(&source, b"owned_blob_objects")?;
    let table = open(source, &fs)?;
    let Err(err) = table.verify_blob_links() else {
        panic!("a table whose owned objects vanished must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "owned_blob_objects disagrees with the objects the table's cell rows own"
            )
        ),
        "the rejection names the owned-objects reason, got {err:?}",
    );
    Ok(())
}

/// `verify_blob_links` must reject a table that still carries indirection
/// entries but advertises NO `linked_blob_files` section (dropped or renamed
/// away): returning `Ok` there lets a healed-digest refresh or salvage accept a
/// table with `ValueHandle`s but no blob references, after which blob GC — which
/// consults `list_blob_file_references()` for liveness — can rewrite or drop a
/// blob file this table still points into.
#[test]
fn verify_blob_links_rejects_a_missing_section_with_live_indirections() -> crate::Result<()> {
    use crate::AbstractTree;

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let crate::AnyTree::Blob(tree) = crate::Config::new(
        dir.path(),
        crate::SequenceNumberCounter::default(),
        crate::SequenceNumberCounter::default(),
    )
    .with_kv_separation(Some(crate::KvSeparationOptions::default()))
    .open()?
    else {
        unreachable!("kv separation configured");
    };
    let big = |i: u32| format!("{i:08}").repeat(512);
    for i in 0u32..10 {
        tree.insert(format!("key{i:05}"), big(i), u64::from(i) + 1);
    }
    tree.flush_active_memtable(10)?;
    let source = {
        let binding = tree.index.version_history.read().latest_version();
        let Some(table) = binding.version.iter_tables().next() else {
            panic!("flush produced one table");
        };
        (*table.path).clone()
    };
    drop(tree);

    // Drop the linked_blob_files section from the TOC: the table keeps its
    // indirection entries but now advertises no blob references.
    crate::test_forge::forge_section_omitted(&source, b"linked_blob_files")?;

    let table = open(source, &fs)?;
    assert!(
        table.list_blob_file_references()?.is_none(),
        "the forge must leave the table with no blob-link section",
    );
    let Err(err) = table.verify_blob_links() else {
        panic!("a table with live indirections but no blob-link section must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "table carries indirection entries but no linked_blob_files section"
            )
        ),
        "the rejection names the missing-blob-link-section reason, got {err:?}",
    );
    Ok(())
}

/// `verify_blob_links` must reject a PRESENT but empty (zero-count)
/// `linked_blob_files` section. The writer omits the section when there are no
/// blob references, so a present-but-blobless section is a forgery: a
/// `delete_bitmap` replaced by a four-byte zero-count `linked_blob_files` leaves
/// both the derived and recorded maps empty, so the equality check passes and
/// the table is kept after its deletion metadata vanished.
#[cfg(feature = "columnar")]
#[test]
fn verify_blob_links_rejects_a_present_empty_section() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A columnar delete-bearing SST with NO blob indirections (so no real
    // linked_blob_files section exists to duplicate).
    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Replace the delete_bitmap section with a four-byte zero-count
    // linked_blob_files (raw shape valid, records no blob references).
    crate::test_forge::forge_rename_and_replace_section(
        &source,
        (b"delete_bitmap", b"linked_blob_files"),
        &[0, 0, 0, 0],
        (0, None),
    )?;

    let table = open(source, &fs)?;
    let Err(err) = table.verify_blob_links() else {
        panic!("a present zero-count linked_blob_files section must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "linked_blob_files section is present but records no blob references"
            )
        ),
        "the rejection names the present-empty blob-link reason, got {err:?}",
    );
    Ok(())
}

/// Standalone salvage must REFUSE a delete-bearing SST whose `delete_bitmap`
/// entry was OMITTED from a re-stamped TOC. The parsed table then reports no
/// deletion (the section's bytes linger unreferenced), no side section
/// degrades, and the mask is not unpositionable, so the relabel and
/// unpositionable guards both pass, and a naive walk would re-emit every
/// physically-present row as live, resurrecting the deleted ones. The repair
/// verifier catches the TOC tiling gap, but `salvage_sst` never runs it: the
/// coverage check must live inside salvage too.
#[cfg(feature = "columnar")]
#[test]
fn salvage_refuses_a_toc_that_omits_a_delete_bitmap() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // OMIT the delete_bitmap TOC entry (bytes remain, nothing references them),
    // re-stamping the trailer so the archive stays internally consistent.
    crate::test_forge::forge_section_omitted(&source, b"delete_bitmap")?;

    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("a TOC that hides a deletion section must be refused, not resurrected");
    };
    assert!(
        matches!(err, crate::Error::FeatureUnsupported(msg) if msg.contains("TOC may hide a deletion section")),
        "the refusal must name the hidden-deletion gate, not any other refusal, got {err:?}",
    );
    assert!(
        !std::path::Path::new(&dest).exists(),
        "no salvaged copy is produced when the deletion section may be hidden",
    );
    Ok(())
}

/// A delete-bearing SST whose `delete_bitmap` section is HIDDEN (renamed away or
/// replaced by another valid optional section) must be rejected by the metadata
/// cross-check: the authenticated `descriptor#delete_bitmap_len` still records
/// the positions, so a `> 0` count with no readable bitmap section is a forgery
/// that would resurrect every positionally-deleted row. Uses omission (a
/// recognized-name rename or a valid filter replacement reaches the same state:
/// the count disagrees with the absent section).
#[cfg(feature = "columnar")]
#[test]
fn verify_metadata_bounds_rejects_a_hidden_delete_bitmap() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Hide the delete_bitmap section while the authenticated meta count still
    // records its 3 positions.
    crate::test_forge::forge_section_omitted(&source, b"delete_bitmap")?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("delete_bitmap count disagrees")),
        "the rejection must name the delete_bitmap count mismatch, got {err:?}",
    );
    Ok(())
}

/// An EQUAL-CARDINALITY `delete_bitmap` substitution (a different, checksum-valid
/// bitmap with the same number of positions) passes the count-only cross-check
/// but must be rejected by the content hash: during manifest repair, with no
/// original whole-file digest, it would otherwise resurrect the originally
/// deleted rows and drop different live ones.
#[cfg(feature = "columnar")]
#[test]
fn verify_metadata_bounds_rejects_an_equal_cardinality_delete_bitmap_substitution()
-> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Replace the bitmap with a DIFFERENT set of three positions (same chunk,
    // same cardinality and encoded length, so the count cross-check still
    // passes) — only the CONTENTS differ.
    crate::test_forge::forge_delete_bitmap_substitute(&source, 0, &[6, 21, 41])?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("delete_bitmap contents disagree")),
        "the rejection must name the delete_bitmap content-hash mismatch, got {err:?}",
    );
    Ok(())
}

/// The SALVAGE walk must authenticate the delete-bitmap CONTENTS, not just its
/// positions: an equal-cardinality substitution (a different, checksum-valid
/// bitmap) is structurally positionable, so without a content check salvage
/// applies it and masks the WRONG rows, resurrecting the authentically-deleted
/// rows and dropping different live ones. Salvage has no original whole-file
/// digest, so with the default (no resurrection opt-in) it must fail closed on an
/// unauthenticated bitmap (#90).
#[cfg(feature = "columnar")]
#[test]
fn salvage_authenticates_the_delete_bitmap_before_masking() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Substitute an equal-cardinality bitmap: same count, different positions,
    // checksum-valid, but the meta content hash still records the original.
    crate::test_forge::forge_delete_bitmap_substitute(&source, 0, &[6, 21, 41])?;

    // Default options do NOT opt into resurrection: salvage must refuse rather
    // than apply the unauthenticated (substituted) mask.
    let result = super::salvage_sst(&source, dest, &fs);
    assert!(
        matches!(&result, Err(crate::Error::InvalidHeader(msg)) if msg.contains("resurrect")),
        "an unauthenticated delete-bitmap substitution must fail closed, got {result:?}",
    );
    Ok(())
}

/// `attempt_owns_temp` must prove ownership from the OUTCOME (a written temp),
/// not infer it from the error kind: a salvage attempt that errors BEFORE
/// `Writer::new` (e.g. a range-tombstone refusal) never created the temp, so the
/// arbitration cleanup must NOT discard that path — on shared storage it could be
/// a concurrent creator's file.
#[test]
fn attempt_owns_temp_tracks_the_written_outcome_not_the_error_kind() {
    let report = |salvaged_path| super::SalvageReport {
        salvaged_path,
        blocks_total: 1,
        blocks_salvaged: 1,
        blocks_copied_verbatim: 0,
        entries_salvaged: 1,
        entries_dropped_by_rewrite: 0,
        dropped: alloc::vec::Vec::new(),
        delete_rows_resurrected: false,
    };

    // Wrote a temp → owned.
    assert!(super::attempt_owns_temp(&Ok(report(Some(
        std::path::PathBuf::from("/x")
    )))));
    // Recovered nothing (empty temp already discarded) → not owned.
    assert!(!super::attempt_owns_temp(&Ok(report(None))));
    // Errored BEFORE create (range tombstones) → temp never created → not owned.
    assert!(!super::attempt_owns_temp(&Err(
        crate::Error::FeatureUnsupported("range tombstones")
    )));
}

/// Divergent-mirror arbitration must not let a TRANSIENT failure on one attempt
/// lose to an INCOMPLETE success on the other: a retry of the transient mirror
/// could recover the blocks the incomplete winner dropped, so the arbitration
/// propagates the transient error. A COMPLETE success still wins (a retry cannot
/// improve on it), and a PERSISTENT failure always loses to any success.
#[test]
fn arbitrate_mirrors_propagates_a_transient_loss_to_an_incomplete_success() {
    use super::{DropReason, DroppedBlock, MirrorArbitration, SalvageReport, arbitrate_mirrors};

    let report = |dropped: usize| SalvageReport {
        salvaged_path: Some(std::path::PathBuf::from("/x")),
        blocks_total: 4,
        blocks_salvaged: 4 - dropped,
        blocks_copied_verbatim: 0,
        entries_salvaged: 4,
        entries_dropped_by_rewrite: 0,
        dropped: (0..dropped)
            .map(|i| DroppedBlock {
                offset: i as u64,
                section: b"data".to_vec(),
                reason: DropReason::ChecksumMismatch,
                key_range: None,
            })
            .collect(),
        delete_rows_resurrected: false,
    };
    let complete = || Ok(report(0));
    let incomplete = || Ok(report(1));
    let transient = || {
        Err(crate::Error::Io(crate::io::Error::from(
            crate::io::ErrorKind::Interrupted,
        )))
    };
    let persistent = || {
        Err(crate::Error::Io(crate::io::Error::from(
            crate::io::ErrorKind::Other,
        )))
    };

    // Transient loss vs an INCOMPLETE success → propagate (either side).
    assert_eq!(
        arbitrate_mirrors(&transient(), &incomplete()),
        MirrorArbitration::Propagate,
    );
    assert_eq!(
        arbitrate_mirrors(&incomplete(), &transient()),
        MirrorArbitration::Propagate,
    );
    // Transient loss vs a COMPLETE success → the complete success wins.
    assert_eq!(
        arbitrate_mirrors(&transient(), &complete()),
        MirrorArbitration::PublishMid,
    );
    assert_eq!(
        arbitrate_mirrors(&complete(), &transient()),
        MirrorArbitration::PublishTail,
    );
    // PERSISTENT failure vs an incomplete success → the success wins (a retry
    // cannot help the persistent failure).
    assert_eq!(
        arbitrate_mirrors(&persistent(), &incomplete()),
        MirrorArbitration::PublishMid,
    );
}

/// `publish_from_temp` must not delete `temp` when the winning attempt ERRORED:
/// a failure BEFORE `Writer::new` never created it, and a failure after already
/// discarded its own partial, so the caller owns nothing to remove. Ownership is
/// proven by the OUTCOME, never inferred from the error kind — otherwise, on
/// shared storage, an attempt whose `temp` path collides with a concurrent
/// process's file (e.g. across PID namespaces with the same numeric id) would
/// delete that process's file. Regression for the winner-publication arm, the
/// sibling of the arbitration loser-cleanup fix.
#[test]
fn publish_from_temp_keeps_a_foreign_temp_on_an_erroring_attempt() -> crate::Result<()> {
    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let temp = dir.path().join("dest.healtmp-0");
    let dest = dir.path().join("dest");

    // A concurrent process owns the temp path; THIS attempt never created it.
    std::fs::write(&temp, b"foreign process output")?;

    // The published attempt errored BEFORE `Writer::new` (a deletion-guard
    // refusal is a FeatureUnsupported, not an AlreadyExists race loss).
    let result = Err(crate::Error::FeatureUnsupported("deletion guard"));
    let Err(err) = super::publish_from_temp(&fs, result, &temp, &dest, &SalvageOptions::default())
    else {
        panic!("an erroring attempt must propagate its error");
    };
    assert!(
        matches!(err, crate::Error::FeatureUnsupported("deletion guard")),
        "the original error propagates unchanged, got {err:?}",
    );
    assert!(
        temp.exists(),
        "the foreign temp must survive: this attempt never created it",
    );
    Ok(())
}

/// A PRESENT `delete_bitmap` section that decodes to an EMPTY bitmap must fail
/// salvage closed. The writer only emits the section when the bitmap is
/// non-empty, so a checksum-consistent corruption to empty is a forge: it keeps
/// the section visible (exempting the concealment guards) yet carries no
/// positions, so the masked columnar path would re-emit every deleted row live
/// without the resurrection opt-in.
#[cfg(feature = "columnar")]
#[test]
fn salvage_refuses_a_present_but_empty_delete_bitmap() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Re-stamp the delete_bitmap section as a valid but EMPTY bitmap.
    crate::test_forge::forge_delete_bitmap_empty(&source, 0)?;

    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("a present-but-empty delete bitmap must fail salvage closed");
    };
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("delete bitmap cannot be applied")),
        "the refusal must name the unpositionable delete mask, got {err:?}",
    );
    assert!(
        !std::path::Path::new(&dest).exists(),
        "no salvaged copy is produced when the delete bitmap is corrupt to empty",
    );
    Ok(())
}

/// A `delete_bitmap` section that is PRESENT and non-empty while the
/// authenticated `descriptor#delete_bitmap_len` records ZERO must be rejected.
/// The writer stamps the count as `0` precisely when no bitmap section is
/// written, so a `0` count paired with a live section is a forge: a
/// checksum-restamped TOC that grafts another table's bitmap (or relabels an
/// optional section to `delete_bitmap`) onto a no-delete table makes reads
/// apply that mask and silently drop live rows. The guard must compare the
/// count and the section state in BOTH directions, not only reject `count > 0`
/// with no section.
#[cfg(feature = "columnar")]
#[test]
fn verify_metadata_bounds_rejects_a_zero_count_with_a_live_bitmap() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Zero the authenticated count in both mirrors while the section keeps its
    // three live positions: the state a TOC forge reaches by grafting a bitmap
    // onto a genuinely no-delete table (recorded count 0).
    crate::test_forge::forge_meta_value_both_mirrors(
        &source,
        b"descriptor#delete_bitmap_len",
        &0u64.to_le_bytes(),
    )?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("delete_bitmap count disagrees")),
        "the rejection must name the delete_bitmap count mismatch, got {err:?}",
    );
    Ok(())
}

/// A forged TLI that REORDERS columnar block handles must fail salvage closed.
/// `delete_block_starts` is built by walking the index, so a reorder rebuilds
/// the starts in that same reordered sequence and `delete_positions_verified`
/// self-validates against them. But the bitmap positions were assigned in the
/// writer's PHYSICAL block order, so the salvage walk (which sorts blocks back
/// to physical order) would mask against the wrong starts, deleting live rows
/// and resurrecting deleted ones without the opt-in. Anchoring the verification
/// to monotonic physical offsets detects the reorder.
#[cfg(feature = "columnar")]
#[test]
fn salvage_refuses_a_reordered_columnar_index_with_deletes() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Small blocks so the SST spills several columnar data blocks; deletes land
    // across them.
    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 60, 120] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Swap the first two block handles in both TLI mirrors. The index and
    // delete_block_starts now agree with each other but not with the physical
    // bitmap positions.
    crate::test_forge::forge_tli_mirrors_swap_first_two(&source, 0, None)?;

    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("a reordered columnar index with deletes must fail salvage closed");
    };
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("delete bitmap cannot be applied")),
        "the refusal must name the unpositionable delete mask, got {err:?}",
    );
    assert!(
        !std::path::Path::new(&dest).exists(),
        "no salvaged copy is produced when the delete bitmap cannot be positioned",
    );
    Ok(())
}

/// `verify_seqno_bounds` must reject a PRESENT `seqno_bounds` section that
/// decodes to an EMPTY map on a table that still holds data blocks: every real
/// writer emits one entry per block, so an empty map (a `delete_bitmap` renamed
/// to `seqno_bounds` and re-stamped empty) means the table's real bounds/deletes
/// metadata was laundered away. Accepting it lets a healed-digest refresh grade
/// the table clean and reopen it without its deletion metadata, resurrecting
/// positionally deleted rows.
#[test]
fn verify_seqno_bounds_rejects_a_present_empty_map_on_a_nonempty_table() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A PLAIN table with the per-block seqno_bounds section enabled.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_seqno_in_index(true)
        .use_data_block_size(128);
    for i in 0u64..64 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Re-stamp the seqno_bounds section as a valid but EMPTY map.
    crate::test_forge::forge_seqno_bounds_empty(&source, 0)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::SeqnoBounds, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader("seqno_bounds is missing a data block's entry")
        ),
        "the rejection names the missing-block-entry reason, got {err:?}",
    );
    Ok(())
}

/// `verify_block_layout` must reject a PRESENT `block_layout` section that
/// decodes to an EMPTY map on a table that carries multi-inner-block frames:
/// the writer only emits the section when it has boundaries to record, so an
/// empty map (a `delete_bitmap` renamed and re-roled to an empty `block_layout`)
/// means the deletion metadata was laundered away. The per-block loop cannot
/// catch it (a block absent from the map is skipped), so accepting it lets a
/// healed-digest refresh keep the table without its delete bitmap.
#[cfg(feature = "zstd")]
#[test]
fn verify_block_layout_rejects_a_present_empty_map_on_a_nonempty_table() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // 256 KiB blocks at zstd L19 split into many inner blocks, so the writer
    // records a real block_layout section.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256 * 1024)
        .use_data_block_compression(crate::CompressionType::Zstd(19));
    for i in 0u64..20_000 {
        writer.write(InternalValue::from_components(
            format!("key-{i:012}").into_bytes(),
            format!("value-{i:08}-payload").into_bytes(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Re-stamp the block_layout section as a valid but EMPTY map.
    crate::test_forge::forge_block_layout_empty(&source, 0, None)?;

    let table = open(source, &fs)?;
    let Err(err) = table.verify_block_layout() else {
        panic!("an empty block_layout on a table with data blocks must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "block_layout section is present but empty on a table with data blocks"
            )
        ),
        "the rejection names the present-empty block_layout reason, got {err:?}",
    );
    Ok(())
}

/// A recorded boundary that no longer marks where an inner zstd block ends
/// must be rejected BY NAME, and by the reconcile walk rather than a pass of
/// its own.
///
/// The forged section stays checksum-clean and structurally plausible: the ends
/// are still strictly increasing and still finish at the block's uncompressed
/// length, so every cheap check passes. Only decoding the frame reveals that an
/// interior boundary moved — and the partial range-read path bounds its
/// decompression by exactly that boundary, so believing it silently omits keys.
///
/// Driven through `verify_reconcile_gates` because that is where the check now
/// lives: it reuses the frame the walk already read, where the standalone pass
/// re-read every data block to get it.
#[cfg(feature = "zstd")]
#[test]
fn reconcile_gates_with_a_shifted_block_layout_boundary_reject_it() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // 256 KiB blocks at zstd L19: each data block splits into several inner
    // zstd blocks, which is the only shape that carries a block_layout.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256 * 1024)
        .use_data_block_compression(crate::CompressionType::Zstd(19));
    for i in 0u64..20_000 {
        writer.write(InternalValue::from_components(
            format!("key-{i:012}").into_bytes(),
            format!("value-{i:08}-payload").into_bytes(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Sanity: the untouched table reconciles clean, so the failure below is
    // the forgery and not the fixture.
    let table = open(source.clone(), &fs)?;
    assert!(
        table.verify_reconcile_gates(None, false).is_ok(),
        "the intact multi-inner-block table must reconcile clean",
    );
    drop(table);

    crate::test_forge::forge_block_layout_shifted_end(&source, 0)?;

    let table = open(source, &fs)?;
    let Err((gate, err)) = table.verify_reconcile_gates(None, false) else {
        panic!("a boundary that does not match the frame's inner blocks must be rejected");
    };
    assert!(
        matches!(gate, crate::table::ReconcileGate::BlockLayout),
        "the failure must be attributed to the block-layout gate, got {gate:?}",
    );
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader("block_layout disagrees with the frames' inner blocks")
        ),
        "the rejection names the frame disagreement, got {err:?}",
    );
    Ok(())
}

/// `verify_block_layout` must run the present-but-empty rejection on builds
/// WITHOUT zstd too. The function used to return `Ok(())` for the whole
/// non-zstd build before reading the section, so a columnar SST with positional
/// deletes whose `delete_bitmap` was relabeled to an empty `block_layout` graded
/// clean and reopened without its deletion metadata. The emptiness check now
/// runs independent of the zstd frame cross-check.
#[cfg(all(feature = "columnar", not(feature = "zstd")))]
#[test]
fn verify_block_layout_rejects_an_empty_map_without_zstd() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Relabel the delete_bitmap section to an empty block_layout (the writer
    // never emits a real one on a non-zstd build, so synthesise the forgery).
    crate::test_forge::forge_delete_bitmap_as_empty_block_layout(&source, 0)?;

    let table = open(source, &fs)?;
    let Err(err) = table.verify_block_layout() else {
        panic!("an empty block_layout on a non-zstd table with data blocks must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "block_layout section is present but empty on a table with data blocks"
            )
        ),
        "the rejection names the present-empty block_layout reason, got {err:?}",
    );
    Ok(())
}

/// `verify_filter` must reject a PRESENT full `filter` section that decodes to
/// the empty "no filter installed" sentinel on a table with data blocks: the
/// writer omits the section when filtering is disabled, so a present-empty
/// filter is a forgery (a `delete_bitmap` renamed and re-roled to an empty
/// filter). The read-path probe treats an empty payload permissively — `Ok(true)`
/// for every key — so without this rejection the relabel passes the whole check,
/// a healed digest keeps the table, and reopening resurrects the deleted rows.
#[test]
fn verify_filter_rejects_a_present_empty_full_filter() -> crate::Result<()> {
    use crate::config::BloomConstructionPolicy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_bloom_policy(BloomConstructionPolicy::BitsPerKey(10.0));
    for i in 0u32..200 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Re-stamp the full filter section as a valid but EMPTY payload.
    crate::test_forge::forge_filter_empty(&source, 0)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::Filter, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "filter section is present but empty on a table with data blocks"
            )
        ),
        "the rejection names the present-empty filter reason, got {err:?}",
    );
    Ok(())
}

/// `verify_filter` must reject an empty PARTITION too, not just an empty full
/// filter. The partitioned path seeks the partition for a key and probes it; on
/// the empty "no filter" sentinel `maybe_contains_hash` answers `Ok(true)` for
/// every key, so a partition the writer would never leave empty (a `delete_bitmap`
/// relabeled to an empty filter partition under a re-stamped `filter_tli`) would
/// pass the probe and keep the table as deletion-free.
#[test]
fn verify_filter_rejects_an_empty_partition_for_an_existing_key() -> crate::Result<()> {
    use crate::config::BloomConstructionPolicy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A partitioned filter with a tiny partition budget so several partitions
    // spill and the writer emits the filter_tli over them.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_partitioned_filter()
        .use_meta_partition_size(3)
        .use_bloom_policy(BloomConstructionPolicy::BitsPerKey(10.0));
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Empty the first partition in place (the one covering the lowest keys).
    crate::test_forge::forge_filter_first_partition_empty(&source)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::Filter, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "filter partition is present but empty for an existing key"
            )
        ),
        "the rejection names the present-empty partition reason, got {err:?}",
    );
    Ok(())
}

/// `verify_zone_map` must reject a PRESENT `zone_map` section that decodes to an
/// EMPTY map on a table with data blocks: the writer emits one entry per data
/// block whenever the section exists, so an empty map (a `delete_bitmap`
/// relabeled and re-roled to an empty `zone_map`) means the deletion metadata
/// was laundered away while every semantic gate still passes.
#[cfg(feature = "columnar")]
#[test]
fn verify_zone_map_rejects_a_present_empty_map_on_a_nonempty_table() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true);
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+zonemap SST is non-empty"
    );

    // Re-stamp the zone_map section as a valid but EMPTY map.
    crate::test_forge::forge_zone_map_empty(&source, 0)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::ZoneMap, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "zone_map section is present but empty on a table with data blocks"
            )
        ),
        "the rejection names the present-empty zone_map reason, got {err:?}",
    );
    Ok(())
}

/// On a ROW table `verify_zone_map` must authenticate the synthetic column's
/// IDENTITY, not only its min / max / row count. The writer stamps every
/// whole-block column with `column_id == 0` and zero type / codec / null
/// fields; a re-stamped map can change that id to a consumer value-column id
/// while leaving the key bounds untouched. The bounds check then passes, repair
/// keeps the table, and `ColumnRangePredicate::can_skip_block` reads those key
/// bounds as value-column statistics and can skip blocks holding matching rows.
#[test]
fn verify_zone_map_rejects_a_forged_synthetic_column_id() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_zone_map(true);
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source row+zonemap SST is non-empty"
    );

    // Repurpose the first block's whole-block key statistic as a value-column
    // statistic by changing its id, leaving min / max / row count intact.
    crate::test_forge::forge_zone_map_column_id(&source, 0)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::ZoneMap, None);
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("zone_map synthetic column")),
        "the rejection must name the synthetic column identity, got {err:?}",
    );
    Ok(())
}

/// On a COLUMNAR table `verify_zone_map` authenticates the FULL per-column map
/// by re-deriving it from each decoded block: a columnar block records one entry
/// per stored column, so a re-stamped id (here the user-key column's id flipped
/// to a consumer value-column id) no longer matches the re-derivation and is
/// rejected. This is the columnar counterpart of the row-block synthetic-column
/// identity check.
#[cfg(feature = "columnar")]
#[test]
fn verify_zone_map_rejects_a_forged_columnar_column_id() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true);
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+zonemap SST is non-empty"
    );

    // Flip the first block's first (user-key) column id, leaving every other
    // recorded field intact: the re-derived per-column map still expects the
    // user-key column at id 0, so the mismatch is caught.
    crate::test_forge::forge_zone_map_column_id(&source, 0)?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::ZoneMap, None);
    assert!(
        matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("per-column statistics")),
        "the rejection must name the columnar per-column mismatch, got {err:?}",
    );
    Ok(())
}

/// A columnar SST of one 64 KiB row group cut into 1 KiB row pages, so the
/// group carries statistics zones in its directory and its zone block.
#[cfg(feature = "columnar")]
fn zoned_source(source: &std::path::Path, fs: &Arc<dyn Fs>) -> crate::Result<()> {
    let mut writer = Writer::new(source.to_path_buf(), 0, 0, Arc::clone(fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(64 * 1_024)
        .use_columnar_page_size(1_024);
    for i in 0u32..400 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");
    Ok(())
}

/// Narrows the zone of row page 1 by raising the last byte of its lower bound,
/// so the zones stay well formed and only disagree with the rows they
/// describe: the bound's first row now falls below it. The byte is one no
/// neighbouring bound shares, so the zones keep their wire length, which the
/// forged block's framing needs.
#[cfg(feature = "columnar")]
fn narrowed(
    zones: &crate::table::column_page::PageZones,
    columns: &[u16],
    row_pages: usize,
) -> crate::table::column_page::PageZones {
    let mut out = crate::table::column_page::PageZones::new(columns.to_vec());
    for (row_page, ordinal) in (0..row_pages).zip(0u16..) {
        for &column_id in columns {
            let Some(zone) = zones.zone(ordinal, column_id) else {
                panic!("a zone per row page");
            };
            let Some(max) = zone.max else {
                panic!("the fixture's zones are bounded");
            };
            let mut min = zone.min.to_vec();
            if row_page == 1 {
                let Some(last) = min.last_mut() else {
                    panic!("the fixture's bounds are not empty");
                };
                *last += 1;
                assert!(min.as_slice() <= max, "the forgery keeps the zone ordered");
            }
            out.push(zone.null_count, Some((&min, max)));
        }
    }
    let (mut before, mut after) = (Vec::new(), Vec::new());
    zones.encode_into(&mut before);
    out.encode_into(&mut after);
    assert_eq!(after.len(), before.len(), "the forgery keeps the length");
    out
}

/// A row page's statistics zones decide whether a read skips it, and nothing
/// on the read path looks at the rows it skipped, so the gates authenticate
/// them by re-deriving them from the decoded group. A zone narrowed past the
/// rows it describes, checksum-consistent, in the directory or in the zone
/// block, is refused.
#[cfg(feature = "columnar")]
#[test]
fn verify_rejects_row_page_zones_that_disagree_with_their_rows() -> crate::Result<()> {
    use crate::table::columnar::{COL_USER_KEY, COL_VALUE};

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // The key column's zones, in the head zone block before the pages, and
    // the value column's, in a zone block after them.
    for (name, column_id) in [("key_zones", COL_USER_KEY), ("value_zones", COL_VALUE)] {
        let source = dir.path().join(name);
        zoned_source(&source, &fs)?;
        let honest = open(source.clone(), &fs)?;
        reconcile_clean(&honest, None);
        let (group_at, directory) = row_group(&source, &fs, 0)?;
        let row_pages = directory.row_pages().len();
        assert!(
            row_pages > 2,
            "the fixture must cut its group into row pages"
        );
        let mut bytes = std::fs::read(&source)?;
        let directory_len = {
            use crate::coding::Decode;
            let Some(frame) = bytes.get(group_at..) else {
                panic!("group within the file");
            };
            crate::table::block::Header::decode_from(&mut &frame[..])?.on_disk_size_with(None)
        };
        let Some((after_directory, _)) = directory.zone_block(column_id) else {
            panic!("column {column_id} has a zone block");
        };
        let zones_at = group_at + directory_len as usize + after_directory as usize;
        let zones = directory.decode_zone_block(column_id, &block_payload(&bytes, zones_at)?)?;
        let mut payload = Vec::new();
        crate::table::column_page::PageDirectory::encode_zone_block(
            directory.group_tag(),
            &narrowed(&zones, &[column_id], row_pages),
            &mut payload,
        );
        restamp_block(&mut bytes, zones_at, &payload)?;
        std::fs::write(&source, &bytes)?;
        let err = reconcile_error(
            &open(source, &fs)?,
            crate::table::ReconcileGate::BlockEntryCounts,
            None,
        );
        assert!(
            matches!(err, crate::Error::InvalidHeader(msg) if msg.contains("row page statistics")),
            "{name}: a forged zone block must be refused, got {err:?}",
        );
    }
    Ok(())
}

/// The compaction scan decodes every zone block of a group it streams against
/// the group's directory, as an indexed read does, so a checksum-consistent
/// zone block naming another group is refused rather than rewritten with fresh
/// statistics: the key column's in the head block and the value column's
/// after the pages.
#[cfg(feature = "columnar")]
#[test]
fn a_scan_refuses_a_zone_block_that_does_not_decode_against_its_directory() -> crate::Result<()> {
    use crate::table::columnar::{COL_USER_KEY, COL_VALUE};

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    for (name, column_id) in [("key_zones", COL_USER_KEY), ("value_zones", COL_VALUE)] {
        let source = dir.path().join(name);
        zoned_source(&source, &fs)?;
        let honest = open(source.clone(), &fs)?.scan()?.count();
        assert!(honest > 0, "{name}: the honest table scans");
        let (group_at, directory) = row_group(&source, &fs, 0)?;
        let mut bytes = std::fs::read(&source)?;
        let directory_len = {
            use crate::coding::Decode;
            let Some(frame) = bytes.get(group_at..) else {
                panic!("group within the file");
            };
            crate::table::block::Header::decode_from(&mut &frame[..])?.on_disk_size_with(None)
        };
        let Some((after_directory, _)) = directory.zone_block(column_id) else {
            panic!("column {column_id} has a zone block");
        };
        let zones_at = group_at + directory_len as usize + after_directory as usize;
        let zones = directory.decode_zone_block(column_id, &block_payload(&bytes, zones_at)?)?;
        let mut payload = Vec::new();
        crate::table::column_page::PageDirectory::encode_zone_block(
            directory.group_tag() + 1,
            &zones,
            &mut payload,
        );
        restamp_block(&mut bytes, zones_at, &payload)?;
        std::fs::write(&source, &bytes)?;

        // The scanner reads its first group as it is made, so the refusal
        // comes from making it or from the iteration.
        let Err(err) = open(source, &fs)?
            .scan()
            .and_then(Iterator::collect::<crate::Result<Vec<_>>>)
        else {
            panic!("{name}: a zone block of another group must be refused");
        };
        assert!(
            matches!(err, crate::Error::InvalidHeader(_)),
            "{name}: expected a framing refusal, got {err:?}",
        );
    }
    Ok(())
}

/// A salvaged COLUMNAR table must keep its per-column zone-map statistics. The
/// clean-block verbatim copy-through re-emits columnar blocks byte-for-byte via
/// `append_verbatim_data_block`; if that path recorded the row-block synthetic
/// column-0 statistic instead of the per-column stats, the salvaged table would
/// fail its own `verify_zone_map` forgery cross-check (the verifier re-derives
/// per-column stats from the decoded columnar block).
#[cfg(feature = "columnar")]
#[test]
fn salvaged_columnar_table_keeps_per_column_zone_statistics() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A clean columnar SST with a zone map and several small data blocks, so at
    // least one clean block is byte-copied verbatim during salvage.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(128);
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.salvaged_path.is_some(),
        "the clean SST salvages: {report:?}",
    );
    assert!(
        report.blocks_copied_verbatim > 0,
        "at least one clean columnar block is byte-copied verbatim: {report:?}",
    );

    let table = open(dest, &fs)?;
    assert!(table.has_zone_map(), "the salvaged copy carries a zone map");
    // The per-column stats the copy-through recorded must equal what the
    // verifier re-derives from each decoded columnar block.
    reconcile_clean(&table, None);
    Ok(())
}

/// A salvaged copy keeps the source's groups as they were cut and encoded,
/// so it records the source's shape rather than the writer's defaults: a
/// later merge would otherwise copy those groups into a table that cuts them
/// another way.
#[cfg(feature = "columnar")]
#[test]
fn salvage_records_the_source_group_shape() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_column_encoding(crate::config::ColumnEncoding::Auto)
        .use_row_group_size(4_096)
        .use_columnar_page_size(256);
    for i in 0u32..64 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );
    let shape = open(source.clone(), &fs)?.metadata.group_shape;
    assert_eq!(
        shape,
        Some(crate::table::meta::GroupShape {
            encoding: crate::config::ColumnEncoding::Auto,
            row_group_size: 4_096,
            page_size: 256,
        }),
    );

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.blocks_copied_verbatim > 0,
        "the source's groups are copied as they are: {report:?}",
    );
    assert_eq!(open(dest, &fs)?.metadata.group_shape, shape);
    Ok(())
}

/// A clean group of row pages is copied verbatim with its zone blocks: the
/// copy is the group whole, directory, pages and zones, so the salvaged
/// table's groups still fill their extents, pass every gate, and prune and
/// read as the source did.
#[cfg(feature = "columnar")]
#[test]
fn salvage_copies_a_group_of_row_pages_with_its_zone_blocks() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    zoned_source(&source, &fs)?;
    let (_, directory) = row_group(&source, &fs, 0)?;
    assert!(
        !directory.zone_blocks().is_empty(),
        "the fixture group has zone blocks"
    );

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.blocks_copied_verbatim > 0 && report.dropped.is_empty(),
        "the clean group is copied verbatim: {report:?}",
    );
    let (_, copied) = row_group(&dest, &fs, 0)?;
    assert_eq!(copied, directory, "the copy keeps the group's directory");

    let table = open(dest, &fs)?;
    reconcile_clean(&table, None);
    for i in 0u32..400 {
        let key = iv(i).key.user_key;
        let Some(got) = table.get(&key, crate::SeqNo::MAX, crate::hash::hash64(&key))? else {
            panic!("key {i} survives the salvage");
        };
        assert_eq!(got.value, iv(i).value, "key {i}");
    }
    Ok(())
}

/// `verify_block_layout` must apply the present-empty rejection to ENCRYPTED
/// tables too. The emptiness check used to sit AFTER the `self.encryption`
/// early return, so an encrypted table's empty `block_layout` (a `delete_bitmap`
/// relabeled to `block_layout`) slipped through. Decoding the section with the
/// provider both runs the check and, for a genuine relabel, fails the AEAD open
/// (the block-type AAD binds the real role).
#[cfg(all(feature = "encryption", feature = "zstd"))]
#[test]
fn verify_block_layout_rejects_an_empty_map_on_an_encrypted_table() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let enc: Arc<dyn crate::encryption::EncryptionProvider> =
        Arc::new(crate::encryption::Aes256GcmProvider::new(&[0x42; 32]));

    // 256 KiB blocks at zstd L19 split into many inner blocks, so the writer
    // records a real block_layout section even under encryption.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256 * 1024)
        .use_data_block_compression(crate::CompressionType::Zstd(19))
        .use_encryption(Some(Arc::clone(&enc)));
    for i in 0u64..20_000 {
        writer.write(InternalValue::from_components(
            format!("key-{i:012}").into_bytes(),
            format!("value-{i:08}-payload").into_bytes(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source encrypted SST is non-empty",
    );

    // Re-stamp the block_layout section as a valid but EMPTY encrypted map.
    crate::test_forge::forge_block_layout_empty(&source, 0, Some(&*enc))?;

    let table = open_encrypted(source, &fs, Arc::clone(&enc))?;
    let Err(err) = table.verify_block_layout() else {
        panic!("an empty block_layout on an encrypted table with data blocks must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "block_layout section is present but empty on a table with data blocks"
            )
        ),
        "the rejection names the present-empty block_layout reason, got {err:?}",
    );
    Ok(())
}

/// `verify_metadata_bounds` must cross-check `seqno#min` against the decoded
/// entries even on a RANGE-TOMBSTONE-bearing table. The synthetic sentinel is
/// only written when the table has NO real KV items, so a table with both point
/// entries AND range tombstones has no sentinel and its decoded seqnos are all
/// real. Skipping the check for every RT-bearing table let both meta mirrors be
/// re-stamped with `seqno#min` raised above a real KV seqno, hiding live
/// versions from snapshots at/below the forged minimum after reopen.
#[test]
fn verify_metadata_bounds_rejects_a_raised_seqno_min_on_a_range_tombstone_table()
-> crate::Result<()> {
    use crate::UserKey;
    use crate::range_tombstone::RangeTombstone;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Point entries at seqnos 1..=8 PLUS a range tombstone: the table carries a
    // range_tombstones section but, having real KV items, writes no sentinel.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..8 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    writer.write_range_tombstone(RangeTombstone::new(
        UserKey::from(b"key-002".as_slice()),
        UserKey::from(b"key-005".as_slice()),
        9,
    ));
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Raise seqno#min above the real minimum KV seqno (1) in BOTH meta mirrors.
    crate::test_forge::forge_meta_value_both_mirrors(&source, b"seqno#min", &5u64.to_le_bytes())?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader("meta seqno#min is above the decoded minimum seqno")
        ),
        "the rejection names the seqno#min branch specifically, got {err:?}",
    );
    Ok(())
}

/// The synthetic-sentinel exclusion must NOT fire on a table with REAL KV
/// entries. Only an RT-ONLY table carries a synthetic sentinel, whose RT-derived
/// seqno the writer keeps ABOVE `highest_kv_seqno`. On an RT+KV table a real weak
/// tombstone whose key and seqno happen to match the RT-minimal `(start, seqno)`
/// contributed to `highest_kv_seqno` (so its seqno is `<=` it) and must be counted
/// toward the seqno bounds — excluding it as if it were the sentinel drops the
/// true minimum, letting a `seqno#min` restamped above that entry pass and hide
/// the live version from snapshots at/below the forged minimum.
#[test]
fn verify_metadata_bounds_keeps_a_real_weak_tombstone_matching_the_rt_sentinel() -> crate::Result<()>
{
    use crate::UserKey;
    use crate::range_tombstone::RangeTombstone;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // key-002 is a REAL weak tombstone at seqno 3 — the TRUE minimum seqno — and
    // its (key, seqno) equals the (seqno, start)-minimal range tombstone's, so
    // the old code mistook it for the synthetic RT-only sentinel and excluded it.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    writer.write(InternalValue::from_components(
        b"key-000".to_vec(),
        b"val-000".to_vec(),
        4,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"key-001".to_vec(),
        b"val-001".to_vec(),
        5,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::new_weak_tombstone(
        UserKey::from(b"key-002".as_slice()),
        3,
    ))?;
    for i in 3u64..8 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 3,
            ValueType::Value,
        ))?;
    }
    // The (seqno, start)-minimal range tombstone is (start = key-002, seqno = 3),
    // so the derived sentinel is exactly the real weak tombstone above. Its end
    // stays within the table key range (max key-007) so coverage passes.
    writer.write_range_tombstone(RangeTombstone::new(
        UserKey::from(b"key-002".as_slice()),
        UserKey::from(b"key-005".as_slice()),
        3,
    ));
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Raise seqno#min to 4 (above the real minimum 3, at the next-lowest seqno):
    // caught only if the seqno-3 weak tombstone is counted. With the sentinel
    // wrongly excluding it, the decoded minimum becomes 4 and 4 > 4 is false, so
    // the forge slips through.
    crate::test_forge::forge_meta_value_both_mirrors(&source, b"seqno#min", &4u64.to_le_bytes())?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader("meta seqno#min is above the decoded minimum seqno")
        ),
        "the rejection names the seqno#min branch specifically, got {err:?}",
    );
    Ok(())
}

/// `verify_metadata_bounds` must reject a `range_tombstones` section whose
/// decoded count disagrees with the recorded `range_tombstone_count`: a dropped
/// section (or a re-stamped block decoding to a subset) passes coverage for its
/// surviving entries, but the missing ranges no longer mask lower-level data —
/// reads resurrect the keys those tombstones deleted. The recorded count lives
/// in the (separately authenticated) meta block, so a mismatch is a forgery.
#[test]
fn verify_metadata_bounds_rejects_a_range_tombstone_count_mismatch() -> crate::Result<()> {
    use crate::UserKey;
    use crate::range_tombstone::RangeTombstone;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..8 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    writer.write_range_tombstone(RangeTombstone::new(
        UserKey::from(b"key-002".as_slice()),
        UserKey::from(b"key-005".as_slice()),
        9,
    ));
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Drop the range_tombstones section: the meta block still records
    // range_tombstone_count = 1, so the decoded count (0) disagrees.
    crate::test_forge::forge_section_omitted(&source, b"range_tombstones")?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "range_tombstones count disagrees with the recorded range_tombstone_count"
            )
        ),
        "the rejection names the RT count-mismatch reason, got {err:?}",
    );
    Ok(())
}

/// `verify_tli_mirrors` must cross-check each PARTITIONED top-level separator
/// against its partition's last data-block separator. Lowering a top-level
/// separator (in both mirrors) keeps the mirrors equal, the handles tiling, and
/// every LEAF separator matching its block — yet `TwoLevelBlockIndex` seeks by
/// the top-level separator and would route reads to the wrong partition,
/// skipping the keys the real partition holds. Only comparing the top-level
/// separator with the partition's decoded last key catches it.
#[test]
fn verify_tli_mirrors_rejects_a_forged_partition_boundary() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A partitioned index with many small data blocks yields >= 2 top-level
    // partitions, so the forge lowers a real partition BOUNDARY (not a leaf).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_partitioned_index()
        .use_data_block_size(128)
        .use_meta_partition_size(2);
    for i in 0u64..512 {
        writer.write(InternalValue::from_components(
            format!("key-{i:05}").into_bytes(),
            format!("val-{i:05}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Lower the FIRST top-level separator in both TLI mirrors.
    crate::test_forge::forge_tli_mirrors_lower_first_separator(&source, 0, None)?;

    let table = open(source, &fs)?;
    let Err(err) = table.verify_tli_mirrors() else {
        panic!("a forged partition boundary must be rejected");
    };
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "tli separator disagrees with its partition's last separator"
            )
        ),
        "the rejection names the partition-boundary mismatch, got {err:?}",
    );
    Ok(())
}

/// A `linked_blob_files` section that PARSES but under-reports its contents
/// (the first record's blob id forged to one nothing references) must not be
/// trusted as the sole source of the recovered copy's links: the recovered entries
/// still hold `ValueHandle` indirections into the blob file, and blob GC /
/// relocation consults the links to decide liveness — an under-reported list
/// would let GC delete or rewrite a blob the copy still references. Salvage
/// derives the links from the recovered indirections and unions them with
/// the source list.
#[test]
fn salvage_rebuilds_blob_links_from_recovered_indirections() -> crate::Result<()> {
    use crate::AbstractTree;

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let crate::AnyTree::Blob(tree) = crate::Config::new(
        dir.path(),
        crate::SequenceNumberCounter::default(),
        crate::SequenceNumberCounter::default(),
    )
    .with_kv_separation(Some(crate::KvSeparationOptions::default()))
    .open()?
    else {
        unreachable!("kv separation configured");
    };
    let big = |i: u32| format!("{i:08}").repeat(512);
    for i in 0u32..10 {
        tree.insert(format!("key{i:05}"), big(i), u64::from(i) + 1);
    }
    tree.flush_active_memtable(10)?;
    let source = {
        let binding = tree.index.version_history.read().latest_version();
        let Some(table) = binding.version.iter_tables().next() else {
            panic!("flush produced one table");
        };
        (*table.path).clone()
    };
    drop(tree);

    // The TRUE links, before the forgery.
    let Some(true_links) = open(source.clone(), &fs)?.list_blob_file_references()? else {
        panic!("the source carries a linked_blob_files section");
    };
    assert!(!true_links.is_empty(), "the source references blob files");

    // Forge the first record's blob id, right after the four-byte count: the
    // section still parses but no longer lists that blob file.
    let pos = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader
            .toc()
            .iter()
            .find(|e| e.name() == b"linked_blob_files")
        else {
            panic!("the source must carry a linked_blob_files section");
        };
        usize::try_from(entry.pos()).unwrap_or(usize::MAX)
    };
    const FORGED_ID: u64 = u64::MAX - 7;
    let Some(hidden) = true_links.first().map(|link| link.blob_file_id) else {
        panic!("the source references a first blob file");
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(id) = bytes.get_mut(pos + 4..pos + 12) else {
        panic!("linked_blob_files first record within the file");
    };
    id.copy_from_slice(&FORGED_ID.to_le_bytes());
    std::fs::write(&source, &bytes)?;

    // Sanity: the forgery took — the source now under-reports.
    let Some(forged) = open(source.clone(), &fs)?.list_blob_file_references()? else {
        panic!("the forged section still parses");
    };
    assert!(
        forged.iter().all(|link| link.blob_file_id != hidden),
        "the forged id hides the first link"
    );

    let dest = dir.path().join("salvaged");
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(report.is_complete(), "data blocks are healthy: {report:?}");

    // The copy's links are derived from its recovered indirections, not
    // parroted from the forged source list.
    let Some(recovered_links) = open(dest, &fs)?.list_blob_file_references()? else {
        panic!("the salvaged copy must carry links derived from its indirections");
    };
    for link in &true_links {
        assert!(
            recovered_links
                .iter()
                .any(|l| l.blob_file_id == link.blob_file_id),
            "blob file {} is referenced by recovered indirections but missing \
             from the copy's links: {recovered_links:?}",
            link.blob_file_id,
        );
    }
    Ok(())
}

/// A parseable `linked_blob_files` section that OVER-reports — carries a blob
/// id no recovered `ValueHandle` actually references — must not have that
/// source-only id copied into the salvaged table: the id may not exist under
/// `blobs/` at all (a forged record), and a manifest whose table links a blob
/// absent from the blob-file list is a corrupt reference downstream consumers
/// must never see. The copy's links are the recovered indirections, exactly.
#[test]
fn salvage_drops_source_only_blob_links() -> crate::Result<()> {
    use crate::AbstractTree;

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let crate::AnyTree::Blob(tree) = crate::Config::new(
        dir.path(),
        crate::SequenceNumberCounter::default(),
        crate::SequenceNumberCounter::default(),
    )
    .with_kv_separation(Some(crate::KvSeparationOptions::default()))
    .open()?
    else {
        unreachable!("kv separation configured");
    };
    let big = |i: u32| format!("{i:08}").repeat(512);
    for i in 0u32..10 {
        tree.insert(format!("key{i:05}"), big(i), u64::from(i) + 1);
    }
    tree.flush_active_memtable(10)?;
    let source = {
        let binding = tree.index.version_history.read().latest_version();
        let Some(table) = binding.version.iter_tables().next() else {
            panic!("flush produced one table");
        };
        (*table.path).clone()
    };
    drop(tree);

    // Forge the FIRST link record's blob id (the leading u64 of the 32-byte
    // record, right after the LE u32 count) to an id that exists nowhere.
    const FORGED_ID: u64 = 9_999_999;
    let pos = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader
            .toc()
            .iter()
            .find(|e| e.name() == b"linked_blob_files")
        else {
            panic!("the source must carry a linked_blob_files section");
        };
        usize::try_from(entry.pos()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(id_slot) = bytes.get_mut(pos + 4..pos + 12) else {
        panic!("first link record within the file");
    };
    id_slot.copy_from_slice(&FORGED_ID.to_le_bytes());
    std::fs::write(&source, &bytes)?;

    // Sanity: the forgery took — the source reports the forged id.
    let Some(forged) = open(source.clone(), &fs)?.list_blob_file_references()? else {
        panic!("the forged section still parses");
    };
    assert!(
        forged.iter().any(|l| l.blob_file_id == FORGED_ID),
        "the source's section carries the forged id",
    );

    let dest = dir.path().join("salvaged");
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(report.is_complete(), "data blocks are healthy: {report:?}");

    let Some(recovered_links) = open(dest, &fs)?.list_blob_file_references()? else {
        panic!("the salvaged copy carries links derived from its indirections");
    };
    assert!(
        !recovered_links.is_empty(),
        "the recovered indirections reference the real blob file",
    );
    assert!(
        !recovered_links.iter().any(|l| l.blob_file_id == FORGED_ID),
        "a source-only id no recovered indirection references must not be \
         copied into the salvaged table: {recovered_links:?}",
    );
    Ok(())
}

/// A KV-separated source whose `linked_blob_files` section is UNREADABLE (a
/// count header claiming more records than the section holds) must not abort
/// the salvage: the section is not authoritative — the walk derives the
/// copy's links from the recovered `ValueHandle` indirections — so the
/// readable data blocks are still recovered and the copy carries the derived
/// links. Failing here would leave the whole table unrecovered over a
/// non-authoritative side section.
#[test]
fn salvage_recovers_with_derived_links_when_the_blob_link_section_is_unreadable()
-> crate::Result<()> {
    use crate::AbstractTree;

    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let crate::AnyTree::Blob(tree) = crate::Config::new(
        dir.path(),
        crate::SequenceNumberCounter::default(),
        crate::SequenceNumberCounter::default(),
    )
    .with_kv_separation(Some(crate::KvSeparationOptions::default()))
    .open()?
    else {
        unreachable!("kv separation configured");
    };
    let big = |i: u32| format!("{i:08}").repeat(512);
    for i in 0u32..10 {
        tree.insert(format!("key{i:05}"), big(i), u64::from(i) + 1);
    }
    tree.flush_active_memtable(10)?;
    let source = {
        let binding = tree.index.version_history.read().latest_version();
        let Some(table) = binding.version.iter_tables().next() else {
            panic!("flush produced one table");
        };
        (*table.path).clone()
    };
    drop(tree);

    // Overwrite the linked_blob_files count header (leading LE u32) with a
    // value far larger than the section can hold, so parsing the records
    // hits EOF and list_blob_file_references errors.
    let pos = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader
            .toc()
            .iter()
            .find(|e| e.name() == b"linked_blob_files")
        else {
            panic!("the source must carry a linked_blob_files section");
        };
        usize::try_from(entry.pos()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(count) = bytes.get_mut(pos..pos + 4) else {
        panic!("linked_blob_files count header within the file");
    };
    count.copy_from_slice(&u32::MAX.to_le_bytes());
    std::fs::write(&source, &bytes)?;

    let dest = dir.path().join("salvaged");
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(report.is_complete(), "data blocks are healthy: {report:?}");

    // The copy's links are derived from its recovered indirections; the
    // unreadable source section contributes nothing.
    let Some(recovered_links) = open(dest, &fs)?.list_blob_file_references()? else {
        panic!("the salvaged copy must carry links derived from its indirections");
    };
    assert!(
        !recovered_links.is_empty(),
        "the recovered indirections reference at least one blob file",
    );
    Ok(())
}

/// Standalone salvage preserves the SOURCE's persisted table id: an
/// unencrypted SST written under a non-zero id salvages WITHOUT the caller
/// supplying that id (the salvage-mode open reads it from the metadata
/// instead of failing the id cross-check against the options default of 0),
/// and the recovered copy is stamped with the source's id — so it keeps its
/// identity when an operator swaps it in for the original.
#[test]
fn salvage_preserves_a_nonzero_source_table_id() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    const TID: crate::TableId = 7;
    let mut writer = Writer::new(source.clone(), TID, 0, Arc::clone(&fs))?.use_data_block_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Default options (table_id = 0): the salvage must still open the source
    // and carry its real id through.
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "a healthy non-zero-id SST salvages completely: {report:?}",
    );

    // The recovered copy reopens under the SOURCE's id (the recover
    // cross-checks the stored table_id against the expected one).
    let recovered = open_with_id(dest, &fs, TID)?;
    assert_eq!(
        recovered.metadata.id, TID,
        "the salvaged copy is stamped with the source's table id",
    );
    assert_eq!(recovered.metadata.item_count, u64::from(n));
    Ok(())
}

#[test]
fn salvage_of_a_healthy_sst_recovers_every_block() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Build a multi-block source SST: small data blocks force several blocks so
    // the per-block walk has more than one block to recover.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    assert!(
        report.is_complete(),
        "a healthy SST salvages with no dropped blocks: {report:?}",
    );
    assert!(
        report.blocks_total >= 2,
        "256-byte blocks over 200 entries should yield several data blocks, got {}",
        report.blocks_total,
    );
    assert_eq!(
        report.blocks_salvaged, report.blocks_total,
        "every block of a healthy SST is salvaged",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every entry is recovered",
    );
    assert_eq!(
        report.salvaged_path.as_deref(),
        Some(dest.as_path()),
        "a salvaged file is written when at least one block is recovered",
    );

    // Every block of a healthy SST reads back clean, so every salvaged block is
    // copied through verbatim — none re-encoded.
    assert_eq!(
        report.blocks_copied_verbatim, report.blocks_salvaged,
        "a healthy SST's blocks are all copied verbatim",
    );

    // The salvaged copy is a valid SST that reopens and holds every key.
    assert_eq!(
        reopen_item_count(dest, &fs)?,
        u64::from(n),
        "the salvaged SST reopens with the full item count",
    );
    Ok(())
}

/// A clean block is byte-copied verbatim, not decoded and re-encoded: its raw
/// on-disk bytes in the salvaged SST are identical to the source's, and the walk
/// reports it under `blocks_copied_verbatim`.
#[test]
fn salvage_copies_a_clean_block_verbatim() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "healthy SST salvages clean: {report:?}"
    );
    assert_eq!(
        report.blocks_copied_verbatim, report.blocks_total,
        "every clean block is copied verbatim, none re-encoded",
    );

    // The first data block's raw on-disk bytes must be byte-identical between the
    // source and the salvaged copy (each resolved through its own intact index).
    let first_block = |path: &std::path::Path| -> crate::Result<(usize, usize)> {
        let table = open(path.to_path_buf(), &fs)?;
        let Some(kh) = table.data_block_handles().find_map(Result::ok) else {
            panic!("a non-empty SST has at least one data block");
        };
        let off = usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX);
        Ok((off, kh.as_ref().size() as usize))
    };
    let (src_off, src_size) = first_block(&source)?;
    let (dst_off, dst_size) = first_block(&dest)?;
    assert_eq!(
        src_size, dst_size,
        "the verbatim copy preserves the block's on-disk size",
    );

    let src_bytes = std::fs::read(&source)?;
    let dst_bytes = std::fs::read(&dest)?;
    let src_block = src_bytes.get(src_off..src_off + src_size);
    let dst_block = dst_bytes.get(dst_off..dst_off + dst_size);
    assert!(
        src_block.is_some() && src_block == dst_block,
        "the clean block is copied byte-for-byte into the salvaged SST",
    );
    Ok(())
}

/// One deliberately corrupted data block: salvage drops exactly that block
/// (naming its key range) and recovers every other block, instead of failing
/// the whole file. This is the core block-granular contract.
#[test]
fn salvage_drops_a_corrupted_block_and_keeps_the_rest() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Resolve the second data block's on-disk offset from the (intact) index,
    // then flip a byte inside its PAYLOAD (past the fixed-size block header) so
    // the header still frames but the block's data checksum fails on load.
    // load_data_block reads the block by the index handle's size, so the
    // corruption surfaces as that one block failing at load — kept via the
    // index, dropped as a checksum mismatch — not as a physical-walk desync.
    let target = {
        let table = open(source.clone(), &fs)?;
        let offsets: alloc::vec::Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        let Some(&second) = offsets.get(1) else {
            panic!("source SST must have at least two data blocks, got {offsets:?}");
        };
        second
    };
    // Land well past the fixed data-block header (magic + type + checksums +
    // lengths) so the frame stays intact and only the payload rots.
    let header_len = crate::table::block::Header::header_len(crate::table::block::BlockType::Data);
    let Ok(target_usize) = usize::try_from(target) else {
        panic!("data block offset {target} does not fit usize on this target");
    };
    let flip = target_usize + header_len + 8;
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    assert!(
        !report.is_complete(),
        "a corrupted block must be reported as dropped: {report:?}",
    );
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the one corrupted block is dropped: {report:?}",
    );
    assert_eq!(
        report.blocks_salvaged,
        report.blocks_total - 1,
        "every block but the corrupted one is recovered",
    );
    assert!(
        report.entries_salvaged > 0 && report.entries_salvaged < u64::from(n),
        "a partial key range is recovered, got {} of {n}",
        report.entries_salvaged,
    );
    assert!(
        report.dropped.first().is_some_and(|d| {
            matches!(d.reason, DropReason::ChecksumMismatch) && d.key_range.is_some()
        }),
        "the dropped block reports a checksum mismatch and names the key range it lost: {report:?}",
    );
    assert_eq!(report.salvaged_path.as_deref(), Some(dest.as_path()));

    // The salvaged copy reopens and holds exactly the recovered entries.
    assert_eq!(
        reopen_item_count(dest, &fs)?,
        report.entries_salvaged,
        "the salvaged SST holds exactly the entries the report counted",
    );
    Ok(())
}

/// A data block whose HEADER no longer frames is dropped by the physical gap
/// walk, not the load path: the index still points at it, but `probe_gap`
/// records the region rather than placing it in `items`. It must still count
/// toward `blocks_total` (the walk INSPECTED it and found it unframeable), so
/// the `blocks_total == recovered + dropped` contract holds. Previously a
/// header-corrupt block dropped without being counted, so `blocks_total`
/// equalled `blocks_salvaged` while a block had been lost: recovery ratios
/// reported full coverage despite the loss.
#[test]
fn salvage_counts_a_header_corrupt_block_in_blocks_total() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Flip a byte INSIDE the second data block's header (its magic), so the
    // physical frame no longer parses: the tiling walk cannot trust the index
    // span and resyncs PAST the block, dropping it through the gap probe rather
    // than the load path: a `dropped` entry that never enters `items`.
    let target = {
        let table = open(source.clone(), &fs)?;
        let offsets: alloc::vec::Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        let Some(&second) = offsets.get(1) else {
            panic!("source SST must have at least two data blocks, got {offsets:?}");
        };
        second
    };
    let Ok(target_usize) = usize::try_from(target) else {
        panic!("data block offset {target} does not fit usize on this target");
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(target_usize) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest, &fs)?;

    assert!(
        !report.dropped.is_empty(),
        "the header-corrupt block must be reported dropped: {report:?}",
    );
    assert_eq!(
        report.blocks_total,
        report.blocks_salvaged + report.dropped.len(),
        "every inspected block is either recovered or dropped, a header-corrupt \
         block must count toward the total, not vanish from it: {report:?}",
    );
    Ok(())
}

/// A data block that needs ECC recovery to read is NOT copied verbatim — its
/// on-disk bytes are faulty, so propagating them would carry the corruption into
/// the recovered copy. Salvage re-encodes the healed payload instead (clean bytes
/// in the copy), while the surrounding clean blocks are still copied verbatim.
#[cfg(feature = "page_ecc")]
#[test]
fn salvage_reencodes_an_ecc_recovered_block_rather_than_copying_it() -> crate::Result<()> {
    use crate::table::block::{EccParams, Header};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Data blocks carry RS(4,2) parity, so a small corruption is healed on read
    // rather than failing the block.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_ecc(Some(EccParams::RS_4_2));
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Flip one payload byte of the FIRST data block so reading it must repair via
    // RS parity (an ECC-recovered read, not a clean one).
    let first_off = {
        let table = open(source.clone(), &fs)?;
        let Some(kh) = table.data_block_handles().find_map(Result::ok) else {
            panic!("a non-empty SST has at least one data block");
        };
        *kh.as_ref().offset()
    };
    let pos = usize::try_from(first_off).unwrap_or(usize::MAX) + Header::MIN_LEN + 3;
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(pos) {
        *b ^= 0x80;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    // The healed block is recovered, not dropped — nothing is lost.
    assert!(
        report.is_complete(),
        "an ECC-recoverable block is healed, not dropped: {report:?}",
    );
    assert_eq!(
        report.blocks_salvaged, report.blocks_total,
        "every block is recovered",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every entry is recovered",
    );
    // Exactly the healed block was re-encoded; the rest were copied verbatim.
    assert_eq!(
        report.blocks_copied_verbatim,
        report.blocks_salvaged - 1,
        "the ECC-recovered block is re-encoded, not copied verbatim",
    );

    // The salvaged copy reopens with every key; its bytes are freshly encoded, so
    // they no longer need ECC repair.
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n));
    Ok(())
}

/// Bit rot confined to a block's PARITY trailer reads as clean (the payload
/// checksum passes and parity is only consulted on a mismatch), so a verbatim
/// copy would carry the rotted parity into the salvaged SST as latent ECC
/// corruption. Salvage must verify the trailer before copying and re-encode
/// (regenerating fresh parity) when it disagrees: every data block of the
/// recovered copy must carry parity that matches its payload.
#[cfg(feature = "page_ecc")]
#[test]
fn salvage_regenerates_a_rotted_parity_trailer_rather_than_copying_it() -> crate::Result<()> {
    use crate::coding::Decode;
    use crate::table::block::{EccParams, Header, expected_parity_len};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_ecc(Some(EccParams::RS_4_2));
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Flip one byte INSIDE the first data block's parity trailer (right after
    // its `data_length` payload). The payload checksum still verifies, so the
    // block reads back clean with no ECC recovery.
    let first_off = {
        let table = open(source.clone(), &fs)?;
        let Some(kh) = table.data_block_handles().find_map(Result::ok) else {
            panic!("a non-empty SST has at least one data block");
        };
        usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(mut cursor) = bytes.get(first_off..) else {
        panic!("first data block within the file");
    };
    let header = Header::decode_from(&mut cursor)?;
    let trailer_pos =
        first_off + Header::header_len(header.block_type) + header.data_length as usize;
    let Some(slot) = bytes.get_mut(trailer_pos) else {
        panic!("parity trailer within the file");
    };
    *slot ^= 0xFF;
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "the payload is intact, every block is recovered: {report:?}",
    );
    assert_eq!(reopen_item_count(dest.clone(), &fs)?, u64::from(n));

    // Every data block of the salvaged copy carries parity that matches its
    // payload — the rotted trailer was regenerated, not byte-copied.
    let dest_bytes = std::fs::read(&dest)?;
    let dest_table = open(dest, &fs)?;
    let (ds, ps) = EccParams::RS_4_2.as_shards();
    for kh in dest_table.data_block_handles() {
        let kh = kh?;
        let off = usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX);
        let Some(mut cursor) = dest_bytes.get(off..) else {
            panic!("block at offset {off} within the file");
        };
        let hdr = Header::decode_from(&mut cursor)?;
        let hlen = Header::header_len(hdr.block_type);
        let dl = hdr.data_length as usize;
        let Some(payload) = dest_bytes.get(off + hlen..off + hlen + dl) else {
            panic!("payload of block at offset {off} within the file");
        };
        let plen = expected_parity_len(hdr.data_length, EccParams::RS_4_2) as usize;
        let Some(trailer) = dest_bytes.get(off + hlen + dl..off + hlen + dl + plen) else {
            panic!("parity trailer of block at offset {off} within the file");
        };
        let fresh = crate::ecc::encode_parity(payload, ds, ps)?;
        assert_eq!(
            trailer,
            fresh.as_slice(),
            "block at offset {off}: the salvaged copy's parity matches its payload",
        );
    }
    Ok(())
}

/// Two row groups of an encrypted columnar table swap their value pages. Each
/// page still authenticates, because the AEAD binds its table and its role but
/// not the group it belongs to, and the two pages have the same length and row
/// count. Unlike a block swap, which only costs a lookup miss because a value
/// cannot leave its key's block, a page swap would serve one key's value under
/// another key; the read must refuse the group instead.
#[cfg(all(feature = "columnar", feature = "encryption"))]
#[test]
fn a_value_page_moved_between_row_groups_of_an_encrypted_table_is_refused() -> crate::Result<()> {
    use crate::coding::Decode;
    use crate::table::columnar::COL_VALUE;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let enc: Arc<dyn crate::encryption::EncryptionProvider> =
        Arc::new(crate::encryption::Aes256GcmProvider::new(&[0x42; 32]));
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_row_group_size(256)
        .use_encryption(Some(Arc::clone(&enc)));
    for i in 0..200_u32 {
        writer.write(InternalValue::from_components(
            format!("key{i:05}").into_bytes(),
            format!("val{i:05}").into_bytes(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source is non-empty");

    // Where the first two groups keep their value pages. The directory is
    // encrypted, so it is read through the table; the page extents are then
    // plain file ranges.
    let pages = {
        let table = open_encrypted(source.clone(), &fs, Arc::clone(&enc))?;
        let handles: Vec<_> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .take(2)
            .collect();
        assert_eq!(handles.len(), 2, "source has at least two row groups");
        let bytes = std::fs::read(&source)?;
        let mut pages = Vec::new();
        for kh in &handles {
            let handle = kh.as_ref();
            let directory = table.salvage_load_row_group(handle)?.group.directory;
            let at = usize::try_from(*handle.offset()).unwrap_or(usize::MAX);
            let Some(frame) = bytes.get(at..) else {
                panic!("row group within the file");
            };
            let header = crate::table::block::Header::decode_from(&mut &frame[..])?;
            let Some(value_page) = directory
                .entries()
                .iter()
                .find(|e| e.id.column_id == COL_VALUE)
            else {
                panic!("row group holds a value page");
            };
            pages.push((
                directory.row_count(),
                at + header.on_disk_size_with(None) as usize
                    + directory.pages_start() as usize
                    + value_page.offset as usize,
                value_page.length as usize,
            ));
        }
        pages
    };
    let [(rows_a, at_a, len_a), (rows_b, at_b, len_b)] = pages[..] else {
        panic!("two value pages located");
    };
    assert_eq!(
        (rows_a, len_a),
        (rows_b, len_b),
        "the swap needs pages of the same length over the same row count",
    );

    let mut bytes = std::fs::read(&source)?;
    let (Some(page_a), Some(page_b)) = (
        bytes.get(at_a..at_a + len_a).map(<[u8]>::to_vec),
        bytes.get(at_b..at_b + len_b).map(<[u8]>::to_vec),
    ) else {
        panic!("both value pages within the file");
    };
    // Each page is re-bound to where it moves, so its block checksum verifies
    // and the encrypted stamp is the check left to refuse it.
    for (at, from, page) in [(at_a, at_b, &page_b), (at_b, at_a, &page_a)] {
        let Some(target) = bytes.get_mut(at..at + page.len()) else {
            panic!("value page within the file");
        };
        target.copy_from_slice(page);
        crate::test_forge::rebind_moved_frame(target, (0, from as u64), (0, at as u64))?;
    }
    std::fs::write(&source, &bytes)?;

    let table = open_encrypted(source, &fs, enc)?;
    let key = b"key00000";
    match table.get(key, crate::MAX_SEQNO, crate::hash::hash64(key)) {
        Err(crate::Error::InvalidHeader(_)) => {}
        other => panic!("a page from another row group must be refused, got {other:?}"),
    }
    Ok(())
}

/// A columnar SST of table `table_id`, encrypted under `enc` when given,
/// holding `key{i}` rows whose values are `prefix{i}`, all values of one
/// length, in one group.
#[cfg(feature = "columnar")]
fn columnar_source_as(
    path: &std::path::Path,
    fs: &Arc<dyn Fs>,
    enc: Option<&Arc<dyn crate::encryption::EncryptionProvider>>,
    table_id: crate::TableId,
    prefix: &str,
) -> crate::Result<()> {
    let mut writer = Writer::new(path.to_path_buf(), table_id, 0, Arc::clone(fs))?
        .use_columnar(true)
        .use_encryption(enc.map(Arc::clone));
    for i in 0..20_u32 {
        writer.write(InternalValue::from_components(
            format!("key{i:05}").into_bytes(),
            format!("{prefix}{i:05}").into_bytes(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source is non-empty");
    Ok(())
}

/// Opens the SST at `path`, encrypted under `enc` when given, as table
/// `table_id`.
#[cfg(any(feature = "columnar", feature = "encryption"))]
fn open_as(
    path: std::path::PathBuf,
    fs: &Arc<dyn Fs>,
    enc: Option<&Arc<dyn crate::encryption::EncryptionProvider>>,
    table_id: crate::TableId,
) -> crate::Result<Table> {
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&**fs, &path)?);
    let mut params = crate::table::RecoverParams::new(
        path,
        checksum,
        table_id,
        Arc::clone(fs),
        default_comparator(),
        Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
    );
    params.encryption = enc.map(Arc::clone);
    Table::recover(params)
}

/// The file extent of the value page of the first group of `table`.
#[cfg(feature = "columnar")]
fn first_value_page(table: &Table, bytes: &[u8]) -> crate::Result<core::ops::Range<usize>> {
    use crate::coding::Decode;
    use crate::table::columnar::COL_VALUE;

    let Some(Ok(first)) = table.data_block_handles().next() else {
        panic!("the table has a row group");
    };
    let directory = table
        .salvage_load_row_group(first.as_ref())?
        .group
        .directory;
    let at = usize::try_from(*first.as_ref().offset()).unwrap_or(usize::MAX);
    let Some(frame) = bytes.get(at..) else {
        panic!("row group within the file");
    };
    let header = crate::table::block::Header::decode_from(&mut &frame[..])?;
    let Some(page) = directory
        .entries()
        .iter()
        .find(|e| e.id.column_id == COL_VALUE)
    else {
        panic!("row group holds a value page");
    };
    let start = at
        + header.on_disk_size_with(None) as usize
        + directory.pages_start() as usize
        + page.offset as usize;
    Ok(start..start + page.length as usize)
}

/// A value page moved into another table, in the same place of a group of the
/// same shape, is refused, encrypted or not and whether or not the two tables
/// share an id. A page's values are separated from their keys, so a page
/// taken for another table's would hand back that table's value under this
/// table's key. Every table starts its group tags from its own identity, so
/// the stamp names the table as well as the group, and under encryption the
/// AEAD binds the table id on top.
#[cfg(feature = "columnar")]
#[test]
fn a_value_page_moved_between_tables_is_refused() -> crate::Result<()> {
    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let mut providers: Vec<Option<Arc<dyn crate::encryption::EncryptionProvider>>> = vec![None];
    #[cfg(feature = "encryption")]
    providers.push(Some(Arc::new(crate::encryption::Aes256GcmProvider::new(
        &[0x42; 32],
    ))));

    for (n, enc) in providers.iter().enumerate() {
        let enc = enc.as_ref();
        // Moves `from`'s first value page into `into`'s; the two tables were
        // written with the same row count and value lengths, so the page fits.
        let transplant = |into: &std::path::Path, into_id, from: &std::path::Path, from_id| {
            let from_bytes = std::fs::read(from)?;
            let (page, from_at) = {
                let table = open_as(from.to_path_buf(), &fs, enc, from_id)?;
                let range = first_value_page(&table, &from_bytes)?;
                (
                    from_bytes.get(range.clone()).map(<[u8]>::to_vec),
                    range.start,
                )
            };
            let mut into_bytes = std::fs::read(into)?;
            let range = {
                let table = open_as(into.to_path_buf(), &fs, enc, into_id)?;
                first_value_page(&table, &into_bytes)?
            };
            let Some(page) = page.filter(|p| p.len() == range.len()) else {
                panic!("the two value pages are of one length");
            };
            let into_at = range.start;
            let Some(target) = into_bytes.get_mut(range) else {
                panic!("value page within the file");
            };
            target.copy_from_slice(&page);
            // Re-bound to its new table and place, so the page verifies as a
            // block and the stamp (and the AEAD) are the checks left.
            crate::test_forge::rebind_moved_frame(
                target,
                (from_id, from_at as u64),
                (into_id, into_at as u64),
            )?;
            std::fs::write(into, &into_bytes)?;
            crate::Result::Ok(())
        };
        let key = b"key00000";

        for (into_id, from_id) in [(3, 3), (3, 4)] {
            let a = dir.path().join(format!("{n}-{into_id}-{from_id}-into"));
            let b = dir.path().join(format!("{n}-{into_id}-{from_id}-from"));
            columnar_source_as(&a, &fs, enc, into_id, "vala")?;
            columnar_source_as(&b, &fs, enc, from_id, "valb")?;
            transplant(&a, into_id, &b, from_id)?;
            let table = open_as(a, &fs, enc, into_id)?;
            let refused = table.get(key, crate::MAX_SEQNO, crate::hash::hash64(key));
            assert!(
                refused.is_err(),
                "encrypted: {}, tables {into_id} and {from_id}: a page of another table \
                 must be refused, got {refused:?}",
                enc.is_some(),
            );
        }
    }
    Ok(())
}

/// A block of another role in a page's slot, its checksum intact, is refused
/// by role, as a data-block handle landing on a block of another role is:
/// relabelling a page as a zone block keeps its bytes valid as a block, and
/// only the role says it is not the page the directory files there.
#[cfg(feature = "columnar")]
#[test]
fn a_block_of_another_role_in_a_page_slot_is_refused() -> crate::Result<()> {
    use crate::coding::{Decode, Encode};
    use crate::table::block::{BlockType, Header};
    use crate::table::columnar::COL_VALUE;
    use crate::table::row_group::PageWant;
    use crate::table::util::ReadCharge;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    zoned_source(&source, &fs)?;
    let (group_at, directory) = row_group(&source, &fs, 0)?;
    let mut bytes = std::fs::read(&source)?;
    let directory_len = {
        let Some(frame) = bytes.get(group_at..) else {
            panic!("group within the file");
        };
        Header::decode_from(&mut &frame[..])?.on_disk_size_with(None)
    };
    let Some(page) = directory
        .entries()
        .iter()
        .find(|e| e.id.column_id == COL_VALUE)
    else {
        panic!("the group has a value page");
    };
    let at =
        group_at + directory_len as usize + directory.pages_start() as usize + page.offset as usize;
    let header = {
        let Some(frame) = bytes.get(at..) else {
            panic!("page within the file");
        };
        Header::decode_from(&mut &frame[..])?
    };
    assert_eq!(header.block_type, BlockType::ColumnPage);
    let mut relabelled = Vec::new();
    Header {
        block_type: BlockType::ColumnZones,
        ..header
    }
    .encode_into(&mut relabelled)?;
    let Some(target) = bytes.get_mut(at..at + relabelled.len()) else {
        panic!("page header within the file");
    };
    target.copy_from_slice(&relabelled);
    std::fs::write(&source, &bytes)?;

    let table = open(source, &fs)?;
    let Some(Ok(first)) = table.data_block_handles().next() else {
        panic!("the table has a row group");
    };
    let Err(err) = table.load_row_group(first.as_ref(), &PageWant::ALL, ReadCharge::Foreground)
    else {
        panic!("a zone block in a page slot must be refused");
    };
    assert!(
        matches!(err, crate::Error::InvalidTag(("BlockType", _))),
        "the refusal names the role, got {err:?}",
    );
    Ok(())
}

/// A columnar source with one corrupted PAX data block: the columnar loader
/// fails to reconstruct that block (a torn sub-column frame), so salvage drops
/// it and recovers every other block, writing the survivors as a plain row SST.
#[cfg(feature = "columnar")]
#[test]
fn salvage_drops_a_corrupted_columnar_block_and_keeps_the_rest() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A columnar SST (PAX blocks + zone map), no deletes so there is no
    // delete-bitmap section to worry about here.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    // Corrupt the second columnar data block's PAYLOAD (offset from the intact
    // index, past the fixed block header) so the frame stays intact but its
    // reconstruction fails on load — kept via the index, dropped at load.
    let target = {
        let table = open(source.clone(), &fs)?;
        let offsets: alloc::vec::Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        let Some(&second) = offsets.get(1) else {
            panic!("source columnar SST must have at least two data blocks, got {offsets:?}");
        };
        second
    };
    // Land well past the fixed data-block header so only the payload rots and
    // the header still frames (otherwise the physical walk resyncs past it).
    let header_len = crate::table::block::Header::header_len(crate::table::block::BlockType::Data);
    let Ok(target_usize) = usize::try_from(target) else {
        panic!("data block offset {target} does not fit usize on this target");
    };
    let flip = target_usize + header_len + 8;
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;

    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the one corrupted columnar block is dropped: {report:?}",
    );
    assert_eq!(
        report.blocks_salvaged,
        report.blocks_total - 1,
        "every columnar block but the corrupted one is recovered",
    );
    assert!(
        report.entries_salvaged > 0 && report.entries_salvaged < u64::from(n),
        "a partial key range is recovered, got {} of {n}",
        report.entries_salvaged,
    );
    assert_eq!(report.salvaged_path.as_deref(), Some(dest.as_path()));

    // The salvaged copy stays COLUMNAR (mirrored from the source) and holds the
    // recovered rows — no longer degraded to a row-major copy.
    let recovered = open(dest, &fs)?;
    assert_eq!(recovered.metadata.item_count, report.entries_salvaged);
    assert!(
        recovered.metadata.columnar,
        "a columnar source salvages into a columnar copy, not a row-major one",
    );
    Ok(())
}

/// A columnar block the salvage cannot write to its OUTPUT is a destination
/// failure, whatever the I/O error's kind: the salvage reports it as one,
/// so repair stops instead of grading the intact source as damage.
#[cfg(feature = "columnar")]
#[test]
fn a_columnar_block_the_output_cannot_take_fails_as_an_output_failure() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Without deletes, every block takes the unmasked path.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256);
    for i in 0..20_000u32 {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    // Each write of the output in turn fails, until the salvage gets
    // through. Bounded, so a fault that keeps firing fails the test instead
    // of running it forever.
    let mut failed = 0;
    let mut succeeded = false;
    for skip in 0u64..10_000 {
        let faulty = FaultFs::new(StdFs);
        faulty.injector().arm(
            FaultRule::new(FaultOp::Write, Fault::Error(crate::io::ErrorKind::Other))
                .on_path("salvaged")
                .skip(skip)
                .once(),
        );
        let faulty: Arc<dyn Fs> = Arc::new(faulty);
        match salvage_sst(&source, dest.clone(), &faulty) {
            Ok(_) => {
                succeeded = true;
                break;
            }
            Err(crate::Error::SalvageOutput(_)) => {
                failed += 1;
                assert!(
                    !dest.exists(),
                    "a failed attempt leaves no partial output behind (write {skip})"
                );
            }
            Err(e) => panic!("write {skip} of the output failed as {e:?}"),
        }
    }
    assert!(
        succeeded,
        "the salvage never got through the injected faults"
    );
    assert!(failed > 1, "the output takes more than one write");
    Ok(())
}

/// A columnar block whose outer `ColumnBatch` frame decodes but whose row
/// materialization fails (an invalid value-type byte in an otherwise
/// checksum-consistent block) is dropped like any other block-local decode
/// failure — one malformed block must not abort the whole salvage and discard
/// the destination while later blocks are still recoverable.
#[cfg(feature = "columnar")]
#[test]
fn salvage_drops_a_columnar_block_with_an_invalid_value_type() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    // Poison the SECOND row group: stamp an invalid value-type tag into the
    // first row of its value-type page, at the same length, and re-stamp the
    // page's checksum. The group stays checksum-consistent, so the failure
    // surfaces in row materialization — not as an ordinary checksum drop.
    let poisoned_rows = u64::from(forge_row_group_column(
        &source,
        &fs,
        1,
        crate::table::columnar::COL_VALUE_TYPE,
        |col, _| {
            // 0xFF is not a defined ValueType tag. Column bytes are an
            // immutable view — rebuild the poisoned column.
            let mut poisoned = col.data.to_vec();
            let Some(first_byte) = poisoned.first_mut() else {
                panic!("value-type column is non-empty");
            };
            *first_byte = 0xFF;
            col.data = poisoned.into();
        },
    )?);

    // Salvage drops exactly the poisoned block and recovers every other one.
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the poisoned block is dropped: {report:?}",
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(DropReason::DecodeError(_))
        ),
        "the invalid value-type tag classifies as a decode error: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n) - poisoned_rows,
        "every row outside the poisoned block is recovered",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n) - poisoned_rows);
    Ok(())
}

/// A checksum-consistent columnar block whose entries are OUT OF internal-key
/// order (two adjacent keys swapped) must be dropped, not emitted: verbatim
/// paths skip the ingest ordering checks, so an unvalidated malformed block
/// would register a wrong last-key in the recovered SST's index and corrupt
/// binary search / scan order. The rest of the SST still salvages.
#[cfg(feature = "columnar")]
#[test]
fn salvage_drops_a_columnar_block_with_out_of_order_keys() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    // Poison the SECOND row group: swap the first two rows' user keys inside
    // its key page and re-stamp the page's checksum. The group stays
    // checksum-consistent and its rows materialize fine — only the ordering
    // invariant is broken.
    let poisoned_rows = u64::from(forge_row_group_column(
        &source,
        &fs,
        1,
        crate::table::columnar::COL_USER_KEY,
        swap_first_two_keys,
    )?);

    // Salvage drops exactly the out-of-order block and recovers every other one.
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the out-of-order block is dropped: {report:?}",
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(DropReason::DecodeError(_))
        ),
        "the ordering violation classifies as a decode error: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n) - poisoned_rows,
        "every row outside the poisoned block is recovered",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n) - poisoned_rows);
    Ok(())
}

/// `verify_point_read_reachability` must enforce the GLOBAL internal-key order
/// for a COLUMNAR table too: it has no in-block key index to probe, but
/// `column_batch_match_entries` binary-searches the key column assuming it is
/// sorted. A checksum-consistent columnar block with two swapped keys — which the
/// other columnar gates (count / zone bounds) tolerate — would otherwise pass
/// salvage-mode repair and be KEPT, then miss the key or return a stale version.
#[cfg(feature = "columnar")]
#[test]
fn verify_point_read_reachability_rejects_a_reordered_columnar_block() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    // Swap the first two rows' user keys inside the second row group's key
    // page and re-stamp its checksum: the group stays checksum-consistent and
    // its rows materialize, only the ordering invariant is broken.
    forge_row_group_column(
        &source,
        &fs,
        1,
        crate::table::columnar::COL_USER_KEY,
        swap_first_two_keys,
    )?;

    let table = open(source, &fs)?;
    let err = reconcile_error(
        &table,
        crate::table::ReconcileGate::PointReadReachability,
        None,
    );
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader(
                "columnar entries are out of order (a user key decreased, or an equal key's \
                 seqno did not strictly decrease) across the walk"
            )
        ),
        "the rejection names the columnar order violation, got {err:?}",
    );
    Ok(())
}

/// A checksum-clean columnar block that decodes as a ZERO-ROW `ColumnBatch`
/// is malformed input (a real writer never emits an empty block): the writer
/// primitive emits nothing for it, so counting it as salvaged would let an
/// SST whose only block is empty report `salvaged_path = Some(dest)` while
/// the empty-table `finish` REMOVES `dest` — and a mixed SST would
/// under-report its dropped key ranges. Such a block must be dropped as a
/// decode error.
#[cfg(feature = "columnar")]
#[test]
fn salvage_drops_a_zero_row_columnar_block() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A single-group columnar SST (few rows, default block size).
    let n = 9u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_columnar(true);
    for i in 0..n {
        writer.write(InternalValue::from_components(
            format!("key{i:05}").into_bytes(),
            format!("val{i:05}").into_bytes(),
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Re-stamp the group's directory to declare zero rows. The directory
    // keeps its length and stays checksum-consistent, so the group is
    // refused for WHAT it declares — an empty group, which no writer emits —
    // and not as an ordinary checksum drop.
    forge_row_group_row_count(&source, &fs, 0, 0)?;

    // The zero-row block is DROPPED, so nothing is recoverable: no destination
    // is left behind and no salvaged path is reported (the pre-fix behavior
    // counted it as salvaged, reporting Some(dest) for a file the empty-table
    // finish had just removed).
    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "the zero-row block is dropped as malformed: {report:?}",
    );
    assert_eq!(report.blocks_salvaged, 0, "{report:?}");
    assert_eq!(
        report.salvaged_path, None,
        "an SST whose only block is empty reports nothing salvaged",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind",
    );
    Ok(())
}

/// The ROW-source twin of the columnar out-of-order drop: a checksum-clean
/// row block with two adjacent keys swapped passes frame decode and row
/// materialization, so the ordering guard before the emit is the only thing
/// standing between it and a recovered SST with a corrupt index order. The
/// rejection is block-local: the block drops, the rest still salvages.
#[test]
fn salvage_drops_a_row_block_with_out_of_order_keys() -> crate::Result<()> {
    use crate::coding::Encode;
    use crate::comparator::default_comparator;
    use crate::table::block::Header;
    use crate::table::block::decoder::ParsedItem as _;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    // Restart interval 1 + no hash index: every entry stores its full key, so
    // swapping two equal-length keys re-encodes to a byte-identical length.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_data_block_restart_interval(1)
        .use_data_block_hash_ratio(0.0);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Poison the SECOND data block: decode its entries, swap the first two
    // (equal-length keys), re-encode under the same block parameters, and
    // re-stamp the header checksum.
    let (block_off, poisoned_rows, new_block) = {
        let table = open(source.clone(), &fs)?;
        let Some(kh) = table.data_block_handles().filter_map(Result::ok).nth(1) else {
            panic!("source must have at least two data blocks");
        };
        let block_off = usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX);
        let sb = table.salvage_load_block(kh.as_ref(), crate::table::block::BlockType::Data)?;
        let header = sb.block.header;
        let db = crate::table::DataBlock::from_loaded(sb.block, false)?;
        let iter = db.try_iter(default_comparator())?;
        let mut entries: alloc::vec::Vec<crate::InternalValue> =
            iter.map(|p| p.materialize(db.as_slice())).collect();
        assert!(entries.len() >= 2, "block holds at least two rows");
        entries.swap(0, 1);
        let mut new_payload: alloc::vec::Vec<u8> = alloc::vec::Vec::new();
        crate::table::DataBlock::encode_into(&mut new_payload, &entries, 1, 0.0)?;
        assert_eq!(
            new_payload.len(),
            header.data_length as usize,
            "an adjacent equal-length key swap re-encodes to the same length",
        );
        let header_len = Header::header_len(header.block_type);
        // Bound where the block is, in the table the writer above wrote as
        // id 0.
        let mut new_header = header;
        new_header.bind_checksum(
            crate::Checksum::from_raw(crate::hash::hash128(&new_payload)),
            crate::table::block::ChecksumAt::table(0, block_off as u64),
        );
        let mut new_block: alloc::vec::Vec<u8> =
            alloc::vec::Vec::with_capacity(header_len + new_payload.len());
        new_header.encode_into(&mut new_block)?;
        new_block.extend_from_slice(&new_payload);
        (block_off, entries.len() as u64, new_block)
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(target) = bytes.get_mut(block_off..block_off + new_block.len()) else {
        panic!("block range within the file");
    };
    target.copy_from_slice(&new_block);
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the out-of-order block is dropped: {report:?}",
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(DropReason::DecodeError(_))
        ),
        "the ordering violation classifies as a decode error: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n) - poisoned_rows,
        "every row outside the poisoned block is recovered",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n) - poisoned_rows);
    Ok(())
}

/// A delete-bearing columnar SST whose TLI mirrors AND zone map both omit
/// the same TRAILING block: the positioning chain over the remaining
/// indexed blocks stays self-consistent, so `delete_positions_verified`
/// passes and the walk takes the masked re-emit — but the hidden block is
/// recovered by the PHYSICAL gap walk with no verified start position,
/// and masking it as "all rows live" would permanently resurrect the rows
/// the delete bitmap marked there, without the explicit resurrection
/// opt-in. The masked path must DROP an index-omitted block instead.
#[cfg(feature = "columnar")]
#[test]
fn salvage_does_not_resurrect_deletes_in_an_index_omitted_block() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    // Deletes land in the LAST block (the one both forges will hide).
    let deleted = [n - 1, n - 2, n - 3];
    for pos in deleted {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Row count of the block both forges hide, for the loss accounting.
    let hidden_rows = {
        let table = open(source.clone(), &fs)?;
        let Some(last) = table.data_block_handles().filter_map(Result::ok).last() else {
            panic!("source must have data blocks");
        };
        let Some(batch) = table.load_columnar_block_masked(last.as_ref())? else {
            panic!("the last block holds live rows");
        };
        // The masked load on the INTACT source applies the bitmap, so add
        // the deletes back for the block's physical row count.
        u64::from(batch.row_count) + deleted.len() as u64
    };

    // Consistent forge: the zone map loses its last entry FIRST (the TLI
    // still addresses the block for the entry lookup), then both TLI
    // mirrors hide the same trailing block.
    crate::test_forge::forge_zone_map_drop_last_entry(&source, 0)?;
    crate::test_forge::forge_tli_mirrors_truncated(&source, 0, None)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "the index-omitted block has no verified delete position and must \
         drop, never re-emit live: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n) - deleted.len() as u64 - (hidden_rows - deleted.len() as u64),
        "every indexed block's surviving rows are recovered: {report:?}",
    );
    for pos in deleted {
        let key = format!("key{pos:05}");
        assert!(
            reopen_get(dest.clone(), &fs, key.as_bytes())?.is_none(),
            "a positionally deleted row must not resurrect via the gap walk",
        );
    }
    Ok(())
}

/// The DELETE-BEARING twin of the columnar out-of-order drop: the swapped
/// keys keep every block's row count intact, so the delete positions still
/// verify and the walk takes the masked re-emit — whose writer then rejects
/// the broken ordering. The rejection stays block-local: only the poisoned
/// block drops, deletes still apply to every other block.
#[cfg(feature = "columnar")]
#[test]
fn salvage_drops_an_out_of_order_columnar_block_in_a_delete_bearing_sst() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    let deletes = [5u32, 50, 150];
    for pos in deletes {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Poison the SECOND row group exactly like the delete-free variant: swap
    // the first two rows' user keys inside its key page and re-stamp the
    // checksum. Row counts stay intact, so the delete positions still verify.
    let poisoned_rows = u64::from(forge_row_group_column(
        &source,
        &fs,
        1,
        crate::table::columnar::COL_USER_KEY,
        swap_first_two_keys,
    )?);

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the out-of-order block is dropped: {report:?}",
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(DropReason::DecodeError(_))
        ),
        "the ordering violation classifies as a decode error: {report:?}",
    );
    // Every row outside the poisoned block is recovered, minus the deletes
    // that fall outside it (none of 5 / 50 / 150 land in the second block,
    // which spans rows ~17..34 at 256-byte blocks).
    assert_eq!(
        report.entries_salvaged,
        u64::from(n) - poisoned_rows - deletes.len() as u64,
        "rows outside the poisoned block are recovered with deletes applied",
    );
    // LOGICAL visibility: the deletes were applied faithfully, so the deleted
    // keys stay masked while a neighbouring live key reads back.
    for pos in deletes {
        assert!(
            reopen_get(dest.clone(), &fs, format!("key{pos:05}").as_bytes())?.is_none(),
            "the deleted key at position {pos} stays masked in the recovered copy",
        );
    }
    assert!(
        reopen_get(dest, &fs, b"key00051")?.is_some(),
        "a neighbouring live key reads back from the recovered copy",
    );
    Ok(())
}

/// A DESTINATION write failure mid-walk is a hard error, not a dropped block:
/// the salvage propagates it and removes the partial destination so a retry
/// or repair caller never sees half-written output.
#[test]
fn salvage_sst_errors_and_discards_the_dest_on_a_write_failure() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    // Values big enough that the destination's buffered writer flushes
    // mid-walk (so the failure surfaces through a block emit, not finish).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..200u32 {
        writer.write(InternalValue::from_components(
            format!("key{i:05}").into_bytes(),
            vec![0xAB; 1_024],
            1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    injector.arm(
        FaultRule::new(FaultOp::Write, Fault::Error(ErrorKind::Other))
            .on_path("salvaged")
            .once(),
    );
    let result = salvage_sst(&source, dest.clone(), &fs);
    injector.clear();

    assert!(
        result.is_err(),
        "a destination write failure errors the whole salvage: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "the partial destination is removed on a write failure",
    );
    Ok(())
}

/// A delete-bearing columnar SST whose `delete_bitmap` section is renamed and
/// re-roled to a full `filter` launders the deletion metadata: the TOC still
/// tiles with unique recognized names (nothing catches the rename), no delete
/// bitmap remains visible, and (unlike range tombstones) there is no persisted
/// count to cross-check. The relabeled block's payload is not a real `BuRR`
/// filter, so salvage must PROBE the full filter even though it never pins it —
/// the unparsable payload trips the rebuildable-section degradation and the
/// guard fails closed instead of re-emitting the deleted rows live.
#[cfg(feature = "columnar")]
#[test]
fn salvage_refuses_a_delete_bitmap_relabeled_to_a_full_filter() -> crate::Result<()> {
    use crate::config::{BloomConstructionPolicy, DeleteStrategy};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Filtering disabled, so renaming delete_bitmap to `filter` yields a UNIQUE
    // recognized name (no existing filter to duplicate).
    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .use_bloom_policy(BloomConstructionPolicy::BitsPerKey(0.0))
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Rename delete_bitmap -> filter and re-role its block to Filter.
    crate::test_forge::forge_duplicate_section_name(
        &source,
        b"delete_bitmap",
        b"filter",
        crate::table::block::BlockType::Filter,
    )?;

    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("a delete_bitmap relabeled to a filter must fail salvage");
    };
    let crate::Error::FeatureUnsupported(reason) = &err else {
        panic!("the refusal must be FeatureUnsupported, got {err:?}");
    };
    assert!(
        reason.contains("rebuildable"),
        "the refusal must name the degraded rebuildable section, got {reason:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );
    Ok(())
}

/// A `delete_bitmap` RENAMED to `filter` in the TOC WITHOUT re-roling its block
/// header still launders the deletion: the block loads (its own checksum is
/// valid) but its role is `DeleteBitmap`, so the filter load returns an
/// `InvalidTag` role mismatch. Salvage must treat that STRUCTURAL failure as a
/// degraded rebuildable section (a byte-flip would break the checksum first, so
/// reaching `InvalidTag` means a valid block of the wrong name), not swallow it as
/// genuine bit-rot, so the guard fails closed rather than re-emitting the rows.
#[cfg(feature = "columnar")]
#[test]
fn salvage_refuses_a_delete_bitmap_renamed_to_a_filter_without_reroling() -> crate::Result<()> {
    use crate::config::{BloomConstructionPolicy, DeleteStrategy};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .use_bloom_policy(BloomConstructionPolicy::BitsPerKey(0.0))
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Rename delete_bitmap -> filter in the TOC but KEEP its DeleteBitmap block
    // role (to_role == the original role re-stamps the same header).
    crate::test_forge::forge_duplicate_section_name(
        &source,
        b"delete_bitmap",
        b"filter",
        crate::table::block::BlockType::DeleteBitmap,
    )?;

    let Err(err) = salvage_sst(&source, dest.clone(), &fs) else {
        panic!("a delete_bitmap renamed to a filter without re-roling must fail salvage");
    };
    let crate::Error::FeatureUnsupported(reason) = &err else {
        panic!("the refusal must be FeatureUnsupported, got {err:?}");
    };
    assert!(
        reason.contains("rebuildable"),
        "the refusal must name the degraded rebuildable section, got {reason:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );
    Ok(())
}

/// By default salvage FAILS CLOSED on a delete-bearing SST whose delete
/// bitmap is unreadable: recovering "all rows live" would resurrect
/// positionally-deleted rows, so that degradation requires the caller's
/// explicit [`SalvageOptions::allow_delete_resurrection`] opt-in. No
/// destination file is left behind.
#[cfg(feature = "columnar")]
#[test]
fn salvage_fails_closed_on_a_corrupt_delete_bitmap_by_default() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Corrupt the `delete_bitmap` SFA section (data blocks stay intact).
    let (db_pos, db_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"delete_bitmap") else {
            panic!("source must carry a delete_bitmap section");
        };
        (entry.pos(), entry.len())
    };
    let flip = usize::try_from(db_pos + db_len / 2).unwrap_or(0);
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "default salvage refuses to resurrect deleted rows: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );
    Ok(())
}

/// Same fail-closed default for the other degradation: a READABLE delete
/// bitmap whose positioning zone map is corrupt cannot be applied, so the
/// default salvage refuses rather than recovering all rows live.
#[cfg(feature = "columnar")]
#[test]
fn salvage_fails_closed_on_an_unpositionable_delete_bitmap_by_default() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Corrupt the `zone_map` section (the bitmap stays readable but can no
    // longer be positioned).
    let (zm_pos, zm_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"zone_map") else {
            panic!("source must carry a zone_map section");
        };
        (entry.pos(), entry.len())
    };
    let flip = usize::try_from(zm_pos + zm_len / 2).unwrap_or(0);
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "default salvage refuses to resurrect deleted rows: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );
    Ok(())
}

/// A columnar source carrying deletes whose `delete_bitmap` section is
/// corrupted (data blocks intact): normal recovery refuses to open it (opening
/// would resurrect deleted rows) and default salvage fails closed, but a
/// caller who explicitly opts into [`SalvageOptions::allow_delete_resurrection`]
/// degrades to "all rows live" and recovers every block.
#[cfg(feature = "columnar")]
#[test]
fn salvage_tolerates_a_corrupt_delete_bitmap_as_all_live() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    // Mark a few positions deleted so a delete-bitmap section is co-written.
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Corrupt the middle of the `delete_bitmap` SFA section (the data blocks
    // stay intact, so only the sidecar is damaged).
    let (db_pos, db_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"delete_bitmap") else {
            panic!("source must carry a delete_bitmap section");
        };
        (entry.pos(), entry.len())
    };
    let flip = usize::try_from(db_pos + db_len / 2).unwrap_or(0);
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    // Normal recovery fails closed: a corrupt bitmap would resurrect deleted rows.
    assert!(
        open(source.clone(), &fs).is_err(),
        "normal recovery must fail closed on a corrupt delete-bitmap",
    );

    // With the explicit opt-in, salvage degrades to "all rows live": every
    // block recovers, nothing masked.
    let options = SalvageOptions {
        allow_delete_resurrection: true,
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert!(
        report.is_complete(),
        "the data blocks are intact; only the sidecar was corrupt: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every row is recovered live, the corrupt bitmap is ignored",
    );
    // The SST was written WITH deletes (it carries a delete-bitmap section), so
    // even though the degraded bitmap reads as empty, salvage must NOT take the
    // verbatim copy-through fast path: that would byte-copy the physical blocks
    // (including positionally-deleted rows) without the bitmap. It re-emits
    // instead, so nothing is copied verbatim here.
    assert_eq!(
        report.blocks_copied_verbatim, 0,
        "a delete-bearing SST is never copied verbatim, even with a degraded bitmap: {report:?}",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n));
    Ok(())
}

/// A PERSISTENT read failure of the delete-bitmap SECTION (a bad sector) must
/// honor the salvage opt-in, not abort recovery unconditionally. It fails closed
/// by default, but with the explicit
/// [`SalvageOptions::allow_delete_resurrection`] opt-in it degrades the mask to
/// "all rows live" and recovers every block — the same outcome as a
/// decode-level corruption, reached through the I/O path. Only a TRANSIENT read
/// still propagates for a retry.
#[cfg(feature = "columnar")]
#[test]
fn salvage_tolerates_a_persistently_unreadable_delete_bitmap_as_all_live() -> crate::Result<()> {
    use crate::config::DeleteStrategy;
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let clean: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&clean))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Resolve the delete-bitmap section's file offset.
    let db_pos = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"delete_bitmap") else {
            panic!("source must carry a delete_bitmap section");
        };
        entry.pos()
    };

    // Every positional read of the delete-bitmap section fails with a persistent
    // `Other`/EIO; the data blocks (at other offsets) stay readable.
    let fault = FaultFs::new(StdFs);
    fault
        .injector()
        .arm(FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Other)).at_offset(db_pos));
    let fs: Arc<dyn Fs> = Arc::new(fault);

    // Default policy fails closed: an unreadable bitmap would resurrect deleted rows.
    assert!(
        salvage_sst(&source, dest.clone(), &fs).is_err(),
        "default salvage must fail closed on a persistently unreadable delete-bitmap",
    );

    // With the opt-in, salvage degrades to "all rows live" and recovers every block.
    let options = SalvageOptions {
        allow_delete_resurrection: true,
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every row is recovered live once the unreadable bitmap degrades: {report:?}",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n));
    Ok(())
}

/// A PERSISTENT read failure of the derived `zone_map` SECTION must not fail the
/// whole `Table::recover`: the zone map is a rebuildable block-skip
/// optimization, so a bad sector in it degrades to an empty map (block-skip
/// disabled) rather than making an otherwise-readable SST unopenable. Only a
/// transient read propagates for a retry.
#[test]
fn recover_degrades_a_persistently_unreadable_zone_map_section() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let clean: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&clean))?.use_zone_map(true);
    for i in 0..64u32 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");
    let zm_pos = section_pos(&source, b"zone_map");

    let fault = FaultFs::new(StdFs);
    fault
        .injector()
        .arm(FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Other)).at_offset(zm_pos));
    let fs: Arc<dyn Fs> = Arc::new(fault);

    // Open must SUCCEED with block-skip disabled, not abort on the bad sector.
    let table = open(source, &fs)?;
    assert!(
        table.zone_map.is_empty(),
        "the persistently unreadable zone map degrades to an empty map",
    );
    Ok(())
}

/// As above for the derived `seqno_bounds` SECTION: a persistent read failure
/// disables the seqno block-skip rather than failing `Table::recover`.
#[test]
fn recover_degrades_a_persistently_unreadable_seqno_bounds_section() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let clean: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer =
        Writer::new(source.clone(), 0, 0, Arc::clone(&clean))?.use_seqno_in_index(true);
    for i in 0..64u32 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");
    let sb_pos = section_pos(&source, b"seqno_bounds");

    let fault = FaultFs::new(StdFs);
    fault
        .injector()
        .arm(FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Other)).at_offset(sb_pos));
    let fs: Arc<dyn Fs> = Arc::new(fault);

    // Open must SUCCEED with the seqno block-skip disabled, not abort.
    let table = open(source, &fs)?;
    assert!(
        table.seqno_bounds.is_empty(),
        "the persistently unreadable seqno-bounds section degrades to an empty map",
    );
    Ok(())
}

/// An UNREADABLE data block inside a delete-bearing columnar SST makes every
/// later delete position unverifiable: the block's actual row count is
/// unknowable, and trusting the zone map's claim for it would let a
/// checksum-repatched count on exactly that block shift the masks of all
/// later readable blocks undetected. Default salvage must fail closed; the
/// explicit [`SalvageOptions::allow_delete_resurrection`] opt-in recovers the
/// readable rows live (never masking against unverified positions).
#[cfg(feature = "columnar")]
#[test]
fn salvage_fails_closed_on_an_unreadable_block_in_a_delete_bearing_sst() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Corrupt the SECOND data block's bytes (a plain checksum break): the
    // block becomes unreadable, so its actual row count is unknowable.
    let target = {
        let table = open(source.clone(), &fs)?;
        let offsets: alloc::vec::Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        let Some(&second) = offsets.get(1) else {
            panic!("source must have at least two data blocks, got {offsets:?}");
        };
        second
    };
    let flip = usize::try_from(target).unwrap_or(0) + 16;
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    // Default: fail closed — the delete positions past the unreadable block
    // cannot be proven faithful.
    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "unverifiable delete positions fail the default salvage: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );

    // Explicit opt-in: the readable rows are recovered LIVE; only the corrupt
    // block's rows are lost.
    let options = SalvageOptions {
        allow_delete_resurrection: true,
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the corrupt block is dropped: {report:?}",
    );
    assert!(
        report.entries_salvaged < u64::from(n),
        "the dropped block's rows are lost: {report:?}",
    );
    assert_eq!(
        reopen_item_count(dest.clone(), &fs)?,
        report.entries_salvaged,
        "the recovered copy reopens with every salvaged row live",
    );
    // LOGICAL visibility of the resurrection: the deleted positions live
    // outside the corrupt block, so under the opt-in their keys read back.
    for pos in [5u32, 50, 150] {
        assert!(
            reopen_get(dest.clone(), &fs, format!("key{pos:05}").as_bytes())?.is_some(),
            "the opt-in resurrects the deleted key at position {pos}",
        );
    }
    Ok(())
}

/// A delete-bearing columnar SST where one block was REPLACED by a zero-row
/// batch and the zone map's claim for it patched to 0: the position verifier
/// would accept the block (decoded count 0 matches the claim) while the walk
/// drops it as malformed — leaving later blocks masked at starts that no
/// longer reflect the ORIGINAL row layout the bitmap was built against.
/// A zero-row batch is malformed input everywhere else in the salvage
/// pipeline, so the verifier must reject it too and fail the salvage closed.
#[cfg(feature = "columnar")]
#[test]
fn salvage_fails_closed_on_a_zero_row_block_in_a_delete_bearing_sst() -> crate::Result<()> {
    use crate::coding::{Decode, Encode};
    use crate::config::DeleteStrategy;
    use crate::table::block::Header;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Re-stamp the SECOND row group's directory to declare zero rows —
    // length-preserving and checksum-consistent, the same forgery the plain
    // zero-row test uses.
    forge_row_group_row_count(&source, &fs, 1, 0)?;
    let mut bytes = std::fs::read(&source)?;

    // Patch the zone map's row_count claim for the second block to 0 (the
    // first column's count drives the derived delete starts) and re-stamp
    // the zone-map block checksum, so the tampered chain is self-consistent.
    let zm_pos = {
        let mut f = std::io::Cursor::new(&bytes);
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"zone_map") else {
            panic!("source must carry a zone_map section");
        };
        usize::try_from(entry.pos()).unwrap_or(usize::MAX)
    };
    let Some(mut zm_cursor) = bytes.get(zm_pos..) else {
        panic!("zone_map section within the file");
    };
    let zm_header = Header::decode_from(&mut zm_cursor)?;
    let zm_header_len = Header::header_len(zm_header.block_type);
    let zm_payload_range =
        zm_pos + zm_header_len..zm_pos + zm_header_len + zm_header.data_length as usize;
    let old = crate::hash::hash128(bytes.get(zm_payload_range.clone()).unwrap_or(&[]));
    {
        let Some(payload) = bytes.get_mut(zm_payload_range.clone()) else {
            panic!("zone_map payload within the file");
        };
        // Walk the wire layout (count u32; per block: block_offset u64 +
        // n_columns u16; per column: id u32 + type u8 + codec u8 +
        // null_count u32 + row_count u32 + min_len u32 + min + max_len u32
        // + max) to the SECOND block's FIRST column row_count — the field
        // the derived delete starts are built from.
        let read_u32 = |data: &[u8], at: usize| -> u32 {
            let Some(b) = data.get(at..at + 4) else {
                panic!("u32 at {at} within the zone map payload");
            };
            u32::from_le_bytes(b.try_into().unwrap_or([0; 4]))
        };
        let read_u16 = |data: &[u8], at: usize| -> u16 {
            let Some(b) = data.get(at..at + 2) else {
                panic!("u16 at {at} within the zone map payload");
            };
            u16::from_le_bytes(b.try_into().unwrap_or([0; 2]))
        };
        let mut at = 4; // past the block count
        // Skip block 1 entirely: offset u64 + n_columns u16, then each
        // column's fixed 14 bytes + variable min/max.
        at += 8;
        let block1_cols = read_u16(payload, at);
        at += 2;
        for _ in 0..block1_cols {
            at += 10; // id + type + codec + null_count
            at += 4; // row_count
            let min_len = read_u32(payload, at) as usize;
            at += 4 + min_len;
            let max_len = read_u32(payload, at) as usize;
            at += 4 + max_len;
        }
        // Block 2: seek to its first column's row_count and zero it.
        at += 8; // block_offset
        at += 2; // n_columns
        at += 10; // first column's id + type + codec + null_count
        let claimed = read_u32(payload, at);
        assert!(claimed > 0, "the second block originally holds rows");
        let Some(rc) = payload.get_mut(at..at + 4) else {
            panic!("second block's first row_count within the zone map payload");
        };
        rc.copy_from_slice(&0u32.to_le_bytes());
    }
    let new_zm_header =
        crate::test_forge::restamp(zm_header, old, bytes.get(zm_payload_range).unwrap_or(&[]));
    let mut zm_hdr_bytes: Vec<u8> = Vec::with_capacity(zm_header_len);
    new_zm_header.encode_into(&mut zm_hdr_bytes)?;
    let Some(zm_dst) = bytes.get_mut(zm_pos..zm_pos + zm_header_len) else {
        panic!("zone_map header within the file");
    };
    zm_dst.copy_from_slice(&zm_hdr_bytes);
    std::fs::write(&source, &bytes)?;

    // Default: fail closed — the zero-row block is unpositionable input (the
    // walk drops it while later blocks would be masked at starts the bitmap
    // was never built against).
    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "a zero-row block in a delete-bearing SST fails the default salvage: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );
    Ok(())
}

/// A footer-bearing row SST (per-KV checksums) whose block checksum was
/// re-stamped over a tampered entry: the BLOCK checksum verifies, but the
/// entry no longer matches its per-KV digest. Salvage must verify the footer
/// before emitting (verbatim or re-encoded) — otherwise it recovers a block
/// the live per-KV scrub would reject, laundering the corruption into a
/// "fully valid" copy.
#[test]
fn salvage_drops_a_row_block_with_a_stale_kv_digest() -> crate::Result<()> {
    use crate::coding::{Decode, Encode};
    use crate::runtime_config::{ChecksumAlgorithm, KvChecksumPolicy};
    use crate::table::block::Header;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_kv_checksums(KvChecksumPolicy::AllLevels, ChecksumAlgorithm::Xxh3_64);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Tamper one entry byte inside the SECOND block's inner payload and
    // re-stamp the BLOCK checksum: the frame verifies clean, but the entry's
    // stored per-KV digest no longer matches its bytes.
    let (block_off, block_size) = {
        let table = open(source.clone(), &fs)?;
        let Some(kh) = table.data_block_handles().filter_map(Result::ok).nth(1) else {
            panic!("source must have at least two data blocks");
        };
        (
            usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX),
            kh.as_ref().size() as usize,
        )
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(block) = bytes.get(block_off..block_off + block_size) else {
        panic!("block range within the file");
    };
    let mut cursor = block;
    let header = Header::decode_from(&mut cursor)?;
    let header_len = Header::header_len(header.block_type);
    let payload_range =
        block_off + header_len..block_off + header_len + header.data_length as usize;
    let old = crate::hash::hash128(bytes.get(payload_range.clone()).unwrap_or(&[]));
    {
        let Some(payload) = bytes.get_mut(payload_range.clone()) else {
            panic!("payload range within the block");
        };
        // A byte inside the first entry's value bytes (past the entry header),
        // well before the per-KV footer at the payload tail.
        let Some(b) = payload.get_mut(12) else {
            panic!("entry byte within the payload");
        };
        *b ^= 0xFF;
    }
    let new_header =
        crate::test_forge::restamp(header, old, bytes.get(payload_range).unwrap_or(&[]));
    let mut hdr_bytes: Vec<u8> = Vec::with_capacity(header_len);
    new_header.encode_into(&mut hdr_bytes)?;
    let Some(hdr_dst) = bytes.get_mut(block_off..block_off + header_len) else {
        panic!("block header within the file");
    };
    hdr_dst.copy_from_slice(&hdr_bytes);
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "the block with a stale per-KV digest is dropped: {report:?}",
    );
    assert!(
        report.entries_salvaged < u64::from(n),
        "the tampered block's rows are not laundered into the copy: {report:?}",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, report.entries_salvaged);
    Ok(())
}

/// A delete-bearing columnar SST with a checksum-clean block whose
/// `ColumnBatch` does NOT decode (a repatched tamper that keeps the leading
/// row-count u32 intact but breaks the column framing): the block's ACTUAL row
/// count is unknowable, so every later block's delete positions are
/// unverifiable — the position verifier must fully decode each block rather
/// than trust the leading four bytes, and the default salvage must fail
/// closed instead of dropping the block and masking later rows at positions
/// it could not prove.
#[cfg(feature = "columnar")]
#[test]
fn salvage_fails_closed_on_an_undecodable_checksum_clean_block_with_deletes() -> crate::Result<()> {
    use crate::coding::{Decode, Encode};
    use crate::config::DeleteStrategy;
    use crate::table::block::Header;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Poison the SECOND data block: stamp an invalid type tag into the first
    // column's header (payload layout: row_count u32, column_count u32, then
    // per column id u16 + type u8 + ... — so the first type byte sits at
    // payload offset 10) and re-stamp the block checksum. The leading
    // row-count u32 stays intact, the frame stays checksum-consistent, but
    // `ColumnBatch::decode` fails on the unknown tag.
    let (block_off, block_size) = {
        let table = open(source.clone(), &fs)?;
        let Some(kh) = table.data_block_handles().filter_map(Result::ok).nth(1) else {
            panic!("source must have at least two data blocks");
        };
        (
            usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX),
            kh.as_ref().size() as usize,
        )
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(block) = bytes.get(block_off..block_off + block_size) else {
        panic!("block range within the file");
    };
    let mut cursor = block;
    let header = Header::decode_from(&mut cursor)?;
    let header_len = Header::header_len(header.block_type);
    let payload_range =
        block_off + header_len..block_off + header_len + header.data_length as usize;
    let old = crate::hash::hash128(bytes.get(payload_range.clone()).unwrap_or(&[]));
    {
        let Some(payload) = bytes.get_mut(payload_range.clone()) else {
            panic!("payload range within the block");
        };
        let Some(tag) = payload.get_mut(10) else {
            panic!("first column's type byte within the payload");
        };
        *tag = 0xEE;
    }
    let new_header =
        crate::test_forge::restamp(header, old, bytes.get(payload_range).unwrap_or(&[]));
    let mut hdr_bytes = Vec::with_capacity(header_len);
    new_header.encode_into(&mut hdr_bytes)?;
    let Some(hdr_dst) = bytes.get_mut(block_off..block_off + header_len) else {
        panic!("block header within the file");
    };
    hdr_dst.copy_from_slice(&hdr_bytes);
    std::fs::write(&source, &bytes)?;

    // Default: fail closed — the block reads back checksum-clean but cannot
    // be decoded, so its actual row count (and every later block's delete
    // positions) cannot be proven faithful.
    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "unverifiable delete positions fail the default salvage: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );

    // Explicit opt-in: the poisoned block is dropped, every other row is
    // recovered LIVE (never masked against unproven positions).
    let options = SalvageOptions {
        allow_delete_resurrection: true,
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the poisoned block is dropped: {report:?}",
    );
    assert!(
        report.entries_salvaged < u64::from(n),
        "the poisoned block's rows are lost: {report:?}",
    );
    assert_eq!(
        reopen_item_count(dest.clone(), &fs)?,
        report.entries_salvaged,
        "the recovered copy reopens with every salvaged row live",
    );
    // LOGICAL visibility of the resurrection: the deleted positions live
    // outside the poisoned block, so under the opt-in their keys read back.
    for pos in [5u32, 50, 150] {
        assert!(
            reopen_get(dest.clone(), &fs, format!("key{pos:05}").as_bytes())?.is_some(),
            "the opt-in resurrects the deleted key at position {pos}",
        );
    }
    Ok(())
}

/// A zone map that DECODES but carries wrong per-block row counts (a
/// checksum-repatched tamper) would misposition the delete bitmap: the masked
/// re-emit derives each block's start row from the zone map, so deletes land
/// on the wrong rows — deleted rows resurrect AND live rows vanish, silently.
/// Salvage must cross-check the claimed positions against the actual decoded
/// row counts and fail closed on a mismatch; with the explicit
/// [`SalvageOptions::allow_delete_resurrection`] opt-in it recovers all rows
/// live instead of masking against the wrong positions.
#[cfg(feature = "columnar")]
#[test]
fn salvage_fails_closed_on_a_zone_map_with_wrong_row_counts() -> crate::Result<()> {
    use crate::coding::{Decode, Encode};
    use crate::config::DeleteStrategy;
    use crate::table::block::Header;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Tamper the FIRST block's row count inside the zone map (wire layout:
    // count u32, then block_offset u64 + n_columns u16, then per column
    // id u32 + type_tag u8 + codec_id u8 + null_count u32 + row_count u32 —
    // so the first row_count sits at payload bytes 24..28) and re-stamp the
    // section block's checksum. The zone map still DECODES — only its claimed
    // positions are shifted for every block after the first.
    let zm_pos = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"zone_map") else {
            panic!("source must carry a zone_map section");
        };
        usize::try_from(entry.pos()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    let Some(mut cursor) = bytes.get(zm_pos..) else {
        panic!("zone_map section within the file");
    };
    let header = Header::decode_from(&mut cursor)?;
    let header_len = Header::header_len(header.block_type);
    let payload_range = zm_pos + header_len..zm_pos + header_len + header.data_length as usize;
    let old = crate::hash::hash128(bytes.get(payload_range.clone()).unwrap_or(&[]));
    {
        let Some(payload) = bytes.get_mut(payload_range.clone()) else {
            panic!("zone_map payload within the file");
        };
        let Some(rc) = payload.get_mut(24..28) else {
            panic!("first row_count within the zone map payload");
        };
        let claimed = u32::from_le_bytes(rc.try_into().unwrap_or([0; 4]));
        assert!(claimed >= 2, "the first block holds at least two rows");
        rc.copy_from_slice(&(claimed - 1).to_le_bytes());
    }
    let new_header =
        crate::test_forge::restamp(header, old, bytes.get(payload_range).unwrap_or(&[]));
    let mut hdr_bytes = Vec::with_capacity(header_len);
    new_header.encode_into(&mut hdr_bytes)?;
    let Some(hdr_dst) = bytes.get_mut(zm_pos..zm_pos + header_len) else {
        panic!("zone_map header within the file");
    };
    hdr_dst.copy_from_slice(&hdr_bytes);
    std::fs::write(&source, &bytes)?;

    // Default: fail closed — masking against the shifted positions would
    // silently corrupt visibility in the recovered SST.
    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.is_err(),
        "a mispositioning zone map fails the default salvage: {result:?}",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "no destination file is left behind by the refused salvage",
    );

    // Explicit opt-in: recover all rows LIVE (never mask against the wrong
    // positions).
    let options = SalvageOptions {
        allow_delete_resurrection: true,
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert!(
        report.is_complete(),
        "the data blocks are intact; only the zone map lies: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every row is recovered live under the opt-in",
    );
    assert_eq!(reopen_item_count(dest.clone(), &fs)?, u64::from(n));
    // LOGICAL visibility of the resurrection: "all rows live" means the keys
    // at the deleted positions read back.
    for pos in [5u32, 50, 150] {
        assert!(
            reopen_get(dest.clone(), &fs, format!("key{pos:05}").as_bytes())?.is_some(),
            "the opt-in resurrects the deleted key at position {pos}",
        );
    }
    Ok(())
}

/// A columnar SST with deletes whose ZONE MAP is corrupt (the bitmap stays
/// readable): the bitmap cannot be positioned without the zone map, so normal
/// recovery and default salvage fail closed, but a caller who explicitly opts
/// into [`SalvageOptions::allow_delete_resurrection`] ignores the bitmap
/// ("all rows live") and recovers every row.
#[cfg(feature = "columnar")]
#[test]
fn salvage_ignores_a_delete_bitmap_without_a_readable_zone_map() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 50, 150] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Corrupt the zone_map section (the bitmap stays intact). The zone map
    // degrades to empty, leaving a readable bitmap that cannot be positioned.
    let (zm_pos, zm_len) = {
        let mut f = std::fs::File::open(&source)?;
        let reader = match crate::sfa::Reader::from_reader(&mut f) {
            Ok(r) => r,
            Err(e) => panic!("reading the SFA trailer failed: {e:?}"),
        };
        let Some(entry) = reader.toc().iter().find(|e| e.name() == b"zone_map") else {
            panic!("source must carry a zone_map section");
        };
        (entry.pos(), entry.len())
    };
    let flip = usize::try_from(zm_pos + zm_len / 2).unwrap_or(0);
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    // Normal recovery fails closed: a bitmap with no positioning zone map.
    assert!(
        open(source.clone(), &fs).is_err(),
        "normal recovery must reject a bitmap with no readable zone map",
    );

    // With the explicit opt-in, salvage ignores the unpositionable bitmap and
    // recovers every row live.
    let options = SalvageOptions {
        allow_delete_resurrection: true,
        ..SalvageOptions::default()
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert!(
        report.is_complete(),
        "the data blocks are intact; only the zone map was corrupt: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every row is recovered live once the unpositionable bitmap is ignored",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n));
    Ok(())
}

/// When the source cannot be opened at all (a corrupt SFA trailer makes even
/// salvage-mode recovery fail), `salvage_sst` returns an error rather than
/// writing a partial file.
#[test]
fn salvage_sst_errors_when_the_source_cannot_be_opened() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..50 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Truncate away the tail (SFA trailer + section mirrors) so the container is
    // unparseable and even salvage-mode recovery cannot open it.
    let mut bytes = std::fs::read(&source)?;
    bytes.truncate(bytes.len() / 2);
    std::fs::write(&source, &bytes)?;

    assert!(
        salvage_sst(&source, dest.clone(), &fs).is_err(),
        "an unparseable container must fail salvage, not write a partial file",
    );
    assert!(
        !dest.exists(),
        "no destination is written on an open failure"
    );
    Ok(())
}

/// A single-block SST whose only data block is corrupt salvages nothing: no
/// destination file is written and the report records the dropped block.
#[test]
fn salvage_sst_recovers_nothing_when_the_only_block_is_corrupt() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A handful of small keys fit in one default-sized data block.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..8 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Corrupt the sole data block (offset from the intact index).
    let target = {
        let table = open(source.clone(), &fs)?;
        let offsets: alloc::vec::Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        let Some(&only) = offsets.first() else {
            panic!("expected a single data block, got {offsets:?}");
        };
        assert_eq!(
            offsets.len(),
            1,
            "expected a single data block, got {offsets:?}"
        );
        only
    };
    let flip = usize::try_from(target).unwrap_or(0) + 16;
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.blocks_salvaged, 0, "the only block was corrupt");
    assert_eq!(report.entries_salvaged, 0, "no entries recovered");
    assert_eq!(report.dropped.len(), 1, "the dropped block is reported");
    assert!(
        report.salvaged_path.is_none(),
        "nothing recoverable means no file is written",
    );
    assert!(!dest.exists(), "no destination file on an empty salvage");
    Ok(())
}

/// A columnar source whose delete-bitmap wholly covers its leading data
/// block(s): those blocks carry no live rows, so salvage skips them (nothing
/// salvaged, nothing dropped) and recovers the live rows of the rest.
#[cfg(feature = "columnar")]
#[test]
fn salvage_skips_a_wholly_deleted_block() -> crate::Result<()> {
    use crate::config::DeleteStrategy;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let n = 200u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_row_group_size(256)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    // Delete the first 60 row positions: with 256-byte blocks this wholly covers
    // the leading data block(s), which then load as "no live rows".
    let deleted = 60u32;
    for pos in 0..deleted {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "wholly-deleted blocks are skipped, not dropped: {report:?}",
    );
    assert!(
        report.blocks_salvaged < report.blocks_total,
        "at least one leading block was wholly deleted and skipped: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n - deleted),
        "every live row is recovered, the deleted prefix is skipped",
    );
    assert_eq!(reopen_item_count(dest, &fs)?, u64::from(n - deleted));
    Ok(())
}

/// An SST carrying range tombstones cannot be salvaged: the positional KV walk
/// re-emits only point entries, so the tombstones would be silently dropped and
/// lower-level keys they cover could reappear after repair. Until the writer
/// path re-emits them, salvage fails closed.
#[test]
fn salvage_rejects_an_sst_with_range_tombstones() -> crate::Result<()> {
    use crate::UserKey;
    use crate::range_tombstone::RangeTombstone;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..20 {
        writer.write(iv(i))?;
    }
    // A range tombstone over part of the key space: the salvaged copy must not
    // silently drop it.
    writer.write_range_tombstone(RangeTombstone::new(
        UserKey::from(b"key00005".as_slice()),
        UserKey::from(b"key00010".as_slice()),
        2,
    ));
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        matches!(result, Err(crate::Error::FeatureUnsupported(_))),
        "an SST with range tombstones must fail closed, got {result:?}",
    );
    assert!(
        !dest.exists(),
        "no salvaged file is written when salvage fails closed",
    );
    Ok(())
}

/// Salvage drives every read and write through the injected `Fs`: an SST that
/// lives only in an in-memory backend (never on the real filesystem) salvages
/// and reopens purely through that backend. A source-digest path that bypassed
/// `fs` and read through `std::fs` would fail to find the file at all.
#[test]
fn salvage_sst_reads_and_writes_through_the_injected_fs() -> crate::Result<()> {
    use crate::fs::MemFs;

    let fs: Arc<dyn Fs> = Arc::new(MemFs::new());
    // `Writer::new` rewrites its path through `std::path::absolute`, which on
    // Windows resolves a `/`-rooted path against the current drive (`/memfs` ->
    // `D:\memfs`). Create the parent under that same absolutized form so the
    // writer's parent-directory check finds it on every platform (on Unix
    // `absolute` is a no-op, so this is just `/memfs`).
    let dir = std::path::absolute("/memfs")?;
    fs.create_dir_all(&dir)?;
    let source = dir.join("source");
    let dest = dir.join("salvaged");

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(256);
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "in-memory source SST is non-empty"
    );

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "a healthy in-memory SST salvages with no dropped blocks: {report:?}",
    );
    assert_eq!(
        report.entries_salvaged,
        u64::from(n),
        "every entry is recovered through the in-memory backend",
    );
    assert_eq!(report.salvaged_path.as_deref(), Some(dest.as_path()));
    assert_eq!(
        reopen_item_count(dest, &fs)?,
        u64::from(n),
        "the salvaged SST reopens through the same in-memory backend",
    );
    Ok(())
}

// --- Forwarded recovery context: encrypted + dictionary-compressed sources ---

/// Reads the second data block's on-disk offset from a context-aware reopen of
/// `source`, then flips a byte just past that block's header so its checksum /
/// AEAD tag fails on load while every other block stays intact.
#[cfg(any(feature = "encryption", zstd_any))]
fn corrupt_second_data_block(
    source: &std::path::Path,
    fs: &Arc<dyn Fs>,
    table_id: crate::table::TableId,
    encryption: Option<Arc<dyn crate::encryption::EncryptionProvider>>,
    #[cfg(zstd_any)] zstd_dictionary: Option<Arc<crate::compression::ZstdDictionary>>,
) -> crate::Result<()> {
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&**fs, source)?);
    let table = {
        // Open under the source's table id so an encrypted index (AAD binds the
        // id) decrypts when reading the block offsets.
        let mut params = crate::table::RecoverParams::new(
            source.to_path_buf(),
            checksum,
            table_id,
            Arc::clone(fs),
            default_comparator(),
            Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
        );
        params.descriptor_table = Some(Arc::new(crate::descriptor_table::DescriptorTable::new(8)));
        params.encryption = encryption;
        #[cfg(zstd_any)]
        {
            params.zstd_dictionaries = zstd_dictionary
                .map_or_else(crate::compression::ZstdDictionaries::new, |dict| {
                    crate::compression::ZstdDictionaries::new().with(dict)
                });
        }
        Table::recover(params)?
    };
    let offsets: alloc::vec::Vec<u64> = table
        .data_block_handles()
        .filter_map(Result::ok)
        .map(|kh| *kh.as_ref().offset())
        .collect();
    let Some(&second) = offsets.get(1) else {
        panic!("source SST must have at least two data blocks, got {offsets:?}");
    };
    let flip = usize::try_from(second).unwrap_or(0) + 16;
    let mut bytes = std::fs::read(source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(source, &bytes)?;
    Ok(())
}

/// An encrypted source: salvage cannot open it without the provider (the gap this
/// closes), but with the provider in `SalvageOptions` it block-salvages like a
/// plain SST and the recovered copy reopens under the same encryption.
#[cfg(feature = "encryption")]
#[test]
fn salvage_recovers_an_encrypted_sst_with_the_provider() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let enc: Arc<dyn crate::encryption::EncryptionProvider> =
        Arc::new(crate::encryption::Aes256GcmProvider::new(&[0x42; 32]));

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_encryption(Some(Arc::clone(&enc)));
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source encrypted SST is non-empty"
    );

    corrupt_second_data_block(
        &source,
        &fs,
        0,
        Some(Arc::clone(&enc)),
        #[cfg(zstd_any)]
        None,
    )?;

    // Without the provider, the encrypted source cannot even be opened.
    assert!(
        salvage_sst(&source, dest.clone(), &fs).is_err(),
        "an encrypted SST must not salvage without the provider",
    );

    // With the provider, it block-salvages: the corrupt block is dropped, the
    // rest recovered, and the copy is written encrypted.
    let options = SalvageOptions {
        encryption: Some(Arc::clone(&enc)),
        #[cfg(zstd_any)]
        zstd_dictionary: None,
        #[cfg(zstd_any)]
        zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
        table_id: 0,
        expected_stored_id: None,
        output_id: None,
        allow_delete_resurrection: false,
        sync_mode: crate::fs::SyncMode::Normal,
        prefix_extractor: None,
        blob_rewrite: None,
        owner_promotions: None,
        progress: None,
        clock: None,
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the corrupt block drops: {report:?}"
    );
    assert!(
        report.entries_salvaged > 0 && report.entries_salvaged < u64::from(n),
        "a partial key range is recovered, got {} of {n}",
        report.entries_salvaged,
    );

    // The salvaged copy reopens UNDER ENCRYPTION (a plaintext copy would fail the
    // encrypted reopen) and holds exactly the recovered entries.
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&*fs, &dest)?);
    let reopened = {
        let mut params = crate::table::RecoverParams::new(
            dest,
            checksum,
            0,
            Arc::clone(&fs),
            default_comparator(),
            Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
        );
        params.descriptor_table = Some(Arc::new(crate::descriptor_table::DescriptorTable::new(8)));
        params.encryption = Some(Arc::clone(&enc));
        Table::recover(params)?
    };
    assert_eq!(
        reopened.metadata.item_count, report.entries_salvaged,
        "the encrypted salvaged copy reopens with exactly the recovered entries",
    );
    Ok(())
}

/// A copy published under a new id reads back: an encrypted block is sealed
/// under its table id, so a byte copy of the source's blocks would not decrypt
/// under the copy's id. A plain source is still copied byte for byte.
#[cfg(feature = "encryption")]
#[test]
fn a_salvaged_copy_under_a_new_id_reads_every_value() -> crate::Result<()> {
    let dir = tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let key_provider: Arc<dyn crate::encryption::EncryptionProvider> =
        Arc::new(crate::encryption::Aes256GcmProvider::new(&[0x42; 32]));
    let (source_id, output_id) = (3, 4);

    let layouts: &[bool] = if cfg!(feature = "columnar") {
        &[false, true]
    } else {
        &[false]
    };
    for &columnar in layouts {
        for enc in [None, Some(&key_provider)] {
            let name = format!("{columnar}-{}", enc.is_some());
            let source = dir.path().join(format!("{name}-source"));
            let dest = dir.path().join(format!("{name}-copy"));
            let mut writer = Writer::new(source.clone(), source_id, 0, Arc::clone(&fs))?
                .use_columnar(columnar)
                .use_encryption(enc.map(Arc::clone));
            for i in 0..20_u32 {
                writer.write(InternalValue::from_components(
                    format!("key{i:05}").into_bytes(),
                    format!("val{i:05}").into_bytes(),
                    1,
                    ValueType::Value,
                ))?;
            }
            assert!(writer.finish()?.is_some(), "source is non-empty");

            let options = SalvageOptions {
                encryption: enc.map(Arc::clone),
                #[cfg(zstd_any)]
                zstd_dictionary: None,
                #[cfg(zstd_any)]
                zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
                table_id: source_id,
                expected_stored_id: None,
                output_id: Some(output_id),
                allow_delete_resurrection: false,
                sync_mode: crate::fs::SyncMode::Normal,
                prefix_extractor: None,
                blob_rewrite: None,
                owner_promotions: None,
                progress: None,
                clock: None,
            };
            let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
            assert!(
                report.dropped.is_empty(),
                "{name}: a clean source: {report:?}"
            );

            let copy = open_as(dest, &fs, enc, output_id)?;
            for i in 0..20_u32 {
                let key = format!("key{i:05}");
                let got = copy.get(
                    key.as_bytes(),
                    crate::MAX_SEQNO,
                    crate::hash::hash64(key.as_bytes()),
                );
                let value = match got {
                    Ok(Some(entry)) => entry.value,
                    other => panic!("{name}: {key} in the copy, got {other:?}"),
                };
                assert_eq!(&*value, format!("val{i:05}").as_bytes(), "{name}: {key}");
            }
            assert_eq!(
                report.blocks_copied_verbatim > 0,
                enc.is_none(),
                "{name}: only a plain source is copied byte for byte: {report:?}",
            );
        }
    }
    Ok(())
}

/// A zstd-dictionary-compressed source: salvage cannot decompress it without the
/// dictionary, but with the dictionary in `SalvageOptions` it block-salvages and
/// the recovered copy reopens under the same dictionary.
#[cfg(zstd_any)]
#[test]
fn salvage_recovers_a_dictionary_sst_with_the_dictionary() -> crate::Result<()> {
    use crate::CompressionType;
    use crate::compression::ZstdDictionary;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A small training corpus so the dictionary has content to match against.
    let samples: alloc::vec::Vec<u8> = (0..4000u32).map(|i| (i % 251) as u8).collect();
    let dict = Arc::new(ZstdDictionary::new(&samples));
    let compression = CompressionType::ZstdDict {
        level: 3,
        dict_id: dict.id(),
    };

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_data_block_compression(compression)
        .use_zstd_dictionary(Some(Arc::clone(&dict)));
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source dictionary SST is non-empty"
    );

    corrupt_second_data_block(&source, &fs, 0, None, Some(Arc::clone(&dict)))?;

    // Without the dictionary, the source cannot even be opened: `recover_inner`
    // fail-fasts on the ZstdDict-id mismatch at open time (before any block
    // walk), so salvage returns `Err`, not a zero-recovered report.
    assert!(
        salvage_sst(&source, dest.clone(), &fs).is_err(),
        "a dictionary SST must not salvage without the dictionary",
    );

    let options = SalvageOptions {
        encryption: None,
        zstd_dictionary: Some(Arc::clone(&dict)),
        zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
        table_id: 0,
        expected_stored_id: None,
        output_id: None,
        allow_delete_resurrection: false,
        sync_mode: crate::fs::SyncMode::Normal,
        prefix_extractor: None,
        blob_rewrite: None,
        owner_promotions: None,
        progress: None,
        clock: None,
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the corrupt block drops: {report:?}"
    );
    assert!(
        report.entries_salvaged > 0 && report.entries_salvaged < u64::from(n),
        "a partial key range is recovered, got {} of {n}",
        report.entries_salvaged,
    );

    // The salvaged copy reopens UNDER THE DICTIONARY with the recovered entries.
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&*fs, &dest)?);
    let reopened = {
        let mut params = crate::table::RecoverParams::new(
            dest,
            checksum,
            0,
            Arc::clone(&fs),
            default_comparator(),
            Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
        );
        params.descriptor_table = Some(Arc::new(crate::descriptor_table::DescriptorTable::new(8)));
        params.zstd_dictionaries =
            crate::compression::ZstdDictionaries::new().with(Arc::clone(&dict));
        Table::recover(params)?
    };
    assert_eq!(
        reopened.metadata.item_count, report.entries_salvaged,
        "the dictionary salvaged copy reopens with exactly the recovered entries",
    );
    Ok(())
}

/// A dictionary-compressed BLOB file salvaged under the WRONG dictionary must
/// fail closed BEFORE the record walk. Without the up-front id check, every
/// frame's decompress fails with a per-record dictionary mismatch that the
/// walk's catch-all records as a `Corrupt` drop — the run "succeeds" with
/// zero records salvaged, and a repair that trusts the report throws away a
/// fully intact file whose only problem was a mis-supplied dictionary.
#[cfg(zstd_any)]
#[test]
fn a_blob_salvage_under_the_wrong_dictionary_fails_closed() -> crate::Result<()> {
    use crate::CompressionType;
    use crate::compression::ZstdDictionary;

    let dir = tempdir()?;
    let source = dir.path().join("dict_blob");
    let dest = dir.path().join("dict_blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let samples_a: alloc::vec::Vec<u8> = (0..4000u32).map(|i| (i % 251) as u8).collect();
    let dict_a = Arc::new(ZstdDictionary::new(&samples_a));
    let compression = CompressionType::ZstdDict {
        level: 3,
        dict_id: dict_a.id(),
    };

    let mut writer = BlobWriter::new(&source, 0, 0, &*fs)?
        .use_compression(compression)
        .use_zstd_dictionary(Some(Arc::clone(&dict_a)));
    for i in 0..8u32 {
        let key = format!("key{i:04}");
        let value: alloc::vec::Vec<u8> = (0..512u32).map(|j| ((i + j) % 251) as u8).collect();
        writer.write(key.as_bytes(), 0, &value)?;
    }
    writer.finish()?;

    // A DIFFERENT training corpus yields a different dictionary id.
    let samples_b: alloc::vec::Vec<u8> = (0..4000u32).map(|i| (i % 13) as u8).collect();
    let dict_b = Arc::new(ZstdDictionary::new(&samples_b));
    assert_ne!(dict_a.id(), dict_b.id(), "the fixture needs distinct ids");

    let Err(err) = salvage_blob_file(
        &source,
        dest,
        &fs,
        0,
        &default_comparator(),
        0,
        Some(&dict_b),
    ) else {
        panic!("a mismatched dictionary must fail the salvage, not empty it");
    };
    assert!(
        matches!(
            &err,
            crate::Error::ZstdDictMismatch { expected, got }
                if *expected == dict_a.id() && *got == Some(dict_b.id())
        ),
        "the failure names both ids: {err:?}",
    );
    Ok(())
}

/// An encrypted source sealed under a NON-ZERO table id: the encrypted-block AAD
/// binds the table id, so salvage must be given that id. With the wrong id the
/// AAD-bound blocks cannot be decrypted (the gap repair hit when it passed a
/// hardcoded `0`); with the right id it block-salvages and the copy reopens.
#[cfg(feature = "encryption")]
#[test]
fn salvage_recovers_an_encrypted_sst_with_a_nonzero_table_id() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let enc: Arc<dyn crate::encryption::EncryptionProvider> =
        Arc::new(crate::encryption::Aes256GcmProvider::new(&[0x37; 32]));
    const TID: crate::table::TableId = 7;

    let mut writer = Writer::new(source.clone(), TID, 0, Arc::clone(&fs))?
        .use_data_block_size(256)
        .use_encryption(Some(Arc::clone(&enc)));
    let n = 200u32;
    for i in 0..n {
        writer.write(iv(i))?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source encrypted SST is non-empty"
    );

    corrupt_second_data_block(
        &source,
        &fs,
        TID,
        Some(Arc::clone(&enc)),
        #[cfg(zstd_any)]
        None,
    )?;

    // Wrong table id (the legacy hardcoded 0): the AAD-bound blocks cannot be
    // decrypted, so nothing is recovered (salvage either fails to open or drops
    // every block).
    let wrong = SalvageOptions {
        encryption: Some(Arc::clone(&enc)),
        #[cfg(zstd_any)]
        zstd_dictionary: None,
        #[cfg(zstd_any)]
        zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
        table_id: 0,
        expected_stored_id: None,
        output_id: None,
        allow_delete_resurrection: false,
        sync_mode: crate::fs::SyncMode::Normal,
        prefix_extractor: None,
        blob_rewrite: None,
        owner_promotions: None,
        progress: None,
        clock: None,
    };
    let recovered_wrong = salvage_sst_with_options(&source, dest.clone(), &fs, &wrong)
        .map_or(0, |r| r.entries_salvaged);
    assert_eq!(
        recovered_wrong, 0,
        "the wrong table id cannot decrypt the AAD-bound encrypted source",
    );

    // Right table id: block-salvages, dropping only the corrupt block.
    let options = SalvageOptions {
        encryption: Some(Arc::clone(&enc)),
        #[cfg(zstd_any)]
        zstd_dictionary: None,
        #[cfg(zstd_any)]
        zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
        table_id: TID,
        expected_stored_id: None,
        output_id: None,
        allow_delete_resurrection: false,
        sync_mode: crate::fs::SyncMode::Normal,
        prefix_extractor: None,
        blob_rewrite: None,
        owner_promotions: None,
        progress: None,
        clock: None,
    };
    let report = salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "exactly the corrupt block drops: {report:?}"
    );
    assert!(
        report.entries_salvaged > 0 && report.entries_salvaged < u64::from(n),
        "a partial key range is recovered, got {} of {n}",
        report.entries_salvaged,
    );

    // The recovered copy reopens under the same table id + encryption.
    let checksum = crate::Checksum::from_raw(crate::repair::compute_table_checksum(&*fs, &dest)?);
    let reopened = {
        let mut params = crate::table::RecoverParams::new(
            dest,
            checksum,
            TID,
            Arc::clone(&fs),
            default_comparator(),
            Arc::new(crate::cache::Cache::with_capacity_bytes(1 << 20)),
        );
        params.descriptor_table = Some(Arc::new(crate::descriptor_table::DescriptorTable::new(8)));
        params.encryption = Some(Arc::clone(&enc));
        Table::recover(params)?
    };
    assert_eq!(
        reopened.metadata.item_count, report.entries_salvaged,
        "the recovered copy reopens under the same table id with the recovered entries",
    );
    Ok(())
}

// --- Blob (vlog) file record-granular salvage ---

use crate::vlog::blob_file::scanner::Scanner as BlobScanner;
use crate::vlog::blob_file::writer::Writer as BlobWriter;

/// Builds a blob file at `path` from `(key, value)` records (seqno 0, no
/// compression).
fn build_blob(
    path: &std::path::Path,
    fs: &Arc<dyn Fs>,
    records: &[(&[u8], &[u8])],
) -> crate::Result<()> {
    let mut writer = BlobWriter::new(path, 0, 0, &**fs)?;
    for (k, v) in records {
        writer.write(k, 0, v)?;
    }
    writer.finish()?;
    Ok(())
}

/// Mirrors that disagree in a NON-DERIVABLE field must not be arbitrated at
/// all. The salvage walk re-derives the entry-backed metadata from the records
/// it re-emits, so a "complete" attempt settles those — but `bulk_ingested`,
/// `recency` and the compaction lineage are copied from whichever mirror is
/// selected, and nothing in the file authenticates them. A forged tail that
/// clears `bulk_ingested` would otherwise win on block completeness alone and
/// republish an ingested SST at global seqno 0, visible to snapshots that
/// never saw it.
#[test]
fn divergent_non_derivable_metadata_is_never_arbitrated() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A bulk-ingested SST: every entry at local seqno 0, provenance recorded.
    let mut writer =
        Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_bulk_ingested(Some(true));
    for i in 0..32u32 {
        writer.write(crate::InternalValue::from_components(
            format!("k{i:05}").into_bytes(),
            b"v".to_vec(),
            0,
            crate::ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "the source SST is non-empty");

    // Forge ONLY the tail mirror's provenance: the blocks stay intact, so the
    // tail attempt recovers everything and would win on completeness.
    crate::test_forge::forge_tail_meta_value(&source, b"descriptor#bulk_ingested", &[0])?;

    let result = salvage_sst(&source, dest, &fs);
    assert!(
        result.is_err(),
        "a disagreement in provenance the walk cannot re-derive must refuse \
         arbitration, not publish the mirror that happens to decode: {result:?}",
    );
    Ok(())
}

/// The PREPASS twin: the physical tiling walk that discovers blocks must
/// propagate an environmental read too. A break in that chain drops the
/// candidate — and, past an untrusted index, the whole unanchored tail — as
/// permanently lost, so repair publishes the partial replacement and removes
/// a source whose bytes were never proven corrupt. The sweep drives a
/// corrupt table (which forces the physical walk) with a one-shot
/// `PermissionDenied` at increasing skip counts.
#[test]
fn an_environmental_read_never_breaks_the_physical_tiling_silently() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let mut fault_reached_the_walk = false;
    for skip in 0..48u64 {
        let dir = tempdir()?;
        let source = dir.path().join("source");
        let dest = dir.path().join("salvaged");
        let plain: Arc<dyn Fs> = Arc::new(StdFs);
        {
            let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&plain))?
                .use_data_block_size(128)
                .use_partitioned_index();
            for i in 0..256u32 {
                writer.write(iv(i))?;
            }
            assert!(writer.finish()?.is_some(), "the source SST is non-empty");
        }
        // Break the INDEX section: enumeration stops partway and the salvage
        // falls back to the physical tiling walk, which is the path under test.
        {
            let (index_pos, index_len) = {
                let mut f = std::fs::File::open(&source)?;
                let reader = crate::sfa::Reader::from_reader(&mut f)?;
                let Some((pos, len)) = reader
                    .toc()
                    .iter()
                    .find(|e| e.name() == b"index")
                    .map(|e| (e.pos(), e.len()))
                else {
                    panic!("the SST carries an index section");
                };
                (pos, len)
            };
            let Ok(flip) = usize::try_from(index_pos + index_len / 2) else {
                panic!("the index-section offset fits usize");
            };
            let mut bytes = std::fs::read(&source)?;
            if let Some(b) = bytes.get_mut(flip) {
                *b ^= 0xFF;
            }
            std::fs::write(&source, &bytes)?;
        }

        let fault = FaultFs::new(StdFs);
        fault.injector().arm(
            FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::PermissionDenied))
                .skip(skip)
                .times(1),
        );
        let fs: Arc<dyn Fs> = Arc::new(fault);
        match salvage_sst(&source, dest, &fs) {
            Ok(report) => {
                assert!(
                    !report
                        .dropped
                        .iter()
                        .any(|d| format!("{:?}", d.reason).contains("PermissionDenied")),
                    "an environmental failure must never be recorded as a \
                     dropped block (skip {skip}): {report:?}",
                );
            }
            Err(e) => {
                assert!(
                    matches!(&e, crate::Error::Io(io) if io.kind() == ErrorKind::PermissionDenied),
                    "the only acceptable failure is the propagated \
                     environmental error (skip {skip}): {e:?}",
                );
                fault_reached_the_walk = true;
            }
        }
    }
    assert!(
        fault_reached_the_walk,
        "no skip count reached the salvage read path; the sweep proves nothing",
    );
    Ok(())
}

/// The SST twin: an ENVIRONMENTAL data-block read failure must PROPAGATE,
/// never be recorded as a dropped block. The metadata and index are readable,
/// so the walk reaches the block reads; a `PermissionDenied` there is an ACL
/// mistake or host pressure, not rotted bytes, and accepting it as a drop
/// finishes a partial replacement that repair installs before removing the
/// source — permanent loss a fixed environment would have avoided. The sweep
/// arms a one-shot fault at increasing skip counts; wherever it lands, the
/// walk either succeeds untouched or fails with the environmental error.
#[test]
fn an_environmental_read_failure_never_becomes_a_lossy_sst_salvage() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let mut fault_reached_the_walk = false;
    for skip in 0..64u64 {
        let dir = tempdir()?;
        let source = dir.path().join("source");
        let dest = dir.path().join("salvaged");
        let plain: Arc<dyn Fs> = Arc::new(StdFs);
        {
            let mut writer =
                Writer::new(source.clone(), 0, 0, Arc::clone(&plain))?.use_data_block_size(256);
            for i in 0..64u32 {
                writer.write(iv(i))?;
            }
            assert!(writer.finish()?.is_some(), "the source SST is non-empty");
        }

        let fault = FaultFs::new(StdFs);
        fault.injector().arm(
            FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::PermissionDenied))
                .skip(skip)
                .times(1),
        );
        let fs: Arc<dyn Fs> = Arc::new(fault);
        match salvage_sst(&source, dest, &fs) {
            Ok(report) => {
                assert!(
                    !report
                        .dropped
                        .iter()
                        .any(|d| format!("{:?}", d.reason).contains("PermissionDenied")),
                    "an environmental failure must never be recorded as a \
                     dropped block (skip {skip}): {report:?}",
                );
            }
            Err(e) => {
                assert!(
                    matches!(&e, crate::Error::Io(io) if io.kind() == ErrorKind::PermissionDenied),
                    "the only acceptable failure is the propagated \
                     environmental error (skip {skip}): {e:?}",
                );
                fault_reached_the_walk = true;
            }
        }
    }
    assert!(
        fault_reached_the_walk,
        "no skip count reached the salvage read path; the sweep proves nothing",
    );
    Ok(())
}

/// An ENVIRONMENTAL read failure mid-walk (`PermissionDenied`, `OutOfMemory`
/// — an ACL mistake or host pressure, not rotted bytes) must PROPAGATE like
/// a transient one, never be recorded as corruption: accepting it as a drop
/// finishes a lossy salvage whose committed loss a fixed environment would
/// have avoided entirely. The sweep arms a one-shot `PermissionDenied` read
/// fault at increasing skip counts; wherever it lands, the walk either
/// succeeds untouched or fails with the environmental error — no run may
/// report a `PermissionDenied` as a corrupt drop.
#[test]
fn an_environmental_read_failure_never_becomes_a_lossy_blob_salvage() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let mut fault_reached_the_walk = false;
    for skip in 0..64u64 {
        let dir = tempdir()?;
        let source = dir.path().join("blob_env");
        let dest = dir.path().join("blob_env_salvaged");
        let plain: Arc<dyn Fs> = Arc::new(StdFs);
        build_blob(
            &source,
            &plain,
            &[
                (b"aaaa", b"AAAAAAAA"),
                (b"bbbb", b"BBBBBBBB"),
                (b"cccc", b"CCCCCCCC"),
            ],
        )?;

        let fault = FaultFs::new(StdFs);
        fault.injector().arm(
            FaultRule::new(FaultOp::Read, Fault::Error(ErrorKind::PermissionDenied))
                .skip(skip)
                .times(1),
        );
        let fs: Arc<dyn Fs> = Arc::new(fault);
        let result = salvage_blob_file(
            &source,
            dest,
            &fs,
            0,
            &default_comparator(),
            0,
            #[cfg(zstd_any)]
            None,
        );
        match result {
            Ok(report) => {
                assert!(
                    !report
                        .dropped
                        .iter()
                        .any(|d| matches!(&d.reason, BlobDropReason::Corrupt(msg) if msg.contains("PermissionDenied"))),
                    "an environmental failure must never be recorded as a \
                     corrupt drop (skip {skip}): {report:?}",
                );
            }
            Err(e) => {
                assert!(
                    matches!(&e, crate::Error::Io(io) if io.kind() == ErrorKind::PermissionDenied),
                    "the only acceptable failure is the propagated \
                     environmental error (skip {skip}): {e:?}",
                );
                fault_reached_the_walk = true;
            }
        }
    }
    assert!(
        fault_reached_the_walk,
        "no skip count reached the salvage read path; the sweep proves nothing",
    );
    Ok(())
}

/// Scans a blob file into its `(key, value)` records (Ok records only).
fn scan_blob(path: &std::path::Path, fs: &Arc<dyn Fs>) -> crate::Result<Vec<(Vec<u8>, Vec<u8>)>> {
    Ok(BlobScanner::new(path, &**fs, 0)?
        .filter_map(Result::ok)
        .map(|e| (e.key.to_vec(), e.value.to_vec()))
        .collect())
}

#[test]
fn salvage_blob_file_recovers_every_record_of_a_healthy_file() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let records: Vec<(&[u8], &[u8])> = vec![
        (b"k0", b"v0"),
        (b"k1", b"v1"),
        (b"k2", b"v2"),
        (b"k3", b"v3"),
    ];
    build_blob(&source, &fs, &records)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert!(
        report.is_complete(),
        "a healthy blob file drops nothing: {report:?}"
    );
    assert_eq!(report.records_salvaged, 4);
    assert_eq!(report.salvaged_path.as_deref(), Some(dest.as_path()));

    let recovered = scan_blob(&dest, &fs)?;
    let expected: Vec<(Vec<u8>, Vec<u8>)> = records
        .iter()
        .map(|(k, v)| (k.to_vec(), v.to_vec()))
        .collect();
    assert_eq!(
        recovered, expected,
        "every record round-trips through salvage"
    );
    Ok(())
}

/// A published salvage whose TEMP-name removal fails persistently must
/// surface the error, not shrug: the repair caller commits a manifest that
/// never references the temp, so a stuck `.healtmp-` name makes the next
/// open's artifact sweep hit the same removal failure — reporting success
/// would describe a tree that cannot open. Arbitration forces the
/// temp-then-publish path; `lz4`-gated for the mirror-divergence forge.
#[cfg(feature = "lz4")]
#[test]
fn publish_fails_when_the_temp_name_cannot_be_dropped() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");

    // The fault filter has no separator, so it matches Windows paths too.
    let fault = FaultFs::new(StdFs);
    fault.injector().arm(
        FaultRule::new(
            FaultOp::RemoveFile,
            Fault::Error(ErrorKind::PermissionDenied),
        )
        .on_path("healtmp"),
    );
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..10 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.as_ref().is_err_and(|e| matches!(
            e,
            crate::Error::SalvageOutput(e)
                if matches!(&**e, crate::Error::Io(e) if e.kind() == ErrorKind::PermissionDenied)
        )),
        "a stuck temp name must fail the salvage as a destination failure: {result:?}",
    );
    // The freshly linked destination must be unwound with the error: left
    // installed, a retry after the filesystem is fixed would bounce off
    // AlreadyExists forever despite this call reporting failure.
    assert!(
        !dest.exists(),
        "the unpublished destination must not survive the failure",
    );
    Ok(())
}

/// A backend WITHOUT `hard_link` (the trait's default `Unsupported` impl)
/// whose `exists` probe races a concurrent creator: the probe reports the
/// destination free, but a full file already sits there by publish time.
/// The publish must still refuse to replace it — `rename` REPLACES an
/// existing destination by contract, so a probe-then-rename fallback would
/// silently overwrite the concurrently published file. The claim must be
/// atomic (`create_new`), never a TOCTOU probe.
///
/// The wrapper models the race deterministically: `exists` always answers
/// `false` (the probe's view), every other operation is the real backend.
/// Divergent meta mirrors force the arbitration path, whose winner publishes
/// through the temp-then-claim helper; `lz4`-gated because the divergence is
/// forged via `compression#data` → Lz4.
#[cfg(feature = "lz4")]
#[test]
fn publish_refuses_to_replace_a_concurrently_created_destination_without_hard_links()
-> crate::Result<()> {
    use crate::fs::{Fs, FsDirEntry, FsFile, FsMetadata, FsOpenOptions};
    use crate::io;
    use std::path::Path;

    /// Delegates to [`StdFs`], but: `hard_link` stays the trait default
    /// (`Unsupported`), and `exists` always reports `false` — the state the
    /// probe saw before a concurrent creator published the destination.
    #[derive(Debug)]
    struct RacingProbeFs(StdFs);
    impl Fs for RacingProbeFs {
        fn open(&self, path: &Path, opts: &FsOpenOptions) -> io::Result<Box<dyn FsFile>> {
            self.0.open(path, opts)
        }
        fn create_dir_all(&self, path: &Path) -> io::Result<()> {
            self.0.create_dir_all(path)
        }
        fn read_dir(&self, path: &Path) -> io::Result<Vec<FsDirEntry>> {
            self.0.read_dir(path)
        }
        fn remove_file(&self, path: &Path) -> io::Result<()> {
            self.0.remove_file(path)
        }
        fn remove_dir_all(&self, path: &Path) -> io::Result<()> {
            self.0.remove_dir_all(path)
        }
        fn rename(&self, from: &Path, to: &Path) -> io::Result<()> {
            self.0.rename(from, to)
        }
        fn metadata(&self, path: &Path) -> io::Result<FsMetadata> {
            self.0.metadata(path)
        }
        fn sync_directory(&self, path: &Path) -> io::Result<()> {
            self.0.sync_directory(path)
        }
        fn exists(&self, _path: &Path) -> io::Result<bool> {
            Ok(false)
        }
    }

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // The concurrently published destination the probe cannot see.
    let concurrent = b"concurrently published content".to_vec();
    std::fs::write(&dest, &concurrent)?;

    let fs: Arc<dyn Fs> = Arc::new(RacingProbeFs(StdFs));
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0u64..10 {
        writer.write(InternalValue::from_components(
            format!("key-{i:03}").into_bytes(),
            format!("val-{i:03}").into_bytes(),
            i + 1,
            ValueType::Value,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");
    // Diverge the meta mirrors so salvage takes the arbitration path, whose
    // winning attempt publishes from a private temp.
    crate::test_forge::forge_tail_meta_value(&source, b"compression#data", &[1])?;

    let result = salvage_sst(&source, dest.clone(), &fs);
    assert!(
        result.as_ref().is_err_and(|e| matches!(
            e,
            crate::Error::SalvageOutput(e)
                if matches!(&**e, crate::Error::Io(e) if e.kind() == crate::io::ErrorKind::AlreadyExists)
        )),
        "the publish must refuse the taken destination, not replace it: {result:?}",
    );
    assert_eq!(
        std::fs::read(&dest)?,
        concurrent,
        "the concurrently published file must survive byte-for-byte",
    );
    Ok(())
}

/// A bare relative destination (`recovered`, no parent component) must
/// salvage cleanly on `MemFs` too: the backend accepts the empty parent as
/// its implicit root at creation, so the post-publication entry sync (which
/// names that root `.`) must not fail and drag the freshly written
/// destination down with it.
///
/// No SST twin exists for this shape: the SST writer absolutizes its path up
/// front (so open / remove / directory fsync all see one name), which means a
/// bare relative SST destination on `MemFs` fails cleanly at CREATION (the
/// absolutized parent is no `MemFs` directory) — the post-publication window
/// this test pins is unreachable on that path.
#[test]
fn salvage_blob_file_accepts_a_bare_relative_destination_on_memfs() -> crate::Result<()> {
    let fs: Arc<dyn Fs> = Arc::new(crate::fs::MemFs::new());
    let source = std::path::Path::new("blob_source");
    let dest = std::path::PathBuf::from("recovered");

    let records: Vec<(&[u8], &[u8])> = vec![(b"k0", b"v0"), (b"k1", b"v1")];
    build_blob(source, &fs, &records)?;

    let report = salvage_blob_file(
        source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert!(report.is_complete(), "nothing to drop: {report:?}");
    assert_eq!(report.records_salvaged, 2);
    assert!(
        fs.exists(&dest)?,
        "the salvaged destination survives its entry sync",
    );
    Ok(())
}

/// A checksum-consistent frame whose key regresses below the previous salvaged
/// record must be DROPPED, not re-emitted: `BlobWriter` requires records in
/// key order, and a salvaged file that violates it corrupts its own key range
/// and later breaks the merge scanner's per-reader sorted-input assumption. The
/// source here is written physically out of order (the writer does not enforce
/// its precondition), so every frame is individually valid but the sequence is
/// not sorted.
#[test]
fn salvage_blob_file_drops_a_frame_whose_key_regresses() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Physically out of order: k2 precedes k1. Each frame is checksum-valid.
    let records: Vec<(&[u8], &[u8])> = vec![(b"k0", b"v0"), (b"k2", b"v2"), (b"k1", b"v1")];
    build_blob(&source, &fs, &records)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(
        report.records_salvaged, 2,
        "the order-regressing frame is dropped, not re-emitted: {report:?}",
    );
    let [dropped] = report.dropped.as_slice() else {
        panic!(
            "exactly the regressing frame drops, got {:?}",
            report.dropped
        );
    };
    assert!(
        matches!(dropped.reason, BlobDropReason::Corrupt(_)),
        "the drop is recorded as corruption: {:?}",
        dropped.reason,
    );

    // The salvaged file is sorted, so it re-opens under the writer's contract.
    let recovered = scan_blob(&dest, &fs)?;
    assert_eq!(
        recovered,
        vec![
            (b"k0".to_vec(), b"v0".to_vec()),
            (b"k2".to_vec(), b"v2".to_vec())
        ],
        "only the in-order prefix survives, still sorted",
    );
    Ok(())
}

/// The blob-order guard must use the TREE comparator, not raw bytes: a blob file
/// written by a reverse-comparator tree is in descending key order, which is
/// perfectly valid there. Salvaging it under that comparator must keep every
/// record; bytewise ordering would wrongly judge each record as regressing and
/// drop healthy data.
#[test]
fn salvage_blob_file_keeps_reverse_ordered_records_under_a_reverse_comparator() -> crate::Result<()>
{
    struct ReverseComparator;
    impl crate::comparator::UserComparator for ReverseComparator {
        fn name(&self) -> &'static str {
            "reverse-lexicographic-test"
        }
        fn compare(&self, a: &[u8], b: &[u8]) -> core::cmp::Ordering {
            b.cmp(a)
        }
    }

    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // Descending key order: correct for a reverse-comparator tree.
    let records: Vec<(&[u8], &[u8])> = vec![(b"k2", b"v2"), (b"k1", b"v1"), (b"k0", b"v0")];
    build_blob(&source, &fs, &records)?;

    let comparator: crate::comparator::SharedComparator = Arc::new(ReverseComparator);
    let report = salvage_blob_file(
        &source,
        dest,
        &fs,
        0,
        &comparator,
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(
        report.records_salvaged, 3,
        "descending order is NOT regressing under a reverse comparator: {report:?}",
    );
    assert!(
        report.dropped.is_empty(),
        "no record is dropped as out-of-order: {report:?}",
    );
    Ok(())
}

/// `salvage_blob_file` must fsync the destination's PARENT DIRECTORY before
/// returning a salvaged path: the writer syncs the blob file's bytes, but
/// without the directory sync a power loss can discard the new directory entry
/// even though the report says recovery succeeded (the SST salvage writer
/// already syncs its parent directory). Fault-inject the directory fsync to
/// prove it happens — a build that never syncs the directory never triggers the
/// fault and wrongly reports the move durable.
#[test]
fn salvage_blob_file_syncs_the_destination_directory() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultInjector, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let out = dir.path().join("blobdest");
    std::fs::create_dir_all(&out)?;
    let dest = out.join("blob_salvaged");

    let injector = Arc::new(FaultInjector::new());
    let fs: Arc<dyn Fs> = Arc::new(FaultFs::with_injector(StdFs, Arc::clone(&injector)));

    let records: Vec<(&[u8], &[u8])> = vec![(b"k0", b"v0"), (b"k1", b"v1")];
    build_blob(&source, &fs, &records)?;

    // Arm AFTER building, so only the salvage's destination-directory sync trips.
    injector.arm(
        FaultRule::new(FaultOp::SyncDirectory, Fault::Error(ErrorKind::Other)).on_path("blobdest"),
    );

    let Err(err) = salvage_blob_file(
        &source,
        dest,
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    ) else {
        panic!("the destination-directory fsync fault must surface");
    };
    // Assert the surfaced error is the INJECTED directory-sync fault, not some
    // unrelated failure: a build that stopped syncing the directory would fail
    // elsewhere (or not at all), and a loose "any error" check would not notice.
    assert!(
        err.to_string().contains("injected fault on SyncDirectory"),
        "the salvage error must be the injected destination-directory sync fault, got {err:?}",
    );
    Ok(())
}

/// The TOCTOU variant of the pre-existing-destination guarantee: a file that
/// appears at `dest` AFTER any existence probe but BEFORE the writer's
/// `create_new` open (a concurrent worker winning the destination) must also
/// survive the failed salvage. Ownership is decided by `create_new` alone —
/// when it fails, this call created nothing and must remove nothing. The
/// injected `Metadata` fault materializes the race window deterministically:
/// the probe cannot see the file, yet `create_new` finds it.
#[test]
fn salvage_blob_file_keeps_a_racing_dest_created_after_the_existence_probe() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_dest");
    let plain: Arc<dyn Fs> = Arc::new(StdFs);
    build_blob(&source, &plain, &[(b"k0", b"v0")])?;

    // The "racing" worker's file is already at dest, but any metadata probe of
    // dest fails — exactly the window where the file lands between a stat and
    // the `create_new` open.
    std::fs::write(&dest, b"racing worker's blob")?;
    let fault = FaultFs::new(StdFs);
    fault.injector().arm(
        FaultRule::new(FaultOp::Metadata, Fault::Error(ErrorKind::NotFound)).on_path("blob_dest"),
    );
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let result = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    );
    assert!(
        result.is_err(),
        "the destination is taken, the salvage fails: {result:?}",
    );
    assert_eq!(
        std::fs::read(&dest)?,
        b"racing worker's blob",
        "the racing worker's file survives the failed salvage",
    );
    Ok(())
}

/// A transient I/O failure on the verbatim REREAD must not drop the block:
/// the first, recovery-aware read has already produced a verified decoded
/// block, so the loader falls back to the re-encode path (`verbatim = None`)
/// exactly like a checksum / parity mismatch on the re-read frame. Reserving
/// `Err` for the initial verified read keeps one flaky pread from discarding
/// a block that is provably recoverable.
#[test]
fn salvage_load_block_reencodes_when_the_verbatim_reread_fails() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;
    use crate::table::block::BlockType;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..50u32 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Collect the first data block's handle BEFORE arming the fault (the
    // open + index walk issue their own reads).
    let table = open(source, &fs)?;
    let Some(kh) = table.data_block_handles().find_map(Result::ok) else {
        panic!("source has at least one data block");
    };
    let handle = *kh.as_ref();

    // Within `salvage_load_block` the FIRST positional read is the verified
    // recovery-aware load; the SECOND is the raw verbatim re-read. Fail
    // exactly that second read, once.
    injector.arm(
        FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Other))
            .on_path("source")
            .skip(1)
            .once(),
    );
    let result = table.salvage_load_block(&handle, BlockType::Data);
    injector.clear();

    let sb = match result {
        Ok(sb) => sb,
        Err(e) => panic!("a failed verbatim re-read falls back to re-encode, got Err({e:?})"),
    };
    assert!(
        sb.verbatim.is_none(),
        "the re-read was never verified, so the block must not be byte-copied",
    );
    assert!(
        !sb.block.data.is_empty(),
        "the verified first read's decoded payload is preserved for re-encoding",
    );
    Ok(())
}

/// `tli_structure_authenticated` must PROPAGATE a TRANSIENT I/O fault from
/// opening / reading the index mirrors rather than fold it into `false` (an
/// untrusted index). The salvage walk consults this to decide whether an indexed
/// offset is a trusted block boundary; reading a flaky open as `false` degrades
/// to physical-chain-only provenance and surrenders every block past the first
/// header break — dropping healthy keys the intact TLI could anchor on retry. A
/// single transient-kind (`Interrupted`) `Open` fault reproduces the flaky
/// failure; the method must return the propagated I/O error.
#[test]
fn tli_structure_authenticated_propagates_a_transient_read_failure() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..50u32 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Open cleanly, THEN fault the dedicated open the authentication read issues
    // with a TRANSIENT (`Interrupted`) kind.
    let table = open(source, &fs)?;
    injector.arm(
        FaultRule::new(FaultOp::Open, Fault::Error(ErrorKind::Interrupted))
            .on_path("source")
            .once(),
    );
    let result = table.tli_structure_authenticated();
    injector.clear();
    assert!(
        matches!(result, Err(crate::Error::Io(_))),
        "a transient failure authenticating the index structure must propagate, not be \
         read as an untrusted index: {result:?}",
    );
    Ok(())
}

/// `tli_structure_authenticated` must DEGRADE a PERSISTENT I/O fault to `false`
/// (an untrusted index) rather than propagate it. When one mirror is
/// persistently unreadable (a bad sector) but the other mirror and the data
/// section stay readable, propagating the error makes the `salvage_blocks`
/// caller's `?` abort before its physical-chain fallback, recovering NONE of the
/// intact data blocks. `false` lets the walk fall back to physical-chain
/// provenance and recover the readable blocks.
#[test]
fn tli_structure_authenticated_degrades_a_persistent_read_failure() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    for i in 0..50u32 {
        writer.write(iv(i))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Open cleanly, THEN fault the authentication read's open with a PERSISTENT
    // (`Other`/EIO) kind.
    let table = open(source, &fs)?;
    injector.arm(FaultRule::new(FaultOp::Open, Fault::Error(ErrorKind::Other)).on_path("source"));
    let result = table.tli_structure_authenticated();
    injector.clear();
    assert!(
        matches!(result, Ok(false)),
        "a persistent authentication failure must degrade to an untrusted index so the \
         salvage walk can still recover the readable data: {result:?}",
    );
    Ok(())
}

/// `delete_positions_verified` must PROPAGATE a transient I/O fault rather than
/// fold it into `false` (an unpositionable mask). Under the default
/// `allow_delete_resurrection == false` a `false` verdict aborts salvage, so a
/// flaky block read during `repair_with_salvage` would drop the table from the
/// rebuilt manifest even though a retry could recover it faithfully. A single
/// `ReadAt` fault on the first positional read the walk issues reproduces the
/// transient read; the method must return the propagated I/O error, not
/// `Ok(false)`.
#[cfg(feature = "columnar")]
#[test]
fn delete_positions_verified_propagates_a_transient_block_read_failure() -> crate::Result<()> {
    use crate::config::DeleteStrategy;
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Open cleanly (footer + block-index reads happen here), THEN fault the
    // first positional read the verification walk issues with a TRANSIENT
    // (`Interrupted`) kind.
    let table = open(source, &fs)?;
    injector.arm(
        FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Interrupted))
            .on_path("source")
            .once(),
    );
    let result = table.delete_positions_verified();
    injector.clear();
    assert!(
        matches!(result, Err(crate::Error::Io(_))),
        "a transient read during delete-position validation must propagate, not be read \
         as a persistent unpositionable mask: {result:?}",
    );
    Ok(())
}

/// An ENVIRONMENTAL failure is not always transient: `PermissionDenied` (a
/// refused mount), `StorageFull`, a missing key, a missing dictionary. None of
/// them say anything about the DATA, and every one of them is fixed by
/// correcting the context and retrying. Folding one into `Ok(false)` makes the
/// mask unpositionable, which under the default resurrection policy aborts the
/// salvage — so `repair_with_salvage` rebuilds the manifest without a table
/// that was never damaged.
#[cfg(feature = "columnar")]
#[test]
fn delete_positions_verified_propagates_an_environmental_block_read_failure() -> crate::Result<()> {
    use crate::config::DeleteStrategy;
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Open cleanly, THEN refuse the first positional read the walk issues.
    // `PermissionDenied` is environmental but NOT transient — the class the
    // narrower gate let fall through to `Ok(false)`.
    let table = open(source, &fs)?;
    injector.arm(
        FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::PermissionDenied))
            .on_path("source")
            .once(),
    );
    let result = table.delete_positions_verified();
    injector.clear();
    assert!(
        matches!(result, Err(crate::Error::Io(ref e)) if e.kind() == ErrorKind::PermissionDenied),
        "a refused read during delete-position validation must propagate, not be read \
         as a persistent unpositionable mask: {result:?}",
    );
    Ok(())
}

/// `delete_positions_verified` must DEGRADE a PERSISTENT positional read failure
/// to `Ok(false)` (an unpositionable mask), not propagate it: with
/// `allow_delete_resurrection` a bad sector under one block still lets the caller
/// re-emit the remaining readable rows unmasked, instead of aborting the whole
/// salvage.
#[cfg(feature = "columnar")]
#[test]
fn delete_positions_verified_degrades_a_persistent_block_read_failure() -> crate::Result<()> {
    use crate::config::DeleteStrategy;
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let n = 64u32;
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for i in 0..n {
        writer.write(iv(i))?;
    }
    for pos in [5u32, 20, 40] {
        writer.delete_bitmap_mut().insert(pos);
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar+deletes SST is non-empty",
    );

    // Open cleanly, THEN fault the first positional read the verification walk
    // issues with a PERSISTENT (`Other`/EIO) kind.
    let table = open(source, &fs)?;
    injector.arm(
        FaultRule::new(FaultOp::ReadAt, Fault::Error(ErrorKind::Other))
            .on_path("source")
            .once(),
    );
    let result = table.delete_positions_verified();
    injector.clear();
    assert!(
        matches!(result, Ok(false)),
        "a persistent read during delete-position validation must degrade to an \
         unpositionable mask so the resurrection opt-in can take effect: {result:?}",
    );
    Ok(())
}

/// `salvage_blob_file` must not delete a pre-existing file at `dest` when the
/// destination cannot be created (the writer's `create_new` open fails because
/// the path already exists): the error-path cleanup is only for a partial file
/// THIS call created. Deleting a pre-existing destination would turn an
/// argument mistake (a stale path collision, or `source == dest`) into data
/// loss.
#[test]
fn salvage_blob_file_keeps_a_preexisting_dest_on_open_failure() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_dest");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    build_blob(&source, &fs, &[(b"k0", b"v0")])?;
    std::fs::write(&dest, b"pre-existing destination bytes")?;

    let result = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    );
    assert!(
        result.is_err(),
        "an already-existing destination fails the salvage: {result:?}",
    );
    assert_eq!(
        std::fs::read(&dest)?,
        b"pre-existing destination bytes",
        "a pre-existing destination file survives the failed salvage",
    );
    Ok(())
}

/// The salvaged blob file is COMPACTED: after a dropped record every later
/// record shifts to a new offset, and existing SST `ValueHandle::offset`
/// values point into the SOURCE. The report's `offset_remap` must map every
/// salvaged record's source frame offset to its offset in the recovered file
/// (and omit the dropped one), so a caller can re-target handles before
/// swapping the file in.
#[test]
fn salvage_blob_file_reports_an_offset_remap_for_every_salvaged_record() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let records: Vec<(&[u8], &[u8])> = vec![
        (b"k0", b"v0-payload"),
        (b"k1", b"v1-payload"),
        (b"k2", b"v2-payload"),
        (b"k3", b"v3-payload"),
    ];
    build_blob(&source, &fs, &records)?;

    // Source frame offsets, in order, from a clean pre-corruption scan.
    let source_offsets: Vec<u64> = BlobScanner::new(&source, &*fs, 0)?
        .filter_map(Result::ok)
        .map(|e| e.offset)
        .collect();
    assert_eq!(source_offsets.len(), 4, "four source records");

    // Corrupt the SECOND record's value bytes (a checksum break): the scanner
    // re-syncs at the next magic. k2, reached by that byte scan through k1's
    // damaged bytes, has an unprovable boundary; the taint is sticky, so the walk
    // STOPS there and reports the surrendered tail (k2, k3) as ONE drop. So only
    // record 0 (before the corruption) survives.
    {
        let Some(&second) = source_offsets.get(1) else {
            panic!("second record offset");
        };
        // Past the frame header, inside key/value bytes.
        let flip = usize::try_from(second).unwrap_or(0) + 45;
        let mut bytes = std::fs::read(&source)?;
        if let Some(b) = bytes.get_mut(flip) {
            *b ^= 0xFF;
        }
        std::fs::write(&source, &bytes)?;
    }

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(report.records_salvaged, 1, "{report:?}");
    assert_eq!(
        report.dropped.len(),
        2,
        "the corrupt record + the surrendered tail (recorded once) drop: {report:?}"
    );

    // The remap covers exactly the salvaged records, keyed by their SOURCE
    // offsets, and its targets are the actual frame offsets AND on-disk value
    // sizes in the recovered file (verified against a scan of the destination —
    // a live read cross-checks the handle's size against the frame header, so
    // the relocation must carry the re-emitted size, not the source's).
    let dest_records: Vec<(u64, u32)> = BlobScanner::new(&dest, &*fs, 0)?
        .filter_map(Result::ok)
        .map(|e| {
            let value_start = e.offset
                + u64::try_from(crate::vlog::blob_file::writer::BLOB_HEADER_LEN).unwrap_or(0)
                + u64::try_from(e.key.len()).unwrap_or(0);
            (
                e.offset,
                u32::try_from(e.frame_end - value_start).unwrap_or(0),
            )
        })
        .collect();
    let expected: Vec<(u64, super::BlobRecordRelocation)> = std::iter::once(&0usize)
        .zip(&dest_records)
        .map(|(&src_idx, &(offset, on_disk_size))| {
            (
                source_offsets.get(src_idx).copied().unwrap_or(u64::MAX),
                super::BlobRecordRelocation {
                    offset,
                    on_disk_size,
                },
            )
        })
        .collect();
    assert_eq!(
        report.offset_remap, expected,
        "the remap maps each surviving source frame to its compacted target",
    );
    // The dropped record's source offset is NOT in the map: its handle is lost.
    let Some(&dropped_src) = source_offsets.get(1) else {
        panic!("second record offset");
    };
    assert!(
        report
            .offset_remap
            .iter()
            .all(|(src, _)| *src != dropped_src),
        "a dropped record has no remap target: {report:?}",
    );
    Ok(())
}

/// When a record write to the destination fails mid-salvage, `salvage_blob_file`
/// must error AND remove the partial destination it created, so a retry / repair
/// caller never finds a half-written blob file.
#[test]
fn salvage_blob_file_removes_the_partial_dest_when_a_write_fails() -> crate::Result<()> {
    use crate::fs::{Fault, FaultFs, FaultOp, FaultRule};
    use crate::io::ErrorKind;

    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fault = FaultFs::new(StdFs);
    let injector = fault.injector();
    let fs: Arc<dyn Fs> = Arc::new(fault);

    let records: Vec<(&[u8], &[u8])> = vec![(b"k0", b"v0"), (b"k1", b"v1")];
    build_blob(&source, &fs, &records)?;

    // Fail every write to the destination file: the first recovered record's
    // write-back errors.
    injector.arm(
        FaultRule::new(FaultOp::Write, Fault::Error(ErrorKind::Other)).on_path("blob_salvaged"),
    );

    let result = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    );
    assert!(
        result.is_err(),
        "a failed destination write must error the salvage",
    );
    assert!(
        !std::path::Path::new(&dest).exists(),
        "the partial destination is removed on a write failure",
    );
    Ok(())
}

#[test]
fn salvage_blob_file_drops_a_corrupt_record_and_keeps_the_rest() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let records: Vec<(&[u8], &[u8])> = vec![
        (b"k0", b"value-zero"),
        (b"k1", b"value-one"),
        (b"k2", b"value-two"),
        (b"k3", b"value-three"),
    ];
    build_blob(&source, &fs, &records)?;

    // Flip the last byte of the second record's value: the checksum (over
    // key + value) fails, but the frame header (lengths, magic) stays intact, so
    // the scanner reports a checksum mismatch and re-syncs at the next record.
    let Some(second_frame_end) = BlobScanner::new(&source, &*fs, 0)?
        .filter_map(Result::ok)
        .nth(1)
        .map(|e| e.frame_end)
    else {
        panic!("source blob must have at least two records");
    };
    let flip = usize::try_from(second_frame_end - 1).unwrap_or(0);
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(flip) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    // The corrupt k1 drops on its checksum, and the scanner then RESYNCS to the
    // next magic. k2's frame is reached by that byte scan through k1's damaged
    // bytes, so its boundary is UNPROVEN. The taint is sticky, so the walk STOPS at
    // k2 and reports the whole surrendered tail (k2 and k3 chained past it) as ONE
    // drop, rather than re-emitting a possibly-fabricated chain. Only k0, before
    // the corruption, is provable.
    assert_eq!(
        report.dropped.len(),
        2,
        "the corrupt record + the surrendered tail (recorded once) drop: {report:?}"
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(BlobDropReason::ChecksumMismatch)
        ),
        "the corrupt record reports a checksum mismatch: {report:?}",
    );
    assert_eq!(
        report
            .dropped
            .iter()
            .filter(
                |d| matches!(&d.reason, BlobDropReason::Corrupt(m) if m.contains("surrendered"))
            )
            .count(),
        1,
        "the surrendered tail after the resync is recorded ONCE: {report:?}",
    );
    assert_eq!(
        report.records_salvaged, 1,
        "only k0 is recovered; k1 (corrupt) and the surrendered tail (k2, k3) are not"
    );

    // The salvaged file holds only k0: k1 (corrupt) and the whole tainted tail
    // (k2, k3) after the resync are absent.
    let recovered = scan_blob(&dest, &fs)?;
    let keys: Vec<Vec<u8>> = recovered.iter().map(|(k, _)| k.clone()).collect();
    assert_eq!(
        keys,
        vec![b"k0".to_vec()],
        "only the provable frame before the corruption survives",
    );
    Ok(())
}

/// A blob file where EVERY record is corrupt salvages nothing: the report
/// carries only drops, `salvaged_path` is `None`, and the empty destination
/// placeholder the writer created is removed (a repair caller would otherwise
/// re-reject a stray zero-record blob file in its place).
#[test]
fn salvage_blob_file_removes_the_empty_dest_when_nothing_is_recoverable() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let records: Vec<(&[u8], &[u8])> = vec![(b"k0", b"value-zero"), (b"k1", b"value-one")];
    build_blob(&source, &fs, &records)?;

    // Flip the last value byte of BOTH records: each frame header stays
    // intact, so the scanner reports one checksum mismatch per record and
    // re-syncs — leaving zero salvageable records.
    let frame_ends: Vec<u64> = BlobScanner::new(&source, &*fs, 0)?
        .filter_map(Result::ok)
        .map(|e| e.frame_end)
        .collect();
    assert_eq!(frame_ends.len(), 2, "source blob holds two records");
    let mut bytes = std::fs::read(&source)?;
    for end in frame_ends {
        let flip = usize::try_from(end - 1).unwrap_or(0);
        if let Some(b) = bytes.get_mut(flip) {
            *b ^= 0xFF;
        }
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(report.records_salvaged, 0, "{report:?}");
    assert_eq!(report.dropped.len(), 2, "both records drop: {report:?}");
    assert_eq!(
        report.salvaged_path, None,
        "nothing recoverable yields no salvaged path",
    );
    assert!(
        fs.metadata(&dest).is_err(),
        "the empty destination placeholder is removed",
    );
    Ok(())
}

/// A STRUCTURAL failure mid-walk (a record frame whose magic bytes are gone,
/// not a checksum miss) terminates the blob walk: the scanner cannot re-sync
/// past it, so the salvage records one `Corrupt` drop for the unreadable tail
/// and keeps everything scanned before it.
#[test]
fn salvage_blob_file_stops_at_a_smashed_frame_and_keeps_the_prefix() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let records: Vec<(&[u8], &[u8])> = vec![
        (b"k0", b"value-zero"),
        (b"k1", b"value-one"),
        (b"k2", b"value-two"),
    ];
    build_blob(&source, &fs, &records)?;

    // Smash the LAST record's frame magic (the file structure and trailer
    // stay intact): the scanner reports it as a structural InvalidHeader it
    // cannot re-sync from, unlike a checksum miss.
    let Some(last_start) = BlobScanner::new(&source, &*fs, 0)?
        .filter_map(Result::ok)
        .nth(1)
        .map(|e| e.frame_end)
    else {
        panic!("source blob must have at least two records");
    };
    let mut bytes = std::fs::read(&source)?;
    let at = usize::try_from(last_start).unwrap_or(0);
    let Some(magic) = bytes.get_mut(at..at + 4) else {
        panic!("last record's frame magic within the file");
    };
    magic.copy_from_slice(b"????");
    std::fs::write(&source, &bytes)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(
        report.records_salvaged, 2,
        "the records before the smashed frame are recovered: {report:?}",
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(BlobDropReason::Corrupt(_))
        ),
        "the truncated tail is recorded as a structural drop: {report:?}",
    );
    let recovered = scan_blob(&dest, &fs)?;
    assert_eq!(recovered.len(), 2, "the salvaged copy holds the prefix");
    Ok(())
}

/// A blob frame whose header CRC and data checksum are internally consistent
/// but whose `key_len` is ZERO yields an Ok scanner entry with an empty key —
/// which the blob writer's ingest asserts against. Salvage must route such a
/// frame through the corrupt-record path (dropped, walk continues) instead of
/// panicking in the writer and leaving a partial destination behind.
#[test]
fn salvage_blob_file_drops_an_empty_key_frame() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    build_blob(&source, &fs, &[(b"k", b"vvvv"), (b"k2", b"second")])?;

    // Re-frame the FIRST record as `key_len = 0`: its key byte becomes the
    // first value byte (the hashed key||value byte span is unchanged), with
    // the header CRC and data checksum recomputed so the frame stays
    // internally consistent. V4 frame layout from offset 0:
    //   magic 4 | checksum 16 | seqno 8 | key_len 2 | real_val_len 4 |
    //   on_disk_val_len 4 | header_crc 4 | key | value.
    let mut bytes = std::fs::read(&source)?;
    let seqno = {
        let Some(b) = bytes.get(20..28) else {
            panic!("seqno within the first frame");
        };
        u64::from_le_bytes(b.try_into().unwrap_or([0; 8]))
    };
    // header_crc = truncated xxh3 over (seqno, key_len, real_val_len,
    // on_disk_val_len), matching the writer's framing.
    let new_hcrc = {
        let mut hasher = xxhash_rust::xxh3::Xxh3::default();
        hasher.update(&seqno.to_le_bytes());
        hasher.update(&0u16.to_le_bytes());
        hasher.update(&5u32.to_le_bytes());
        hasher.update(&5u32.to_le_bytes());
        #[expect(
            clippy::cast_possible_truncation,
            reason = "intentionally truncated to the 4-byte header CRC"
        )]
        {
            hasher.digest() as u32
        }
    };
    // data checksum = xxh3_128(key || value || header_crc_le); with the empty
    // key the hashed span is the same "kvvvv" bytes plus the NEW header CRC.
    let new_checksum = {
        let mut hasher = xxhash_rust::xxh3::Xxh3::default();
        hasher.update(b"kvvvv");
        hasher.update(&new_hcrc.to_le_bytes());
        hasher.digest128()
    };
    let patch = |bytes: &mut Vec<u8>, range: core::ops::Range<usize>, val: &[u8]| {
        let Some(slot) = bytes.get_mut(range) else {
            panic!("patch range within the first frame");
        };
        slot.copy_from_slice(val);
    };
    patch(&mut bytes, 4..20, &new_checksum.to_le_bytes());
    patch(&mut bytes, 28..30, &0u16.to_le_bytes());
    patch(&mut bytes, 30..34, &5u32.to_le_bytes());
    patch(&mut bytes, 34..38, &5u32.to_le_bytes());
    patch(&mut bytes, 38..42, &new_hcrc.to_le_bytes());
    std::fs::write(&source, &bytes)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(
        report.dropped.len(),
        1,
        "the empty-key frame drops as corrupt: {report:?}",
    );
    assert!(
        matches!(
            report.dropped.first().map(|d| &d.reason),
            Some(BlobDropReason::Corrupt(_))
        ),
        "the drop reason names the malformed frame: {report:?}",
    );
    assert_eq!(
        report.records_salvaged, 1,
        "the record after the malformed frame is still recovered",
    );
    let recovered = scan_blob(&dest, &fs)?;
    assert_eq!(
        recovered,
        vec![(b"k2".to_vec(), b"second".to_vec())],
        "the salvaged copy holds exactly the healthy record",
    );
    Ok(())
}

/// A blob frame whose header CRC and data checksum are internally consistent
/// but whose `real_val_len` field disagrees with the bytes actually stored
/// (`on_disk_val_len`, an uncompressed source) is REJECTED by the live blob
/// reader — salvage must not launder it: re-emitting through the writer would
/// restamp the frame with a consistent length and count a record the live
/// read path treats as corrupt as salvaged. It must drop through the
/// corrupt-record path instead.
#[test]
fn salvage_blob_file_drops_a_frame_with_a_forged_value_length() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    build_blob(&source, &fs, &[(b"k", b"vvvv"), (b"k2", b"second")])?;

    // Re-stamp the FIRST record's `real_val_len` (4 → 5) while the stored
    // bytes stay 4 (`on_disk_val_len` unchanged), with the header CRC and
    // data checksum recomputed so the frame stays internally consistent —
    // exactly the shape the scanner accepts and `Reader::get` rejects.
    // V4 frame layout from offset 0:
    //   magic 4 | checksum 16 | seqno 8 | key_len 2 | real_val_len 4 |
    //   on_disk_val_len 4 | header_crc 4 | key | value.
    let mut bytes = std::fs::read(&source)?;
    let seqno = {
        let Some(b) = bytes.get(20..28) else {
            panic!("seqno within the first frame");
        };
        u64::from_le_bytes(b.try_into().unwrap_or([0; 8]))
    };
    let new_hcrc = {
        let mut hasher = xxhash_rust::xxh3::Xxh3::default();
        hasher.update(&seqno.to_le_bytes());
        hasher.update(&1u16.to_le_bytes());
        hasher.update(&5u32.to_le_bytes());
        hasher.update(&4u32.to_le_bytes());
        #[expect(
            clippy::cast_possible_truncation,
            reason = "intentionally truncated to the 4-byte header CRC"
        )]
        {
            hasher.digest() as u32
        }
    };
    let new_checksum = {
        let mut hasher = xxhash_rust::xxh3::Xxh3::default();
        hasher.update(b"kvvvv");
        hasher.update(&new_hcrc.to_le_bytes());
        hasher.digest128()
    };
    let patch = |bytes: &mut Vec<u8>, range: core::ops::Range<usize>, val: &[u8]| {
        let Some(slot) = bytes.get_mut(range) else {
            panic!("patch range within the first frame");
        };
        slot.copy_from_slice(val);
    };
    patch(&mut bytes, 4..20, &new_checksum.to_le_bytes());
    patch(&mut bytes, 30..34, &5u32.to_le_bytes());
    patch(&mut bytes, 38..42, &new_hcrc.to_le_bytes());
    std::fs::write(&source, &bytes)?;

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(
        report.dropped.len(),
        1,
        "the forged-length frame drops as corrupt: {report:?}",
    );
    assert_eq!(
        report.records_salvaged, 1,
        "the record after the malformed frame is still recovered",
    );
    let recovered = scan_blob(&dest, &fs)?;
    assert_eq!(
        recovered,
        vec![(b"k2".to_vec(), b"second".to_vec())],
        "the salvaged copy holds exactly the healthy record",
    );
    Ok(())
}

/// The salvaged copy holds the source's values and keeps their lifetime class,
/// so a later relocation still knows how long they have lived.
#[test]
fn salvage_blob_file_keeps_the_lifetime_class() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    {
        let mut writer = BlobWriter::new(&source, 0, 0, &*fs)?;
        writer.set_lifetime_class(2, None);
        writer.write(b"k0", 0, b"a value that has lived a while")?;
        writer.finish()?;
    }

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(report.records_salvaged, 1);

    let handle = crate::vlog::recover_blob_file(
        &dest,
        0,
        crate::Checksum::from_raw(crate::repair::compute_table_checksum(&*fs, &dest)?),
        0,
        &fs,
        #[cfg(zstd_any)]
        &crate::compression::ZstdDictionaries::new(),
    )?;
    assert_eq!(handle.lifetime_class(), 2);
    Ok(())
}

/// A compressed blob source is rejected (fail-closed): the scanner yields on-disk
/// compressed bytes that this path cannot faithfully re-emit yet.
#[cfg(feature = "lz4")]
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn salvage_blob_file_recovers_a_compressed_source() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let value = b"some compressible value aaaaaaaaaaaaaaaa";

    {
        let mut writer =
            BlobWriter::new(&source, 0, 0, &*fs)?.use_compression(crate::CompressionType::Lz4);
        writer.write(b"k0", 0, value)?;
        writer.finish()?;
    }

    // A compressed source is salvaged by DECOMPRESSING each record (proving it
    // round-trips) and re-emitting it under the same compression descriptor —
    // never copied through verbatim, which would store undecodable bytes.
    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        None,
    )?;
    assert_eq!(report.records_salvaged, 1, "the record is recovered");
    assert!(
        report.is_complete(),
        "nothing dropped: {:?}",
        report.dropped
    );
    assert_eq!(report.salvaged_path.as_ref(), Some(&dest));

    // The recovered copy reads back through the live blob reader.
    let handle = crate::vlog::recover_blob_file(
        &dest,
        0,
        crate::Checksum::from_raw(crate::repair::compute_table_checksum(&*fs, &dest)?),
        0,
        &fs,
        // Lz4, so there is no dictionary for the handle to pin.
        #[cfg(zstd_any)]
        &crate::compression::ZstdDictionaries::new(),
    )?;
    assert_eq!(
        handle.compression(),
        crate::CompressionType::Lz4,
        "the copy keeps the source's compression descriptor",
    );
    // A live SST handle is rebuilt straight from the relocation: BOTH its
    // fields come from the salvaged file (the reader cross-checks the size
    // against the frame header, so the relocation's size must be the
    // re-emitted one).
    let (_, relocation) = *report.offset_remap.first().expect("one remap entry");
    let file = std::fs::File::open(&dest)?;
    let reader = crate::vlog::blob_file::reader::Reader::new(&handle, &file);
    let vhandle = crate::vlog::ValueHandle {
        blob_file_id: 0,
        offset: relocation.offset,
        on_disk_size: relocation.on_disk_size,
    };
    assert_eq!(
        reader.get(b"k0", &vhandle)?.as_ref(),
        value,
        "the salvaged record decompresses to its original value",
    );
    Ok(())
}

/// With the matching dictionary supplied (manifest repair passes the tree's
/// configured one), a dictionary-compressed source salvages like any other
/// compressed blob: intact records decompress, re-emit under the same
/// descriptor, and damaged records drop — instead of the whole file (and every
/// dependent SST) being set aside for want of dictionary context.
#[cfg(zstd_any)]
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn salvage_blob_file_recovers_a_dictionary_source_with_the_dictionary() -> crate::Result<()> {
    use std::io::{Seek, SeekFrom, Write};

    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let dict = Arc::new(crate::compression::ZstdDictionary::new(
        b"sample sample sample sample payload payload payload",
    ));
    let compression = crate::CompressionType::ZstdDict {
        level: 3,
        dict_id: dict.id(),
    };

    {
        let mut w = BlobWriter::new(&source, 0, 0, &*fs)?
            .use_compression(compression)
            .use_zstd_dictionary(Some(Arc::clone(&dict)));
        w.write(b"a", 1, b"sample payload sample payload sample payload")?;
        w.write(b"b", 2, b"sample payload sample payload sample sample")?;
        w.finish()?;
    }

    // Rot the SECOND record's payload (checksum fails, the record drops; the
    // first record is the intact survivor the dictionary must decode).
    let entries: Vec<_> =
        crate::vlog::BlobFileScanner::new(&source, &*fs, 0)?.collect::<crate::Result<Vec<_>>>()?;
    let second = entries.get(1).expect("two records");
    {
        let mut f = std::fs::OpenOptions::new().write(true).open(&source)?;
        f.seek(SeekFrom::Start(second.frame_end - 4))?;
        f.write_all(&[0xFF, 0xFF, 0xFF, 0xFF])?;
    }

    let report = salvage_blob_file(
        &source,
        dest.clone(),
        &fs,
        0,
        &default_comparator(),
        0,
        #[cfg(zstd_any)]
        Some(&dict),
    )?;
    assert_eq!(
        report.records_salvaged, 1,
        "the intact dictionary-compressed record is recovered: {report:?}",
    );
    assert_eq!(report.dropped.len(), 1, "the rotted record drops");

    // The re-emitted record round-trips under the SAME dictionary descriptor.
    let salvaged = crate::vlog::BlobFileScanner::new(&dest, &*fs, 0)?
        .next()
        .expect("one salvaged record")?;
    let value = super::decompress_blob_value(
        compression,
        &salvaged.value,
        salvaged.uncompressed_len as usize,
        #[cfg(zstd_any)]
        Some(&dict),
    )?;
    assert_eq!(
        value.as_ref(),
        b"sample payload sample payload sample payload",
        "the salvaged record decodes to its original value under the dictionary",
    );
    Ok(())
}

#[cfg(all(feature = "lz4", zstd_any))]
#[test]
fn salvage_blob_file_rejects_a_dictionary_compressed_source() -> crate::Result<()> {
    let dir = tempdir()?;
    let source = dir.path().join("blob_source");
    let dest = dir.path().join("blob_salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    // A dictionary descriptor with no dictionary context on this standalone
    // path: fail closed rather than re-emit undecodable values.
    {
        let mut writer =
            BlobWriter::new(&source, 0, 0, &*fs)?.use_compression(crate::CompressionType::Lz4);
        writer.metadata_compression_override = Some(crate::CompressionType::ZstdDict {
            level: 3,
            dict_id: 7,
        });
        writer.write(b"k0", 0, b"some compressible value aaaaaaaaaaaaaaaa")?;
        writer.finish()?;
    }

    assert!(
        matches!(
            salvage_blob_file(
                &source,
                dest,
                &fs,
                0,
                &default_comparator(),
                0,
                #[cfg(zstd_any)]
                None
            ),
            // The SAME error a wrong dictionary raises, with `got: None`:
            // both are a mis-supplied recovery context, and repair
            // propagates that class instead of grading the file damaged.
            Err(crate::Error::ZstdDictMismatch {
                expected: 7,
                got: None
            }),
        ),
        "a dictionary-compressed blob file must be rejected rather than mis-salvaged",
    );
    Ok(())
}

/// The blob-handle rewrite must install BOTH relocation fields into a remapped
/// indirection: the salvaged file is compacted (offset moves) AND its records
/// are re-compressed (the on-disk size may change — compressor output is not
/// stable across versions), while `Reader::get` cross-checks the handle's size
/// against the frame header and rejects a mismatch. Carrying only the offset
/// would leave otherwise-salvaged values unreadable.
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn blob_handle_rewrite_installs_the_relocated_size() -> crate::Result<()> {
    use crate::blob_tree::handle::BlobIndirection;
    use crate::coding::{Decode, Encode};
    use crate::vlog::ValueHandle;
    use crate::{InternalValue, ValueType};

    let ind = BlobIndirection {
        vhandle: ValueHandle {
            blob_file_id: 7,
            offset: 400,
            on_disk_size: 64,
        },
        size: 100,
    };
    let mut value = Vec::new();
    ind.encode_into(&mut value)?;
    let entries = vec![InternalValue::from_components(
        b"k".to_vec(),
        value,
        1,
        ValueType::Indirection,
    )];

    // The salvaged replacement re-emitted this record at a new offset AND a
    // new (re-compressed) on-disk size.
    let mut map = crate::HashMap::default();
    map.insert(
        400u64,
        super::BlobRecordRelocation {
            offset: 16,
            on_disk_size: 61,
        },
    );
    let mut rewrite = crate::HashMap::default();
    // The replacement is a FRESH blob file, so the handle must be retargeted
    // at its id as well as at the new offset.
    rewrite.insert(
        7u64,
        super::BlobFileRewrite::Remap {
            new_id: 9,
            offsets: map,
        },
    );

    let mut dropped = 0u64;
    let (out, carry) = super::rewrite_block_indirections(
        entries,
        super::RowRewrites::of_blobs(&rewrite),
        &mut dropped,
    )?;
    assert_eq!(dropped, 0, "the record survived, nothing drops");
    assert!(carry.is_none(), "nothing was beheaded, nothing to suppress");
    let entry = out.first().expect("one rewritten entry");
    let rewritten = BlobIndirection::decode_from(&mut &entry.value[..])?;
    assert_eq!(
        rewritten.vhandle.blob_file_id, 9,
        "the handle names the salvaged replacement, not the damaged original",
    );
    assert_eq!(rewritten.vhandle.offset, 16, "the offset is re-targeted");
    assert_eq!(
        rewritten.vhandle.on_disk_size, 61,
        "the on-disk size is the RE-EMITTED record's, not the source's: the \
         reader rejects a handle whose size disagrees with the frame header",
    );
    Ok(())
}

/// Losing a blob record removes the HEAD of a key's version chain, so the older
/// versions behind it must go too: keeping them republishes an overwritten value
/// as current, or undoes a newer write by exposing an older tombstone.
#[test]
fn blob_handle_rewrite_drops_older_versions_when_the_head_record_is_lost() -> crate::Result<()> {
    use crate::blob_tree::handle::BlobIndirection;
    use crate::coding::Encode;
    use crate::vlog::ValueHandle;
    use crate::{InternalValue, ValueType};

    let indirection = |offset: u64| -> crate::Result<Vec<u8>> {
        let mut value = Vec::new();
        BlobIndirection {
            vhandle: ValueHandle {
                blob_file_id: 7,
                offset,
                on_disk_size: 64,
            },
            size: 100,
        }
        .encode_into(&mut value)?;
        Ok(value)
    };

    // The salvaged replacement re-emitted NOTHING of the original: every
    // handle into file 7 has lost its record.
    let mut rewrite = crate::HashMap::default();
    rewrite.insert(
        7u64,
        super::BlobFileRewrite::Remap {
            new_id: 9,
            offsets: crate::HashMap::default(),
        },
    );

    // `k`'s newest version is the lost indirection; its older INLINE value sits
    // right behind it, and `z` ends the run.
    let entries = vec![
        InternalValue::from_components(
            b"k".to_vec(),
            indirection(400)?,
            10,
            ValueType::Indirection,
        ),
        InternalValue::from_components(b"k".to_vec(), b"old".to_vec(), 5, ValueType::Value),
        InternalValue::from_components(b"z".to_vec(), b"v".to_vec(), 1, ValueType::Value),
    ];
    let mut dropped = 0u64;
    let (out, carry) = super::rewrite_block_indirections(
        entries,
        super::RowRewrites::of_blobs(&rewrite),
        &mut dropped,
    )?;
    let keys: Vec<_> = out.iter().map(|e| e.key.user_key.to_vec()).collect();
    assert_eq!(
        keys,
        vec![b"z".to_vec()],
        "the beheaded key goes entirely; the next key is untouched",
    );
    assert_eq!(dropped, 2, "both the lost head and the version behind it");
    assert!(
        carry.is_none(),
        "a surviving entry with a different key ended the run",
    );

    // The same chain, but the block ENDS inside it: the suppression has to
    // continue into the next block.
    let entries = vec![
        InternalValue::from_components(
            b"k".to_vec(),
            indirection(400)?,
            10,
            ValueType::Indirection,
        ),
        InternalValue::from_components(b"k".to_vec(), b"old".to_vec(), 5, ValueType::Value),
    ];
    let mut dropped = 0u64;
    let (out, carry) = super::rewrite_block_indirections(
        entries,
        super::RowRewrites::of_blobs(&rewrite),
        &mut dropped,
    )?;
    assert!(out.is_empty(), "the whole block was one beheaded key");
    assert_eq!(
        carry.as_deref(),
        Some(b"k".as_slice()),
        "nothing proved the run ended, so the key keeps suppressing",
    );
    Ok(())
}

/// A cell row's references take a blob-file rewrite the way an indirection
/// does: a reference into a remapped file names the re-emitted record, one
/// into an untouched file and a value cell stay as written, and a row whose
/// referenced record is gone loses its key's head, taking the older versions
/// behind it.
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn blob_handle_rewrite_remaps_and_beheads_cell_rows() -> crate::Result<()> {
    use crate::blob_tree::field_row::{RowCell, RowField, decode_row, encode_row};
    use crate::blob_tree::handle::BlobIndirection;
    use crate::vlog::ValueHandle;
    use crate::{InternalValue, ValueType};

    let reference = |blob_file_id: u64, offset: u64, owner: bool| RowCell::Ref {
        indirection: BlobIndirection {
            vhandle: ValueHandle {
                blob_file_id,
                offset,
                on_disk_size: 64,
            },
            size: 100,
        },
        owner,
    };
    let row = |cells: Vec<RowCell<'_>>| -> crate::Result<Vec<u8>> {
        let fields: Vec<RowField<'_>> = cells
            .into_iter()
            .zip(3u16..)
            .map(|(cell, column)| RowField::bytes(column, cell))
            .collect();
        encode_row(&fields)
    };
    let cell_row = |key: &[u8], seqno, value: Vec<u8>| {
        InternalValue::from_components(key.to_vec(), value, seqno, ValueType::CellRow)
    };

    let mut offsets = crate::HashMap::default();
    offsets.insert(
        400u64,
        super::BlobRecordRelocation {
            offset: 16,
            on_disk_size: 61,
        },
    );
    let mut rewrite = crate::HashMap::default();
    rewrite.insert(7u64, super::BlobFileRewrite::Remap { new_id: 9, offsets });
    rewrite.insert(8u64, super::BlobFileRewrite::DropBelow(500));

    let entries = vec![
        // Remapped, beside an untouched file and a value cell.
        cell_row(
            b"a",
            3,
            row(vec![
                reference(7, 400, true),
                reference(5, 10, false),
                RowCell::Value(b"inline"),
            ])?,
        ),
        // Untouched entirely.
        cell_row(b"b", 3, row(vec![reference(5, 20, true)])?),
        // Its record is gone from the remapped file: the key's head is lost,
        // and its older version goes with it.
        cell_row(b"c", 4, row(vec![reference(7, 999, true)])?),
        InternalValue::from_components(b"c".to_vec(), b"old".to_vec(), 2, ValueType::Value),
        // Below the punched frontier: lost the same way.
        cell_row(b"d", 3, row(vec![reference(8, 100, false)])?),
        // At or past the frontier: kept.
        cell_row(b"e", 3, row(vec![reference(8, 600, false)])?),
    ];
    let untouched = entries.get(1).expect("row b").value.clone();
    let mut dropped = 0u64;
    let (out, carry) = super::rewrite_block_indirections(
        entries,
        super::RowRewrites::of_blobs(&rewrite),
        &mut dropped,
    )?;
    let keys: Vec<_> = out.iter().map(|e| e.key.user_key.to_vec()).collect();
    assert_eq!(keys, [b"a".to_vec(), b"b".to_vec(), b"e".to_vec()]);
    assert_eq!(dropped, 3, "the two lost heads and the version behind one");
    assert!(carry.is_none(), "the last entry is a kept key");

    let a = decode_row(&out.first().expect("row a").value)?;
    let field = |i: usize| a.get(i).expect("a field of row a");
    let RowCell::Ref { indirection, owner } = field(0).cell else {
        panic!("the first field stays a reference");
    };
    assert_eq!(
        (
            indirection.vhandle.blob_file_id,
            indirection.vhandle.offset,
            indirection.vhandle.on_disk_size,
            owner
        ),
        (9, 16, 61, true),
        "remapped to the re-emitted record, ownership kept"
    );
    assert!(
        matches!(field(1).cell, RowCell::Ref { indirection, .. } if indirection.vhandle.blob_file_id == 5)
    );
    assert!(matches!(field(2).cell, RowCell::Value(b"inline")));
    let b = out.get(1).expect("row b");
    assert_eq!(b.value, untouched, "an untouched row is copied as is");
    assert_eq!(b.key.value_type, ValueType::CellRow);
    Ok(())
}

/// A recovered block's references include a cell row's, each with its owner
/// bit: an owned reference adds its object to the file's link and to the
/// copy's owned objects, a borrowed one only links the file.
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn recovered_cell_row_references_link_their_files_by_ownership() -> crate::Result<()> {
    use crate::blob_tree::field_row::{RowCell, RowField, encode_row};
    use crate::blob_tree::handle::BlobIndirection;
    use crate::vlog::ValueHandle;
    use crate::{InternalValue, ValueType};

    let reference = |blob_file_id: u64, owner: bool| RowCell::Ref {
        indirection: BlobIndirection {
            vhandle: ValueHandle {
                blob_file_id,
                offset: 0,
                on_disk_size: 64,
            },
            size: 100,
        },
        owner,
    };
    let value = encode_row(&[
        RowField::bytes(3, reference(7, true)),
        RowField::bytes(4, reference(8, false)),
    ])?;
    let entries = vec![InternalValue::from_components(
        b"k".to_vec(),
        value,
        1,
        ValueType::CellRow,
    )];
    let refs = super::collect_indirections(&entries)?;
    assert_eq!(
        refs.iter()
            .map(|(_, ind, holding)| (ind.vhandle.blob_file_id, holding.owns()))
            .collect::<Vec<_>>(),
        [(7, true), (8, false)]
    );

    let mut derived = crate::HashMap::default();
    let mut owned_cells = Vec::new();
    super::fold_blob_links(&mut derived, &mut owned_cells, &refs);
    assert_eq!(
        owned_cells,
        [(7, 0)],
        "only the owned cell object is listed"
    );
    let owned = derived.get(&7).expect("the owned file is linked");
    assert_eq!((owned.len, owned.bytes, owned.on_disk_bytes), (1, 100, 64));
    let borrowed = derived.get(&8).expect("the borrowed file is linked");
    assert_eq!(
        (borrowed.len, borrowed.bytes, borrowed.on_disk_bytes),
        (0, 0, 0),
        "a borrowed reference charges nothing to its file"
    );
    Ok(())
}

/// A columnar source carrying a per-field value sub-column salvages into a copy
/// that KEEPS the sub-column (verbatim `ColumnBatch` re-emit), instead of
/// collapsing it into a single value column via a row round-trip.
#[cfg(feature = "columnar")]
#[test]
fn salvage_preserves_columnar_value_subcolumns() -> crate::Result<()> {
    use crate::table::columnar::{Column, TypeTag, entries_to_column_batch};

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let cmp = default_comparator();

    // Two columnar blocks whose value is a single fixed-4 sub-column (id 3),
    // written verbatim through the ingest batch path (per-row seqno 0).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true);
    for block in 0..2u32 {
        let entries: Vec<InternalValue> = (0..4u32)
            .map(|i| {
                let k = format!("k{:04}", block * 4 + i);
                InternalValue::from_components(k.into_bytes(), b"x".to_vec(), 0, ValueType::Value)
            })
            .collect();
        let mut batch = entries_to_column_batch(&entries)?;
        batch.columns.pop();
        let mut data = Vec::new();
        for i in 0..4u32 {
            data.extend_from_slice(&(block * 4 + i).to_le_bytes());
        }
        batch.columns.push(Column {
            column_id: 3,
            type_tag: TypeTag::Fixed(4),
            validity: None,
            data: data.into(),
        });
        writer.write_columnar_batch(&batch, &cmp)?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "a healthy columnar SST drops nothing: {report:?}"
    );
    // No deletes + clean blocks: each columnar block is copied through verbatim,
    // which is exactly why the per-field sub-columns survive byte-for-byte.
    assert_eq!(
        report.blocks_copied_verbatim, report.blocks_salvaged,
        "clean columnar blocks are copied verbatim",
    );

    // Reopen and project sub-column 3 via the per-SST scan: it survives as a
    // sub-column. A row round-trip would have collapsed it into the value column.
    let recovered = open(dest, &fs)?;
    assert!(
        recovered.metadata.columnar,
        "the recovered copy stays columnar"
    );
    let batches = recovered
        .columnar_scan(&[3], None)?
        .collect::<crate::Result<Vec<_>>>()?;
    let rows: u32 = batches.iter().map(|b| b.row_count).sum();
    assert_eq!(rows, 8, "every row's sub-column is recovered");
    assert!(
        batches
            .iter()
            .all(|b| b.columns.iter().all(|c| c.column_id == 3)),
        "the value sub-column (id 3) is preserved verbatim, not collapsed",
    );
    Ok(())
}

/// A columnar Page-ECC SST with a single-byte RS-recoverable fault in a data
/// block (no deletes): salvage recovers the block from parity and **re-encodes**
/// the healed batch rather than copying the faulty on-disk bytes verbatim, so the
/// recovered copy carries clean bytes. The clean block around it is still copied
/// verbatim.
#[cfg(all(feature = "columnar", feature = "page_ecc"))]
#[test]
fn salvage_reencodes_an_ecc_recovered_columnar_block() -> crate::Result<()> {
    use crate::table::block::{EccParams, Header};
    use crate::table::columnar::entries_to_column_batch;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let cmp = default_comparator();

    // Two columnar blocks under RS(4,2) parity, no deletes (so the no-deletes
    // copy-through / recover path is taken, not the delete-masked one).
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_zone_map(true)
        .use_ecc(Some(EccParams::RS_4_2));
    for block in 0..2u32 {
        let entries: Vec<InternalValue> = (0..4u32)
            .map(|i| {
                let k = format!("k{:04}", block * 4 + i);
                InternalValue::from_components(k.into_bytes(), b"x".to_vec(), 0, ValueType::Value)
            })
            .collect();
        let batch = entries_to_column_batch(&entries)?;
        writer.write_columnar_batch(&batch, &cmp)?;
    }
    assert!(
        writer.finish()?.is_some(),
        "source columnar SST is non-empty"
    );

    // Flip one byte of the first columnar data block (RS(4,2) recovers a single
    // byte error).
    let first_off = {
        let table = open(source.clone(), &fs)?;
        let Some(kh) = table.data_block_handles().find_map(Result::ok) else {
            panic!("a non-empty SST has at least one data block");
        };
        usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX)
    };
    let pos = first_off + Header::MIN_LEN + 3;
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(pos) {
        *b ^= 0x80;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert!(
        report.is_complete(),
        "an RS-recoverable columnar block is healed, not dropped: {report:?}",
    );
    assert_eq!(
        report.blocks_salvaged, report.blocks_total,
        "every block is recovered",
    );
    // The recovered block was re-encoded (verbatim:None), so fewer verbatim copies
    // than salvaged blocks; the other (clean) block is copied verbatim.
    assert!(
        report.blocks_copied_verbatim < report.blocks_salvaged,
        "the ECC-recovered columnar block is re-encoded, not copied verbatim: {report:?}",
    );

    let recovered = open(dest, &fs)?;
    assert!(
        recovered.metadata.columnar,
        "the recovered copy stays columnar"
    );
    assert_eq!(recovered.metadata.item_count, 8, "every row is recovered");
    Ok(())
}

/// A salvage whose output mixes copied and re-encoded row groups after a lost
/// one. A copy keeps its pages' group tag, so the copy's tags no longer match
/// its ordinals, and a re-encoded group must take a tag no copy already holds:
/// two groups sharing a tag would let their pages pass for each other's. Here
/// group 0 is lost, group 1 is re-encoded because the loss suppresses its
/// boundary key, group 2 is copied with its tag, and group 3 is healed from
/// parity and re-encoded after it. Group 1, the copy's first, takes the tag it
/// replaces, so group 2 can still be copied above it; group 3 takes the next
/// tag above group 2's.
#[cfg(all(feature = "columnar", feature = "page_ecc"))]
#[test]
fn salvage_keeps_row_group_tags_unique_when_copies_and_reencodes_mix() -> crate::Result<()> {
    use crate::table::block::EccParams;
    use crate::table::columnar::entries_to_column_batch;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let dest = dir.path().join("salvaged");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let cmp = default_comparator();
    let key = |i: u32| format!("k{i:04}").into_bytes();

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_ecc(Some(EccParams::RS_4_2));
    for group in 0..4u32 {
        let entries: Vec<InternalValue> = (0..4u32)
            .map(|i| {
                InternalValue::from_components(
                    key(group * 4 + i),
                    b"x".to_vec(),
                    0,
                    ValueType::Value,
                )
            })
            .collect();
        writer.write_columnar_batch(&entries_to_column_batch(&entries)?, &cmp)?;
    }
    assert!(writer.finish()?.is_some(), "source is non-empty");

    let offsets: Vec<usize> = {
        let table = open(source.clone(), &fs)?;
        table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| usize::try_from(*kh.as_ref().offset()).unwrap_or(usize::MAX))
            .collect()
    };
    let [lost, lost_end, _, healed] = offsets[..] else {
        panic!("source has four row groups, got {offsets:?}");
    };
    let mut bytes = std::fs::read(&source)?;
    // Past what RS(4,2) corrects: every byte of the lost group after its
    // directory's header, parity included.
    let header_len = crate::table::block::Header::MIN_LEN;
    let Some(lost_bytes) = bytes.get_mut(lost + header_len..lost_end) else {
        panic!("lost group within the file");
    };
    for b in lost_bytes {
        *b ^= 0xFF;
    }
    // One byte, which parity restores: the healed group is re-encoded.
    let Some(b) = bytes.get_mut(healed + header_len + 3) else {
        panic!("healed group within the file");
    };
    *b ^= 0x80;
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "only the first group is lost: {report:?}"
    );
    assert_eq!(report.blocks_salvaged, 3, "{report:?}");
    assert_eq!(
        report.blocks_copied_verbatim, 1,
        "the clean group keeps its bytes, so the tag rule did not force a re-encode: {report:?}",
    );

    let recovered = open(dest, &fs)?;
    let tags: Vec<u64> = recovered
        .data_block_handles()
        .filter_map(Result::ok)
        .map(|kh| {
            recovered
                .salvage_load_row_group(kh.as_ref())
                .map(|g| g.group.directory.group_tag())
        })
        .collect::<crate::Result<_>>()?;
    let distinct: crate::HashSet<u64> = tags.iter().copied().collect();
    assert_eq!(
        distinct.len(),
        tags.len(),
        "row group tags must stay unique in the copy, got {tags:?}",
    );
    // Row 4 is the suppressed boundary key; every row after it survives.
    for i in 5..16u32 {
        let k = key(i);
        assert!(
            recovered
                .get(&k, crate::MAX_SEQNO, crate::hash::hash64(&k))?
                .is_some(),
            "row {i} of a recovered group must read back",
        );
    }
    Ok(())
}

/// Salvage decides key identity the way the rest of the engine does, so a
/// version whose bytes differ is a DIFFERENT key and survives the loss of its
/// neighbour. That is not a shortcut for ordering: a comparator that called two
/// different spellings equal would break point lookups outright — filters and
/// the locator hash the raw bytes — which is why the trait's contract forbids
/// it, and why every path can share one relation.
#[test]
fn salvage_suppresses_only_the_shadowed_key_itself() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // One entry per block: a deletion for `K` in block 1, then a DIFFERENT key
    // `k` in block 2 (a byte-distinct key, hence a distinct key), then `z`.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(1);
    writer.write(InternalValue::new_tombstone(b"K".as_slice(), 10))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"old",
        5,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"z",
        b"v",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Corrupt the FIRST block: the deletion of `K` is lost.
    let first_off = {
        let table = open_with_id(source.clone(), &fs, 0)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert!(
            offsets.len() >= 3,
            "one block per entry, got {}",
            offsets.len()
        );
        usize::try_from(offsets.first().copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(first_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.dropped.len(), 1, "one block is lost: {report:?}");

    let recovered = open_with_id(dest, &fs, 0)?;
    assert!(
        recovered
            .get(b"k", crate::SeqNo::MAX, crate::hash::hash64(b"k"))?
            .is_some(),
        "`k` is a different key from the lost `K`: its own newest version is \
         intact and must survive",
    );
    Ok(())
}

/// The blob-handle rewrite suppresses a beheaded chain the same way, and asks
/// the same question: losing a key's newest record takes every OLDER version of
/// THAT key with it, and leaves a different key alone.
#[test]
fn the_blob_rewrite_suppresses_only_the_beheaded_key() -> crate::Result<()> {
    use crate::coding::Encode;
    use crate::{InternalValue, ValueType};

    let indirection = |offset: u64| -> crate::Result<Vec<u8>> {
        let ind = crate::blob_tree::handle::BlobIndirection {
            vhandle: crate::vlog::ValueHandle {
                blob_file_id: 7,
                offset,
                on_disk_size: 16,
            },
            size: 16,
        };
        let mut buf = Vec::new();
        ind.encode_into(&mut buf)?;
        Ok(buf)
    };

    // `a` (newest) points at a record the salvage LOST, and its own older
    // version follows; `z` is a different key whose record survived.
    let entries = vec![
        InternalValue::from_components(b"a", indirection(100)?, 10, ValueType::Indirection),
        InternalValue::from_components(b"a", indirection(200)?, 5, ValueType::Indirection),
        InternalValue::from_components(b"z", indirection(200)?, 1, ValueType::Indirection),
    ];

    // The remap holds only offset 200: the record behind `a`'s newest is gone.
    let mut offsets = crate::HashMap::default();
    offsets.insert(
        200u64,
        super::BlobRecordRelocation {
            offset: 300,
            on_disk_size: 16,
        },
    );
    let mut rewrite = crate::HashMap::default();
    rewrite.insert(7u64, super::BlobFileRewrite::Remap { new_id: 9, offsets });

    let mut dropped = 0u64;
    let (kept, carry) = super::rewrite_block_indirections(
        entries,
        super::RowRewrites::of_blobs(&rewrite),
        &mut dropped,
    )?;

    let keys: Vec<_> = kept.iter().map(|e| e.key.user_key.to_vec()).collect();
    assert_eq!(
        keys,
        vec![b"z".to_vec()],
        "the beheaded key's older version goes with it; a different key stays",
    );
    assert_eq!(
        dropped, 2,
        "both the lost head and its orphaned older version"
    );
    assert!(
        carry.is_none(),
        "a surviving different key ended the run inside this block",
    );
    Ok(())
}

/// A columnar table of a blob tree's cell rows takes a blob-file rewrite as
/// it lays its rows out: a reference into the remapped file names the
/// re-emitted record in the copy, which keeps the cells layout, and a row whose
/// record is gone is dropped.
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn salvage_remaps_the_references_of_a_columnar_cells_table() -> crate::Result<()> {
    use crate::blob_tree::field_row::{RowCell, RowField, decode_row, encode_row};
    use crate::blob_tree::handle::BlobIndirection;
    use crate::table::Writer;
    use crate::vlog::ValueHandle;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    let row = |offset: u64| -> crate::Result<Vec<u8>> {
        encode_row(&[
            RowField::bytes(3, RowCell::Value(b"status")),
            RowField::bytes(
                4,
                RowCell::Ref {
                    indirection: BlobIndirection {
                        vhandle: ValueHandle {
                            blob_file_id: 7,
                            offset,
                            on_disk_size: 64,
                        },
                        size: 100,
                    },
                    owner: true,
                },
            ),
        ])
    };
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_cell_rows(true)
        .use_row_group_size(1);
    for (key, offset) in [(b"a", 400u64), (b"b", 999), (b"c", 500)] {
        writer.write(InternalValue::from_components(
            key.as_slice(),
            row(offset)?,
            1,
            ValueType::CellRow,
        ))?;
    }
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let mut offsets = crate::HashMap::default();
    for (from, to) in [(400u64, 16u64), (500, 80)] {
        offsets.insert(
            from,
            super::BlobRecordRelocation {
                offset: to,
                on_disk_size: 61,
            },
        );
    }
    let mut rewrite = crate::HashMap::default();
    rewrite.insert(7u64, super::BlobFileRewrite::Remap { new_id: 9, offsets });
    let options = SalvageOptions {
        encryption: None,
        #[cfg(zstd_any)]
        zstd_dictionary: None,
        #[cfg(zstd_any)]
        zstd_dictionaries: crate::compression::ZstdDictionaries::new(),
        table_id: 0,
        expected_stored_id: None,
        output_id: None,
        allow_delete_resurrection: false,
        sync_mode: crate::fs::SyncMode::Normal,
        prefix_extractor: None,
        blob_rewrite: Some(Arc::new(rewrite)),
        owner_promotions: None,
        progress: None,
        clock: None,
    };
    salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;

    let recovered = open(dest, &fs)?;
    let field_ref = |key: &[u8]| -> crate::Result<Option<(u64, u64)>> {
        let Some(entry) = recovered.get(key, crate::SeqNo::MAX, crate::hash::hash64(key))? else {
            return Ok(None);
        };
        assert_eq!(entry.key.value_type, ValueType::CellRow);
        let fields = decode_row(&entry.value)?;
        let RowCell::Ref { indirection, .. } = fields.get(1).expect("the body field").cell else {
            panic!("the body stays a reference");
        };
        Ok(Some((
            indirection.vhandle.blob_file_id,
            indirection.vhandle.offset,
        )))
    };
    assert_eq!(field_ref(b"a")?, Some((9, 16)));
    assert_eq!(field_ref(b"b")?, None, "its record is gone, so is the row");
    assert_eq!(field_ref(b"c")?, Some((9, 80)));
    Ok(())
}

/// The same rewrite over a cells table whose delete bitmap masks a row: the
/// masked row stays gone, the live ones are remapped, and a row whose record
/// is gone is dropped, all through the masked columnar arm.
#[cfg(feature = "columnar")]
#[test]
#[expect(clippy::expect_used, reason = "test code")]
fn salvage_remaps_the_references_of_a_delete_masked_columnar_cells_table() -> crate::Result<()> {
    use crate::blob_tree::field_row::{RowCell, RowField, decode_row, encode_row};
    use crate::blob_tree::handle::BlobIndirection;
    use crate::config::DeleteStrategy;
    use crate::table::Writer;
    use crate::vlog::ValueHandle;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    let row = |offset: u64| -> crate::Result<Vec<u8>> {
        encode_row(&[
            RowField::bytes(3, RowCell::Value(b"status")),
            RowField::bytes(
                4,
                RowCell::Ref {
                    indirection: BlobIndirection {
                        vhandle: ValueHandle {
                            blob_file_id: 7,
                            offset,
                            on_disk_size: 64,
                        },
                        size: 100,
                    },
                    owner: true,
                },
            ),
        ])
    };
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_cell_rows(true)
        .use_zone_map(true)
        .use_row_group_size(4_096)
        .delete_strategy(DeleteStrategy::MergeOnRead);
    for (key, offset) in [(b"a", 400u64), (b"b", 450), (b"c", 500), (b"d", 999)] {
        writer.write(InternalValue::from_components(
            key.as_slice(),
            row(offset)?,
            1,
            ValueType::CellRow,
        ))?;
    }
    // Row position 1 is "b".
    writer.delete_bitmap_mut().insert(1);
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let mut offsets = crate::HashMap::default();
    for (from, to) in [(400u64, 16u64), (450, 48), (500, 80)] {
        offsets.insert(
            from,
            super::BlobRecordRelocation {
                offset: to,
                on_disk_size: 61,
            },
        );
    }
    let mut rewrite = crate::HashMap::default();
    rewrite.insert(7u64, super::BlobFileRewrite::Remap { new_id: 9, offsets });
    let options = SalvageOptions {
        blob_rewrite: Some(Arc::new(rewrite)),
        ..SalvageOptions::default()
    };
    salvage_sst_with_options(&source, dest.clone(), &fs, &options)?;

    let recovered = open(dest, &fs)?;
    let field_ref = |key: &[u8]| -> crate::Result<Option<(u64, u64)>> {
        let Some(entry) = recovered.get(key, crate::SeqNo::MAX, crate::hash::hash64(key))? else {
            return Ok(None);
        };
        assert_eq!(entry.key.value_type, ValueType::CellRow);
        let fields = decode_row(&entry.value)?;
        let RowCell::Ref { indirection, .. } = fields.get(1).expect("the body field").cell else {
            panic!("the body stays a reference");
        };
        Ok(Some((
            indirection.vhandle.blob_file_id,
            indirection.vhandle.offset,
        )))
    };
    assert_eq!(field_ref(b"a")?, Some((9, 16)));
    assert_eq!(field_ref(b"b")?, None, "the masked row stays deleted");
    assert_eq!(field_ref(b"c")?, Some((9, 80)));
    assert_eq!(field_ref(b"d")?, None, "its record is gone, so is the row");
    Ok(())
}

#[test]
fn salvage_keeps_tied_seqno_merge_operands() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // One write batch adds several merge operands for a key, so they share the
    // batch's seqno and a flush stores every one of them. The pair below sits in
    // ONE block; `z` gets its own so a corruption elsewhere can trigger salvage.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(100);
    writer.write(InternalValue::new_merge_operand(
        b"k".as_slice(),
        [b'A'; 64],
        10,
    ))?;
    writer.write(InternalValue::new_merge_operand(
        b"k".as_slice(),
        [b'B'; 64],
        10,
    ))?;
    writer.write(InternalValue::from_components(
        b"z",
        [b'z'; 128],
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Corrupt the LAST block so salvage runs while the operand block is clean.
    let last_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert!(
            offsets.len() >= 2,
            "the fixture needs the operands and `z` in separate blocks, got {}",
            offsets.len(),
        );
        usize::try_from(offsets.last().copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(last_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "only `z`'s block is lost: {report:?}"
    );

    let recovered = open(dest, &fs)?;
    let operands: Vec<_> = recovered
        .scan()?
        .filter_map(Result::ok)
        .filter(|e| e.key.user_key.as_ref() == b"k")
        .map(|e| e.value.to_vec())
        .collect();
    assert_eq!(
        operands.len(),
        2,
        "both operands are valid input to the merge, so a clean block holding \
         them must not be rejected as out of order: {operands:?}",
    );
    Ok(())
}

/// The same run can straddle a block boundary — the batch's operands do not have
/// to land in one block — so the cross-edge check has to accept the tie too.
#[test]
fn salvage_keeps_tied_seqno_merge_operands_across_a_block_boundary() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // One entry per block, so the two operands sit on either side of an edge.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(1);
    writer.write(InternalValue::new_merge_operand(b"k".as_slice(), b"A", 10))?;
    writer.write(InternalValue::new_merge_operand(b"k".as_slice(), b"B", 10))?;
    writer.write(InternalValue::from_components(
        b"z",
        b"v",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let last_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert_eq!(offsets.len(), 3, "one block per entry");
        usize::try_from(offsets.last().copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(last_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(
        report.dropped.len(),
        1,
        "only `z`'s block is lost: {report:?}"
    );

    let recovered = open(dest, &fs)?;
    let operands = recovered
        .scan()?
        .filter_map(Result::ok)
        .filter(|e| e.key.user_key.as_ref() == b"k")
        .count();
    assert_eq!(
        operands, 2,
        "the tie spanning the block edge is a valid run, not a violation",
    );
    Ok(())
}

/// A key's MVCC versions are contiguous, so they can straddle a block
/// boundary: the newest can be the last entry of one block and an older one the
/// first entry of the next. When the block holding the newest version is
/// dropped, emitting the next block as-is republishes the older version as
/// current — a delete resurrects, or an overwritten value comes back. The
/// boundary key's entries must be dropped with it unless resurrection is on.
#[test]
fn salvage_drops_the_boundary_key_when_its_newest_version_is_lost() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // One entry per block, so `k`'s newest version (a deletion) is the whole of
    // block 1 and its older value the whole of block 2.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(1);
    writer.write(InternalValue::from_components(
        b"a",
        b"v",
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::new_tombstone(b"k".as_slice(), 10))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"old",
        5,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"z",
        b"v",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Corrupt the block whose last key is `k` and whose seqno is the newest:
    // the second data block in offset order.
    let second_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert!(
            offsets.len() >= 3,
            "the fixture needs one block per entry, got {}",
            offsets.len(),
        );
        usize::try_from(offsets.get(1).copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(second_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.dropped.len(), 1, "one block is lost: {report:?}");

    let recovered = open(dest, &fs)?;
    assert!(
        recovered
            .get(b"k", crate::SeqNo::MAX, crate::hash::hash64(b"k"))?
            .is_none(),
        "the older version of the boundary key must not be republished as \
         current when the newest version was lost",
    );
    assert!(
        recovered
            .get(b"a", crate::SeqNo::MAX, crate::hash::hash64(b"a"))?
            .is_some(),
        "unaffected keys are still recovered",
    );
    Ok(())
}

/// A version chain can span MORE than one surviving block. Clearing the
/// suppression on the first block after the drop lets the second one publish an
/// even older version of the same key, so the boundary has to hold until a
/// surviving entry with a different key proves the run ended.
#[test]
fn salvage_keeps_suppressing_a_boundary_key_across_a_whole_shadowed_block() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // One entry per block: `k`'s newest version (a deletion) is block 1, and its
    // two older values are blocks 2 and 3 — a chain of whole shadowed blocks.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(1);
    writer.write(InternalValue::new_tombstone(b"k".as_slice(), 10))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"mid",
        5,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"old",
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"z",
        b"v",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Corrupt the FIRST data block: the one holding `k`'s newest version.
    let first_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert!(
            offsets.len() >= 4,
            "the fixture needs one block per entry, got {}",
            offsets.len(),
        );
        usize::try_from(offsets.first().copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(first_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.dropped.len(), 1, "one block is lost: {report:?}");

    let recovered = open(dest, &fs)?;
    assert!(
        recovered
            .get(b"k", crate::SeqNo::MAX, crate::hash::hash64(b"k"))?
            .is_none(),
        "the chain continues past the first surviving block, so every older \
         version of the boundary key has to stay suppressed",
    );
    assert!(
        recovered
            .get(b"z", crate::SeqNo::MAX, crate::hash::hash64(b"z"))?
            .is_some(),
        "the key that ends the run is recovered",
    );
    Ok(())
}

/// A lost block whose KEY RANGE is unknown (no index separator survived) must
/// arm the suppression as unknown, not with the last key seen. Naming the
/// previous surviving block's key instead suppresses a key nothing shadowed
/// while the key the lost block really covered is republished from its older
/// version.
#[test]
fn salvage_arms_an_unknown_boundary_when_the_lost_block_has_no_range() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // Block 1: `a`. Block 2: `k`'s newest version, lost with its index entry.
    // Block 3: `k`'s older value — the version that must NOT come back.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(1);
    writer.write(InternalValue::from_components(
        b"a",
        b"v",
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::new_tombstone(b"k".as_slice(), 10))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"old",
        5,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // Smash the second block's header MAGIC so the physical walk cannot frame
    // it. The index stays intact, so the walk still trusts the following
    // handle's boundary and emits it — the block itself is recorded as a lost
    // REGION by the gap probe, carrying no key range at all.
    let second_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert!(
            offsets.len() >= 3,
            "the fixture needs one block per entry, got {}",
            offsets.len(),
        );
        usize::try_from(offsets.get(1).copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(second_off) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    let recovered = open(dest, &fs)?;
    assert!(
        recovered
            .get(b"k", crate::SeqNo::MAX, crate::hash::hash64(b"k"))?
            .is_none(),
        "the lost region'\u{2019}s range is unknown, so the FOLLOWING block'\u{2019}s \
         first key is the one it may have shadowed: {report:?}",
    );
    Ok(())
}

/// Suppression rewrites what the block CONTAINS, so the block can no longer be
/// byte-copied: a verbatim append would carry the shadowed key's bytes into the
/// recovered copy while the filtered entry list quietly claims it is gone.
#[test]
fn salvage_does_not_byte_copy_a_block_it_suppressed_a_key_from() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // Block 1: `k`'s newest version, big enough to fill a block on its own and
    // lost below. Block 2: `k`'s older value TOGETHER WITH `m`, so suppression
    // leaves the block non-empty and the emit path still has something to write.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?.use_data_block_size(100);
    writer.write(InternalValue::from_components(
        b"k",
        [b'n'; 200],
        10,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"old",
        5,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"m",
        b"v",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let first_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert_eq!(
            offsets.len(),
            2,
            "the fixture needs the lost version alone in block 1 and the \
             shadowed one sharing block 2",
        );
        usize::try_from(offsets.first().copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(first_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.dropped.len(), 1, "one block is lost: {report:?}");

    let recovered = open(dest, &fs)?;
    assert!(
        recovered
            .get(b"k", crate::SeqNo::MAX, crate::hash::hash64(b"k"))?
            .is_none(),
        "a verbatim copy would carry the suppressed key'\u{2019}s bytes through \
         untouched, republishing the version its newer one had replaced",
    );
    assert!(
        recovered
            .get(b"m", crate::SeqNo::MAX, crate::hash::hash64(b"m"))?
            .is_some(),
        "the block'\u{2019}s other key is still recovered",
    );
    Ok(())
}

/// A columnar source publishes its recovered blocks through its own branch, so
/// it carries the same no-resurrection obligation: losing the block that held a
/// key's newest version must not let the columnar path re-emit an older one.
#[cfg(feature = "columnar")]
#[test]
fn salvage_drops_a_columnar_boundary_key_when_its_newest_version_is_lost() -> crate::Result<()> {
    use crate::table::Writer;
    use crate::{InternalValue, ValueType};

    let dir = tempfile::tempdir()?;
    let fs: Arc<dyn Fs> = Arc::new(StdFs);
    let source = dir.path().join("source");
    let dest = dir.path().join("dest");

    // One entry per block, so `k`'s newest version (a deletion) is the whole of
    // block 2 and its older value the whole of block 3.
    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?
        .use_columnar(true)
        .use_row_group_size(1);
    writer.write(InternalValue::from_components(
        b"a",
        b"v",
        1,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::new_tombstone(b"k".as_slice(), 10))?;
    writer.write(InternalValue::from_components(
        b"k",
        b"old",
        5,
        ValueType::Value,
    ))?;
    writer.write(InternalValue::from_components(
        b"z",
        b"v",
        1,
        ValueType::Value,
    ))?;
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    let second_off = {
        let table = open(source.clone(), &fs)?;
        let offsets: Vec<u64> = table
            .data_block_handles()
            .filter_map(Result::ok)
            .map(|kh| *kh.as_ref().offset())
            .collect();
        assert!(
            offsets.len() >= 3,
            "the fixture needs one block per entry, got {}",
            offsets.len(),
        );
        usize::try_from(offsets.get(1).copied().unwrap_or_default()).unwrap_or(usize::MAX)
    };
    let mut bytes = std::fs::read(&source)?;
    if let Some(b) = bytes.get_mut(second_off + crate::table::block::Header::MIN_LEN + 1) {
        *b ^= 0xFF;
    }
    std::fs::write(&source, &bytes)?;

    let report = salvage_sst(&source, dest.clone(), &fs)?;
    assert_eq!(report.dropped.len(), 1, "one block is lost: {report:?}");

    let recovered = open(dest, &fs)?;
    assert!(
        recovered
            .get(b"k", crate::SeqNo::MAX, crate::hash::hash64(b"k"))?
            .is_none(),
        "the columnar branch must suppress the boundary key too, or salvage \
         republishes a value the deletion had removed",
    );
    assert!(
        recovered
            .get(b"a", crate::SeqNo::MAX, crate::hash::hash64(b"a"))?
            .is_some(),
        "unaffected keys are still recovered",
    );
    Ok(())
}

/// The narrow shape the multi-entry sibling test does not reach: a table whose
/// ONLY point entry is a real weak tombstone that coincides with the
/// `(seqno, start)`-minimal range tombstone. `item_count == 1` then holds, so
/// the sentinel gate fires and the real entry is excluded, leaving NO decoded
/// seqno to cross-check. A re-stamped `seqno#min` therefore passes unchallenged
/// and `point_read` skips the table at snapshots below the forged minimum,
/// exposing an older value from a lower level.
///
/// One write batch produces this shape: entries in a batch share one seqno, so
/// `remove_weak(k)` beside `remove_range(k..)` lands both at the same seqno with
/// the same start key.
#[test]
fn verify_metadata_bounds_keeps_a_lone_weak_tombstone_matching_the_rt_sentinel() -> crate::Result<()>
{
    use crate::UserKey;
    use crate::range_tombstone::RangeTombstone;

    let dir = tempdir()?;
    let source = dir.path().join("source");
    let fs: Arc<dyn Fs> = Arc::new(StdFs);

    let mut writer = Writer::new(source.clone(), 0, 0, Arc::clone(&fs))?;
    writer.write(InternalValue::new_weak_tombstone(
        UserKey::from(b"key-002".as_slice()),
        3,
    ))?;
    writer.write_range_tombstone(RangeTombstone::new(
        UserKey::from(b"key-002".as_slice()),
        UserKey::from(b"key-005".as_slice()),
        3,
    ));
    assert!(writer.finish()?.is_some(), "source SST is non-empty");

    // An honest one-key table cannot hold this RT: coverage pins the RT inside
    // the key range, and an empty RT is rejected at decode. But `key#max` lives
    // in the SAME meta block as `seqno#min`, so the adversary these checks
    // exist to stop re-stamps both, and the shape becomes reachable.
    crate::test_forge::forge_meta_value_both_mirrors(&source, b"key#max", b"key-005")?;

    // Raise seqno#min above the real entry's seqno: caught only if that entry
    // still counts toward the decoded minimum.
    crate::test_forge::forge_meta_value_both_mirrors(&source, b"seqno#min", &9u64.to_le_bytes())?;

    let table = open(source, &fs)?;
    let err = reconcile_error(&table, crate::table::ReconcileGate::MetadataBounds, None);
    assert!(
        matches!(
            err,
            crate::Error::InvalidHeader("meta seqno#min is above the decoded minimum seqno")
        ),
        "the rejection names the seqno#min branch specifically, got {err:?}",
    );
    Ok(())
}

/// A verbatim copy re-checks the source bytes it carries: damage found there
/// keeps its own class, so a repair grades the source instead of stopping on
/// a destination that is fine.
#[test]
fn output_failure_of_carried_source_damage_keeps_its_class() {
    let mismatch = crate::Error::ChecksumMismatch {
        got: crate::Checksum::from_raw(1),
        expected: crate::Checksum::from_raw(2),
    };
    assert!(matches!(
        super::output_failure(mismatch),
        crate::Error::ChecksumMismatch { .. }
    ));
    assert!(matches!(
        super::output_failure(crate::Error::InvalidHeader("carried")),
        crate::Error::InvalidHeader("carried"),
    ));
}

/// An I/O failure of the destination stops the repair as an output failure.
#[test]
fn output_failure_of_destination_io_is_marked() {
    assert!(matches!(
        super::output_failure(crate::io::Error::other("device")),
        crate::Error::SalvageOutput(_),
    ));
}

/// Sealing the destination is part of writing it: its failure is marked,
/// with the cause kept.
#[test]
fn output_failure_of_sealing_is_marked_with_its_cause() {
    assert!(matches!(
        super::output_failure(crate::Error::Encrypt("seal")),
        crate::Error::SalvageOutput(inner) if matches!(*inner, crate::Error::Encrypt(_)),
    ));
}

/// A failure already marked passes through unwrapped once more.
#[test]
fn output_failure_of_a_marked_failure_is_not_marked_twice() {
    let marked = super::output_failure(crate::Error::Encrypt("seal"));
    assert!(matches!(
        super::output_failure(marked),
        crate::Error::SalvageOutput(inner) if matches!(*inner, crate::Error::Encrypt(_)),
    ));
}