znippy_plugin_git/

git_ops.rs

//! **`GitOps` — ours, typed, and base znippy never sees it.**
//!
//! §13.2: `ArchiveTypePlugin` is base znippy's contract and
//! [`crate::NativeGitPlugin`] implements it — base stores, the plugin annotates,
//! exactly as maven, python, rust-toolchain and media already work. `GitOps` is
//! the *other* trait: the one gunnar calls, with git's own vocabulary in it.
//!
//! §13.3: `ArchiveTypePlugin::run_command(cmd, args)` is **CLI-only string
//! dispatch** and delegates here. A server's hot path does not go through
//! stringly-typed dispatch, so nothing below takes a `&str` command or an
//! `args: &[String]`.
//!
//! ```text
//!   gunnar ──typed──► GitOps  (this file, twelve methods)
//!                       │
//!      ┌────────────────┼──────────────────┬─────────────┐
//!      ▼                ▼                  ▼             ▼
//!   ArchiveWrite    ObjectIndex          RefLog         Gc
//!   SafeWriter      ObjectReadStack      push log       NewGeneration
//!   (blob fsync,    (stree → Arrow       (Arrow IPC     (link, compact,
//!    journal row)    → redb tail)         frame = tx)    verify, rename)
//! ```
//!
//! Every arm under that diagram already existed and was already measured. This
//! file adds no storage, no index and no durability mechanism of its own — if a
//! method here looks like it is inventing one, that is a bug.
//!
//! # The one thing that is NOT here
//!
//! `run_command` still lives on the plugin in [`crate::lib`], and it now
//! delegates. Nothing git-shaped went into base znippy for any of this.

use std::collections::{BTreeMap, HashMap};
use std::fs::File;
use std::path::{Path, PathBuf};
use std::sync::{Arc, Mutex, RwLock};

use anyhow::{anyhow, bail, Context, Result};
use znippy_common::ReservedSection;

use crate::arms::StoreConfig;
use crate::exploded::ExplodedStats;
use crate::exploded_cache::ExplodedCache;
use crate::gc::{Gc, GcReport};
use crate::graph::{assign_generations, CommitNode};
use crate::index_layout::{ObjectIndex, OneTableFourColumns};
use crate::indexer::{IndexJob, ObjectAbsorb, PushPath};
use crate::object::{GitHashKind, GitObjectKind};
use crate::pack_walk::PackWalk;
use crate::reach::ReachEntry;
use crate::read_stack::{ObjectReadStack, RebuildTriggers};
use crate::refs::{RefLog, RefUpdate};
use crate::resolve::BaseSource;

// ── the contract, now owned by `git-storage-trait` ────────────────────────────
//
// `GitOps` (the neutral ELEVEN) and its value types live in the
// `git-storage-trait` crate at this repo's root, so a gix backend
// (edda/storage-git-gix) can implement the same contract without linking Arrow
// or OpenZL. The twelfth method, `seal()`, returns `znippy_common::
// ReservedSection` (an Arrow type a gix backend has no analog for), so it is an
// **inherent** method on [`GitStore`] / [`SelectedStore`], not part of the
// trait. The re-exports below keep every existing path
// (`znippy_plugin_git::git_ops::{GitOps, Oid, …}`) alive.
pub use git_storage_trait::{Extent, GitOps, Oid, RefRow, Stored, TxId};


// ── the batch shape, which is measured and must not be papered over ───────────

/// Which path a lookup of `n` oids takes.
///
/// The number this dispatch was written for is `index_layout_bench`'s: on the bare
/// Arrow arms, `lookup_batch` with a **single** element costs 818 ns against
/// `lookup`'s 482 ns — 1.7x worse — because the batch path allocates a result
/// vector, a key vector and a second pass over them to serve one oid.
///
/// **RE-MEASURED against the shipping read stack, and the magnitude does not
/// carry over.** oden 2026-08-07, release, 1-min loadavg 4.94 (another tenant's
/// work — see LAW: never measure on a busy box, and read these as a ratio rather
/// than as absolutes), `ObjectReadStack<OneTableFourColumns>` over a real
/// repository's 2687-object pack, 20 000 iterations
/// ([`crate::store::tests::the_batch_of_one_really_is_slower_than_the_serial_path`]):
///
/// | call | per oid |
/// |---|---:|
/// | `lookup` | **628 ns** |
/// | `lookup_batch` with 1 oid | **639 ns** — 1.02x worse |
/// | `lookup_batch` with 100 oids | **161 ns** — 3.9x better |
///
/// So: the *direction* holds and the dispatch is right — a batch of one is never
/// faster — but on this stack at this size it costs 2%, not 70%. The real finding
/// is the third row: the batch path is worth **3.9x** at
/// [`BATCH_SATURATES_AT`], which is a much stronger reason to keep the two paths
/// apart than the batch-of-one penalty ever was. Both numbers are stated because
/// quoting only the 1.7x would be quoting a figure this stack does not reproduce.
///
/// The dispatch is a named function with its own guard precisely so it cannot
/// quietly become `lookup_batch` for every n.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum LookupPath {
    /// One oid: `ObjectIndex::lookup`.
    Serial,
    /// Two or more: `ObjectIndex::lookup_batch`, one pass.
    Batch,
}

/// Oids at which the batch path stops getting faster. Reported, not enforced:
/// splitting a bigger batch would cost a pass and buy nothing.
pub const BATCH_SATURATES_AT: usize = 100;

/// The dispatch, in one place.
pub fn lookup_path(n: usize) -> LookupPath {
    if n == 1 {
        LookupPath::Serial
    } else {
        LookupPath::Batch
    }
}

// ── the store ─────────────────────────────────────────────────────────────────

/// A pack whose bytes are durable but whose objects the index has not absorbed.
///
/// §13.12's `indexed` bit, in the form this store needs it: membership in
/// [`PackState::pending`] *is* the bit being clear. A read that arrives while it
/// is non-empty falls back to absorbing it — slower, never wrong.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
struct PendingPack {
    pack_id: u64,
    offset: u64,
    len: u64,
    /// The journal queued it on open (a restart), not a push in this process.
    /// What makes the progress row say `resumed 0%` after a reboot.
    resumed: bool,
}

impl PendingPack {
    fn from_job(j: IndexJob) -> Self {
        Self {
            pack_id: j.pack_id,
            offset: j.offset,
            len: j.len,
            resumed: false,
        }
    }
}

/// **§13.12's `indexed` bit — one bit per pack, indexed by pack ordinal.**
///
/// A plain bitset, and the two rejected alternatives say why:
///
/// * **not a hash set.** Pack ordinals are dense and small: they are assigned
///   `0..N` within **one repository** (§13.14 scopes a negotiation, and therefore
///   this store, to one repository — account-wide they would be sparse and the
///   array would grow with the account's total packs). Dense means the bit is one
///   shift and one AND with no hashing, and 10 000 packs is **1250 bytes** of
///   `words`, which lives in L1.
/// * **not a bloom filter.** A bloom's false positive would be *actively wrong*
///   here: it would report a pack absorbed when it is not, the read path would
///   stop falling back, and negotiation would be told `absent` about an object
///   whose bytes are durable — which makes a client withhold objects and lose
///   data. The bitset has no such mode. `1` means absorbed, `0` means fall back,
///   and both are exact.
///
/// **The bit is never stored.** It is derived on open by diffing two facts that
/// are already durable — the journal's extents against the index's rows — in
/// [`Absorber::adopt_journal`]. There is nothing for a derived bit to drift from.
///
/// **Why the extents sit beside it.** A `0` bit is only actionable with the
/// extent to absorb, so `extent[i]` carries it; `None` means "this store has not
/// been told about ordinal `i`", which is not the same as "not absorbed" and must
/// not be counted as work.
///
/// **Why `absorbed` is a separate fact from `extent`.** The ack path does two
/// things in this order: record the pack pending, then hand its extent to the
/// account's channel. The indexer's worker is a different thread and can reach
/// `absorb` *before* the first of those returns. With only a pending list, that
/// race ends in a bit that is set after the rows are already in and is never
/// cleared, so every later read pays a pointless fallback for the life of the
/// process. The bit is what the two writers agree on: [`PackState::note`] will
/// not queue a pack that is already done, and `absorb` will not do the work
/// twice.
///
/// Touched only under short critical sections — never across an absorb, which is
/// what the separate [`Absorber::gate`] is for. **The ack path must never block
/// behind index work.**
#[derive(Default)]
struct PackState {
    /// One bit per pack ordinal, 64 to a word. `1` = its rows are in the index.
    words: Vec<u64>,
    /// The extent of each known ordinal, so a `0` bit can be acted on.
    extent: Vec<Option<Extent>>,
    /// Known ordinals whose bit is `0` — what [`GitStore::unindexed_packs`]
    /// reports, kept as a count so the read path's check is a field load rather
    /// than a scan.
    unabsorbed: usize,
}

impl PackState {
    /// The bit.
    fn is_absorbed(&self, pack_id: u64) -> bool {
        let i = pack_id as usize;
        self.words
            .get(i / 64)
            .is_some_and(|w| w >> (i % 64) & 1 == 1)
    }

    /// Record a durable pack whose rows are not in yet. The bit **going clear**.
    ///
    /// **Idempotent, and the extent slot is the single thing that makes it so.**
    /// A second `note` for an ordinal that already has one — the ack path
    /// arriving after the drain has already absorbed the pack, or a re-queue on
    /// open racing the same pack's push — finds the slot filled and counts no new
    /// work. [`PackState::mark_absorbed`] fills that same slot, so a pack that is
    /// already done cannot be queued behind its own completion, and there is no
    /// second condition here that could disagree with the first.
    fn note(&mut self, pack_id: u64, extent: Extent) {
        let i = pack_id as usize;
        self.grow_to(i);
        if self.extent[i].is_none() {
            self.extent[i] = Some(extent);
            self.unabsorbed += 1;
        }
    }

    /// Room for ordinal `i` in both arrays. Dense ordinals are what makes this a
    /// `resize` and not an insert.
    fn grow_to(&mut self, i: usize) {
        if self.words.len() <= i / 64 {
            self.words.resize(i / 64 + 1, 0);
        }
        if self.extent.len() <= i {
            self.extent.resize(i + 1, None);
        }
    }

    /// The bit **going up**: this pack's rows are in the index.
    fn mark_absorbed(&mut self, pack_id: u64, extent: Extent) {
        let i = pack_id as usize;
        self.grow_to(i);
        if self.words[i / 64] >> (i % 64) & 1 == 0 {
            self.words[i / 64] |= 1 << (i % 64);
            if self.extent[i].is_some() {
                self.unabsorbed -= 1;
            }
        }
        self.extent[i] = Some(extent);
    }

    /// Every known pack whose bit is clear, in ordinal order.
    fn pending(&self) -> Vec<PendingPack> {
        self.extent
            .iter()
            .enumerate()
            .filter(|(i, e)| e.is_some() && !self.is_absorbed(*i as u64))
            .map(|(i, e)| {
                let (offset, len) = e.expect("filtered on Some");
                PendingPack {
                    pack_id: i as u64,
                    offset,
                    len,
                    resumed: true,
                }
            })
            .collect()
    }
}

/// The derived tables: the commit graph, the tree payloads the reachability
/// bitmaps are built from, and this store's own ordinal space.
///
/// **The ordinal space is this struct's, not the projection's.** An
/// [`crate::index_layout::IndexRow::ordinal`] is a row address inside one
/// projection generation and is re-derived by every rebuild, so a bitmap built
/// over projection ordinals would silently address different objects after any
/// rebuild. The bitmaps here are over `ordinal`, which is rebuilt in the same
/// fold that rebuilds them — the two cannot drift because one call produces both.
/// **Nothing in here is a source of truth.** Every field is folded from the
/// `objects` table and §14's exploded table by [`Absorber::refold`], and both of
/// those are on disk — which is what makes a clean reopen come back with the
/// same graph it shut down with. Before the exploded table existed, `graph` and
/// `trees` were *accumulated* here as packs were absorbed and stored nowhere, so
/// a clean shutdown lost them silently: MEASURED on oden 2026-08-08, a reopened
/// store over a fully absorbed 2687-object pack had all 2687 rows and **0 of 551
/// commits**, with `reachable()` returning a commit alone instead of its closure
/// and `gc` seeing an empty live set.
#[derive(Default)]
struct Derived {
    /// Parents-before-children, generations assigned. Folded from the exploded
    /// table's commit payloads.
    graph: Vec<CommitNode>,
    // **`trees` is GONE (lane Q, 2026-08-30).** It held every tree payload in
    // the repository, resident, keyed by a hex `String`, and was rebuilt from
    // the exploded table on every fold. MEASURED on the real `torvalds/linux`
    // mirror: 28.24 GB of tree payloads, 3.24× the 8.13 GiB pack, kept eagerly
    // on disk so this map could be rebuilt into a 50.9 GiB RSS — on a box that
    // ships with 16 GB. `crate::reach::TreeSource` replaced it with a lookup
    // ([`Absorber::tree_source`]), so a tree is read the way every other object
    // is: from the cache if it is there, from its own delta chain in the
    // mmapped pack if it is not.
    /// **The ordinal space, and the ONLY copy of it: every distinct oid in the
    /// store, raw, fixed-width, back to back, in oid-lexicographic order.**
    ///
    /// Ordinal `o` is the slice `oids_raw[o * oid_len .. (o + 1) * oid_len]`, and
    /// because the buffer is **sorted** the inverse (oid → ordinal) is a binary
    /// search over it — [`crate::reach::OrdinalSpace`] is that pair of lookups
    /// and owns nothing but a borrow of this field.
    ///
    /// # Why this is now the only representation
    ///
    /// It used to be one of THREE. `refold` also built `oids: Vec<String>` (the
    /// same list, hex) and `ordinal: HashMap<String, u32>` (the inverse, hex
    /// keys) — a 40-byte hex `String` per oid in the vector, a second as each
    /// map key, plus 20 raw bytes here. MEASURED, lane MEMORY, oden 2026-08-31
    /// (`examples/rss_per_open_store`): a cached open store cost **382 bytes of
    /// live RSS per indexed object** on a blob-only corpus and **744–966 B** on
    /// the two real repository packs — about half of it that pair. And it was
    /// paid twice at every fold: REFOLD measured the hex `Vec` at ~0.13 s and
    /// the `HashMap` at ~0.08 s of a 136k-object reopen, ~0.21 s spent turning
    /// bytes the tail already held into `String`-keyed tables.
    ///
    /// The flat buffer already answered both questions — a slice and a binary
    /// search — so the two `String`-per-oid tables were LAW 5's exact shape: a
    /// second and third spelling of one list that a comment had to *promise*
    /// agreed. They are gone; the serve path reads oid ⇄ ordinal straight off
    /// this buffer. See `gunnar/.nornir/stinky.md` S-082.
    ///
    /// # Why raw rather than hex
    ///
    /// Because every *serving* answer is built out of raw bytes.
    /// `reachable_raw` answers in oid bytes and each object is a
    /// `copy_from_slice` of a slice of this buffer — nothing is encoded,
    /// decoded or allocated per object. `reachable_oids` (the maintenance path)
    /// encodes on the way out, at one `String` per object, and nothing on the
    /// serve path calls it. The tail hands oids in oid order, which over
    /// fixed-width oids is byte order, so this is already the sorted
    /// concatenation with no sort of its own — see [`Absorber::refold`].
    oids_raw: Vec<u8>,
    /// The bitmaps, or empty if nothing has needed them since the last fold.
    ///
    /// **Behind an `Arc` since 2026-08-14, and that is a measurement and not a
    /// tidy-up.** [`Absorber::reach_bitmaps`] used to hand out `d.reach.clone()`
    /// — a deep copy of every commit's `RoaringBitmap` plus its oid `String` —
    /// and `crate::serve::GitServe::select` calls
    /// [`GitStore::reachable_oids`] **twice** per request (once for `want`, once
    /// to close over `have`). MEASURED on oden with `stage-probe`, the
    /// znippy-served `nornir` mirror (12 455 objects), a 4-object negotiated
    /// fetch: the `select.bitmap` stage was **75.2 % of the request's CPU** and
    /// **645 876 allocations**; at four requests sharing one warm store it was
    /// still 62 762 allocations *per request*. None of that copying is work the
    /// answer depends on — the table is immutable between folds — so the `Arc`
    /// removes it rather than dividing it across cores.
    reach: Arc<Vec<ReachEntry>>,
    /// Every commit's oid as **raw bytes**, the form
    /// [`crate::serve::GitServe::select`] tests a `want` against.
    ///
    /// `None` means a graph row's oid did not parse as hex, which `select`
    /// answers by declining the whole request. That check is why this is
    /// `Option` rather than a plain set: it used to run **per request** —
    /// `graph_snapshot()` cloned the whole `Vec<CommitNode>` and then
    /// `hex::decode`d every commit into a fresh `Vec<u8>` — and the answer
    /// cannot change between folds, so it is folded once here and borrowed
    /// after. The refusal is preserved exactly; only the arithmetic moved.
    commit_raw: Option<Arc<std::collections::HashSet<Vec<u8>>>>,
    /// **hex oid → index into [`Derived::graph`].** The graph-by-oid lookup
    /// [`crate::reach::accumulate`]'s walk needs, folded once by the same call
    /// that folds `graph` and `commit_raw` rather than rebuilt on every
    /// negotiation. `reachable` used to `d.graph.iter().map(...).collect()` a
    /// `HashMap<&str,&CommitNode>` per call — O(commits) per fetch, on every
    /// fetch and clone — for a table that cannot change between folds. MEASURED
    /// on oden 2026-09-01 (`bench_reachable_negotiation`, a 20 000-commit linear
    /// history): a negotiation over a one-object closure spent **895 µs**
    /// building this map and the reach one below, work the answer never used.
    /// It is keyed by hex, not raw, because the walk keys by hex — parents and
    /// tree oids in a [`CommitNode`] are hex — so a raw key would force a decode
    /// per walked edge to save the O(w+h) tip encode, a bad trade at any size.
    graph_index: HashMap<String, u32>,
    /// **hex commit oid → index into the [`Derived::reach`] table**, built by
    /// the same call that builds `reach` and cleared with it, so the pair is
    /// always in step. Replaces the per-call `HashMap<&str,&RoaringBitmap>`
    /// `reachable` rebuilt over every bitmap on every request. Behind an `Arc`
    /// for the same reason `reach` is: it is handed to the negotiation with a
    /// refcount bump, never a copy.
    reach_index: Arc<HashMap<String, u32>>,
    /// How many commits the graph held when [`Derived::reach`] was last built
    /// from scratch — the denominator of the carry rule in
    /// [`Absorber::refold`]. Bitmaps carried across a fold are exact for the
    /// commits they cover, but every commit added since is served by walking
    /// down to the nearest carried one, so once the additions outgrow
    /// [`REACH_CARRY_MAX`] of this the next reader rebuilds the table whole.
    reach_full_len: usize,
    /// Verbatim packs, so any absorbed object's content can be re-derived from
    /// the truth on demand. `(archive offset, len)`.
    packs: Vec<(u64, u64)>,
}

/// **The carry rule's ceiling**: a fold carries the reachability bitmaps
/// across when the graph has grown by at most this fraction of the commits the
/// last full build covered (and never fewer than 64 commits' worth). Beyond it
/// the table is rebuilt from scratch on the next read — which is what every
/// fold used to do, unconditionally.
///
/// MEASURED 2026-09-05 on oden (`gunnar serve`, a 10 000-commit / 51 023-object
/// repository over ssh): the FIRST clone after any push took 10.2 s and the next
/// 1.3 s, and after one more 4-object push the pattern repeated (10.9 s, then
/// 1.0 s). The fold had dropped `Derived::reach` and the next `reachable`
/// rebuilt every commit's tree closure — 0.2 ms per object of server CPU for a
/// push that changed nothing about any existing commit's reachability. A bitmap
/// is a fact about an immutable commit; only its ordinal numbering moves when
/// the sorted oid space grows, and a monotone remap costs one pass over the
/// bits. Sized at a tenth so the amortised rebuild stays bounded by the pushes
/// that earned it, like the persisted index's own fold threshold.
///
/// MEASURED AGAIN with the carry in place, same box, same repository, same
/// sequence (seed push, clone, clone, 4-object push, clone, clone): 10.5 s,
/// 1.15 s, then **1.12 s** and 1.15 s. The server logged the fold carrying all
/// 501 sampled entries, and the one 9.3 s build from scratch is the first
/// clone's, which no carry can serve. `reach_source_stats` is the same fact for
/// a test to assert.
pub(crate) const REACH_CARRY_MAX_FRACTION: usize = 10;
/// The floor under the rule, in commits: a repository small enough that a tenth
/// of it is a handful still carries across pushes of up to this many commits.
pub(crate) const REACH_CARRY_MIN_COMMITS: usize = 64;

/// **The object-level index, and the one thing that writes into it.**
///
/// Its own allocation rather than a field group inside [`GitStore`], and §13.9-12
/// is the only reason: the drain runs on the account indexer's worker thread, so
/// whatever it absorbs into cannot be reachable solely through a `&GitStore`.
/// Splitting it out is what lets the background worker and a falling-back read
/// call **the same** absorb instead of two copies of it (LAW 5) — the alternative
/// was a second implementation on the worker side, which is precisely the
/// twinning that law forbids.
///
/// It is a **leaf**: nothing in here points back at the store, so the worker's
/// `Arc` closes no cycle and the store still drops (and joins its worker)
/// normally.
/// Entries below which [`GitStore::resolve_emit_payloads`] does **not** fan out.
///
/// A `std::thread::scope` spawn is ~10–20 µs a worker and the per-entry work it
/// would divide is a bounds check plus a varint; a 100-object fetch would pay
/// four spawns to save four microseconds of parsing. The threshold is what keeps
/// `crate::serve`'s "one core's worth per transfer" true for every request that
/// is not a bulk clone — and a bulk clone is the only request the copy this
/// replaces was ever measured on.
const FAN_OUT_AT: usize = 4096;

/// The folded commit graph, cached beside the verbatim pack so a reopen *reads*
/// it instead of re-deriving it from every commit payload (S-065).
const GRAPH_CACHE_FILE: &str = "objects.graph";
/// `objects.graph` framing: 4-byte magic + 1-byte version + 32-byte token, then
/// the Arrow IPC commit graph.
const GRAPH_CACHE_MAGIC: &[u8; 4] = b"GRPH";
const GRAPH_CACHE_HEADER: usize = 4 + 1 + 32;

/// **`objects.ordinal` is no longer written, and an existing one is removed at
/// open.** It was S-082's memo of the ordinal space — every oid in sorted order,
/// with a `(epoch, count)` header validated against the redb tail — so that a
/// cold open could *read* the space instead of scanning the whole redb tail to
/// rebuild it. Since 2026-09-04 the read stack persists its projection
/// (`objects.index`) and the sorted oid list IS that index's oid column, mapped;
/// [`ObjectReadStack::oids_raw`] hands it back with a merge of the few redb rows
/// since the last fold. The scan the memo spared no longer exists, and a second
/// file carrying the same bytes under a bespoke header is exactly the kind of
/// file this store should not have. On the kernel mirror it was 278 964 902 B —
/// 20.0 B/object, the oid column to the byte. The name is kept only so the
/// stale file can be found and deleted.
const ORDINAL_CACHE_FILE: &str = "objects.ordinal";

/// The validity token for a persisted commit graph: blake3 over the sorted raw
/// oids of **every** object in the store.
///
/// It is both necessary and sufficient. The graph is a pure, deterministic
/// function of the commit payloads, which are content-addressed by their oids,
/// so an identical object set yields a byte-identical graph — and this hashes
/// the whole object set, in the one order `refold` already sorts it into, with
/// no read the fold did not already make. A mismatch (any push, any gc) forces a
/// recompute; a 256-bit blake3 makes an accidental match of two different object
/// sets not a case worth engineering against.
/// **`seen − live`: what a GC may drop.** `seen` is every oid the index held
/// when the GC started (`oids_in_order`), `live` the reachable set computed
/// after it. The difference is the dead set, and it is a POSITIVE set on
/// purpose: `retain(|o| live.contains(o))` would also drop every row that
/// arrived between the snapshot and the drop — a push in flight, whose refs
/// are not updated yet — because to `live` those are simply unknown. A row in
/// neither set is kept. `live` is hex, `seen` raw; the decode happens once
/// here rather than once per row in the retain predicate.
pub(crate) fn dead_oids(
    seen: Vec<Vec<u8>>,
    live: &std::collections::HashSet<String>,
) -> std::collections::HashSet<Vec<u8>> {
    let live_raw: std::collections::HashSet<Vec<u8>> =
        live.iter().filter_map(|h| hex::decode(h).ok()).collect();
    seen.into_iter().filter(|o| !live_raw.contains(o)).collect()
}

fn graph_token(oids_raw: &[u8]) -> [u8; 32] {
    let mut h = znippy_common::blake3::Hasher::new();
    h.update(oids_raw);
    *h.finalize().as_bytes()
}

/// Commits that get a reachability bitmap on the **live** selection path.
///
/// # It was `usize::MAX`, and that was the single most expensive line in a fetch
///
/// The comment that stood here said a sampled table *"is for the sealed archive
/// where the cost is disk"* and that every commit must be bitmapped on the live
/// path. That was not a preference, it was forced:
/// [`GitStore::reachable_oids`] had no walk, so a `want` with no bitmap could
/// only contribute itself — a silent under-send — and `crate::serve`'s `select`
/// therefore had to refuse unless *"in the graph"* and *"has a bitmap"* were the
/// same fact.
///
/// MEASURED on oden 2026-08-14, one 4-object negotiated fetch out of the
/// znippy-served `nornir` mirror (1 581 commits, 12 455 objects): the
/// `select.bitmap` stage was **75 % of the request's CPU**, and **530 218 of its
/// 575 349 allocations were this build** — which `refold` throws away again on
/// the next push. The gix arm answered the identical fetch in 6.8 ms because it
/// *reads* a `.bitmap` instead of building one.
///
/// [`crate::reach::accumulate`] is the walk that makes the cap legal, and 512 is
/// the number [`crate::reach::ReachPolicy::default`] already uses for the sealed
/// archive — the same figure git's own bitmap selection is of the order of, and
/// deliberately not a second, different constant to keep in step.
const LIVE_REACH_COMMITS: usize = 512;

/// [`LIVE_REACH_COMMITS`], or `ZNIPPY_GIT_REACH_COMMITS`.
///
/// **An operator knob, not an arm.** It selects no implementation and changes no
/// byte of any answer: the walk covers whatever the table does not, so the cap
/// trades build cost against walk cost and nothing else. That property is the
/// whole subject of
/// `the_bounded_walk_answers_exactly_what_a_full_bitmap_table_answers`, which
/// drives this very knob to its two extremes over one store.
///
/// Read **once per process** through [`crate::arms::read_env`], so it is
/// counted with every other environment read this crate makes, and so a
/// reachability request does not cost a `getenv` — until 2026-08-21 it did,
/// once per `reach_bitmaps`/`reachable_oids` call, uncounted.
fn live_reach_policy() -> crate::reach::ReachPolicy {
    static MAX: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    let max_commits = *MAX.get_or_init(|| {
        crate::arms::read_env(crate::arms::ENV_REACH_COMMITS)
            .and_then(|v| v.parse::<usize>().ok())
            .unwrap_or(LIVE_REACH_COMMITS)
    });
    crate::reach::ReachPolicy { max_commits }
}

/// Ceiling on phase-1 workers **per request**.
///
/// The serving tier admits ~32 concurrent transfers, and the fan-out primitive
/// has no reentrancy detection, so an uncapped `gatling_for_each` here is
/// 32 × ncores threads on a 128-core box. Four is a bound that leaves the
/// admission gate in charge of the machine.
const MAX_RESOLVE_WORKERS: usize = 4;

/// Workers for phase 1: [`MAX_RESOLVE_WORKERS`], or `ZNIPPY_GIT_EMIT_WORKERS`.
///
/// **An operator knob, not an arm.** It selects no implementation and changes no
/// byte of the emitted pack — a pack emitted at one worker and at sixteen is
/// byte-identical, because phase 1 only decides *where each entry's bytes are*
/// and phase 2 writes them in one serial order either way. `0` is read as "as
/// many as this box has cores", which is `gatling_for_each`'s own convention.
///
/// Read **once per process** through [`crate::arms::read_env`] — see
/// [`live_reach_policy`] for why; this one was a `getenv` per emitted request
/// until 2026-08-21.
fn emit_workers() -> usize {
    static WORKERS: std::sync::OnceLock<usize> = std::sync::OnceLock::new();
    *WORKERS.get_or_init(|| {
        crate::arms::read_env(crate::arms::ENV_EMIT_WORKERS)
            .and_then(|v| v.parse::<usize>().ok())
            .unwrap_or(MAX_RESOLVE_WORKERS)
    })
}

pub(crate) struct Absorber<S: ObjectIndex = OneTableFourColumns> {
    blobs: PathBuf,
    /// Read handle on the verbatim packs, for `pread` of an extent.
    ///
    /// The **fallback**, since 2026-08-14, not the ordinary path: `map` below
    /// answers every extent the mapping covers, and this reads the ones past its
    /// end (a push that landed after the snapshot was taken).
    ///
    /// **It is the writer's own open file description** (2026-08-30, lane AH),
    /// not a second `File::open`: this reads with `pread` at an explicit offset
    /// and shares no cursor with `append`, so a private descriptor bought a
    /// kernel object and a page-cache alias per open repository and no
    /// isolation. See [`crate::archive_write::open_blobs`] and
    /// `tests/fd_per_store.rs`.
    reader: Arc<File>,
    /// The same file, mapped read-only. See [`crate::archive_map`] for why an
    /// append-only file `gc` never rewrites is safe to map, and for the ~20 % of
    /// serve cycles the `pread`-per-object shape was spending on copies.
    map: crate::archive_map::ArchiveMap,
    hash: GitHashKind,
    /// The `objects` table: stree → Arrow → redb tail.
    ///
    /// `S` is the payload layout the projection is built in
    /// ([`crate::arms::IndexArm`]) and it is a **type parameter, not a field**:
    /// this is the hot path, every lookup goes through it, and the whole reason
    /// the arms exist is to measure the differences between them. A `dyn` here
    /// would put a virtual call on every single oid.
    objects: ObjectReadStack<S>,
    /// **§14's exploded objects table** — oid → resolved type and content, built
    /// eagerly by this absorb and by nothing else.
    ///
    /// Beside the `objects` table rather than inside it because the two have
    /// different lifetimes: `objects` is the index and losing it loses the
    /// repository's addressing, while this one is derived, verifiable against the
    /// verbatim bytes and deletable at any time. They are separate files so that
    /// "droppable" is a `rm`, not a schema migration.
    ///
    /// SWAPPED 2026-08-13 from the redb `ExplodedTable` to **one Arrow IPC
    /// table**, payload in the row. redb held 16.8 GB of kernel payload in a
    /// 204 GB file — 12× — and serialised the gatling fan-out behind its one
    /// writer at 96.8% of a single core. See [`crate::exploded_arrow`].
    ///
    /// **BUDGETED AND EVICTABLE since lane Q (2026-08-30)**: the two-pool
    /// [`ExplodedCache`] — the indexer's eager pool (commits and tags) beside a
    /// lazy pool the FIRST reader of a blob or tree fills and a byte budget
    /// empties. The unbounded eager tier it replaces MEASURED **141.7 GiB** for
    /// the 8.12 GiB `torvalds/linux` pack on the rig, 17× the pack, which is
    /// more than one appliance holds for all its repositories together.
    exploded: Arc<ExplodedCache>,
    derived: RwLock<Derived>,
    /// How many folds since open read the persisted commit graph off disk vs
    /// recomputed it from the exploded table. Counters, not clocks — the only
    /// thing that can say which path a fold took (LAW 2). See
    /// [`Absorber::graph_source_stats`].
    graph_read: std::sync::atomic::AtomicU64,
    graph_recomputed: std::sync::atomic::AtomicU64,
    /// **Is `objects.graph` read and written at all** ([`crate::arms::ENV_GRAPH_CACHE`],
    /// read once at open). `false` folds from the exploded table every time and
    /// leaves no file: a store that serves no reads holds no derived memo.
    graph_cache: std::sync::atomic::AtomicBool,
    /// How many folds CARRIED the reachability table across (renumbered, no
    /// tree opened) vs how many reads had to build it from scratch. The only
    /// witness that the carry ran — the bytes a carried and a rebuilt table
    /// answer with are identical by construction (LAW 2). See
    /// [`Absorber::reach_source_stats`].
    reach_carried: std::sync::atomic::AtomicU64,
    reach_rebuilt: std::sync::atomic::AtomicU64,
    /// Where the last fold's **ordinal space** came from: how many oids straight
    /// off the persisted index's column, how many off a scan of the redb tail.
    /// The same shape as [`Self::graph_read`]/[`Self::graph_recomputed`], and for
    /// the same reason: which source produced the space is not visible in the
    /// (identical) bytes, only in a counter (LAW 2). See
    /// [`Absorber::ordinal_source_stats`].
    ordinal_folded: std::sync::atomic::AtomicU64,
    ordinal_tail: std::sync::atomic::AtomicU64,
    packs: Mutex<PackState>,
    /// Serialises **absorption**, which is long. Held across a whole pack.
    ///
    /// Two absorbs of one pack are duplicated *work* and nothing worse: every
    /// writer under it is keyed by oid — [`ObjectReadStack::append`] swallows an
    /// identical second row, the exploded table overwrites its own key, and the
    /// commit graph is folded from that table rather than pushed at, so it cannot
    /// double. (It could before §14's table existed, when the graph was a `Vec`
    /// accumulated per absorb; that is why this paragraph used to say the gate
    /// was load-bearing for correctness and now does not.) The gate is still what
    /// makes the ordering below hold: **the rows go in, and only then is the bit
    /// cleared**, so a read either finds the bit still clear (and waits here) or
    /// finds the rows. Neither order can produce "absent".
    gate: Mutex<()>,
    /// **The progress counter** ([`crate::progress`]) a dashboard or a stats
    /// call reads. One per store, registered process-wide.
    progress: Arc<crate::progress::IndexProgress>,
    /// Set by [`GitStore::cancel_index`]: the repository is being deleted. The
    /// resolver checks it every few thousand objects and aborts the pack; the
    /// drain skips every pack still queued.
    cancel: std::sync::atomic::AtomicBool,
    /// Names the store in log lines: its root directory.
    label: String,
    /// The memory watchdog this store's worker pauses on. The process-wide one
    /// unless a test hands in a fake ([`GitStore::set_mem_watchdog`]).
    watchdog: RwLock<Arc<crate::memguard::MemWatchdog>>,
    /// **Pending packs a reader has already scanned**, keyed by pack offset:
    /// the walk and an oid-sorted row table, built by ONE read-only pass and
    /// kept until the pack is absorbed. Every later lookup into that pack is
    /// a binary search; a content read inflates one chain. Without this,
    /// every `has` during the index window re-resolved the whole pack —
    /// MEASURED on the rig 2026-08-30: a `push --mirror` of linux stalled
    /// pre-ack, ~800 ref updates × one 8 GiB resolve each.
    scanned: Mutex<HashMap<u64, Arc<ScannedPack>>>,
    /// One scan build at a time: a second reader waits for the first reader's
    /// pass rather than starting its own. That is waiting on a *reader*, not
    /// on the indexer.
    scan_build: Mutex<()>,
    /// **Packs this store has stopped trying to index**, by ordinal, with how
    /// many attempts it took to stop and the reason the last one gave.
    ///
    /// 🔴 The state this replaces was *no state*. A pack whose absorb failed
    /// stayed pending, every absorb loop hit it FIRST (they run in ordinal
    /// order) and aborted on it with `?`, so the packs behind it were never
    /// reached — and the next push, and the one after that, re-did the same
    /// doomed resolve and was refused for it. MEASURED on oden 2026-08-30
    /// (lane AI): one pack that could not be absorbed refused **every** push to
    /// that repository from then on, 21 of 140, and lane AD's standby re-asked
    /// for it every 300 s for 21 minutes. Work that cannot succeed has to stop
    /// and say why; see [`MAX_ABSORB_ATTEMPTS`] and [`Absorber::absorb_each`].
    unindexable: Mutex<HashMap<u64, Unindexable>>,
}

/// A pack this store will not try to index again, and why.
#[derive(Debug, Clone)]
pub struct Unindexable {
    /// The pack's ordinal.
    pub pack_id: u64,
    /// Its extent in the verbatim archive. The bytes are still there.
    pub extent: Extent,
    /// How many absorbs were spent before it stopped.
    pub attempts: u32,
    /// The last attempt's root cause — the backend's own sentence, which is
    /// what a client and an operator both get told.
    pub reason: String,
}

/// How many times one pack's absorb is attempted before the store stops.
///
/// Three, and the reason it is not one: an absorb can fail for a reason outside
/// the pack — the memory watchdog gave up, the repository was being deleted, a
/// transient read error — and a single bad moment must not condemn durable
/// bytes. It is not thirty either, because the resolve is deterministic in the
/// store's own state: a pack that failed three times with nothing else changing
/// fails the fourth time for the same reason, and every one of those attempts
/// is paid for by a client waiting on a push.
pub const MAX_ABSORB_ATTEMPTS: u32 = 3;

// **The pack this thread is inside an absorb of**, if any.
//
// `absorb_gated` holds `Absorber::gate` for the whole pack, and the resolver it
// runs asks this very store for delta bases through `BaseSource::content`. That
// re-entry must not take the gate again — a `std::sync::Mutex` is not
// reentrant, so it would not be a wait, it would be a deadlock — and it must
// not treat the pack it is halfway through as something it can absorb. One
// `Cell` answers both questions, and it is the same guard shape
// `scan_pending`'s `SCANNING` uses for the same reason.
thread_local! {
    static IN_ABSORB: std::cell::Cell<Option<u64>> = const { std::cell::Cell::new(None) };
}

/// Sets [`IN_ABSORB`] for the duration of one absorb and restores whatever was
/// there before — absorbs nest (a pack whose base is in another pending pack).
struct InAbsorb(Option<u64>);

impl InAbsorb {
    fn enter(pack_id: u64) -> Self {
        InAbsorb(IN_ABSORB.with(|c| c.replace(Some(pack_id))))
    }
}

impl Drop for InAbsorb {
    fn drop(&mut self) {
        IN_ABSORB.with(|c| c.set(self.0));
    }
}

/// One pending pack, scanned read-only: what a lookup needs and nothing that
/// the index will hold. O(objects) memory — for linux ~56 B of walk + 32 B of
/// row per object, ~1.2 GB — and gone when the pack's bit goes up.
///
/// **Nothing on the production read path builds one of these any more.** A
/// reader that needs an un-indexed pack indexes it instead
/// ([`Absorber::index_pending_now`]); this remains for the one caller that
/// cannot — a resolver already inside `absorb_gated`, which may consult a scan
/// somebody else already built but must never start one.
struct ScannedPack {
    walk: crate::pack_walk::PackWalk,
    /// `(oid, entry index, resolved size)`, sorted by oid.
    rows: Vec<(Vec<u8>, u32, u64)>,
    archive_offset: u64,
}

impl ScannedPack {
    fn find(&self, oid: &[u8]) -> Option<(usize, u64)> {
        self.rows
            .binary_search_by(|(o, _, _)| o.as_slice().cmp(oid))
            .ok()
            .map(|i| (self.rows[i].1 as usize, self.rows[i].2))
    }

    fn row(&self, oid: &[u8]) -> Option<crate::index_layout::IndexRow> {
        let (entry, size) = self.find(oid)?;
        let e = &self.walk.entries[entry];
        Some(crate::index_layout::IndexRow {
            // Not ranked: the row is not in the index yet.
            ordinal: u32::MAX,
            offset: e.offset + self.archive_offset,
            len: e.len,
            obj_type: e.obj_type,
            uncompressed_size: size,
            delta_base: match &e.delta_base {
                crate::pack_walk::DeltaBase::Offset(o) => o + self.archive_offset,
                _ => 0,
            },
        })
    }
}

/// One repository's git store.
///
/// | on disk | what it is | who owns the format |
/// |---|---|---|
/// | `<root>/objects.pack` | the verbatim pack bytes, appended | [`SafeWriter`] |
/// | `<root>/objects.pack.journal` | the durable extent rows | [`SafeWriter`] |
/// | `<root>/objects.tail` | the un-sealed index tail | [`ObjectReadStack`] (redb) |
/// | `<root>/objects.exploded` | §14's resolved content, **droppable** | [`ExplodedArchive`] (one Arrow IPC table) |
/// | `<root>/refs.log` | the ref transaction log | [`RefLog`] (Arrow IPC frames) |
/// | `<root>/<name>.znippy` | the sealed archive GC compacts | base znippy |
///
/// Nothing in that table is new. The store is the composition, not a fifth
/// format.
/// # Which implementation of each trait, and how a caller picks
///
/// Three of the four boxes in the diagram at the top of this file have more than
/// one implementation. [`StoreConfig`] is the choice, [`crate::arms`] documents
/// what each arm promises, and the dispatch is split deliberately:
///
/// | trait | selected by | dispatch | cost |
/// |---|---|---|---|
/// | [`ObjectIndex`] | the type parameter `S` | **static** | none — the hot path is monomorphised |
/// | [`ArchiveWrite`](crate::archive_write::ArchiveWrite) | [`StoreConfig::writer`] | `dyn`, inside [`PushPath`] as it already was | one virtual call **per pack append**, not per object |
/// | [`Gc`] | [`StoreConfig::gc`] | `dyn` | one virtual call per `gc()` run |
///
/// `S` defaults to [`OneTableFourColumns`], so `GitStore` still names exactly
/// what it named before and every existing caller compiles unchanged.
pub struct GitStore<S: ObjectIndex = OneTableFourColumns> {
    root: PathBuf,
    /// The verbatim pack bytes. The **truth** (§14).
    blobs: PathBuf,
    /// The sealed znippy archive `gc` compacts. Only GC and `seal` touch it.
    archive: PathBuf,
    /// The selected [`ArchiveWrite`](crate::archive_write::ArchiveWrite) arm
    /// plus the per-account index channel. The ack path.
    push: PushPath,
    account: String,
    hash: GitHashKind,
    /// The object-level index and the absorb behind it — **shared with the
    /// account indexer's worker**, which is how a push ends in object rows
    /// without a read having to ask for them.
    absorber: Arc<Absorber<S>>,
    refs: RefLog,
    /// Serialises the ref log's read-compare-append. A CAS that read the
    /// namespace, then appended, without holding this could be overtaken between
    /// the two and would overwrite the update it did not see.
    ref_gate: Mutex<()>,
    /// The selected [`Gc`] arm. Built once, here, and run by
    /// [`GitOps::gc`] — a `Box<dyn>` because a GC is a whole-archive
    /// compaction that happens on a maintenance timer, so one virtual call
    /// against it is unmeasurable and a third type parameter on every signature
    /// in the crate is not free to read.
    gc: Box<dyn Gc + Send + Sync>,
    /// What this store was built with, for a bench row or a config dump. A
    /// description of the arms already constructed — **nothing consults it on a
    /// serving path**.
    arms: StoreConfig,
}

impl GitStore<OneTableFourColumns> {
    /// Open (creating) the store for one repository under `root`.
    ///
    /// **The default arms, and it reads no environment to get them.** An
    /// operator who exported `ZNIPPY_GIT_WRITER=fast` does not change what this
    /// builds — a caller that never asked for a selector must not have its
    /// durability contract moved out from under it. Use
    /// [`open_from_env`] when the selection is wanted.
    ///
    /// ⚠ One variable *is* read here, and it is not an arm:
    /// `ZNIPPY_GIT_REDB_CACHE_BYTES` (see
    /// [`crate::arms::redb_cache_bytes`]). It selects no implementation,
    /// changes no durability contract and leaves the bytes on disk identical —
    /// it caps redb's page cache, whose own default is 1 GiB *per database*, of
    /// which this store opens two. The sentence above is about arms, and a
    /// memory ceiling is not one.
    pub fn open(root: &Path, account: &str) -> Result<Self> {
        Self::open_with(root, account, GitHashKind::Sha1)
    }

    /// Same, with an explicit oid width. sha256 repositories exist and the store
    /// does not get to assume otherwise; every oid it stores is
    /// [`GitHashKind::oid_len`] bytes wide and the tail refuses a mixed batch.
    ///
    /// # What a reopen recovers, and why nothing had to be stored for it
    ///
    /// A store opened over an `objects.pack` some earlier process was still
    /// indexing has to end up in the same state that process would have reached.
    /// §13.12's `indexed` bit is what says so, and it is **derived here rather
    /// than persisted**: a pack is unabsorbed *iff* its extent is in the journal
    /// and its rows are not in the index, and both of those are already durable
    /// (see [`Absorber::adopt_journal`]). The packs that come back clear are put
    /// straight onto the account's channel, so the drain finishes the interrupted
    /// work with no read having to ask — and until it does, every read falls back
    /// exactly as it would have before the crash. **Slower, never wrong**, across
    /// a restart as well as within one.
    ///
    /// The bits are set **before** the channel is fed, which is the same ordering
    /// the ack path uses for the same reason: a read that arrives in between finds
    /// the bit clear and waits, never an index that does not mention the pack.
    ///
    /// # What a reopen recovers of the DERIVED tables, and how that was fixed
    ///
    /// [`Derived::graph`] and [`Derived::trees`] come back **in full**, for every
    /// pack, whether or not anything is re-queued. They are folded from §14's
    /// exploded table ([`Absorber::refold`]), which is on disk, so a clean
    /// shutdown carries them across.
    ///
    /// This was a live correctness hole until 2026-08-08 and it is recorded
    /// rather than quietly repaired. The graph was *accumulated in RAM* as packs
    /// were absorbed and stored nowhere; a store killed mid-absorb got it back
    /// because the interrupted pack was re-queued and re-resolved, but a store
    /// that shut down **cleanly** re-queued nothing and came back with an empty
    /// graph. MEASURED on oden 2026-08-08: 2687 rows and **0 of 551 commits**,
    /// `reachable` on a commit returning that commit alone instead of its
    /// closure — a *quiet* wrong answer — and `gc` seeing an empty live set. The
    /// fix is the table, because the payloads the graph is folded from had no
    /// home; [`tests::a_clean_shutdown_and_reopen_keeps_the_graph_and_reachable`]
    /// is what proves it, over a real process that exits normally.
    ///
    /// The exploded table is still **droppable**: if its row count is short of
    /// the `objects` table's, [`Absorber::adopt_journal`] declares no pack
    /// absorbed and every one of them is re-queued and re-exploded. Absent means
    /// rebuild, never wrong (§13.12, applied unchanged to a derived table).
    ///
    /// **Also the default arms, and also no environment read** — see
    /// [`open`](GitStore::open).
    pub fn open_with(root: &Path, account: &str, hash: GitHashKind) -> Result<Self> {
        // The default ARMS, and the environment's CEILING: the one variable
        // this constructor reads, because an operator capping a server's
        // footprint must not have to pick a different constructor to do it.
        // See [`crate::arms::redb_cache_bytes`].
        Self::open_with_arms(
            root,
            account,
            hash,
            StoreConfig::DEFAULT.with_redb_cache_bytes(crate::arms::redb_cache_bytes()?),
        )
    }
}

/// Where a stored entry's delta base sits, relative to the request being served.
///
/// The `Outside` arm carries the base in **both** namings because the two delta
/// forms know only one each — an `OFS_DELTA` its archive offset, a `REF_DELTA`
/// its oid — and [`GitStore::emit_set`]'s thin arm has to ask "does the client
/// hold this" in whichever one the entry speaks.
enum BaseOf {
    /// Not a delta at all. There is nothing to place it after and nothing to
    /// re-head; the stored entry is copied as it stands.
    Whole,
    /// A delta whose base **is** in the request, at this archive offset.
    ///
    /// The offset is carried rather than recomputed because a `REF_DELTA` does
    /// not know it — it names its base by oid, and its own `delta_base` column
    /// is `0` by construction — and the emitter needs it twice: as the
    /// `EmitEntry::delta_base` that [`crate::pack_walk::topological_order`]
    /// orders on, and as the base an in-pack ref-delta is re-headed to point at.
    Inside { at: u64 },
    /// A delta whose base the request does not carry.
    Outside {
        /// The base's archive offset, `0` when the entry named an oid instead.
        offset: u64,
        /// The base's oid, `None` when the entry named an offset instead — or
        /// when the entry is too short to read one out of.
        oid: Option<Vec<u8>>,
    },
}

/// The bases a **thin** pack may name without carrying them: what the receiver
/// already held before this transfer, keyed both ways.
///
/// Built by [`GitStore::client_bases`], which is where the argument for trusting
/// it is written down.
struct ExternalBases {
    /// The held object's oid, keyed by its archive offset — the coordinate an
    /// `OFS_DELTA` names its base in, and the only reason this map exists.
    by_offset: HashMap<u64, OidKey>,
    /// The same objects keyed the way a `REF_DELTA` names them.
    held: std::collections::HashSet<OidKey>,
}

/// **A raw oid as a fixed-size, `Copy` hash key — 20 bytes for SHA-1, 32 for
/// SHA-256, and no heap allocation either way.**
///
/// The two maps in [`ExternalBases`] are repository-sized: the client's `have`
/// closure on an incremental fetch is very nearly every object in the store.
/// Keying them by `Vec<u8>` is one heap allocation per entry per request —
/// 12 112 of them for a 69-object fetch, measured on oden 2026-08-15 — for a
/// question (*does the client hold this base?*) that never looks at more than
/// `oid_len` bytes.
///
/// Padded to the widest hash rather than parameterised by width: a store holds
/// exactly one hash kind, so every key in one set has the same real length and
/// two different oids can never collide through the padding. The length is
/// carried anyway, because a key that silently compared equal across widths is
/// the kind of wrong that only shows up in a mixed-hash mirror.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
struct OidKey {
    bytes: [u8; 32],
    len: u8,
}

impl OidKey {
    /// The key for `oid`, or a refusal for one wider than any git hash.
    fn new(oid: &[u8]) -> Result<Self> {
        let len = oid.len();
        if len == 0 || len > 32 {
            bail!("{len} is not the width of any git object id (1..=32 bytes)");
        }
        let mut bytes = [0u8; 32];
        bytes[..len].copy_from_slice(oid);
        Ok(OidKey {
            bytes,
            len: len as u8,
        })
    }

    /// The oid back, at its real width.
    fn as_slice(&self) -> &[u8] {
        &self.bytes[..self.len as usize]
    }
}

impl<S: ObjectIndex + 'static> GitStore<S> {
    /// **The one constructor**, with every arm named.
    ///
    /// `S` picks the [`ObjectIndex`] layout at compile time; `arms` picks the
    /// [`ArchiveWrite`](crate::archive_write::ArchiveWrite) and [`Gc`]
    /// implementations at run time. [`open`](GitStore::open) and
    /// [`open_with`](GitStore::open_with) are this function with
    /// [`StoreConfig::DEFAULT`], which is why nothing about them moved.
    ///
    /// `arms.index` is **not** consulted here — `S` is the index arm, and a
    /// value that disagreed with the type would be a second source of truth.
    /// [`open_from_env`] is the one place the two are tied together, and it does
    /// it by choosing `S` from the value.
    ///
    /// ```no_run
    /// use znippy_plugin_git::arms::{GcArm, StoreConfig, WriterArm};
    /// use znippy_plugin_git::index_layout::PackedPayload;
    /// use znippy_plugin_git::{GitHashKind, GitStore};
    ///
    /// // A rebuildable import mirror: the ceiling writer, the packed payload
    /// // layout, and an in-place compaction.
    /// let arms = StoreConfig::DEFAULT
    ///     .with_writer(WriterArm::Fast)
    ///     .with_gc(GcArm::CompactInPlace);
    /// let store = GitStore::<PackedPayload>::open_with_arms(
    ///     std::path::Path::new("/srv/repo"),
    ///     "rickard",
    ///     GitHashKind::Sha1,
    ///     arms,
    /// )?;
    /// # Ok::<(), anyhow::Error>(())
    /// ```
    pub fn open_with_arms(
        root: &Path,
        account: &str,
        hash: GitHashKind,
        arms: StoreConfig,
    ) -> Result<Self> {
        std::fs::create_dir_all(root)
            .with_context(|| format!("creating the store root {}", root.display()))?;
        let blobs = root.join("objects.pack");
        let writer = arms.writer.create(&blobs)?;
        // `None` for an arm that keeps no durable ack log — see
        // [`crate::arms::WriterArm::journal`]. Everything below that touches the
        // journal is written against that `Option` rather than assuming one.
        let journal = arms.writer.journal(&blobs);
        // **The redb ceiling**, carried in on the config rather
        // than read here: `StoreConfig::from_env` reads it with the arms, and
        // `open`/`open_with` put the environment's value on `DEFAULT` before
        // getting here, so every constructor still honours the variable and
        // `arms.to_string()` prints what was actually applied. Not an arm — see
        // [`crate::arms::redb_cache_bytes`] — redb's own default is 1 GiB *per
        // database*, and this store opens exactly ONE of them, the
        // `ObjectReadStack` below. The `let _ = cache_bytes` two lines down is
        // what is left of the second.
        let cache_bytes = arms.redb_cache_bytes;
        let objects = ObjectReadStack::<S>::open(
            &root.join("objects.tail"),
            RebuildTriggers::default(),
            cache_bytes,
        )?;
        let _ = cache_bytes; // no page cache to size: the blob is pread, the index is a slice
        // A store written before 2026-09-04 carries `objects.ordinal`, the memo
        // of the sorted oid list the read stack's persisted index now IS. Its
        // bytes are never read again, so it goes on the first open by this
        // build — 278 964 902 B on the kernel mirror. Absent is the normal case
        // and not an event; anything else that stops the delete is logged, not
        // fatal, because the store does not depend on the file being gone.
        match std::fs::remove_file(root.join(ORDINAL_CACHE_FILE)) {
            Ok(()) => tracing::info!(
                target: "znippy.git",
                path = %root.join(ORDINAL_CACHE_FILE).display(),
                "removed a stale objects.ordinal — the persisted index's oid column is the \
                 ordinal space now"
            ),
            Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
            Err(e) => tracing::warn!(
                target: "znippy.git",
                path = %root.join(ORDINAL_CACHE_FILE).display(),
                error = %e,
                "could not remove a stale objects.ordinal; it is dead weight, not a fault"
            ),
        }
        let exploded = Arc::new(ExplodedCache::open(
            &root.join("objects.exploded"),
            crate::exploded_arrow::ExplodePolicy::from_env()?,
            crate::exploded_cache::budget_from_env(root)?,
        )?)
        .registered();
        // **One descriptor on the blob file, not three.** The writer opened it
        // read/write and `flock`ed it; the absorber's `pread` fallback and the
        // index workers' `read_exact_at` both address it positionally, so they
        // take the same open file description rather than two more of their
        // own. 5 descriptors per open store → 3, which is the divisor of
        // gunnar's `open_repo_budget` (`tests/fd_per_store.rs`).
        let absorber = Arc::new(Absorber::<S>::open(
            &blobs,
            writer.blobs_shared(),
            hash,
            objects,
            exploded,
        )?);

        // **The `indexed` bit, derived.** Every extent this archive has ever
        // acked, diffed against the rows the index holds. `SafeWriter::create`
        // above appends to that journal and never truncates it, which is what
        // makes the older extents still be here to diff against.
        //
        // An arm with no journal has nothing to derive the bit from, and that is
        // the honest consequence of choosing it rather than a gap: a
        // `FastWriter` store has no durable record that a pack was acked, so a
        // reopen re-queues nothing and there is nothing for it to re-queue.
        let acked = match journal.as_deref() {
            Some(p) if p.exists() => crate::archive_write::read_journal(p).with_context(|| {
                format!(
                    "reading {} — it is the durable half of the indexed bit and a store \
                         cannot be opened without knowing which packs it owes index work for",
                    p.display()
                )
            })?,
            _ => Vec::new(),
        };
        let requeue = absorber.adopt_journal(&acked)?;

        // **The wiring.** The absorber goes into the push path, which hands it to
        // every account indexer it starts, so the drain behind `push_pack` ends
        // in object rows instead of one pack row. Nothing about the ack changes:
        // `append` still returns after the two fsyncs and this is all downstream
        // of that (§13.9).
        let push = PushPath::with_absorber(
            writer,
            &blobs,
            journal,
            absorber.clone() as Arc<dyn ObjectAbsorb>,
        )?;
        let refs = RefLog::new(root.join("refs.log"));
        let archive = root.join("repository.znippy");

        let store = Self {
            root: root.to_path_buf(),
            blobs,
            archive,
            push,
            account: account.to_string(),
            hash,
            absorber,
            refs,
            ref_gate: Mutex::new(()),
            gc: arms.gc.create(),
            arms,
        };
        // The interrupted work, back on the channel it fell off. Nothing reads
        // here and nothing blocks: this is the same 24-byte handoff a push makes,
        // and the drain picks it up on its own thread (§13.9-11).
        let indexer = store.push.indexer(&store.account);
        for p in &requeue {
            indexer.submit(IndexJob {
                pack_id: p.pack_id,
                offset: p.offset,
                len: p.len,
            })?;
        }
        // A store reopened over an existing tail already knows its objects; the
        // derived tables are rebuilt from them rather than from a side-car,
        // because a side-car is a second truth.
        store.refold()?;
        Ok(store)
    }

    /// The arms this store was built with. A description of what was
    /// constructed, not a switch — nothing on a serving path reads it.
    pub fn arms(&self) -> StoreConfig {
        self.arms
    }

    /// The selected writer's name, as it goes on a bench row.
    pub fn writer_name(&self) -> &'static str {
        self.push.name()
    }

    /// What the selected writer's `append` actually promises. Printed next to a
    /// throughput, because a throughput without it is not a comparison.
    pub fn writer_durability(&self) -> &'static str {
        self.push.durability()
    }

    /// The selected [`Gc`] arm.
    pub(crate) fn gc_arm(&self) -> &dyn Gc {
        self.gc.as_ref()
    }

    /// Where the verbatim packs are.
    pub fn blobs_path(&self) -> &Path {
        &self.blobs
    }

    /// The sealed archive GC operates on.
    pub fn archive_path(&self) -> &Path {
        &self.archive
    }

    pub fn root(&self) -> &Path {
        &self.root
    }

    pub fn hash_kind(&self) -> GitHashKind {
        self.hash
    }

    /// The `objects` table, for a caller that wants the index directly (the
    /// bench does).
    pub fn index(&self) -> &ObjectReadStack<S> {
        &self.absorber.objects
    }

    /// Objects in the `objects` table.
    pub fn object_count(&self) -> usize {
        self.absorber.objects.len()
    }

    /// Commits in the graph.
    pub fn commit_count(&self) -> usize {
        self.absorber
            .derived
            .read()
            .expect("derived lock")
            .graph
            .len()
    }

    /// `(graph_read_from_disk, graph_recomputed)` since this store was opened —
    /// how many folds read the persisted `objects.graph` vs re-derived it.
    pub fn graph_source_stats(&self) -> (u64, u64) {
        self.absorber.graph_source_stats()
    }

    /// `(reach_carried, reach_rebuilt)` since this store was opened — how many
    /// folds carried the reachability table across (renumbered, no tree
    /// opened) vs how many reads built it from scratch. The witness for the
    /// clone-after-push lever; see `REACH_CARRY_MAX_FRACTION`.
    pub fn reach_source_stats(&self) -> (u64, u64) {
        self.absorber.reach_source_stats()
    }

    /// `(folded_rows, tail_rows)` of the ordinal space the last fold built — how
    /// many oids came off the persisted index's oid column and how many off a
    /// scan of the redb tail. Replaces S-082's `(read, recomputed)`: there is no
    /// memo to read any more, the index IS the space.
    pub fn ordinal_source_stats(&self) -> (u64, u64) {
        self.absorber.ordinal_source_stats()
    }

    /// The folded ordinal space — every oid in ordinal order, hex. For the tests
    /// that assert a memo read reproduces the scanned space exactly (same oids,
    /// same order, same ordinals — the ordinal is the index into this vector).
    ///
    /// Formatted on demand from `oids_raw`: there is no resident hex `Vec` any
    /// more (S-082), and a test that wants hex encodes the flat buffer here
    /// rather than the store carrying a second copy for it.
    #[cfg(test)]
    pub(crate) fn ordinal_oids(&self) -> Vec<String> {
        let oid_len = self.hash.oid_len();
        self.absorber
            .derived
            .read()
            .expect("derived lock")
            .oids_raw
            .chunks_exact(oid_len)
            .map(hex::encode)
            .collect()
    }

    /// **One object's resolved content** — §14's exploded table, in one lookup.
    ///
    /// Not one of the twelve and deliberately not on [`GitOps`]: the twelve are
    /// the wire vocabulary and they hand back **stored** bytes ([`GitOps::get`]),
    /// which for a delta entry is a delta. This is the other question — *what
    /// does this object contain* — and it is the one §14's table exists to make
    /// a point lookup instead of a chain walk.
    ///
    /// It absorbs any pending pack first, for the same reason
    /// [`GitOps::has`] does: a store with durable bytes it has not indexed cannot
    /// answer "absent" without lying.
    pub fn content(&self, oid: Oid<'_>) -> Result<Option<(GitObjectKind, Vec<u8>)>> {
        // No absorb here any more: a pack the worker has not reached is read
        // for this one oid by `Absorber::resolved`'s third arm — on this thread,
        // from the mmapped pack, without waiting for the worker.
        self.absorber.resolved(oid)
    }

    /// **One object, with the bill for producing it.**
    ///
    /// [`Self::content`] plus what the read cost: `None` for a cost when the
    /// exploded pools answered (nothing was rebuilt), `Some(cost)` when the
    /// object was inflated out of the pack. What a measurement needs and a
    /// server does not — the cache admission decision is made inside, not by
    /// the caller.
    pub fn content_costed(
        &self,
        oid: Oid<'_>,
    ) -> Result<Option<(GitObjectKind, Vec<u8>, Option<crate::resolve::ReadCost>)>> {
        self.absorber.resolved_costed(oid)
    }

    /// §14's table, counted: rows, rows written, reads it served, reads that fell
    /// through to re-deriving from the verbatim truth.
    ///
    /// Applied output. Both content paths return identical bytes, so `served`
    /// against `rederived` is the only thing that can say which one ran.
    pub fn exploded_stats(&self) -> ExplodedStats {
        self.absorber.exploded.engine_stats()
    }

    /// **What the exploded tier is holding and what it is allowed to hold** —
    /// the row a dashboard prints (`explode 3.1G/8G`) and `Admin.StorageInfo`
    /// carries, per repository.
    pub fn explode_cache_use(&self) -> crate::exploded_cache::CacheUse {
        self.absorber.exploded.use_now()
    }

    /// Rows in the EAGER pool only — the graph the push wrote, with nothing a
    /// reader has since cached. What `adopt_journal` compares against the index
    /// and what a test asking "did the push explode the graph" must ask,
    /// because [`Self::exploded_stats`]'s `rows` counts both pools and a read
    /// moves it.
    pub fn exploded_eager_rows(&self) -> Result<u64> {
        self.absorber.exploded.eager_rows()
    }

    /// **Set this repository's explode budget** — what the control plane calls
    /// when an account's quota changes. Takes effect at the next rotation; it
    /// never refuses a read, never refuses a write and never deletes on the
    /// caller's thread. A bigger budget buys a bigger cache and therefore faster
    /// first reads; a smaller one costs re-inflation, never an error.
    pub fn set_explode_budget(&self, bytes: u64) {
        self.absorber.exploded.set_budget(bytes);
    }

    /// Flush the lazy pool, rotate it and evict it back under budget — on
    /// **this** thread. What a test and a `gc` call; a reader never does, it
    /// starts a write-back thread and walks away.
    pub fn explode_cache_maintain(&self) -> Result<u64> {
        self.absorber.exploded.maintain()
    }

    /// **`repo gc` reclaims the cache**: every lazy segment unlinked, the graph
    /// pool kept. Returns the bytes that left the filesystem.
    pub fn drop_explode_cache(&self) -> Result<u64> {
        self.absorber.exploded.drop_cache()
    }

    /// Where §14's table lives. Deleting this file while no store holds it is the
    /// sanctioned way to drop it; the next open re-explodes every pack.
    pub fn exploded_path(&self) -> &Path {
        self.absorber.exploded.path()
    }

    /// Every live row of one kind, with its payload — what the graph fold reads,
    /// exposed so a measurement can ask **the table** how many bytes it holds
    /// rather than dividing a file size by what the pack claimed.
    pub fn exploded_of_kind(&self, kind: GitObjectKind) -> Result<Vec<(Vec<u8>, Vec<u8>)>> {
        self.absorber.exploded.of_kind(kind)
    }

    /// The commit graph as it stands, generations included. A snapshot: the
    /// caller holds no lock and the next fold replaces it wholesale.
    pub fn graph_snapshot(&self) -> Vec<CommitNode> {
        self.absorber
            .derived
            .read()
            .expect("derived lock")
            .graph
            .clone()
    }

    /// Packs whose objects are not in the index yet — §13.12's bit, counted.
    pub fn unindexed_packs(&self) -> usize {
        self.absorber.unindexed_packs()
    }

    /// This account's indexer: the thread that drains the channel a push queues
    /// on. `is_indexed(pack_id)` on it is §13.12's bit at the pack level, and it
    /// is set only once that pack's **objects** are in the index.
    pub fn indexer(&self) -> std::sync::Arc<crate::indexer::AccountIndexer> {
        self.push.indexer(&self.account)
    }

    /// Block until the background drain has absorbed everything this store has
    /// pushed. What a maintenance tick, a seal or a test uses instead of racing
    /// it.
    pub fn wait_indexed(&self) {
        self.push.pool().wait_caught_up();
    }

    /// **Stop this store's index work — the repository is being deleted.**
    ///
    /// Sets the cancel flag the resolver polls every few thousand objects and
    /// the drain checks before every pack. The pack in flight is **aborted at
    /// its next checkpoint** (milliseconds), not finished: its rows are keyed
    /// by oid and the exploded tier is droppable, and the whole directory is
    /// about to be removed anyway. Every pack still queued is skipped. After
    /// this the store can be dropped without the drop joining a worker that is
    /// hours into a kernel pack — which is what `Drop for AccountIndexer` would
    /// otherwise do.
    pub fn cancel_index(&self) {
        self.absorber.cancel.store(true, std::sync::atomic::Ordering::SeqCst);
        self.push.pool().cancel_all();
    }

    /// Put a different memory watchdog under this store's worker — a test's
    /// fake `/proc/meminfo`. Takes effect from the next pack.
    pub fn set_mem_watchdog(&self, dog: Arc<crate::memguard::MemWatchdog>) {
        *self.absorber.watchdog.write().unwrap_or_else(|e| e.into_inner()) = dog;
    }

    /// Has [`cancel_index`](Self::cancel_index) been called?
    pub fn index_cancelled(&self) -> bool {
        self.absorber.cancel.load(std::sync::atomic::Ordering::SeqCst)
    }

    /// This store's index progress counter, for a dashboard or a stats call.
    pub fn index_progress(&self) -> crate::progress::Snapshot {
        self.absorber.progress.snapshot()
    }

    /// **The indexer's half of the split**, and not one of the twelve.
    ///
    /// Runs on the account indexer's worker (§13.9), and on a read that arrives
    /// before that worker got there — the same function either way, which is why
    /// there is no such thing as a background index that behaves differently from
    /// the fallback one.
    pub fn absorb_pending(&self) -> Result<usize> {
        self.absorber.absorb_pending()
    }

    /// **Make every durable object addressable, on this thread, now** —
    /// [`git_storage_trait::GitServe::index_now`]'s implementation and the ack
    /// path's one call. See [`Absorber::index_pending_now`] for why the
    /// connectivity check is the only caller allowed to wait here.
    pub fn index_now(&self) -> Result<usize> {
        self.absorber.index_pending_now()
    }

    /// **The packs this store has stopped trying to index**, with the reason
    /// each one stopped for.
    ///
    /// Empty is the ordinary answer and the only good one. A non-empty list is
    /// a repository that is missing objects whose bytes are durable, and it is
    /// the fact an operator needs before wondering why a push was refused or a
    /// standby is behind — see [`MAX_ABSORB_ATTEMPTS`].
    pub fn unindexable_packs(&self) -> Vec<Unindexable> {
        self.absorber.unindexable_packs()
    }

    /// **The object ids a stored pack introduced**, and not one of the twelve.
    ///
    /// [`GitOps::put_pack`] cannot answer this and never will: the walk it does
    /// reads pack *entries*, and a pack entry does not carry its oid — that is
    /// why [`crate::pack_walk::Closure`] can only name the `REF_DELTA` bases and
    /// says so. The oids exist once the indexer has absorbed the pack, so this
    /// absorbs first rather than racing the drain, and is therefore the one
    /// caller that deliberately pays §13.9's deferral back.
    ///
    /// **Only call it when something needs the set.** A push whose policy has no
    /// gate over objects must not pay this, which is the whole reason it is a
    /// separate call and not a field on `TxId`.
    ///
    /// The extent *is* the identity: §14 stores a pack verbatim as one
    /// contiguous append, so "introduced by this pack" and "indexed at an offset
    /// inside this pack's extent" are the same set, with no pack column to keep
    /// in step. Bases the push deltas against live in *other* extents and are
    /// excluded by construction — which is what a thin push needs, and what
    /// reading a written `.idx` back could never do.
    ///
    /// Cost is one pass over the index, so it is O(objects in the repository),
    /// not O(objects pushed). Recorded rather than hidden: a per-pack ordinal
    /// range would make it O(pushed), and needs the indexer to keep one.
    pub fn oids_in_extent(&self, extent: Extent) -> Result<Vec<Vec<u8>>> {
        let (start, len) = extent;
        // **Never waits on the indexer.** If the gate is free this thread
        // absorbs the pending packs (the same work, done now, by the caller
        // that needs it); if the worker holds it, the extent is resolved
        // read-only on this thread and its oids are the answer — the same
        // resolver, no sink, no rows written.
        if self.try_absorb_pending()?.is_none() {
            let snap = self.archive_snapshot()?;
            let bytes = self.extent(&snap, start, len)?;
            let walked = crate::pack_walk::walk(&bytes, self.hash.oid_len())?;
            let mut oids: Vec<Vec<u8>> = Vec::with_capacity(walked.entries.len());
            let limits = crate::resolve::ResolveLimits {
                base_budget: (crate::resolve::base_budget_from_env() / 4).max(1),
                ..crate::resolve::ResolveLimits::none()
            };
            crate::resolve::resolve_streaming(
                &bytes,
                &walked,
                self.hash,
                start,
                &*self.absorber,
                &crate::exploded::NoSink,
                limits,
                &mut |rows| {
                    oids.extend(rows.drain(..).map(|r| r.oid));
                    Ok(())
                },
            )
            .context("resolving the pushed pack read-only to enumerate its objects")?;
            return Ok(oids);
        }
        let end = start.saturating_add(len);
        let all = self.index().oids_in_order()?;
        let refs: Vec<&[u8]> = all.iter().map(Vec::as_slice).collect();
        let extents = self.index().extents_batch(&refs);
        Ok(all
            .into_iter()
            .zip(extents)
            .filter_map(|(oid, at)| match at {
                Some((offset, _)) if offset >= start && offset < end => Some(oid),
                _ => None,
            })
            .collect())
    }

    /// **Everything needed to emit a pack for EXACTLY `oids`** — nothing added,
    /// nothing dropped.
    ///
    /// Not one of the twelve. It sits beside [`absorb_pending`](Self::absorb_pending)
    /// for the same reason those do: it is about how storage is *shaped*, not
    /// about what a caller stores or reads.
    ///
    /// # 🔴 It used to add the delta bases, and that shipped broken clones
    ///
    /// Until 2026-08-11 this closed the request over its delta bases: an
    /// `OFS_DELTA` names its base by position, so a base outside the set cannot
    /// be encoded, and *adding the base* looked like the answer that is always
    /// safe. It is not, and the failure is silent on the server:
    ///
    /// ```text
    /// git clone --bare --single-branch --branch base <url>
    ///     fatal: did not receive expected object 8601ec33920b7d701c9887fa04916501136d6e90
    ///     fatal: fetch-pack: invalid index-pack output
    /// event="git.upload_pack.served" objects=33773      ← the server said SUCCESS
    /// ```
    ///
    /// Measured on the `h2h-linear-sha1-2048c-1024f-16k` fixture: `base` reaches
    /// **31 805** objects, the delta-base closure added **1 968** more — 33 773,
    /// exactly the count the server logged — and 507 of those additions were
    /// **trees**. A tree that is not reachable from the wants drags its own
    /// children in as a *requirement*, because `git index-pack
    /// --check-self-contained-and-connected` (which is what a clone runs, and
    /// which sets `strict`) walks every received object's links and then demands
    /// each one exist. Those 507 trees named **204** objects the pack did not
    /// contain. `8601ec33…` is one of them.
    ///
    /// So a delta base pulled into a narrowed request is not free: it satisfies
    /// the *pack format* and violates *connectivity*. Closing over the added
    /// bases' children in turn does terminate — the same fixture reaches a fixed
    /// point at 34 534 objects — but it is an over-send by construction, and on
    /// a `--filter=blob:none` fetch it adds back exactly the blobs the filter
    /// excluded, which is the defect
    /// `emit_pack_emits_exactly_the_set_it_is_given_and_never_its_closure`
    /// exists to forbid.
    ///
    /// # What it does instead: the base decides how the entry is encoded
    ///
    /// The set is never widened. For each requested object:
    ///
    /// * **base inside the request** → the stored entry is copied byte for byte,
    ///   which is the whole point of this engine and is what a full clone does
    ///   for every single object;
    /// * **base outside the request** → the entry is **rebuilt** from §14's
    ///   resolved content and marked
    ///   [`EmitEntry::recompressed`](crate::pack_walk::EmitEntry::recompressed) —
    ///   as a computed delta against a base the request *does* carry where one
    ///   can be found (see [`Self::in_request_chain_ancestor`] and
    ///   [`crate::delta`]), and whole where it cannot.
    ///
    /// That is what stock `pack-objects` does — it reuses a stored delta only
    /// when the base is also being packed — and it keeps the receipt a fact:
    /// `copied` and `recompressed` add up to the objects written, and
    /// `recompressed` is **0 for a whole-repository clone** because such a
    /// selection contains every base. The cost lands on the boundary a narrowed
    /// request cuts through and nowhere else: 1 968 of 33 773 entries, 5.9 %, on
    /// the fixture above.
    ///
    /// `ofs_delta_ok` is the client's `ofs-delta` capability. An old client that
    /// cannot parse one is not refused any more: the entry is re-headed as a
    /// `REF_DELTA` naming its base by oid, **carrying the same compressed delta
    /// payload**, so that arm stays a copy too. Only a base outside the request
    /// costs a rebuild.
    ///
    /// # `thin_haves` — the boundary entry that does NOT have to be rebuilt
    ///
    /// `Some(tips)` says the client consented to a thin pack *and* the
    /// negotiation found common ground; [`crate::serve::GitStore::emit_oids`] is
    /// where the two halves are required together, because either alone is a
    /// failed fetch rather than a smaller one. Given it, a base outside the
    /// request is checked against [`the client's own objects`](GitStore::client_bases),
    /// and one the client already holds is named by oid in a `REF_DELTA` that
    /// carries the **stored delta stream unchanged**. That entry counts as
    /// `copied`, because it is: nothing was inflated and nothing was deflated.
    ///
    /// `None` — a clone, or a client that did not ask for a thin pack — behaves
    /// exactly as before, and a **full clone is byte-identical either way**,
    /// because a whole-repository request has no base outside itself for any of
    /// this to apply to.
    ///
    /// It does nothing for the *clone* half of the boundary cost, and cannot: a
    /// clone's receiver holds nothing, so there is no external base to name.
    /// That half is the arm below it — [`crate::delta`] computes a delta against
    /// something the pack **does** carry, which is the only move left when there
    /// is no receiver to lean on. The two do not overlap: `thin_haves` reaches a
    /// boundary entry first and costs nothing when it fires, so a fetch stays a
    /// pure copy and only a clone pays for a computed delta.
    ///
    /// # An oid this repository does not hold is REFUSED
    ///
    /// It used to be skipped, which is the same silent-under-send shape as the
    /// closure defect above wearing a different hat: N asked, N-1 emitted,
    /// `PackStats` reporting success, and the client the first to know. gunnar's
    /// in-memory arm has always refused it, so the skip also meant the two
    /// engines behind one contract disagreed about what a missing object means.
    /// A partial-clone filter does not depend on the skip and cannot: a filter
    /// *removes* oids from the selection, it never adds one the store lacks.
    ///
    /// # Cost
    ///
    /// O(objects requested). The previous shape read `oids_in_order()` and
    /// `lookup_batch` over the **whole repository** on every call, to build an
    /// offset→oid map it needed only for the closure; with no closure the
    /// selection's own offsets answer "is this base in the request", so that
    /// scan is gone and no reverse index has to exist.
    pub fn emit_set(
        &self,
        oids: &[Oid<'_>],
        ofs_delta_ok: bool,
        thin_haves: Option<&[Oid<'_>]>,
    ) -> Result<Vec<crate::pack_walk::EmitEntry>> {
        use crate::index_layout::ObjType;
        use std::collections::hash_map::Entry;

        // If the gate is free, absorb now; if the worker holds it, the rows
        // come from the index plus a read-only pass over the pending packs
        // (`lookup_batch_never_wait` below). Never a wait.
        let _ = self
            .try_absorb_pending()
            .context("absorbing pending index work before emitting a pack")?;

        // Deduplicate, preserving the caller's order. A repeated oid in the
        // request must not become a repeated entry in the pack.
        //
        // **A map and not a set**, keyed oid → position in `asked` (which is the
        // position in `rows`). The membership question is the one it always was;
        // what the position buys is the base's *archive offset* for a
        // `REF_DELTA`, which names its base by oid and can learn where that base
        // sits no other way. It is a `usize` per requested object on a vector
        // this function already builds — not a second pass and not a second
        // index lookup.
        let mut asked: Vec<&[u8]> = Vec::with_capacity(oids.len());
        let mut seen: HashMap<&[u8], usize> = HashMap::with_capacity(oids.len());
        for oid in oids {
            if let Entry::Vacant(slot) = seen.entry(*oid) {
                slot.insert(asked.len());
                asked.push(*oid);
            }
        }

        let rows = self.lookup_batch_never_wait(&asked)?;

        // **One mapping for the whole emit set.** Only the header probes below
        // read through it — the entries themselves are recorded as extents and
        // resolved by [`Self::resolve_emit_payloads`], which takes its own
        // snapshot of the same [`ArchiveMap`](crate::archive_map::ArchiveMap) and
        // therefore the same `Arc` unless a push landed in between.
        let snap = self.archive_snapshot()?;

        // The two questions "is this base in the request" are asked in the two
        // coordinate systems the two delta forms use, and both are answered off
        // the request itself rather than off the index.
        //
        // **Offset → position in the request**, not a bare set of offsets. The
        // membership question is the same one it always was; what the position
        // buys is the base's *oid*, which the re-delta arm below needs in order
        // to read the base's content. It is the same single pass over `rows`
        // that built the set, so it costs a `usize` per requested object and no
        // extra work — and it is emphatically not the whole-repository
        // offset→oid scan the old delta-base closure used to do.
        let in_set_offsets: HashMap<u64, usize> = rows
            .iter()
            .enumerate()
            .filter_map(|(i, r)| r.as_ref().map(|r| (r.offset, i)))
            .collect();
        let oid_len = self.hash_kind().oid_len();

        // The base of the last boundary entry that was re-deltified, kept so a
        // run of entries hanging off one ancestor inflates it once. One slot and
        // not an LRU: measured on `h2h-linear-sha1-2048c-1024f-16k`, 961
        // boundary entries name 855 distinct stored bases, so there is no reuse
        // worth a cache that could hold a repository's worth of objects — and
        // this crate is under a live RSS budget.
        let mut base_content: Option<(u64, Vec<u8>)> = None;

        // What the client already holds, built **lazily and at most once**. A
        // clone never asks for it (`thin_haves` is `None`), and a fetch every one
        // of whose bases is inside the request never reaches the line that builds
        // it either — only a request that actually cuts a delta chain pays, which
        // is the only request it can help.
        let mut client_bases: Option<ExternalBases> = None;

        let mut out = Vec::with_capacity(asked.len());
        for (oid, row) in asked.iter().zip(rows.iter()) {
            // **An oid this repository does not hold is REFUSED, not dropped.**
            //
            // This used to `continue`, and that is the same silent-under-send
            // shape as the closure defect above: the request asks for N objects,
            // the pack carries N-1, the receipt says success and the client is
            // the first to find out. It also made the two engines disagree about
            // what a missing object *means* — gunnar's in-memory arm has always
            // said *"emit_pack was asked for X, which this store does not hold;
            // a pack emitter that skipped it would send a pack no client can
            // close"* — so a benchmark across the two was comparing two
            // contracts.
            //
            // No legitimate caller relies on the skip. Every set that reaches
            // here is built out of this store's own index: `select` answers from
            // the reachability bitmaps, and gunnar's own walk resolves each id
            // against the store before it selects it. A partial-clone filter is
            // no exception and is the case worth naming, because it is the one
            // that *looks* like it might be: `--filter=blob:none` **removes**
            // oids from the selection, it never adds one the store lacks, so a
            // filtered request is a smaller set of objects that are all present
            // and this refusal never fires on it.
            let Some(row) = row.as_ref() else {
                bail!(
                    "this pack was asked for {}, which this repository does not hold; emitting \
                     the rest would be a short pack reported as a success, so this refuses \
                     instead",
                    hex::encode(oid)
                );
            };

            // Which base this entry names, if any, and whether the request holds
            // it. A `REF_DELTA`'s `delta_base` column is 0 by construction —
            // `DeltaBase::as_offset` reports 0 for a ref base and 0 is also the
            // no-base sentinel — so its base is read out of the entry's own
            // bytes, directly after the type/size varint.
            //
            // A base that is *outside* the request is carried in **both**
            // coordinate systems rather than as a bare `false`, because that is
            // what the thin arm below needs: an `OFS_DELTA` knows only where its
            // base sits, a `REF_DELTA` knows only what it is called, and
            // "does the client hold it" has to be asked in whichever one the
            // entry speaks.
            let base = match row.obj_type {
                ObjType::OfsDelta => match in_set_offsets.get(&row.delta_base) {
                    Some(_) => BaseOf::Inside {
                        at: row.delta_base,
                    },
                    None => BaseOf::Outside {
                        offset: row.delta_base,
                        oid: None,
                    },
                },
                ObjType::RefDelta => {
                    let head = self.extent(&snap, row.offset, row.len.min(64))?;
                    let (_, _, n) = crate::pack_walk::type_and_size_of(&head)?;
                    match head.get(n..n + oid_len) {
                        // In the request — **and where**. `rows[i]` is `None`
                        // only for an oid this store does not hold, which the
                        // refusal above turns into an error the moment the loop
                        // reaches it; until then the base is treated as outside,
                        // which is the answer that cannot be wrong.
                        Some(b) => match seen
                            .get(b)
                            .and_then(|i| rows[*i].as_ref())
                            .map(|r| r.offset)
                        {
                            Some(at) => BaseOf::Inside { at },
                            None => BaseOf::Outside {
                                offset: 0,
                                oid: Some(b.to_vec()),
                            },
                        },
                        // A truncated entry names no base this can read. It is
                        // not "inside", and it is not offered to the client
                        // either — it falls through to the whole rebuild, which
                        // is the answer that cannot be wrong.
                        None => BaseOf::Outside {
                            offset: 0,
                            oid: None,
                        },
                    }
                }
                _ => BaseOf::Whole,
            };

            if let BaseOf::Whole | BaseOf::Inside { .. } = base {
                let inside_at = match base {
                    BaseOf::Inside { at } => Some(at),
                    _ => None,
                };
                // **The whole-clone arm, and it reads not one byte here.** The
                // entry is recorded as its address in the archive; the bytes are
                // resolved against the mapping at emit time and go straight from
                // the page cache to the wire. This line used to be a `pread`
                // into a fresh `Vec` — 13.8 M of each on a `linux.git` clone, all
                // of them held at once, which is the 2314 MB peak RSS
                // [`crate::pack_walk::EntryBytes`] documents.
                let extent = crate::pack_walk::EntryBytes::Extent {
                    offset: row.offset,
                    len: row.len,
                };
                let (stored, obj_type, delta_base) = match (row.obj_type, inside_at) {
                    // An `OFS_DELTA` for a client that cannot read one: re-head it
                    // as a `REF_DELTA` naming the same base by oid. The compressed
                    // payload is the delta stream either way, so this is still a
                    // copy in the `recompressed` sense — nothing is inflated — but
                    // the concatenation of a new header and an old payload exists
                    // nowhere on disk, so it cannot be an extent and is `Owned`.
                    (ObjType::OfsDelta, Some(_)) if !ofs_delta_ok => {
                        let base_oid = self.oid_at_offset(row.delta_base, &asked, &rows)?;
                        let bytes = self.extent(&snap, row.offset, row.len)?;
                        (
                            crate::pack_walk::EntryBytes::Owned(Self::as_ref_delta(
                                &bytes, &base_oid,
                            )?),
                            ObjType::RefDelta,
                            0,
                        )
                    }
                    // 🚨 **A stored `REF_DELTA` whose base is IN THIS PACK is
                    // re-headed as an `OFS_DELTA`, and that is not an
                    // optimisation.**
                    //
                    // gitoxide — which is `gunnar_client`, `gunnar`'s own
                    // `receive-pack`, and every other gix-based receiver — reads
                    // every `OBJ_REF_DELTA` in an incoming pack as naming an
                    // object *the receiver already has*.
                    // `gix_pack::data::input::LookupRefDeltaObjectsIter` consults
                    // the local object database and the bases it has already
                    // spliced in, and **nowhere else**: never inside the pack it
                    // is reading. A base that is in the pack and not in the
                    // receiver's store is `input::Error::NotFound`, and the fetch
                    // dies with `The object <base> could not be decoded or wasn't
                    // found` — with the whole pack rejected, on a clone whose
                    // target is empty by definition, so *every* such entry is
                    // fatal. That is `S-003` / gitoxide#2882, and `gunnar-wire`'s
                    // `Selection::client_has` carries the same rule for gunnar's
                    // own emitter.
                    //
                    // Stock git resolves it happily, which is why this survived
                    // five green `git clone` runs against the store that could
                    // not serve one `gunnar_client` fetch (`gunnar.rewrite`
                    // round 0 and `gunnar.year` week 0, all four znippy columns,
                    // 2026-08-14 — the server logged the same request `served`).
                    //
                    // The entry keeps its compressed delta stream byte for byte;
                    // only the header changes, from "base named by oid" to "base
                    // named by distance", which is the naming every receiver in
                    // existence resolves inside the pack. This is where such an
                    // entry comes from in the first place: `git index-pack
                    // --fix-thin` appends the base into the pack and leaves the
                    // ref-delta naming it by oid, so any repository pushed to
                    // more than once holds them.
                    //
                    // # What it costs, stated rather than glossed
                    //
                    // `Owned`, so this is the one arm that gives up the zero-copy
                    // property for an entry it does **not** rebuild: the
                    // compressed delta stream is memcpy'd once, because a new
                    // header concatenated with an old payload exists nowhere on
                    // disk and no extent can address it. Bounded by the
                    // compressed size of the ref-deltas whose base is in the
                    // request — not by the repository, and not by the pack.
                    //
                    // It could be avoided: `emit_pack` re-encodes every header
                    // from `EmitEntry::obj_type` while `header_len` finds the
                    // payload from the STORED bytes' own type, so an `Extent`
                    // over the untouched ref-delta declared as an `OfsDelta`
                    // would emit the identical pack with no copy at all. That is
                    // deliberately not done: it makes an `EmitEntry` disagree
                    // with the bytes it points at, and the entry model being
                    // trustworthy is worth more than the memcpy. If the RSS
                    // budget ever says otherwise, the honest form is a third
                    // `EntryBytes` variant that carries the new header beside
                    // the extent, not a silent disagreement.
                    (ObjType::RefDelta, Some(at)) if ofs_delta_ok => {
                        let bytes = self.extent(&snap, row.offset, row.len)?;
                        (
                            crate::pack_walk::EntryBytes::Owned(Self::as_ofs_delta(
                                &bytes, oid_len, row.offset, at,
                            )?),
                            ObjType::OfsDelta,
                            at,
                        )
                    }
                    // Everything else copies. A `REF_DELTA` left as one here is
                    // the `!ofs_delta_ok` client, which cannot be handed a
                    // distance at all — and which is stock git older than 1.4.4,
                    // never a gix receiver, since gix has always advertised
                    // `ofs-delta`.
                    _ => (extent, row.obj_type, row.delta_base),
                };
                out.push(crate::pack_walk::EmitEntry {
                    oid: oid.to_vec(),
                    stored,
                    obj_type,
                    uncompressed_size: row.uncompressed_size,
                    delta_base,
                    offset: row.offset,
                    recompressed: false,
                    deltified: false,
                });
                continue;
            }

            // ── the base is outside the request ──────────────────────────────
            //
            // **Thin first.** If the client consented to a thin pack and the
            // negotiation vouches that it already holds this base, the entry
            // goes out as a `REF_DELTA` naming that base by oid, carrying the
            // **same compressed delta stream**. Nothing is inflated, nothing is
            // deflated, and the pack is smaller than the whole object by exactly
            // the margin the delta was worth. That is what `caps.thin` has always
            // been the allowance for.
            //
            // An `OFS_DELTA` cannot survive this: it names its base by a
            // backwards distance *within the pack*, and the base is not in the
            // pack. Re-heading it as a `REF_DELTA` is the same header swap the
            // no-`ofs-delta` client path does, over an unchanged payload.
            if let Some(haves) = thin_haves {
                if client_bases.is_none() {
                    client_bases = Some(self.client_bases(haves)?);
                }
                let ext = client_bases
                    .as_ref()
                    .expect("the client's bases were just built");
                let base_oid: Option<Vec<u8>> = match &base {
                    // Unreachable: the arm above `continue`s on both. Matched
                    // rather than `unreachable!()` so that a future arm falling
                    // through here declines the thin path instead of panicking
                    // inside a serving thread.
                    BaseOf::Whole | BaseOf::Inside { .. } => None,
                    BaseOf::Outside {
                        oid: Some(named), ..
                    } => ext
                        .held
                        .contains(&OidKey::new(named)?)
                        .then(|| named.clone()),
                    BaseOf::Outside {
                        offset,
                        oid: None,
                    } => ext.by_offset.get(offset).map(|k| k.as_slice().to_vec()),
                };
                if let Some(base_oid) = base_oid {
                    // A stored `REF_DELTA` already names its base by oid, so its
                    // bytes go out untouched and stay an extent. An `OFS_DELTA`
                    // has to be re-headed, and the result is new bytes.
                    let stored = if row.obj_type == ObjType::OfsDelta {
                        let bytes = self.extent(&snap, row.offset, row.len)?;
                        crate::pack_walk::EntryBytes::Owned(Self::as_ref_delta(&bytes, &base_oid)?)
                    } else {
                        crate::pack_walk::EntryBytes::Extent {
                            offset: row.offset,
                            len: row.len,
                        }
                    };
                    out.push(crate::pack_walk::EmitEntry {
                        oid: oid.to_vec(),
                        stored,
                        obj_type: ObjType::RefDelta,
                        uncompressed_size: row.uncompressed_size,
                        // The base is NOT in this pack, so nothing may be ordered
                        // after it: `topological_order` reads 0 as "names no base
                        // in here", which for an external base is the truth.
                        delta_base: 0,
                        offset: row.offset,
                        // Copied. The payload never left the archive's bytes.
                        recompressed: false,
                        deltified: false,
                    });
                    continue;
                }
            }

            // The base is not being sent and the client cannot supply it, so the
            // stored bytes cannot go out as they are. Two answers remain, and
            // the first is worth ~4× the second on the wire.
            let (kind, body) = self.content(oid)?.ok_or_else(|| {
                anyhow!(
                    "{} is a delta whose base is outside this request, so it must be sent whole, \
                     and this repository cannot produce its content — neither §14's exploded \
                     table nor a re-derivation from the verbatim packs answered",
                    hex::encode(oid)
                )
            })?;

            // ── re-delta against a base the request DOES carry ────────────────
            //
            // The candidate comes from the entry's own stored delta chain: the
            // nearest ancestor that is inside the request. That is the packer's
            // original similarity judgement, already made and already written
            // down — no window, no sort, no scan of the repository.
            let candidate = if crate::delta::enabled() {
                self.in_request_chain_ancestor(row.delta_base, row.offset, &in_set_offsets)?
            } else {
                None
            };
            if let Some(base_at) = candidate {
                let at = *in_set_offsets
                    .get(&base_at)
                    .expect("the ancestor was found by lookup in this very map");
                // Reuse the last base if this entry hangs off the same one.
                if base_content.as_ref().is_none_or(|(o, _)| *o != base_at) {
                    let base_oid = asked[at];
                    let (_, base_body) = self.content(base_oid)?.ok_or_else(|| {
                        anyhow!(
                            "{} is in this request and is the delta base chosen for {}, and this \
                             repository cannot produce its content",
                            hex::encode(base_oid),
                            hex::encode(oid)
                        )
                    })?;
                    base_content = Some((base_at, base_body));
                }
                let (_, base_body) = base_content.as_ref().expect("just filled");

                // `None` is the encoder declining — an unrelated base, or a
                // delta that came out no smaller than the object. Falling
                // through to the whole rebuild is then strictly right.
                if let Some(d) = crate::delta::delta(base_body, &body)? {
                    let stored = Self::ofs_delta_entry(&d, row.offset, base_at)?;
                    // A client with no `ofs-delta` is served the same computed
                    // delta under a `REF_DELTA` head naming the base by oid —
                    // the identical header swap the base-inside arm above does,
                    // over a payload that is already built.
                    // **The `Owned` exception, in its primary shape.** These
                    // bytes were computed here and exist on no disk, so no
                    // extent addresses them; carrying them on the entry is
                    // correct and is bounded by the boundary a narrowed request
                    // cuts through (961 of 33 773 entries on the fixture below),
                    // never by the repository.
                    let (stored, obj_type, delta_base) = if ofs_delta_ok {
                        (
                            crate::pack_walk::EntryBytes::Owned(stored),
                            ObjType::OfsDelta,
                            base_at,
                        )
                    } else {
                        (
                            crate::pack_walk::EntryBytes::Owned(Self::as_ref_delta(
                                &stored, asked[at],
                            )?),
                            ObjType::RefDelta,
                            0,
                        )
                    };
                    out.push(crate::pack_walk::EmitEntry {
                        oid: oid.to_vec(),
                        stored,
                        obj_type,
                        uncompressed_size: row.uncompressed_size,
                        delta_base,
                        offset: row.offset,
                        // Inflated and re-deflated, so NOT a copy — and the
                        // receipt says which of the two rebuilds ran.
                        recompressed: true,
                        deltified: true,
                    });
                    continue;
                }
            }

            // Ship the object whole rather than dragging an unreachable base —
            // and its children — into a clone.
            out.push(crate::pack_walk::EmitEntry {
                oid: oid.to_vec(),
                // Inflated and re-deflated from §14's resolved content: the
                // `Owned` exception again, and again bounded by the boundary.
                stored: crate::pack_walk::EntryBytes::Owned(Self::whole_entry(kind, &body)?),
                obj_type: crate::serve::resolved_type(kind),
                uncompressed_size: body.len() as u64,
                // Whole: it names no base, and nothing may order it after one.
                delta_base: 0,
                offset: row.offset,
                recompressed: true,
                deltified: false,
            });
        }
        Ok(out)
    }

    /// **PHASE 1 of emitting: resolve every entry's stored bytes, in parallel.**
    ///
    /// Emission splits in two, and the split is not a preference — it is where
    /// the ordering dependency actually is:
    ///
    /// | phase | what it does | why |
    /// |---|---|---|
    /// | 1, **here**, parallel | address → bytes for every entry | the archive is immutable, so N workers over disjoint (or shared) extents need **no lock at all** |
    /// | 2, [`crate::pack_walk::emit_pack`], serial | assign output offsets, encode `OFS_DELTA` distances, write, hash | an `OFS_DELTA` names its base by **distance back in the output pack**, so entry *n* cannot be encoded until every earlier entry's byte length is known |
    ///
    /// Phase 2 is not parallelisable and no attempt is made: the distance chain
    /// is a genuine serial dependency, and a running sha1 is another.
    ///
    /// # No lock, and why that is a fact rather than an intention
    ///
    /// [`Self::adopt_journal`] states the storage property this rests on — *"the
    /// blob file is append-only and `gc` truncates nothing, so a retired pack's
    /// bytes are still exactly where the journal says they are"*. Every byte a
    /// snapshot covers is therefore immutable for the life of the file, so the
    /// workers below share one `&Mapped` and read it concurrently with no
    /// synchronisation of any kind. There is no mutex here; if one were needed
    /// the data model would be wrong.
    ///
    /// # 🚨 This is a fan-out inside the serving tier, and `crate::serve`'s
    /// header says that tier is serial
    ///
    /// It did, and the reason it gave is still true and still respected: the
    /// constellation's fan-out primitive has no reentrancy detection, so an
    /// unbounded fan-out per request inside an admission gate that already runs
    /// 32 transfers spawns W² threads. That is why this is **bounded twice**:
    ///
    ///  * it does not fan out at all below [`FAN_OUT_AT`] entries, so every fetch
    ///    small enough for thread-spawn to dominate runs on the calling thread
    ///    exactly as before, and
    ///  * it never exceeds [`MAX_RESOLVE_WORKERS`] (override with
    ///    `ZNIPPY_GIT_EMIT_WORKERS`), so the worst case is 32 × 4 and not
    ///    32 × ncores.
    ///
    /// # What the fan-out is worth — stated honestly
    ///
    /// Resolving a mapped extent is a bounds check and a pointer, so the work
    /// this parallelises is **not** the resolve itself. It is the two things
    /// beside it: the first-touch **page fault** on each entry's header (the
    /// grammar is parsed here, which is what forces that fault, and it validates
    /// the entry before phase 2 commits a byte to the wire), and the `pread`
    /// fallback for any extent the mapping does not cover. On a warm page cache
    /// with the whole archive mapped, phase 1 is nearly free and the fan-out
    /// buys correspondingly little — the 20 % of cycles this change is aimed at
    /// is recovered by *not copying*, not by threads. I could not measure the
    /// fan-out's own contribution separately: oden was at loadavg 25.9 while
    /// this was written (2026-08-14), and a shared box cannot answer a question
    /// that fine.
    pub(crate) fn resolve_emit_payloads<'a>(
        &self,
        entries: &'a [crate::pack_walk::EmitEntry],
        snap: &'a crate::archive_map::Mapped,
    ) -> Result<Vec<std::borrow::Cow<'a, [u8]>>>
    where
        Self: Sync,
    {
        let workers = if entries.len() < FAN_OUT_AT {
            1
        } else {
            emit_workers()
        };
        self.resolve_with(entries, snap, workers)
    }

    /// [`Self::resolve_emit_payloads`] with the worker count named rather than
    /// decided — `1` runs inline on the calling thread.
    ///
    /// Split out so the fan-out is **testable at a size a test fixture can
    /// reach**. The production threshold is 4096 entries and the corpus these
    /// guards run on holds 2687, so a test that went through the front door
    /// would exercise the serial arm every time and the gatling arm never;
    /// `the_fan_out_resolves_the_identical_bytes_the_serial_pass_does` calls
    /// this with 1 and with 4 over the same set and requires the two to agree
    /// byte for byte.
    pub(crate) fn resolve_with<'a>(
        &self,
        entries: &'a [crate::pack_walk::EmitEntry],
        snap: &'a crate::archive_map::Mapped,
        workers: usize,
    ) -> Result<Vec<std::borrow::Cow<'a, [u8]>>>
    where
        Self: Sync,
    {
        use znippy_zoomies::gatling_forkjoin::gatling_for_each;

        let one = |i: usize| -> Result<std::borrow::Cow<'a, [u8]>> {
            let e = &entries[i];
            let bytes = match &e.stored {
                // The bytes live on the entry — the rebuilt-delta exception.
                // Borrowed, never cloned: `EntryBytes::Owned` is already the
                // only allocation this path makes and copying it would double it.
                crate::pack_walk::EntryBytes::Owned(v) => std::borrow::Cow::Borrowed(v.as_slice()),
                crate::pack_walk::EntryBytes::Extent { offset, len } => {
                    self.extent(snap, *offset, *len)?
                }
            };
            // Parse the entry's own header now. Two things fall out: a malformed
            // entry is refused **before** phase 2 has written any of the pack,
            // and the first cache line of every entry is touched here — on
            // whichever worker got it — rather than one at a time on the serial
            // writer.
            crate::pack_walk::type_and_size_of(&bytes).with_context(|| {
                format!(
                    "resolving the stored bytes of {} for emission",
                    hex::encode(&e.oid)
                )
            })?;
            Ok(bytes)
        };

        if workers <= 1 {
            return (0..entries.len()).map(one).collect();
        }
        gatling_for_each(entries.len(), workers, one)
            .into_iter()
            .collect()
    }

    /// **Both phases, one call** — the only way a caller in this crate turns an
    /// ordered emit set into pack bytes.
    ///
    /// It is three lines, and it is a function rather than three lines because
    /// LAW 5's *fix by construction* applies exactly here: the snapshot has to
    /// outlive the slices taken from it, and a caller that took its own snapshot
    /// per entry, or dropped it between the phases, would be writing a pack out
    /// of a mapping that no longer exists. Routing `emit_oids` and every test
    /// through one writer makes that impossible to get wrong twice.
    pub(crate) fn emit_ordered(
        &self,
        ordered: &[crate::pack_walk::EmitEntry],
        out: &mut dyn std::io::Write,
    ) -> Result<crate::pack_walk::EmitReport>
    where
        Self: Sync,
    {
        let snap = self
            .archive_snapshot()
            .context("mapping the archive to emit")?;
        let payloads = self
            .resolve_emit_payloads(ordered, &snap)
            .context("resolving the emit set's stored bytes")?;
        crate::pack_walk::emit_pack(
            ordered,
            self.hash_kind(),
            out,
            &|i| Ok(payloads[i].as_ref()),
        )
    }

    /// **The nearest ancestor of `from`'s stored delta chain that is inside the
    /// request**, or `None` when the chain leaves the request and never comes
    /// back.
    ///
    /// `from` is the archive offset of a base already known to be *outside* the
    /// request; `entry_at` is the offset of the entry that named it.
    ///
    /// # Why this cannot make a cycle, and why that is not a comment but an
    /// invariant
    ///
    /// The answer is required to satisfy `answer < entry_at`. In a well-formed
    /// pack a base always precedes the delta that names it, so every edge in the
    /// stored graph already points from a higher offset to a lower one; a new
    /// edge with the same property keeps the emitted graph a DAG **by
    /// construction**, whatever else the request contains. That matters because
    /// the obvious better candidate — the entry's own delta *children*, one
    /// revision away instead of two to nine — is exactly the set that would
    /// close a cycle: a child's stored base is this entry, so pointing this
    /// entry at the child makes `topological_order` unable to order either, and
    /// `emit_pack` refuses the set. Measured cost of the restriction on
    /// `h2h-linear-sha1-2048c-1024f-16k`: ~1 470 bytes per boundary entry
    /// against ~314 for the cyclic candidate, and 5 800 for shipping it whole.
    ///
    /// An `OFS_DELTA` decreases the offset every hop, so the bound is free
    /// there; a `REF_DELTA` names its base by oid and could point anywhere, so
    /// the comparison is made rather than assumed.
    ///
    /// # Cost
    ///
    /// One 64-byte `pread` per hop, no inflate, no allocation sized by an
    /// object — the same walk and the same ceiling as
    /// [`crate::serve::GitStore::resolved_type_at`], for the same reason: a
    /// pushed pack's back-references are not under this server's control, so a
    /// cycle must terminate rather than spin. Measured hop counts on the fixture
    /// above: 2 for 392 of the 859 that resolve, 3 for 260, and 9 at the worst.
    fn in_request_chain_ancestor(
        &self,
        from: u64,
        entry_at: u64,
        in_request: &HashMap<u64, usize>,
    ) -> Result<Option<u64>> {
        use crate::index_layout::{ObjType, ObjectIndex as _};

        /// git's own `pack.depth` ceiling, doubled — the same bound
        /// `resolved_type_at` carries and for the same reason.
        const MAX_LINKS: usize = 100;
        /// Enough of an entry to read its type/size varint plus either an
        /// `OFS_DELTA` distance or a `REF_DELTA` oid.
        const HEADER_PROBE: u64 = 64;

        let oid_len = self.hash_kind().oid_len();
        let mut at = from;
        for _ in 0..MAX_LINKS {
            if at != 0 && at < entry_at && in_request.contains_key(&at) {
                return Ok(Some(at));
            }
            if at == 0 {
                return Ok(None);
            }
            let head = self.read_extent(at, HEADER_PROBE)?;
            let (t, _, n) = crate::pack_walk::type_and_size_of(&head)?;
            at = match t {
                // A whole object with no base: the chain ends here, and it was
                // not in the request.
                ObjType::Commit | ObjType::Tree | ObjType::Blob | ObjType::Tag => {
                    return Ok(None)
                }
                ObjType::OfsDelta => {
                    let (distance, _) = crate::pack_walk::ofs_distance_of(&head[n..])?;
                    match at.checked_sub(distance) {
                        Some(next) if next != 0 => next,
                        // A base before the start of the archive is a corrupt
                        // back-reference. The whole rebuild is always a correct
                        // answer, so this declines rather than failing a clone
                        // over an entry it was only ever going to optimise.
                        _ => return Ok(None),
                    }
                }
                ObjType::RefDelta => {
                    let Some(base_oid) = head.get(n..n + oid_len) else {
                        return Ok(None);
                    };
                    match self.index().lookup(base_oid) {
                        Some(row) => row.offset,
                        None => return Ok(None),
                    }
                }
            };
        }
        Ok(None)
    }

    /// An `OFS_DELTA` pack entry carrying `d`, deflated.
    ///
    /// The distance written here is the **archive** distance, which is the same
    /// coordinate every stored entry's distance is in;
    /// [`crate::pack_walk::emit_pack`] rewrites it into output coordinates for
    /// every entry it emits, copied or not, so this only has to be well-formed
    /// and self-consistent — [`crate::pack_walk::header_len`] parses it back to
    /// find where the payload starts.
    ///
    /// `Compression::fast` for the same reason [`Self::whole_entry`] uses it,
    /// and the trade is far better here: the buffer being deflated is the delta,
    /// which on the measured fixture is ~1.4 KB against the object's 16 KiB, so
    /// this path deflates roughly a **tenth** of the bytes the whole rebuild it
    /// replaces does.
    fn ofs_delta_entry(d: &[u8], entry_at: u64, base_at: u64) -> Result<Vec<u8>> {
        use crate::index_layout::ObjType;
        use std::io::Write as _;

        let distance = entry_at.checked_sub(base_at).ok_or_else(|| {
            anyhow!(
                "a computed delta at archive offset {entry_at} names a base at {base_at}, which \
                 is after it — an ofs-delta distance is backwards and this would not encode"
            )
        })?;
        let mut out = Vec::with_capacity(d.len() / 2 + 32);
        crate::pack_walk::encode_type_and_size(&mut out, ObjType::OfsDelta, d.len() as u64);
        crate::pack_walk::encode_ofs_distance(&mut out, distance);
        let mut z = flate2::write::ZlibEncoder::new(out, flate2::Compression::fast());
        z.write_all(d)?;
        Ok(z.finish()?)
    }

    /// **The objects the receiver already held before this transfer**, in both
    /// coordinate systems a delta can name a base in.
    ///
    /// `haves` is the negotiated common **tips**, exactly as
    /// [`git_storage_trait::GitServe::emit_pack`] defines them, and the closure
    /// over them is this store's own — the contract deliberately passes tips
    /// rather than the walk's exclusion set, because the closure is
    /// repository-sized and the engine can rebuild it from its own bitmaps.
    ///
    /// # This is a VOUCHER, and the caller has to be able to stand behind it
    ///
    /// Every object named here may be pointed at by a delta the pack does not
    /// carry. If the receiver turns out **not** to hold one, its
    /// `index-pack --fix-thin` dies with `pack has N unresolved deltas`, so a
    /// caller that cannot vouch must not pass `caps.thin`. The case that matters
    /// is a **partial clone**: a `blob:none` client's `have` tips reach blobs it
    /// was deliberately never sent, so the closure over-states what it holds and
    /// a filtered fetch must not be marked thin.
    ///
    /// # Cost
    ///
    /// One bitmap union plus one point lookup per held object, and it happens at
    /// most once per request — and only on a request that has a base outside
    /// itself to place. The `by_offset` map exists because an `OFS_DELTA` knows
    /// its base only as an archive offset; building it over what the *client*
    /// holds rather than over the repository is what keeps this off the
    /// whole-repository scan the closure used to do.
    ///
    /// **And it is spent in raw bytes now, end to end.** This used to ask
    /// [`GitOps::reachable`] — hex out of the ordinal space, `hex::decode`d
    /// straight back to bytes, then a third `to_vec` per object to key the
    /// offset map, three heap allocations per held object for a set the size of
    /// the client's history. It asks [`GitStore::reachable_raw`] instead: one
    /// buffer for the closure, one hash table of [`OidKey`]s over it, and
    /// nothing per object.
    fn client_bases(&self, haves: &[Oid<'_>]) -> Result<ExternalBases> {
        use crate::index_layout::ObjectIndex as _;

        let held = self.reachable_raw(haves, &[]).context(
            "closing over the negotiated common tips, to learn which bases the client can supply \
             for itself",
        )?;
        let refs: Vec<&[u8]> = held.iter().collect();
        let rows = self.index().lookup_batch(&refs);
        let mut by_offset = HashMap::with_capacity(refs.len());
        let mut keys = std::collections::HashSet::with_capacity(refs.len());
        for (oid, row) in refs.iter().zip(rows.iter()) {
            let key = OidKey::new(oid)?;
            if let Some(row) = row {
                by_offset.insert(row.offset, key);
            }
            keys.insert(key);
        }
        Ok(ExternalBases {
            by_offset,
            held: keys,
        })
    }

    /// The oid of the entry at `offset`, looked up **in the request** — which is
    /// the only place it can be, because this is called for a base the request
    /// was just shown to contain.
    ///
    /// Linear over the request rather than a map, because it runs only on the
    /// `ofs-delta`-less client path: a capability no git since 1.4.4 omits, and
    /// one that has to be *advertised absent* to reach here at all.
    fn oid_at_offset(
        &self,
        offset: u64,
        asked: &[&[u8]],
        rows: &[Option<crate::index_layout::IndexRow>],
    ) -> Result<Vec<u8>> {
        asked
            .iter()
            .zip(rows)
            .find(|(_, r)| r.as_ref().is_some_and(|r| r.offset == offset))
            .map(|(oid, _)| oid.to_vec())
            .ok_or_else(|| {
                anyhow!(
                    "the entry at archive offset {offset} was checked to be in this request and \
                     then could not be found in it"
                )
            })
    }

    /// Re-head a stored `OFS_DELTA` as a `REF_DELTA` naming `base_oid`.
    ///
    /// The compressed delta stream is copied unchanged; only the header differs,
    /// because the two forms differ in **how the base is named** and in nothing
    /// else. This is why a client without `ofs-delta` costs bytes and not CPU.
    fn as_ref_delta(stored: &[u8], base_oid: &[u8]) -> Result<Vec<u8>> {
        use crate::index_layout::ObjType;
        let (t, stated_size, n) = crate::pack_walk::type_and_size_of(stored)?;
        if t != ObjType::OfsDelta {
            bail!("only an ofs-delta can be re-headed as a ref-delta, this entry is {t:?}");
        }
        let (_, d) = crate::pack_walk::ofs_distance_of(stored.get(n..).unwrap_or(&[]))?;
        let mut out = Vec::with_capacity(stored.len() + base_oid.len());
        crate::pack_walk::encode_type_and_size(&mut out, ObjType::RefDelta, stated_size);
        out.extend_from_slice(base_oid);
        out.extend_from_slice(stored.get(n + d..).unwrap_or(&[]));
        Ok(out)
    }

    /// Re-head a stored `REF_DELTA` as an `OFS_DELTA` naming the base that sits
    /// at `base_at` — **the inverse of [`Self::as_ref_delta`]**, and the fix for
    /// the one pack shape gitoxide cannot read.
    ///
    /// The compressed delta stream is copied unchanged; only the header differs,
    /// because the two forms differ in how the base is *named* and in nothing
    /// else. The reason this has to exist is in the caller: a `REF_DELTA` whose
    /// base is inside the same pack is resolved by stock git and refused by
    /// every gix-based receiver.
    ///
    /// # The distance written here is a placeholder, and that is by design
    ///
    /// Exactly as [`Self::ofs_delta_entry`] documents: the distance an
    /// `OFS_DELTA` carries is relative to a position in the pack it is *in*, so
    /// [`crate::pack_walk::emit_pack`] re-encodes it for every entry from the
    /// output offsets it is assigning, and [`crate::pack_walk::header_len`]
    /// parses whatever is written here only to find where the payload starts. It
    /// therefore has to be well-formed and self-consistent, and nothing more.
    ///
    /// It cannot simply be `entry_at - base_at`: a `REF_DELTA` is allowed to
    /// name a base that comes **after** it in the pack, and `git index-pack
    /// --fix-thin` produces exactly that — it appends the base at the end. So a
    /// forward reference encodes as `1`, and `emit_pack` writes the real
    /// backwards distance once `topological_order` has put the base first.
    fn as_ofs_delta(
        stored: &[u8],
        oid_len: usize,
        entry_at: u64,
        base_at: u64,
    ) -> Result<Vec<u8>> {
        use crate::index_layout::ObjType;
        let (t, stated_size, n) = crate::pack_walk::type_and_size_of(stored)?;
        if t != ObjType::RefDelta {
            bail!("only a ref-delta can be re-headed as an ofs-delta, this entry is {t:?}");
        }
        let payload = stored.get(n + oid_len..).ok_or_else(|| {
            anyhow!(
                "a ref-delta entry at archive offset {entry_at} is {} bytes, which is not even its \
                 header plus a {oid_len}-byte base oid",
                stored.len()
            )
        })?;
        let mut out = Vec::with_capacity(payload.len() + 32);
        crate::pack_walk::encode_type_and_size(&mut out, ObjType::OfsDelta, stated_size);
        crate::pack_walk::encode_ofs_distance(
            &mut out,
            entry_at.checked_sub(base_at).filter(|d| *d > 0).unwrap_or(1),
        );
        out.extend_from_slice(payload);
        Ok(out)
    }

    /// A whole (non-delta) pack entry for `body`: type/size header, then the
    /// body deflated.
    ///
    /// The **only** place in this crate that deflates on a serving path, and the
    /// only reason it exists is a delta whose base the request does not contain.
    /// `Compression::fast` deliberately: these bytes are re-compressed because
    /// they *cannot* be copied, so the trade is wire size against a clone's
    /// latency, and the entries this runs on are a small boundary fraction of any
    /// real request.
    fn whole_entry(kind: crate::object::GitObjectKind, body: &[u8]) -> Result<Vec<u8>> {
        use std::io::Write as _;
        let mut out = Vec::with_capacity(body.len() / 2 + 32);
        crate::pack_walk::encode_type_and_size(
            &mut out,
            crate::serve::resolved_type(kind),
            body.len() as u64,
        );
        let mut z = flate2::write::ZlibEncoder::new(out, flate2::Compression::fast());
        z.write_all(body)?;
        Ok(z.finish()?)
    }

    /// Take the absorb gate and hold it.
    ///
    /// The only way to observe "a read that arrived before the drain" as a
    /// **state** rather than as a race: with this held, the worker is parked at
    /// the top of its absorb, so the pack is provably durable, provably queued
    /// and provably not indexed. Its guard is in this module's tests, and nothing
    /// on a serving path takes it.
    #[cfg(test)]
    pub(crate) fn hold_absorb_gate(&self) -> std::sync::MutexGuard<'_, ()> {
        self.absorber.gate.lock().expect("absorb gate poisoned")
    }

    /// `pread` an extent out of the verbatim archive, into a fresh `Vec`.
    ///
    /// **The fallback and the owning path.** A caller that can hold a borrow
    /// wants [`Self::archive_snapshot`] plus [`Self::extent`] instead: this one
    /// costs a syscall, an allocation and a kernel→user copy, which is the shape
    /// [`crate::archive_map`] exists to get off the emit path.
    pub(crate) fn read_extent(&self, offset: u64, len: u64) -> Result<Vec<u8>> {
        self.absorber.read_extent(offset, len)
    }

    /// One mapping of the verbatim archive, to be taken **once per operation**
    /// and sliced N times. See [`crate::archive_map::ArchiveMap::snapshot`].
    pub(crate) fn archive_snapshot(&self) -> Result<Arc<crate::archive_map::Mapped>> {
        self.absorber.map.snapshot()
    }

    /// **The one resolver**: an extent's bytes, borrowed from `snap` when it
    /// covers them and `pread` into a fresh buffer when it does not.
    ///
    /// `Cow::Borrowed` is the ordinary answer and the whole point — no syscall,
    /// no allocation, no copy, just a bounds-checked slice of the page cache.
    /// `Cow::Owned` happens for an extent past the mapped end, which means the
    /// blob file grew after the snapshot was taken (a push landed mid-clone).
    /// `Mapped::get` returning `None` is *"go and pread it"* and never *"there
    /// are no bytes"*, so the two arms return the **same bytes** and differ only
    /// in what they cost — which is the property
    /// `an_extent_past_the_mapping_falls_back_and_returns_the_same_bytes`
    /// asserts.
    pub(crate) fn extent<'m>(
        &self,
        snap: &'m crate::archive_map::Mapped,
        offset: u64,
        len: u64,
    ) -> Result<std::borrow::Cow<'m, [u8]>> {
        match snap.get(offset, len) {
            Some(b) => Ok(std::borrow::Cow::Borrowed(b)),
            None => Ok(std::borrow::Cow::Owned(self.read_extent(offset, len)?)),
        }
    }

    /// Record a durable pack as pending index work.
    ///
    /// The **bit going clear**, and the ack path's last act before it returns.
    pub(crate) fn queue(&self, pack_id: u64, extent: Extent) -> Result<()> {
        self.absorber.queue(pack_id, extent)
    }

    fn reach_bitmaps(&self) -> Result<Arc<Vec<ReachEntry>>> {
        self.absorber.reach_bitmaps()
    }

    /// [`Absorber::reach_bitmaps_with`], for the differential guard.
    pub(crate) fn reach_bitmaps_with(
        &self,
        policy: crate::reach::ReachPolicy,
        cache: bool,
    ) -> Result<Arc<Vec<ReachEntry>>> {
        self.absorber.reach_bitmaps_with(policy, cache)
    }

    /// [`Absorber::reach_tables_with`] — the reach table paired with its oid
    /// index, for the negotiation.
    pub(crate) fn reach_tables_with(
        &self,
        policy: crate::reach::ReachPolicy,
        cache: bool,
    ) -> Result<(Arc<Vec<ReachEntry>>, Arc<HashMap<String, u32>>)> {
        self.absorber.reach_tables_with(policy, cache)
    }

    /// The commit oid set, raw, shared. See [`Derived::commit_raw`].
    pub(crate) fn commit_oids_raw(
        &self,
    ) -> Result<Option<Arc<std::collections::HashSet<Vec<u8>>>>> {
        self.absorber.commit_oids_raw()
    }

    pub(crate) fn refold(&self) -> Result<()> {
        self.absorber.refold()
    }
}

impl<S: ObjectIndex> Absorber<S> {
    /// `shared` is the writer's descriptor on the same file
    /// ([`crate::archive_write::ArchiveWrite::blobs_shared`]); `None` opens one,
    /// which is what a caller with no writer does.
    fn open(
        blobs: &Path,
        shared: Option<Arc<File>>,
        hash: GitHashKind,
        objects: ObjectReadStack<S>,
        exploded: Arc<ExplodedCache>,
    ) -> Result<Self> {
        let reader = match shared {
            Some(f) => f,
            None => Arc::new(
                File::open(blobs)
                    .with_context(|| format!("opening {} for reads", blobs.display()))?,
            ),
        };
        let label = blobs
            .parent()
            .map(|p| p.display().to_string())
            .unwrap_or_else(|| blobs.display().to_string());
        let graph_cache = crate::arms::graph_cache_from_env()?;
        if !graph_cache {
            // A memo left by an earlier open with the cache on: this store keeps none.
            let stale = blobs.with_file_name(GRAPH_CACHE_FILE);
            match std::fs::remove_file(&stale) {
                Ok(()) => {}
                Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
                Err(e) => tracing::warn!(
                    target: "znippy.git",
                    path = %stale.display(),
                    error = %e,
                    "the graph cache is off but a stale objects.graph could not be removed"
                ),
            }
        }
        Ok(Self {
            blobs: blobs.to_path_buf(),
            reader,
            map: crate::archive_map::ArchiveMap::new(blobs),
            hash,
            objects,
            exploded,
            derived: RwLock::new(Derived::default()),
            graph_read: std::sync::atomic::AtomicU64::new(0),
            graph_recomputed: std::sync::atomic::AtomicU64::new(0),
            graph_cache: std::sync::atomic::AtomicBool::new(graph_cache),
            reach_carried: std::sync::atomic::AtomicU64::new(0),
            reach_rebuilt: std::sync::atomic::AtomicU64::new(0),
            ordinal_folded: std::sync::atomic::AtomicU64::new(0),
            ordinal_tail: std::sync::atomic::AtomicU64::new(0),
            packs: Mutex::new(PackState::default()),
            gate: Mutex::new(()),
            progress: crate::progress::IndexProgress::new(&label),
            cancel: std::sync::atomic::AtomicBool::new(false),
            label,
            watchdog: RwLock::new(crate::memguard::global()),
            scanned: Mutex::new(HashMap::new()),
            scan_build: Mutex::new(()),
            unindexable: Mutex::new(HashMap::new()),
        })
    }

    /// **Tree payloads, on demand** — [`crate::reach::TreeSource`] over this
    /// store.
    ///
    /// One lookup per tree the walk actually visits, instead of a resident map
    /// of every tree in the repository. Where the bytes come from is
    /// [`Absorber::resolved`]'s business and nobody else's: the exploded cache
    /// if the tree is in it, its own delta chain out of the mmapped pack if it
    /// is not — and the read then writes it back, so a second walk finds it
    /// warm.
    ///
    /// It must not take the `derived` lock, and it does not: `resolved` reads
    /// the exploded pools, the `objects` index and the pending packs, none of
    /// which is `derived`. Both callers build one of these **inside** their
    /// `derived.write()` section, and that is only sound because of this
    /// sentence.
    fn tree_source(&self) -> TreesFromStore<'_, S> {
        TreesFromStore { abs: self }
    }

    /// Packs whose objects are not in the index yet — §13.12's bit, counted.
    fn unindexed_packs(&self) -> usize {
        self.packs.lock().expect("pack state lock").unabsorbed
    }

    /// **Index everything durable, here, now — the ack path's half of §13.9.**
    ///
    /// # The two hours, in one function
    ///
    /// MEASURED by lane P2 on the oden rig 2026-08-30. `git.receive_pack.checking`
    /// fired at 07:43:49; its very first `peel` of a pushed tip missed the
    /// exploded table and missed the `objects` index (the worker was minutes
    /// into a 13 948 244-object pack), and fell into [`Self::scan_pending`],
    /// which ran **a second, read-only, throw-away resolve of the whole 8.12 GiB
    /// pack on the connection thread**:
    ///
    /// ```text
    /// 09:20:21  index.scanned  "a reader scanned a pending pack for itself:
    ///                           13948244 objects in 5792442 ms"
    /// ```
    ///
    /// 1h36m32s on the 16 GB box and 1h44m14s on the 8 GB one, for rows it then
    /// threw away with the comment *"keeping them would be the index, and the
    /// worker is building that"*. That sentence is the defect: two resolvers
    /// over the same bytes, each holding its own O(pack) `PackWalk::entries` and
    /// oid table, on four vCPUs — which is also **why the watchdog paused**. The
    /// reader's second copy is what ate the `MemAvailable` the indexer tripped
    /// over, and neither pass ever finished the check.
    ///
    /// # Who may call this
    ///
    /// **The ack path, and nothing else.** It takes [`Absorber::gate`], so it
    /// waits for whoever is absorbing — and *"nobody waits on the indexer"* is a
    /// law with its own red test
    /// (`a_read_before_the_drain_is_served_from_the_pack_and_never_waits`): a
    /// clone, an `ls-remote`, a `control status` must be answered while the
    /// drain is parked, and they still are, through [`Self::scan_pending`].
    ///
    /// The connectivity check is the one caller for which that is the wrong
    /// trade. It has to prove that every object reachable from 2874 new tips is
    /// present, which means reading them; an entry's oid exists only once its
    /// delta chain has been applied; so the pack must be resolved before the
    /// check can answer, and the only question is whether it is resolved **once
    /// or twice**. This makes it once. `git receive-pack` makes the same choice
    /// and calls it `index-pack`.
    ///
    /// If the account worker already holds the gate, this blocks on it and then
    /// finds `is_absorbed` set and returns: whoever got there first did the work.
    ///
    /// Returns how many packs this call absorbed; `0` means it could not (it is
    /// nested inside an absorb of the only pending pack, or there was nothing
    /// to do).
    fn index_pending_now(&self) -> Result<usize> {
        if self.unindexed_packs() == 0 {
            return Ok(0);
        }
        // **Nested**: this thread is already inside `absorb_gated` and holds the
        // gate, so taking it again is a deadlock, not a wait. The other pending
        // packs are still ours to absorb inline — that is lane U's case, a thin
        // push whose base sits in a pack pushed a moment earlier — and the one
        // we are in the middle of is not.
        if let Some(current) = IN_ABSORB.with(|c| c.get()) {
            let jobs: Vec<PendingPack> = {
                let s = self.packs.lock().map_err(|_| anyhow!("pack state poisoned"))?;
                s.pending()
            };
            return self.absorb_each(&jobs, Some(current));
        }
        let _gate = self.gate.lock().map_err(|_| anyhow!("absorb gate poisoned"))?;
        let jobs: Vec<PendingPack> = {
            let s = self.packs.lock().map_err(|_| anyhow!("pack state poisoned"))?;
            s.pending()
        };
        // `absorb_gated` is a no-op for a pack somebody absorbed while we were
        // queued on the gate, so the loser of the race pays nothing.
        self.absorb_each(&jobs, None)
    }

    /// **Absorb the pending packs only if nobody else is absorbing.**
    ///
    /// `Some(n)` = the gate was free and this caller did the work; `None` = the
    /// worker (or another reader) holds the gate — possibly paused by the
    /// memory watchdog for as long as the box is short — and this caller did
    /// **not** wait for it. Rickard's rule, 2026-08-30: *nobody waits on the
    /// indexer.* A read that needs the index now answers from the pack for
    /// itself ([`Self::scan_pending`]) and a read that needs the whole graph
    /// says so by name rather than blocking behind a kernel pack.
    fn try_absorb_pending(&self) -> Result<Option<usize>> {
        if self.unindexed_packs() == 0 {
            return Ok(Some(0));
        }
        let _gate = match self.gate.try_lock() {
            Ok(g) => g,
            Err(std::sync::TryLockError::WouldBlock) => return Ok(None),
            Err(std::sync::TryLockError::Poisoned(_)) => bail!("absorb gate poisoned"),
        };
        let jobs: Vec<PendingPack> = {
            let s = self
                .packs
                .lock()
                .map_err(|_| anyhow!("pack state poisoned"))?;
            s.pending()
        };
        self.absorb_each(&jobs, None).map(Some)
    }

    /// **A read served from the pack on the reader's own request, never by
    /// waiting for the worker.**
    ///
    /// Every oid in `wants` is looked up in the `objects` table first — the
    /// rows the worker has already landed, batch by batch, are visible there
    /// with no lock in common with it. The ones still missing, if any pack is
    /// still un-indexed, are found by **resolving the pending packs read-only
    /// on this thread**: the same resolver the worker runs
    /// ([`crate::resolve::resolve_streaming`]), over the same mmapped bytes,
    /// with a bounded base set, but with **no sink, no gate and no rows
    /// written** — it produces rows for the caller and nothing else. The pass
    /// stops the moment every wanted oid has been seen.
    ///
    /// `capture` names one oid whose **content** the caller wants as well (a
    /// [`GitStore::content`] read); it rides through a
    /// [`crate::exploded::CaptureOne`] sink and comes back in the second slot.
    ///
    /// # Cost, stated
    ///
    /// An oid that is in a pending pack costs one read-only pass up to where it
    /// sits; an oid that is **absent** costs a full pass over every pending
    /// pack — for a kernel pack, minutes of one core. That is the price of
    /// never answering "absent" for an object whose bytes are durable, paid by
    /// the reader that asked and by nobody else. The rows it computes are not
    /// kept: keeping them would be the index, and the worker is building that.
    fn scan_pending(
        &self,
        wants: &[&[u8]],
        capture: Option<&[u8]>,
    ) -> Result<(Vec<Option<crate::index_layout::IndexRow>>, Option<(GitObjectKind, Vec<u8>)>)> {
        use crate::index_layout::IndexRow;
        let mut out: Vec<Option<IndexRow>> = vec![None; wants.len()];
        let mut missing: HashMap<&[u8], Vec<usize>> = HashMap::with_capacity(wants.len());
        for (i, w) in wants.iter().enumerate() {
            missing.entry(*w).or_default().push(i);
        }
        if missing.is_empty() && capture.is_none() {
            return Ok((out, None));
        }
        // **No nested BUILD.** The resolver asks this absorber (as its
        // `BaseSource`) for a thin pack's external base, and that base may
        // itself be un-indexed — a second *scan build* started from inside the
        // first, and a third from inside that, without end.
        //
        // The guard used to answer `None` outright when re-entered, and that
        // was wider than the recursion it was written for: a base sitting in
        // **another pack that is already scanned** is one binary search away,
        // and refusing it made a thin push on top of a just-acked one fail
        // `[remote rejected] … missing necessary objects` — lane U's rig
        // finding, 2026-08-30. So a nested call now consults the scans that
        // exist and builds none: a lookup terminates, a build is what does not.
        thread_local! { static SCANNING: std::cell::Cell<bool> = const { std::cell::Cell::new(false) }; }
        let nested = SCANNING.with(|s| s.get());
        struct Unset(bool);
        impl Drop for Unset {
            fn drop(&mut self) {
                if !self.0 {
                    SCANNING.with(|s| s.set(false));
                }
            }
        }
        SCANNING.with(|s| s.set(true));
        let _unset = Unset(nested);
        let pending: Vec<PendingPack> = {
            let s = self
                .packs
                .lock()
                .map_err(|_| anyhow!("pack state poisoned"))?;
            s.pending()
        };
        if pending.is_empty() {
            return Ok((out, None));
        }
        let snap = self.map.snapshot()?;
        let mut content: Option<(GitObjectKind, Vec<u8>)> = None;
        for job in &pending {
            if missing.is_empty() && (capture.is_none() || content.is_some()) {
                break;
            }
            let scanned = match self.cached_scan(job.offset) {
                Some(s) => s,
                // Inside another scan: no build, so the recursion the guard
                // exists for cannot happen, and a pack nobody has scanned yet
                // is simply not consulted for this base.
                None if nested => continue,
                None => self.scanned_pack(&snap, job)?,
            };
            missing.retain(|oid, slots| match scanned.row(oid) {
                Some(r) => {
                    for &i in slots.iter() {
                        out[i] = Some(r.clone());
                    }
                    false
                }
                None => true,
            });
            if let Some(want) = capture
                && content.is_none()
                && let Some((entry, _)) = scanned.find(want)
            {
                let bytes: std::borrow::Cow<'_, [u8]> = match snap.get(job.offset, job.len) {
                    Some(b) => std::borrow::Cow::Borrowed(b),
                    None => std::borrow::Cow::Owned(self.read_extent(job.offset, job.len)?),
                };
                let sp = scanned.clone();
                let offset_of = move |oid: &[u8]| sp.find(oid).map(|(e, _)| sp.walk.entries[e].offset);
                content = Some(crate::resolve::materialize_entry(
                    &bytes,
                    &scanned.walk,
                    entry,
                    self.hash.oid_len(),
                    &offset_of,
                    self,
                )?);
            }
        }
        Ok((out, content))
    }

    /// The scan of one pending pack **if it is already built** — a lookup, and
    /// nothing that could recurse. What a nested `scan_pending` is allowed.
    fn cached_scan(&self, offset: u64) -> Option<Arc<ScannedPack>> {
        self.scanned
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .get(&offset)
            .cloned()
    }

    /// The scan of one pending pack: cached, or built now by one read-only
    /// pass (the same resolver, bounded bases, no sink, no rows written).
    fn scanned_pack(&self, snap: &crate::archive_map::Mapped, job: &PendingPack) -> Result<Arc<ScannedPack>> {
        if let Some(s) = self.scanned.lock().unwrap_or_else(|e| e.into_inner()).get(&job.offset) {
            return Ok(s.clone());
        }
        let _one_at_a_time = self.scan_build.lock().unwrap_or_else(|e| e.into_inner());
        if let Some(s) = self.scanned.lock().unwrap_or_else(|e| e.into_inner()).get(&job.offset) {
            return Ok(s.clone());
        }
        let t0 = std::time::Instant::now();
        let bytes: std::borrow::Cow<'_, [u8]> = match snap.get(job.offset, job.len) {
            Some(b) => std::borrow::Cow::Borrowed(b),
            None => std::borrow::Cow::Owned(self.read_extent(job.offset, job.len)?),
        };
        let walk = crate::pack_walk::walk(&bytes, self.hash.oid_len())?;
        let mut rows: Vec<(Vec<u8>, u32, u64)> = Vec::with_capacity(walk.entries.len());
        // A reader's own budget: a quarter of the worker's, no watchdog —
        // the pause is the worker's and only the worker's.
        let limits = crate::resolve::ResolveLimits {
            base_budget: (crate::resolve::base_budget_from_env() / 4).max(1),
            cancel: Some(&self.cancel),
            label: &self.label,
            ..crate::resolve::ResolveLimits::none()
        };
        // Entry index = position in offset order; the rows come out in
        // resolution order, so map offset → index by binary search.
        let report = crate::resolve::resolve_streaming(
            &bytes,
            &walk,
            self.hash,
            job.offset,
            self,
            &crate::exploded::NoSink,
            limits,
            &mut |batch| {
                for r in batch.drain(..) {
                    let pack_off = r.offset - job.offset;
                    let idx = walk
                        .entries
                        .binary_search_by_key(&pack_off, |e| e.offset)
                        .map_err(|_| anyhow!("resolved row at {pack_off} has no walk entry"))?;
                    rows.push((r.oid, idx as u32, r.uncompressed_size));
                }
                Ok(())
            },
        )
        .context("resolving a pending pack read-only for lookups")?;
        rows.sort_unstable_by(|a, b| a.0.cmp(&b.0));
        let scanned = Arc::new(ScannedPack {
            walk,
            rows,
            archive_offset: job.offset,
        });
        tracing::info!(
            target: "gunnar::index",
            event = "index.scanned",
            repo = %self.label,
            pack_id = job.pack_id,
            objects = scanned.rows.len(),
            reinflations = report.reinflations,
            elapsed_ms = t0.elapsed().as_millis() as u64,
            "a reader scanned a pending pack for itself: {} objects in {} ms",
            scanned.rows.len(),
            t0.elapsed().as_millis()
        );
        self.scanned
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .insert(job.offset, scanned.clone());
        Ok(scanned)
    }

    /// **Drop everything this store holds that can be rebuilt** — what the
    /// watchdog's flush reaches: the readers' scanned packs (`scanned`, ~1.2
    /// GB on linux, rebuilt by the next lookup's pass) and the exploded
    /// table's oid tree and batch cache. The rows on disk are untouched.
    ///
    /// **Returns the rows it actually dropped**, and that number is the point.
    /// Lane P2's rig logged this event 190 and 627 times with
    /// `scanned_rows_dropped=0` on every single one and kept waiting anyway,
    /// because the caller could not tell a spill that freed a gigabyte from one
    /// that freed nothing. Three zeros in a row now end the pause
    /// ([`crate::memguard::PauseEnd::NothingToGive`]).
    ///
    /// The exploded table's own caches are an mmap index and a small batch
    /// cache; they are dropped here too and deliberately not counted, because
    /// they are not where a kernel-sized pack's memory is.
    fn spill(&self) -> u64 {
        let dropped = {
            let mut s = self.scanned.lock().unwrap_or_else(|e| e.into_inner());
            let n: usize = s.values().map(|p| p.rows.len()).sum();
            s.clear();
            n
        };
        self.exploded.drop_caches();
        tracing::info!(
            target: "gunnar::index",
            event = "index.spilled",
            repo = %self.label,
            scanned_rows_dropped = dropped,
            "the store dropped its rebuildable caches for the memory watchdog"
        );
        dropped as u64
    }

    /// This store's watchdog (see the field).
    fn watchdog(&self) -> Arc<crate::memguard::MemWatchdog> {
        self.watchdog.read().unwrap_or_else(|e| e.into_inner()).clone()
    }

    /// One oid, never waiting: the index, then the pending packs.
    fn lookup_never_wait(&self, oid: &[u8]) -> Result<Option<crate::index_layout::IndexRow>> {
        if let Some(r) = self.objects.lookup(oid) {
            return Ok(Some(r));
        }
        if self.unindexed_packs() == 0 {
            return Ok(None);
        }
        let (mut rows, _) = self.scan_pending(&[oid], None)?;
        Ok(rows.pop().flatten())
    }

    /// Many oids, never waiting: one index pass, then one read-only pass over
    /// the pending packs for whatever the index did not have.
    fn lookup_batch_never_wait(&self, oids: &[&[u8]]) -> Result<Vec<Option<crate::index_layout::IndexRow>>> {
        let mut rows = self.objects.lookup_batch(oids);
        if self.unindexed_packs() == 0 || rows.iter().all(Option::is_some) {
            return Ok(rows);
        }
        let missing: Vec<&[u8]> = oids
            .iter()
            .zip(&rows)
            .filter(|(_, r)| r.is_none())
            .map(|(o, _)| *o)
            .collect();
        let (found, _) = self.scan_pending(&missing, None)?;
        let mut by_oid: HashMap<&[u8], crate::index_layout::IndexRow> = HashMap::with_capacity(found.len());
        for (o, r) in missing.iter().zip(found) {
            if let Some(r) = r {
                by_oid.insert(*o, r);
            }
        }
        for (o, slot) in oids.iter().zip(rows.iter_mut()) {
            if slot.is_none()
                && let Some(r) = by_oid.get(*o)
            {
                *slot = Some(r.clone());
            }
        }
        Ok(rows)
    }

    /// The bit going clear: this pack is durable and its objects are not in.
    ///
    /// Skips a pack the drain already absorbed — the ack path queues *after* it
    /// has handed the extent to the channel, so the worker can legitimately be
    /// finished before this runs, and clearing a bit for a pack that is already
    /// indexed would leave it clear for ever.
    fn queue(&self, pack_id: u64, extent: Extent) -> Result<()> {
        let mut s = self
            .packs
            .lock()
            .map_err(|_| anyhow!("pack state poisoned"))?;
        s.note(pack_id, extent);
        Ok(())
    }

    /// **The bit, derived from the two durable facts, on open.**
    ///
    /// `journal` is every row this archive's journal holds, in append order.
    /// [`acked_packs`](crate::archive_write::acked_packs) is what turns that into
    /// the packs: its index *is* the pack ordinal — dense by construction, and
    /// re-derived the same way by every later open. For each one, the question is
    /// only *does this pack have any rows in the index*, which
    /// [`ObjectReadStack::extents_with_rows`] answers in one early-terminating
    /// scan rather than per pack.
    ///
    /// Returns the packs whose bit came up **clear**: durable bytes, no rows. The
    /// caller re-queues them on the account's channel, which is what makes a
    /// pack that was mid-absorb when the process died get absorbed by the next
    /// one instead of being durable-but-invisible for ever.
    ///
    /// The verbatim extents are also handed to [`Derived::packs`] — *all* of
    /// them, not only the unabsorbed ones. That is what lets [`BaseSource`]
    /// re-derive a delta base's content from the truth after a restart; without
    /// it a reopened store knows the rows of a pack it can no longer find the
    /// bytes of.
    ///
    /// # The one case where "no rows" does not mean "never absorbed" — and the
    /// third durable fact that now tells them apart
    ///
    /// [`GitOps::gc`] is the only thing that removes rows. A pack **all** of
    /// whose objects it finds dead therefore ends with a journal extent and no
    /// rows, which is byte for byte the state a pack that was in flight when the
    /// machine died leaves behind. This diff cannot tell them apart, and it must
    /// not guess: re-queueing the first resurrects every dead object, and *not*
    /// re-queueing the second loses a pack whose bytes were acked.
    ///
    /// So `gc` says which it is, durably, before it drops a row:
    /// [`retire_packs`](crate::archive_write::retire_packs) appends a tombstone
    /// naming the retired pack's offset, and a pack named by one is never
    /// re-queued here. The bit is still derived from durable facts only — there
    /// are now three of them (the journal's pack rows, the journal's tombstones,
    /// the index's rows) and none of them is a cache of another.
    ///
    /// A **partly** dead pack is not tombstoned and still has rows, so it takes
    /// the `absorbed` branch exactly as it always did.
    ///
    /// A tombstoned pack that *does* still have rows is a `gc` that was killed
    /// between the tombstone and the drop. It is treated as absorbed, which is
    /// what its rows say: nothing is lost, and the next `gc` finishes the job.
    ///
    /// Its extent still goes to [`Derived::packs`] either way. The blob file is
    /// append-only and `gc` truncates nothing, so a retired pack's bytes are
    /// still exactly where the journal says they are — and a live object may
    /// legitimately be a delta against a base that is itself unreachable, so
    /// dropping the extent from that list would break the resolve rather than
    /// tidy it.
    ///
    /// # The second durable fact: §14's exploded table
    ///
    /// A pack counts as absorbed only if the exploded table is **whole**, which
    /// is `exploded.rows() >= objects.len()` — one row per object, because §14's
    /// table is eager and skips nothing. If it is short, the table has been
    /// dropped (or a process died between the two commits) and **no** pack counts
    /// as absorbed: every one is re-queued and re-exploded from the verbatim
    /// truth. That is the whole of "droppable": deleting the file is a supported
    /// operation whose only cost is one re-ingest, and §13.12's rule holds
    /// unchanged over it — absent means fall back, never means wrong.
    ///
    /// `>=` and not `==` on purpose. A resolve that failed part way leaves rows
    /// for the objects it got to, and a `gc` retires both tables together, so the
    /// table can legitimately hold *more* than the index names; only being
    /// **short** is evidence of a gap.
    fn adopt_journal(&self, journal: &[Extent]) -> Result<Vec<PendingPack>> {
        if journal.is_empty() {
            return Ok(Vec::new());
        }
        // The packs, in ordinal order, and the offsets a `gc` retired. A
        // tombstone is a row and not a pack, so it holds no ordinal.
        let packs = crate::archive_write::acked_packs(journal);
        let retired = crate::archive_write::retired_offsets(journal);
        if packs.is_empty() {
            return Ok(Vec::new());
        }
        // Under `Graph` the table is short **by design** — commits and trees
        // only — so the row-count comparison would call every store dropped and
        // re-queue every pack on every open. It is weakened to "not empty",
        // which still catches the case this rule exists for (the file deleted),
        // and is stated rather than hidden: making it exact needs a durable
        // commit-and-tree count, which the index does not publish today.
        // Only `Full` explodes one row per object, so only `Full` can compare
        // counts. Under `Commits` (the shipped default) and `Graph` the eager
        // pool is short **by design**, and the comparison would call every store
        // dropped and re-queue every pack on every open — a Linux-sized re-index
        // per restart. The weakened test still catches what the rule exists for
        // (the file deleted) and is stated rather than hidden: making it exact
        // needs a durable per-kind count, which the index does not publish.
        //
        // The LAZY pool is deliberately not consulted. Its size is a function of
        // what has been READ, so it says nothing about whether a pack was
        // absorbed, and an eviction must never look like a lost index.
        let exploded_is_whole = match self.exploded.policy() {
            crate::exploded_arrow::ExplodePolicy::Full => {
                self.exploded.eager_rows()? >= self.objects.len() as u64
            }
            crate::exploded_arrow::ExplodePolicy::Off => true,
            _ => self.objects.len() == 0 || self.exploded.eager_rows()? > 0,
        };
        let has_rows = if exploded_is_whole {
            self.objects.extents_with_rows(&packs)?
        } else {
            vec![false; packs.len()]
        };
        {
            let mut d = self
                .derived
                .write()
                .map_err(|_| anyhow!("derived poisoned"))?;
            d.packs = packs.clone();
        }
        let mut s = self
            .packs
            .lock()
            .map_err(|_| anyhow!("pack state poisoned"))?;
        let mut requeue = Vec::new();
        for (i, (&extent, &absorbed)) in packs.iter().zip(&has_rows).enumerate() {
            let pack_id = i as u64;
            if absorbed {
                s.mark_absorbed(pack_id, extent);
            } else if retired.contains(&extent.0) {
                // A `gc` dropped every row this pack had and said so durably
                // *before* dropping them. There is no index work owed for it,
                // ever: re-queueing it is precisely how its dead objects would
                // come back. The bit goes **up** rather than being left clear,
                // because a clear bit is what makes a read fall back to
                // absorbing — which is the same resurrection by another door.
                s.mark_absorbed(pack_id, extent);
            } else {
                s.note(pack_id, extent);
                requeue.push(PendingPack {
                    pack_id,
                    offset: extent.0,
                    len: extent.1,
                    resumed: true,
                });
            }
        }
        Ok(requeue)
    }

    /// **The indexer's half of the split**, and not one of the twelve.
    ///
    /// Reads each pending pack's verbatim bytes back out of the archive,
    /// resolves them to oids ([`crate::resolve`]), appends the rows to the
    /// `objects` table and re-folds the derived tables. This is what §13.9-12
    /// puts behind the channel; it runs after the ack, never before one.
    ///
    /// Two callers, one function: the account indexer's worker (through
    /// [`ObjectAbsorb`]) and a read that arrived before that worker did. The
    /// second is §13.12's fallback in its literal form — *slower, never wrong*.
    fn absorb_pending(&self) -> Result<usize> {
        // Nothing to do, and — this is the point — **no gate taken**: the common
        // case is a store whose drain has kept up, and it must not queue behind
        // one that is running.
        if self.unindexed_packs() == 0 {
            return Ok(0);
        }
        let _gate = self
            .gate
            .lock()
            .map_err(|_| anyhow!("absorb gate poisoned"))?;
        let jobs: Vec<PendingPack> = {
            let s = self
                .packs
                .lock()
                .map_err(|_| anyhow!("pack state poisoned"))?;
            s.pending()
        };
        self.absorb_each(&jobs, None)
    }

    /// **Every pack it can, skipping the ones it has already given up on.**
    ///
    /// The one loop all four absorb entry points run (LAW 5): they were four
    /// copies of `for job in &jobs { self.absorb_gated(job)?; }`, and that `?`
    /// is the bug this function exists to remove. Two rules, and both are
    /// consequences of the same sentence — *work that can never succeed must
    /// stop and say why*:
    ///
    /// 1. **A failure does not abandon the packs behind it.** The loop runs to
    ///    the end and reports the FIRST failure afterwards, so a pack that
    ///    cannot be indexed no longer stops the push that arrived after it from
    ///    being indexed and acked.
    /// 2. **A pack that has been given up on is not attempted again.** After
    ///    [`MAX_ABSORB_ATTEMPTS`] it is recorded in
    ///    [`Absorber::unindexable`] with the reason, logged once as
    ///    `index.unindexable`, and skipped from then on. Its objects are then
    ///    genuinely absent from the index, which every read path already knows
    ///    how to say — a push that needs one is refused by name at
    ///    `external_bases_exist`, and a clone that needs one fails the
    ///    connectivity check. What no longer happens is the store re-resolving
    ///    it on every request for the rest of the process's life.
    ///
    /// `skip` is the pack this thread is standing inside an absorb of
    /// ([`IN_ABSORB`]).
    fn absorb_each(&self, jobs: &[PendingPack], skip: Option<u64>) -> Result<usize> {
        let mut absorbed = 0usize;
        let mut first: Option<anyhow::Error> = None;
        for job in jobs {
            if skip == Some(job.pack_id) || self.given_up_on(job.pack_id) {
                continue;
            }
            match self.absorb_gated(job) {
                Ok(()) => absorbed += 1,
                Err(e) => {
                    // Cancellation is the repository being deleted: there is no
                    // queue left worth draining and nothing to record.
                    if self.cancel.load(std::sync::atomic::Ordering::Relaxed) {
                        return Err(e);
                    }
                    self.note_absorb_failure(job, &e);
                    if first.is_none() {
                        first = Some(e);
                    }
                }
            }
        }
        match first {
            None => Ok(absorbed),
            Some(e) => Err(e),
        }
    }

    /// Every pack this store has given up on, ordinal order.
    fn unindexable_packs(&self) -> Vec<Unindexable> {
        let mut v: Vec<Unindexable> = self
            .unindexable
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .values()
            .cloned()
            .collect();
        v.sort_by_key(|u| u.pack_id);
        v
    }

    /// Has this store stopped trying to index that pack?
    fn given_up_on(&self, pack_id: u64) -> bool {
        self.unindexable
            .lock()
            .unwrap_or_else(|e| e.into_inner())
            .contains_key(&pack_id)
    }

    /// Count one failed absorb, and stop at [`MAX_ABSORB_ATTEMPTS`].
    ///
    /// The reason recorded is `root_cause`, not the outermost wrapper: that is
    /// the sentence the backend actually has to offer, and lane Z's rule is that
    /// the client gets it verbatim rather than a layer's paraphrase of it.
    fn note_absorb_failure(&self, job: &PendingPack, err: &anyhow::Error) {
        let giving_up;
        {
            let mut map = self.unindexable.lock().unwrap_or_else(|e| e.into_inner());
            // A pack already given up on is not attempted, so an entry here is
            // always a fresh count.
            let e = map.entry(job.pack_id).or_insert_with(|| Unindexable {
                pack_id: job.pack_id,
                extent: (job.offset, job.len),
                attempts: 0,
                reason: String::new(),
            });
            e.attempts += 1;
            e.reason = format!("{}", err.root_cause());
            if e.attempts < MAX_ABSORB_ATTEMPTS {
                let attempts = e.attempts;
                map.remove(&job.pack_id);
                tracing::warn!(
                    target: "znippy.git",
                    store = %self.label,
                    pack = job.pack_id,
                    attempts,
                    max = MAX_ABSORB_ATTEMPTS,
                    error = %err.root_cause(),
                    "a pack could not be indexed; it stays pending and will be tried again"
                );
                return;
            }
            giving_up = Some(e.clone());
        }
        if let Some(u) = giving_up {
            tracing::error!(
                target: "znippy.git",
                store = %self.label,
                event = "index.unindexable",
                pack = u.pack_id,
                offset = u.extent.0,
                len = u.extent.1,
                attempts = u.attempts,
                error = %u.reason,
                "this store has STOPPED trying to index this pack after {} attempts. Its bytes \
                 are durable and untouched, and every object it introduced is absent from the \
                 index — a push that deltas against one is refused by name and a clone that \
                 needs one fails its connectivity check, both immediately. Nothing will retry \
                 it: re-push the pack, or delete and re-create the repository.",
                u.attempts
            );
        }
    }

    /// One pack, **with [`Absorber::gate`] already held**.
    ///
    /// The order at the end is the correctness argument: the rows are in the
    /// index *before* the bit is cleared. A read that finds the bit still clear
    /// blocks on the gate and then finds the rows; a read that finds it set finds
    /// the rows too. There is no interleaving in which a durable object answers
    /// "absent" — which during negotiation would make a client withhold objects
    /// and lose data.
    fn absorb_gated(&self, job: &PendingPack) -> Result<()> {
        if self.cancel.load(std::sync::atomic::Ordering::Relaxed) {
            // The repository is being deleted: no pack still queued is worth a
            // byte of work, and the drain must get through its queue fast so
            // the store can be dropped and the directory removed.
            return Err(crate::resolve::ResolveError::cancelled());
        }
        {
            let s = self
                .packs
                .lock()
                .map_err(|_| anyhow!("pack state poisoned"))?;
            if s.is_absorbed(job.pack_id) {
                // Already in, by whoever reached it first. Re-resolving is now
                // *correct* — the exploded table is keyed by oid and the graph is
                // folded from it rather than pushed at, so a second absorb of one
                // pack changes nothing (that idempotence is what makes dropping
                // the table and re-queueing every pack a safe operation). It is
                // still a whole pack re-read, re-inflated and re-hashed for
                // nothing, so it is skipped.
                return Ok(());
            }
        }
        // From here to the end of `absorb_one` this thread holds the gate, and
        // the resolver is about to ask this same store for delta bases. See
        // [`IN_ABSORB`]: the re-entry must not take the gate again and must not
        // try to absorb the pack it is standing in.
        let _in = InAbsorb::enter(job.pack_id);
        self.absorb_one(job).map_err(|e| {
            // The entry stays pending by construction — it was never removed —
            // so the bytes are durable, the pack is still un-indexed, and reads
            // keep falling back rather than answering absent.
            e.context(format!(
                "absorbing pack {} at ({}, {}) — its bytes are durable and it stays \
                 un-indexed; reads will keep falling back rather than answer absent",
                job.pack_id, job.offset, job.len
            ))
        })?;
        let mut s = self
            .packs
            .lock()
            .map_err(|_| anyhow!("pack state poisoned"))?;
        s.mark_absorbed(job.pack_id, (job.offset, job.len));
        drop(s);
        // The index answers for it now; the reader's scan is dead weight.
        self.scanned.lock().unwrap_or_else(|e| e.into_inner()).remove(&job.offset);
        Ok(())
    }

    /// One pack: verbatim bytes in, `objects` rows **and** §14's exploded rows
    /// out, from **one** walk and **one** resolve.
    ///
    /// The exploded rows are not a second pass and not a second resolver (LAW 5):
    /// [`crate::resolve::resolve_walked`] hands every object it produces to the
    /// sink as it produces it, including the blob payloads it is about to drop,
    /// so the side table falls out of the pass that was already being made.
    ///
    /// Ordering is the correctness argument, and it is the same one the drain
    /// makes one level up: **the exploded rows are committed before the fold that
    /// reads them**, and the fold runs before this returns, so there is no
    /// interleaving in which the graph is folded over a table that does not yet
    /// hold this pack's commits.
    fn absorb_one(&self, job: &PendingPack) -> Result<()> {
        let t0 = std::time::Instant::now();
        self.progress.begin(job.pack_id, job.len, job.resumed);
        // **A live row every 10 s for as long as this pack takes**, carrying
        // `peak_rss` and the re-inflation count. Those two rode only on
        // `index.finished`, so lane P2's two rig boxes — which ran for 2h14m
        // and 2h12m and never finished a pack — produced neither, in exactly
        // the failure they were built for. The ticker is a thread and not a
        // hook inside the resolve loop on purpose: it prints while the resolver
        // is PAUSED too, which is when an operator most needs it.
        let _tick = crate::progress::ProgressTicker::start(self.progress.clone());
        let result = self.absorb_one_inner(job, t0);
        self.progress.sample_memory();
        self.progress.finish();
        result
    }

    /// The body of [`absorb_one`](Self::absorb_one), so `finish` runs on every
    /// exit and the progress row never sticks at 99 %.
    fn absorb_one_inner(&self, job: &PendingPack, t0: std::time::Instant) -> Result<()> {
        use crate::progress::Phase;
        let progress = &self.progress;
        // **The pack is read where it already is.** `read_extent` was a
        // `pread` of the whole extent into a fresh `Vec` — 8.12 GiB of
        // anonymous heap for torvalds/linux, before one byte was inflated, on a
        // 16 GB box. The mapping costs nothing the kernel cannot reclaim
        // ([`crate::archive_map`]), and `MemAvailable` — what the watchdog
        // reads — knows that.
        let snap = self.map.snapshot()?;
        let bytes: std::borrow::Cow<'_, [u8]> = match snap.get(job.offset, job.len) {
            Some(b) => std::borrow::Cow::Borrowed(b),
            // Past the mapping: the file grew after the snapshot. Same bytes,
            // one `pread` — the exception, not the shape.
            None => std::borrow::Cow::Owned(self.read_extent(job.offset, job.len)?),
        };
        progress.set_phase(Phase::Walk);
        let walked = crate::pack_walk::walk(&bytes, self.hash.oid_len())?;
        let t_walk = t0.elapsed();
        let dog = self.watchdog();
        let spill = || self.spill();
        let limits = crate::resolve::ResolveLimits::for_indexer(
            &self.label,
            &self.cancel,
            progress,
            dog,
            &spill,
        );
        let mut batches = 0u64;
        let mut appended = 0u64;
        let mut entries: Vec<crate::index_layout::IndexEntry> = Vec::with_capacity(limits.row_batch);
        let report = crate::resolve::resolve_streaming(
            &bytes,
            &walked,
            self.hash,
            job.offset,
            self,
            &*self.exploded,
            limits,
            &mut |rows| {
                // **One batch of rows lands, then the next.** The exploded
                // tier already flushed itself every 64 MiB; the index rows
                // used to wait for the whole pack. Now the batch is the unit:
                // rows go to the `objects` table as they are produced, and the
                // pack's own "indexed" bit still goes up only after the LAST
                // batch — `mark_absorbed` is the caller's, after this returns.
                entries.clear();
                entries.extend(rows.iter().map(|r| r.index_entry()));
                rows.clear();
                self.objects.append(&entries)?;
                appended += entries.len() as u64;
                batches += 1;
                Ok(())
            },
        )?;
        drop(entries);
        let t_resolve = t0.elapsed() - t_walk;
        progress.set_phase(Phase::Flush);
        // Down before anything folds over it.
        self.exploded.flush()?;

        {
            let mut d = self
                .derived
                .write()
                .map_err(|_| anyhow!("derived poisoned"))?;
            // A reopen already listed every acked extent (`adopt_journal`), so
            // this is the first absorb of a pack pushed in *this* process — or a
            // re-queued one that is already listed. `find` is what reads this
            // list, so a duplicate would not be wrong, only unbounded.
            if !d.packs.contains(&(job.offset, job.len)) {
                d.packs.push((job.offset, job.len));
            }
        }
        progress.set_phase(Phase::Fold);
        let t_fold0 = std::time::Instant::now();
        // The graph and the tree payloads are folded from the exploded table, not
        // pushed at from here. That is what makes a re-absorb idempotent — which
        // it has to be, because dropping the table re-queues every pack.
        self.refold()?;
        let _ = self.objects.maybe_rebuild()?;
        let t_fold = t_fold0.elapsed();
        // **The audit line**: every number is applied output of this absorb.
        // `peak_rss` is the process's VmHWM off /proc/self/status — the high
        // water mark of the whole server, which on the appliance is the box.
        tracing::info!(
            target: "gunnar::index",
            event = "index.finished",
            repo = %self.label,
            pack_id = job.pack_id,
            pack_bytes = job.len,
            peak_rss = crate::memguard::peak_rss_bytes(),
            mem_available = self.watchdog().available().unwrap_or(0),
            objects = report.rows,
            batches = batches,
            rows_appended = appended,
            reinflations = report.reinflations,
            evictions = report.evictions,
            watchdog_drops = report.watchdog_drops,
            peak_base_bytes = report.peak_base_bytes,
            pauses = report.pauses,
            paused_ms = report.paused_ms,
            walk_ms = t_walk.as_millis() as u64,
            resolve_ms = t_resolve.as_millis() as u64,
            fold_ms = t_fold.as_millis() as u64,
            total_ms = t0.elapsed().as_millis() as u64,
            resumed = job.resumed,
            "indexed {} objects in {} batches: walk {} ms, resolve {} ms, fold {} ms; peak bases {}, {} evictions, {} re-inflations, {} pauses ({} ms); peak rss {}",
            report.rows,
            batches,
            t_walk.as_millis(),
            t_resolve.as_millis(),
            t_fold.as_millis(),
            crate::memguard::fmt_bytes(report.peak_base_bytes),
            report.evictions,
            report.reinflations,
            report.pauses,
            report.paused_ms,
            crate::memguard::fmt_bytes(crate::memguard::peak_rss_bytes())
        );
        Ok(())
    }

    /// Recompute everything derived from the object set: the commit graph, the
    /// tree payloads, the ordinal space, the generation numbers, and (lazily) the
    /// bitmaps.
    ///
    /// **The graph is folded from §14's exploded table, not accumulated.** That
    /// is the whole fix for the clean-reopen hole: commit and tree payloads are
    /// on disk, so this call produces the same graph after a restart as before
    /// one, and it is called at the end of [`GitStore::open_with_arms`] for
    /// exactly that reason. It also makes a re-absorb idempotent — the old
    /// accumulate-into-a-`Vec` shape doubled every commit if a pack was absorbed
    /// twice, which is why dropping the table and re-queueing everything was not
    /// a safe operation before this.
    ///
    /// **Generations are stored and must be recomputed in any fold** (§13,
    /// decided). A generation number is `1 + max(parents)`, so it is invalidated
    /// the moment history's shape changes — which is exactly what absorbing a
    /// pack does. [`assign_generations`] is the one writer of that number and it
    /// is called here, over the whole graph, rather than incrementally: an
    /// incremental update would have to know which descendants moved, and being
    /// wrong about that produces an ancestry cutoff that skips real ancestors.
    ///
    /// # Cost
    ///
    /// One range scan of the exploded table's commits and one of its trees, per
    /// absorbed pack. It is the same order as the `tail_oids` scan below, which
    /// this function already made on every absorb, and it reads **no blob
    /// payload** — that is what the kind index in
    /// [`crate::exploded`] is for.
    ///
    /// ## And that order is the whole repository, on every push
    ///
    /// Said plainly, because "the same order as the scan below" reads like a
    /// reassurance and is not one: **nothing here is proportional to the push.**
    /// Every call rebuilds every tree payload, every commit node, every
    /// generation number, a `HashMap<String, u32>` over every oid in the
    /// repository and a sorted `Vec<String>` of the same — for a push of one
    /// commit exactly as for a push of ten thousand.
    ///
    /// MEASURED on oden 2026-08-11, release, quiet box (`some avg10` 0.00). The
    /// **same** 100-commit increment, pushed onto the same history seeded to
    /// four different depths, server CPU for the push:
    ///
    /// | objects already in the repository | CPU |
    /// |---:|---:|
    /// | 9 614 | 0.12 s |
    /// | 35 350 | 0.23 s |
    /// | 103 988 | 0.45 s |
    /// | 206 925 | 0.77 s |
    ///
    /// That is a straight line: **≈88 ms fixed, plus ≈3.3 µs per object already
    /// stored**, and the fit predicts the four points to within 20 ms. It is
    /// also why a push is slower than the ack path suggests — §13.9 puts the
    /// index behind a channel, but `put_refs` checks every ref target with
    /// `has`, `has` calls `lookup_one`, and `lookup_one` absorbs pending packs
    /// first. The drain is off the *ack* and squarely on the *push*.
    ///
    /// This is not fixed here and the numbers are recorded so the next attempt
    /// starts from them rather than from a guess. The obvious lever — fold
    /// incrementally — is the one the paragraph above rules out for
    /// generations, so it needs its own argument and its own gate.
    fn refold(&self) -> Result<()> {
        // Everything derived is computed **before** the write lock, which is then
        // held only for the swap: the fold is the long part and nothing that
        // reads `derived` should queue behind it.
        //
        // Ordinals first — **read, not re-scanned, when the tail has not moved.**
        // S-082's fix, mirroring the commit-graph memo one paragraph down:
        // `objects.ordinal` beside the pack holds `oids_raw` (every oid in the
        // store, in ordinal = sorted-raw order), and a cold open reads its couple
        // of megabytes back sequentially in milliseconds instead of walking the
        // whole redb B-tree — REFOLD measured that scan at ~65% (0.45 s) of a
        // 136k-object reopen, and proved a flat buffer measured the same, so the
        // cost is the SCAN, not the allocation. It is a memo, not a second truth:
        // the redb tail stays authoritative and the file is discarded unless its
        // (epoch, count) both match the live tail, so a push, a gc or a torn
        // write all fall back to the scan and re-persist.
        //
        // The raw concatenation is BOTH the ordinal space and the validity token
        // for the commit graph below, so the graph is read whenever the object
        // set is unchanged; the hex form and the ordinal map fold out of the same
        // bytes, so ordinal `o` names the same object in every representation by
        // construction rather than by agreeing computations (LAW 5). The tail
        // hands oids in oid order, which is byte order over fixed-width oids and
        // therefore already the sorted concatenation the memo stores.
        //
        // **2026-09-04: the memo is gone, because the scan it spared is gone.**
        // The read stack persists its projection now (`objects.index`), and the
        // sorted oid list IS that index's oid column — one contiguous copy out
        // of a mapped buffer, plus a merge of the few rows redb still holds
        // since the last fold. No file of its own, no header, no epoch to
        // validate: the projection is current by construction. What is kept is
        // the provenance — how many oids came off the index and how many off
        // the tail — because that is the only thing that can prove the space
        // was not a scan (both paths yield identical bytes).
        let space = self.objects.oids_raw()?;
        self.ordinal_folded
            .store(space.folded_rows, std::sync::atomic::Ordering::Relaxed);
        self.ordinal_tail
            .store(space.tail_rows, std::sync::atomic::Ordering::Relaxed);
        let oids_raw = space.raw;
        let token = graph_token(&oids_raw);

        // The commit graph — **read, not recomputed, when the object set has not
        // moved.** S-065's fix: `objects.graph` beside the pack holds the folded
        // graph with generations, written by this same call at derive time and
        // read here. It is not a second truth — it is a memoized projection
        // validated on every read against the authoritative object set (the
        // token is a blake3 over every oid in the store), and an absent file, a
        // stale token or a decode failure all fall back to the fold and
        // re-persist. Same shape as lane REFCAS persisting a tag's peeled target.
        //
        // On a fold, commit payloads are read out of §14's exploded table (no
        // tree payload — that was lane Q's 28.24 GB and most of the 50.9 GiB RSS
        // on `torvalds/linux`), parsed, and generation-numbered. That is the
        // whole of the O(commits) open cost this persistence removes.
        let cached = if self.graph_cache.load(std::sync::atomic::Ordering::Relaxed) {
            self.load_persisted_graph()?
        } else {
            None
        };
        let graph: Vec<CommitNode> = match cached {
            Some((tok, g)) if tok == token => {
                self.graph_read.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                g
            }
            _ => {
                self.graph_recomputed
                    .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
                let commits = self.exploded.of_kind(GitObjectKind::Commit)?;
                let g: Vec<CommitNode> = commits
                    .iter()
                    .map(|(oid, payload)| {
                        let h = crate::object::parse_commit(payload);
                        CommitNode {
                            oid: hex::encode(oid),
                            parents: h.parents,
                            tree: h.tree,
                            committer_time: h.committer_time,
                            generation: 0,
                        }
                    })
                    .collect();
                let g = assign_generations(g);
                // Persist for the next open. A write failure is not fatal to the
                // fold — the graph is correct in memory and the next open simply
                // recomputes — so it is logged, not propagated.
                if !self.graph_cache.load(std::sync::atomic::Ordering::Relaxed) {
                    // Switched off: no memo on this disk.
                } else if let Err(e) = self.persist_graph(&token, &g) {
                    tracing::warn!(store = %self.label, "could not persist objects.graph: {e:#}");
                }
                g
            }
        };

        // The same commit set `select` refuses a `want` against, in the raw form
        // it tests. Folded once here instead of once per request; see the field.
        // A row that does not decode makes the whole set `None`, which is the
        // decline `select` used to reach by returning early out of its own loop.
        let commit_raw = graph
            .iter()
            .map(|n| hex::decode(&n.oid).ok())
            .collect::<Option<std::collections::HashSet<Vec<u8>>>>()
            .map(Arc::new);
        // No hex `Vec<String>` and no `HashMap<String, u32>` are folded here any
        // more (S-082): `oids_raw` is the whole ordinal space, and oid ⇄ ordinal
        // is served off it by [`crate::reach::OrdinalSpace`] — a slice one way, a
        // binary search the other. That removed the ~0.13 s + ~0.08 s this fold
        // spent building those two per-oid `String` tables, and their resident
        // cost with them.
        // The graph-by-oid lookup, folded here so `reachable` borrows it instead
        // of rebuilding it per negotiation. Keyed by hex, the form the walk
        // keys by. Same lifecycle as `commit_raw`: rebuilt on every fold, so it
        // cannot outlive the graph it indexes.
        let graph_index: HashMap<String, u32> = graph
            .iter()
            .enumerate()
            .map(|(i, n)| (n.oid.clone(), i as u32))
            .collect();

        let mut d = self
            .derived
            .write()
            .map_err(|_| anyhow!("derived poisoned"))?;
        // ── carry the reachability table across the fold ──────────────────
        //
        // The table describes immutable commits; the fold only renumbers the
        // ordinal space. Carry it (renumbered) while the graph has not outgrown
        // the last full build by more than the carry rule allows — else drop it
        // and let the next reader rebuild, which is what every fold used to do.
        // See `REACH_CARRY_MAX_FRACTION` for the measurement behind this.
        let carried: Option<Vec<ReachEntry>> = if d.reach.is_empty() {
            None
        } else {
            let grown = graph.len().saturating_sub(d.reach_full_len);
            let allowed = (d.reach_full_len / REACH_CARRY_MAX_FRACTION).max(REACH_CARRY_MIN_COMMITS);
            if grown > allowed {
                None
            } else {
                let oid_len = self.hash.oid_len();
                let old_space = crate::reach::OrdinalSpace::new(&d.oids_raw, oid_len);
                let new_space = crate::reach::OrdinalSpace::new(&oids_raw, oid_len);
                crate::reach::carry_reach(&d.reach, &old_space, &new_space, |c| graph_index.contains_key(c))
            }
        };
        tracing::info!(
            store = %self.label,
            reach_len = d.reach.len(),
            reach_full_len = d.reach_full_len,
            graph_len = graph.len(),
            carried = carried.as_ref().map(Vec::len),
            "fold: reachability table carry decision"
        );
        d.graph = graph;
        d.graph_index = graph_index;
        d.commit_raw = commit_raw;
        d.oids_raw = oids_raw;
        match carried {
            Some(entries) => {
                let index: HashMap<String, u32> = entries
                    .iter()
                    .enumerate()
                    .map(|(i, e)| (e.commit.clone(), i as u32))
                    .collect();
                d.reach = Arc::new(entries);
                d.reach_index = Arc::new(index);
                self.reach_carried
                    .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
            }
            None => {
                // Over the ordinal space that just changed and not carriable;
                // both are rebuilt lazily by `reach_bitmaps_with`.
                d.reach = Arc::default();
                d.reach_index = Arc::default();
            }
        }
        Ok(())
    }

    /// `(reach_carried, reach_rebuilt)` — how many folds since open carried the
    /// reachability table across (renumbered) vs how many reads built it from
    /// scratch. See [`tests::a_push_carries_the_reach_table_across_the_fold`].
    fn reach_source_stats(&self) -> (u64, u64) {
        (
            self.reach_carried.load(std::sync::atomic::Ordering::Relaxed),
            self.reach_rebuilt.load(std::sync::atomic::Ordering::Relaxed),
        )
    }

    /// Where the folded commit graph is cached, beside the verbatim pack:
    /// `<root>/objects.graph`.
    fn graph_cache_path(&self) -> PathBuf {
        self.blobs.with_file_name(GRAPH_CACHE_FILE)
    }

    /// The persisted graph and the token it was written under, or `None` when
    /// the cache is absent, truncated, an unknown version, or fails to decode —
    /// every one of which is a signal to recompute and re-persist rather than an
    /// error, because the cache is a memo and the exploded table is the truth.
    fn load_persisted_graph(&self) -> Result<Option<([u8; 32], Vec<CommitNode>)>> {
        let bytes = match std::fs::read(self.graph_cache_path()) {
            Ok(b) => b,
            Err(e) if e.kind() == std::io::ErrorKind::NotFound => return Ok(None),
            Err(e) => {
                return Err(e).with_context(|| {
                    format!("reading {}", self.graph_cache_path().display())
                });
            }
        };
        // Frame: b"GRPH" (4) + version (1) + token (32) + Arrow IPC graph.
        if bytes.len() < GRAPH_CACHE_HEADER || &bytes[..4] != GRAPH_CACHE_MAGIC || bytes[4] != 1 {
            return Ok(None);
        }
        let mut token = [0u8; 32];
        token.copy_from_slice(&bytes[5..GRAPH_CACHE_HEADER]);
        match crate::graph::decode_graph(&bytes[GRAPH_CACHE_HEADER..]) {
            Ok(g) => Ok(Some((token, g))),
            Err(_) => Ok(None),
        }
    }

    /// Write the folded graph and its token durably: encode, write a temp file,
    /// fsync it, then rename over the cache — the rename is the atomic publish,
    /// so a reader never sees a half-written graph.
    fn persist_graph(&self, token: &[u8; 32], nodes: &[CommitNode]) -> Result<()> {
        use std::io::Write as _;
        let ipc = crate::graph::encode_graph(nodes)?;
        let mut out = Vec::with_capacity(GRAPH_CACHE_HEADER + ipc.len());
        out.extend_from_slice(GRAPH_CACHE_MAGIC);
        out.push(1);
        out.extend_from_slice(token);
        out.extend_from_slice(&ipc);

        let path = self.graph_cache_path();
        let tmp = path.with_extension("graph.tmp");
        {
            let mut f = File::create(&tmp)
                .with_context(|| format!("creating {}", tmp.display()))?;
            f.write_all(&out)
                .with_context(|| format!("writing {}", tmp.display()))?;
            f.sync_all()
                .with_context(|| format!("fsync {}", tmp.display()))?;
        }
        std::fs::rename(&tmp, &path)
            .with_context(|| format!("publishing {}", path.display()))?;
        Ok(())
    }

    /// `(graph_read_from_disk, graph_recomputed)` — how many folds since open
    /// read the persisted graph vs recomputed it. The only thing that can prove
    /// the persisted path ran; see
    /// [`tests::a_reopen_reads_the_persisted_graph_it_does_not_recompute`].
    fn graph_source_stats(&self) -> (u64, u64) {
        (
            self.graph_read.load(std::sync::atomic::Ordering::Relaxed),
            self.graph_recomputed
                .load(std::sync::atomic::Ordering::Relaxed),
        )
    }

    /// `(folded_rows, tail_rows)` of the ordinal space the last fold built — how
    /// many oids came straight off the persisted index's oid column and how many
    /// off a scan of the redb tail. The only thing that can prove the space was
    /// not a scan (LAW 2), because both sources yield identical bytes; see
    /// [`tests::the_ordinal_space_comes_off_the_index_not_a_scan`].
    fn ordinal_source_stats(&self) -> (u64, u64) {
        (
            self.ordinal_folded.load(std::sync::atomic::Ordering::Relaxed),
            self.ordinal_tail.load(std::sync::atomic::Ordering::Relaxed),
        )
    }

    /// The bitmaps, built on first use after a fold and cached until the next.
    fn reach_bitmaps(&self) -> Result<Arc<Vec<ReachEntry>>> {
        self.reach_bitmaps_with(live_reach_policy(), true)
    }

    /// [`Absorber::reach_bitmaps`] with the policy **named** and the cache
    /// optional. The bitmaps alone; [`Absorber::reach_tables_with`] is the one
    /// builder and this drops the index it also returns, for callers that seal
    /// the table or compare two of them and never key into one by oid.
    fn reach_bitmaps_with(
        &self,
        policy: crate::reach::ReachPolicy,
        cache: bool,
    ) -> Result<Arc<Vec<ReachEntry>>> {
        Ok(self.reach_tables_with(policy, cache)?.0)
    }

    /// **The reach table and its oid index, as one pair that is always in
    /// step.** [`crate::reach::accumulate`] keys the table by commit oid, and
    /// building that index per negotiation is exactly the O(commits) waste this
    /// lane removes — so it is built once, here, beside the table, and handed
    /// back with it. The two travel together (both cached, both returned) so a
    /// caller can never pair a fresh index against a stale table or the reverse.
    ///
    /// `cache` is false for the differential guard, which builds the table at two
    /// very different caps over one store and requires the two to answer
    /// identically; letting either poison `d.reach` would make the second arm
    /// read the first one's table and the comparison would be of a thing against
    /// itself.
    fn reach_tables_with(
        &self,
        policy: crate::reach::ReachPolicy,
        cache: bool,
    ) -> Result<(Arc<Vec<ReachEntry>>, Arc<HashMap<String, u32>>)> {
        {
            let d = self
                .derived
                .read()
                .map_err(|_| anyhow!("derived poisoned"))?;
            if cache && !d.reach.is_empty() {
                // Two refcount bumps. This was `d.reach.clone()` — one deep copy
                // of every commit's roaring bitmap, per call, twice per request
                // — plus, before this lane, a whole `HashMap<&str,_>` rebuilt
                // from the table on every call at the negotiation site.
                return Ok((Arc::clone(&d.reach), Arc::clone(&d.reach_index)));
            }
            if d.graph.is_empty() {
                return Ok((Arc::default(), Arc::default()));
            }
        }
        let mut d = self
            .derived
            .write()
            .map_err(|_| anyhow!("derived poisoned"))?;
        // Built before the facts so it outlives them, and deliberately NOT
        // holding the `derived` guard: `Absorber::resolved` does not take that
        // lock, which is what makes a live tree lookup legal from inside this
        // write section at all.
        let trees = self.tree_source();
        let oid_len = self.hash.oid_len();
        let facts = crate::reach::ObjectFacts {
            ordinal: crate::reach::OrdinalSpace::new(&d.oids_raw, oid_len),
            trees: &trees,
            oid_len,
        };
        let started = std::time::Instant::now();
        let built = Arc::new(crate::reach::build_reach(&d.graph, &facts, policy));
        self.reach_rebuilt
            .fetch_add(1, std::sync::atomic::Ordering::Relaxed);
        tracing::info!(
            store = %self.label,
            entries = built.len(),
            commits = d.graph.len(),
            elapsed_ms = started.elapsed().as_millis() as u64,
            cache,
            "built the reachability table from scratch"
        );
        // The commit count this build covered — the denominator the next fold's
        // carry rule divides by (`REACH_CARRY_MAX_FRACTION`).
        d.reach_full_len = d.graph.len();
        // The index over the table just built — commit oid → row — so the
        // negotiation looks a bitmap up by a hash rather than by rebuilding this
        // map first.
        let index: Arc<HashMap<String, u32>> = Arc::new(
            built
                .iter()
                .enumerate()
                .map(|(i, e)| (e.commit.clone(), i as u32))
                .collect(),
        );
        if cache {
            // `Arc::new` once each, then a refcount bump into the cache — the
            // second full copy the table line used to make
            // (`d.reach = built.clone()`) is gone, and the index is cached in
            // lockstep so the read arm above can hand out both.
            d.reach = Arc::clone(&built);
            d.reach_index = Arc::clone(&index);
        }
        Ok((built, index))
    }

    /// The commit oid set, raw, as folded. See [`Derived::commit_raw`].
    fn commit_oids_raw(&self) -> Result<Option<Arc<std::collections::HashSet<Vec<u8>>>>> {
        Ok(self
            .derived
            .read()
            .map_err(|_| anyhow!("derived poisoned"))?
            .commit_raw
            .clone())
    }

    // `ordinal_space()` lived here: `d.oids.clone()`, a fresh `Vec<String>` of
    // every oid in the store, handed to `reachable_oids` twice per request. It
    // is gone, and so is `d.oids` itself (S-082): the serve path borrows
    // `d.oids_raw` as an `OrdinalSpace` under the read guard it was already
    // going to take, and reads oid ⇄ ordinal straight off it.

    /// `pread` an extent out of the verbatim archive, **into uninitialised
    /// capacity** — no zero-fill before the read.
    ///
    /// This used to be `vec![0u8; len]` + `read_exact_at`, and under emit
    /// concurrency the zeroing was not a detail: profiled on oden 2026-08-12
    /// (gunnar serving 32 concurrent clones, perf on the server pid), ~44 % of
    /// serve CPU was memory zeroing/copying — `memset` 18.1 % plus kernel
    /// page-zeroing 7.8 % — against ~12 % for one clone, because 32 threads
    /// churning pack-scale buffers recycle mimalloc freelist blocks that must
    /// each be memset before `pread` immediately overwrites every byte.
    ///
    /// Soundness: no reference to uninitialised memory is ever formed. The
    /// bytes are written through a raw pointer into the `Vec`'s spare
    /// capacity, `EINTR` retries, a short read refuses, and `set_len` runs
    /// only after every one of `len` bytes has been written.
    fn read_extent(&self, offset: u64, len: u64) -> Result<Vec<u8>> {
        use std::os::fd::AsRawFd;
        let want = usize::try_from(len).context("extent length overflows usize")?;
        let mut buf: Vec<u8> = Vec::with_capacity(want);
        let fd = self.reader.as_raw_fd();
        let mut filled = 0usize;
        while filled < want {
            let n = unsafe {
                libc::pread(
                    fd,
                    buf.as_mut_ptr().add(filled).cast(),
                    want - filled,
                    i64::try_from(offset + filled as u64).context("extent offset overflows off_t")?,
                )
            };
            if n < 0 {
                let err = std::io::Error::last_os_error();
                if err.kind() == std::io::ErrorKind::Interrupted {
                    continue;
                }
                return Err(err).with_context(|| {
                    format!("reading ({offset}, {len}) out of {}", self.blobs.display())
                });
            }
            if n == 0 {
                bail!(
                    "short read at ({offset}, {len}) out of {}: {filled} of {want} bytes before \
                     EOF — the archive is truncated relative to its index",
                    self.blobs.display()
                );
            }
            filled += n as usize;
        }
        // Every byte of `want` is now initialised.
        unsafe { buf.set_len(want) };
        Ok(buf)
    }
}

/// **The drain's object-level ingress**, and the reason a push ends in object
/// rows without anybody reading first.
///
/// The account indexer holds this as an `Arc<dyn ObjectAbsorb>` and calls it once
/// per drained job, on its own thread, after the ack. It is the *same* absorb a
/// falling-back read takes, gate and all (LAW 5): there is no background copy of
/// this logic to drift.
impl<S: ObjectIndex> ObjectAbsorb for Absorber<S> {
    fn absorb(&self, job: IndexJob) -> Result<()> {
        // **The memory watchdog, BEFORE the pack and BEFORE the gate**
        // ([`crate::memguard`]). Inside the pack the resolver checks again
        // every few thousand objects; this is the check that keeps a burst of
        // packs from starting one more absorb on a box that is already short.
        // Taken before the gate on purpose: a paused worker holds nothing —
        // not the gate, not a row, not a base — so a reader that wants to
        // absorb for itself can, and every other reader scans the pack.
        if self.cancel.load(std::sync::atomic::Ordering::Relaxed) {
            return Err(crate::resolve::ResolveError::cancelled());
        }
        let dog = self.watchdog();
        if dog.breached().is_some() {
            self.progress.begin(job.pack_id, job.len, false);
            self.progress.set_phase(crate::progress::Phase::Paused);
            self.progress.add_pause();
            let r = dog.pause_until_ok(&self.label, Some(&self.cancel), &mut || Ok(self.spill()));
            self.progress.finish();
            let paused = r?;
            // The pause before the pack has the same rule as the one inside it:
            // a give-up is not a resume. Starting an 8 GiB pack on a box that
            // just told us it has nothing left is how the parked hours began.
            if paused.end.gave_up() {
                return Err(crate::resolve::ResolveError::out_of_memory(
                    paused.end.as_str(),
                    paused.available,
                    paused.budget,
                    0,
                    0,
                ));
            }
        }
        let _gate = self
            .gate
            .lock()
            .map_err(|_| anyhow!("absorb gate poisoned"))?;
        self.absorb_gated(&PendingPack::from_job(job))
    }
}

impl<S: ObjectIndex + 'static> GitStore<S> {
    /// The push path, for the two `put` halves.
    pub(crate) fn push_path(&self) -> &PushPath {
        &self.push
    }

    pub(crate) fn account(&self) -> &str {
        &self.account
    }

    pub(crate) fn ref_log(&self) -> &RefLog {
        &self.refs
    }

    pub(crate) fn ref_gate(&self) -> &Mutex<()> {
        &self.ref_gate
    }

    /// The closure check's index half: a `REF_DELTA`'s base has to exist —
    /// **in the store, or in this very pack**.
    ///
    /// This is the one part of receive-pack that reads `objects.oid` (§13's
    /// table), and on the path a real push takes it reads nothing else.
    ///
    /// # The second half of that sentence used to be missing, and it refused
    /// packs git itself writes
    ///
    /// [`crate::pack_walk::PackWalk::closure`] cannot do better on its own:
    /// a walk reports what the bytes say, and the bytes of a `REF_DELTA` name a
    /// base by **oid**, which is a fact about resolved content that no walk can
    /// know. So every ref base came back on `external_refs` and every one of
    /// them was required to be in the store already.
    ///
    /// A `REF_DELTA` naming a base *inside the same pack* is ordinary and git
    /// writes it constantly: `git index-pack --fix-thin` completes a pushed thin
    /// pack by **appending the base object to the pack** and leaving the delta
    /// naming it by oid. Every such pack was refused with *"this pack deltas
    /// against X, which this repository does not have"*, on a store that was
    /// being handed a perfectly ordinary, self-contained packfile — the base was
    /// entry 1 of 25 in the pack's own `.idx` (`gunnar.multi_pack_serve`,
    /// 2026-08-14). The same missing check is `S-003` / gitoxide#2882 in
    /// `LookupRefDeltaObjectsIter`; it is the same defect wearing the receiving
    /// hat instead of the sending one.
    ///
    /// # What it costs, and where
    ///
    /// **Nothing on any push that was going to succeed before.** Every base
    /// found in the store short-circuits exactly as it did, and a pack with no
    /// ref-delta at all never gets past the first loop.
    ///
    /// The pack is resolved only when the alternative is *refusing it*, and then
    /// the resolve is the cheapest honest answer available: an entry's oid is
    /// knowable only by applying its delta chain, which is precisely what
    /// [`crate::resolve::resolve_walked`] does — the same resolver the absorb is
    /// about to run over the same pack anyway, reused rather than twinned
    /// (LAW 5). A push that is genuinely thin against a base nobody has still
    /// fails, one resolve later, and still names the oid.
    pub(crate) fn external_bases_exist(&self, bytes: &[u8], w: &PackWalk) -> Result<()> {
        let c = w.closure();
        if !c.broken_offsets.is_empty() {
            bail!(
                "this pack is corrupt: {} delta base offset(s) do not land on an entry — first {}",
                c.broken_offsets.len(),
                c.broken_offsets[0]
            );
        }
        // The fast path, unchanged: a base the store already holds is settled
        // without reading one byte of the pack.
        let mut unheld: Vec<&[u8]> = Vec::new();
        for oid in &c.external_refs {
            if self.lookup_one(oid)?.is_none() {
                unheld.push(oid.as_slice());
            }
        }
        if unheld.is_empty() {
            return Ok(());
        }

        // Only now, and only because the alternative is refusing the push: does
        // the pack supply these itself? `resolve_walked` answers with the
        // store behind it, so a base that really is external is still found the
        // cheap way and only a genuinely absent one fails.
        let resolved = crate::resolve::resolve_walked(
            bytes,
            w,
            self.hash_kind(),
            0,
            &*self.absorber,
            &crate::exploded::NoSink,
        )
        .with_context(|| {
            format!(
                "this pack deltas against {} object(s) this repository does not have, so it was \
                 resolved to find out whether the pack carries them itself — first {}",
                unheld.len(),
                hex::encode(unheld[0])
            )
        })?;
        let in_pack: std::collections::HashSet<&[u8]> =
            resolved.iter().map(|r| r.oid.as_slice()).collect();
        for oid in unheld {
            if !in_pack.contains(oid) {
                bail!(
                    "this pack deltas against {}, which this repository does not have and which \
                     the pack does not carry either — the push is refused rather than stored with \
                     a dangling base",
                    hex::encode(oid)
                );
            }
        }
        Ok(())
    }

    /// One oid, on the **serial** path. See [`lookup_path`].
    pub(crate) fn lookup_one(&self, oid: Oid<'_>) -> Result<Option<crate::index_layout::IndexRow>> {
        // **Never waits on the indexer.** The rows the worker has landed so far
        // are in the index; an oid it has not reached is resolved from the
        // pending pack on this thread (`Absorber::scan_pending`).
        self.absorber.lookup_never_wait(oid)
    }

    /// Many oids, on the **batch** path, never waiting on the indexer.
    pub(crate) fn lookup_batch_never_wait(
        &self,
        oids: &[Oid<'_>],
    ) -> Result<Vec<Option<crate::index_layout::IndexRow>>> {
        self.absorber.lookup_batch_never_wait(oids)
    }

    /// Absorb the pending packs on this thread **if the gate is free**;
    /// `None` = the worker holds it and this caller did not wait. See
    /// `Absorber::try_absorb_pending`.
    pub fn try_absorb_pending(&self) -> Result<Option<usize>> {
        self.absorber.try_absorb_pending()
    }

    /// The named refusal for a request that needs the whole derived graph
    /// while a pack is still being indexed: `reachable` cannot be answered
    /// from a graph that does not hold the pack's commits yet, and waiting for
    /// the worker is forbidden. The message carries the progress row's
    /// numbers so the client knows when to come back.
    fn index_in_progress(&self) -> anyhow::Error {
        let p = self.index_progress();
        anyhow!(
            "index in progress for this repository: pack {} {}% ({}/{} objects, {}) — reachability \
             needs the whole graph; retry when the index row reaches 100 %",
            p.pack_id,
            p.percent(),
            p.objects_done,
            p.objects_total,
            p.phase.as_str()
        )
    }

    /// The current ref namespace, as the log folds to it. The whole-snapshot
    /// accessor; [`ref_get`](Self::ref_get) is the keyed one.
    pub(crate) fn ref_state(&self) -> Result<BTreeMap<String, crate::refs::RefState>> {
        self.refs.current()
    }

    /// One ref by name, O(log r), without cloning the namespace — for `head`
    /// and a single-ref CAS. See [`crate::refs::RefLog::get`].
    pub(crate) fn ref_get(&self, name: &str) -> Result<Option<crate::refs::RefState>> {
        self.refs.get(name)
    }

    /// Reachability over the graph, as roaring bitmaps, in **this store's**
    /// ordinal space.
    ///
    /// Returns `(want ∪ closure) − (have ∪ closure)` as oid hex. Every object —
    /// commit, tree and blob — is in the bitmaps, because
    /// [`crate::reach::build_reach`] walks the trees; that is why this can answer
    /// with objects rather than only with commits.
    ///
    /// **Hex is for the maintenance path, not the serving one.** The answer is
    /// computed in raw bytes by [`GitStore::reachable_raw_with`] and encoded
    /// here, at one `String` per object; every caller that serves a request asks
    /// for the raw form instead. See [`Derived::oids_raw`] for what that cost.
    pub(crate) fn reachable_oids(&self, want: &[Oid<'_>], have: &[Oid<'_>]) -> Result<Vec<String>> {
        Ok(self
            .reachable_raw_with(want, have, live_reach_policy(), true)?
            .iter()
            .map(hex::encode)
            .collect())
    }

    /// [`GitStore::reachable_oids`] with the bitmap policy **named** and the
    /// cache optional.
    ///
    /// One body, not two (LAW 5): both are [`GitStore::reachable_raw_with`], so
    /// the guarded path and the shipped path are the same code.
    #[cfg(test)]
    pub(crate) fn reachable_oids_with(
        &self,
        want: &[Oid<'_>],
        have: &[Oid<'_>],
        policy: crate::reach::ReachPolicy,
        cache: bool,
    ) -> Result<Vec<String>> {
        Ok(self
            .reachable_raw_with(want, have, policy, cache)?
            .iter()
            .map(hex::encode)
            .collect())
    }

    /// `(want ∪ closure) − (have ∪ closure)`, in **raw oid bytes**, in one
    /// allocation.
    ///
    /// The answer every serving caller wants and the one the bitmaps naturally
    /// produce: an ordinal indexes straight into [`Derived::oids_raw`], so
    /// nothing is encoded, decoded or allocated per object on the way out.
    pub(crate) fn reachable_raw(
        &self,
        want: &[Oid<'_>],
        have: &[Oid<'_>],
    ) -> Result<git_storage_trait::OidList> {
        self.reachable_raw_with(want, have, live_reach_policy(), true)
    }

    /// [`GitStore::reachable_raw`] with the bitmap policy **named** and the
    /// cache optional.
    ///
    /// It exists because the answer is defined to be **independent of the
    /// policy** — [`crate::reach::accumulate`] covers whatever the table does
    /// not — and a property like that can only be tested by driving the policy.
    /// A test that could not reach it would be asserting that one code path
    /// equals itself. `cache: false` keeps one arm's table from being read by
    /// the next, which is the same trap in a different hat.
    pub(crate) fn reachable_raw_with(
        &self,
        want: &[Oid<'_>],
        have: &[Oid<'_>],
        policy: crate::reach::ReachPolicy,
        cache: bool,
    ) -> Result<git_storage_trait::OidList> {
        // The graph is folded by the absorb, so a pending pack has to be
        // absorbed before reachability can be right. If the gate is free this
        // thread does it; if the worker holds it — hours into a kernel pack,
        // or paused by the watchdog — this is refused by name, not waited on.
        if self.try_absorb_pending()?.is_none() {
            return Err(self.index_in_progress());
        }
        // `reach_bitmaps` may take the derived WRITE lock to build, so it runs
        // to completion before the read guard below is taken. Getting that order
        // wrong is a self-deadlock, not a slow path.
        // The reach table AND its oid index, one pair from the same fold. Both
        // used to be rebuilt at this site on every call: the bitmap map
        // (`e.commit → &bitmap`) here, and the graph map below — O(commits)
        // apiece, per request, for tables that cannot change between folds. They
        // are folded once now and borrowed through a [`crate::reach::ReachView`].
        let (bitmaps, reach_index) = self.reach_tables_with(policy, cache)?;
        // **One read guard, held across the whole answer, and nothing cloned out
        // of it.** This used to be `ordinal_space()?` (a `Vec<String>` of every
        // oid in the store) plus `.ordinal.clone()` (a `HashMap<String, u32>` of
        // the same, again) — ~2 × 12 455 `String` allocations per call and two
        // calls per request, for two tables that cannot change between folds.
        // Borrowing them is not a lock held longer than the work: the guard is a
        // reader, folds take the writer, and a fold that lands mid-answer would
        // invalidate the ordinals this loop is translating.
        let d = self
            .absorber
            .derived
            .read()
            .map_err(|_| anyhow!("derived poisoned"))?;
        let space = &d.oids_raw;
        let oid_len = self.hash.oid_len();
        // oid ⇄ ordinal off the flat buffer — a binary search, no map to clone
        // out of the guard. `want`/`have` arrive as raw bytes, so those lookups
        // take the raw form directly (`ordinal.ordinal_of_raw`) and never encode.
        let ordinal = crate::reach::OrdinalSpace::new(space, oid_len);

        // The graph-by-oid lookup for [`crate::reach::accumulate`]'s commit walk
        // is `d.graph_index`, folded beside `d.graph` and borrowed here — not
        // rebuilt per call. The reach table's index (`reach_index`) is paired
        // with `bitmaps` above by the same fold, so this view never mixes a
        // table with an index from a different one.
        let view = crate::reach::ReachView {
            reach: &bitmaps,
            reach_index: &reach_index,
            graph: &d.graph,
            graph_index: &d.graph_index,
        };
        let trees = self.absorber.tree_source();
        let facts = crate::reach::ObjectFacts {
            ordinal,
            trees: &trees,
            oid_len,
        };

        let mut union = roaring::RoaringBitmap::new();
        for oid in want {
            let hex = hex::encode(oid);
            match view.bitmap(hex.as_str()) {
                // A commit with a bitmap: its whole closure, in one OR.
                Some(bm) => union |= bm,
                // **No bitmap. This is the case that used to be a wrong answer.**
                //
                // It contributed the object itself and nothing else — see
                // `crate::reach::accumulate`'s header for what that cost. Now a
                // commit the graph holds is WALKED, bounded by the first
                // bitmapped commit behind it, and only a `want` the store does
                // not hold at all is still refused.
                None => {
                    if view.is_commit(hex.as_str()) {
                        crate::reach::accumulate(hex.as_str(), &view, &facts, &mut union);
                    } else if let Some(o) = ordinal.ordinal_of_raw(oid) {
                        // Not a commit at all — a TAG or a blob asked for
                        // directly.
                        union.insert(o);
                        // **A tag contributes what it points AT, and until
                        // 2026-08-31 it did not.**
                        //
                        // The bitmaps are built over the commit graph and no
                        // tree points at a tag, so a tag is in nobody's closure
                        // and this arm was the only place its target could
                        // enter. It did not, so an annotated tag shipped as one
                        // object with its commit missing, and the far end said
                        // `missing necessary objects … (peeling a tag)`. That
                        // is exactly what gunnar's replication hit on six
                        // repositories: 25 tags on one commit no branch carried.
                        //
                        // gunnar's own serve paths never had the bug because
                        // they peel BEFORE they ask a store
                        // (`gunnar-wire`'s `select_objects`, step 1). A store
                        // that answers `reachable` has to peel here, because
                        // its caller cannot know which oids are tags.
                        //
                        // A tag object is only ever reachable as a REF TIP —
                        // nothing inside a commit or a tree points at one — so
                        // peeling in the `want` loop is complete, not partial.
                        peel_tag_into(
                            self.absorber.as_ref(),
                            &hex,
                            &view,
                            &facts,
                            &mut union,
                        );
                    } else {
                        bail!(
                            "want {hex} is not in this repository — a negotiation must not be \
                             answered from a partial set"
                        );
                    }
                }
            }
        }
        let mut had = roaring::RoaringBitmap::new();
        for oid in have {
            let hex = hex::encode(oid);
            if let Some(bm) = view.bitmap(hex.as_str()) {
                had |= bm;
            } else if view.is_commit(hex.as_str()) {
                // The same walk on the exclude side, and it must be the same
                // walk: an under-counted `have` over-sends (harmless, wasteful)
                // but an under-counted `want` under-sends, and answering the two
                // sides by different rules is how a `want − have` stops being a
                // subtraction of like for like.
                crate::reach::accumulate(hex.as_str(), &view, &facts, &mut had);
            } else if let Some(o) = ordinal.ordinal_of_raw(oid) {
                had.insert(o);
            }
            // An unknown `have` is normal: the client may have objects we do not.
            // It contributes nothing, which is the safe direction — we send more,
            // never less.
        }

        let delta = union - had;
        // One allocation for the whole answer, sized before the loop. The
        // ordinal space is a flat buffer, so each object is a `copy_from_slice`
        // of `oid_len` bytes into it and nothing else.
        let mut out = git_storage_trait::OidList::with_capacity(delta.len() as usize, oid_len);
        for o in delta {
            let at = o as usize * oid_len;
            let raw = space
                .get(at..at + oid_len)
                .ok_or_else(|| anyhow!("ordinal {o} is outside this store's ordinal space"))?;
            out.push(raw)?;
        }
        Ok(out)
    }

    /// Every object reachable from every ref: what a GC must keep.
    pub(crate) fn live_set(&self) -> Result<std::collections::HashSet<String>> {
        let refs = self.ref_state()?;
        let tips: Vec<Vec<u8>> = refs
            .values()
            .filter_map(|s| s.target.as_deref())
            .filter_map(|t| hex::decode(t).ok())
            .collect();
        if tips.is_empty() {
            bail!(
                "this repository has no ref pointing at an object, so every object in it would be \
                 dead. A GC that would delete everything is refused: name a ref first"
            );
        }
        let borrowed: Vec<Oid<'_>> = tips.iter().map(|t| t.as_slice()).collect();
        let mut live: std::collections::HashSet<String> =
            self.reachable_oids(&borrowed, &[])?.into_iter().collect();
        // The peeled targets of annotated tags are reachable too, and a tag
        // object is not a commit so it has no bitmap of its own.
        for s in refs.values() {
            if let Some(p) = &s.peeled {
                live.insert(p.clone());
            }
            if let Some(t) = &s.target {
                live.insert(t.clone());
            }
        }
        Ok(live)
    }

    /// Drop every row the live set does not name. The index is append-only for
    /// everything except this.
    ///
    /// **Both tables, in one call.** §14's exploded table is retired with the
    /// `objects` table because its row count against that one's is what says
    /// whether it is whole ([`Absorber::adopt_journal`]): dropping from one and
    /// not the other would leave every later open re-exploding the entire
    /// repository. The count returned is still the index's — that is what a
    /// [`GcReport`] means by a dropped row.
    /// Drop exactly the rows in `dead` — a set the GC computed as "seen at its
    /// start and not reachable", never "not in `live`". The distinction is a
    /// row that arrived after the GC took its snapshot: it is in neither set,
    /// and the predicate below keeps it. See [`dead_oids`].
    pub(crate) fn drop_dead_rows(&self, dead: &std::collections::HashSet<Vec<u8>>) -> Result<u64> {
        let dropped = self
            .absorber
            .objects
            .retain(&|oid: &[u8]| !dead.contains(oid))?;
        self.absorber
            .exploded
            .retain(&|oid: &[u8]| !dead.contains(oid))?;
        // **Everything folded from those two tables, refolded.** The commit
        // graph, the tree payloads, the ordinal space and the bitmaps are all
        // derivations of the rows that just went, and a derivation that still
        // names a dropped oid is *wrong* rather than merely stale — a `reachable`
        // over a graph holding dead commits selects objects the index can no
        // longer serve. This is the same [`Absorber::refold`] every absorb and
        // every open already calls, over the table this call just changed;
        // nothing about how the graph is rebuilt is touched here.
        //
        // The bitmaps go with it: they address **ordinals**, and dropping rows
        // renumbers the ordinal space, so `refold` clears them and the next
        // `reach_bitmaps()` builds them over the new numbering. Carrying them
        // across would be carrying an index into a different array.
        self.absorber.refold()?;
        Ok(dropped)
    }

    /// **The journal's half of a GC: retire every pack whose objects are now all
    /// dead, before a single row is dropped.**
    ///
    /// Returns the offsets retired.
    ///
    /// # Why this exists
    ///
    /// §13.12's `indexed` bit is derived on open as *extent in the journal, rows
    /// not in the index* ([`Absorber::adopt_journal`]). Dropping every row of a
    /// pack produces that state exactly, so without this call the next open
    /// re-queues the pack and re-absorbs the objects a GC just decided were dead.
    /// A **partly** dead pack keeps rows and was never affected; this is only
    /// about the all-dead case.
    ///
    /// # What it does not do
    ///
    /// It does not truncate `objects.pack`. The blob file is append-only and a
    /// pack in the middle of it cannot be cut out without moving every extent
    /// after it — which would invalidate every index row in the archive. The dead
    /// pack's bytes stay; what changes is that nothing will ever index them
    /// again. Reclaiming those bytes is a rewrite of the blob file, and it is not
    /// this function.
    ///
    /// An arm with no journal ([`FastWriter`](crate::archive_write::FastWriter))
    /// has nothing to retire and nothing that re-queues, so it returns empty.
    ///
    /// # Cost
    ///
    /// One [`ObjectReadStack::extents_with_rows`] scan plus one batch lookup of
    /// the live set. Both are already the shape a GC pays elsewhere, and both run
    /// once per `gc()`, not per pack.
    /// Where the journal's acked extents end right now — the GC's snapshot of
    /// "packs that existed when I started". A pack acked past this point during
    /// the GC belongs to a push in flight and is not the GC's to judge.
    pub(crate) fn acked_end(&self) -> Result<u64> {
        let Some(journal) = self.arms.writer.journal(&self.blobs) else {
            return Ok(u64::MAX);
        };
        if !journal.exists() {
            return Ok(0);
        }
        let rows = crate::archive_write::read_journal(&journal)?;
        Ok(crate::archive_write::acked_packs(&rows)
            .iter()
            .map(|(o, l)| o + l)
            .max()
            .unwrap_or(0))
    }

    /// Tombstone every acked pack that holds no live object — among the packs
    /// whose extent ends at or before `acked_end`, the GC's snapshot from
    /// [`acked_end`](Self::acked_end). A pack pushed after that snapshot has
    /// objects `live` never saw, and without the bound it would be retired for
    /// being unknown, which is how a GC would delete a push in flight.
    pub(crate) fn retire_dead_packs(
        &self,
        live: &std::collections::HashSet<String>,
        acked_end: u64,
    ) -> Result<Vec<u64>> {
        let Some(journal) = self.arms.writer.journal(&self.blobs) else {
            return Ok(Vec::new());
        };
        if !journal.exists() {
            return Ok(Vec::new());
        }
        let rows = crate::archive_write::read_journal(&journal)?;
        let packs: Vec<(u64, u64)> = crate::archive_write::acked_packs(&rows)
            .into_iter()
            .filter(|(o, l)| o + l <= acked_end)
            .collect();
        let already = crate::archive_write::retired_offsets(&rows);
        if packs.is_empty() {
            return Ok(Vec::new());
        }

        // Which packs hold any row at all right now. A pack with none is either
        // already retired or was never absorbed — neither is this call's
        // business, and calling the second one dead would lose it.
        let occupied = self.absorber.objects.extents_with_rows(&packs)?;

        // Which packs hold a row that survives. Placed by binary search over the
        // pack starts, the same way the scan above places a row.
        let mut order: Vec<usize> = (0..packs.len()).filter(|&i| packs[i].1 > 0).collect();
        order.sort_unstable_by_key(|&i| packs[i].0);
        let starts: Vec<u64> = order.iter().map(|&i| packs[i].0).collect();
        let live_oids: Vec<Vec<u8>> = live.iter().filter_map(|h| hex::decode(h).ok()).collect();
        let borrowed: Vec<&[u8]> = live_oids.iter().map(|o| o.as_slice()).collect();
        let mut has_live = vec![false; packs.len()];
        for (offset, _) in self
            .absorber
            .objects
            .extents_batch(&borrowed)
            .into_iter()
            .flatten()
        {
            let p = starts.partition_point(|&s| s <= offset);
            if p == 0 {
                continue;
            }
            let i = order[p - 1];
            if offset < packs[i].0 + packs[i].1 {
                has_live[i] = true;
            }
        }

        let dead: Vec<u64> = (0..packs.len())
            .filter(|&i| occupied[i] && !has_live[i] && !already.contains(&packs[i].0))
            .map(|i| packs[i].0)
            .collect();
        // Durable before the rows go. The one ordering this function has.
        crate::archive_write::retire_packs(&journal, &dead)?;
        Ok(dead)
    }

    /// **Write generation 0** at [`archive_path`](Self::archive_path), carrying
    /// the verbatim packs this store has acked plus `reserved`.
    ///
    /// The whole of the decision-making lives in
    /// [`seal_generation_zero`](crate::archive_write::seal_generation_zero); this
    /// is the three paths it needs, taken off the same fields the push path and
    /// the GC read, so a seal cannot address a different blob file or a different
    /// journal from the one that acked the packs.
    pub(crate) fn seal_archive(
        &self,
        reserved: Vec<ReservedSection>,
    ) -> Result<crate::archive_write::SealReport> {
        crate::archive_write::seal_generation_zero(
            &self.blobs,
            self.arms.writer.journal(&self.blobs).as_deref(),
            &self.archive,
            reserved,
        )
    }

    /// The sections a sealed archive carries: the ref log, the commit graph and
    /// the bitmaps.
    pub(crate) fn reserved_sections(&self) -> Result<Vec<ReservedSection>> {
        use znippy_common::{GUNNAR_GRAPH_MODULE, GUNNAR_REACH_MODULE};
        let mut out = vec![self.refs.seal_section()?];
        let d = self
            .absorber
            .derived
            .read()
            .map_err(|_| anyhow!("derived poisoned"))?;
        if !d.graph.is_empty() {
            out.push(ReservedSection::arrow(
                GUNNAR_GRAPH_MODULE,
                crate::graph::graph_schema(),
                vec![crate::graph::build_graph_batch(&d.graph)?],
            ));
        }
        drop(d);
        let reach = self.reach_bitmaps()?;
        if !reach.is_empty() {
            out.push(ReservedSection::arrow(
                GUNNAR_REACH_MODULE,
                crate::reach::reach_schema(),
                vec![crate::reach::build_reach_batch(&reach)?],
            ));
        }
        Ok(out)
    }
}

// ── the runtime selector ──────────────────────────────────────────────────────

/// **A [`GitStore`] whose index arm was chosen at run time.**
///
/// The [`ObjectIndex`] arm is a *type*, so a value can only choose it by naming
/// every monomorphisation. This enum is that list, and it is an enum rather than
/// a `Box<dyn GitOps>` for two reasons:
///
/// * a boxed trait object can only offer the twelve, and the arms differ in
///   things the twelve deliberately do not expose — how many Arrow bytes the
///   projection actually materialised, which writer is on the ack path. A bench
///   or an operator dump needs those, and a guard needs them to assert that the
///   selector selected **on applied output** rather than on a label;
/// * there is no allocation and no vtable: the `match` is one predictable branch
///   at the top of a call, and everything under it — the `stree` probe, the
///   Arrow gather, the redb tail — is the same fully static code
///   [`GitStore::<S>::open_with_arms`](GitStore::open_with_arms) produces,
///   because `S` is known inside each arm.
///
/// # What it costs, stated rather than hidden
///
/// One branch per [`GitOps`] call — **not per object**. `extents(&[1000 oids])`
/// is one branch and a thousand monomorphised lookups behind it. The binary
/// carries three copies of the store, which is the price of picking a type at
/// run time and there is no version of this that does not pay it.
///
/// A caller that knows its arm at compile time should **not** come through here:
/// [`GitStore::open`] and [`GitStore::open_with_arms`] hand back a concrete
/// store with no branch at all, and that includes every caller that wants the
/// default.
pub enum SelectedStore {
    OneTableFourColumns(GitStore<OneTableFourColumns>),
    FourTables(GitStore<crate::index_layout::FourTables>),
    PackedPayload(GitStore<crate::index_layout::PackedPayload>),
}

/// One expression, evaluated against whichever concrete store this is. The
/// delegation below is generated from it so there is no second copy of any
/// method body (LAW 5).
macro_rules! on_arm {
    ($self:ident, $s:ident => $body:expr) => {
        match $self {
            SelectedStore::OneTableFourColumns($s) => $body,
            SelectedStore::FourTables($s) => $body,
            SelectedStore::PackedPayload($s) => $body,
        }
    };
}

impl SelectedStore {
    /// The arms this store was built with.
    pub fn arms(&self) -> StoreConfig {
        on_arm!(self, s => s.arms())
    }

    /// **The explode band** — bytes held in each pool, the budget, and what has
    /// been evicted since this store was opened. `Admin.StorageInfo` and the
    /// dashboard's `explode 3.1G/8G` cell.
    pub fn explode_cache_use(&self) -> crate::exploded_cache::CacheUse {
        on_arm!(self, s => s.explode_cache_use())
    }

    /// **This repository's explode quota, set live.** What the control plane
    /// calls when an account's plan changes. It never refuses a read and never
    /// deletes on the caller's thread: over budget evicts at the next rotation,
    /// and an account at its limit still serves every object — from the pack,
    /// which is slower and never wrong.
    pub fn set_explode_budget(&self, bytes: u64) {
        on_arm!(self, s => s.set_explode_budget(bytes))
    }

    /// **`repo gc` reclaims the cache**: every lazy segment unlinked, the graph
    /// pool kept. Returns the bytes that left the filesystem.
    pub fn drop_explode_cache(&self) -> anyhow::Result<u64> {
        on_arm!(self, s => s.drop_explode_cache())
    }

    /// The selected writer's name, as it goes on a bench row.
    pub fn writer_name(&self) -> &'static str {
        on_arm!(self, s => s.writer_name())
    }

    /// What the selected writer's `append` promises.
    pub fn writer_durability(&self) -> &'static str {
        on_arm!(self, s => s.writer_durability())
    }

    /// **Arrow IPC bytes the selected index arm actually materialised.**
    ///
    /// The applied output that distinguishes one layout from another: a packed
    /// 25-byte column, four columns in one section and four independent
    /// sections are three different numbers for the same objects. This is what
    /// a guard reads to prove the selector selected, because a *name* would
    /// prove only that a name was copied.
    pub fn index_ipc_bytes(&self) -> usize {
        on_arm!(self, s => s.index().ipc_bytes())
    }

    /// **The name the built projection reports about itself** —
    /// `OneTableFourColumns`, `FourTables` or `PackedPayload`.
    ///
    /// The other half of the applied-output pair [`index_ipc_bytes`] starts:
    /// the byte count separates the arms once objects are in the store, and
    /// this separates them from the instant it is built, including on an empty
    /// store where all three materialise nothing. It is the value
    /// [`crate::arms::IndexArm::projection_name`] exists to be compared with —
    /// what the layout calls itself, never what the selector asked for — so a
    /// server logging it is stating what it built rather than repeating its own
    /// environment back.
    ///
    /// [`index_ipc_bytes`]: SelectedStore::index_ipc_bytes
    pub fn index_name(&self) -> &'static str {
        on_arm!(self, s => s.index().projection_name())
    }

    /// The oids a pushed extent introduced — [`GitStore::oids_in_extent`],
    /// delegated.
    ///
    /// Not one of the twelve (a pack's extent is a znippy concept), so it does
    /// not arrive with the [`GitOps`] impl below and has to be forwarded by
    /// hand. A server that reports the objects a push introduced needs it, and
    /// a server that picked its index arm at run time still needs it.
    pub fn oids_in_extent(&self, extent: Extent) -> Result<Vec<Vec<u8>>> {
        on_arm!(self, s => s.oids_in_extent(extent))
    }

    /// Objects in the `objects` table.
    pub fn object_count(&self) -> usize {
        on_arm!(self, s => s.object_count())
    }

    /// Block until the background drain has absorbed everything pushed.
    pub fn wait_indexed(&self) {
        on_arm!(self, s => s.wait_indexed())
    }

    /// Stop this store's index work; see [`GitStore::cancel_index`].
    pub fn cancel_index(&self) {
        on_arm!(self, s => s.cancel_index())
    }

    /// This store's index progress; see [`GitStore::index_progress`].
    pub fn index_progress(&self) -> crate::progress::Snapshot {
        on_arm!(self, s => s.index_progress())
    }

    /// Rebuild the projection, so the layout under test is what answers rather
    /// than the redb tail.
    pub fn rebuild_projection(&self) -> Result<()> {
        on_arm!(self, s => s.index().rebuild())
    }

    /// The twelfth method — inherent, not on [`GitOps`], because it returns
    /// Arrow `ReservedSection`s a gix backend has no analog for.
    pub fn seal(&self) -> Result<Vec<ReservedSection>> {
        on_arm!(self, s => s.seal())
    }
}

impl GitOps for SelectedStore {
    fn put(&self, pack: &[u8], refs: &[RefUpdate]) -> Result<TxId> {
        on_arm!(self, s => s.put(pack, refs))
    }
    fn put_pack(&self, bytes: &[u8]) -> Result<TxId> {
        on_arm!(self, s => s.put_pack(bytes))
    }
    // Delegated like every other method, and that is the point: without these
    // two lines a `SelectedStore` would silently inherit the trait's *default*
    // `put_pack_observed` — the one that publishes nothing — and every store
    // gunnar actually opens goes through this enum. The count would have been
    // implemented in `store.rs` and unreachable in production.
    fn put_pack_observed(
        &self,
        bytes: &[u8],
        progress: &dyn git_storage_trait::PackProgress,
    ) -> Result<TxId> {
        on_arm!(self, s => s.put_pack_observed(bytes, progress))
    }
    fn publishes_pack_progress(&self) -> bool {
        on_arm!(self, s => s.publishes_pack_progress())
    }
    fn put_refs(&self, updates: &[RefUpdate]) -> Result<TxId> {
        on_arm!(self, s => s.put_refs(updates))
    }
    fn get(&self, oid: Oid<'_>) -> Result<Option<Stored>> {
        on_arm!(self, s => s.get(oid))
    }
    fn has(&self, oid: Oid<'_>) -> Result<bool> {
        on_arm!(self, s => s.has(oid))
    }
    fn size(&self, oid: Oid<'_>) -> Result<Option<u64>> {
        on_arm!(self, s => s.size(oid))
    }
    fn extents(&self, oids: &[Oid<'_>]) -> Result<Vec<Option<Extent>>> {
        on_arm!(self, s => s.extents(oids))
    }
    fn refs(&self) -> Result<Vec<RefRow>> {
        on_arm!(self, s => s.refs())
    }
    fn update_ref(&self, name: &str, old: Option<Oid<'_>>, new: Option<Oid<'_>>) -> Result<TxId> {
        on_arm!(self, s => s.update_ref(name, old, new))
    }
    fn put_refs_cas(&self, edits: &[git_storage_trait::RefCas<'_>]) -> Result<TxId> {
        on_arm!(self, s => s.put_refs_cas(edits))
    }
    fn reachable(&self, want: &[Oid<'_>], have: &[Oid<'_>]) -> Result<Vec<Vec<u8>>> {
        on_arm!(self, s => s.reachable(want, have))
    }
    fn gc(&self) -> Result<GcReport> {
        on_arm!(self, s => s.gc())
    }
}

/// The reading contract, delegated the same way [`GitOps`] is.
///
/// Every method is one predictable branch over a monomorphised body — the same
/// arrangement, and the same cost argument, [`SelectedStore`] makes for the
/// eleven. It is here rather than in [`crate::serve`] only because the macro
/// that generates it lives in this module.
impl crate::serve::GitServe for SelectedStore {
    fn read(&self, oid: Oid<'_>) -> Result<Option<(crate::index_layout::ObjType, Vec<u8>)>> {
        on_arm!(self, s => crate::serve::GitServe::read(s, oid))
    }
    fn header(&self, oid: Oid<'_>) -> Result<Option<(crate::index_layout::ObjType, u64)>> {
        on_arm!(self, s => crate::serve::GitServe::header(s, oid))
    }
    fn sizes(&self, oids: &[Oid<'_>]) -> Result<Vec<Option<u64>>> {
        on_arm!(self, s => crate::serve::GitServe::sizes(s, oids))
    }
    fn index_now(&self) -> Result<usize> {
        on_arm!(self, s => crate::serve::GitServe::index_now(s))
    }
    fn head(&self) -> Result<Option<RefRow>> {
        on_arm!(self, s => crate::serve::GitServe::head(s))
    }
    fn set_head(&self, target: &str) -> Result<TxId> {
        on_arm!(self, s => crate::serve::GitServe::set_head(s, target))
    }
    /// `objects` is the **exact set to emit**, not tips to close over — the
    /// contract's word, and this forwarder uses it so that the name cannot drift
    /// back into `want` one arm at a time. See [`crate::serve::GitServe::emit_pack`].
    fn emit_pack(
        &self,
        objects: &[Oid<'_>],
        have: &[Oid<'_>],
        caps: &git_storage_trait::Caps,
        out: &mut dyn std::io::Write,
    ) -> Result<git_storage_trait::PackStats> {
        on_arm!(self, s => crate::serve::GitServe::emit_pack(s, objects, have, caps, out))
    }
    fn select(
        &self,
        want: &[Oid<'_>],
        have: &[Oid<'_>],
    ) -> Result<Option<git_storage_trait::ReachSet>> {
        on_arm!(self, s => crate::serve::GitServe::select(s, want, have))
    }
}

/// **Open a store with all three arms chosen at run time.**
///
/// The one place [`crate::arms::IndexArm`] — a value — is turned into `S` — a
/// type. See [`SelectedStore`] for what the resulting handle costs.
pub fn open_selected(
    root: &Path,
    account: &str,
    hash: GitHashKind,
    arms: StoreConfig,
) -> Result<SelectedStore> {
    use crate::arms::IndexArm;
    use crate::index_layout::{FourTables, PackedPayload};
    Ok(match arms.index {
        IndexArm::OneTableFourColumns => {
            SelectedStore::OneTableFourColumns(GitStore::<OneTableFourColumns>::open_with_arms(
                root, account, hash, arms,
            )?)
        }
        IndexArm::FourTables => SelectedStore::FourTables(GitStore::<FourTables>::open_with_arms(
            root, account, hash, arms,
        )?),
        IndexArm::PackedPayload => SelectedStore::PackedPayload(
            GitStore::<PackedPayload>::open_with_arms(root, account, hash, arms)?,
        ),
    })
}

/// **The env-driven front door**: read the three variables **once**, here, and
/// build the store they name.
///
/// This is the whole runtime selector a server or a bench needs — one call, one
/// triple of `getenv`s, and from then on the store holds built objects and reads
/// nothing. There is deliberately no per-operation lookup of any kind: a
/// `std::env::var` inside a serving loop would be one syscall per object served,
/// which is a real defect that was found and fixed in gunnar the day before this
/// was written. [`crate::arms::env_reads`] counts every read this crate makes so
/// that "once, at construction" is an assertion rather than a claim.
///
/// An unset variable takes the shipping default, so a process with a clean
/// environment gets exactly [`GitStore::open`]'s combination.
pub fn open_from_env(root: &Path, account: &str, hash: GitHashKind) -> Result<SelectedStore> {
    let arms = StoreConfig::from_env()?;
    open_selected(root, account, hash, arms)
}

impl<S: ObjectIndex> Absorber<S> {
    /// **One object's resolved content, and §14's table is what answers.**
    ///
    /// Two paths, and they are not interchangeable even though they return
    /// identical bytes:
    ///
    /// 1. **the exploded table** — one point lookup, no inflate, no delta chain;
    /// 2. **re-derived from the verbatim truth** — find which pack holds the
    ///    object, read the whole pack back, resolve it. Correct and slow, and it
    ///    exists because the table is droppable: a store whose table has been
    ///    deleted, or whose re-explosion has not caught up, must still answer.
    ///
    /// The counters are the only way to tell them apart. Identical bytes come
    /// back either way, so a byte comparison proves nothing about which path ran
    /// — that is the identity-value trap in its exact form, and it is why
    /// [`crate::exploded::ExplodedStats::rederived`] is bumped here rather than
    /// inferred.
    ///
    /// On the [`Absorber`] rather than on [`GitStore`] because that is where the
    /// absorb runs: the background worker has no store handle, and a base lookup
    /// that needed one would put the whole store behind the drain.
    fn resolved(&self, oid: &[u8]) -> Result<Option<(GitObjectKind, Vec<u8>)>> {
        Ok(self
            .resolved_costed(oid)?
            .map(|(kind, payload, _)| (kind, payload)))
    }

    /// [`Self::resolved`] and the bill. `None` for the cost means a pool
    /// answered and nothing was rebuilt.
    #[allow(clippy::type_complexity)]
    fn resolved_costed(
        &self,
        oid: &[u8],
    ) -> Result<Option<(GitObjectKind, Vec<u8>, Option<crate::resolve::ReadCost>)>> {
        // (1) §14's table.
        if let Some((kind, payload)) = self.exploded.content(oid)? {
            return Ok(Some((kind, payload, None)));
        }
        // (2) the verbatim truth, re-derived.
        let Some(row) = self.objects.lookup(oid) else {
            // (3) not indexed yet: the pending packs, read-only, on this
            // reader's own thread — never by waiting for the worker.
            //
            // **This arm is the ack path's trap and it is not the ack path's
            // fix.** `a_read_before_the_drain_is_served_from_the_pack_and_never_waits`
            // is the law here: a clone must be answered while the drain is
            // parked, so a read may not take the absorb gate. What must not
            // happen is the ACK arriving here, and it no longer does —
            // `GitServe::index_now` runs before the connectivity check, so by
            // the time the check reads there is nothing pending to scan. See
            // [`Absorber::index_pending_now`].
            if self.unindexed_packs() == 0 {
                return Ok(None);
            }
            self.exploded.note_rederived();
            let (_, content) = self.scan_pending(&[], Some(oid))?;
            // The pending scan resolves through a `PackWalk` rather than a
            // chain walk, so it does not produce a bill. `None` is honest here
            // and is not "it was free".
            return Ok(content.map(|(k, p)| (k, p, None)));
        };
        // ── ONE CHAIN, not one pack ─────────────────────────────────────
        //
        // The row's `offset` addresses the entry in the archive directly, so the
        // object is materialised by following its own delta chain out of the
        // mapping — `d` inflates and `d − 1` delta applications, nothing else
        // touched. This used to `read_extent` the WHOLE containing pack and
        // re-resolve it: for `torvalds/linux` an 8.12 GiB read and a 13.9 M
        // object resolve **per object served**, which is the cost that forced
        // the exploded tier to be built eagerly at push time — and that tier
        // MEASURED 141.7 GiB for an 8.12 GiB pack (17×). Lazy explode is only
        // affordable because of this function
        // (`ideas/lazy-explode-index-only-at-push.md`).
        //
        // It also does not care whether the containing pack has been ABSORBED:
        // the bytes were fsynced before the ack, so an object whose row is in
        // the index is readable the instant the row is. The pending-pack scan
        // below is for the window before the row exists at all.
        self.exploded.note_rederived();
        let snap = self.map.snapshot()?;
        let src = ArchiveWindow { abs: self, snap: &snap };
        let offset_of = |base: &[u8]| self.objects.lookup(base).map(|r| r.offset);
        match crate::resolve::materialize_at(
            &src,
            row.offset,
            self.hash.oid_len(),
            &offset_of,
            self,
        ) {
            Ok((kind, payload, cost)) => {
                self.keep_for_the_next_reader(oid, kind, &payload, cost);
                Ok(Some((kind, payload, Some(cost))))
            }
            Err(e) => {
                // A chain that cannot be followed out of the mapping is not a
                // wrong answer to hand back: fall through to the pending scan,
                // which re-derives from the pack the row came in on.
                //
                // **And if that finds nothing either, the ORIGINAL error is
                // what comes back — never `Ok(None)`.** The row says this store
                // holds the object; a failed read of an object we hold is a
                // store defect, and answering "absent" for it turned that defect
                // into a refusal of somebody else's push (lane AI, 2026-08-30 —
                // see `crate::resolve::MAX_DELTA_LINKS`). Absence is a fact the
                // index states, not a shape an error decays into.
                if self.unindexed_packs() == 0 {
                    return Err(e);
                }
                match self.scan_pending(&[], Some(oid))? {
                    (_, Some((k, p))) => Ok(Some((k, p, None))),
                    (_, None) => Err(e),
                }
            }
        }
    }
}

impl<S: ObjectIndex> Absorber<S> {
    /// **The write-back, and it is off the reader's path.**
    ///
    /// The reader has its bytes already — this buffers a copy into the lazy
    /// pool's newest segment (a `memcpy` under one lock, no I/O, no fsync, no
    /// eviction) and, once 4 MiB have accumulated, starts a thread to do the
    /// writing. That is Rickard's law from the design note in one function:
    /// *nobody waits*. A failure here is not the read's failure — the bytes
    /// were correct before the cache was consulted and are correct after it
    /// refuses — so it is counted and swallowed rather than propagated.
    ///
    /// **It does not always write.** A shallow, small object is cheaper to
    /// inflate again than to keep — a whole object is one `inflate` of one zlib
    /// stream at one offset — so the cache declines it and counts the read as
    /// `direct_reads` (`exploded_cache::worth_caching`). That is Rickard's
    /// refinement, 2026-08-30: *do not explode uniformly*. The reader that just
    /// paid the bill is the only thing that knows what it was, which is why the
    /// cost comes back out of [`crate::resolve::materialize_at`] rather than
    /// being predicted from an index column.
    ///
    /// Two readers racing on one oid is benign: identical bytes, and the table
    /// keeps the first row it filed under that key.
    fn keep_for_the_next_reader(
        &self,
        oid: &[u8],
        kind: GitObjectKind,
        payload: &[u8],
        cost: crate::resolve::ReadCost,
    ) {
        match self.exploded.write_back(oid, kind, payload, cost) {
            Ok(true) => self.exploded.kick(),
            Ok(false) => {}
            Err(e) => {
                tracing::debug!(
                    target: "znippy.git",
                    store = %self.label,
                    error = %e,
                    "the exploded cache would not take a write-back; the read is unaffected"
                );
            }
        }
    }
}

/// [`crate::reach::TreeSource`] over a live store — see
/// [`Absorber::tree_source`].
/// The oid a tag object points at, out of its header.
///
/// A tag object's payload begins `object <hex>\n` — that is the format, not a
/// heuristic, and it is the whole parse. Deliberately NOT reached for with a
/// git object library: this crate is a shipped archiver and `gix-object` is a
/// dev-dependency here, so pulling it into the read path to read one line would
/// put the pack pipeline back in an archive-only build (S-071).
fn tag_target_hex(payload: &[u8], oid_len: usize) -> Option<String> {
    let line = payload.split(|b| *b == b'\n').next()?;
    let rest = line.strip_prefix(b"object ")?;
    // Exactly the hash width this store uses, and hex. A shorter or longer
    // field is a malformed tag, and guessing at one would put a wrong oid into
    // a reachability answer — the one place a wrong answer is silent.
    if rest.len() != oid_len * 2 || !rest.iter().all(u8::is_ascii_hexdigit) {
        return None;
    }
    std::str::from_utf8(rest).ok().map(str::to_owned)
}

/// Add what a tag points at — and what THAT reaches — to a reachability union.
///
/// See the call site in `reachable_raw_with` for why this exists and why doing
/// it in the `want` loop is complete rather than partial.
///
/// A tag of a tag is legal git, so this follows the chain; `MAX_TAG_HOPS`
/// bounds it because a corrupt archive could point one at itself and a
/// reachability walk must terminate on bad data, not hang on it.
fn peel_tag_into<S: ObjectIndex>(
    abs: &Absorber<S>,
    tag_hex: &str,
    view: &crate::reach::ReachView<'_>,
    facts: &crate::reach::ObjectFacts<'_>,
    union: &mut roaring::RoaringBitmap,
) {
    const MAX_TAG_HOPS: usize = 8;
    let mut cursor = tag_hex.to_owned();
    for _ in 0..MAX_TAG_HOPS {
        let Ok(raw) = hex::decode(&cursor) else { return };
        // Not a tag, unreadable, or absent: nothing to peel. An unreadable
        // object makes the closure SHORT, and a short closure is refused
        // downstream by name (`missing necessary objects`) rather than served
        // as though it were complete — the same contract `TreesFromStore::tree`
        // documents.
        let Ok(Some((GitObjectKind::Tag, payload))) = abs.resolved(&raw) else { return };
        let Some(target) = tag_target_hex(&payload, facts.oid_len) else { return };
        if let Some(bm) = view.bitmap(target.as_str()) {
            *union |= bm;
            return;
        }
        if view.is_commit(target.as_str()) {
            crate::reach::accumulate(target.as_str(), view, facts, union);
            return;
        }
        let Some(o) = facts.ordinal.ordinal_of_hex(target.as_str()) else { return };
        union.insert(o);
        // A tag may point at a TREE, whose subtrees and blobs are then just as
        // necessary as a commit's would be. Inserting the tree alone would be
        // the same bug one level down.
        match abs.resolved(&hex::decode(&target).unwrap_or_default()) {
            Ok(Some((GitObjectKind::Tree, _))) => {
                crate::reach::accumulate_tree(target.as_str(), facts, union);
                return;
            }
            Ok(Some((GitObjectKind::Tag, _))) => cursor = target,
            // A blob, or unreadable: it is in, and it has no closure.
            _ => return,
        }
    }
}

struct TreesFromStore<'a, S: ObjectIndex> {
    abs: &'a Absorber<S>,
}

impl<S: ObjectIndex> crate::reach::TreeSource for TreesFromStore<'_, S> {
    fn tree(&self, oid_hex: &str) -> Option<std::borrow::Cow<'_, [u8]>> {
        let raw = hex::decode(oid_hex).ok()?;
        // A read that fails is `None`, which is what an object this archive
        // does not hold has always answered here — the walk stops at it. It is
        // NOT silently dropping an error that matters: an unreadable tree makes
        // the closure short, and a short closure is refused downstream by name
        // (`missing necessary objects`) rather than served as if it were
        // complete.
        match self.abs.resolved(&raw) {
            Ok(Some((GitObjectKind::Tree, payload))) => Some(std::borrow::Cow::Owned(payload)),
            // A blob, a tag, a gitlink, or absent: all "not a tree here".
            Ok(_) => None,
            Err(e) => {
                tracing::debug!(
                    target: "znippy.git",
                    store = %self.abs.label,
                    oid = oid_hex,
                    error = %e,
                    "a reachability walk could not read a tree; the walk stops there"
                );
                None
            }
        }
    }
}

/// The verbatim archive as [`crate::resolve::ArchiveBytes`]: **the mapping
/// where it reaches, a `pread` past its end.**
///
/// One snapshot is taken per read and sliced as many times as the chain is
/// deep, which is [`crate::archive_map::ArchiveMap::snapshot`]'s contract. The
/// `pread` arm is the one a push that landed after the snapshot takes; both
/// arms return the same bytes and differ only in what they cost.
struct ArchiveWindow<'a, S: ObjectIndex> {
    abs: &'a Absorber<S>,
    snap: &'a crate::archive_map::Mapped,
}

impl<S: ObjectIndex> crate::resolve::ArchiveBytes for ArchiveWindow<'_, S> {
    fn at(&self, at: u64, len: u64) -> Result<std::borrow::Cow<'_, [u8]>> {
        // A probe near the end of the archive legitimately gets fewer bytes
        // than it asked for, so the window is clamped rather than refused —
        // the parsers read a grammar, not a fixed width.
        let mapped = self.snap.len();
        if at < mapped {
            let have = (mapped - at).min(len);
            if let Some(b) = self.snap.get(at, have) {
                return Ok(std::borrow::Cow::Borrowed(b));
            }
        }
        let end = self
            .abs
            .reader
            .metadata()
            .map(|m| m.len())
            .unwrap_or(mapped);
        if at >= end {
            return Ok(std::borrow::Cow::Owned(Vec::new()));
        }
        let have = (end - at).min(len);
        Ok(std::borrow::Cow::Owned(self.abs.read_extent(at, have)?))
    }
}

/// A thin pack's external delta base, answered from §14's exploded table.
///
/// Before that table existed this had to read and re-resolve a whole pack per
/// base, and it deliberately cached nothing — because a cache here *would* have
/// been the undecided table. It is decided now (eager, 2026-08-08), so the cache
/// is the table and this is one point lookup.
///
/// **`None` means absent and `Err` means unreadable, and they are not the same
/// answer.** A base this store does not hold still comes back `Ok(None)` and
/// [`crate::resolve::ResolveError::missing_base`] refuses the pack by name.
///
/// # 🔴 This was `.ok().flatten()`
///
/// A store that HELD the base and could not reconstruct it answered exactly
/// like one that never had it, so the resolver blamed the pushed pack: *"this
/// repository genuinely does not have that object"*, about an object one index
/// lookup away. That sentence went to the client on the `ng` line and to the
/// operator in the log, and it named the wrong party — lane AI, 2026-08-30, 21
/// of 140 ordinary pushes. The error travels now.
impl<S: ObjectIndex> BaseSource for Absorber<S> {
    fn content(&self, oid: &[u8]) -> Result<Option<(GitObjectKind, Vec<u8>)>> {
        self.resolved(oid).with_context(|| {
            format!(
                "reading the delta base {} out of this store — it is not absent, it could not \
                 be reconstructed",
                hex::encode(oid)
            )
        })
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::archive_write::{read_journal, SafeWriter};
    use crate::object::GitObjectKind;
    use crate::store::tests::{one_blob_pack, corpus_pack, tmpdir};
    use std::collections::HashSet;
    use std::sync::atomic::{AtomicBool, AtomicU64, Ordering as AtomicOrdering};
    use std::time::{Duration, Instant};
    use znippy_zoomies::background::Job;
    use znippy_zoomies::gatling_forkjoin::gatling_for_each;

    fn loadavg() -> String {
        std::fs::read_to_string("/proc/loadavg")
            .unwrap_or_default()
            .split_whitespace()
            .take(3)
            .collect::<Vec<_>>()
            .join(" ")
    }

    /// **The one representation answers exactly what the two it replaced did.**
    ///
    /// Before S-082 the store carried the ordinal space three times: `oids_raw`
    /// (flat bytes), `oids` (hex `Vec<String>`) and `ordinal`
    /// (`HashMap<String, u32>`). The last two are gone; oid ⇄ ordinal is now a
    /// slice and a binary search over `oids_raw`, through
    /// [`crate::reach::OrdinalSpace`]. This rebuilds those two OLD tables inline
    /// from the SAME `oids_raw` the store folded, and asserts the new lookups
    /// return the identical answer **for every oid in a real corpus** — the
    /// property that lets the field be removed without a server ever naming the
    /// wrong object in a pack, which no exit code could see.
    ///
    /// Seen RED by an off-by-one binary search (`Equal => Some(mid as u32 + 1)`
    /// in [`crate::reach::OrdinalSpace::ordinal_of_raw`]): every
    /// `ordinal_of_raw` / `ordinal_of_hex` below disagreed with the old
    /// `HashMap`, and the closure comparison at the end disagreed as well.
    #[test]
    fn the_one_ordinal_space_answers_what_the_two_it_replaced_did() {
        let dir = tmpdir("ordinal-space-one-representation");
        let (pack, rows) = corpus_pack();
        let store = GitStore::open(&dir, "rickard").unwrap();
        store.put_pack(&pack).unwrap();
        store.wait_indexed();

        let oid_len = store.hash_kind().oid_len();
        let raw = {
            let d = store.absorber.derived.read().unwrap();
            d.oids_raw.clone()
        };
        // The two tables `refold` used to build, rebuilt here from the same
        // bytes — this IS the "old" answer the new lookups must equal.
        let old_oids: Vec<String> = raw.chunks_exact(oid_len).map(hex::encode).collect();
        let old_ordinal: std::collections::HashMap<String, u32> = old_oids
            .iter()
            .enumerate()
            .map(|(i, o)| (o.clone(), i as u32))
            .collect();

        assert!(
            !old_oids.is_empty(),
            "the fixture pack ({} entries) folded to an EMPTY ordinal space, so every \
             comparison below would hold vacuously",
            rows.len()
        );
        assert_eq!(
            raw.len(),
            old_oids.len() * oid_len,
            "the raw ordinal space is {} bytes for {} oids of {oid_len} bytes",
            raw.len(),
            old_oids.len()
        );

        let space = crate::reach::OrdinalSpace::new(&raw, oid_len);
        assert_eq!(space.len(), old_oids.len(), "the space holds a different count");
        let mut checked = 0usize;
        for (o, h) in old_oids.iter().enumerate() {
            let want_ord = old_ordinal[h];
            // oid → ordinal, hex and raw, both == the old HashMap answer.
            assert_eq!(
                space.ordinal_of_hex(h),
                Some(want_ord),
                "ordinal_of_hex({h}) != the old HashMap answer {want_ord}"
            );
            let raw_oid = hex::decode(h).unwrap();
            assert_eq!(
                space.ordinal_of_raw(&raw_oid),
                Some(want_ord),
                "ordinal_of_raw disagrees with the old HashMap at {h}"
            );
            // ordinal → oid == the old hex Vec's slice, in both spellings.
            assert_eq!(
                space.oid(o as u32),
                Some(raw_oid.as_slice()),
                "oid({o}) != the old oids_raw slice"
            );
            assert_eq!(
                space.oid(o as u32).map(hex::encode).as_deref(),
                Some(h.as_str()),
                "oid({o}) hex != the old hex Vec entry {h}"
            );
            checked += 1;
        }
        assert_eq!(checked, old_oids.len());

        // A non-member answers None both ways, exactly as `HashMap::get` did for
        // an absent key — the binary search must not land on a neighbour.
        let absent = vec![0xffu8; oid_len];
        assert_eq!(space.ordinal_of_raw(&absent), None, "an absent oid resolved to an ordinal");
        assert_eq!(
            space.ordinal_of_hex(&hex::encode(&absent)),
            None,
            "an absent hex oid resolved to an ordinal"
        );
        // A wrong-width key is absent, not a panic.
        assert_eq!(space.ordinal_of_raw(&absent[..oid_len - 1]), None);
        assert_eq!(space.oid(old_oids.len() as u32), None, "an out-of-range ordinal named an oid");

        // And the answer built out of it: the raw closure IS the hex closure,
        // for a real tip rather than for an empty request.
        let tip = store
            .graph_snapshot()
            .into_iter()
            .max_by_key(|c| c.generation)
            .expect("a graph");
        let tip_raw = hex::decode(&tip.oid).unwrap();
        let flat = store.reachable_raw(&[&tip_raw], &[]).unwrap();
        let hexed = store.reachable_oids(&[&tip_raw], &[]).unwrap();
        assert!(
            flat.len() > 1,
            "the tip's closure is {} object(s) — too small to tell an ordering bug from a \
             coincidence",
            flat.len()
        );
        assert_eq!(
            flat.iter().map(hex::encode).collect::<Vec<_>>(),
            hexed,
            "the serving answer and the maintenance answer name different objects"
        );
        assert!(flat.contains(&tip_raw), "a closure without its own tip");
        eprintln!(
            "load {}; {} ordinals agree in both spellings, tip closure {} objects",
            loadavg(),
            checked,
            flat.len()
        );
    }

    /// Distinct commits in a resolved pack — what the graph must hold exactly
    /// once.
    fn commits_in(rows: &[crate::resolve::Resolved]) -> usize {
        rows.iter()
            .filter(|r| r.kind == GitObjectKind::Commit)
            .map(|r| r.oid.clone())
            .collect::<HashSet<Vec<u8>>>()
            .len()
    }

    /// **A push ends in object rows, and no read asked for them.**
    ///
    /// This is the seam this file wires: before it, `put` produced a *pack* row
    /// on the account indexer's channel and the object rows appeared only when
    /// somebody read (or called `absorb_pending` by hand). Here nothing is read
    /// between `put_pack` and `wait_indexed` — the only thing that could have
    /// built the rows is the drain.
    ///
    /// Asserted on applied output: one row per pack entry, every oid resolving
    /// out of the index, and the pack's indexed bit up. `absorb_failures` is
    /// asserted too, because a drain that absorbed nothing and recorded the
    /// failure would otherwise look like a drain that was never called.
    ///
    /// Seen RED by making the drain skip the absorber
    /// (`match absorber.as_deref()` → `match None::<&dyn ObjectAbsorb>` in
    /// `indexer::index_worker`): "the drain built no object rows: 0 of 2687".
    /// Restored.
    ///
    /// MEASURED on oden, 2026-08-08, release, three runs at **1-minute loadavg
    /// 5.77** (another tenant's work — oden is shared, so this is not an idle
    /// box): 2687 objects out of a 5 653 302-byte real pack absorbed in
    /// 161–164 ms, **16 421–16 673 object rows/s**, behind an ack of 29.1–29.4 ms.
    /// An earlier set at loadavg 16.94 gave 15 528 rows/s, so the figure moves
    /// about 7% with the box.
    ///
    /// Read it as a *floor* rather than as the engine's speed: the drain here is
    /// one pack, so it never uses more than one absorb, and the 162 ms covers a
    /// full inflate-and-delta-apply of every entry, a SHA-1 per object, the redb
    /// append and a whole-tail re-fold. Nothing about it has been optimised and
    /// no arm of it has been isolated. The rate is printed with the loadavg of
    /// its own run so a later number is never compared against a different
    /// machine state by accident.
    #[test]
    fn a_push_ends_in_object_rows_and_no_read_asked_for_them() {
        let dir = tmpdir("drain-rows");
        let store = GitStore::open(&dir, "rickard").unwrap();
        let (pack, rows) = corpus_pack();

        let t = Instant::now();
        let tx = store.put_pack(&pack).unwrap();
        let ack = t.elapsed();
        let pack_id = tx.pack_id.expect("a pack push assigns an id");

        // Nothing is read here. The drain is the only thing that can move.
        store.wait_indexed();
        let drained = t.elapsed();

        assert_eq!(
            store.object_count(),
            rows.len(),
            "the drain built no object rows: {} of {}",
            store.object_count(),
            rows.len()
        );
        assert_eq!(store.unindexed_packs(), 0, "the bit is still clear");
        assert!(
            store.indexer().is_indexed(pack_id),
            "the pack's indexed bit is not up after the drain"
        );
        assert_eq!(
            store.indexer().absorb_failures(),
            0,
            "the drain recorded an absorb failure: {:?}",
            store.indexer().last_absorb_error()
        );

        // Every oid answers **out of the rows**, not out of a fallback: the
        // index is asked directly, so nothing can absorb behind this loop.
        for (i, r) in rows.iter().enumerate() {
            let row = store
                .index()
                .lookup(&r.oid)
                .unwrap_or_else(|| panic!("object {i} {} has no row", hex::encode(&r.oid)));
            assert_eq!(
                (row.offset, row.len),
                (tx.extent.unwrap().0 + r.offset, r.len)
            );
            assert_eq!(row.uncompressed_size, r.uncompressed_size);
        }
        assert_eq!(
            store.absorb_pending().unwrap(),
            0,
            "the drain left work for a read to do"
        );
        assert_eq!(
            store.commit_count(),
            commits_in(&rows),
            "the commit graph does not match the pack"
        );

        let secs = (drained - ack).as_secs_f64();
        eprintln!(
            "load {}; {} objects in {} bytes: ack {:.0} µs, drain {:.1} ms, {:.0} object rows/s",
            loadavg(),
            rows.len(),
            pack.len(),
            ack.as_secs_f64() * 1e6,
            (drained - ack).as_secs_f64() * 1e3,
            rows.len() as f64 / secs.max(1e-9),
        );
    }

    // ── the bit across a restart ─────────────────────────────────────────────

    /// Where the child of the kill guard finds its fixture and leaves its marker.
    const KILL_DIR: &str = "GUNNAR_KILL_MID_ABSORB_DIR";

    /// **The other half of [`a_store_reopened_after_a_kill_mid_absorb_requeues_the_pack`]:
    /// the process that dies.**
    ///
    /// It runs as a child of that test and it **aborts on purpose**, which is why
    /// it is `#[ignore]`d: `SIGABRT` from inside a test binary is a failing test
    /// run, and the point is for the *parent* to see the signal.
    ///
    /// The kill lands where it has to land — after `put_pack` has returned, so
    /// the bytes and the journal row are on the platter, and before any row is in
    /// the index, which the held absorb gate makes a **state** rather than a
    /// hope. Nothing is simulated: no injected error, no fault flag, no early
    /// return. The process is gone.
    #[test]
    #[ignore = "spawned by a_store_reopened_after_a_kill_mid_absorb_requeues_the_pack; it aborts on purpose"]
    fn the_child_that_dies_between_durable_bytes_and_indexed_rows() {
        let Ok(dir) = std::env::var(KILL_DIR) else {
            // Run by hand with `--ignored` and no fixture: there is nothing to
            // kill over, so do nothing rather than abort somebody's test run.
            return;
        };
        let dir = PathBuf::from(dir);
        let pack = std::fs::read(dir.join("fixture.pack")).expect("the parent's fixture");

        let store = GitStore::open(&dir, "rickard").expect("open");
        // The drain is parked in front of its absorb for as long as this lives,
        // so no row can land before the abort below.
        let _gate = store.hold_absorb_gate();
        let tx = store.put_pack(&pack).expect("the ack path");
        assert_eq!(tx.pack_id, Some(0), "a fresh archive starts at ordinal 0");
        assert_eq!(store.unindexed_packs(), 1, "the pack is not queued");
        assert_eq!(store.index().len(), 0, "a row landed before the kill");
        std::fs::write(dir.join("ready"), b"durable, not indexed").expect("marker");

        // Not `panic!`, not `exit`: the process stops here with a live gate, a
        // live worker and a live redb handle, which is what a machine that dies
        // mid-absorb leaves behind.
        std::process::abort();
    }

    /// **A store reopened over a pack that was still being absorbed re-queues
    /// it, answers correctly the whole way through, and gets its derived tables
    /// back.**
    ///
    /// The interruption is a **real** one: a child process is spawned, pushes the
    /// pack, and is killed by `SIGABRT` between "bytes durable" and "rows
    /// indexed". The parent then asserts what the kill left on disk — one journal
    /// row, the pack verbatim, **zero** index rows — before opening a store over
    /// it, so the recovery is being tested against the state it claims to
    /// recover from and not against a lucky one.
    ///
    /// Three things are then asserted, all on applied output:
    ///
    /// 1. **the pack is re-queued** — the account indexer of the *new* process
    ///    publishes ordinal 0, which it can only do for a job somebody submitted;
    /// 2. **a read is correct throughout** — `has` is asked the instant the store
    ///    is open, before any drain has been waited for, and must answer `true`;
    /// 3. **`commit_count()` and `reachable()` come back** — the derived tables
    ///    are rebuilt by the same absorb, so the graph holds exactly the pack's
    ///    commits and a tip's closure is non-empty.
    ///
    /// Seen RED by making `open_with` re-queue nothing (`absorber
    /// .adopt_journal(&acked)?` → an empty `Vec`, which is what this store did
    /// before this change): "a durable object read `absent` from a store reopened
    /// over the pack that holds it — the crash-recovery bit did not re-queue
    /// anything". Restored.
    ///
    /// Seen RED a second time by inverting the diff in
    /// [`Absorber::adopt_journal`] (`if absorbed` → `if !absorbed`), so a pack
    /// with **no** rows has its bit set and a pack with rows is re-queued: the
    /// same first assertion fires, because a set bit over an empty index is
    /// exactly the wrong answer a bloom filter's false positive would have
    /// produced — the read stops falling back and says `absent` about an object
    /// whose bytes are durable. Restored.
    ///
    /// **Why the assertions are not immediately after the reopen.** The trap the
    /// drain guards already record applies here too, mirrored: right after
    /// `GitStore::open` the drain may legitimately have finished the re-queued
    /// pack already, so `unindexed_packs() == 1` is not assertable and is not
    /// asserted. What *is* assertable at that instant is the answer to a read,
    /// which must be `true` whichever side of the drain it lands on — and that is
    /// exactly the property §13.12 promises.
    ///
    /// **And it still holds now that a `gc` can retire a pack's journal row.**
    /// The tombstone tells "indexed then garbage-collected" from "never
    /// indexed", and this pack is the second: nothing retired it, so it must
    /// still be re-queued. Seen RED by making `adopt_journal` treat *every* pack
    /// as retired (`retired.contains(&extent.0)` → `true`): "a durable object
    /// read `absent` from a store reopened over the pack that holds it — the
    /// crash-recovery bit did not re-queue anything". Restored.
    #[test]
    fn a_store_reopened_after_a_kill_mid_absorb_requeues_the_pack() {
        use std::os::unix::process::ExitStatusExt;

        let dir = tmpdir("kill-mid-absorb");
        let (pack, rows) = corpus_pack();
        std::fs::write(dir.join("fixture.pack"), &pack).unwrap();

        let status = std::process::Command::new(std::env::current_exe().unwrap())
            .args([
                "--exact",
                "--ignored",
                "--nocapture",
                "git_ops::tests::the_child_that_dies_between_durable_bytes_and_indexed_rows",
            ])
            .env(KILL_DIR, &dir)
            .status()
            .expect("spawn the child that dies");
        assert_eq!(
            status.signal(),
            Some(6),
            "the child did not die by SIGABRT — it {status:?}, so nothing was interrupted"
        );
        assert!(
            dir.join("ready").exists(),
            "the child never reached the kill point"
        );

        // ── what the kill left on disk ───────────────────────────────────────
        let blobs = dir.join("objects.pack");
        let acked = read_journal(&SafeWriter::journal_path(&blobs)).unwrap();
        assert_eq!(
            acked.len(),
            1,
            "the pack's extent is not durable: {acked:?}"
        );
        assert_eq!(acked[0].1, pack.len() as u64);
        let on_disk = std::fs::read(&blobs).unwrap();
        assert_eq!(
            &on_disk[acked[0].0 as usize..(acked[0].0 + acked[0].1) as usize],
            &pack[..],
            "the verbatim bytes did not survive the kill"
        );
        {
            // Opened and dropped before the store takes redb's file lock.
            let tail = ObjectReadStack::<OneTableFourColumns>::open(
                &dir.join("objects.tail"),
                RebuildTriggers::default(),
                crate::arms::DEFAULT_REDB_CACHE_BYTES,
            )
            .unwrap();
            assert_eq!(
                tail.len(),
                0,
                "the kill landed after the rows were indexed, not between the bytes and the rows \
                 — this guard would then be testing nothing"
            );
        }

        // ── the reopen ───────────────────────────────────────────────────────
        let store = GitStore::open(&dir, "rickard").unwrap();
        assert!(
            store.has(&rows[0].oid).unwrap(),
            "a durable object read `absent` from a store reopened over the pack that holds it — \
             the crash-recovery bit did not re-queue anything"
        );

        store.wait_indexed();
        assert_eq!(
            store.object_count(),
            rows.len(),
            "the re-queued pack's objects never reached the index — left: {}, right: {}",
            store.object_count(),
            rows.len()
        );
        assert!(
            store.indexer().is_indexed(0),
            "the re-queued pack's indexed bit never went up in the new process"
        );
        assert_eq!(store.unindexed_packs(), 0);
        assert_eq!(store.absorb_pending().unwrap(), 0);
        for r in rows.iter().take(512) {
            let row = store
                .index()
                .lookup(&r.oid)
                .unwrap_or_else(|| panic!("{} has no row after recovery", hex::encode(&r.oid)));
            assert_eq!((row.offset, row.len), (acked[0].0 + r.offset, r.len));
        }

        // ── the derived tables, which were the second symptom ────────────────
        assert_eq!(
            store.commit_count(),
            commits_in(&rows),
            "the commit graph did not come back: {} of {}",
            store.commit_count(),
            commits_in(&rows)
        );
        let tip = store
            .graph_snapshot()
            .into_iter()
            .max_by_key(|c| c.generation)
            .expect("a graph");
        let tip_raw = hex::decode(&tip.oid).unwrap();
        let closure = store.reachable(&[&tip_raw], &[]).unwrap();
        assert!(
            !closure.is_empty(),
            "reachable() came back empty after the reopen"
        );
        assert!(closure.contains(&tip_raw));
        eprintln!(
            "load {}; killed mid-absorb: {} journal extent(s) recovered {} object rows, {} \
             commits, closure of the tip {} objects",
            loadavg(),
            acked.len(),
            store.object_count(),
            store.commit_count(),
            closure.len(),
        );
    }

    /// **A clean reopen re-queues nothing, and the next push gets a fresh
    /// ordinal.**
    ///
    /// The other side of the diff. A store whose pack was fully absorbed before
    /// it closed must come back with the bit **set**: re-absorbing would be a
    /// whole pack re-resolved for nothing, and — because the commit graph is a
    /// `Vec` — it is the same doubling the drain's own guard exists to prevent.
    ///
    /// "Re-queued nothing" is asserted deterministically rather than by a timing
    /// window: the account indexer in this process publishes a pack only for a
    /// job somebody submitted, so `is_indexed(0)` being **false** after
    /// `wait_indexed` is proof that open submitted none.
    ///
    /// The second half is the ordinal. Pack ids are the journal's row numbers, so
    /// a reopened store must resume after them. Asserted on applied output rather
    /// than on the id alone: the newly pushed pack's object has to be findable,
    /// and it is exactly what goes missing when a fresh pack is handed an
    /// absorbed pack's ordinal — `queue` skips it, the drain skips it, and its
    /// durable bytes are never indexed by anybody.
    ///
    /// Seen RED by making `Absorber::adopt_journal` treat every journal extent as
    /// unabsorbed (`let has_rows = vec![false; journal.len()]`): "a fully
    /// absorbed pack was re-queued on open — the diff is not exact".
    ///
    /// Seen RED a second time by `AtomicU64::new(packs_already_acked(...)?)` →
    /// `AtomicU64::new(0)` in `PushPath::with_absorber`: "a reopened store handed
    /// a fresh pack the ordinal of an absorbed one — left: Some(0), right:
    /// Some(1)". With that same edit **and the id assertion deleted**, the
    /// applied-output one goes too — "the second push's object never reached the
    /// index — its ordinal collided with an absorbed pack's, so every writer
    /// skipped it" — which is what says this guard is about the objects and not
    /// about a counter. Restored.
    #[test]
    fn a_clean_reopen_requeues_nothing_and_the_next_push_gets_a_fresh_ordinal() {
        let dir = tmpdir("clean-reopen");
        let (pack, rows) = corpus_pack();
        let (second, blob_oid) = one_blob_pack(b"a push that arrives after the restart");

        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
            assert_eq!(
                store.object_count(),
                rows.len(),
                "the first run did not index"
            );
        }

        let store = GitStore::open(&dir, "rickard").unwrap();
        store.wait_indexed();
        assert!(
            !store.indexer().is_indexed(0),
            "a fully absorbed pack was re-queued on open — the diff is not exact"
        );
        assert_eq!(store.unindexed_packs(), 0);
        assert_eq!(
            store.absorb_pending().unwrap(),
            0,
            "the reopen left index work"
        );
        assert_eq!(
            store.object_count(),
            rows.len(),
            "the reopened store lost rows"
        );
        assert!(store.has(&rows[0].oid).unwrap());

        let tx = store.put_pack(&second).unwrap();
        assert_eq!(
            tx.pack_id,
            Some(1),
            "a reopened store handed a fresh pack the ordinal of an absorbed one"
        );
        store.wait_indexed();
        assert!(
            store.has(&blob_oid).unwrap(),
            "the second push's object never reached the index — its ordinal collided with an \
             absorbed pack's, so every writer skipped it"
        );
        assert_eq!(store.object_count(), rows.len() + 1);
        assert!(store.indexer().is_indexed(1));
        // And the journal is a log: both packs' extents are in it, in order.
        let acked = read_journal(&SafeWriter::journal_path(store.blobs_path())).unwrap();
        assert_eq!(
            acked.len(),
            2,
            "the reopen truncated the journal: {acked:?}"
        );
        assert_eq!(acked[1], tx.extent.unwrap());
    }

    /// **What the derive-on-open diff costs at a realistic pack count.**
    ///
    /// Not a guard — a measurement, with two assertions on it that keep it from
    /// measuring the wrong thing: every absorbed pack must come back `true`, and
    /// the one pack that has no rows must come back `false`.
    ///
    /// Both cases are timed because they are different algorithms in practice.
    /// The **clean reopen** — every pack absorbed — terminates as soon as each
    /// extent has been hit once, which happens after a few rows per pack because
    /// the tail is in oid order and an oid says nothing about which pack its
    /// object came from. The **crash** case has one pack with no rows at all, and
    /// proving an absence means reading the tail to the end.
    ///
    /// MEASURED on oden 2026-08-08, release, 1000 real pushed packs / 3686 object
    /// rows, 1-minute loadavg 2.77: **0.38 ms** clean and 0.39 ms with one pack
    /// un-absorbed (0.40 / 0.39 ms at loadavg 13.06, so the figure barely moves
    /// with the box). At this row count the two cases are the same, because a
    /// 3686-row tail is scanned to the end faster than the early exit saves
    /// anything; they diverge with the object count, and the table on
    /// [`ObjectReadStack::extents_with_rows`] carries that at a million rows
    /// (0.74 ms against 109 ms).
    ///
    /// The printed line carries the loadavg of its own run, the pack count and
    /// the row count, because none of the three is comparable across machines or
    /// across a busy box.
    #[test]
    fn the_derive_on_open_diff_costs_milliseconds_at_a_realistic_pack_count() {
        const PACKS: usize = 1000;
        let dir = tmpdir("diff-cost");
        let (real, rows) = corpus_pack();

        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&real).unwrap();
            for i in 0..PACKS - 1 {
                let (p, _) = one_blob_pack(format!("pack number {i}").as_bytes());
                store.put_pack(&p).unwrap();
            }
            store.wait_indexed();
            assert_eq!(
                store.object_count(),
                rows.len() + PACKS - 1,
                "the fixture did not absorb"
            );
        }

        let acked = read_journal(&SafeWriter::journal_path(&dir.join("objects.pack"))).unwrap();
        assert_eq!(acked.len(), PACKS, "one journal row per pack");
        // The tail on its own, so what is timed is the diff rather than redb's
        // open and the projection build that a store's open also pays for.
        let tail = ObjectReadStack::<OneTableFourColumns>::open(
            &dir.join("objects.tail"),
            RebuildTriggers::default(),
            crate::arms::DEFAULT_REDB_CACHE_BYTES,
        )
        .unwrap();

        let t = Instant::now();
        let hit = tail.extents_with_rows(&acked).unwrap();
        let clean = t.elapsed();
        assert!(
            hit.iter().all(|&h| h),
            "an absorbed pack was reported unabsorbed"
        );

        // One pack in flight when the machine died: durable bytes, no rows.
        let mut in_flight = acked.clone();
        let end = acked[PACKS - 1].0 + acked[PACKS - 1].1;
        in_flight.push((end, 4096));
        let t = Instant::now();
        let hit = tail.extents_with_rows(&in_flight).unwrap();
        let crashed = t.elapsed();
        assert_eq!(
            hit.iter().filter(|h| !**h).count(),
            1,
            "the pack with no rows is not the only one reported unabsorbed"
        );

        eprintln!(
            "load {}; {PACKS} packs / {} object rows: derive-on-open diff {:.2} ms clean \
             (early exit), {:.2} ms with one pack un-absorbed (full tail scan)",
            loadavg(),
            tail.len(),
            clean.as_secs_f64() * 1e3,
            crashed.as_secs_f64() * 1e3,
        );
    }

    /// **The bit is a bitset over dense ordinals, and one bit is one pack.**
    ///
    /// Ordinals that straddle word boundaries (63, 64, 65) are the whole point:
    /// a shift that ignores the word split, or a word index that does not, sets
    /// a neighbour's bit — and a neighbour's bit set means a durable pack is
    /// declared absorbed with no rows behind it.
    ///
    /// Seen RED by `>> (i % 64) & 1` → `>> (i % 63) & 1` in
    /// `PackState::is_absorbed`: "the bit for ordinal 64 did not go up". Restored.
    ///
    /// Seen RED a second time by `if self.extent[i].is_none()` → `if
    /// self.extent[i].is_none() || true` in `PackState::note`, i.e. letting a
    /// re-queue of a pack that is already recorded count as new work — which is
    /// the race the ack path really produces, since the drain can absorb before
    /// `queue` runs: "an absorbed pack was queued again — left: 6, right: 5".
    /// Restored.
    #[test]
    fn the_indexed_bit_is_a_bitset_over_dense_pack_ordinals() {
        let ids = [0u64, 1, 63, 64, 65, 4095];
        let mut s = PackState::default();
        for id in ids {
            s.note(id, (id * 4096 + 12, 4096));
        }
        assert_eq!(s.unabsorbed, ids.len());
        assert!(
            ids.iter().all(|&i| !s.is_absorbed(i)),
            "a bit is up already"
        );

        s.mark_absorbed(64, (64 * 4096 + 12, 4096));
        assert!(s.is_absorbed(64), "the bit for ordinal 64 did not go up");
        for other in [0u64, 1, 63, 65, 4095] {
            assert!(
                !s.is_absorbed(other),
                "setting ordinal 64 also set ordinal {other} — one bit spilled onto a neighbour"
            );
        }
        assert_eq!(s.unabsorbed, ids.len() - 1);
        assert_eq!(
            s.pending().iter().map(|p| p.pack_id).collect::<Vec<_>>(),
            vec![0, 1, 63, 65, 4095],
            "the pending list is not the clear bits"
        );
        assert_eq!(
            s.pending()[0],
            PendingPack {
                pack_id: 0,
                offset: 12,
                len: 4096,
                resumed: true,
            },
            "a pending pack lost the extent it must be absorbed from"
        );

        // Both writers are idempotent, because both of them race.
        s.mark_absorbed(64, (64 * 4096 + 12, 4096));
        s.note(64, (64 * 4096 + 12, 4096));
        assert!(s.is_absorbed(64));
        assert_eq!(
            s.unabsorbed,
            ids.len() - 1,
            "an absorbed pack was queued again"
        );

        // One bit per pack, and that is the whole cost: 4096 ordinals is 64
        // words. A hash set of the same ordinals is an order of magnitude more,
        // and a bloom would be smaller and wrong.
        assert_eq!(s.words.len(), 64);
        assert_eq!(s.words.len() * 8, 512, "4096 packs is 512 bytes of bits");
    }

    /// **A read that arrives before the drain is served from the pack, on its
    /// own thread, and never waits for the worker.**
    ///
    /// Rickard's rule, 2026-08-30: *nobody waits on the indexer.* Until then
    /// this test asserted the opposite — that a read parked behind the absorb
    /// gate stayed unanswered until the drain released it, which on a kernel
    /// pack is hours, and under the memory watchdog is *as long as the box is
    /// short*. Now the gate is the absorbers' and only the absorbers': the read
    /// finds no row, sees a pending pack, and resolves that pack read-only
    /// from the mmapped bytes for the one oid it wants
    /// (`Absorber::scan_pending`). The drain is still parked, the index is
    /// still empty, and the answer is **present** — from the verbatim truth.
    ///
    /// Seen RED by making `lookup_never_wait` return `Ok(None)` when the index
    /// misses: "a read with the drain parked answered absent for a durable
    /// object". And by putting `self.gate.lock()` back in front of it: the
    /// 2-second bound below trips with the reader still parked.
    #[test]
    fn a_read_before_the_drain_is_served_from_the_pack_and_never_waits() {
        let dir = tmpdir("drain-before");
        let store = Arc::new(GitStore::open(&dir, "rickard").unwrap());
        let (pack, rows) = corpus_pack();
        let (second, blob_oid) = one_blob_pack(b"a second push while the drain is parked");

        // The drain is parked here, before it can absorb anything.
        let gate = store.hold_absorb_gate();
        let tx = store.put_pack(&pack).unwrap();
        let pack_id = tx.pack_id.unwrap();
        assert_eq!(store.unindexed_packs(), 1, "the pack is not queued");
        assert_eq!(store.index().len(), 0, "the index already holds rows");
        assert!(!store.indexer().is_indexed(pack_id), "the bit is already up");

        // (1) a second push does not wait either: the ack path never meets the gate.
        let t = Instant::now();
        let tx2 = store.put_pack(&second).unwrap();
        assert_eq!(tx2.pack_id, Some(pack_id + 1));
        assert!(t.elapsed() < Duration::from_secs(2), "the push waited on the parked drain");

        // (2) reads into exactly that state: present, from the pack, bounded.
        let oid = rows[0].oid.clone();
        let last = rows[rows.len() - 1].oid.clone();
        let t = Instant::now();
        assert!(store.has(&oid).unwrap(), "a read with the drain parked answered absent for a durable object");
        assert!(store.has(&last).unwrap(), "the last object of the pack answered absent");
        assert!(store.has(&blob_oid).unwrap(), "the second (still pending) pack's object answered absent");
        let (kind, body) = store.content(&blob_oid).unwrap().expect("content from the pending pack");
        assert_eq!(kind, GitObjectKind::Blob);
        assert_eq!(body, b"a second push while the drain is parked");
        let stored = store.get(&oid).unwrap().expect("get from the pending pack");
        assert_eq!(stored.extent.0, rows[0].offset + tx.extent.unwrap().0);
        assert_eq!(
            store.size(&last).unwrap(),
            Some(rows[rows.len() - 1].uncompressed_size),
            "size from the pending pack"
        );
        let absent = vec![0xEEu8; 20];
        assert!(!store.has(&absent).unwrap(), "an oid in no pack must still be absent");
        let elapsed = t.elapsed();
        assert!(elapsed < Duration::from_secs(20), "reads took {elapsed:?} with the drain parked");
        // The drain is STILL parked and the index STILL empty: nothing above
        // absorbed anything, and nothing waited.
        assert_eq!(store.index().len(), 0, "a read absorbed the pack behind the parked gate");
        assert!(!store.indexer().is_indexed(pack_id));
        assert_eq!(store.unindexed_packs(), 2);
        eprintln!("{} reads served from two pending packs with the drain parked, in {elapsed:?}", 6);

        // (3) a whole-graph request is refused by name, not waited on.
        let err = store.reachable(&[oid.as_slice()], &[]).unwrap_err();
        assert!(err.to_string().contains("index in progress"), "{err:#}");

        drop(gate);
        // And afterwards it is served from the rows.
        store.wait_indexed();
        assert_eq!(store.unindexed_packs(), 0);
        assert!(store.indexer().is_indexed(pack_id));
        assert_eq!(store.index().len(), rows.len() + 1);
        assert_eq!(store.absorb_pending().unwrap(), 0);
        for r in rows.iter().take(256) {
            assert!(
                store.index().lookup(&r.oid).is_some(),
                "{} is not in the rows after the drain",
                hex::encode(&r.oid)
            );
        }
        assert!(store.reachable(&[oid.as_slice()], &[]).is_ok(), "reachable after the drain");
    }

    /// **The memory watchdog pauses the indexer worker and nobody else.**
    ///
    /// A fake `/proc/meminfo` under the floor is put under the process-wide
    /// watchdog; the worker parks in front of its batch (`index_worker`'s
    /// pre-batch check) and stays parked. Meanwhile a push is acked, every
    /// kind of read answers from the pack, and the progress snapshot returns —
    /// all bounded, none of them behind the pause. Then the fake is raised,
    /// the worker resumes, and the pack is indexed normally.
    ///
    /// Seen RED with the pre-batch check taking the absorb gate before pausing
    /// (the reads' scan does not need the gate, but `absorb_pending`-based
    /// reads did): every read parked for the length of the pause.
    #[test]
    fn a_paused_indexer_blocks_no_push_and_no_read() {
        let dir = tmpdir("paused-indexer");
        let meminfo = dir.join("meminfo");
        let write = |avail_kib: u64| {
            std::fs::write(&meminfo, format!("MemTotal:       16000000 kB\nMemAvailable:   {avail_kib} kB\n"))
                .unwrap()
        };
        write(1_000_000);
        let dog = Arc::new(crate::memguard::MemWatchdog::with(
            &meminfo,
            3_200_000 * 1024,
            Duration::from_millis(10),
        ));
        let store = Arc::new(GitStore::open(&dir, "rickard").unwrap());
        // This store's worker only: every other store in the process keeps
        // the real watchdog.
        store.set_mem_watchdog(dog.clone());
        let (pack, rows) = corpus_pack();
        let (second, blob_oid) = one_blob_pack(b"pushed while the indexer is paused");

        let t_push = Instant::now();
        let tx = store.put_pack(&pack).unwrap();
        let pack_id = tx.pack_id.unwrap();
        // Give the worker time to reach its pre-batch check and park.
        let mut parked = false;
        for _ in 0..200 {
            if dog.is_paused() {
                parked = true;
                break;
            }
            std::thread::sleep(Duration::from_millis(10));
        }
        assert!(parked, "the worker never paused under a meminfo 1 GB < 3.2 GB");
        assert!(!store.indexer().is_indexed(pack_id), "indexed while paused");
        assert_eq!(store.index().len(), 0);

        // The ack path: a push while paused.
        let tx2 = store.put_pack(&second).unwrap();
        assert_eq!(tx2.pack_id, Some(pack_id + 1));
        assert!(t_push.elapsed() < Duration::from_secs(5), "a push waited on the paused worker");

        // Reads while paused: present, from the pack.
        let t = Instant::now();
        assert!(store.has(&rows[0].oid).unwrap());
        assert!(store.has(&blob_oid).unwrap());
        let (kind, body) = store.content(&blob_oid).unwrap().unwrap();
        assert_eq!((kind, body.as_slice()), (GitObjectKind::Blob, &b"pushed while the indexer is paused"[..]));
        assert_eq!(
            store.extents(&[rows[0].oid.as_slice(), rows[1].oid.as_slice(), blob_oid.as_slice()])
                .unwrap()
                .iter()
                .filter(|e| e.is_some())
                .count(),
            3,
            "the batch path answered absent while paused"
        );
        assert!(t.elapsed() < Duration::from_secs(20), "reads waited on the paused worker");
        // The progress row: information, never a gate.
        let snap = store.index_progress();
        assert_eq!(snap.root, store.root().display().to_string());
        assert!(dog.is_paused(), "the worker resumed on its own — the fake still says 1 GB");
        assert!(!store.indexer().is_indexed(pack_id), "indexed while still paused");

        // A whole-graph request while the worker is paused: the gate is free
        // (a paused worker holds nothing), so the READER absorbs the pending
        // packs on its own thread and answers — the worker is still parked.
        let reach = store.reachable(&[rows[0].oid.as_slice()], &[]).unwrap();
        assert!(!reach.is_empty());
        assert_eq!(store.unindexed_packs(), 0, "the reader did not absorb for itself");
        assert!(dog.is_paused(), "the worker is still parked; the reader did the work");
        assert_eq!(store.index().len(), rows.len() + 1);

        // Memory comes back: the worker resumes, finds nothing to do, and the
        // bits go up through the ordinary drain.
        write(9_000_000);
        store.wait_indexed();
        assert!(store.indexer().is_indexed(pack_id));
        assert!(store.indexer().is_indexed(pack_id + 1));
        assert_eq!(dog.pauses(), 1, "one pause, one resume");
        assert!(!dog.is_paused());
    }

    /// **The bit is cleared after the rows go in, never before.**
    ///
    /// The ordering inside `absorb_gated` is the whole reason a read cannot be
    /// told "absent" while the drain is running, and the only way to see it is to
    /// read *during* one. Four gatling workers (LAW 3 — never rayon, and never a
    /// raw thread pool) hammer `has` over the pack's oids from the moment
    /// `put_pack` returns until the pack's bit is up.
    ///
    /// Asserted on applied output: not one of those reads may answer absent for
    /// an object whose bytes are already durable.
    ///
    /// Seen RED by clearing the bit *before* absorbing (a
    /// `s.pending.retain(|p| p.pack_id != job.pack_id)` added in `absorb_gated`
    /// above the `self.absorb_one(job)` call): "4 reader(s) were told a durable
    /// object was absent — **32 865 593** absent answers over 2687 objects".
    /// Restored. The count is that large because the readers spin until the bit
    /// goes up and, with the bit cleared first, every pass over the sample missed
    /// for the whole length of the absorb.
    ///
    /// The honest limit: this is a race, so a green run proves the window was not
    /// hit, not that no window exists. That is why the printed line carries how
    /// many of the reads actually landed while the pack was still un-indexed —
    /// **a run where that number is 0 proved nothing**, and it is reported rather
    /// than asserted because forcing it would mean pausing the drain, which is
    /// the one thing this test must not do. MEASURED 2026-08-08 on oden at
    /// loadavg 16.94: 336 of 336 reads landed early.
    #[test]
    fn no_read_is_told_absent_while_the_drain_is_running() {
        let dir = tmpdir("drain-during");
        let store = GitStore::open(&dir, "rickard").unwrap();
        let (pack, rows) = corpus_pack();
        let oids: Vec<&[u8]> = rows.iter().map(|r| r.oid.as_slice()).collect();

        let tx = store.put_pack(&pack).unwrap();
        let pack_id = tx.pack_id.unwrap();

        let early = AtomicU64::new(0);
        let absent = AtomicU64::new(0);
        let reads = AtomicU64::new(0);
        let workers = 4usize;
        gatling_for_each(workers, workers, |w| loop {
            let still_pending = store.unindexed_packs() > 0;
            for oid in oids.iter().skip(w).step_by(workers * 8) {
                if still_pending {
                    early.fetch_add(1, AtomicOrdering::Relaxed);
                }
                reads.fetch_add(1, AtomicOrdering::Relaxed);
                if !store.has(oid).unwrap() {
                    absent.fetch_add(1, AtomicOrdering::Relaxed);
                }
            }
            if store.indexer().is_indexed(pack_id) {
                break;
            }
        });

        let absent = absent.load(AtomicOrdering::Acquire);
        assert_eq!(
            absent,
            0,
            "{workers} reader(s) were told a durable object was absent — {absent} absent answers \
             over {} objects",
            rows.len()
        );
        assert_eq!(store.object_count(), rows.len());
        eprintln!(
            "load {}; {} reads across {workers} gatling workers, {} of them while the pack was \
             still un-indexed",
            loadavg(),
            reads.load(AtomicOrdering::Acquire),
            early.load(AtomicOrdering::Acquire),
        );
    }

    /// **One pack is absorbed once, whichever of the two callers gets there
    /// first.**
    ///
    /// Both callers are made to attempt it, deterministically: with the gate
    /// held, the drain is parked inside `ObjectAbsorb::absorb` and a reader is
    /// parked inside `absorb_pending`, both past their own early exits. Releasing
    /// the gate lets exactly one do the work; the other must take the skip.
    ///
    /// The observable is the **commit graph**, not the object table:
    /// `ObjectReadStack::append` is keyed by oid and would swallow a second
    /// identical row, but the graph is a `Vec` and a second absorb pushes every
    /// commit into it again — which silently doubles the fold every reachability
    /// bitmap is built from.
    ///
    /// Seen RED by disabling the `s.absorbed.contains(&job.pack_id)` skip in
    /// `absorb_gated`: "the commit graph carries **1102** commits, the pack has
    /// **551**" — exactly double, both callers having done the whole fold.
    /// Restored.
    #[test]
    fn a_pack_is_absorbed_once_even_when_both_callers_race_for_it() {
        let dir = tmpdir("drain-once");
        let store = Arc::new(GitStore::open(&dir, "rickard").unwrap());
        let (pack, rows) = corpus_pack();

        let gate = store.hold_absorb_gate();
        store.put_pack(&pack).unwrap();

        let s = store.clone();
        let reader = Job::spawn(move || s.absorb_pending().unwrap());
        // Both are now blocked on the gate: the drain inside `absorb`, the reader
        // inside `absorb_pending`, each having already decided there is work.
        std::thread::sleep(Duration::from_millis(100));
        drop(gate);

        let by_read = reader.join().unwrap();
        store.wait_indexed();

        assert_eq!(
            store.commit_count(),
            commits_in(&rows),
            "the commit graph carries {} commits, the pack has {}",
            store.commit_count(),
            commits_in(&rows)
        );
        assert_eq!(store.object_count(), rows.len());
        assert_eq!(store.unindexed_packs(), 0);
        eprintln!(
            "load {}; the read absorbed {by_read} pack(s), the drain absorbed the rest",
            loadavg()
        );
    }

    // ── §14's exploded table ─────────────────────────────────────────────────

    /// **A push produces an exploded row for every object in the pack, and each
    /// one hashes back to its own oid.**
    ///
    /// Eager, decided 2026-08-08: no threshold, no size cap, no type skipped —
    /// so the row count is the *object* count and not some subset of it. Nothing
    /// is read between `put_pack` and `wait_indexed`, so the only thing that
    /// could have built the rows is the background drain, which is the half of
    /// §14 that says "built by the same background indexer".
    ///
    /// The assertion is applied output in its strongest available form: every
    /// object's content is read back out of the table, re-serialised canonically
    /// and **re-hashed**, and the hash has to be the oid it was filed under. A
    /// row count alone could be satisfied by 2687 rows of the wrong bytes; a
    /// re-hash cannot.
    ///
    /// Seen RED by deleting the `sink.explode(&oid, kind, &payload)?` call in
    /// `resolve::resolve_walked`: "the drain built no exploded rows: 0 of 2687".
    /// Restored.
    ///
    /// Seen RED a second time by making the sink skip blobs in
    /// `ExplodedTable::explode` (`if kind == GitObjectKind::Blob { return Ok(()) }`
    /// at the top — which is what a *lazy* or size-capped table would look like
    /// from here): "the drain built no exploded rows: 1420 of 2687". Restored,
    /// because §14's open question was closed EAGER and disk is explicitly not a
    /// consideration — and because the 1267 rows that edit removes are exactly
    /// the blobs a content read wants most.
    #[test]
    fn a_push_produces_an_exploded_row_for_every_object_in_the_pack() {
        let dir = tmpdir("exploded-every-object");
        let store = GitStore::open(&dir, "rickard").unwrap();
        let (pack, rows) = corpus_pack();

        let t = Instant::now();
        store.put_pack(&pack).unwrap();
        let ack = t.elapsed();
        store.wait_indexed();
        // **Drain, printed beside ack**, because the two answer different
        // questions and only one of them was ever in doubt. Ack is unchanged by
        // construction — nothing on it touches this table. Drain is where the
        // eager resolve and the sink both live, and it is the column redb's
        // 2026-08-08 measurement recorded as 316.2–329.8 ms against 166.7–168.6
        // ms with the sink stubbed out. Without it here, "did the medium buy
        // anything but disk" has no answer in this repository.
        let drain = t.elapsed();

        let distinct: HashSet<Vec<u8>> = rows.iter().map(|r| r.oid.clone()).collect();
        let stats = store.exploded_stats();
        // **The push explodes the GRAPH, not the content.** Under the shipped
        // policy the drain writes commits and trees and leaves blobs in the
        // pack; the assertion is therefore two-sided, because "some rows" would
        // pass on a table that wrote one row and on one that wrote all of them,
        // and it is exactly the second that costs 17× the pack.
        let graph_rows = store.exploded_of_kind(GitObjectKind::Commit).unwrap().len();
        assert_eq!(
            store.exploded_eager_rows().unwrap(),
            graph_rows as u64,
            "the drain wrote {} eager rows and there are {} commits — the eager pool \
             holds something other than the graph",
            store.exploded_eager_rows().unwrap(),
            graph_rows
        );
        assert!(
            stats.rows > 0 && stats.rows < distinct.len() as u64,
            "the eager pool holds {} of {} objects; lazy explode means neither 0 \
             (the graph would be blind) nor all of them (that is the 22.4x tier)",
            stats.rows,
            distinct.len()
        );

        // Every one of them, re-hashed. Bytes that came back wrong cannot pass.
        let hash = store.hash_kind();
        let mut checked = 0usize;
        for oid in &distinct {
            let (kind, payload) = store
                .content(oid)
                .unwrap()
                .unwrap_or_else(|| panic!("{} has no exploded row", hex::encode(oid)));
            let rehashed = hash.oid_of(&crate::object::canonical(kind, &payload));
            assert_eq!(
                &rehashed,
                oid,
                "the exploded row filed under {} contains an object that hashes to {}",
                hex::encode(oid),
                hex::encode(&rehashed)
            );
            checked += 1;
        }
        assert_eq!(checked, distinct.len());
        // **Every object came back byte-correct**, and the ones the eager pool
        // does not hold came back by inflating their own delta chain out of the
        // mapped pack ([`crate::resolve::materialize_at`]). So the re-derivation
        // count is the blob count and not zero — asserting zero here is what a
        // test written for the eager tier does, and it would now be asserting
        // that lazy explode is off.
        let after_reads = store.exploded_stats();
        assert!(
            after_reads.rederived > 0,
            "no read re-derived anything, so either every object is eager or the \
             lazy path never ran"
        );
        assert_eq!(
            after_reads.rederived + after_reads.served,
            checked as u64,
            "reads {} = served {} + rederived {} must account for every object",
            checked,
            after_reads.served,
            after_reads.rederived
        );
        // **What it costs on disk, measured rather than assumed.** §14 says the
        // price is "both copies are held" and the decision was taken with "we
        // don't care if disk is tripled" — so the real multiple belongs in the
        // output where anybody can read it, not in a sentence. It is not 3x on a
        // real pack: the verbatim bytes are deflated *and* delta-encoded, the
        // table holds raw inflated content, and redb pages it.
        let inflated: u64 = rows.iter().map(|r| r.uncompressed_size).sum();
        let table = std::fs::metadata(store.exploded_path())
            .map(|m| m.len())
            .unwrap_or(0);
        eprintln!(
            "load {}; {} pack entries → {} exploded rows, all {} re-hashed to their own oid; \
             ack {:.1} ms, drain {:.1} ms, {:.0} rows/s; disk: pack {:.1} MiB verbatim, objects \
             inflate to {:.1} MiB, table file {:.1} MiB ({:.0}x the pack, {:.3}x the payload)",
            loadavg(),
            rows.len(),
            stats.rows,
            checked,
            ack.as_secs_f64() * 1e3,
            drain.as_secs_f64() * 1e3,
            stats.rows as f64 / drain.as_secs_f64(),
            pack.len() as f64 / (1 << 20) as f64,
            inflated as f64 / (1 << 20) as f64,
            table as f64 / (1 << 20) as f64,
            table as f64 / pack.len() as f64,
            table as f64 / inflated.max(1) as f64,
        );
    }

    /// **The amplification measurement, at whatever scale you point it at.**
    ///
    /// [`a_push_produces_an_exploded_row_for_every_object_in_the_pack`] prints
    /// the same ratio on a 5.4 MiB fixture and is a *guard*, so it has to stay
    /// fast. This is the same arithmetic with no assertion about size and no
    /// fixture of its own: give it a pack and it reports what the table cost.
    ///
    /// ```text
    /// ZNIPPY_EXPLODE_BENCH_PACK=/path/to/pack-….pack \
    ///   cargo test -p znippy-plugin-git --lib the_table_costs_what_it_holds \
    ///   -- --ignored --nocapture
    /// ```
    ///
    /// It exists because the 12× figure that killed redb — 16.8 GB of resolved
    /// `linux.git` content in a 204 GB file — was taken by hand off a bench run
    /// that no longer exists, and a number nobody can re-take is a number that
    /// rots. `#[ignore]`d and pack-less by default: with no pack named it
    /// returns, so it never fails a suite for the absence of a corpus nobody
    /// promised to keep.
    #[test]
    #[ignore = "needs a pack: ZNIPPY_EXPLODE_BENCH_PACK=<path> cargo test … -- --ignored --nocapture"]
    fn the_table_costs_what_it_holds() {
        let Ok(path) = std::env::var("ZNIPPY_EXPLODE_BENCH_PACK") else {
            eprintln!("no ZNIPPY_EXPLODE_BENCH_PACK named; nothing measured");
            return;
        };
        let pack = std::fs::read(&path).unwrap_or_else(|e| panic!("reading {path}: {e}"));
        let dir = tmpdir("exploded-cost");
        let store = GitStore::open(&dir, "rickard").unwrap();

        let t = Instant::now();
        store.put_pack(&pack).unwrap();
        let ack = t.elapsed();
        store.wait_indexed();
        let drain = t.elapsed();

        let stats = store.exploded_stats();
        let table = std::fs::metadata(store.exploded_path())
            .map(|m| m.len())
            .unwrap_or(0);
        // The payload the table actually holds, asked of the table rather than
        // of the pack — a pack entry's `uncompressed_size` is what git recorded,
        // and this is what we stored. Only the second one can be divided into
        // the file size and mean anything.
        let mut held = 0u64;
        for kind in [
            GitObjectKind::Commit,
            GitObjectKind::Tree,
            GitObjectKind::Blob,
            GitObjectKind::Tag,
        ] {
            held += store
                .exploded_of_kind(kind)
                .unwrap()
                .iter()
                .map(|(_, p)| p.len() as u64)
                .sum::<u64>();
        }
        let mib = |n: u64| n as f64 / (1 << 20) as f64;
        eprintln!(
            "load {}; pack {} — {:.1} MiB verbatim → {} rows holding {:.1} MiB, table file \
             {:.1} MiB = {:.3}x the payload and {:.1}x the pack; ack {:.0} ms, drain {:.0} ms",
            loadavg(),
            path,
            mib(pack.len() as u64),
            stats.rows,
            mib(held),
            mib(table),
            table as f64 / held.max(1) as f64,
            table as f64 / pack.len() as f64,
            ack.as_secs_f64() * 1e3,
            drain.as_secs_f64() * 1e3,
        );
    }

    /// Where the child of the clean-shutdown guard finds its fixture and leaves
    /// its markers.
    const CLEAN_DIR: &str = "GUNNAR_CLEAN_SHUTDOWN_DIR";

    /// **The other half of
    /// [`a_clean_shutdown_and_reopen_keeps_the_graph_and_reachable`]: the process
    /// that exits normally.**
    ///
    /// A real process, a real push, a real drain, and then a real `drop` — every
    /// worker joined, every redb handle closed, no signal, exit status 0. That is
    /// what makes the parent's reopen a *clean* reopen rather than a recovery,
    /// and it is the case the crash guard cannot reach: after a clean shutdown
    /// every pack's bit is legitimately set and **nothing is re-queued**, so the
    /// derived tables have to have been on disk or they are gone.
    ///
    /// `#[ignore]`d because it is a fixture, not a guard — run on its own it has
    /// no directory to work in and returns.
    #[test]
    #[ignore = "spawned by a_clean_shutdown_and_reopen_keeps_the_graph_and_reachable"]
    fn the_child_that_pushes_and_shuts_down_cleanly() {
        let Ok(dir) = std::env::var(CLEAN_DIR) else {
            return;
        };
        let dir = PathBuf::from(dir);
        let pack = std::fs::read(dir.join("fixture.pack")).expect("the parent's fixture");

        let store = GitStore::open(&dir, "rickard").expect("open");
        store.put_pack(&pack).expect("the ack path");
        store.wait_indexed();
        assert_eq!(store.unindexed_packs(), 0, "the child shut down mid-index");

        let tip = store
            .graph_snapshot()
            .into_iter()
            .max_by_key(|c| c.generation)
            .expect("the child built no graph at all");
        let tip_raw = hex::decode(&tip.oid).unwrap();
        let closure = store.reachable(&[&tip_raw], &[]).unwrap();
        assert!(!closure.is_empty(), "the child's own reachable() was empty");

        std::fs::write(dir.join("tip"), &tip.oid).expect("marker");
        std::fs::write(dir.join("closure"), closure.len().to_string()).expect("marker");
        std::fs::write(dir.join("commits"), store.commit_count().to_string()).expect("marker");
        std::fs::write(dir.join("objects"), store.object_count().to_string()).expect("marker");

        // **The clean shutdown.** Not a signal and not an abort: the store drops,
        // which closes the channel, joins the account indexer's worker and closes
        // redb. The process then returns normally with status 0.
        drop(store);
        std::fs::write(dir.join("clean"), b"closed").expect("marker");
    }

    /// **`commit_count()` and `reachable()` are correct after a CLEAN shutdown
    /// and reopen.** This is the guard the exploded table exists to make pass.
    ///
    /// It was the live bug: the commit graph was folded from commit and tree
    /// *payloads* which were stored nowhere, accumulated in RAM as packs were
    /// absorbed. A store killed mid-absorb got them back only because the
    /// interrupted pack was re-queued and re-resolved
    /// ([`a_store_reopened_after_a_kill_mid_absorb_requeues_the_pack`]); a store
    /// that shut down **cleanly** re-queued nothing, and MEASURED on oden
    /// 2026-08-08 came back with 2687 rows and **0 of 551 commits** —
    /// `reachable()` returning a commit alone instead of its closure, and `gc`
    /// seeing an empty live set. Both are quiet wrong answers.
    ///
    /// The shape is the crash guard's: a **real** child process, which here exits
    /// normally, and the parent asserts against what is on disk. Two assertions
    /// carry it and they have to be in this order:
    ///
    /// 1. **`commit_count()` immediately after `open`**, before anything is
    ///    waited on — so the graph can only have come from the fold over the
    ///    exploded table that `open` performs, not from a re-absorb;
    /// 2. **`is_indexed(0)` is false after `wait_indexed()`** — the account
    ///    indexer in *this* process publishes a pack only for a job somebody
    ///    submitted, so a clear bit proves open re-queued nothing and the graph
    ///    above is not a recovery artefact.
    ///
    /// `reachable()` is then asserted against the closure the child computed on
    /// the same tip, so it is the same answer and not merely a non-empty one.
    ///
    /// Seen RED by deleting `store.refold()?` from the end of
    /// `GitStore::open_with_arms`, so nothing folds the table on open: "the
    /// commit graph did not survive a clean shutdown: 0 of 551 commits — the
    /// exploded table is a warm cache, not a durable table", left 0 right 551.
    /// That is the recorded bug's own number, reproduced. Restored.
    ///
    /// Seen RED a second time by deleting `self.exploded.flush()?` from
    /// `Absorber::absorb_one`, which is the **durability** edit: the rows then
    /// live only in the un-flushed buffer, which answers perfectly inside the
    /// child and is gone when the process ends. It fires assertion (2) rather
    /// than (1) — "the reopen re-queued the pack, so the graph above is a
    /// recovery artefact rather than the table's" — and that is worth writing
    /// down, because assertion (1) *passed*: the 64 MiB flush threshold had
    /// tripped once mid-pack, leaving 1798 of 2687 rows on disk which happened
    /// to include all 551 commits. A guard with only assertion (1) would have
    /// called that edit green. Restored.
    /// **An annotated tag reaches the commit it points at.**
    ///
    /// # The bug this is the guard for
    ///
    /// Reachability bitmaps are built over the commit graph, and no tree points
    /// at a tag, so a tag is in nobody's closure. `reachable`'s `want` loop had
    /// one arm for "not a commit": insert the object itself, and stop. So a
    /// `want` that was an annotated tag answered with **one object** — the tag —
    /// and the commit it names, plus that commit's whole history, was left out.
    ///
    /// It stayed invisible because a tag is nearly always pushed alongside the
    /// branch that carries the same commit, and then the branch's closure
    /// covers it. It surfaces the moment a tag is the ONLY thing naming a
    /// commit — which is what gunnar's replication does, one ref at a time. Six
    /// repositories on the live fleet could not replicate on 2026-08-31, one of
    /// them with 25 tags hanging off a single commit no branch carried, all
    /// refused by the far end as `missing necessary objects … (peeling a tag)`.
    ///
    /// So the fixture asserts the premise it needs: the tag is the only `want`.
    ///
    /// # 🔴 Seen red
    ///
    /// Against the arm as it was (`union.insert(o)` and nothing more):
    /// `the tag's closure is 1 object, so it carries the tag and nothing it
    /// points at`. With the peel: the commit, its trees and its blobs.
    #[test]
    fn an_annotated_tag_reaches_the_commit_it_points_at() {
        let dir = tmpdir("tag-peel");
        let store = GitStore::open(&dir, "rickard").unwrap();
        let (pack, rows) = corpus_pack();
        store.put(&pack, &[]).unwrap();
        store.wait_indexed();

        // Any commit will do: the tag is the only `want`, so nothing else can
        // pull this commit's closure in behind our backs.
        let commit = rows
            .iter()
            .find(|r| r.kind == GitObjectKind::Commit)
            .expect("the corpus has commits");
        let commit_hex = hex::encode(&commit.oid);
        let commit_closure = store.reachable(&[&commit.oid], &[]).unwrap();
        assert!(
            commit_closure.len() > 1,
            "the fixture's commit closes over nothing, so this test could not tell a peel from a              no-op"
        );

        let (tag_pack, tag_oid) = crate::store::tests::one_tag_pack(&commit_hex, "v1.0");
        store.put(&tag_pack, &[]).unwrap();
        store.wait_indexed();

        let closure = store.reachable(&[&tag_oid], &[]).unwrap();
        assert!(
            closure.contains(&tag_oid),
            "the tag itself is not in its own closure"
        );
        assert!(
            closure.contains(&commit.oid),
            "the tag's closure is {} object(s), so it carries the tag and nothing it points at —              the commit {} it names is missing and a mirror built from this answer cannot be              read",
            closure.len(),
            &commit_hex[..12]
        );
        // Not merely the commit: everything the commit reaches, or a clone from
        // this answer is short by a tree.
        for oid in &commit_closure {
            assert!(
                closure.contains(oid),
                "the tag's closure is missing {} which the commit it points at reaches",
                hex::encode(oid)
            );
        }
    }

    #[test]
    fn a_clean_shutdown_and_reopen_keeps_the_graph_and_reachable() {
        let dir = tmpdir("clean-shutdown-graph");
        let (pack, rows) = corpus_pack();
        std::fs::write(dir.join("fixture.pack"), &pack).unwrap();

        let status = std::process::Command::new(std::env::current_exe().unwrap())
            .args([
                "--exact",
                "--ignored",
                "--nocapture",
                "git_ops::tests::the_child_that_pushes_and_shuts_down_cleanly",
            ])
            .env(CLEAN_DIR, &dir)
            .status()
            .expect("spawn the child that shuts down cleanly");
        assert!(
            status.success(),
            "the child did not exit cleanly — it {status:?}, so this is a crash guard and not a \
             clean-shutdown one"
        );
        assert!(
            dir.join("clean").exists(),
            "the child never reached its clean shutdown"
        );

        let tip_hex = std::fs::read_to_string(dir.join("tip")).unwrap();
        let closure_before: usize = std::fs::read_to_string(dir.join("closure"))
            .unwrap()
            .parse()
            .unwrap();
        let commits_before: usize = std::fs::read_to_string(dir.join("commits"))
            .unwrap()
            .parse()
            .unwrap();
        assert_eq!(
            commits_before,
            commits_in(&rows),
            "the child's own graph was already wrong, so the reopen proves nothing"
        );

        // ── the reopen, and (1): the graph before anything is waited on ──────
        let store = GitStore::open(&dir, "rickard").unwrap();
        assert_eq!(
            store.commit_count(),
            commits_in(&rows),
            "the commit graph did not survive a clean shutdown: {} of {} commits — the exploded \
             table is a warm cache, not a durable table",
            store.commit_count(),
            commits_in(&rows)
        );
        let tip_raw = hex::decode(&tip_hex).unwrap();
        let closure = store.reachable(&[&tip_raw], &[]).unwrap();
        assert_eq!(
            closure.len(),
            closure_before,
            "reachable() answers differently after a clean reopen: {} objects against the {} the \
             same tip closed over before the shutdown",
            closure.len(),
            closure_before
        );
        assert!(closure.contains(&tip_raw));

        // ── (2): and none of it came from a re-absorb ────────────────────────
        store.wait_indexed();
        assert!(
            !store.indexer().is_indexed(0),
            "the reopen re-queued the pack, so the graph above is a recovery artefact rather \
             than the table's"
        );
        assert_eq!(store.unindexed_packs(), 0);
        assert_eq!(
            store.absorb_pending().unwrap(),
            0,
            "the reopen left index work"
        );
        assert_eq!(store.object_count(), rows.len());
        assert_eq!(
            store.exploded_stats().written,
            0,
            "this process wrote exploded rows, so the table was rebuilt rather than read"
        );

        eprintln!(
            "load {}; clean shutdown → reopen: {} rows, {} commits, tip closure {} objects \
             (rebuilt nothing)",
            loadavg(),
            store.object_count(),
            store.commit_count(),
            closure.len(),
        );
    }

    /// **A reopen READS `objects.graph`; it does not recompute it — and the
    /// read graph equals the recomputed one.** This is S-065's fix, and the only
    /// thing that can say which path a fold took is a counter, so the assertion
    /// is on [`GitStore::graph_source_stats`], not on the (identical) graph
    /// bytes — the identity-value trap, since both paths yield the same graph.
    ///
    /// One test exercises **both** arms so the counter cannot be blindly stuck:
    ///
    /// 1. A first store builds the graph, persists it, and shuts down cleanly.
    /// 2. `objects.graph` is deleted, so the reopen has no cache: it **recomputes**
    ///    (`(read, recomputed) == (0, 1)`) and re-persists. Its graph and its
    ///    `reachable()` are captured — this is the recomputed truth.
    /// 3. A second reopen finds the cache the step above wrote: it **reads**
    ///    (`(1, 0)`), and its graph and `reachable()` must equal step 2's exactly
    ///    — read == recomputed.
    ///
    /// # 🔴 Seen red
    ///
    /// With the persistence removed (open always folds), step 3 is `(0, 1)`:
    /// "the reopen recomputed the graph instead of reading objects.graph", left
    /// `(0, 1)` right `(1, 0)`. With the token check inverted so a stale cache is
    /// accepted, step 3's `reachable()` diverges from step 2's and the read==
    /// recomputed assertion fires. Restored.
    #[test]
    fn a_reopen_reads_the_persisted_graph_it_does_not_recompute() {
        let dir = tmpdir("persisted-graph");
        let (pack, rows) = corpus_pack();
        let tip_hex;
        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
            tip_hex = store
                .graph_snapshot()
                .into_iter()
                .max_by_key(|c| c.generation)
                .expect("a graph")
                .oid;
            assert!(
                dir.join("objects.graph").exists(),
                "an absorb never persisted the graph, so a reopen has nothing to read"
            );
        }
        let tip_raw = hex::decode(&tip_hex).unwrap();

        // ── (2) no cache → recompute, and capture the recomputed truth ───────
        std::fs::remove_file(dir.join("objects.graph")).unwrap();
        let (recomputed_graph, recomputed_closure) = {
            let store = GitStore::open(&dir, "rickard").unwrap();
            assert_eq!(
                store.graph_source_stats(),
                (0, 1),
                "with no objects.graph the reopen should have recomputed the graph"
            );
            assert_eq!(store.commit_count(), commits_in(&rows));
            let g = store.graph_snapshot();
            let c = store.reachable(&[&tip_raw], &[]).unwrap();
            assert!(!c.is_empty());
            (g, c)
        };

        // ── (3) cache present → read, and it must equal the recomputed truth ──
        let store = GitStore::open(&dir, "rickard").unwrap();
        assert_eq!(
            store.graph_source_stats(),
            (1, 0),
            "the reopen recomputed the graph instead of reading objects.graph"
        );
        assert_eq!(
            store.graph_snapshot(),
            recomputed_graph,
            "the graph READ from objects.graph differs from the one refold RECOMPUTES"
        );
        assert_eq!(
            store.reachable(&[&tip_raw], &[]).unwrap(),
            recomputed_closure,
            "reachable() over the persisted graph differs from reachable() over the recomputed one"
        );
        assert_eq!(store.commit_count(), commits_in(&rows));
    }

    /// **The graph cache switched off leaves no `objects.graph`** — and the store
    /// still answers the same graph and closure, folded from the exploded table.
    /// The switch is flipped on the absorber directly: `ZNIPPY_GIT_GRAPH_CACHE`
    /// is read at open, and setting it process-wide would race the test above.
    ///
    /// Seen red with the persist guard removed: "the graph cache is off and
    /// objects.graph was written anyway".
    #[test]
    fn with_the_graph_cache_off_no_objects_graph_is_written_and_the_graph_is_the_same() {
        let on_dir = tmpdir("graph-cache-on");
        let off_dir = tmpdir("graph-cache-off");
        let (pack, rows) = corpus_pack();
        let on = GitStore::open(&on_dir, "rickard").unwrap();
        on.put_pack(&pack).unwrap();
        on.wait_indexed();
        let off = GitStore::open(&off_dir, "rickard").unwrap();
        off.absorber.graph_cache.store(false, std::sync::atomic::Ordering::Relaxed);
        std::fs::remove_file(off_dir.join("objects.graph")).ok();
        off.put_pack(&pack).unwrap();
        off.wait_indexed();
        assert!(on_dir.join("objects.graph").exists(), "the default still persists the graph");
        assert!(
            !off_dir.join("objects.graph").exists(),
            "the graph cache is off and objects.graph was written anyway"
        );
        assert_eq!(off.graph_source_stats().0, 0, "a store with the cache off never reads one");
        assert_eq!(on.graph_snapshot(), off.graph_snapshot());
        assert_eq!(off.commit_count(), commits_in(&rows));
    }

    /// **The token gates the read: a cache whose token does not match the
    /// object set is rejected and the graph is recomputed.** This is what makes
    /// `objects.graph` a memo rather than a second truth — and a push, a gc, or
    /// a torn write all reach the store through exactly this door.
    ///
    /// The object set is left unchanged and the token in the cache is corrupted
    /// directly, so the ONE thing under test is the token comparison and not
    /// some incidental object-count difference.
    ///
    /// # 🔴 Seen red
    ///
    /// With the token comparison dropped (`Some((_, g)) => g`), the corrupted
    /// cache is read and the reopen is `(1, 0)`: "a cache with the wrong token
    /// was trusted". Restored.
    #[test]
    fn a_graph_whose_token_does_not_match_is_rejected() {
        let dir = tmpdir("wrong-token-graph");
        let (pack, _rows) = corpus_pack();
        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
        }
        // Flip a byte inside the 32-byte token (offset 5..37) without touching
        // the graph body, so only the token is wrong.
        let path = dir.join("objects.graph");
        let mut bytes = std::fs::read(&path).unwrap();
        assert!(bytes.len() > 37 && &bytes[..4] == b"GRPH");
        bytes[6] ^= 0xff;
        std::fs::write(&path, &bytes).unwrap();

        let store = GitStore::open(&dir, "rickard").unwrap();
        assert_eq!(
            store.graph_source_stats(),
            (0, 1),
            "a cache with the wrong token was trusted instead of being recomputed"
        );
    }

    /// **The ordinal space comes off the persisted index, not off a scan — and
    /// no `objects.ordinal` is ever written.** This replaces S-082's memo test:
    /// the memo spared a cold open the redb scan that rebuilt the sorted oid
    /// list, and since 2026-09-04 that list IS the persisted index's oid column,
    /// so there is nothing to spare and no file to validate. What is asserted:
    ///
    /// 1. A first store absorbs the corpus; its ordinal space equals the tail in
    ///    oid order and every oid is accounted to one of the two sources. Then
    ///    it folds explicitly, so redb is empty and the index holds everything.
    /// 2. The reopen is warm: its space comes **entirely off the index** —
    ///    `(folded, tail) == (n, 0)` — and equals step 1's space, closure and
    ///    counts exactly. `objects.ordinal` does not exist after either session.
    ///
    /// # 🔴 Seen red
    ///
    /// With `ObjectReadStack::oids_raw` returning only the redb tail on a
    /// persisting stack, step 2's space is empty (redb holds nothing after the
    /// fold) and `reachable()` over it is empty: "the ordinal space after the
    /// warm reopen is not the tail in oid order". Restored.
    #[test]
    fn the_ordinal_space_comes_off_the_index_not_a_scan() {
        let dir = tmpdir("ordinal-off-the-index");
        let (pack, rows) = corpus_pack();
        let tip_hex;
        let (first_oids, first_closure, first_objects, first_commits);
        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
            tip_hex = store
                .graph_snapshot()
                .into_iter()
                .max_by_key(|c| c.generation)
                .expect("a graph")
                .oid;
            let tip_raw = hex::decode(&tip_hex).unwrap();
            let n = store.object_count() as u64;
            let (folded, tail) = store.ordinal_source_stats();
            assert_eq!(folded + tail, n, "every oid is accounted to the index or the tail");
            let oids = store.ordinal_oids();
            let scanned: Vec<String> =
                store.index().oids_in_order().unwrap().iter().map(hex::encode).collect();
            assert_eq!(oids, scanned, "the ordinal space is not the repository in oid order");
            assert!(
                !dir.join("objects.ordinal").exists(),
                "no memo may be written: the index's oid column is the ordinal space"
            );
            // Fold, so the reopen has everything on the index and nothing in redb.
            store.index().rebuild().unwrap();
            first_closure = store.reachable(&[&tip_raw], &[]).unwrap();
            assert!(!first_closure.is_empty());
            first_oids = oids;
            first_objects = store.object_count();
            first_commits = store.commit_count();
        }
        let tip_raw = hex::decode(&tip_hex).unwrap();

        let store = GitStore::open(&dir, "rickard").unwrap();
        let n = store.object_count() as u64;
        assert_eq!(
            store.ordinal_source_stats(),
            (n, 0),
            "after a fold the warm reopen's ordinal space must come entirely off the index"
        );
        assert_eq!(
            store.ordinal_oids(),
            first_oids,
            "the ordinal space after the warm reopen is not the tail in oid order"
        );
        assert_eq!(
            store.reachable(&[&tip_raw], &[]).unwrap(),
            first_closure,
            "reachable() over the index's ordinal space differs from the first session's"
        );
        assert_eq!(store.object_count(), first_objects);
        assert_eq!(store.commit_count(), first_commits);
        assert_eq!(store.commit_count(), commits_in(&rows));
        assert!(!dir.join("objects.ordinal").exists(), "still no memo after the reopen");
    }

    /// **A stale `objects.ordinal` left by an older binary is removed at open.**
    /// The file is dead weight now (279 MB on the kernel mirror) and its
    /// contents are never read, so a store that carries one loses it the first
    /// time this build opens it — and keeps working.
    ///
    /// # 🔴 Seen red
    ///
    /// With the `remove_file` in `open_with_arms` dropped, the file survives the
    /// open: "a stale objects.ordinal must be removed at open". Restored.
    #[test]
    fn a_stale_objects_ordinal_from_an_older_binary_is_removed_at_open() {
        let dir = tmpdir("stale-ordinal");
        let (pack, _rows) = corpus_pack();
        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
        }
        // The old memo's shape: magic, version, oid_len, epoch, count, oids.
        let path = dir.join("objects.ordinal");
        let mut bytes = b"ORDL".to_vec();
        bytes.push(1);
        bytes.push(20);
        bytes.extend_from_slice(&7u64.to_le_bytes());
        bytes.extend_from_slice(&1u64.to_le_bytes());
        bytes.extend_from_slice(&[0xabu8; 20]);
        std::fs::write(&path, &bytes).unwrap();
        assert!(path.exists(), "precondition");

        let store = GitStore::open(&dir, "rickard").unwrap();
        assert!(!path.exists(), "a stale objects.ordinal must be removed at open");
        assert!(store.object_count() > 0, "and the store still answers");
    }

    /// **A content read is served from the table, not re-derived** — and the only
    /// thing that can say so is a counter.
    ///
    /// Both paths return the **identical** bytes: the table hands back what the
    /// resolver produced, and the fallback re-resolves the same pack and produces
    /// it again. A byte comparison therefore proves nothing at all about which
    /// one ran — this is the identity-value trap in its exact form, and it fired
    /// on this codebase once already. So the assertion is on
    /// [`ExplodedStats::served`] against [`ExplodedStats::rederived`], both of
    /// which are bumped where the work happens.
    ///
    /// The fallback is not asserted away: it is exercised in the same test, on a
    /// store whose table has been dropped, and the counters move the other way.
    /// A guard that only saw the fast path could not tell a working counter from
    /// one that is never incremented.
    ///
    /// Seen RED by deleting the `if let Some(hit) = self.exploded.content(oid)?`
    /// early return in `Absorber::resolved`, so every read re-derives: "512 content
    /// reads re-resolved a whole pack instead of hitting the table — served 0,
    /// rederived 512". The bytes were still correct on every one of them, which
    /// is the point. Restored.
    ///
    /// Seen RED a second time by bumping `served` unconditionally at the top of
    /// `ExplodedTable::content`, before the `pending`/redb probe — the counter
    /// then reports a hit whether or not the table had the object. The first half
    /// fires on the double count, "512 content reads re-resolved a whole pack
    /// instead of hitting the table — served 1024, rederived 0", and
    /// `exploded::tests::a_row_round_trips_and_the_kind_index_finds_it` fires with
    /// it ("a content read was not counted", left 5 right 2). Restored — a
    /// counter that cannot be wrong about a miss is the only kind worth
    /// asserting on.
    #[test]
    fn a_content_read_is_served_from_the_table_and_not_re_derived() {
        const READS: usize = 512;
        let dir = tmpdir("exploded-served");
        let (pack, rows) = corpus_pack();
        let oids: Vec<Vec<u8>> = rows
            .iter()
            .map(|r| r.oid.clone())
            .collect::<HashSet<Vec<u8>>>()
            .into_iter()
            .take(READS)
            .collect();

        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();

            // ── FIRST read: the lazy path, timed ────────────────────────
            let before = store.exploded_stats();
            let t = Instant::now();
            let mut first: Vec<(GitObjectKind, Vec<u8>)> = Vec::with_capacity(oids.len());
            for oid in &oids {
                first.push(
                    store
                        .content(oid)
                        .unwrap()
                        .expect("the store lost an object"),
                );
            }
            let cold = t.elapsed();
            let after = store.exploded_stats();
            let rederived = after.rederived - before.rederived;
            assert_eq!(
                rederived + (after.served - before.served),
                oids.len() as u64,
                "{} reads = served {} + rederived {} must account for every one",
                oids.len(),
                after.served - before.served,
                rederived
            );
            assert!(
                rederived > 0,
                "not one read re-derived: either every object is eager or the lazy \
                 path never ran, and this test then proves nothing about either"
            );

            // The write-back is off the read path, so it has to be waited for
            // HERE and only here — a reader never does.
            store.explode_cache_maintain().unwrap();
            // **How many of those reads the cache DECLINED.** A shallow, small
            // object is cheaper to inflate again than to keep, so it is read
            // directly and nothing is written
            // (`exploded_cache::worth_caching`). Those objects are re-derived
            // on every read for ever, on purpose, and the second pass below
            // has to expect exactly that many — not zero, which is what this
            // test asserted before the admission policy existed and would now
            // be asserting that the policy is off.
            let declined = store.explode_cache_use().direct_reads;
            assert!(
                declined > 0,
                "not one of {} reads was declined as cheap; either every object in \
                 this corpus is expensive or the admission policy is not running",
                oids.len()
            );

            // ── SECOND read: the design note's "= today" column ──────────
            let mid = store.exploded_stats();
            let t = Instant::now();
            for (i, oid) in oids.iter().enumerate() {
                let again = store.content(oid).unwrap().expect("the store lost an object");
                assert_eq!(again, first[i], "the second read of an object disagreed with the first");
            }
            let warm = t.elapsed();
            let end = store.exploded_stats();
            assert_eq!(
                end.rederived - mid.rederived,
                declined,
                "the second pass re-derived {} objects and the cache declined {declined} — \
                 every EXPENSIVE object must come out of the table the second time, and \
                 every cheap one must be read directly again",
                end.rederived - mid.rederived
            );
            assert_eq!(
                end.served - mid.served,
                oids.len() as u64 - declined,
                "the expensive objects did not come out of the table on the second pass"
            );
            eprintln!(
                "load {}; {} content reads: FIRST {:.1} ms ({:.1} µs/read, {} re-derived from \
                 the pack, {declined} of them too cheap to cache) → SECOND {:.1} ms \
                 ({:.1} µs/read) = {:.1}x",
                loadavg(),
                oids.len(),
                cold.as_secs_f64() * 1e3,
                cold.as_secs_f64() * 1e6 / oids.len() as f64,
                rederived,
                warm.as_secs_f64() * 1e3,
                warm.as_secs_f64() * 1e6 / oids.len() as f64,
                cold.as_secs_f64() / warm.as_secs_f64().max(f64::MIN_POSITIVE),
            );
        }

        // ── and the fallback, so the counter is not a constant ───────────────
        //
        // The table is dropped and the *drain* is what would rebuild it, so the
        // reads have to happen before it gets there. `absorb_pending` is called
        // by hand first so the fallback under test is the content path's and not
        // the index's.
        std::fs::remove_file(dir.join("objects.exploded")).unwrap();
        // **And the lazy pool with it.** Dropping only the eager table leaves
        // the write-backs of the pass above sitting in `objects.exploded.cN`,
        // and those would answer — the fallback would then be measured on a
        // store that still has a cache, which is the exact shape of a test that
        // cannot fail. One `drop_explode_cache` is the sanctioned door, and it
        // is what `repo gc` calls.
        for e in std::fs::read_dir(&dir).unwrap().flatten() {
            if e.file_name().to_string_lossy().starts_with("objects.exploded.c") {
                std::fs::remove_file(e.path()).unwrap();
            }
        }
        let store = GitStore::open(&dir, "rickard").unwrap();
        let gate = store.hold_absorb_gate();
        let before = store.exploded_stats();
        let mut rederived = 0u64;
        for oid in oids.iter().take(8) {
            // Straight at the absorber: the store's `content` would absorb the
            // re-queued pack first and rebuild the very table this is testing
            // the absence of.
            assert!(
                store.absorber.resolved(oid).unwrap().is_some(),
                "the verbatim truth could not re-derive {}",
                hex::encode(oid)
            );
            rederived += 1;
        }
        let after = store.exploded_stats();
        // The drain starts the instant the store opens; on a loaded box it can
        // rebuild the small corpus's table before the gate above is taken, and
        // then the reads are legitimately served from it. Seen 2026-08-30 under
        // a parallel suite. That run proves nothing about the fallback and says
        // so, rather than failing on a race the test does not control.
        let drain_won = store.unindexed_packs() == 0;
        drop(gate);
        if drain_won {
            eprintln!("the drain rebuilt the table before the gate was held; the fallback was not observed this run");
            return;
        }
        assert_eq!(
            after.served - before.served,
            0,
            "a dropped table still reported {} reads served",
            after.served - before.served
        );
        assert_eq!(
            after.rederived - before.rederived,
            rederived,
            "the fallback ran but was not counted"
        );
    }

    /// ★ **THE READ-AFTER-ACK GAP (lane U → lane Q): every read path answers
    /// with the indexer PAUSED, and none of them waits for it.**
    ///
    /// MEASURED on the rig 2026-08-30 (lane U, both ISOs, timing-dependent):
    /// right after an ACKED push a `clone` came back
    /// `remote error: object … is not in this repository (peeling a tag)` and a
    /// `push --mirror` on top got `[remote rejected] … missing necessary
    /// objects`. Both of those are gunnar's `Graph::load` — `GitServe::read` —
    /// answering `None` for an object whose bytes were fsynced before the client
    /// was told the push had landed. §13.12 says an index miss falls back to the
    /// journal; the tag-peel and connectivity-check paths were not observed to.
    ///
    /// So this holds the absorb gate — the indexer paused, exactly the state
    /// lane P's watchdog puts it in — and requires **`has`, `header` and `read`
    /// to answer for every object of a just-acked pack**, byte-correct, and to
    /// do it inside a deadline. The deadline is not decoration: "nobody waits"
    /// (the design note's rule) is violated by a read that blocks on the gate
    /// just as badly as by one that answers absent, and a blocking read looks
    /// exactly like a correct one to an assertion that has no clock.
    ///
    /// Seen RED by returning `Ok(None)` from `Absorber::resolved`'s pending arm:
    /// 1352 of 1352 objects report "the store said absent for an object it had
    /// already acked".
    #[test]
    fn every_read_path_answers_a_just_acked_push_with_the_indexer_paused() {
        use crate::serve::GitServe as _;
        let dir = tmpdir("read-after-ack");
        let store = Arc::new(GitStore::open(&dir, "rickard").unwrap());
        let (pack, rows) = corpus_pack();

        // The ack: `put_pack` returning is what the client is told.
        store.put_pack(&pack).unwrap();

        // The indexer, paused. A watchdog pause (lane P) pauses the indexer and
        // nothing else, so this is the shape the appliance is really in.
        let gate = store.hold_absorb_gate();
        assert!(
            store.unindexed_packs() > 0,
            "the drain absorbed the pack before the gate was taken, so this run \
             read an INDEXED store and proves nothing about the window"
        );

        let distinct: Vec<Vec<u8>> = rows
            .iter()
            .map(|r| r.oid.clone())
            .collect::<HashSet<Vec<u8>>>()
            .into_iter()
            .collect();
        let hash = store.hash_kind();
        let n = distinct.len();

        // On another thread, so the deadline measures the reads and not this
        // test's own patience.
        let (tx, rx) = std::sync::mpsc::channel::<Result<(usize, usize, usize)>>();
        let reader = {
            let store = store.clone();
            let distinct = distinct.clone();
            std::thread::spawn(move || {
                let mut had = 0usize;
                let mut headed = 0usize;
                let mut readback = 0usize;
                let r = (|| -> Result<()> {
                    for oid in &distinct {
                        if store.has(oid)? {
                            had += 1;
                        }
                        if crate::serve::GitServe::header(&*store, oid)?.is_some() {
                            headed += 1;
                        }
                        // The tag-peel / connectivity-check path, byte for byte.
                        let Some((kind, payload)) = crate::serve::GitServe::read(&*store, oid)? else {
                            continue;
                        };
                        let k = match kind {
                            crate::index_layout::ObjType::Commit => GitObjectKind::Commit,
                            crate::index_layout::ObjType::Tree => GitObjectKind::Tree,
                            crate::index_layout::ObjType::Blob => GitObjectKind::Blob,
                            crate::index_layout::ObjType::Tag => GitObjectKind::Tag,
                            // A read resolves the chain, so a delta type coming
                            // back out of it is the store contradicting itself.
                            other => anyhow::bail!("`read` answered with a {other:?}"),
                        };
                        let rehashed = hash.oid_of(&crate::object::canonical(k, &payload));
                        if &rehashed == oid {
                            readback += 1;
                        }
                    }
                    Ok(())
                })();
                let _ = tx.send(r.map(|_| (had, headed, readback)));
            })
        };

        // The deadline. Generous — a cold pending scan of the corpus is
        // milliseconds — and it is against a WALL clock while the gate is held,
        // so a read that parks on the absorber trips it.
        let out = rx
            .recv_timeout(Duration::from_secs(20))
            .expect(
                "a read path blocked on the paused indexer for 20 s: 'nobody waits' is the \
                 design note's rule and a read that waits for the drain is a bug by definition",
            );
        let (had, headed, readback) = out.expect("a read path errored inside the index window");
        drop(gate);
        let _ = reader.join();

        assert_eq!(
            (had, headed, readback),
            (n, n, n),
            "the store answered for {had}/{n} `has`, {headed}/{n} `header` and {readback}/{n} \
             byte-correct `read`s of a pack it had already ACKED — an object the client was \
             told had landed must never read as absent"
        );
    }

    /// ★ **An account at its explode budget serves EVERY object, correctly —
    /// only slower.**
    ///
    /// The quota is a cache size, not a storage limit: over budget the server
    /// EVICTS, it does not refuse. So a repository squeezed to a budget far
    /// under what it holds must still answer every read byte-for-byte, and the
    /// only thing that may change is where the bytes came from — the table, or
    /// the pack via [`crate::resolve::materialize_at`].
    ///
    /// This is requirement (c) of the quota: *"an account at its budget must
    /// still serve every read correctly, only slower"*. The `rederived`
    /// counter is what proves the eviction really happened, because identical
    /// bytes come back either way and a byte comparison cannot tell a working
    /// cache from a dead one — the identity-value trap in its exact form.
    ///
    /// Seen RED by having `write_back` return the payload's bytes from a
    /// half-written segment (`out.truncate(1)`): "the store returned the wrong
    /// bytes for an object under a squeezed budget".
    #[test]
    fn an_account_at_its_budget_serves_every_object_only_slower() {
        let dir = tmpdir("explode-budget");
        let store = GitStore::open(&dir, "rickard").unwrap();
        let (pack, rows) = corpus_pack();
        store.put_pack(&pack).unwrap();
        store.wait_indexed();

        let oids: Vec<Vec<u8>> = rows
            .iter()
            .map(|r| r.oid.clone())
            .collect::<HashSet<Vec<u8>>>()
            .into_iter()
            .collect();

        // Every object once, unsqueezed: the truth to compare against.
        let mut truth: Vec<(GitObjectKind, Vec<u8>)> = Vec::with_capacity(oids.len());
        for oid in &oids {
            truth.push(store.content(oid).unwrap().expect("the store lost an object"));
        }
        store.explode_cache_maintain().unwrap();

        // Now the plan changes: a budget far under what this repository holds.
        store.set_explode_budget(64 << 10);
        let before = store.exploded_stats();
        for _ in 0..3 {
            for (i, oid) in oids.iter().enumerate() {
                let got = store
                    .content(oid)
                    .unwrap()
                    .expect("an object went absent when the budget was lowered");
                assert_eq!(
                    got, truth[i],
                    "the store returned the wrong bytes for an object under a squeezed budget"
                );
            }
            store.explode_cache_maintain().unwrap();
        }
        let after = store.exploded_stats();
        let band = store.explode_cache_use();

        assert!(
            band.evicted_bytes > 0 || band.lazy_bytes <= 64 << 10,
            "the budget was lowered to 64 KiB and the lazy pool holds {} B with nothing \\
             evicted — the quota is not enforced, so this run proves nothing about reading \\
             under one: {band:?}",
            band.lazy_bytes
        );
        assert!(
            after.rederived > before.rederived,
            "not one read re-derived from the pack under a 64 KiB budget, so the cache was \\
             never squeezed and 'only slower' was never tested"
        );
        eprintln!(
            "load {}; {} objects read 3x under a 64 KiB budget: {} re-derived, {} served, \
{} evicted bytes over {} segments; every byte identical",
            loadavg(),
            oids.len(),
            after.rederived - before.rederived,
            after.served - before.served,
            band.evicted_bytes,
            band.evicted_segments,
        );
    }

    /// ★ **A THIN push on top of a just-acked one is accepted with the indexer
    /// PAUSED** — lane U's second rig symptom.
    ///
    /// MEASURED on the rig 2026-08-30 (lane U): right after an acked push,
    /// `git push --mirror` of the next pack came back
    /// `[remote rejected] … missing necessary objects`, on a varying set of
    /// repositories run to run — a race, not a repository. A thin pack names
    /// its delta bases by oid and expects the server to have them; the server
    /// **did** have them, in a pack it had acked and not yet indexed, and
    /// `scan_pending`'s re-entrancy guard answered `None` for every base looked
    /// up from inside a scan rather than only for the ones that would have
    /// started a second scan.
    ///
    /// Seen RED by restoring that guard (`if nested { return Ok((out, None)) }`):
    /// *"reading the thin pack's object errored inside the index window:
    /// resolving a pending pack read-only for lookups — this pack is thin: it
    /// deltas against 42088154… which is not in it, and no base source could
    /// produce that object's content"*. The **push** is accepted either way —
    /// `external_bases_exist` finds the base with a top-level `lookup_one`
    /// before it ever resolves — so the assertion that discriminates is the
    /// READ with the gate still held, and an earlier draft of this test that
    /// dropped the gate first passed under both arms.
    #[test]
    fn a_thin_push_on_top_of_an_unindexed_one_is_accepted() {
        let dir = tmpdir("thin-on-unindexed");
        let store = GitStore::open(&dir, "rickard").unwrap();
        let (pack, rows) = corpus_pack();
        store.put_pack(&pack).unwrap();

        // The indexer, paused — so the base can only be found in the pack that
        // was acked a moment ago.
        let gate = store.hold_absorb_gate();
        assert!(
            store.unindexed_packs() > 0,
            "the drain absorbed the first pack before the gate was taken; this run \
             pushed onto an INDEXED store and proves nothing"
        );

        // A base with a payload, and its real size: a `REF_DELTA` declares the
        // source size and the resolver checks it, so a made-up number would fail
        // for a reason that has nothing to do with the window under test.
        let base = rows
            .iter()
            .find(|r| r.kind == GitObjectKind::Blob && r.uncompressed_size > 0)
            .expect("the corpus has a blob");
        let thin = crate::pack_fixture::thin_ref_delta_pack(
            &base.oid,
            base.uncompressed_size,
            b"lane Q was here\n",
        );

        let tx = store.put_pack(&thin).unwrap_or_else(|e| {
            panic!("the thin push was refused: {e:#}");
        });
        assert!(tx.pack_id.is_some(), "the thin pack was not stored");

        // **And its object reads back with the gate STILL HELD.** This is the
        // nested case and the one the guard was too wide for: the read scans
        // pack B, B's entry is a ref-delta, its base is in pack A, and
        // resolving it is a second `scan_pending` from inside the first. A
        // lookup into A's already-built scan terminates; only a *build* would
        // not.
        let hash = store.hash_kind();
        let oid = hash.oid_of(&crate::object::canonical(
            GitObjectKind::Blob,
            b"lane Q was here\n",
        ));
        let got = store.content(&oid);
        drop(gate);
        let (kind, payload) = got
            .expect("reading the thin pack's object errored inside the index window")
            .expect("the thin pack's own object is absent after its push was acked");
        assert_eq!(kind, GitObjectKind::Blob);
        assert_eq!(payload, b"lane Q was here\n");
    }

    /// **Dropping the table and reopening still answers correctly — fall back,
    /// then rebuild.**
    ///
    /// §14 calls the resolved table *droppable*: derived, verifiable against the
    /// verbatim truth, and deletable at any time without consulting a client.
    /// This is that sentence as an operation — `rm objects.exploded` between two
    /// opens — and §13.12's rule applied to it unchanged: **absent means fall
    /// back, never means wrong.**
    ///
    /// Three things are asserted, in this order:
    ///
    /// 1. the file really is gone and the reopened store really does start with
    ///    an empty table — otherwise the rest is testing a warm one;
    /// 2. **`reachable()` is already correct**, asked before any drain has been
    ///    waited on: the fall-back absorb runs inline and rebuilds what the
    ///    answer needs;
    /// 3. the table is **rebuilt** — one row per object again, the graph back,
    ///    and the rows written by *this* process, which is what says a rebuild
    ///    happened rather than a survival.
    ///
    /// Seen RED by `let exploded_is_whole = self.exploded.rows()? >= self.objects.len()
    /// as u64;` → `= true;` in `Absorber::adopt_journal`, so a dropped table is
    /// never noticed and nothing is re-queued. Assertion (2) fires first:
    /// **"reachable() after dropping the table: 1 objects, against 2683
    /// before"** — one object, the tip alone, which is the recorded bug's exact
    /// symptom ("`reachable` on a commit returns that commit alone rather than
    /// its closure") arrived at from the other direction. Restored.
    ///
    /// Seen RED a second time by `>=` → `>` in the same expression, which makes a
    /// *whole* table look short. This guard stays green — the rebuild is
    /// correct — and two others fire instead:
    /// [`a_clean_shutdown_and_reopen_keeps_the_graph_and_reachable`]'s "the
    /// reopen re-queued the pack, so the graph above is a recovery artefact
    /// rather than the table's", and the pre-existing
    /// [`a_clean_reopen_requeues_nothing_and_the_next_push_gets_a_fresh_ordinal`]'s
    /// "a fully absorbed pack was re-queued on open — the diff is not exact".
    /// That pair is the point: one guard catches the table not being rebuilt when
    /// it should be, the others catch it being rebuilt when it should not.
    /// Restored.
    #[test]
    fn dropping_the_exploded_table_and_reopening_still_answers() {
        let dir = tmpdir("exploded-dropped");
        let (pack, rows) = corpus_pack();
        let tip_hex;
        let closure_before;
        {
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
            let tip = store
                .graph_snapshot()
                .into_iter()
                .max_by_key(|c| c.generation)
                .expect("a graph");
            let tip_raw = hex::decode(&tip.oid).unwrap();
            closure_before = store.reachable(&[&tip_raw], &[]).unwrap().len();
            tip_hex = tip.oid;
        }

        // (1) — the drop.
        let table = dir.join("objects.exploded");
        assert!(table.exists(), "the store never built an exploded table");
        std::fs::remove_file(&table).unwrap();
        {
            let fresh = crate::exploded_arrow::ExplodedArchive::open(&table).unwrap();
            assert_eq!(fresh.rows().unwrap(), 0, "the drop did not drop anything");
        }
        // Opening an absent table does not create one — so this only has
        // anything to remove if the assertion above was inspecting a live file.
        let _ = std::fs::remove_file(&table);

        // (2) — a correct answer with the table gone.
        let store = GitStore::open(&dir, "rickard").unwrap();
        let tip_raw = hex::decode(&tip_hex).unwrap();
        let closure = store.reachable(&[&tip_raw], &[]).unwrap();
        assert_eq!(
            closure.len(),
            closure_before,
            "reachable() after dropping the table: {} objects, against {} before",
            closure.len(),
            closure_before
        );
        assert!(store.has(&rows[0].oid).unwrap());

        // (3) — and it is back. The GRAPH is what comes back at push time; a
        // blob comes back when somebody reads it, which is the whole of lazy
        // explode, so the count to compare against is the graph's and not the
        // pack's.
        store.wait_indexed();
        let stats = store.exploded_stats();
        let distinct: HashSet<Vec<u8>> = rows.iter().map(|r| r.oid.clone()).collect();
        // Commits, and NOT `of_kind(Tree)`: under the shipped policy a tree is
        // lazy, so the trees this test's own reads cached are in the lazy pool
        // and `of_kind` unions both.
        let graph_rows = store.exploded_of_kind(GitObjectKind::Commit).unwrap().len();
        // The EAGER pool, not `stats.rows`: step (2) above read objects through
        // the store, and every one of those reads wrote its object back into the
        // lazy pool, which `stats.rows` also counts. Comparing the sum against
        // the graph would be comparing "what the push wrote" against "what the
        // push wrote plus what this test happened to read".
        let eager = store.exploded_eager_rows().unwrap();
        assert_eq!(
            eager, graph_rows as u64,
            "the dropped table came back holding {eager} rows against a {}-object graph",
            graph_rows
        );
        assert!(
            eager > 0 && eager < distinct.len() as u64,
            "the rebuilt pool holds {eager} of {} objects — neither 0 nor all of them is \
             the lane Q shape",
            distinct.len()
        );
        assert!(
            stats.written > 0,
            "the table came back without this process writing a row — it was never dropped"
        );
        assert_eq!(
            store.commit_count(),
            commits_in(&rows),
            "the graph did not come back with the table: {} of {}",
            store.commit_count(),
            commits_in(&rows)
        );
        assert_eq!(
            store.object_count(),
            rows.len(),
            "the rebuild lost index rows"
        );
        eprintln!(
            "load {}; dropped and rebuilt: {} rows written, {} commits, tip closure {} objects",
            loadavg(),
            stats.written,
            store.commit_count(),
            closure.len(),
        );
    }

    // ─────────────────────────────────────────────────────────────────────
    // lane ENGINE-OPT scaffold: a graded linear-history generator, the O(C)
    // negotiation timing test, and the cross-push memo-invalidation guard.
    // ─────────────────────────────────────────────────────────────────────

    /// A linear commit history of `n` commits as a clone pack, optionally
    /// chained onto `start_parent` (an oid already in the store). Each commit
    /// carries its own one-entry tree and one unique blob, so the object count
    /// is `3n` and the graph is a single parent chain `n` deep — the shape that
    /// makes a negotiation's cost O(commits), which is what this lane memoizes.
    fn linear_history_pack_from(
        n: usize,
        start_parent: Option<&[u8]>,
    ) -> (Vec<u8>, Vec<Vec<u8>>) {
        use crate::object::canonical;
        use crate::pack_fixture::{build_pack, Item};
        let hash = GitHashKind::Sha1;
        let mut items: Vec<Item> = Vec::with_capacity(n * 3);
        let mut commits: Vec<Vec<u8>> = Vec::with_capacity(n);
        let mut prev: Option<Vec<u8>> = start_parent.map(|p| p.to_vec());
        for c in 0..n {
            let body = format!("blob for commit {c} of {n}, parent {prev:?}\n").into_bytes();
            let blob_oid = hash.oid_of(&canonical(GitObjectKind::Blob, &body));
            items.push(Item::Whole(GitObjectKind::Blob, body));

            let mut tpayload = Vec::new();
            tpayload.extend_from_slice(b"100644 f.bin\0");
            tpayload.extend_from_slice(&blob_oid);
            let tree_oid = hash.oid_of(&canonical(GitObjectKind::Tree, &tpayload));
            items.push(Item::Whole(GitObjectKind::Tree, tpayload));

            let mut cbody = format!("tree {}\n", hex::encode(&tree_oid));
            if let Some(parent) = &prev {
                cbody.push_str(&format!("parent {}\n", hex::encode(parent)));
            }
            cbody.push_str(
                "author A U Thor <a@example.com> 1700000000 +0000\n\
                 committer A U Thor <a@example.com> 1700000000 +0000\n\n\
                 commit\n",
            );
            let cbody = cbody.into_bytes();
            let commit_oid = hash.oid_of(&canonical(GitObjectKind::Commit, &cbody));
            items.push(Item::Whole(GitObjectKind::Commit, cbody));
            prev = Some(commit_oid.clone());
            commits.push(commit_oid);
        }
        (build_pack(&items), commits)
    }

    /// The O(commits) negotiation cost, graded. `reachable` rebuilt the
    /// `by_commit` and graph maps on every call — O(C) regardless of how few
    /// objects the request touched — so the `root` query (closure ~1) isolates
    /// that map-build cost from the walk, and the `tip` query (full closure)
    /// pays both. Median of nine, min–max, with the loadavg of its own run.
    #[test]
    #[ignore = "perf measurement; run with --release --ignored --nocapture"]
    fn bench_reachable_negotiation() {
        for &n in &[48usize, 2_000, 20_000] {
            let dir = tmpdir(&format!("bench-reach-{n}"));
            let (pack, commits) = linear_history_pack_from(n, None);
            let store = GitStore::open(&dir, "rickard").unwrap();
            store.put_pack(&pack).unwrap();
            store.wait_indexed();
            let root = commits.first().unwrap().clone();
            let tip = commits.last().unwrap().clone();
            // warm the reach bitmaps once so the loop times the negotiation, not
            // the one-time bitmap build.
            let _ = store.reachable(&[&tip], &[]).unwrap();

            let reps = 9usize;
            let (mut root_t, mut tip_t) = (Vec::new(), Vec::new());
            for _ in 0..reps {
                let t = Instant::now();
                let _ = store.reachable(&[&root], &[]).unwrap();
                root_t.push(t.elapsed());
                let t = Instant::now();
                let _ = store.reachable(&[&tip], &[]).unwrap();
                tip_t.push(t.elapsed());
            }
            root_t.sort();
            tip_t.sort();
            let tip_closure = store.reachable(&[&tip], &[]).unwrap().len();
            eprintln!(
                "load {}; n={n}: reachable(root, closure~1) med {:?} [{:?}..{:?}]  |  \
                 reachable(tip, closure {}) med {:?} [{:?}..{:?}]",
                loadavg(),
                root_t[reps / 2],
                root_t[0],
                root_t[reps - 1],
                tip_closure,
                tip_t[reps / 2],
                tip_t[0],
                tip_t[reps - 1],
            );
        }
    }

    /// **A negotiation answered from the memoized reach view is the answer a
    /// cold fold gives — including across a push that changes the fold.**
    ///
    /// Lane ENGINE-OPT memoizes the `by_commit` and graph lookup maps that
    /// `reachable` used to rebuild on every call, in `Derived`, folded by the
    /// same call that folds `commit_raw`. The bug this discipline exists to
    /// catch is a memo that SURVIVES a fold it should not: after a push adds
    /// commits, the warm store must answer with the new closure, byte-identical
    /// (order included — both spaces sort the same oids) to a fresh store
    /// opened cold on the same bytes.
    ///
    /// Seen RED by skipping the memo rebuild in `refold` (leaving `graph_index`
    /// from before the push): the post-push tip is then not a known commit,
    /// `reachable` contributes it alone, and the warm answer is a handful of
    /// objects where the cold fold gives the whole chain — the assertion names
    /// both lengths.
    #[test]
    fn reachable_memo_matches_a_cold_fold_across_a_push() {
        let dir = tmpdir("reach-memo-across-push");
        let (pack_a, commits_a) = linear_history_pack_from(40, None);
        let store = GitStore::open(&dir, "rickard").unwrap();
        store.put_pack(&pack_a).unwrap();
        store.wait_indexed();
        let tip_a = commits_a.last().unwrap().clone();
        let warm_a = store.reachable(&[&tip_a], &[]).unwrap();
        assert!(
            warm_a.len() > 10,
            "A's closure is too small to tell a stale memo from a coincidence: {}",
            warm_a.len()
        );

        // Push B: a second linear history chained ONTO A's tip, so B's tip
        // closes over A too. This is a fold that genuinely grows the graph.
        let (pack_b, commits_b) = linear_history_pack_from(40, Some(&tip_a));
        store.put_pack(&pack_b).unwrap();
        store.wait_indexed();
        let tip_b = commits_b.last().unwrap().clone();
        let warm_b_full = store.reachable(&[&tip_b], &[]).unwrap();
        let warm_b_minus_a = store.reachable(&[&tip_b], &[&tip_a]).unwrap();
        assert!(
            warm_b_full.len() > warm_a.len(),
            "the push did not grow the closure ({} !> {}), so the memo could be \
             stale and this test would not know",
            warm_b_full.len(),
            warm_a.len()
        );

        // The cold oracle: a fresh handle on the same bytes, folded from disk.
        drop(store);
        let cold = GitStore::open(&dir, "rickard").unwrap();
        cold.wait_indexed();
        let cold_b_full = cold.reachable(&[&tip_b], &[]).unwrap();
        let cold_b_minus_a = cold.reachable(&[&tip_b], &[&tip_a]).unwrap();

        assert_eq!(
            warm_b_full, cold_b_full,
            "warm memo != cold fold for want=tip_b (a memo that survived the push)"
        );
        assert_eq!(
            warm_b_minus_a, cold_b_minus_a,
            "warm memo != cold fold for want=tip_b have=tip_a"
        );
        // tip_a is still answered identically after the push (its closure did
        // not change; only the store grew around it).
        let cold_a = cold.reachable(&[&tip_a], &[]).unwrap();
        assert_eq!(
            warm_a, cold_a,
            "tip_a's closure changed across the push"
        );
        eprintln!(
            "load {}; A closure {}, B closure {}, B−A {}",
            loadavg(),
            warm_a.len(),
            warm_b_full.len(),
            warm_b_minus_a.len(),
        );
    }

    /// **The clone-after-push lever, witnessed.** A push folds; the fold used
    /// to drop the reachability table and the next reader rebuilt it by
    /// walking every commit's trees — 0.2 ms per object, 10 s on a 51k-object
    /// repository, measured 2026-09-05. Now the fold CARRIES the table
    /// (renumbered) when the graph grew by no more than the carry rule allows.
    /// RED before the carry: `reach_rebuilt` went 1 → 2 across the push.
    #[test]
    fn a_push_carries_the_reach_table_across_the_fold() {
        let dir = tmpdir("reach-carry-across-push");
        let (pack_a, commits_a) = linear_history_pack_from(40, None);
        let store = GitStore::open(&dir, "rickard").unwrap();
        store.put_pack(&pack_a).unwrap();
        store.wait_indexed();
        let tip_a = commits_a.last().unwrap().clone();
        let warm_a = store.reachable(&[&tip_a], &[]).unwrap();
        assert_eq!(store.reach_source_stats(), (0, 1), "the first read builds the table once");

        // A push of 30 commits onto A's tip: within the carry rule (64 or a
        // tenth of 40, whichever is larger), so the fold carries.
        let (pack_b, commits_b) = linear_history_pack_from(30, Some(&tip_a));
        store.put_pack(&pack_b).unwrap();
        store.wait_indexed();
        let tip_b = commits_b.last().unwrap().clone();
        let warm_b = store.reachable(&[&tip_b], &[]).unwrap();
        let warm_b_minus_a = store.reachable(&[&tip_b], &[&tip_a]).unwrap();
        assert_eq!(
            store.reach_source_stats(),
            (1, 1),
            "the fold carried the table; no read rebuilt it"
        );
        assert!(warm_b.len() > warm_a.len(), "the push grew the closure");
        assert_eq!(warm_b_minus_a.len(), warm_b.len() - warm_a.len(), "tip_b − tip_a is exactly the pushed objects");
        assert_eq!(store.reachable(&[&tip_a], &[]).unwrap(), warm_a, "tip_a's closure survived the renumbering");

        // The cold oracle: the same bytes folded from disk with a fresh table.
        drop(store);
        let cold = GitStore::open(&dir, "rickard").unwrap();
        cold.wait_indexed();
        assert_eq!(cold.reachable(&[&tip_b], &[]).unwrap(), warm_b, "carried != rebuilt for want=tip_b");
        assert_eq!(
            cold.reachable(&[&tip_b], &[&tip_a]).unwrap(),
            warm_b_minus_a,
            "carried != rebuilt for want=tip_b have=tip_a"
        );
        assert_eq!(cold.reach_source_stats(), (0, 1), "the cold handle built once and carried nothing");
    }

    /// The carry rule's other arm: a push larger than the allowance drops the
    /// table and the next read rebuilds it whole — bounded staleness, never a
    /// walk that grows without limit.
    #[test]
    fn a_push_past_the_carry_allowance_rebuilds_the_reach_table() {
        let dir = tmpdir("reach-carry-rebuild");
        let (pack_a, commits_a) = linear_history_pack_from(20, None);
        let store = GitStore::open(&dir, "rickard").unwrap();
        store.put_pack(&pack_a).unwrap();
        store.wait_indexed();
        let tip_a = commits_a.last().unwrap().clone();
        let _ = store.reachable(&[&tip_a], &[]).unwrap();
        assert_eq!(store.reach_source_stats(), (0, 1));
        // 100 commits > max(64, 20/10): over the allowance.
        let (pack_b, commits_b) = linear_history_pack_from(100, Some(&tip_a));
        store.put_pack(&pack_b).unwrap();
        store.wait_indexed();
        let tip_b = commits_b.last().unwrap().clone();
        let warm_b = store.reachable(&[&tip_b], &[]).unwrap();
        assert_eq!(store.reach_source_stats(), (0, 2), "over the allowance: not carried, rebuilt once more");
        drop(store);
        let cold = GitStore::open(&dir, "rickard").unwrap();
        cold.wait_indexed();
        assert_eq!(cold.reachable(&[&tip_b], &[]).unwrap(), warm_b, "rebuilt warm != cold fold");
    }
}