liminal-server 0.24.1

//! DUR-024: a checkpoint published in batches under the store's own
//! admission bound, off the owner's write path, with the locator last.
//!
//! On 10 October 2026 the door's room crossed its checkpoint boundary and
//! published its retained history as one request of 3,338,540,899 bytes;
//! haematite admits 67,108,864 per shard and refused it by name, the
//! refusal tore every connection down, and the room was write-dead for 91
//! minutes. Here a publication whose history exceeds the batch bound is
//! split into successive commits of at most the bound each, in node key
//! order; the checkpoint body and its locator ride in the final commit only,
//! so no reader loads a locator whose history is not wholly present. Every
//! history node is immutable and keyed by its content, so a commit that
//! landed is never rewritten and an interrupted publication resumes at the
//! next crossing by staging only the nodes still absent.
//!
//! The publication runs on its own thread (R3): the owner's lock is taken
//! once, to capture the history and stage the body, and is released before
//! any batch is submitted; a live write never waits behind a batch, and the
//! publisher never takes the owner's lock. A refusal of any batch is logged
//! once by name, marks the checkpoint owed in the [`PublicationLedger`], and
//! changes nothing about the owner: it keeps serving from its replay and the
//! next crossing retries. The owner's retained history root stays the one
//! whose nodes are all present; the trie is canonical for a key set, so the
//! retry's capture finds the committed nodes and stages the rest.
//!
//! The bound is read from the store ([`DurableStore::max_request_bytes`]),
//! never assumed or learnt from a refusal; liminal pins its own batch bound
//! at a quarter of haematite's default, so three quarters of a shard's queue
//! stay free for live writes while a batch is admitted (R2).
//!
//! The census (R5) prepares a conversation's capture exactly as a
//! publication would and counts it by kind without writing.

use std::collections::{BTreeMap, HashMap};
use std::fmt::Write as _;
use std::sync::{
    Arc, Mutex,
    atomic::{AtomicBool, Ordering},
};

use liminal::durability::bridge::block_on;
use liminal::durability::settle::apply_routed_batch_settled;
use liminal::durability::{DurabilityError, DurableStore, RoutedBatchItem};

use super::log::OperationLog;
use super::operation_batch::OperationBatch;
use super::outbox::ConversationOutboxLimits;
use super::outbox_log::OutboxLog;
use super::restore_checkpoint_history::{
    CheckpointHistoryNode, CheckpointHistoryRoot, HistoryNodeKind, ROOT_ENCODED_BYTES,
};
use super::restore_checkpoint_locator::CheckpointLocator;
use super::restore_from_checkpoint::{CheckpointDecision, load_checkpoint};
use super::state::{ConversationAuthority, checkpoint_owner::ConversationAuthorityRestoreContext};

/// The most value bytes one publication commit carries: 16 MiB, a quarter of
/// haematite's default 64 MiB per-shard admission bound, so three quarters of
/// a shard's queue stays free for live writes while a batch is admitted.
/// Waffles' number by Tom's rule of 10 October 2026; pinned in
/// `tests/ceilings/checkpoint.toml`. The store's own bound, when smaller,
/// governs instead ([`publication_bound`]).
pub(super) const CHECKPOINT_BATCH_BYTES: usize = 16 * 1024 * 1024;

/// The bound a publication packs its batches under: the pinned batch bound,
/// or the store's own admission bound when that is smaller.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct PublicationBound {
    /// [`CHECKPOINT_BATCH_BYTES`].
    pub pinned: usize,
    /// What the store admits, when it names a bound.
    pub store: Option<usize>,
    /// The smaller of the two: what every batch is kept under.
    pub bound: usize,
}

impl PublicationBound {
    /// The largest single item, or final batch, a publication may submit:
    /// what the store admits in one request, and no limit for a store that
    /// names none. The batch bound is a packing target; an item over it is
    /// committed on its own (DUR-024, 11 Oct 2026: a checkpoint body of a
    /// large conversation is one item larger than the 16 MiB target and well
    /// under the store's 64 MiB, and refusing it meant a full replay at
    /// every boot).
    #[must_use]
    pub fn item_limit(&self) -> usize {
        self.store.unwrap_or(usize::MAX)
    }
}

/// A test's batch target in place of [`CHECKPOINT_BATCH_BYTES`], so a
/// fixture of a few hundred operations can publish in several batches while
/// its store still admits the whole body. nextest runs one test per process,
/// so a test that sets it owns it. Zero is unset.
#[cfg(test)]
pub(super) static TEST_BATCH_BYTES: std::sync::atomic::AtomicUsize =
    std::sync::atomic::AtomicUsize::new(0);

#[cfg(test)]
fn pinned_batch_bytes() -> usize {
    match TEST_BATCH_BYTES.load(Ordering::SeqCst) {
        0 => CHECKPOINT_BATCH_BYTES,
        set => set,
    }
}

#[cfg(not(test))]
const fn pinned_batch_bytes() -> usize {
    CHECKPOINT_BATCH_BYTES
}

/// Reads the store's bound and settles the publication's.
pub fn publication_bound(store: &dyn DurableStore) -> PublicationBound {
    let store_bound = store.max_request_bytes();
    let pinned = pinned_batch_bytes();
    PublicationBound {
        pinned,
        store: store_bound,
        bound: store_bound.map_or(pinned, |admitted| admitted.min(pinned)),
    }
}

/// A publication refused before anything was submitted (R2): one item alone
/// is larger than the store admits, so no batch could ever carry it.
#[derive(Debug, Clone, PartialEq, Eq, thiserror::Error)]
pub(super) enum CheckpointPublishError {
    #[error("checkpoint_item_exceeds_store_bound: key={key} bytes={bytes} bound={bound}")]
    ItemExceedsStoreBound {
        /// The item's routed key, as text where it is text and hex where it
        /// is not.
        key: String,
        bytes: usize,
        bound: usize,
    },
}

/// A key as the refusal names it.
fn key_name(key: &[u8]) -> String {
    match std::str::from_utf8(key) {
        Ok(text) if text.chars().all(|c| !c.is_control()) => text.to_owned(),
        _ => key.iter().fold(String::new(), |mut hex, byte| {
            if write!(hex, "{byte:02x}").is_err() {
                hex.push_str("??");
            }
            hex
        }),
    }
}

/// Splits `nodes` into batches of at most `bound` value bytes each, in key
/// order, filled greedily. A node over `bound` but within `item_limit` is a
/// batch of its own; a node over `item_limit` is the named refusal; an empty
/// map is no batches.
///
/// # Errors
/// [`CheckpointPublishError::ItemExceedsStoreBound`], before anything is
/// assembled past the item.
pub(super) fn split_history(
    nodes: BTreeMap<Vec<u8>, Vec<u8>>,
    bound: usize,
    item_limit: usize,
) -> Result<Vec<Vec<RoutedBatchItem>>, CheckpointPublishError> {
    let mut batches: Vec<Vec<RoutedBatchItem>> = Vec::new();
    let mut current: Vec<RoutedBatchItem> = Vec::new();
    let mut current_bytes = 0_usize;
    for (key, value) in nodes {
        let bytes = value.len();
        if bytes > item_limit {
            return Err(CheckpointPublishError::ItemExceedsStoreBound {
                key: key_name(&key),
                bytes,
                bound: item_limit,
            });
        }
        if current_bytes.saturating_add(bytes) > bound && !current.is_empty() {
            batches.push(std::mem::take(&mut current));
            current_bytes = 0;
        }
        current_bytes = current_bytes.saturating_add(bytes);
        current.push(RoutedBatchItem {
            key,
            expected: None,
            value,
            ttl: None,
        });
    }
    if !current.is_empty() {
        batches.push(current);
    }
    Ok(batches)
}

/// Value bytes of one batch.
pub(super) fn batch_bytes(items: &[RoutedBatchItem]) -> usize {
    items.iter().map(|item| item.value.len()).sum()
}

/// How many commits a publication of `nodes` makes at `bound`, with `tail`
/// value bytes (the body, the locator and the receipts) that ride in the
/// final commit when it has room and in a commit of their own when it does
/// not. Zero for nothing to write. The census answers with this, and R1's
/// publication commits exactly this many ([`CheckpointPublication::new`]).
pub(super) fn commit_count(
    node_bytes: impl IntoIterator<Item = usize>,
    tail: usize,
    bound: usize,
) -> usize {
    let mut batches = 0_usize;
    let mut current_bytes = 0_usize;
    let mut open = false;
    for bytes in node_bytes {
        if open && current_bytes.saturating_add(bytes) > bound {
            batches += 1;
            current_bytes = 0;
        }
        open = true;
        current_bytes = current_bytes.saturating_add(bytes);
    }
    if open {
        batches += 1;
    }
    if tail == 0 {
        return batches;
    }
    if open && current_bytes.saturating_add(tail) <= bound {
        batches
    } else {
        batches + 1
    }
}

/// What a crossing hands to the publisher: the history batches, in order,
/// and the final batch carrying the body, the locator, the receipts and
/// whatever history fitted beside them.
pub(super) struct CheckpointPublication {
    pub(super) conversation_id: u64,
    pub(super) route: String,
    pub(super) history: Vec<Vec<RoutedBatchItem>>,
    pub(super) final_batch: OperationBatch,
    pub(super) bound: PublicationBound,
    pub(super) bytes: usize,
}

impl CheckpointPublication {
    /// Lays out a publication: `history` split under the bound, the last
    /// history batch folded into `final_batch` when it fits beside what the
    /// final batch already carries, kept as its own commit when it does not.
    ///
    /// # Errors
    /// [`CheckpointPublishError::ItemExceedsStoreBound`] for a node or the
    /// final batch's own items over the bound; the batch's refusal of a key.
    pub(super) fn new(
        conversation_id: u64,
        route: &str,
        mut nodes: BTreeMap<Vec<u8>, Vec<u8>>,
        mut final_batch: OperationBatch,
        bound: PublicationBound,
    ) -> Result<Self, PublicationLayoutError> {
        // A body's parts are content-addressed like history nodes and are
        // committed with them, before the locator that names the body.
        nodes.extend(
            final_batch
                .take_body_parts()
                .into_iter()
                .map(|part| (part.key, part.value)),
        );
        let tail = final_batch.value_bytes();
        if tail > bound.item_limit() {
            return Err(CheckpointPublishError::ItemExceedsStoreBound {
                key: "checkpoint body".to_owned(),
                bytes: tail,
                bound: bound.item_limit(),
            }
            .into());
        }
        let node_bytes: usize = nodes.values().map(Vec::len).sum();
        let mut history = split_history(nodes, bound.bound, bound.item_limit())?;
        if let Some(last) = history.last() {
            if batch_bytes(last).saturating_add(tail) <= bound.bound {
                if let Some(last) = history.pop() {
                    for item in last {
                        final_batch.stage(item)?;
                    }
                }
            }
        }
        Ok(Self {
            conversation_id,
            route: route.to_owned(),
            history,
            final_batch,
            bound,
            bytes: node_bytes.saturating_add(tail),
        })
    }

    /// How many commits this publication makes.
    pub(super) fn commits(&self) -> usize {
        self.history.len().saturating_add(1)
    }
}

/// A publication that could not be laid out.
#[derive(Debug, thiserror::Error)]
pub(super) enum PublicationLayoutError {
    #[error(transparent)]
    Refused(#[from] CheckpointPublishError),
    #[error(transparent)]
    Store(#[from] DurabilityError),
}

/// One batch the store refused, by its 1-based index in the publication.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct PublicationRefusal {
    pub batch_index: usize,
    pub bytes: usize,
    pub bound: usize,
    pub refusal: String,
}

impl std::fmt::Display for PublicationRefusal {
    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(
            formatter,
            "checkpoint publication refused at batch {}: {} bytes against a bound of {}: {}",
            self.batch_index, self.bytes, self.bound, self.refusal
        )
    }
}

/// What a completed publication came to.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub struct PublishedCheckpoint {
    pub commits: usize,
    pub bytes: usize,
}

/// Commits the history batches of `publication` in order, then its final
/// batch, and runs the cut its receipts promised. Stops at the first
/// refusal, named by batch; nothing is retried here (R3: the next crossing
/// retries).
///
/// # Errors
/// The refusing batch's index, bytes and bound, with the store's own words.
pub(super) async fn commit_publication(
    store: &dyn DurableStore,
    publication: CheckpointPublication,
) -> Result<PublishedCheckpoint, PublicationRefusal> {
    let CheckpointPublication {
        conversation_id,
        route,
        history,
        mut final_batch,
        bound,
        bytes,
    } = publication;
    let total = history.len().saturating_add(1);
    for (index, items) in history.into_iter().enumerate() {
        let batch_bytes = batch_bytes(&items);
        if let Err(error) = apply_routed_batch_settled(store, &route, items).await {
            return Err(PublicationRefusal {
                batch_index: index.saturating_add(1),
                bytes: batch_bytes,
                bound: bound.bound,
                refusal: error.to_string(),
            });
        }
    }
    let final_bytes = final_batch.value_bytes();
    let prune = final_batch.take_prune();
    // The publication is not a live write: an outcome the store could not
    // answer is settled by the helper's own readback, never told to a client.
    if let Err(error) = final_batch.commit(store, || ()).await {
        return Err(PublicationRefusal {
            batch_index: total,
            bytes: final_bytes,
            bound: bound.bound,
            refusal: error.to_string(),
        });
    }
    if let Some(plan) = prune {
        super::prune::report_cut(store, conversation_id, &plan).await;
    }
    super::restore_checkpoint_history::retire_v1_history(store, conversation_id).await;
    Ok(PublishedCheckpoint {
        commits: total,
        bytes,
    })
}

/// Where one conversation's publication stands, for the crossing that asks
/// before it captures and for the test that asks whether a checkpoint is
/// owed.
#[derive(Debug, Clone, PartialEq, Eq)]
pub enum PublicationState {
    /// Handed to the publisher and not yet answered.
    InFlight,
    /// Refused: the checkpoint is owed, and the next crossing retries.
    Owed(PublicationRefusal),
    /// Committed whole.
    Published(PublishedCheckpoint),
}

/// The per-conversation publication states, shared by the handler and every
/// publisher thread. It is not the owner's lock and is never held across a
/// submission.
#[derive(Debug, Clone, Default)]
pub struct PublicationLedger {
    states: Arc<Mutex<HashMap<u64, PublicationState>>>,
    bound_logged: Arc<AtomicBool>,
}

impl PublicationLedger {
    /// The states, or `None` when the ledger's lock is poisoned, said once
    /// per read by name. Every caller fails closed on `None`: no new
    /// publication is handed off, so the owner keeps serving from its
    /// replay and nothing is published over an unreadable ledger.
    fn states(&self) -> Option<std::sync::MutexGuard<'_, HashMap<u64, PublicationState>>> {
        self.states.lock().map_or_else(
            |_| {
                tracing::warn!(
                    "checkpoint publication ledger lock is poisoned; no publication is handed off and the owner keeps serving from its replay"
                );
                None
            },
            Some,
        )
    }

    /// Claims the conversation for a publication; `false` when one is in
    /// flight already.
    pub(super) fn begin(&self, conversation_id: u64) -> bool {
        let Some(mut states) = self.states() else {
            return false;
        };
        if matches!(
            states.get(&conversation_id),
            Some(PublicationState::InFlight)
        ) {
            return false;
        }
        states.insert(conversation_id, PublicationState::InFlight);
        true
    }

    /// Whether a publication is in flight; a poisoned ledger answers yes,
    /// so the crossing stages no capture it could not account for.
    pub(super) fn in_flight(&self, conversation_id: u64) -> bool {
        self.states().is_none_or(|states| {
            matches!(
                states.get(&conversation_id),
                Some(PublicationState::InFlight)
            )
        })
    }

    pub(super) fn finish(&self, conversation_id: u64, state: PublicationState) {
        if let Some(mut states) = self.states() {
            states.insert(conversation_id, state);
        }
    }

    /// Whether the conversation's last publication was refused and its
    /// checkpoint is owed to the next crossing.
    #[cfg(test)]
    pub fn checkpoint_owed(&self, conversation_id: u64) -> bool {
        self.states().is_some_and(|states| {
            matches!(
                states.get(&conversation_id),
                Some(PublicationState::Owed(_))
            )
        })
    }

    /// The ledger's record of the conversation's last publication.
    #[cfg(test)]
    pub fn state(&self, conversation_id: u64) -> Option<PublicationState> {
        self.states()
            .and_then(|states| states.get(&conversation_id).cloned())
    }

    /// Says once, at the owner's first publication, which bound governs
    /// when the store's is smaller than the pinned one (R2).
    pub(super) fn log_bound_once(&self, bound: PublicationBound) {
        if bound.bound >= bound.pinned {
            return;
        }
        if self
            .bound_logged
            .compare_exchange(false, true, Ordering::SeqCst, Ordering::SeqCst)
            .is_ok()
        {
            tracing::info!(
                pinned_batch_bytes = bound.pinned,
                store_max_request_bytes = bound.bound,
                "checkpoint batch bound: the store admits less than the pinned batch bound, so the store's bound is used"
            );
        }
    }
}

/// Runs publications off the owner's write path, one thread each, and keeps
/// their ledger.
#[derive(Debug, Default)]
pub struct CheckpointPublisher {
    ledger: PublicationLedger,
    threads: Mutex<Vec<std::thread::JoinHandle<()>>>,
    /// The layer above, told of every committed crossing.
    crossing_observer: std::sync::OnceLock<Arc<dyn crate::server::connection::ConnectionNotifier>>,
}

impl CheckpointPublisher {
    /// Installs the notifier told of every committed crossing; the first
    /// installed stays.
    pub fn install_crossing_observer(
        &self,
        observer: Arc<dyn crate::server::connection::ConnectionNotifier>,
    ) {
        if self.crossing_observer.set(observer).is_err() {
            tracing::warn!("checkpoint crossing observer was already installed; the first stays");
        }
    }

    /// Tells the installed notifier, if any, of one committed crossing.
    pub fn crossing_committed(&self, crossing: &crate::server::connection::CheckpointCrossing) {
        if let Some(observer) = self.crossing_observer.get() {
            observer.on_checkpoint_crossing(crossing);
        }
    }

    pub const fn ledger(&self) -> &PublicationLedger {
        &self.ledger
    }

    /// Hands `publication` to its own thread. The ledger already holds the
    /// conversation in flight ([`PublicationLedger::begin`]).
    ///
    /// # Errors
    /// The thread could not be started: the publication is answered as
    /// refused at batch 0 so the next crossing retries.
    pub(super) fn spawn(
        &self,
        store: Arc<dyn DurableStore>,
        publication: CheckpointPublication,
    ) -> Result<(), std::io::Error> {
        let conversation_id = publication.conversation_id;
        let ledger = self.ledger.clone();
        let commits = publication.commits();
        let bytes = publication.bytes;
        let bound = publication.bound;
        tracing::info!(
            conversation_id,
            batches = commits,
            bytes,
            bound = bound.bound,
            "checkpoint publication handed off the owner's write path"
        );
        let spawned = std::thread::Builder::new()
            .name(format!("checkpoint-publish-{conversation_id}"))
            .spawn(move || {
                let outcome = block_on(commit_publication(store.as_ref(), publication));
                let state = match outcome {
                    Ok(Ok(published)) => {
                        tracing::info!(
                            conversation_id,
                            commits = published.commits,
                            bytes = published.bytes,
                            "checkpoint published in batches"
                        );
                        PublicationState::Published(published)
                    }
                    Ok(Err(refusal)) => {
                        log_refusal(conversation_id, &refusal);
                        PublicationState::Owed(refusal)
                    }
                    Err(bridge) => {
                        let refusal = PublicationRefusal {
                            batch_index: 0,
                            bytes,
                            bound: bound.bound,
                            refusal: bridge.to_string(),
                        };
                        log_refusal(conversation_id, &refusal);
                        PublicationState::Owed(refusal)
                    }
                };
                ledger.finish(conversation_id, state);
            });
        match spawned {
            Ok(handle) => {
                // A poisoned handle list only loses the test seam's join;
                // the thread runs and records its own outcome either way.
                if let Ok(mut threads) = self.threads.lock() {
                    threads.push(handle);
                } else {
                    tracing::warn!(
                        conversation_id,
                        "checkpoint publisher handle list lock is poisoned; the publication runs unjoined"
                    );
                }
                Ok(())
            }
            Err(error) => {
                let refusal = PublicationRefusal {
                    batch_index: 0,
                    bytes,
                    bound: bound.bound,
                    refusal: format!("publisher thread could not start: {error}"),
                };
                log_refusal(conversation_id, &refusal);
                self.ledger
                    .finish(conversation_id, PublicationState::Owed(refusal));
                Err(error)
            }
        }
    }

    /// Records a publication refused before anything was submitted (R2's
    /// named refusal): the checkpoint is owed, the owner keeps serving.
    pub(super) fn refuse_before_submission(
        &self,
        conversation_id: u64,
        refusal: PublicationRefusal,
    ) {
        log_refusal(conversation_id, &refusal);
        self.ledger
            .finish(conversation_id, PublicationState::Owed(refusal));
    }

    /// Waits for every publication handed off so far: a test's seam, so a
    /// count is read after the thread is done and never raced.
    ///
    /// # Errors
    /// A publisher thread that panicked.
    #[cfg(test)]
    pub fn join_all(&self) -> Result<(), String> {
        let handles: Vec<_> = std::mem::take(
            &mut *self
                .threads
                .lock()
                .map_err(|_| "the publisher handle list lock is poisoned".to_owned())?,
        );
        for handle in handles {
            handle
                .join()
                .map_err(|_| "a checkpoint publisher thread panicked".to_owned())?;
        }
        Ok(())
    }
}

/// The one line a refusal gets (R3): the conversation, the batch, the bytes
/// and the bound, with the store's own words, and what happens next.
fn log_refusal(conversation_id: u64, refusal: &PublicationRefusal) {
    tracing::warn!(
        conversation_id,
        batch_index = refusal.batch_index,
        bytes = refusal.bytes,
        bound = refusal.bound,
        refusal = %refusal.refusal,
        "checkpoint publication refused by the store; the owner keeps serving from its replay, the checkpoint is owed, and the next crossing retries"
    );
}

/// What one conversation's checkpoint holds, counted without writing (R5).
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct CheckpointCensus {
    pub conversation_id: u64,
    /// Absent history nodes by kind: how many, and their bytes.
    pub nodes: BTreeMap<&'static str, (usize, usize)>,
    pub history_nodes: usize,
    pub history_bytes: usize,
    pub tokens: usize,
    pub committed_admissions: usize,
    pub slots: usize,
    pub attach_provenance: usize,
    pub retired: usize,
    /// The latest locator's covered log sequence and extension boundary,
    /// when one is published.
    pub locator: Option<(u64, u64)>,
    pub log_rows_behind: u64,
    pub log_rows_ahead: u64,
    pub outbox_rows_behind: u64,
    pub outbox_rows_ahead: u64,
    /// The body and locator bytes a publication would stage, or why the
    /// owner cannot be captured.
    pub body: Result<usize, String>,
    pub bound: PublicationBound,
    pub batches: usize,
}

impl CheckpointCensus {
    /// The figures as plain lines, one per kind; no row body is ever printed.
    pub fn lines(&self) -> Vec<String> {
        let mut lines = vec![format!("conversation: {}", self.conversation_id)];
        for (kind, (count, bytes)) in &self.nodes {
            lines.push(format!("{kind}: {count} nodes, {bytes} bytes"));
        }
        lines.push(format!(
            "history to publish: {} nodes, {} bytes",
            self.history_nodes, self.history_bytes
        ));
        lines.push(format!("tokens: {}", self.tokens));
        lines.push(format!(
            "committed admissions: {}",
            self.committed_admissions
        ));
        lines.push(format!("slots: {}", self.slots));
        lines.push(format!(
            "attach provenance entries: {}",
            self.attach_provenance
        ));
        lines.push(format!("retired identities: {}", self.retired));
        match self.locator {
            Some((covered, extension)) => lines.push(format!(
                "latest locator: covered log sequence {covered}, extension boundary {extension}"
            )),
            None => lines.push("latest locator: absent".to_owned()),
        }
        lines.push(format!(
            "operation log rows: {} behind the locator, {} ahead",
            self.log_rows_behind, self.log_rows_ahead
        ));
        lines.push(format!(
            "outbox rows: {} behind the locator, {} ahead",
            self.outbox_rows_behind, self.outbox_rows_ahead
        ));
        match &self.body {
            Ok(bytes) => lines.push(format!("checkpoint body and locator: {bytes} bytes")),
            Err(refusal) => lines.push(format!("checkpoint body: not capturable: {refusal}")),
        }
        lines.push(format!(
            "batch bound: {} bytes (pinned {}, store {})",
            self.bound.bound,
            self.bound.pinned,
            self.bound
                .store
                .map_or_else(|| "unbounded".to_owned(), |bound| bound.to_string())
        ));
        lines.push(format!("batches a publication commits: {}", self.batches));
        lines
    }
}

/// A census that could not be taken.
#[derive(Debug, thiserror::Error)]
pub enum CensusError {
    #[error("the conversation could not be replayed: {0}")]
    Replay(String),
    #[error("the checkpoint locator could not be read: {0}")]
    Locator(String),
    #[error("the history could not be captured: {0}")]
    Capture(String),
    #[error("a captured node could not be decoded: {0}")]
    Node(String),
    #[error("the outbox limits could not be built: {0}")]
    Limits(String),
    #[error("the store could not be read: {0}")]
    Store(String),
}

/// Prepares `conversation_id`'s capture exactly as a publication would and
/// counts it. Reads only: nothing is staged against the store and nothing
/// is committed.
///
/// # Errors
/// [`CensusError`], by step.
pub fn checkpoint_census(
    store: &Arc<dyn DurableStore>,
    config: &crate::config::types::ParticipantConfig,
    conversation_id: u64,
) -> Result<CheckpointCensus, CensusError> {
    let limits =
        ConversationOutboxLimits::try_new(config.max_retained_record_rows, config.identity_slots)
            .map_err(|error| CensusError::Limits(error.to_string()))?;
    let log = OperationLog::new(Arc::clone(store), conversation_id);
    let outbox_log = OutboxLog::new(Arc::clone(store), conversation_id);
    let (mut owner, audit, schema, prefix) = block_on(ConversationAuthority::replay_for_rebuild(
        conversation_id,
        &log,
        &outbox_log,
        config,
        limits,
    ))
    .map_err(|error| CensusError::Replay(error.to_string()))?
    .map_err(|error| CensusError::Replay(error.to_string()))?;
    let (root, locator) = latest_root_and_locator(store, config, conversation_id)?;
    let history = block_on(
        super::restore_checkpoint_history_capture::capture_identity_history(
            store.as_ref(),
            &owner,
            root,
        ),
    )
    .map_err(|error| CensusError::Capture(error.to_string()))?
    .map_err(|error| CensusError::Capture(error.to_string()))?;
    let nodes = count_nodes_by_kind(&history)?;
    let bound = publication_bound(store.as_ref());
    // The body is framed exactly as a rebuild frames it: against the audit
    // the replay composed and the root the capture answered, retained on
    // the replayed owner first (restore_checkpoint_rebuild.rs does the same
    // before it publishes).
    let body = owner
        .retain_checkpoint_metadata(history.root, &audit)
        .map_err(|error| error.to_string())
        .and_then(|()| checkpoint_body_bytes(store.as_ref(), &owner, schema, history.root, prefix));
    let batches = commit_count(
        history.nodes().values().map(Vec::len),
        body.clone().unwrap_or(0),
        bound.bound,
    );
    let log_head = owner.next_log_sequence;
    let outbox_head = owner.outbox.as_ref().map_or(
        0,
        super::outbox::ConversationOutbox::next_extension_sequence,
    );
    let (log_rows_behind, outbox_rows_behind) = locator.map_or((0, 0), |(covered, extension)| {
        (
            covered.saturating_add(1).min(log_head),
            extension.min(outbox_head),
        )
    });
    Ok(CheckpointCensus {
        conversation_id,
        nodes,
        history_nodes: history.node_count(),
        history_bytes: history.byte_total(),
        tokens: owner.tokens.len(),
        committed_admissions: owner.committed_admissions.len(),
        slots: owner.slots.len(),
        attach_provenance: owner
            .slots
            .values()
            .map(|slot| slot.attach_provenance.len())
            .sum(),
        retired: owner.retired.len(),
        locator,
        log_rows_behind,
        log_rows_ahead: log_head.saturating_sub(log_rows_behind),
        outbox_rows_behind,
        outbox_rows_ahead: outbox_head.saturating_sub(outbox_rows_behind),
        body,
        bound,
        batches,
    })
}

/// The root the latest published checkpoint names and the locator's
/// (covered log sequence, extension boundary), or the empty root and no
/// locator when a load would rebuild.
fn latest_root_and_locator(
    store: &Arc<dyn DurableStore>,
    config: &crate::config::types::ParticipantConfig,
    conversation_id: u64,
) -> Result<(CheckpointHistoryRoot, Option<(u64, u64)>), CensusError> {
    let loaded = block_on(
        load_checkpoint::<ConversationAuthority, ROOT_ENCODED_BYTES>(
            store.as_ref(),
            conversation_id,
            |positions, root| ConversationAuthorityRestoreContext::new(*positions, root, config),
            |locator, root| {
                CheckpointHistoryRoot::decode_canonical(root, locator.conversation_id())
                    .map(|_| ())
                    .map_err(
                        |source| super::restore_checkpoint::CheckpointError::LiveState {
                            source: Box::new(source),
                        },
                    )
            },
        ),
    )
    .map_err(|error| CensusError::Locator(error.to_string()))?
    .map_err(|error| CensusError::Locator(error.to_string()))?;
    Ok(match loaded.decision {
        CheckpointDecision::Restore(checkpoint) => {
            let root =
                CheckpointHistoryRoot::decode_canonical(checkpoint.history_root(), conversation_id)
                    .map_err(|error| CensusError::Locator(error.to_string()))?;
            let locator = loaded
                .observed_locator
                .as_deref()
                .map(|bytes| CheckpointLocator::decode_canonical(bytes, conversation_id))
                .transpose()
                .map_err(|error| CensusError::Locator(error.to_string()))?
                .map(|locator| (locator.covered_sequence(), locator.extension_boundary()));
            (root, locator)
        }
        CheckpointDecision::Rebuild(_) => (CheckpointHistoryRoot::empty(conversation_id), None),
    })
}

/// Every captured node by kind: (count, bytes), with every kind present.
fn count_nodes_by_kind(
    history: &super::restore_checkpoint_history_capture::PreparedCheckpointHistory,
) -> Result<BTreeMap<&'static str, (usize, usize)>, CensusError> {
    let mut nodes: BTreeMap<&'static str, (usize, usize)> = BTreeMap::new();
    for kind in [
        HistoryNodeKind::RetainedValue,
        HistoryNodeKind::RetiredHistory,
        HistoryNodeKind::IndexPage,
    ] {
        nodes.insert(kind.name(), (0, 0));
    }
    for (key, body) in history.nodes() {
        let node_root = CheckpointHistoryRoot::from_routed_node_key(key).ok_or_else(|| {
            CensusError::Node(format!("not a history node key: {}", key_name(key)))
        })?;
        let node = CheckpointHistoryNode::decode_canonical(body, node_root)
            .map_err(|error| CensusError::Node(error.to_string()))?;
        let entry = nodes.entry(node.kind().name()).or_insert((0, 0));
        entry.0 = entry.0.saturating_add(1);
        entry.1 = entry.1.saturating_add(body.len());
    }
    Ok(nodes)
}

/// The bytes the body and locator would add to the final batch, staged into
/// a scratch batch that is never committed; or why the owner cannot be
/// captured, in the checkpoint's own words.
fn checkpoint_body_bytes(
    store: &dyn DurableStore,
    owner: &ConversationAuthority,
    schema: super::restore_checkpoint::CheckpointSchema,
    root: CheckpointHistoryRoot,
    prefix: [u8; 32],
) -> Result<usize, String> {
    let extension = owner.outbox.as_ref().map_or(
        0,
        super::outbox::ConversationOutbox::next_extension_sequence,
    );
    let positions = owner
        .checkpoint_positions(schema, owner.next_log_sequence, extension)
        .map_err(|error| error.to_string())?;
    let route =
        super::restore_checkpoint_locator::checkpoint_conversation_route(owner.conversation_id);
    let mut scratch = OperationBatch::new(&route);
    block_on(scratch.stage_borrowed_checkpoint(
        store,
        positions,
        &root.encode_canonical(),
        owner,
        prefix,
    ))
    .map_err(|error| error.to_string())?
    .map_err(|error| error.to_string())?;
    Ok(scratch.value_bytes())
}