dsh-usage-badge 0.1.4 → 0.1.5

package/lib/sync.js

@@ -0,0 +1,1179 @@
+/**
+ * dsh-usage-badge — the shard format and the two directions of a sync.
+ *
+ * The cache's own store stays local; this module is the **export/import boundary**. It lives apart
+ * from the store because it needs two things the store must not know about:
+ *
+ *   - **a wire format.** One shard is one calendar month of one device's rows, as NDJSON compressed
+ *     with `zstd` (`node:zlib`, already used to read `.jsonl.zstd` logs). A month is the granularity
+ *     because a day would mean hundreds of files and a request each, while a year would mean
+ *     re-uploading ~27 MB whenever any single day inside it changed.
+ *   - **a manifest.** Each device publishes **its own** `index.json`, listing the shards it wrote
+ *     with a sha256 each. There is deliberately no shared manifest: a file two devices both wrote
+ *     would need locking, and this design has no locking.
+ *
+ * ## What a device must never publish
+ *
+ * Only rows with `source = local` are exported. A row that *arrived* from another device is stored so
+ * the panel can count it, but exporting it back would make this device claim another's history — and
+ * if that other device then deleted it, the copy here would resurrect it forever.
+ *
+ * `fallback` rows are not exported either: they are copies of a log version this device has already
+ * moved past, kept as local insurance rather than as history to publish.
+ *
+ * ## The import side
+ *
+ * A shard's rows carry their own `source` (the id of the device that wrote them) and their own
+ * stamp, and are handed to the store as they are. The store then applies the rule it applies to any
+ * weak copy: a row that is not strictly fuller than the one already held for that `(session, date,
+ * source, slot)` is dropped. That is what stops a machine whose logs were pruned from overwriting the
+ * fuller history another machine still has.
+ *
+ * @module dsh-usage-badge/sync
+ */
+
+import { createHash } from 'node:crypto'
+import { mkdirSync, readFileSync, renameSync, statSync, unlinkSync, writeFileSync } from 'node:fs'
+import { hostname } from 'node:os'
+import { join } from 'node:path'
+import { zstdCompressSync, zstdDecompressSync } from 'node:zlib'
+
+import { CURRENT, FALLBACK, LOCAL_SOURCE, rowKind, rowSource } from './buckets.js'
+import { WebDavClient } from './webdav.js'
+
+/** The shard format this build writes. A shard of another version is refused, not guessed at. */
+export const SHARD_VERSION = 1
+
+/** The manifest format this build writes. */
+export const MANIFEST_VERSION = 1
+
+/** The file extension a shard carries; the file name is its month. */
+export const SHARD_EXTENSION = '.ndjson.zst'
+
+/** Everything this plugin owns on the server lives under this prefix. */
+export const REMOTE_PREFIX = 'usage-badge/'
+
+/**
+ * A file beside each device's `index.json`.
+ *
+ * Apache's `PROPFIND` lists the collections under `usage-badge/`, which is how a device discovers
+ * the others — but a collection holding only `index.json` may list without its `index.json` (the
+ * server is free to omit a member), and discovery is the one thing that must not depend on which of
+ * the two a reader happens to see. So each device directory also gets this small marker, created
+ * right after the collection and never rewritten: if a reader can see the directory at all, it sees
+ * the marker, and a directory without it is treated as half-created rather than as a device.
+ */
+export const DEVICE_MARKER = '.device'
+
+const sha256 = (value) => createHash('sha256').update(value).digest('hex')
+
+/** The month a day key belongs to: `2026-09-14` → `2026-09`. */
+export const monthOf = (date) => String(date ?? '').slice(0, 7)
+
+/** One shard's path, given a device id and a month. */
+export const shardPath = (device, key) => `${REMOTE_PREFIX}${device}/${key}${SHARD_EXTENSION}`
+
+/** One device's manifest path. */
+export const manifestPath = (device) => `${REMOTE_PREFIX}${device}/index.json`
+
+/**
+ * A device id as it may appear in a path: no slashes, no traversal, nothing exotic.
+ *
+ * `local` is refused on purpose. A device's own rows are stored with `source = 'local'` until they are
+ * exported, and `chunkRows` decides ownership by that exact value — so a device *named* `local` would
+ * make every other device mistake its rows for their own and re-publish them under their own name, after
+ * which this device's own fold could delete them. The name is a path segment and an ownership tag; it
+ * cannot be the tag that means "not from the sync".
+ */
+export const safeDeviceId = (device) => {
+  const id = String(device ?? '').trim()
+  if (id === LOCAL_SOURCE) throw new Error(`\`${LOCAL_SOURCE}\` is reserved for this machine's own rows and cannot be a device id`)
+  if (!/^[A-Za-z0-9._-]{1,64}$/.test(id) || id === '.' || id === '..') {
+    throw new Error(`a device id must be 1-64 of [A-Za-z0-9._-], got ${JSON.stringify(device)}`)
+  }
+  return id
+}
+
+/** A shard key as it may appear in a path: exactly one `YYYY-MM`. */
+export const safeShardKey = (key) => {
+  const value = String(key ?? '').trim()
+  if (!/^\d{4}-\d{2}$/.test(value)) throw new Error(`a shard key must be YYYY-MM, got ${JSON.stringify(key)}`)
+  return value
+}
+
+/**
+ * One device's local rows grouped into month shards.
+ *
+ * Only `current` rows this device folded are exported (`owns` selects them; it is a parameter so a
+ * test can exercise the filter without pretending to be another device). Each row is copied out with
+ * its `sessionId` attached, because a shard has to be able to rebuild a session record on the other
+ * side.
+ *
+ * @param {object[]} records - records as `store.scan()` returns them, each with a `rows` array.
+ * @param {{owns?:string}} [options]
+ * @returns {Map<string, object[]>} shard key → rows, sorted by date then session.
+ */
+export function chunkRows(records, { owns = LOCAL_SOURCE } = {}) {
+  const shards = new Map()
+  for (const record of records ?? []) {
+    for (const row of record?.rows ?? []) {
+      if (rowSource(row) !== owns) continue
+      if (rowKind(row) === FALLBACK) continue
+      const key = monthOf(row.date)
+      if (!/^\d{4}-\d{2}$/.test(key)) continue
+      let list = shards.get(key)
+      if (!list) {
+        list = []
+        shards.set(key, list)
+      }
+      list.push({
+        date: row.date,
+        source: rowSource(row),
+        kind: CURRENT,
+        seq: Number(row.seq) || 0,
+        fileSize: Number(row.fileSize) || 0,
+        fileMtimeMs: Number(row.fileMtimeMs) || 0,
+        at: Number(row.at) || 0,
+        sessionId: String(record.sessionId ?? ''),
+        path: String(record.path ?? ''),
+        routes: row.routes ?? {},
+      })
+    }
+  }
+  for (const list of shards.values()) {
+    list.sort((a, b) => (a.date === b.date ? (a.sessionId < b.sessionId ? -1 : a.sessionId > b.sessionId ? 1 : 0) : a.date < b.date ? -1 : 1))
+  }
+  return shards
+}
+
+/**
+ * One shard's bytes: a header line, then one JSON row per line, compressed as a whole.
+ *
+ * Two properties are load-bearing:
+ *
+ *   - **the bytes are deterministic.** Nothing about *when* the shard was written goes into them, so
+ *     the upload plan can decide "this month did not change" by comparing a content hash. If the
+ *     timestamp were part of the payload, every re-plan would look like a new revision and every sync
+ *     would re-upload everything. When a month was written is manifest metadata, not shard data.
+ *   - **a row's `source` is rewritten to the publishing device.** A row stored here says
+ *     `source = local`, meaning "this device folded it"; on the receiving side the same bytes must
+ *     mean "device `<id>` published it". Carrying `local` across would make the receiver treat
+ *     another machine's history as **its own** — it would then publish it onward under its own name,
+ *     and a later fold here could delete rows it does not own.
+ */
+export function encodeShard(rows, { device }) {
+  const id = safeDeviceId(device)
+  const header = JSON.stringify({ v: SHARD_VERSION, device: id })
+  const body = [
+    header,
+    ...rows.map((row) => JSON.stringify({ ...row, source: rowSource(row) === LOCAL_SOURCE ? id : rowSource(row) })),
+  ].join('\n')
+  const compressed = zstdCompressSync(Buffer.from(body, 'utf8'))
+  return { body: compressed, sha256: sha256(compressed), rows: rows.length }
+}
+
+/**
+ * Decode one shard.
+ *
+ * A shard whose version is not this build's is refused **by name** rather than parsed leniently: a
+ * half-understood shard would be merged as if it were understood, and the result would be wrong
+ * numbers that look right.
+ *
+ * @param {Buffer|Uint8Array} buffer
+ * @returns {{version:number, device:string, exportedAt:number, rows:object[]}}
+ */
+export function decodeShard(buffer) {
+  let text
+  try {
+    text = zstdDecompressSync(buffer).toString('utf8')
+  } catch (error) {
+    throw new Error(`a shard could not be decompressed: ${error?.message ?? error}`)
+  }
+  const lines = text.split('\n')
+  if (!lines[0]) throw new Error('a shard is empty')
+  let header
+  try {
+    header = JSON.parse(lines[0])
+  } catch (error) {
+    throw new Error(`a shard header is not JSON: ${error?.message ?? error}`)
+  }
+  if (header?.v !== SHARD_VERSION) {
+    throw new Error(`a shard is version ${String(header?.v)}, but this build writes version ${SHARD_VERSION}`)
+  }
+  const rows = []
+  let malformed = 0
+  for (let i = 1; i < lines.length; i++) {
+    if (!lines[i]) continue
+    let row
+    try {
+      row = JSON.parse(lines[i])
+    } catch {
+      malformed++
+      continue
+    }
+    if (typeof row?.date !== 'string' || !/^\d{4}-\d{2}-\d{2}$/.test(row.date)) {
+      malformed++
+      continue
+    }
+    rows.push({
+      date: row.date,
+      source: String(row.source ?? header.device ?? LOCAL_SOURCE),
+      kind: row.kind === FALLBACK ? FALLBACK : CURRENT,
+      seq: Number(row.seq) || 0,
+      fileSize: Number(row.fileSize) || 0,
+      fileMtimeMs: Number(row.fileMtimeMs) || 0,
+      at: Number(row.at) || 0,
+      sessionId: String(row.sessionId ?? ''),
+      routes: row.routes && typeof row.routes === 'object' ? row.routes : {},
+    })
+  }
+  return { version: header.v, device: String(header.device ?? ''), exportedAt: Number(header.at) || 0, rows, malformed }
+}
+
+/** One chunk's manifest entry, describing the bytes that were actually written. */
+export function chunkEntry({ device, key, body, rows, at }) {
+  const dates = rows.map((row) => row.date).sort()
+  return {
+    key,
+    path: shardPath(device, key),
+    from: dates[0] ?? '',
+    to: dates[dates.length - 1] ?? '',
+    bytes: body.length,
+    sha256: sha256(body),
+    rows: rows.length,
+    at,
+  }
+}
+
+/** A device's manifest: what this device has published. */
+export function buildManifest({ device, label = '', chunks = {}, at = Date.now(), instanceId = '' }) {
+  // `instanceId` is written so that the *other* installations can tell whether the name in this
+  // directory is theirs. Two machines that both call themselves `office` would otherwise each overwrite
+  // the other's month — the shard path is the same — and the rows would be gone with no warning at all.
+  return { v: MANIFEST_VERSION, device, label, generatedAt: at, instanceId, chunks }
+}
+
+/** Parse a manifest, refusing a version this build does not write and a device id it cannot use. */
+export function parseManifest(text) {
+  let raw
+  try {
+    raw = typeof text === 'string' || Buffer.isBuffer(text) ? JSON.parse(text.toString('utf8')) : text
+  } catch (error) {
+    throw new Error(`the manifest is not JSON: ${error?.message ?? error}`)
+  }
+  if (raw?.v !== MANIFEST_VERSION) {
+    throw new Error(`the manifest is version ${String(raw?.v)}, but this build writes version ${MANIFEST_VERSION}`)
+  }
+  const device = safeDeviceId(raw.device)
+  const chunks = {}
+  for (const [key, entry] of Object.entries(raw.chunks ?? {})) {
+    if (!entry || typeof entry !== 'object' || typeof entry.sha256 !== 'string') continue
+    chunks[key] = {
+      key,
+      path: String(entry.path ?? shardPath(device, key)),
+      from: String(entry.from ?? ''),
+      to: String(entry.to ?? ''),
+      bytes: Number(entry.bytes) || 0,
+      sha256: entry.sha256,
+      rows: Number(entry.rows) || 0,
+      at: Number(entry.at) || 0,
+    }
+  }
+  return { version: raw.v, device, label: String(raw.label ?? ''), generatedAt: Number(raw.generatedAt) || 0, instanceId: String(raw.instanceId ?? ''), chunks }
+}
+
+/**
+ * Decide what to upload: the shards whose bytes differ from what this device last published.
+ *
+ * The comparison is by **content hash**, not by date or mtime: the sync is incremental and has to
+ * skip a month that has not changed, and a re-fold that reproduces identical bytes must not become
+ * an upload just because it happened.
+ *
+ * @param {Map<string, object[]>} shards - as {@link chunkRows} returns.
+ * @param {object} [previous] - the manifest this device published last time, or `{}`.
+ * @returns {{pending: Array<{key:string, body:Buffer, entry:object, rows:number}>, manifest:object, unchanged:number}}
+ */
+export function planUpload(shards, previous = {}, { device, label = '', at = Date.now(), instanceId = '' } = {}) {
+  const safeDevice = safeDeviceId(device)
+  const chunks = { ...(previous?.chunks ?? {}) }
+  const pending = []
+  let unchanged = 0
+  for (const [key, rows] of [...(shards ?? new Map())].sort(([a], [b]) => (a < b ? -1 : 1))) {
+    const { body } = encodeShard(rows, { device: safeDevice })
+    const hash = sha256(body)
+    const known = chunks[key]
+    if (known?.sha256 === hash) {
+      // Left exactly as published, timestamp included: re-stamping it would say the month changed.
+      unchanged++
+      continue
+    }
+    const entry = chunkEntry({ device: safeDevice, key, body, rows, at })
+    chunks[key] = entry
+    pending.push({ key, body, entry, rows: rows.length })
+  }
+  return { pending, unchanged, manifest: buildManifest({ device: safeDevice, label, chunks, at, instanceId }) }
+}
+
+/**
+ * The shards a reader should fetch.
+ *
+ * `windowYears` bounds the **first** pull: a decade of history is real, but downloading all of it
+ * before the panel shows anything would be the worst first impression this feature could make. The
+ * bound applies to fetching, never to keeping — shards already held stay held, and each device's
+ * manifest still lists everything it has.
+ *
+ * @param {{manifests:object[], held?:object, now?:number, windowYears?:number, except?:string}} options
+ *   `held` maps `<device>/<key>` to the sha256 this reader already applied. `except` is the reader's
+ *   own device id: its own manifest lists what it published, which is already in its store.
+ */
+export function planDownload({ manifests, held = {}, now = Date.now(), windowYears = 5, except = '' }) {
+  const cutoff = `${new Date(now).getFullYear() - Math.max(0, windowYears) + 1}-01`
+  const wanted = []
+  for (const manifest of manifests ?? []) {
+    if (except && manifest?.device === except) continue
+    for (const [key, entry] of Object.entries(manifest?.chunks ?? {})) {
+      if (!entry?.path || !entry?.sha256) continue
+      if (windowYears > 0 && key < cutoff) continue
+      if (held[`${manifest.device}/${key}`] === entry.sha256) continue
+      wanted.push({ device: manifest.device, key, entry })
+    }
+  }
+  wanted.sort((a, b) => (a.key === b.key ? (a.device < b.device ? -1 : 1) : a.key < b.key ? -1 : 1))
+  return wanted
+}
+
+/**
+ * Turn decoded shard rows into the records the store's import takes.
+ *
+ * Grouped by `(source, sessionId)` because that is the store's write unit, and every row keeps the
+ * stamp it was written with: the receiving store compares those stamps to decide whether the
+ * incoming copy is worth storing at all. The cursor fields on the record are the best stamp among
+ * its rows, so a caller that only needs "the newest log version this device knows" can read them.
+ */
+export function rowsToRecords(rows) {
+  const bySource = new Map()
+  for (const row of rows ?? []) {
+    if (!row?.sessionId) continue
+    const key = `${row.source}\u0000${row.sessionId}`
+    let record = bySource.get(key)
+    if (!record) {
+      record = { sessionId: row.sessionId, path: '', source: row.source, rows: [] }
+      bySource.set(key, record)
+    }
+    record.rows.push(row)
+  }
+  for (const record of bySource.values()) {
+    const best = record.rows.reduce((top, row) => (top === null || row.seq > top.seq || (row.seq === top.seq && row.fileSize > top.fileSize) ? row : top), null)
+    record.seq = best?.seq ?? -1
+    record.fileSize = best?.fileSize ?? 0
+    record.fileMtimeMs = best?.fileMtimeMs ?? 0
+  }
+  return [...bySource.values()]
+}
+
+/** A short, display-safe prefix of a hash, for logs and the panel. */
+export const shortHash = (hash) => String(hash ?? '').slice(0, 12)
+
+// ── the price table: one shared file, and one copy per device ────────────────
+
+/** The shared price table, which every device reads and the last writer wins. */
+export const PRICING_FILE = `${REMOTE_PREFIX}pricing.json`
+
+/** The price table this device wrote, kept so a conflict can always be resolved by hand. */
+export const pricingDevicePath = (device) => `${REMOTE_PREFIX}${safeDeviceId(device)}/pricing-${safeDeviceId(device)}.json`
+
+/** Byte-exact comparison, since the plan is about *whether* the file changed, not how. */
+const sameBytes = (a, b) => (a === null || b === null ? a === b : Buffer.compare(Buffer.from(a), Buffer.from(b)) === 0)
+
+/**
+ * Decide what to do about the shared price table.
+ *
+ * The price table is the one thing here that a human writes by hand, so it cannot be merged by a
+ * rule — two people editing the same table is a decision only a person can make. What this can do is
+ * tell the three cases apart without ever losing a version:
+ *
+ *   - **nothing changed** (`remote === baseline`) → do nothing, whichever device wrote last;
+ *   - **only the remote moved** → take it;
+ *   - **only we moved** → publish ours;
+ *   - **both moved, differently** → a conflict: the remote is kept beside ours as
+ *     `.conflict-pricing-<device>-<at>.json` before either wins, and the later `at` wins. Losing a
+ *     price row silently is exactly the outcome this whole design is built to avoid.
+ *
+ * @param {{localBytes?:Buffer|null, remoteBytes?:Buffer|null, baseline?:string|null,
+ *   localAt?:number, remoteAt?:number, device?:string, at?:number}} options
+ * @returns {{action:'none'|'pull'|'push'|'conflict', winner:'local'|'remote'|null, conflictFile:string|null}}
+ */
+export function planPricing({ localBytes = null, remoteBytes = null, baseline = null, localAt = 0, remoteAt = 0, device = '', at = 0 } = {}) {
+  const localHash = localBytes === null ? null : sha256(localBytes)
+  const remoteHash = remoteBytes === null ? null : sha256(remoteBytes)
+  // A missing local file means "we have not published anything yet", not "we changed something":
+  // otherwise a fresh install calls the server's very first table a conflict. An *empty* file is a
+  // different thing — it exists, and it is not what the server has.
+  const localChanged = localHash === null ? baseline !== null : localHash !== baseline
+  const remoteChanged = remoteHash !== null && remoteHash !== baseline
+
+  // Nothing to take and nothing to publish: either both sides agree, or the file we published is
+  // still the one on the server.
+  if (!localChanged && !remoteChanged) return { action: 'none', winner: null, conflictFile: null }
+  if (!localChanged && remoteChanged) return { action: 'pull', winner: 'remote', conflictFile: null }
+  if (localChanged && !remoteChanged) return { action: 'push', winner: 'local', conflictFile: null }
+  if (sameBytes(localBytes, remoteBytes)) return { action: 'none', winner: 'remote', conflictFile: null }
+
+  // Both sides changed, and they disagree. Keep the one that would have been lost.
+  const winner = remoteAt >= localAt ? 'remote' : 'local'
+  const loserPath = remoteAt >= localAt ? localBytes : remoteBytes
+  const conflictFile = device && loserPath !== null ? `.conflict-pricing-${safeDeviceId(device)}-${at || Date.now()}.json` : null
+  return { action: 'conflict', winner, conflictFile }
+}
+
+// ── the data directory is owned by one process at a time ─────────────────────
+
+/** The lock file a host holds while it may write this data directory. */
+export const LOCK_FILE = 'host.lock'
+
+/**
+ * Whether the process that wrote a lock file is still alive.
+ *
+ * `process.kill(pid, 0)` is the portable probe: it sends no signal, and its failure modes separate
+ * "no such process" (ESRCH — the lock is stale, take it) from "not allowed to signal it" (EPERM —
+ * there *is* a process, and on Windows that is the normal answer for a process owned by another
+ * session). Anything else is treated as "cannot tell", which means **not** taking the lock: a
+ * second writer is worse than a read-only instance.
+ */
+export function processLooksAlive(pid) {
+  const value = Number(pid)
+  if (!Number.isInteger(value) || value <= 0) return false
+  if (value === process.pid) return true
+  try {
+    process.kill(value, 0)
+    return true
+  } catch (error) {
+    const code = error?.code
+    if (code === 'ESRCH') return false
+    if (code === 'EPERM') return true
+    return true
+  }
+}
+
+/**
+ * Write the lock, and read back what a previous holder left.
+ *
+ * `process.pid` alone is not enough: pid reuse after a crash would make a live-looking lock out of a
+ * dead process. Comparing the recorded process start time catches that, and the comparison is
+ * deliberately lax — a mismatch in the *started* field only downgrades the lock to "someone else
+ * holds it", which is the safe direction.
+ */
+export function readLock(dir) {
+  const path = join(dir, LOCK_FILE)
+  try {
+    const raw = JSON.parse(readFileSync(path, 'utf8'))
+    return {
+      pid: Number(raw?.pid) || 0,
+      startedAt: Number(raw?.startedAt) || 0,
+      instanceId: String(raw?.instanceId ?? ''),
+      hostname: String(raw?.hostname ?? ''),
+      at: Number(raw?.at) || 0,
+    }
+  } catch {
+    return null
+  }
+}
+
+/** Whether a lock belongs to a process that is no longer running. */
+export const lockIsStale = (lock) => lock === null || !processLooksAlive(lock.pid)
+
+/** Just this process's own start, in epoch ms. */
+const PROCESS_STARTED_AT = Date.now() - Math.round(process.uptime() * 1000)
+
+/**
+ * Try to take ownership of a data directory.
+ *
+ * Ownership decides who may *write*: the cache, the fold, and the sync all need one writer, and the
+ * sync needs it most — two processes sharing a `deviceId` would each believe the same remote shard is
+ * theirs to overwrite. The caller that fails here is not broken and does not refuse to start: it
+ * uses the cache read-only and skips syncing, which is what the design promises for this case.
+ *
+ * @returns {{acquired:boolean, owner?:object, reason?:string}}
+ */
+export function acquireLock(dir, { instanceId, log = () => {} } = {}) {
+  try {
+    // Creating the data directory is part of taking the lock, and it can fail for reasons that have
+    // nothing to do with a lock: a read-only home, or a `storages` that turned out to be a file. Those are
+    // reported like any other "cannot write here" — an exception escaping here reached the panel as a 500
+    // with no explanation.
+    mkdirSync(dir, { recursive: true })
+  } catch (error) {
+    return { acquired: false, owner: null, reason: `the data directory cannot be created: ${error?.message ?? error}` }
+  }
+  const path = join(dir, LOCK_FILE)
+  const existing = readLock(dir)
+  if (existing && !lockIsStale(existing)) {
+    return {
+      acquired: false,
+      owner: existing,
+      reason: `this data directory is open by another DSH instance (pid ${existing.pid} on ${existing.hostname || 'this host'})`,
+    }
+  }
+  if (existing) log(`taking over a stale ${LOCK_FILE} left by pid ${existing.pid}`)
+  const body = { pid: process.pid, startedAt: PROCESS_STARTED_AT, instanceId, hostname: hostname(), at: Date.now() }
+  try {
+    writeFileSync(path, JSON.stringify(body))
+  } catch (error) {
+    return { acquired: false, owner: null, reason: `could not write ${LOCK_FILE}: ${error?.message ?? error}` }
+  }
+  return { acquired: true, owner: body }
+}
+
+/** Release the lock, but only if this process is the one holding it. */
+export function releaseLock(dir) {
+  const existing = readLock(dir)
+  if (!existing || existing.pid !== process.pid) return false
+  try {
+    unlinkSync(join(dir, LOCK_FILE))
+    return true
+  } catch {
+    return false
+  }
+}
+
+// ── sync.json: what the user configured, and what the last run learned ───────
+
+/** The file holding the sync configuration and the last run's state. */
+export const SYNC_FILE = 'sync.json'
+
+/** The defaults a sync configuration starts from. */
+export const defaultSyncConfig = () => ({
+  enabled: false,
+  url: '',
+  username: '',
+  realm: '',
+  ha1: '',
+  allowInsecureTls: false,
+  deviceId: '',
+  deviceLabel: hostname(),
+  instanceId: '',
+  windowYears: 5,
+})
+
+/** Read `sync.json`, or a default document when it is missing or unreadable. */
+export function readSyncDocument(dir, log = () => {}) {
+  const path = join(dir, SYNC_FILE)
+  try {
+    const raw = JSON.parse(readFileSync(path, 'utf8'))
+    if (!raw || typeof raw !== 'object') return { ...defaultSyncConfig(), state: {}, known: {} }
+    return { ...defaultSyncConfig(), ...raw }
+  } catch (error) {
+    if (error?.code !== 'ENOENT') log(`${SYNC_FILE} could not be read (${error?.message ?? error}); using the defaults`)
+    return { ...defaultSyncConfig(), state: {}, known: {} }
+  }
+}
+
+/** Write `sync.json` atomically, so a crash cannot leave a half-written configuration. */
+export function writeSyncDocument(dir, document) {
+  mkdirSync(dir, { recursive: true })
+  const path = join(dir, SYNC_FILE)
+  const temp = `${path}.tmp-${process.pid}`
+  writeFileSync(temp, JSON.stringify(document, null, 2))
+  try {
+    renameSync(temp, path)
+  } catch {
+    try {
+      unlinkSync(path)
+    } catch {
+      // absent is fine
+    }
+    renameSync(temp, path)
+  }
+  return path
+}
+
+// ── the orchestration: discover, pull, push ──────────────────────────────────
+
+/**
+ * One device's view of the whole remote, kept between runs.
+ *
+ * `known.shards` maps `<device>/<key>` to the sha256 already applied, which is what makes a second
+ * run cheap: it is also the record that proves a shard was read rather than merely listed. It is
+ * rebuilt from the remote on every run, so losing it costs a re-download and nothing else.
+ */
+export const defaultSyncState = () => ({
+  lastSyncAt: 0,
+  lastResult: 'never',
+  message: '',
+  held: {},
+  published: {},
+})
+
+/**
+ * The remote orchestration.
+ *
+ * It owns the order of operations, because the order is the correctness argument: take the lock,
+ * **pull before pushing** (so a shard this device is about to overwrite is already accounted for),
+ * import into the store, publish only what changed, and write the manifest last — a manifest that
+ * names a shard which is not there yet would send every other device after a 404.
+ */
+export class SyncEngine {
+  /**
+   * @param {{dataDir:string, store?:object, log?:Function, client?:object, now?:Function,
+   *   config?:object, state?:object, instanceId?:string}} options
+   *   `store` may be null: a sync without a store still publishes and reads manifests, which is what
+   *   the first step's live check exercises.
+   */
+  constructor({ dataDir, store = null, log = () => {}, client = null, now = Date.now, config = null, state = null, instanceId = '' }) {
+    this.dataDir = dataDir
+    this.store = store
+    this.log = log
+    this.now = now
+    const document = config ? { ...defaultSyncConfig(), ...config } : readSyncDocument(dataDir, log)
+    this.config = document
+    this.state = state ?? { ...defaultSyncState(), ...(document.state ?? {}) }
+    this.instanceId = instanceId || document.instanceId || `${hostname()}-${process.pid}`
+    this.client = client
+    this.lock = null
+    this.ownsLock = false
+  }
+
+  /**
+   * Give up ownership of the data directory.
+   *
+   * Called on unload by the host half, and by anything embedding this engine for one run. It only
+   * removes a lock this engine took: a lock handed in by the caller belongs to the caller.
+   */
+  release() {
+    if (!this.ownsLock) return false
+    const released = releaseLock(this.dataDir)
+    this.ownsLock = false
+    return released
+  }
+
+  get deviceId() {
+    return this.config.deviceId
+  }
+
+  /** Persist the configuration (and the latest state) to `sync.json`. */
+  save({ config = {}, state = {} } = {}) {
+    this.config = { ...this.config, ...config, instanceId: this.instanceId }
+    this.state = { ...this.state, ...state }
+    writeSyncDocument(this.dataDir, { ...this.config, state: this.state })
+    return this.config
+  }
+
+  /** A WebDAV client for the configured server, or a clear failure. */
+  connect() {
+    if (this.client) return this.client
+    const { url, username, ha1 } = this.config
+    if (!url) throw new Error('no WebDAV URL is configured')
+    if (!username) throw new Error('no WebDAV username is configured')
+    if (!ha1) throw new Error('no WebDAV credential is configured')
+    return new WebDavClient({
+      baseUrl: url,
+      username,
+      // The stored secret is the digest HA1 — `MD5(user:realm:pass)`, exactly what the server keeps in
+      // its own password file. Passing it as `passwordHash` means the plaintext never has to be
+      // written down here, and the response is still the one the server expects.
+      passwordHash: ha1,
+      allowInsecureTls: Boolean(this.config.allowInsecureTls),
+    })
+  }
+
+  /**
+   * The device directories under `usage-badge/`.
+   *
+   * It lists the `usage-badge/` collection itself rather than the root: the root's children include
+   * the collection's own name, and treating that as a device would send this sync after
+   * `usage-badge/usage-badge/index.json`. A collection that does not exist yet means an empty list —
+   * the first device to sync creates it.
+   */
+  async discover(client) {
+    let entries
+    try {
+      entries = await client.propfind(REMOTE_PREFIX)
+    } catch (error) {
+      // `PROPFIND` answers 404 for a path that has no collection yet; that is "nothing here", not a
+      // failure to talk to the server.
+      if (/answered 404/.test(String(error?.message ?? ''))) return []
+      throw new Error(`could not list ${REMOTE_PREFIX}: ${error?.message ?? error}`)
+    }
+    const devices = []
+    for (const entry of entries) {
+      if (!entry.collection) continue
+      const name = String(entry.path ?? entry.name).replace(/\/+$/, '').split('/').pop() ?? ''
+      if (!/^[A-Za-z0-9._-]{1,64}$/.test(name) || name === '.' || name === '..') continue
+      devices.push(name)
+    }
+    return devices
+  }
+
+  /** Read one device's manifest, or null when it has none yet. */
+  async readManifest(client, device) {
+    const body = await client.get(manifestPath(device))
+    if (body === null) return null
+    try {
+      return parseManifest(body)
+    } catch (error) {
+      // A manifest that cannot be parsed is reported and skipped: refusing the whole sync over one
+      // unreadable device would stop the devices that are fine.
+      this.log(`manifest of ${device} could not be read: ${error?.message ?? error}`)
+      return null
+    }
+  }
+
+  /** Everything the remote knows, as manifests. */
+  async readManifests(client) {
+    const devices = await this.discover(client)
+    const manifests = []
+    for (const device of devices) {
+      const manifest = await this.readManifest(client, device)
+      if (manifest) manifests.push(manifest)
+    }
+    return manifests
+  }
+
+  /**
+   * Pull: fetch the shards this device does not already hold, and store their rows.
+   *
+   * @returns {{fetched:number, rows:number, failed:Array<{path:string, error:string}>}}
+   */
+  async pull(client, manifests) {
+    const wanted = planDownload({
+      manifests,
+      held: this.state.held ?? {},
+      now: this.now(),
+      windowYears: Number(this.config.windowYears) || 0,
+      // This device's own manifest is not something to fetch: it lists the shards this device
+      // published, and its rows are already in this store under their local source. Reading them back
+      // would re-attribute them to this device's id and import duplicates of what is already here.
+      //
+      // Unless that directory is **someone else's**, which is the name clash: the shards in it are then a
+      // history this machine does not have, and it is exactly the history a reinstalled machine came
+      // back for. Importing them is safe — the merge key is the session, not the device — and the
+      // alternative is an empty panel until the user resolves a naming question they may not understand.
+      //
+      // The decision comes from **this run's** finding, never from the persisted `state.collision`: a
+      // refusal can stay recorded for months, and reading one's own directory because of a stale note
+      // means re-fetching one's own shards — including the half-written one that caused the refusal.
+      except: this.pullOwnDirectory ? '' : this.deviceId,
+    })
+    let rows = 0
+    let fetched = 0
+    const failed = []
+    for (const item of wanted) {
+      let body
+      try {
+        body = await client.get(item.entry.path)
+      } catch (error) {
+        failed.push({ path: item.entry.path, error: String(error?.message ?? error) })
+        continue
+      }
+      if (body === null) {
+        failed.push({ path: item.entry.path, error: 'the manifest lists a shard the server does not have' })
+        continue
+      }
+      let decoded
+      try {
+        decoded = decodeShard(body)
+      } catch (error) {
+        failed.push({ path: item.entry.path, error: String(error?.message ?? error) })
+        continue
+      }
+      // A shard whose hash does not match its manifest is not stored: the store keeps whatever is
+      // newer, so importing a truncated month could not be undone by re-importing the good one.
+      if (item.entry.sha256 && item.entry.sha256 !== sha256(body)) {
+        failed.push({ path: item.entry.path, error: 'the shard does not match the hash its manifest advertises' })
+        continue
+      }
+      if (this.store) {
+        this.store.intakeRows(decoded.rows)
+        // `held` is only advanced once the rows are in the store: if the import throws, the shard
+        // must be fetched again rather than remembered as applied.
+        this.state.held = { ...(this.state.held ?? {}), [`${item.device}/${item.key}`]: item.entry.sha256 }
+      }
+      rows += decoded.rows.length
+      fetched++
+    }
+    return { fetched, rows, failed }
+  }
+
+  /**
+   * Push: publish this device's changed months, then its manifest.
+   *
+   * The manifest is written **last** on purpose. A manifest naming a shard that has not landed yet
+   * would send every other device after a 404 on their next run.
+   */
+  async push(client) {
+    if (!this.store) return { uploaded: 0, unchanged: 0, bytes: 0, shards: [], refused: null }
+    // Whose name is this? The shard path is `usage-badge/<device>/<month>`, so a second installation
+    // using the same device name does not merge with the first — it **overwrites** its months, and the
+    // rows are gone. Nothing about the bytes reveals that, so it has to be caught here, where the answer
+    // is "do not write" rather than "write and hope".
+    //
+    // The owner is recorded in the directory's own manifest. A directory with no recorded owner is one
+    // published by a build that did not write the field (or the first publish of anything), and adopting
+    // it is what makes an upgrade work; an owner that is a *different* installation is a real clash.
+    const remote = (await this.readManifests(client)).find((manifest) => manifest.device === this.deviceId)
+    if (remote && remote.instanceId && remote.instanceId !== this.instanceId) {
+      this.state.collision = { device: this.deviceId, owner: remote.instanceId, at: this.now() }
+      this.log(
+        `not publishing: usage-badge/${this.deviceId}/ was published by another installation ` +
+          `(${remote.instanceId}, this one is ${this.instanceId}). Pulling is unaffected — everything on the ` +
+          `server is merged into this device's history. To publish from this machine, give it a device name ` +
+          `that is not already there; "usage-badge/${this.deviceId}/" only decides where new months are written.`,
+      )
+      return { uploaded: 0, unchanged: 0, bytes: 0, shards: [], refused: 'collision' }
+    }
+    if (remote && !remote.instanceId && !this.instanceId) {
+      // No owner on either side: there is no way to tell the two installations apart, and the whole point
+      // of the check is not to guess.
+      this.state.collision = { device: this.deviceId, owner: null, at: this.now() }
+      this.log(`refusing to publish: usage-badge/${this.deviceId}/ exists and neither side has an identity to compare`)
+      return { uploaded: 0, unchanged: 0, bytes: 0, shards: [], refused: 'unidentified' }
+    }
+    delete this.state.collision
+
+    const rows = chunkRows(this.store.scan())
+    // `state.published` is the manifest's **chunks map**, so it is wrapped back into a manifest shape
+    // here. Passing the map itself would leave `previous.chunks` empty, and every run would re-upload
+    // every month while looking perfectly healthy.
+    const { pending, unchanged, manifest } = planUpload(rows, { chunks: this.state.published ?? {} }, {
+      device: this.deviceId,
+      label: this.config.deviceLabel ?? '',
+      at: this.now(),
+      instanceId: this.instanceId,
+    })
+
+    // Is the directory still saying what we last put there? An identity check cannot answer this: a
+    // directory copied to another machine carries `sync.json` with it, `instanceId` included, so that
+    // machine passes every owner check and then `MOVE`s its months over this one's. The **bytes** answer
+    // it — a month we recorded at one hash and the server now holds at another was written by something
+    // that was not us, and we have nothing of our own to put there.
+    const diverged = Object.entries(this.state.published ?? {}).find(([key, known]) => {
+      const remoteEntry = remote?.chunks?.[key]
+      if (!remoteEntry || remoteEntry.sha256 === known.sha256) return false
+      return !pending.some((item) => item.key === key)
+    })
+    if (diverged) {
+      const [key] = diverged
+      this.state.collision = { device: this.deviceId, owner: remote?.instanceId || 'unknown', kind: 'diverged', month: key, at: this.now() }
+      this.log(
+        `not publishing: usage-badge/${this.deviceId}/${key}.ndjson.zst is not the shard this device last published, ` +
+          `and this device has no change of its own for that month — another installation is writing into this directory`,
+      )
+      return { uploaded: 0, unchanged: 0, bytes: 0, shards: [], refused: 'diverged' }
+    }
+
+    // A month this device published and no longer has any rows for: a session folder was deleted, or a
+    // log version that no longer covers that day superseded the rows. The manifest is republished from
+    // what is actually held, so **new** devices must stop being told about it — otherwise they import
+    // usage its own machine no longer has, which is precisely the "the numbers came back" surprise. The
+    // bytes stay on the server; only the claim that this device still holds them is withdrawn.
+    const retired = Object.keys(this.state.published ?? {}).filter((key) => !rows.has(key))
+
+    // The collection comes **before** anything is written into it. `mod_dav_fs` refuses a write whose
+    // parent does not exist (403/409, not the 404 an HTTP client might expect), so creating it on the
+    // way past would fail the first publish of a device — and mkcol is idempotent anyway, so it costs
+    // nothing to do it here.
+    await client.mkcol(`${REMOTE_PREFIX}${this.deviceId}/`)
+    let bytes = 0
+    const shards = []
+    for (const item of pending) {
+      // Upload under a temporary name and rename: a crash mid-upload leaves a name nobody reads,
+      // instead of a shard that looks complete and is half a month.
+      const temp = `${item.entry.path}.upload-${process.pid}-${Math.random().toString(16).slice(2, 8)}`
+      await client.put(temp, item.body, { contentType: 'application/octet-stream' })
+      try {
+        await client.move(temp, item.entry.path, { overwrite: true })
+      } catch (error) {
+        // Leave no debris behind if the rename fails; a leftover temp file is harmless but untidy.
+        await client.delete(temp).catch(() => {})
+        throw error
+      }
+      bytes += item.body.length
+      shards.push({ key: item.key, bytes: item.body.length, rows: item.rows })
+    }
+    // The marker exists before the manifest names the directory a device, and the manifest is written
+    // **last** on purpose: a manifest naming a shard that has not landed yet would send every other
+    // device after a 404 on their next run.
+    await client.put(`${REMOTE_PREFIX}${this.deviceId}/${DEVICE_MARKER}`, Buffer.from(`${this.deviceId}\n`))
+    let withdrawn = 0
+    const withdrawnKeys = []
+    for (const key of retired) {
+      try {
+        // Gone from the manifest *and* from the server: leaving the bytes behind while dropping the claim
+        // would be tidier for a manual rescue and worse for everything else — the next reader that finds
+        // the manifest is authoritative would see a shard nobody lists. The shard is *this device's own*,
+        // which is the one direction where deleting is not the remote deciding anything.
+        await client.delete(shardPath(this.deviceId, key))
+        delete manifest.chunks[key]
+        withdrawn++
+        withdrawnKeys.push(key)
+        this.log(`withdrew ${key}: this device no longer holds any rows for it`)
+      } catch (error) {
+        // The month could not be withdrawn, so the manifest is not written at all: half-recording it
+        // would leave the claim in place while the shard is gone. Reported, and retried next run.
+        return { uploaded: pending.length, unchanged, bytes, shards, manifest: null, withdrawn: 0, refused: null, withdrawError: `could not withdraw ${key}: ${error?.message ?? error}` }
+      }
+    }
+    await client.put(manifestPath(this.deviceId), Buffer.from(JSON.stringify(manifest, null, 2)), { contentType: 'application/json' })
+    return { uploaded: pending.length, unchanged, bytes, shards, manifest, withdrawn, withdrawnKeys }
+  }
+
+  /**
+   * Reconcile the shared price table.
+   *
+   * Runs after the shards, because losing a price row is worse than a delayed month: whatever this
+   * does, it happens while the shard sync is already recorded as done. The device copy is published
+   * first, so that even a conflict the user never looks at leaves both versions on the server.
+   *
+   * @returns {{action:string, status:string}}
+   */
+  async syncPricing(client) {
+    const localPath = join(this.dataDir, 'pricing.json')
+    // A table this device pulled is still "ours" for the next comparison, so local edits are stamped
+    // by the file's own mtime rather than by when the last sync happened — otherwise taking a remote
+    // copy would make the *next* run see a local change that nobody made.
+    const localStamp = () => {
+      try {
+        return Math.round(statSync(localPath).mtimeMs) || 0
+      } catch {
+        return 0
+      }
+    }
+    let localBytes = null
+    try {
+      localBytes = readFileSync(localPath)
+    } catch {
+      // No price table of our own: the built-in template is in force, and there is nothing to publish.
+      localBytes = null
+    }
+    let remoteBytes = null
+    try {
+      remoteBytes = await client.get(PRICING_FILE)
+    } catch (error) {
+      this.log(`the shared price table could not be read (${error?.message ?? error}); leaving ours alone`)
+      return { action: 'none', status: 'unreadable' }
+    }
+
+    const known = this.state.pricing ?? {}
+    const remoteHash = remoteBytes === null ? null : sha256(remoteBytes)
+    const plan = planPricing({
+      localBytes,
+      remoteBytes,
+      baseline: known.baseline ?? null,
+      localAt: localStamp(),
+      remoteAt: known.remoteAt ?? 0,
+      device: this.deviceId,
+      at: this.now(),
+    })
+
+    if (plan.action === 'none') {
+      // `baseline` records what the server had when this device last looked. It is only meaningful
+      // when there *is* a remote file: writing the hash of nothing would make the next run treat the
+      // baseline as a real version and call a fresh install's first sight of the table a conflict.
+      this.state.pricing = remoteHash === null ? { ...known, remoteHash } : { ...known, baseline: remoteHash, remoteHash }
+      // In step, but the file on disk may not match — a device that has no table of its own, or whose
+      // last run left the server's copy as a conflict file, still needs it written out.
+      if (remoteBytes !== null && !sameBytes(localBytes, remoteBytes)) {
+        writeFileSync(localPath, remoteBytes)
+        return { action: 'pull', status: 'took the server’s copy' }
+      }
+      return { action: plan.action, status: plan.winner ?? 'in step' }
+    }
+
+    if (plan.conflictFile) {
+      // Written locally, not remotely: a conflict copy is for whoever opens this machine. The losing
+      // side is always written, and the winner is never applied over a person's edit — the table is
+      // hand-written, so a silent overwrite is the one outcome this must not produce.
+      const losing = plan.winner === 'remote' ? localBytes ?? Buffer.alloc(0) : remoteBytes ?? Buffer.alloc(0)
+      writeFileSync(join(this.dataDir, plan.conflictFile), losing)
+      this.log(`the price table changed on both sides; kept the copy that would have been lost as ${plan.conflictFile}`)
+      if (plan.winner === 'local') {
+        // Ours is the later write, so it is the one that gets published — and the server's copy has
+        // just been saved beside it, so nothing was lost either way.
+        if (localBytes !== null) {
+          await client.put(pricingDevicePath(this.deviceId), localBytes, { contentType: 'application/json' })
+          await client.put(PRICING_FILE, localBytes, { contentType: 'application/json' })
+          const hash = sha256(localBytes)
+          this.state.pricing = { ...known, baseline: hash, remoteHash: hash, remoteAt: this.now() }
+          return { action: plan.action, status: 'published ours' }
+        }
+      }
+      // The server's version is the later write. It is **not** applied over the user's file: their
+      // edit stays where it is, and the server's copy is the conflict file that was just written.
+      this.state.pricing = { ...known, baseline: remoteHash, remoteHash, remoteAt: this.now() }
+      return { action: plan.action, status: 'kept the server’s copy as a conflict file' }
+    }
+
+    if (plan.winner === 'remote') {
+      writeFileSync(localPath, remoteBytes ?? Buffer.alloc(0))
+      this.state.pricing = { ...known, baseline: remoteHash, remoteHash, remoteAt: this.now() }
+      return { action: plan.action, status: 'took the server’s copy' }
+    }
+
+    if (localBytes !== null) {
+      await client.put(pricingDevicePath(this.deviceId), localBytes, { contentType: 'application/json' })
+      await client.put(PRICING_FILE, localBytes, { contentType: 'application/json' })
+      const hash = sha256(localBytes)
+      this.state.pricing = { ...known, baseline: hash, remoteHash: hash, remoteAt: this.now() }
+      return { action: plan.action, status: 'published ours' }
+    }
+    return { action: plan.action, status: 'nothing to publish' }
+  }
+
+  /**
+   * Publish only the price table, taking and giving the lock itself.
+   *
+   * This is what the host half calls a moment after an edit. It is deliberately **not** a full run: a
+   * price edit has nothing to do with shards, and a full run would drag the whole remote listing along
+   * with it. The lock is taken because the edit can land while a full run is in flight, and two writers
+   * on one data directory is the case the lock exists for.
+   *
+   * @param {{lock?:object, instanceId?:string}} [options]
+   * @returns {{result:string, status:string, action?:string, message:string}}
+   */
+  async pushPricing({ lock = null, instanceId = this.instanceId } = {}) {
+    if (!this.config.enabled || !this.config.url || !this.config.ha1) {
+      return { result: 'skipped', status: 'sync is off or unconfigured', message: 'sync is off or unconfigured' }
+    }
+    const held = lock ?? acquireLock(this.dataDir, { instanceId, log: this.log })
+    if (!held.acquired) return { result: 'locked', status: held.reason, message: held.reason }
+    let client
+    try {
+      client = this.connect()
+    } catch (error) {
+      const message = String(error?.message ?? error)
+      return { result: 'failed', status: message, message }
+    }
+    try {
+      const pricing = await this.syncPricing(client)
+      const message = `prices: ${pricing.status}`
+      // `unreadable` is not a successful reconcile: the shared table could not be read, so whether this
+      // device's edit is safe to publish is unknown — and it is deliberately not published. Recording it
+      // as `ok` would tell the panel everything went fine at the exact moment nothing could be checked.
+      const failed = pricing.status === 'unreadable' || pricing.action === 'failed'
+      this.state.lastSyncAt = this.now()
+      this.state.lastResult = failed ? 'failed' : 'ok'
+      this.state.message = message
+      writeSyncDocument(this.dataDir, { ...this.config, instanceId, state: this.state })
+      return { result: failed ? 'failed' : 'ok', status: pricing.status, action: pricing.action, message }
+    } catch (error) {
+      const message = String(error?.message ?? error)
+      this.log(`pricing push failed: ${message}`)
+      this.state.lastResult = 'failed'
+      this.state.message = `prices: ${message}`
+      writeSyncDocument(this.dataDir, { ...this.config, instanceId, state: this.state })
+      return { result: 'failed', status: message, message }
+    }
+  }
+
+  /**
+   * One sync: lock, discover, pull, push, price table, record.
+   *
+   * Every failure is caught and turned into `state.lastResult`, never into a thrown error the caller
+   * has to interpret: the panel shows the sentence, and the local cache is untouched either way.
+   */
+  async run({ lock = null } = {}) {
+    const started = this.now()
+    if (!this.config.enabled) {
+      return this.finish('skipped', 'sync is turned off', { started })
+    }
+    this.lock = lock ?? acquireLock(this.dataDir, { instanceId: this.instanceId, log: this.log })
+    this.ownsLock = lock === null
+    if (!this.lock.acquired) {
+      // `locked` is not a failure of the sync; it is this process not being the writer. The previous
+      // run's timestamp is deliberately left alone, so the panel does not claim a sync just happened.
+      this.state.lastResult = 'locked'
+      this.state.message = this.lock.reason
+      try {
+        writeSyncDocument(this.dataDir, { ...this.config, instanceId: this.instanceId, state: this.state })
+      } catch (error) {
+        this.log(`sync state could not be saved: ${error?.message ?? error}`)
+      }
+      return { result: 'locked', message: this.lock.reason, ms: this.now() - started }
+    }
+    let client
+    try {
+      client = this.connect()
+    } catch (error) {
+      return this.finish('failed', String(error?.message ?? error), { started })
+    }
+
+    try {
+      const manifests = await this.readManifests(client)
+      // **Pull first, refuse only the push.** An installation whose device name is taken still wants the
+      // history back, and that is the case the feature exists for (reinstall, another machine): the name
+      // is what is in the way, not the data. So the clash is noted here — which is what tells `pull` to
+      // take that directory's shards instead of skipping them as its own — and the refusal itself is
+      // `push`'s, where the decision to write is actually made.
+      const clash = manifests.find((manifest) => manifest.device === this.deviceId && manifest.instanceId && manifest.instanceId !== this.instanceId)
+      if (clash) this.state.collision = { device: this.deviceId, owner: clash.instanceId, at: this.now() }
+      else if (this.state.collision && this.state.collision.device !== this.deviceId) delete this.state.collision
+      // Read the clashing directory only when the clash is a *name* clash. A directory that merely
+      // disagrees with what we recorded is our own, and pulling it back would import a shard we are
+      // refusing to publish — the one case where the two readings of `collision` differ.
+      this.pullOwnDirectory = Boolean(clash)
+      const pulled = await this.pull(client, manifests)
+      this.pullOwnDirectory = false
+      const pushed = await this.push(client)
+      if (pushed.manifest) this.state.published = pushed.manifest.chunks
+      // The price table last: it is the only hand-written thing here, and it should not be touched by
+      // a run that failed to move the usage it was about to move.
+      const pricing = await this.syncPricing(client)
+      const parts = [
+        `pulled ${pulled.fetched} shard(s) / ${pulled.rows} row(s)`,
+        `pushed ${pushed.uploaded} shard(s) / ${Math.round(pushed.bytes / 1024)} KB`,
+        `${pushed.unchanged} month(s) unchanged`,
+      ]
+      if (pushed.refused) {
+        // Not `ok`: the months this device folded are not on the server. It is not `failed` either —
+        // the pull worked and nothing was lost. Naming it is what lets the panel say what to do.
+        parts.push(`nothing published: device name "${this.deviceId}" is being written by another installation`)
+      }
+      if (pushed.withdrawError) parts.push(pushed.withdrawError)
+      if (pushed.withdrawn) parts.push(`withdrew ${pushed.withdrawn} month(s) this device no longer holds`)
+      if (pricing.action !== 'none') parts.push(`prices: ${pricing.status}`)
+      if (pulled.failed.length > 0) parts.push(`${pulled.failed.length} shard(s) failed`)
+      const message = parts.join(', ')
+      this.log(`sync: ${message}`)
+      // Four outcomes, each meaning a different thing to the person reading the panel: `ok` moved
+      // everything, `partial` moved almost everything, `clash` deliberately published nothing and needs a
+      // decision from them, `failed` is a problem with the server or the network.
+      const result = pushed.refused ? 'clash' : pulled.failed.length > 0 || pushed.withdrawError ? 'partial' : 'ok'
+      return this.finish(result, message, {
+        started,
+        extra: {
+          pulled: pulled.fetched,
+          rows: pulled.rows,
+          pushed: pushed.uploaded,
+          bytes: pushed.bytes,
+          pricing: pricing.action,
+          failed: pulled.failed,
+          refused: pushed.refused ?? null,
+          withdrawn: pushed.withdrawn ?? 0,
+        },
+      })
+    } catch (error) {
+      const message = String(error?.message ?? error)
+      this.log(`sync failed: ${message}`)
+      return this.finish('failed', message, { started })
+    }
+  }
+
+  /**
+   * Record the outcome, persist the state, and return it.
+   *
+   * The state object is **mutated, not replaced**: {@link pull} and {@link push} hold a reference to
+   * it while they work, so assigning a fresh object here would silently drop everything they recorded
+   * — which is exactly how a sync ends up re-importing every shard on every run.
+   */
+  finish(result, message, { started, extra = {} } = {}) {
+    this.state.lastSyncAt = this.now()
+    this.state.lastResult = result
+    this.state.message = message
+    try {
+      writeSyncDocument(this.dataDir, { ...this.config, instanceId: this.instanceId, state: this.state })
+    } catch (error) {
+      this.log(`sync state could not be saved: ${error?.message ?? error}`)
+    }
+    return { result, message, ms: this.now() - started, ...extra }
+  }
+}