dsh-usage-badge 0.1.3 → 0.1.5

package/lib/cache-store.js

@@ -4,27 +4,46 @@
  * The folder hands this layer one session at a time and asks nothing about storage. Two
  * backends sit behind that interface:
  *
- *   - **sqlite** (default). One row per (session, day). A fold compares the session against its
- *     stored rows and writes only the days whose payload actually changed — measured 0.12 ms for a
- *     one-day session, and 4.98 ms for a 400-day one (the comparison still reads those rows, so the
- *     cost tracks the changed session's size, not the table's). Loading takes only the window the
- *     panel can show, which bounds resident memory the same way. Older rows stay on disk, which is
- *     the point: history is never deleted, it is just not all held in memory at once.
- *     It uses `node:sqlite`, which ships inside Node itself — no dependency to install, nothing
- *     to compile, no external tool for a user to run.
+ *   - **sqlite** (default). One row per (session, date) **copy**. A fold compares the session
+ *     against its stored rows and writes only the days whose payload actually changed — measured
+ *     0.12 ms for a one-day session, and 4.98 ms for a 400-day one (the comparison still reads
+ *     those rows, so the cost tracks the changed session's size, not the table's). Loading takes
+ *     only the window the panel can show, which bounds resident memory the same way. Older rows
+ *     stay on disk, which is the point: history is never deleted, it is just not all held in
+ *     memory at once. It uses `node:sqlite`, which ships inside Node itself — no dependency to
+ *     install, nothing to compile, no external tool for a user to run.
  *   - **json**: the original single-document cache, kept as the fallback for a runtime without
  *     `node:sqlite` and reachable on purpose with `DSH_USAGE_BADGE_CACHE=json`. It rewrites the
  *     whole document on every save and loads all of it, which is exactly the shape SQLite exists
  *     to avoid — so it is a compatibility path, not the recommended one.
  *
- * Both backends keep whatever they are given: neither ever deletes history on its own. The only
- * thing that removes a day is the folder folding a session whose log no longer contains it.
+ * Both backends keep whatever they are given: neither ever deletes a day on its own. A day goes
+ * away only when the folder folds a session whose log no longer contains it — and even then only
+ * if the fold is not *smaller* than what is stored, because a shorter fold is the signature of a
+ * shortened log or a restored backup and must not erase a fuller copy. A shorter fold is stored
+ * **beside** the fuller row (the rule and its reasoning live in `buckets.js`), and `resolveRows`
+ * picks the winner when reading.
  *
- * Beyond `put`/`load` they answer two questions the panel needs: `years()` — which calendar
- * years the stored history covers, so the year picker is built from the data and a new year
- * appears the first day it has a request, with nothing to configure — and a bounded range read
- * (`load({ until })`) for reproducing one of those years. A range read is a reporting path: it
- * never feeds a write back, which is what keeps the window optimisation safe.
+ * ## What a record looks like on the way in and out
+ *
+ * In: `{sessionId, path, fileSize, fileMtimeMs, seq, provider, model, days: Map<date, Map<key, bucket>>}`
+ * as the folder builds it.
+ *
+ * Out of `load()`: the same record with its rows **resolved** — one bucket map per date, the
+ * fuller log version having won — plus `rows`, every copy the store holds, each with its own
+ * `date`, `seq`, `fileSize`, `fileMtimeMs`, `source` and `at`. `scan()` is the same shape without
+ * resolution: it exists for the sync layer, which exports the local copies and must not export one
+ * it merely downloaded.
+ *
+ * Resolving inside the store rather than at the call site is what keeps the copies invisible to
+ * everything that only wants usage numbers: the folder, the snapshot and the year view all read
+ * `record.days` and know nothing about sources.
+ *
+ * Beyond `put`/`load` they answer two questions the panel needs: `years()` — which calendar years
+ * the stored history covers, so the year picker is built from the data and a new year appears the
+ * first day it has a request, with nothing to configure — and a bounded range read
+ * (`load({ since, until })`) for reproducing one of those years. A range read is a reporting path:
+ * it never feeds a write back, which is what keeps the window optimisation safe.
  *
  * @module dsh-usage-badge/cache-store
  */
@@ -33,7 +52,7 @@ import { createRequire } from 'node:module'
 import { existsSync, mkdirSync, readFileSync, renameSync, unlinkSync, writeFileSync } from 'node:fs'
 import { basename, join } from 'node:path'
 
-import { deserializeDay, deserializeDays, serializeDay, serializeDays } from './buckets.js'
+import { CURRENT, FALLBACK, LOCAL_SOURCE, legacyDayRows, normalizeRows, resolveRows, rowIsNewerOrEqual, rowIsStrictlyNewer, rowKind, rowSource, rowsOf } from './buckets.js'
 import { dayKey } from './pricing.js'
 
 /**
@@ -46,11 +65,19 @@ import { dayKey } from './pricing.js'
  */
 export const CACHE_WINDOW_DAYS = 430
 
-/** The document version the JSON backend writes; the original cache format, unchanged. */
-const JSON_CACHE_VERSION = 1
+/** The document version this backend writes. Version 1 is the pre-row shape, still read. */
+const JSON_CACHE_VERSION = 2
 
-/** The SQLite schema version, kept in `PRAGMA user_version`. */
-const SQLITE_SCHEMA_VERSION = 1
+/**
+ * The SQLite schema version, kept in `PRAGMA user_version`.
+ *
+ * Version 2 added the row model: `source` (which device a row came from), `kind` (whether the row
+ * is the fold the cursor stands on or a superseded copy kept beside it), `at` (the fold time, the
+ * tie-break between two copies of the same log version) and `seq` / `file_size` / `file_mtime_ms`
+ * on the day row itself, so a **fallback** row keeps the stamp of the log version it was folded
+ * from even after the session cursor has moved on.
+ */
+const SQLITE_SCHEMA_VERSION = 2
 
 const SCHEMA = `
 CREATE TABLE IF NOT EXISTS sessions (
@@ -65,8 +92,14 @@ CREATE TABLE IF NOT EXISTS sessions (
 CREATE TABLE IF NOT EXISTS days (
   session_id TEXT NOT NULL,
   date TEXT NOT NULL,
+  source TEXT NOT NULL,
+  kind TEXT NOT NULL,
+  seq INTEGER NOT NULL,
+  file_size INTEGER NOT NULL,
+  file_mtime_ms REAL NOT NULL,
+  at REAL NOT NULL,
   payload TEXT NOT NULL,
-  PRIMARY KEY (session_id, date)
+  PRIMARY KEY (session_id, date, source, kind)
 ) WITHOUT ROWID;
 CREATE INDEX IF NOT EXISTS days_by_date ON days (date);
 `
@@ -93,16 +126,27 @@ function resolveBounds({ windowDays, since, until } = {}) {
   }
 }
 
-/** A session's days restricted to `[since, until]`, both optional. */
-function filterDays(days, since, until) {
-  if (since === undefined && until === undefined) return days
-  const kept = new Map()
-  for (const [date, dayObj] of days) {
-    if (since !== undefined && date < since) continue
-    if (until !== undefined && date > until) continue
-    kept.set(date, dayObj)
+/** A session's rows restricted to `[since, until]`, both optional. */
+function within(rows, since, until) {
+  if (since === undefined && until === undefined) return rows
+  return rows.filter((row) => (since === undefined || row.date >= since) && (until === undefined || row.date <= until))
+}
+
+/**
+ * Parse one stored payload into a row, or null when it is not readable as one.
+ *
+ * The stored payload **is** the route map, so it is handed to `normalizeRows` under `routes`
+ * explicitly. Passing it as the row itself would send the normalizer looking for a `date` field
+ * inside a bucket, find none, and quietly drop every day.
+ */
+function parseRow(source, date, raw, meta = {}) {
+  try {
+    const parsed = typeof raw === 'string' ? JSON.parse(raw) : raw
+    return normalizeRows(source, [{ date, routes: parsed, ...meta }], { at: meta.at ?? 0 })[0] ?? null
+  } catch {
+    // A malformed row is skipped rather than losing the whole session.
+    return null
   }
-  return kept
 }
 
 /** Read a JSON cache document, or null when it is missing or of an unknown version. */
@@ -113,7 +157,7 @@ function readJsonDocument(path) {
   } catch {
     return null
   }
-  if (!raw || raw.version !== JSON_CACHE_VERSION || !raw.sessions || typeof raw.sessions !== 'object') return null
+  if (!raw || (raw.version !== JSON_CACHE_VERSION && raw.version !== 1) || !raw.sessions || typeof raw.sessions !== 'object') return null
   return raw
 }
 
@@ -123,15 +167,23 @@ const storedNumber = (value, fallback) => (Number.isFinite(Number(value)) ? Numb
 /** One stored session document → the record shape the folder works with. */
 function documentToRecord(sessionId, source) {
   if (!source || typeof source !== 'object') return null
+  const stamp = {
+    seq: storedNumber(source.seq, -1),
+    fileSize: Number(source.fileSize) || 0,
+    fileMtimeMs: Number(source.fileMtimeMs) || 0,
+  }
+  // Version 1 kept a single `{date: {key: bucket}}` map with one stamp per session. Its days become
+  // rows carrying that stamp, which is what lets an upgrade keep reading them.
+  const rows = Array.isArray(source.days)
+    ? source.days.flatMap((row) => normalizeRows(stamp, [row]))
+    : legacyDayRows(source.days ?? {}, stamp)
   return {
     sessionId,
     path: String(source.path ?? ''),
-    fileSize: Number(source.fileSize) || 0,
-    fileMtimeMs: Number(source.fileMtimeMs) || 0,
-    seq: storedNumber(source.seq, -1),
+    ...stamp,
     provider: source.provider ?? null,
     model: source.model ?? null,
-    days: deserializeDays(source.days),
+    rows,
   }
 }
 
@@ -203,58 +255,205 @@ function createJsonStore({ path, log }) {
     }
   }
 
+  /**
+   * The document this store holds, read from disk the first time.
+   *
+   * Kept in memory across `put` calls while a write is pending: re-reading here would discard the
+   * fold that `put` just accepted, because the pending timer would then write that stale copy
+   * back. This is also why an absolute range read (the year view) must not reload — it is a
+   * reporting read, not a refresh.
+   */
+  const ensureDocument = () => {
+    if (document && dirty) return document
+    const loaded = readJsonDocument(path)
+    if (loaded) document = loaded
+    else {
+      quarantine()
+      document = { version: JSON_CACHE_VERSION, sessions: {} }
+    }
+    return document
+  }
+
   return {
     kind: 'json',
     path,
     load(bounds = {}) {
       const { from, until, absolute } = resolveBounds(bounds)
-      // Keep the document this store already holds while a write is pending: re-reading the file
-      // here would silently discard the fold that `put` just accepted, because the pending timer
-      // would then write that stale copy back. This is why an absolute range read (the year view)
-      // must not reload either — it is a reporting read, not a refresh.
-      if (!document || !dirty) {
-        const loaded = readJsonDocument(path)
-        if (loaded) document = loaded
-        else {
-          quarantine()
-          document = { version: JSON_CACHE_VERSION, sessions: {} }
-        }
-      }
-      const records = Object.entries(document.sessions)
+      const doc = ensureDocument()
+      const records = Object.entries(doc.sessions)
         .map(([sessionId, source]) => documentToRecord(sessionId, source))
         .filter(Boolean)
       // This backend writes back exactly what it holds, so it cannot honour a *relative* window
       // without dropping history on the next save: it returns everything. An absolute range is a
       // reporting read and safe to filter.
-      if (!absolute) return records
-      return records.map((record) => ({ ...record, days: filterDays(record.days, from, until) }))
+      if (!absolute) return resolveRows(records)
+      return resolveRows(records.map((record) => ({ ...record, rows: within(record.rows, from, until) })))
+    },
+    scan() {
+      const doc = ensureDocument()
+      return Object.entries(doc.sessions)
+        .map(([sessionId, source]) => documentToRecord(sessionId, source))
+        .filter(Boolean)
     },
     years() {
       const years = new Set()
       for (const source of Object.values(document?.sessions ?? {})) {
-        for (const date of Object.keys(source?.days ?? {})) years.add(String(date).slice(0, 4))
+        const dates = Array.isArray(source?.days) ? source.days.map((row) => row?.date) : Object.keys(source?.days ?? {})
+        for (const date of dates) if (date) years.add(String(date).slice(0, 4))
       }
       return [...years].filter((year) => /^\d{4}$/.test(year)).sort().reverse()
     },
-    put(record) {
-      if (!document) document = { version: JSON_CACHE_VERSION, sessions: {} }
-      document.sessions[record.sessionId] = {
+    /**
+     * Store one fold and hand back the resolved view of that session.
+     *
+     * The return value matters: a fold that covered **less** than what was already stored leaves
+     * the fuller copy in place, and the caller must go on showing the fuller one rather than the
+     * short fold it just produced. Returning the resolved record is what makes "the cache kept
+     * it" true in memory as well as on disk.
+     *
+     * The rules are the SQLite backend's, expressed on the document: a fold replaces its own
+     * `current` rows, a weaker fold lands in the `fallback` slot instead of over the fuller copy,
+     * another source's rows are never a fold's to judge, and a date the log no longer covers is
+     * dropped only when the log version is not itself weaker than what is stored (a shortened log is
+     * a rollback, not a deletion).
+     */
+    put(record, options = {}) {
+      const doc = ensureDocument()
+      const at = options.at ?? Date.now()
+      const previous = documentToRecord(record.sessionId, doc.sessions[record.sessionId])
+      const incoming = rowsOf(record, { source: LOCAL_SOURCE, at })
+      const incomingKeys = new Set(incoming.map((row) => `${row.date}\u0000${row.source}`))
+      const cursor = { seq: record.seq ?? -1, fileSize: record.fileSize ?? 0 }
+      const kept = []
+      for (const row of previous?.rows ?? []) {
+        const key = `${row.date}\u0000${rowSource(row)}`
+        // A row survives the fold unless the fold is entitled to remove it:
+        //   - another source's copy is never a fold's to judge, and neither is a `fallback` copy —
+        //     that is a log version the cursor has moved past, kept on purpose;
+        //   - its own `current` row the fold still covers is replaced by the incoming copy below;
+        //   - its own `current` row the fold does **not** cover goes only if this log version still
+        //     reaches back to it. A row older than the incoming cursor was folded from a fuller log
+        //     — a shorter one now, or a restored backup — and it is all that is left of those days.
+        const theirs = rowSource(row) !== LOCAL_SOURCE || rowKind(row) === FALLBACK
+        const keep = theirs || incomingKeys.has(key) || !rowIsNewerOrEqual(cursor, row)
+        if (keep) kept.push(row)
+      }
+      // Grouped by `(date, source, kind)`: the `current` and `fallback` copies of one date share a
+      // date and a source but are different rows, and a map keyed without the slot would let one
+      // silently overwrite the other.
+      const group = (rows) => new Map(rows.map((row) => [`${row.date}\u0000${row.source}\u0000${rowKind(row)}`, row]))
+      const merged = group(kept)
+      for (const row of incoming) {
+        const key = `${row.date}\u0000${row.source}\u0000${CURRENT}`
+        const existing = merged.get(key) ?? null
+        if (!rowIsStrictlyNewer(row, existing)) {
+          // A short fold goes into the `fallback` slot beside the fuller copy rather than replacing
+          // it — that copy is the log's own current truth, and all that is left if the fuller one
+          // ever goes away.
+          merged.set(`${row.date}\u0000${row.source}\u0000${FALLBACK}`, { ...row, kind: FALLBACK })
+          continue
+        }
+        merged.set(key, row)
+      }
+      kept.length = 0
+      kept.push(...merged.values())
+      doc.sessions[record.sessionId] = {
         path: record.path,
         fileSize: record.fileSize,
         fileMtimeMs: record.fileMtimeMs,
         seq: record.seq,
         provider: record.provider,
         model: record.model,
-        days: serializeDays(record.days),
+        days: kept.map((row) => ({
+          date: row.date,
+          routes: row.routes,
+          seq: row.seq,
+          fileSize: row.fileSize,
+          fileMtimeMs: row.fileMtimeMs,
+          source: rowSource(row),
+          kind: rowKind(row),
+          at: row.at ?? 0,
+        })),
       }
       dirty = true
-      if (timer) return
-      // Batched, because this backend pays for the whole history on every save.
-      timer = setTimeout(() => {
-        timer = null
-        if (dirty) write()
-      }, 2000)
-      timer.unref?.()
+      if (!timer) {
+        // Batched, because this backend pays for the whole history on every save.
+        timer = setTimeout(() => {
+          timer = null
+          if (dirty) write()
+        }, 2000)
+        timer.unref?.()
+      }
+      return resolveRows([documentToRecord(record.sessionId, doc.sessions[record.sessionId])])[0] ?? null
+    },
+    /**
+     * Store rows that came from another device, and hand back how many were new.
+     *
+     * The SQLite backend's rules on the document: never delete, and store a row only when it is
+     * strictly fuller than the one already held for that `(session, date, source, slot)`. Grouped by
+     * session because a shard's rows arrive together but a document is edited per session.
+     */
+    intakeRows(rows) {
+      const doc = ensureDocument()
+      const bySession = new Map()
+      for (const row of rows ?? []) {
+        if (!row?.sessionId || typeof row.date !== 'string') continue
+        let list = bySession.get(row.sessionId)
+        if (!list) {
+          list = []
+          bySession.set(row.sessionId, list)
+        }
+        list.push(row)
+      }
+      let written = 0
+      for (const [sessionId, incoming] of bySession) {
+        const existing = documentToRecord(sessionId, doc.sessions[sessionId])
+        const merged = new Map((existing?.rows ?? []).map((row) => [`${row.date}\u0000${rowSource(row)}\u0000${rowKind(row)}`, row]))
+        for (const row of incoming) {
+          const kind = rowKind(row) === FALLBACK ? FALLBACK : CURRENT
+          const key = `${row.date}\u0000${row.source}\u0000${kind}`
+          const known = merged.get(key)
+          if (known && !rowIsStrictlyNewer(row, known)) continue
+          merged.set(key, { ...row, kind })
+          written++
+        }
+        const rows_ = [...merged.values()]
+        const fullest = rows_.reduce((top, row) => (top === null || rowIsStrictlyNewer(row, top) ? row : top), null)
+        const previous = doc.sessions[sessionId]
+        // The cursor has to exist for the days to be readable at all (`readRecords` walks the
+        // sessions and hangs the days off them), and it must never regress: it is the stamp the fold
+        // path compares its own log version against.
+        const keepCursor = previous && !rowIsStrictlyNewer(fullest ?? {}, previous)
+        doc.sessions[sessionId] = {
+          path: previous?.path ?? '',
+          fileSize: keepCursor ? previous.fileSize : fullest?.fileSize ?? 0,
+          fileMtimeMs: keepCursor ? previous.fileMtimeMs : fullest?.fileMtimeMs ?? 0,
+          seq: keepCursor ? previous.seq : fullest?.seq ?? -1,
+          provider: previous?.provider ?? null,
+          model: previous?.model ?? null,
+          days: rows_.map((row) => ({
+            date: row.date,
+            routes: row.routes,
+            seq: row.seq,
+            fileSize: row.fileSize,
+            fileMtimeMs: row.fileMtimeMs,
+            source: rowSource(row),
+            kind: rowKind(row),
+            at: row.at ?? 0,
+          })),
+        }
+      }
+      if (written > 0) {
+        dirty = true
+        if (!timer) {
+          timer = setTimeout(() => {
+            timer = null
+            if (dirty) write()
+          }, 2000)
+          timer.unref?.()
+        }
+      }
+      return written
     },
     flush,
     // The same thing here, and a plain function rather than `this.flush()`, so a caller that
@@ -314,6 +513,7 @@ function openDatabase(DatabaseSync, path, log) {
     // A cache does not need full durability: the WAL survives a process crash, and a power loss
     // can only cost the last few folded days, which the session logs can supply again.
     probe.exec('PRAGMA synchronous = NORMAL')
+    migrateSchema(probe, version, log)
     probe.exec(SCHEMA)
     if (version !== SQLITE_SCHEMA_VERSION) probe.exec(`PRAGMA user_version = ${SQLITE_SCHEMA_VERSION}`)
     return { db: probe, version, writable: true }
@@ -323,6 +523,66 @@ function openDatabase(DatabaseSync, path, log) {
   }
 }
 
+/**
+ * Bring a database written by an older build up to this schema.
+ *
+ * Checked by asking the file which columns it has, not by trusting `user_version` alone: a file
+ * whose version number was rewritten, or whose `days` table predates the stamp columns, must be
+ * migrated rather than hit with an INSERT naming columns it does not have. The old rows are
+ * brought across rather than dropped — an upgrade that silently emptied the cache would look
+ * exactly like a bug.
+ */
+function migrateSchema(db, version, log) {
+  const columns = (table) => {
+    try {
+      return db.prepare(`PRAGMA table_info(${table})`).all().map((row) => String(row.name))
+    } catch {
+      return []
+    }
+  }
+  const dayColumns = columns('days')
+  if (dayColumns.length === 0 || dayColumns.includes('kind')) return
+
+  log(`cache schema ${version} predates the row model; migrating its days table`)
+  const previous = []
+  try {
+    for (const row of db.prepare('SELECT session_id, date, payload FROM days').all()) {
+      const cursor = db.prepare('SELECT seq, file_size, file_mtime_ms FROM sessions WHERE id = ?').get(row.session_id) ?? {}
+      previous.push({
+        sessionId: row.session_id,
+        date: row.date,
+        source: LOCAL_SOURCE,
+        kind: CURRENT,
+        seq: storedNumber(cursor.seq, -1),
+        fileSize: Number(cursor.file_size) || 0,
+        fileMtimeMs: Number(cursor.file_mtime_ms) || 0,
+        payload: row.payload,
+      })
+    }
+  } catch (error) {
+    log(`cache migration could not read the old rows (${error?.message ?? error}); the days table starts fresh`)
+    previous.length = 0
+  }
+
+  db.exec('BEGIN IMMEDIATE')
+  try {
+    db.exec('DROP TABLE IF EXISTS days')
+    db.exec(SCHEMA)
+    const insert = db.prepare('INSERT OR REPLACE INTO days (session_id, date, source, kind, seq, file_size, file_mtime_ms, at, payload) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)')
+    for (const row of previous) {
+      insert.run(row.sessionId, row.date, row.source, row.kind, row.seq, row.fileSize, row.fileMtimeMs, 0, row.payload)
+    }
+    db.exec('COMMIT')
+  } catch (error) {
+    try {
+      db.exec('ROLLBACK')
+    } catch {
+      // the transaction is already gone
+    }
+    throw error
+  }
+}
+
 /**
  * Whether a failure means "these bytes are not a database", as opposed to "this process may not
  * use this database right now". Only the first is worth repairing by moving the file aside.
@@ -358,7 +618,6 @@ function createSqliteStore({ path, jsonPath, log, DatabaseSync, db: given = null
   const { db } = opened
   const writable = opened.writable !== false
 
-  /** Prepare, marking any failure as "do not touch this file" — it opened, so it is a database. */
   const prepare = (sql) => {
     try {
       return db.prepare(sql)
@@ -373,53 +632,187 @@ function createSqliteStore({ path, jsonPath, log, DatabaseSync, db: given = null
     }
   }
 
-  const selectSessions = prepare('SELECT id, path, file_size, file_mtime_ms, seq, provider, model FROM sessions')
-  const selectDays = prepare('SELECT session_id, date, payload FROM days WHERE date >= ? AND date <= ? ORDER BY date')
+  /**
+   * A statement that may be impossible against this file's schema, or null.
+   *
+   * This exists for the read-only path: a database written by a newer version may have moved a
+   * column this build's queries name, and the alternative would be to give up on the file
+   * entirely — which the design refuses to do, because a downgrade must not cost the user their
+   * history. A query that cannot run is skipped, so such a file reads as far as it can be read
+   * and is then left untouched.
+   */
+  const tryPrepare = (sql) => {
+    try {
+      return db.prepare(sql)
+    } catch {
+      return null
+    }
+  }
+  const all = (statement, ...args) => (statement ? statement.all(...args) : [])
+  const get = (statement, ...args) => (statement ? statement.get(...args) : undefined)
+
+  const selectSessions = tryPrepare('SELECT id, path, file_size, file_mtime_ms, seq, provider, model FROM sessions')
+  const selectDays = tryPrepare('SELECT session_id, date, source, kind, seq, file_size, file_mtime_ms, at, payload FROM days WHERE date >= ? AND date <= ? ORDER BY date')
+  const selectSessionDays = tryPrepare('SELECT date, source, kind, seq, file_size, file_mtime_ms, at, payload FROM days WHERE session_id = ?')
+  const selectEveryDay = tryPrepare('SELECT session_id, date, source, kind, seq, file_size, file_mtime_ms, at, payload FROM days')
   // Which years exist is a question only the store can answer: the folder holds a window, not
   // the whole history, and the panel's year list has to include years older than that window.
-  const selectYears = prepare("SELECT DISTINCT substr(date, 1, 4) AS year FROM days WHERE date GLOB '[0-9][0-9][0-9][0-9]-*' ORDER BY year DESC")
-  const countSessions = prepare('SELECT COUNT(*) AS n FROM sessions')
-  const upsertSession = prepare(
-    `INSERT INTO sessions (id, path, file_size, file_mtime_ms, seq, provider, model)
-     VALUES (?, ?, ?, ?, ?, ?, ?)
-     ON CONFLICT(id) DO UPDATE SET
-       path = excluded.path, file_size = excluded.file_size, file_mtime_ms = excluded.file_mtime_ms,
-       seq = excluded.seq, provider = excluded.provider, model = excluded.model`,
-  )
-  const selectSessionDays = prepare('SELECT date, payload FROM days WHERE session_id = ?')
-  const deleteDay = prepare('DELETE FROM days WHERE session_id = ? AND date = ?')
-  const insertDay = prepare('INSERT INTO days (session_id, date, payload) VALUES (?, ?, ?)')
+  const selectYears = tryPrepare("SELECT DISTINCT substr(date, 1, 4) AS year FROM days WHERE date GLOB '[0-9][0-9][0-9][0-9]-*' ORDER BY year DESC")
+  const countSessions = tryPrepare('SELECT COUNT(*) AS n FROM sessions')
+  const selectCursor = tryPrepare('SELECT seq, file_size, file_mtime_ms FROM sessions WHERE id = ?')
+  const insertRow = writable
+    ? prepare('INSERT OR REPLACE INTO days (session_id, date, source, kind, seq, file_size, file_mtime_ms, at, payload) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)')
+    : null
+  const deleteRow = writable ? prepare('DELETE FROM days WHERE session_id = ? AND date = ? AND source = ? AND kind = ?') : null
+  const upsertSession = writable
+    ? prepare(
+        `INSERT INTO sessions (id, path, file_size, file_mtime_ms, seq, provider, model)
+         VALUES (?, ?, ?, ?, ?, ?, ?)
+         ON CONFLICT(id) DO UPDATE SET
+           path = excluded.path, file_size = excluded.file_size, file_mtime_ms = excluded.file_mtime_ms,
+           seq = excluded.seq, provider = excluded.provider, model = excluded.model`,
+      )
+    : null
+
+  /**
+   * Every row a session holds, or every row in the store when no session is named, in whatever
+   * schema the file turned out to have.
+   */
+  const readSessionRows = (sessionId, { source, kind } = {}) =>
+    (sessionId === undefined ? all(selectEveryDay) : all(selectSessionDays, sessionId))
+      .filter((row) => (source === undefined || row.source === source) && (kind === undefined || (row.kind ?? CURRENT) === kind))
+      .map((row) => ({
+        sessionId: row.session_id ?? sessionId,
+        date: row.date,
+        source: row.source,
+        kind: row.kind ?? CURRENT,
+        seq: storedNumber(row.seq, 0),
+        fileSize: Number(row.file_size) || 0,
+        fileMtimeMs: Number(row.file_mtime_ms) || 0,
+        at: Number(row.at) || 0,
+        payload: String(row.payload ?? ''),
+      }))
 
   /**
-   * Write one session: its cursor row, plus only the days whose stored payload really differs.
+   * Store one row.
    *
-   * The folder hands over a whole session, so the day *set* is authoritative — a day the record no
-   * longer has is deleted, which is how a log that shrank behaves (exactly as it did when this was
-   * one JSON document). But a day that did not change is not rewritten: a long-lived session is
-   * re-folded on every poll that sees its log grow, and re-inserting its whole history each time
-   * would make a save cost grow with that session's age.
+   * `INSERT OR REPLACE` rather than a delete-then-insert pair: the primary key is
+   * `(session, date, source)`, and a second copy for the same key is the local fold's own
+   * superseded row set — exactly the case a plain `INSERT` fails on. Failing there would roll the
+   * whole fold back, which is the loudest possible way to lose the shortened fold this model exists
+   * to keep.
    */
-  const writeSession = (record) => {
+  const writeRow = (sessionId, row) =>
+    insertRow.run(sessionId, row.date, row.source, row.kind ?? CURRENT, row.seq, row.fileSize, row.fileMtimeMs, row.at ?? 0, JSON.stringify(row.routes ?? {}))
+
+  /** Store a session's cursor: the log version this device last folded it from. */
+  const writeCursor = (record) =>
     upsertSession.run(
       record.sessionId,
-      record.path,
-      record.fileSize,
-      record.fileMtimeMs,
-      record.seq,
-      record.provider,
-      record.model,
+      record.path ?? '',
+      record.fileSize ?? 0,
+      record.fileMtimeMs ?? 0,
+      record.seq ?? -1,
+      record.provider ?? null,
+      record.model ?? null,
     )
-    const stored = new Map(selectSessionDays.all(record.sessionId).map((row) => [row.date, row.payload]))
-    for (const [date, dayObj] of record.days ?? []) {
-      const payload = JSON.stringify(serializeDay(dayObj))
-      const known = stored.get(date)
-      stored.delete(date)
-      if (known === payload) continue
-      if (known !== undefined) deleteDay.run(record.sessionId, date)
-      insertDay.run(record.sessionId, date, payload)
+
+  /** A session's stored cursor, as a stamp the row comparisons understand. */
+  const readCursor = (sessionId) => {
+    const row = get(selectCursor, sessionId)
+    return row ? { seq: storedNumber(row.seq, -1), fileSize: Number(row.file_size) || 0, fileMtimeMs: Number(row.file_mtime_ms) || 0 } : null
+  }
+
+  /**
+   * Write one session's fold.
+   *
+   * Three rules, all about not losing history:
+   *
+   *   - a fold owns **its own** rows only. A day the log no longer has is deleted from this device's
+   *     `current` rows and from nowhere else — a row that arrived from another device is not the
+   *     folder's to remove;
+   *   - a `current` row is replaced only when the new fold describes **at least as much** of the
+   *     session (higher `seq`, or equal `seq` and no smaller file). A shorter fold goes to the
+   *     `fallback` slot beside the fuller copy instead of over it;
+   *   - when the fold is shorter *and* drops a day, the day is rolled back rather than deleted —
+   *     this is what a restored backup looks like, and it is the case the whole row model exists for.
+   */
+  const writeSession = (record, { source = LOCAL_SOURCE, at = Date.now() } = {}) => {
+    writeCursor(record)
+    const cursor = readCursor(record.sessionId)
+    const incoming = rowsOf(record, { source, at })
+    const stored = new Map(readSessionRows(record.sessionId, { source, kind: CURRENT }).map((row) => [row.date, row]))
+    for (const row of incoming) {
+      const known = stored.get(row.date)
+      if (known) stored.delete(row.date)
+      if (known && !rowIsNewerOrEqual(row, known)) {
+        // The fold covers less than the copy already stored for that day. The fuller copy keeps the
+        // `current` slot — the log has been shortened, but that is no reason to forget what it said —
+        // and this short fold is written as the `fallback`: it is what is left if the fuller copy is
+        // ever gone (the machine is restored from the same backup, say).
+        if (known.payload !== JSON.stringify(row.routes)) writeRow(record.sessionId, { ...row, kind: FALLBACK })
+        continue
+      }
+      if (known && known.payload === JSON.stringify(row.routes)) continue
+      if (known) deleteRow.run(record.sessionId, row.date, source, CURRENT)
+      writeRow(record.sessionId, row)
+    }
+    // A date this device folded before, that the record no longer covers at all, is gone from the
+    // log. Its `current` row goes with it — unless this log version is itself weaker than the copy
+    // already stored, in which case a shortened log is treated as a rollback rather than as a
+    // deletion, and the day is kept.
+    for (const row of stored.values()) {
+      if (cursor && !rowIsNewerOrEqual(cursor, row)) continue
+      deleteRow.run(record.sessionId, row.date, source, CURRENT)
     }
-    // Whatever is left was stored and is not in the record any more.
-    for (const date of stored.keys()) deleteDay.run(record.sessionId, date)
+  }
+
+  /**
+   * Take rows that came from somewhere other than this device's own fold.
+   *
+   * This is the **sync import path**, and it takes rows rather than records because an incoming row
+   * carries its own stamp and its own slot: a shard is a set of rows, not a session's current fold.
+   * Two rules, both about not losing history:
+   *
+   *   - it never deletes. Only this device's own fold may remove this device's rows; a machine whose
+   *     logs were pruned must not be able to erase what another machine published;
+   *   - a row is stored only when it is **strictly fuller** than the one already held for that
+   *     `(date, source, slot)`. An equal-stamp copy is the same bytes and is left as it is; a weaker
+   *     one is a worse answer to the same question.
+   *
+   * @returns {number} how many rows were actually written.
+   */
+  const intakeRows = (rows) => {
+    const known = new Map(readSessionRows(undefined).map((row) => [`${row.sessionId}\u0000${row.date}\u0000${row.source}\u0000${row.kind}`, row]))
+    let written = 0
+    for (const row of rows ?? []) {
+      if (!row?.sessionId || typeof row.date !== 'string') continue
+      const kind = rowKind(row) === FALLBACK ? FALLBACK : CURRENT
+      // The day rows are only reachable through a session row (`readRecords` walks `sessions` and
+      // hangs the days off it), so an import has to make sure one exists — otherwise the history is
+      // written into a table nothing reads.
+      const existing = known.get(`${row.sessionId}\u0000${row.date}\u0000${row.source}\u0000${kind}`)
+      if (existing && !rowIsStrictlyNewer(row, existing)) continue
+      const payload = JSON.stringify(row.routes ?? {})
+      if (existing && existing.payload === payload) continue
+      if (existing) deleteRow.run(row.sessionId, row.date, row.source, kind)
+      writeRow(row.sessionId, { ...row, kind })
+      written++
+    }
+    // One cursor per session that arrived, from the fullest row that arrived for it.
+    const bySession = new Map()
+    for (const row of rows ?? []) {
+      if (!row?.sessionId) continue
+      const best = bySession.get(row.sessionId)
+      if (!best || rowIsStrictlyNewer(row, best)) bySession.set(row.sessionId, row)
+    }
+    for (const [sessionId, best] of bySession) {
+      const cursor = readCursor(sessionId)
+      if (!cursor || rowIsStrictlyNewer(best, cursor)) {
+        writeCursor({ sessionId, path: '', fileSize: best.fileSize, fileMtimeMs: best.fileMtimeMs, seq: best.seq })
+      }
+    }
+    return written
   }
 
   /** Bring an existing JSON cache across, once, keeping the old file for inspection. */
@@ -428,7 +821,7 @@ function createSqliteStore({ path, jsonPath, log, DatabaseSync, db: given = null
     // database that already has rows is never re-read from the JSON file, so a failed rename
     // below cannot cause the same history to be folded in twice.
     if (!writable) return
-    if ((countSessions.get()?.n ?? 0) > 0) {
+    if ((get(countSessions)?.n ?? 0) > 0) {
       // Rows already exist, so this JSON file is a leftover — from a run that used the JSON
       // backend deliberately, from a period spent on the fallback, or from an older host that
       // wrote it after the migration. Those folds are not merged (a day-by-day merge of two
@@ -449,9 +842,15 @@ function createSqliteStore({ path, jsonPath, log, DatabaseSync, db: given = null
       .map(([sessionId, source]) => documentToRecord(sessionId, source))
       .filter(Boolean)
     if (records.length === 0) return
+    // The rows are written directly rather than through `writeSession`: a migrated record is
+    // already row-shaped (version 1's day map was turned into rows with the session's stamp on
+    // the way in), and re-deriving rows from a `days` Map there is not would import nothing.
     try {
       db.exec('BEGIN IMMEDIATE')
-      for (const record of records) writeSession(record)
+      for (const record of records) {
+        writeCursor(record)
+        for (const row of record.rows) writeRow(record.sessionId, row)
+      }
       db.exec('COMMIT')
     } catch (error) {
       try {
@@ -480,52 +879,98 @@ function createSqliteStore({ path, jsonPath, log, DatabaseSync, db: given = null
   }
   migrateFromJson()
 
+  /** Every session, with the rows inside `[from, until]`. */
+  const readRecords = (from, until) => {
+    const byId = new Map()
+    for (const row of all(selectSessions)) {
+      byId.set(row.id, {
+        sessionId: row.id,
+        path: row.path,
+        fileSize: Number(row.file_size) || 0,
+        fileMtimeMs: Number(row.file_mtime_ms) || 0,
+        seq: storedNumber(row.seq, -1),
+        provider: row.provider ?? null,
+        model: row.model ?? null,
+        rows: [],
+      })
+    }
+    for (const row of all(selectDays, from, until)) {
+      const record = byId.get(row.session_id)
+      if (!record) continue
+      const parsed = parseRow(record, row.date, row.payload, {
+        seq: Number(row.seq) || 0,
+        fileSize: Number(row.file_size) || 0,
+        fileMtimeMs: Number(row.file_mtime_ms) || 0,
+        source: row.source,
+        kind: row.kind ?? CURRENT,
+        at: Number(row.at) || 0,
+      })
+      if (!parsed) continue
+      record.rows.push(parsed)
+    }
+    return [...byId.values()]
+  }
+
   return {
     kind: 'sqlite',
     path,
     writable,
     load(bounds = {}) {
       const { from, until } = resolveBounds(bounds)
-      const byId = new Map()
-      for (const row of selectSessions.all()) {
-        byId.set(row.id, {
-          sessionId: row.id,
-          path: row.path,
-          fileSize: Number(row.file_size) || 0,
-          fileMtimeMs: Number(row.file_mtime_ms) || 0,
-          seq: storedNumber(row.seq, -1),
-          provider: row.provider ?? null,
-          model: row.model ?? null,
-          days: new Map(),
-        })
-      }
-      for (const row of selectDays.all(from, until)) {
-        const record = byId.get(row.session_id)
-        if (!record) continue
+      return resolveRows(readRecords(from, until))
+    },
+    scan() {
+      return readRecords('0000-01-01', '9999-12-31')
+    },
+    years() {
+      return all(selectYears).map((row) => String(row.year))
+    },
+    /**
+     * Store one fold and hand back the resolved view of that session.
+     *
+     * The return value matters: a fold that covered **less** than what was already stored leaves
+     * the fuller copy in place, and the caller must go on showing the fuller one rather than the
+     * short fold it just produced. Returning the resolved record is what makes "the cache kept it"
+     * true in memory as well as on disk.
+     */
+    put(record, { at = Date.now() } = {}) {
+      if (!writable) return null
+      try {
+        db.exec('BEGIN IMMEDIATE')
+        writeSession(record, { source: LOCAL_SOURCE, at })
+        db.exec('COMMIT')
+      } catch (error) {
         try {
-          record.days.set(row.date, deserializeDay(JSON.parse(row.payload)))
+          db.exec('ROLLBACK')
         } catch {
-          // A malformed row is skipped rather than losing the whole session.
+          // the transaction is already gone
         }
+        log(`cache write failed for ${record.sessionId}: ${error?.message ?? error}`)
+        return null
       }
-      return [...byId.values()]
+      return resolveRows(readRecords('0000-01-01', '9999-12-31').filter((row) => row.sessionId === record.sessionId))[0] ?? null
     },
-    years() {
-      return selectYears.all().map((row) => String(row.year))
-    },
-    put(record) {
-      if (!writable) return
+    /**
+     * Store rows that came from another device, and hand back how many were new.
+     *
+     * Wrapped in one transaction: an imported shard is a month of somebody's history, and a partial
+     * month would leave the two devices disagreeing about which days exist.
+     */
+    intakeRows(rows) {
+      if (!writable) return 0
       try {
         db.exec('BEGIN IMMEDIATE')
-        writeSession(record)
+        const written = intakeRows(rows)
         db.exec('COMMIT')
+        return written
       } catch (error) {
         try {
           db.exec('ROLLBACK')
         } catch {
           // the transaction is already gone
         }
-        log(`cache write failed for ${record.sessionId}: ${error?.message ?? error}`)
+        log(`cache import failed: ${error?.message ?? error}`)
+        return 0
       }
     },
     flush() {
@@ -559,7 +1004,7 @@ function createSqliteStore({ path, jsonPath, log, DatabaseSync, db: given = null
  *     must not destroy data it does not understand.
  *
  * @param {{dir:string, log?:Function, env?:object}} options
- * @returns {{kind:string, path:string, load:Function, years:Function, put:Function, flush:Function, close:Function}}
+ * @returns {{kind:string, path:string, load:Function, scan:Function, years:Function, put:Function, flush:Function, close:Function}}
  */
 export function createCacheStore({ dir, log = () => {}, env = process.env }) {
   const jsonPath = join(dir, 'cache.json')