Skip to main content

mnemo_engine/store/
maintenance.rs

1//! Maintenance: export, rekey, verify, stats, and compaction (which also upgrades v8 files to v9).
2
3use super::*;
4
5impl Mnemo {
6    /// Return the KDF (Argon2id) parameters this file currently uses to
7    /// wrap its DEK. Needed by callers that want to [`rekey`] while
8    /// keeping the file's original protection level (e.g. the Python
9    /// binding's `rekey(fast=False)`, whose docstring promises exactly
10    /// that). FIX-1 R-001.5.
11    pub fn kdf_params(&self) -> KdfParams {
12        self.kdf
13    }
14
15    /// Flush any pending writes, then stream the (encrypted) file to
16    /// `dest` atomically: writes to a sibling temp file in `dest`'s
17    /// parent directory, `sync_all`s it, renames onto `dest`, and
18    /// fsyncs the parent directory on Unix so a crash never leaves a
19    /// partial export.
20    ///
21    /// Refuses when `dest` resolves to the file this handle has open
22    /// — that case would otherwise `File::create`-truncate the live
23    /// database to 0 bytes (`fs::copy` on Unix reads *after* the
24    /// `O_TRUNC`, so a naive export-to-self destroys the source).
25    /// Unix: compares `(dev, ino)` of both paths via `MetadataExt`,
26    /// catching symlinks and hard links. Windows: compares
27    /// canonicalized paths. (The rename is no second line of defense:
28    /// renaming over a file another handle holds open and locked
29    /// succeeds on Windows too, as a 2026-11-13 CI probe showed.)
30    ///
31    /// Reads the bytes from the already-open file handle, not by
32    /// `fs::copy`. On Windows that matters: `LockFileEx` locks are
33    /// mandatory, so a second handle opened by `fs::copy` would fail
34    /// while this writer still holds the lock.
35    ///
36    /// FIX-1 R-001.5 + R-002.2 part 3.
37    pub fn export_encrypted(&mut self, dest: impl AsRef<Path>) -> Result<()> {
38        let dest = dest.as_ref();
39
40        // Step 1 — refuse export to the open file itself. We compare
41        // by identity, not string, so symlinks and hard links to
42        // `self.path` are caught too. The check runs before any flush
43        // so a self-targeted export is a pure no-op.
44        if paths_refer_to_same_file(&self.path, dest).unwrap_or(false) {
45            return Err(MnemoError::Invalid(format!(
46                "export_encrypted refuses to overwrite the open database at {}",
47                dest.display()
48            )));
49        }
50
51        // Step 2 — flush, so on-disk bytes match the live state.
52        if !self.read_only {
53            self.flush()?;
54        }
55
56        // Step 3 — stream from the open handle into a sibling temp.
57        // Using the open handle (instead of a second open on `self.path`)
58        // means a Windows mandatory lock can't block the copy.
59        let parent = dest.parent().ok_or_else(|| {
60            MnemoError::Invalid(format!(
61                "export_encrypted: dest {} has no parent directory",
62                dest.display()
63            ))
64        })?;
65        // A new file under a fresh name: `create_new` refuses anything
66        // already there, a symlink included, and a taken name gets another.
67        let (tmp, mut out) = {
68            let mut attempt = 0;
69            loop {
70                let tmp = temp_sibling(dest, "export");
71                match owner_only(OpenOptions::new().write(true).create_new(true)).open(&tmp) {
72                    Ok(out) => break (tmp, out),
73                    Err(e)
74                        if e.kind() == std::io::ErrorKind::AlreadyExists
75                            && attempt + 1 < TEMP_NAME_ATTEMPTS =>
76                    {
77                        attempt += 1;
78                    }
79                    Err(e) => return Err(MnemoError::Io(e)),
80                }
81            }
82        };
83        // Scope the temp handle so it closes before the rename on
84        // Windows (which refuses to rename a file this process has open).
85        {
86            let src = self.pager.file_mut();
87            src.seek(SeekFrom::Start(0))?;
88            let mut buf = vec![0u8; 1 << 20]; // 1 MiB
89            loop {
90                let n = src.read(&mut buf)?;
91                if n == 0 {
92                    break;
93                }
94                out.write_all(&buf[..n])?;
95            }
96            out.sync_all()?;
97            crate::bump_fsync();
98            drop(out);
99        }
100
101        // Step 4 — atomic rename onto `dest`.
102        if let Err(e) = std::fs::rename(&tmp, dest) {
103            // Best-effort cleanup; propagate the rename error.
104            let _ = std::fs::remove_file(&tmp);
105            return Err(MnemoError::Io(e));
106        }
107
108        // Step 5 — fsync the parent directory on Unix so the rename
109        // itself is durable. On Windows directory syncs aren't
110        // supported; the rename above is atomic enough on NTFS.
111        #[cfg(unix)]
112        {
113            if let Ok(dir) = File::open(parent) {
114                let _ = dir.sync_all();
115                crate::bump_fsync();
116            }
117        }
118        #[cfg(not(unix))]
119        {
120            let _ = parent;
121        }
122
123        Ok(())
124    }
125
126    /// Change the passphrase. Cheap: re-derives the KEK and re-wraps the DEK;
127    /// the encrypted pages are never rewritten.
128    ///
129    /// **This changes the secret that opens this file from now on; anyone who
130    /// had the file and the old secret can still read it and later writes.**
131    /// The data key stays the same, so an earlier copy of the file (or of
132    /// just its header) and the old passphrase still unlock it. The old wrap
133    /// is scrubbed from this file (the previous WAL region is zeroed, and any
134    /// stale copy of the old header's key wrap is overwritten); that's
135    /// hygiene, not revocation. **To revoke, rotate:** rekey, then
136    /// [`Mnemo::compact_file`] (`mnemo rekey --rotate`), which rewrites the
137    /// file under a fresh data key. Rotation can't erase old ciphertext the storage hardware still holds: flash keeps rewritten blocks until it reuses them, and that's out of the engine's hands.
138    ///
139    /// **An error means the old secret is still in force,** for this handle
140    /// and for the file, including after later writes and a close (REKEY-1).
141    /// The one exception is when the new header's write fails and putting
142    /// the old one back fails too: the error then says so, and the file may
143    /// open with either secret.
144    ///
145    /// `Ok` means the new secret is in force. Scrubbing the old wrap
146    /// afterwards is hygiene; if part of it fails, the rekey still succeeds
147    /// and [`Mnemo::rekey_cleanup_error`] says what was left.
148    pub fn rekey(&mut self, new_passphrase: &str, kdf: KdfParams) -> Result<()> {
149        self.rekey_secret(new_passphrase.as_bytes(), SecretKind::Passphrase, kdf)
150    }
151
152    /// PERF-1 step 7: switch to (or rotate) a raw 256-bit key, recording
153    /// [`KdfParams::raw_key`] in the header. Like [`Mnemo::rekey`], only the
154    /// wrapped DEK changes; pages aren't rewritten, and it doesn't revoke
155    /// the old key (see [`Mnemo::rekey`]).
156    ///
157    /// Errors and the cleanup report work as in [`Mnemo::rekey`].
158    pub fn rekey_with_key(&mut self, new_key: &[u8; crypto::KEY_LEN]) -> Result<()> {
159        self.rekey_secret(new_key, SecretKind::Key, KdfParams::raw_key())
160    }
161
162    pub(super) fn rekey_secret(
163        &mut self,
164        secret: &[u8],
165        kind: SecretKind,
166        kdf: KdfParams,
167    ) -> Result<()> {
168        if self.shared {
169            // RB-007.2: the shared handle would keep the old secret, and
170            // fail at the next compaction or recovery.
171            return Err(MnemoError::Unsupported(
172                "inside a shared-mode window, change the passphrase with SharedMnemo::rekey, \
173                 which also updates the secret the handle keeps. Nothing was changed"
174                    .into(),
175            ));
176        }
177        self.rekey_unchecked(secret, kind, kdf, None)
178    }
179
180    /// 12.3 (the keys note, section 4): [`Mnemo::rekey`] or
181    /// [`Mnemo::rekey_with_key`], with `before_commit` run once the new salt
182    /// is drawn and before anything is written, so a caller can store the new
183    /// secret under that salt first. An error from it stops the rekey with
184    /// nothing changed. Used by [`crate::keys::rekey_stored`].
185    pub(crate) fn rekey_staged(
186        &mut self,
187        secret: &[u8],
188        kind: SecretKind,
189        kdf: KdfParams,
190        before_commit: &mut crypto::SaltHook<'_>,
191    ) -> Result<()> {
192        if self.shared {
193            return Err(MnemoError::Unsupported(
194                "a staged rekey runs on a plain handle only".into(),
195            ));
196        }
197        self.rekey_unchecked(secret, kind, kdf, Some(before_commit))
198    }
199
200    /// The rekey itself; [`SharedMnemo::rekey`] calls it in its window. 12.3:
201    /// where every rekey meets, so a passphrase with raw-key parameters is
202    /// refused here (the keys note, section 1, rule 2).
203    pub(super) fn rekey_unchecked(
204        &mut self,
205        secret: &[u8],
206        kind: SecretKind,
207        kdf: KdfParams,
208        before_commit: Option<&mut crypto::SaltHook<'_>>,
209    ) -> Result<()> {
210        if self.read_only {
211            return Err(MnemoError::ReadOnly);
212        }
213        crypto::check_kdf_for(kind, kdf)?;
214        self.rekey_cleanup = None;
215        self.flush()?;
216        // v0.14.0 step 1: `flush` checks only when it has something to
217        // write; the rekey writes either way.
218        self.ensure_not_replaced()?;
219        let new_salt = crypto::random_salt();
220        let new_kek = crypto::derive_kek(secret, &new_salt, kdf)?;
221        let new_nonce = crypto::random_nonce();
222        let wrapped = crypto::wrap_dek(&new_kek, &new_nonce, self.pager.dek())?;
223        if wrapped.len() != WRAPPED_DEK_LEN {
224            return Err(MnemoError::Crypto("unexpected wrapped DEK length".into()));
225        }
226        let mut wrapped_dek = [0u8; WRAPPED_DEK_LEN];
227        wrapped_dek.copy_from_slice(&wrapped);
228        // 12.3: the caller stores the new secret under `new_salt` before
229        // anything is written.
230        if let Some(hook) = before_commit {
231            hook(&new_salt)?;
232        }
233
234        // SEC-2: the old header's key wrap (salt, DEK nonce, wrapped DEK, as
235        // laid out at header bytes 86..162), to scrub once the new header
236        // is committed.
237        let mut old_wrap =
238            Vec::with_capacity(crypto::SALT_LEN + crypto::NONCE_LEN + WRAPPED_DEK_LEN);
239        old_wrap.extend_from_slice(&self.header.salt);
240        old_wrap.extend_from_slice(&self.header.dek_nonce);
241        old_wrap.extend_from_slice(&self.header.wrapped_dek);
242
243        // REKEY-1 (RB-004.2): until the new page 0 is on disk, a failure puts
244        // these back, so an `Err` means "still the old secret".
245        let before = self.header.clone();
246        let before_kdf = self.kdf;
247
248        self.header.salt = new_salt;
249        self.header.dek_nonce = new_nonce;
250        self.header.wrapped_dek = wrapped_dek;
251        self.header.m_cost = kdf.m_cost;
252        self.header.t_cost = kdf.t_cost;
253        self.header.p_cost = kdf.p_cost;
254        self.kdf = kdf;
255
256        // DUR-1 (R-002.1): WAL-log the new page 0 before touching it.
257        // Allocate a fresh WAL region at the tail so that a crash during
258        // the rekey leaves either the old region (old passphrase works)
259        // or the new region (new passphrase works) fully committed —
260        // overwriting the existing region would risk destroying both.
261        let old_wal_start = self.header.wal_start;
262        let new_pages = self.header.wal_pages;
263        let new_start = self.header.next_page;
264        self.header.next_page += new_pages;
265        self.header.wal_start = new_start;
266        self.header.wal_pages = new_pages;
267        self.header.wal_seq += 1;
268
269        // The steps of `relocate_header_through_wal`, apart, so a failure
270        // can be told from one before page 0 and one after it.
271        self.header.write_counter = self.pager.write_counter;
272        let mut page = self.header.to_page();
273        if let Err(e) = self.header.apply_seal(&mut page, self.pager.dek()) {
274            return Err(self.undo_rekey(before, before_kdf, None, false, e));
275        }
276        let frames = vec![(0u64, page.to_vec())];
277        let txn_id = self.header.wal_seq;
278        if let Err(e) = wal::commit(self.pager.file_mut(), new_start, new_pages, txn_id, &frames) {
279            let region = Some((new_start, new_pages));
280            return Err(self.undo_rekey(before, before_kdf, region, false, e));
281        }
282        if let Err(e) = self
283            .pager
284            .write_raw(0, &page)
285            .and_then(|_| self.pager.sync())
286        {
287            let region = Some((new_start, new_pages));
288            return Err(self.undo_rekey(before, before_kdf, region, true, e));
289        }
290
291        // The new page 0 is on disk: the new secret is in force, and nothing
292        // below fails the rekey. SEC-2 (RB-001.30): zero the old region's
293        // first frame header (so no scan reads it as a WAL), then all of it
294        // (its frames hold page-0 images with the old wrap), then any other
295        // stale copy of the old wrap in the file (an earlier, grown or
296        // relocated region). Each step runs even if the one before failed.
297        // Hygiene only: a copy of the file taken before the rekey still
298        // opens with the old secret (see `rekey`).
299        let mut left = Vec::new();
300        if let Err(e) = wal::invalidate_region(self.pager.file_mut(), old_wal_start) {
301            left.push(format!("clearing the old WAL region's first frame: {e}"));
302        }
303        if let Err(e) = wal::scrub_region(self.pager.file_mut(), old_wal_start, new_pages) {
304            left.push(format!("zeroing the old WAL region: {e}"));
305        }
306        if let Err(e) = scrub_bytes(self.pager.file_mut(), &old_wrap) {
307            left.push(format!("overwriting other copies of the old key wrap: {e}"));
308        }
309        self.rekey_cleanup = (!left.is_empty()).then(|| left.join("; "));
310        // FMT-9 (a): the old region is freed by the next commit (see
311        // `ensure_wal_capacity`).
312        self.release_pages(old_wal_start, new_pages);
313        Ok(())
314    }
315
316    /// REKEY-1: a rekey failed before its new page 0 was on disk. Put the
317    /// header and the KDF back, so the handle (and every later write) uses
318    /// the old secret. Then, best effort: zero the new WAL region, whose
319    /// commit wraps the data key under the new secret, so a later
320    /// torn-header scan can't find it; and, if page 0 may have been
321    /// written, write the old one back.
322    fn undo_rekey(
323        &mut self,
324        before: crate::format::Header,
325        before_kdf: KdfParams,
326        new_region: Option<(u64, u64)>,
327        page0_written: bool,
328        err: MnemoError,
329    ) -> MnemoError {
330        self.header = before;
331        self.kdf = before_kdf;
332        // RB-005.2: a failed zeroing is reported, not dropped. The old
333        // secret is still in force either way; what's left is a committed
334        // copy of the new key wrap in the region past the file's end,
335        // which a torn-header scan could use with the new secret until a
336        // compaction clears it (a later write over those pages also does).
337        let scrub_failed = new_region.and_then(|(start, pages)| {
338            wal::scrub_region(self.pager.file_mut(), start, pages).err()
339        });
340        let with_scrub_note = |err: MnemoError| match &scrub_failed {
341            None => err,
342            Some(e) => MnemoError::Io(std::io::Error::other(format!(
343                "{err}; the old secret is still in force, but zeroing the new key wrap's \
344                 copy failed ({e}), so a copy may be left in the file until `mnemo compact` \
345                 clears it (a later write over those pages also does)"
346            ))),
347        };
348        if !page0_written {
349            return with_scrub_note(err);
350        }
351        self.header.write_counter = self.pager.write_counter;
352        let mut page = self.header.to_page();
353        let put_back = self
354            .header
355            .apply_seal(&mut page, self.pager.dek())
356            .and_then(|_| self.pager.write_raw(0, &page))
357            .and_then(|_| self.pager.sync());
358        match put_back {
359            Ok(()) => with_scrub_note(err),
360            // `mnemo rekey` looks for "may open with either secret" to leave
361            // out its "the passphrase was not changed": keep the phrase.
362            Err(again) => MnemoError::Io(std::io::Error::other(format!(
363                "rekey failed writing the new header ({err}), and the old header \
364                 couldn't be written back ({again}): the file {EITHER_SECRET}; close it \
365                 without writing and try both"
366            ))),
367        }
368    }
369
370    /// 12.3: whether `e` is REKEY-1's one exception, the error that says the
371    /// file may open with either secret. Kept beside the message it reads.
372    pub(crate) fn rekey_left_either_secret(e: &MnemoError) -> bool {
373        e.to_string().contains(EITHER_SECRET)
374    }
375
376    /// REKEY-1 (RB-004.2): after a successful [`Mnemo::rekey`] or
377    /// [`Mnemo::rekey_with_key`], what couldn't be scrubbed of the old key
378    /// wrap, or `None` if it all was. The rekey took effect either way; what
379    /// is left is plaintext key-wrap bytes the old secret could still use, so
380    /// to be sure, rotate (`compact_file`) as [`Mnemo::rekey`] describes.
381    /// Each rekey resets it.
382    pub fn rekey_cleanup_error(&self) -> Option<&str> {
383        self.rekey_cleanup.as_deref()
384    }
385
386    /// SEC-2 (RB-001.30): write a copy of this database at `dest` under a
387    /// **fresh data key**, protected by `passphrase` through `kdf`. It goes
388    /// through the compaction path (live memories, the result cache and
389    /// the index, no history), so the copy shares no key material with
390    /// this file: neither this file's passphrase nor its data key opens
391    /// it, and its passphrase opens nothing here. This file is flushed
392    /// first (unless read-only) and is otherwise unchanged; `dest` must not
393    /// exist. The copy is v9.
394    pub fn export_rekeyed(
395        &mut self,
396        dest: impl AsRef<Path>,
397        passphrase: &str,
398        kdf: KdfParams,
399    ) -> Result<CompactReport> {
400        self.export_rekeyed_secret(
401            dest.as_ref(),
402            passphrase.as_bytes(),
403            SecretKind::Passphrase,
404            kdf,
405        )
406    }
407
408    /// [`Mnemo::export_rekeyed`] protected by a raw 256-bit key (no KDF).
409    pub fn export_rekeyed_with_key(
410        &mut self,
411        dest: impl AsRef<Path>,
412        key: &[u8; crypto::KEY_LEN],
413    ) -> Result<CompactReport> {
414        self.export_rekeyed_secret(dest.as_ref(), key, SecretKind::Key, KdfParams::raw_key())
415    }
416
417    pub(super) fn export_rekeyed_secret(
418        &mut self,
419        dest: &Path,
420        secret: &[u8],
421        kind: SecretKind,
422        kdf: KdfParams,
423    ) -> Result<CompactReport> {
424        crypto::check_kdf_for(kind, kdf)?;
425        if paths_refer_to_same_file(&self.path, dest).unwrap_or(false) {
426            return Err(MnemoError::Invalid(format!(
427                "export_rekeyed refuses to overwrite the open database at {}",
428                dest.display()
429            )));
430        }
431        if dest.exists() {
432            return Err(MnemoError::Invalid(format!(
433                "{} already exists",
434                dest.display()
435            )));
436        }
437        if !self.read_only {
438            self.flush()?;
439        }
440        let config = MnemoConfig {
441            dimensions: self.dimensions,
442            kdf,
443            store_float_vectors: self.float_vectors(),
444            ..Default::default()
445        };
446        let (tmp, mut new) = create_temp(dest, "export", secret, kind, config, None)?;
447        let result = (|| {
448            let report = self.copy_live_into(&mut new)?;
449            new.flush()?;
450            drop(new);
451            std::fs::rename(&tmp, dest)?;
452            Ok(report)
453        })();
454        if result.is_err() {
455            let _ = std::fs::remove_file(&tmp);
456        }
457        result
458    }
459
460    /// The compaction copy: every live, unexpired memory and cache entry of
461    /// `self` into `new` (a freshly created file), and the index if `self`
462    /// has one. Counts as [`CompactReport`].
463    pub(super) fn copy_live_into(&mut self, new: &mut Mnemo) -> Result<CompactReport> {
464        // FMT-9e: compaction rewrites everything, so it compresses every
465        // thread turn, recent or not.
466        new.turn_compress_after = 0;
467        // ANN-1.1: compaction is "re-tune for the current size" by
468        // default. Pass IndexConfig::default() so partitions,
469        // subspaces, n_probe and n_rerank are all re-derived from the
470        // compacted vector count. Previously we froze the pre-compact
471        // (n_probe, n_rerank), which left the new index on stale
472        // numbers for the smaller post-compact corpus.
473        let want_index = self.has_index();
474        let now = memory::now_secs();
475        let all = self.memories()?;
476        let before = all.len();
477        let mut after = 0;
478        let mut zero_vector_manifests = 0;
479        let mut zero_vector_memories = 0;
480        for mut m in all {
481            if m.is_expired(now) {
482                continue;
483            }
484            // A file from before the zero-vector rule (Phase 3 PR 10a) can
485            // hold all-zero vectors. A legacy onboarding or manifest memory
486            // gets the sentinel unit vector a current manifest has, so it
487            // behaves like a new one; any other is carried unchanged.
488            if embed::is_zero_vector(&m.vector) {
489                if crud::is_onboarding(&m) || is_manifest(&m) {
490                    m.vector = Memory::scaffold_manifest(self.dimensions).vector;
491                    zero_vector_manifests += 1;
492                } else {
493                    zero_vector_memories += 1;
494                }
495            }
496            // The engine's own copy: `_mnemo.*` keys come through as they
497            // are, and the write-time zero-vector check doesn't apply.
498            new.put_existing(m)?;
499            after += 1;
500        }
501        if want_index && after > 0 {
502            new.build_index_with(IndexConfig::default())?;
503        }
504
505        // Count + copy cache entries. We walk `self.cache_dir` by index
506        // so each hop reads the directory's live `accessed_at` /
507        // `access_count` directly rather than going through
508        // `cache_get` (which would also treat read-only semantics,
509        // tombstones, etc. — we want the raw, authoritative values).
510        let mut cache_before = 0usize;
511        let mut cache_after = 0usize;
512        let positions: Vec<usize> = (0..self.cache_dir.len()).collect();
513        for pos in positions {
514            let (keep, reserved) = {
515                let e = &self.cache_dir[pos];
516                (
517                    !e.deleted && !e.is_expired(now),
518                    embed::is_reserved_namespace(&e.namespace),
519                )
520            };
521            if keep {
522                // EMBED-ID-1: the engine's own entries (the embedder
523                // identity) are copied but not counted.
524                new.copy_cache_entry_from(self, pos)?;
525                if !reserved {
526                    cache_before += 1;
527                    cache_after += 1;
528                }
529            }
530        }
531        Ok(CompactReport {
532            before,
533            after,
534            cache_before,
535            cache_after,
536            zero_vector_manifests,
537            zero_vector_memories,
538        })
539    }
540
541    /// Decrypt and validate every live record. Returns the count verified
542    /// across memories **and** cache entries (FIX-1 R-001.1: now that the
543    /// cache survives compaction, `verify()` has to walk it too — the
544    /// pre-FIX-1 implementation only checked `self.catalog`, so a
545    /// corrupted cache record went unnoticed).
546    pub fn verify(&mut self) -> Result<usize> {
547        Ok(self.verify_report()?.records)
548    }
549
550    /// [`Mnemo::verify`], also counting the memories whose vector is all
551    /// zeros. A file written before the zero-vector rule (Phase 3 PR 10a)
552    /// can hold them; they decode fine but score 0.0 on every vector query.
553    /// `mnemo verify` warns about them, and compaction carries them
554    /// (giving a legacy manifest the sentinel vector).
555    pub fn verify_report(&mut self) -> Result<VerifyReport> {
556        let entries: Vec<CatalogEntry> = self
557            .catalog
558            .iter()
559            .filter(|e| !e.deleted)
560            .cloned()
561            .collect();
562        let mut n = entries.len();
563        let mut zero_vector_memories = 0;
564        for e in &entries {
565            let bytes = self.read_record(e)?;
566            let m: Memory = self.decode_memory(&bytes)?;
567            // FMT-9b: and the row's vector column entry.
568            let zero = if e.v9.is_some() {
569                let (h, _) = self.column_entry(e)?;
570                if h.len() != self.dimensions {
571                    return Err(MnemoError::Invalid(
572                        "vector entry has the wrong width".into(),
573                    ));
574                }
575                // fp16 +0.0 and -0.0.
576                !h.is_empty() && h.iter().all(|b| b & 0x7fff == 0)
577            } else {
578                embed::is_zero_vector(&m.vector)
579            };
580            if zero {
581                zero_vector_memories += 1;
582            }
583        }
584
585        // Also decode every live cache entry. Tombstoned entries are
586        // skipped — their pages are orphan ciphertext awaiting
587        // `compact_file`, same as deleted memories. Expired entries
588        // are *still* on disk until compaction, so we verify them too.
589        let cache_positions: Vec<(u64, u32, u32, u64, bool)> = self
590            .cache_dir
591            .iter()
592            .filter(|e| !e.deleted)
593            .map(|e| {
594                let reserved = embed::is_reserved_namespace(&e.namespace);
595                (e.start_page, e.page_count, e.len, e.gen, reserved)
596            })
597            .collect();
598        for (start_page, page_count, len, gen, reserved) in cache_positions {
599            let buf = self.read_record_run(start_page, page_count, len, gen)?;
600            let _c: CacheEntry = rmp_serde::from_slice(&buf)
601                .map_err(|err| MnemoError::Serialize(err.to_string()))?;
602            // EMBED-ID-1: the engine's own entries are checked, not counted.
603            if !reserved {
604                n += 1;
605            }
606        }
607        // FMT-9 (a): no free run may overlap a page the committed state
608        // references (records, cache entries, control runs, the WAL).
609        if self.is_v9() {
610            self.check_free_list_disjoint()?;
611        }
612        Ok(VerifyReport {
613            records: n,
614            zero_vector_memories,
615            keyword_tokenizer: self.keyword_tokenizer()?,
616        })
617    }
618
619    /// FMT-9 (a): fail if the free list (or a pending free) overlaps any
620    /// page the current state references.
621    pub(super) fn check_free_list_disjoint(&self) -> Result<()> {
622        let h = &self.header;
623        let mut live: Vec<(u64, u64, &str)> = vec![
624            (h.wal_start, h.wal_pages, "WAL region"),
625            (h.catalog_start, h.catalog_pages, "catalog run"),
626            (h.index_start, h.index_pages, "index run"),
627            (h.manifest_start, h.manifest_pages, "manifest run"),
628            (h.cache_start, h.cache_pages, "cache directory run"),
629            (h.free.start, h.free.pages, "free-list run"),
630            (h.keys.start, h.keys.pages, "key-slot run"),
631        ];
632        live.extend(
633            self.catalog
634                .iter()
635                .filter(|e| !e.deleted)
636                .map(|e| (e.start_page, e.page_count as u64, "record")),
637        );
638        live.extend(
639            self.cache_dir
640                .iter()
641                .filter(|e| !e.deleted)
642                .map(|e| (e.start_page, e.page_count as u64, "cache entry")),
643        );
644        live.extend(self.col.page_rows.keys().map(|&p| (p, 1, "vector page")));
645        live.extend(
646            self.delta_chain
647                .iter()
648                .map(|r| (r.start, r.pages, "catalog delta run")),
649        );
650        live.extend(
651            self.index_chain
652                .iter()
653                .map(|r| (r.start, r.pages, "index delta run")),
654        );
655        live.push((
656            self.header.bm25.start,
657            self.header.bm25.pages,
658            "BM25 manifest",
659        ));
660        live.extend(
661            self.bm25
662                .manifest
663                .segments
664                .iter()
665                .map(|t| (t.0, t.1, "BM25 segment")),
666        );
667        for (p, c, what) in live {
668            if c == 0 {
669                continue;
670            }
671            if self.free.overlaps(p, c) {
672                return Err(MnemoError::Invalid(format!(
673                    "free list overlaps a live {what} at page {p} (+{c})"
674                )));
675            }
676            if self
677                .pending_free
678                .iter()
679                .any(|&(q, n)| q < p + c && p < q + n)
680            {
681                return Err(MnemoError::Invalid(format!(
682                    "a pending free overlaps a live {what} at page {p} (+{c})"
683                )));
684            }
685        }
686        Ok(())
687    }
688
689    /// The statistics that need no record decrypts and no index load
690    /// (v0.12.2, RB-002.3): counts, sizes, the format, whether an index
691    /// exists, the recall mode and the bytes the vector column needs. The
692    /// first [`Mnemo::stats`] on a handle decrypts every record (for its
693    /// agent list) and loads a persisted index; this is what a phone app
694    /// polling the recall mode should call.
695    pub fn stats_summary(&self) -> StatsSummary {
696        let deleted = self.catalog.iter().filter(|e| e.deleted).count();
697        let v9 = self.is_v9();
698        let column = if v9 {
699            (self.scored_rows() * crate::veccol::entry_len(self.dimensions)) as u64
700        } else {
701            0
702        };
703        let recall_mode = if self.has_index() {
704            RecallMode::Indexed
705        } else if v9 && column <= self.vector_cache_budget as u64 {
706            RecallMode::ExactResident
707        } else {
708            RecallMode::ExactPaged
709        };
710        let rp = self.restore_points();
711        StatsSummary {
712            memories: self.len(),
713            deleted,
714            dimensions: self.dimensions,
715            file_bytes: self.header.next_page * PAGE_SIZE as u64,
716            format_version: self.header.version,
717            wal_pages: self.header.wal_pages,
718            has_index: self.has_index(),
719            recall_mode,
720            vector_column_bytes: column,
721            snapshots_retained: rp.retained,
722            oldest_restorable_at: rp.oldest_at,
723            snapshot_pinned_bytes: rp.pinned_bytes,
724        }
725    }
726
727    /// Summary statistics for the open database. The first call on a
728    /// handle decrypts every record (to list the agents) and loads a
729    /// persisted index; [`Mnemo::stats_summary`] does neither.
730    pub fn stats(&mut self) -> Result<Stats> {
731        let deleted = self.catalog.iter().filter(|e| e.deleted).count();
732        // FIX-1 R-001.8 part 2: build the agent set lazily on first
733        // call per handle, then maintain in `put` / invalidate in
734        // `delete`. Pre-FIX-1 every `stats()` call decrypted every
735        // live memory (O(N) AES-GCM decrypts just to list a usually-
736        // tiny set of distinct agent IDs). On a 1M-memory database
737        // this was in the multi-second range on hardware AES.
738        if self.agent_set.is_none() {
739            let built: std::collections::BTreeSet<String> =
740                self.memories()?.into_iter().map(|m| m.agent_id).collect();
741            self.agent_set = Some(built);
742        }
743        let agents: Vec<String> = self
744            .agent_set
745            .as_ref()
746            .expect("agent_set just filled above")
747            .iter()
748            .cloned()
749            .collect();
750        let file_bytes = self.header.next_page * PAGE_SIZE as u64;
751        let recall_mode = self.recall_mode()?;
752        let rp = self.restore_points();
753        Ok(Stats {
754            memories: self.len(),
755            deleted,
756            dimensions: self.dimensions,
757            file_bytes,
758            agents,
759            encrypted: self.header.flags & FLAG_ENCRYPTED != 0,
760            created_at: self.header.created_at,
761            index: {
762                self.ensure_ann()?;
763                self.ann.as_ref().map(|a| a.info())
764            },
765            wal_pages: self.header.wal_pages,
766            fsyncs_last_flush: self.fsyncs_last_flush,
767            format_version: self.header.version,
768            recall_mode,
769            snapshots_retained: rp.retained,
770            oldest_restorable_at: rp.oldest_at,
771            snapshot_pinned_bytes: rp.pinned_bytes,
772            keyword_tokenizer: self.keyword_tokenizer()?,
773        })
774    }
775
776    /// Rewrite the file, dropping tombstoned and expired memories and
777    /// reclaiming stale pages left by updates. Done out-of-place via a temp
778    /// file and an atomic rename.
779    ///
780    /// FIX-1 R-001.1 + R-002.2 part 2: the result cache now survives
781    /// compaction. Every live, non-expired directory entry is copied
782    /// into the new file verbatim — `created_at`, `ttl_secs`,
783    /// `cost_hint_ms`, `key_preview`, `size_bytes`, `content_type`,
784    /// `vector` and `model` come from the record body; `accessed_at`
785    /// and `access_count` come from the *directory* (the record body
786    /// only holds the insert-time snapshot — the directory is the
787    /// source of truth for the live counts, as of v5's recall trick).
788    /// Tombstoned and TTL-expired entries are dropped. Pre-FIX-1
789    /// compaction kept only the memory catalog and silently emptied
790    /// the cache; the Planner's interim README caveat (c417d2e) called
791    /// that out — this PR removes the surprise.
792    ///
793    /// The output is always v9, so this is also how a v4–v8 file is
794    /// converted: the only read-write open that accepts one (the v8
795    /// sunset, v0.13). It keeps every live memory.
796    pub fn compact_file(path: &str, passphrase: &str) -> Result<CompactReport> {
797        Self::compact_file_secret(
798            path,
799            passphrase.as_bytes(),
800            SecretKind::Passphrase,
801            0,
802            &crate::SharedOptions::default(),
803        )
804    }
805
806    /// PERF-1 step 7: [`Mnemo::compact_file`] for a raw-key database.
807    pub fn compact_file_with_key(path: &str, key: &[u8; crypto::KEY_LEN]) -> Result<CompactReport> {
808        Self::compact_file_secret(
809            path,
810            key,
811            SecretKind::Key,
812            0,
813            &crate::SharedOptions::default(),
814        )
815    }
816
817    /// [`Mnemo::compact_file`] with the lock layer's options: how long to
818    /// wait for the file's lock (`opts.deadline`; `compact_file` waits 10 s),
819    /// a cancel, and `allow_network_fs`. `mnemo compact --wait` uses it.
820    pub fn compact_file_waiting(
821        path: &str,
822        passphrase: &str,
823        opts: crate::SharedOptions,
824    ) -> Result<CompactReport> {
825        Self::compact_file_secret(
826            path,
827            passphrase.as_bytes(),
828            SecretKind::Passphrase,
829            0,
830            &opts,
831        )
832    }
833
834    /// [`Mnemo::compact_file_waiting`] for a raw-key database.
835    pub fn compact_file_waiting_with_key(
836        path: &str,
837        key: &[u8; crypto::KEY_LEN],
838        opts: crate::SharedOptions,
839    ) -> Result<CompactReport> {
840        Self::compact_file_secret(path, key, SecretKind::Key, 0, &opts)
841    }
842
843    pub(super) fn compact_file_secret(
844        path: &str,
845        secret: &[u8],
846        kind: SecretKind,
847        format_version: u16,
848        opts: &crate::SharedOptions,
849    ) -> Result<CompactReport> {
850        // RB-007.3: compact the file the path names, through any symlink,
851        // and rename over that file, so a link stays a link and every path
852        // to the brain names the new file.
853        let real = std::fs::canonicalize(path)?;
854        // The one read-write open that takes a v4–v8 file (the v8 sunset).
855        // RB-007.4: through the lock layer, so shared handles' busy reads
856        // can't keep it out, and a shared caller that times out reads
857        // "compact" in the lock-info note.
858        let old = Mnemo::open_exclusive_secret(&real, secret, "compact", opts, true)?;
859        // RB-007.8 item 8: with the exclusive lock held no compaction of
860        // this file is running, so temp files left by dead ones can go.
861        remove_dead_compaction_temps(&real);
862        let dims = old.dimensions;
863        // 12.3 (the keys note, section 1, rule 6): never lower the
864        // parameters, and raise a passphrase file that's below the floor.
865        let kdf = if kind == SecretKind::Passphrase && old.kdf.below_passphrase_floor() {
866            old.kdf.for_passphrase()
867        } else {
868            old.kdf
869        };
870        // 12.3 (the salt rule): the new file keeps the salt, so whatever
871        // found the secret for the old file (a keychain entry) finds it.
872        let salt = old.header.salt;
873        let config = MnemoConfig {
874            dimensions: dims,
875            kdf,
876            // FMT-9f: compaction writes v9 (upgrading a v8 file),
877            // unless `compact_file_to` asked for v8. A v9 file that
878            // keeps float vectors keeps them.
879            format_version,
880            store_float_vectors: old.float_vectors(),
881            ..Default::default()
882        };
883        // v0.14.0 step 1: a new file under a name no other compaction shares.
884        let (tmp, new) = create_temp(&real, "compact", secret, kind, config, Some(salt))?;
885        // Created owner-only; it keeps the old file's mode.
886        if let Err(e) = copy_mode(old.pager.file(), new.pager.file()) {
887            drop(new);
888            let _ = std::fs::remove_file(&tmp);
889            return Err(e);
890        }
891        // `compact_into` releases the old file's lock only after the rename.
892        let result = Self::compact_into(old, new, &real, &tmp, |old, new| old.copy_live_into(new));
893        if result.is_err() {
894            let _ = std::fs::remove_file(&tmp);
895        }
896        result
897    }
898
899    /// Write the compacted copy of `old` to `tmp`, then rename it over
900    /// `path`, all while `old` holds the old file's exclusive lock.
901    ///
902    /// v0.14.0 step 1 (12.4's design note, "Gaps in today's code"): the lock
903    /// used to be released before the rename. A process that opened the
904    /// path in that window locked the old file and wrote to it, and the
905    /// rename then replaced it, losing those writes. Holding the lock
906    /// through the rename and the directory sync means a process either
907    /// waits on the old file's lock (and then finds, by `open_with_lock`'s
908    /// identity check, that the path names the new file) or opens the new
909    /// file.
910    ///
911    /// **Windows and the toolchain.** Renaming over a file another handle
912    /// holds open and locked works on Linux and macOS with every toolchain,
913    /// and on Windows when the engine is built with Rust 1.85 or later,
914    /// whose `std::fs::rename` uses POSIX semantics. Built with 1.75 to 1.84
915    /// (the MSRV is 1.75), Windows refuses it with "Access is denied" (os
916    /// error 5) while the old file is open, so the rename is retried once
917    /// after the old handle is closed, as before v0.14.0. That leaves the
918    /// old window open for those builds; each commit's identity check
919    /// (`ensure_not_replaced`) then turns nearly every write that would have
920    /// been lost into `Replaced`. (Probed on CI, 2026-11-13: 1.75 to 1.84
921    /// refuse, 1.85 to 1.100 succeed; the released binaries are built with
922    /// stable.)
923    fn compact_into<R>(
924        mut old: Mnemo,
925        mut new: Mnemo,
926        path: &Path,
927        tmp: &Path,
928        copy: impl FnOnce(&mut Mnemo, &mut Mnemo) -> Result<R>,
929    ) -> Result<R> {
930        let report = copy(&mut old, &mut new)?;
931
932        new.flush()?;
933        // The new file is closed before the rename: Windows won't rename a
934        // file this process still has open.
935        drop(new);
936
937        match std::fs::rename(tmp, path) {
938            Ok(()) => {}
939            #[cfg(windows)]
940            Err(e) if e.kind() == std::io::ErrorKind::PermissionDenied => {
941                // A toolchain before 1.85: see above.
942                drop(old);
943                std::fs::rename(tmp, path)?;
944                return Ok(report);
945            }
946            Err(e) => return Err(MnemoError::Io(e)),
947        }
948        // Make the rename durable before the old file's lock goes.
949        #[cfg(unix)]
950        if let Some(parent) = path.parent() {
951            let parent = if parent.as_os_str().is_empty() {
952                Path::new(".")
953            } else {
954                parent
955            };
956            if let Ok(dir) = File::open(parent) {
957                let _ = dir.sync_all();
958                crate::bump_fsync();
959            }
960        }
961        // Only now does the old file's lock go.
962        drop(old);
963        Ok(report)
964    }
965
966    /// 12.3 step 4 (v0.15): **embed every live memory again with
967    /// `embedder`** and record its identity, so a file moves to a new model
968    /// (or a new query convention, or a new width) in one step.
969    ///
970    /// It's a compaction with new vectors: it runs under the file's
971    /// exclusive lock (waiting 10 s for it, like [`Mnemo::compact_file`]),
972    /// writes a temp file beside the brain with the same salt and KDF
973    /// parameters, and renames it over the brain while the lock is held. An
974    /// interrupted run leaves the old file whole (and a temp file the next
975    /// compaction or reembed removes), so the file never opens with two
976    /// identities or mixed vectors.
977    ///
978    /// - Every live memory is embedded with [`Embedder::embed`] (the
979    ///   document side), in batches of 64: facts, superseded versions (so
980    ///   history reads stay consistent), essence items and thread turns.
981    ///   A manifest and a thread link (`_mnemo.kind = "thread_link"`) keep
982    ///   their sentinel vector, at the new width, and aren't embedded.
983    /// - The new identity is recorded in the `_mnemo.embedder` entry and in
984    ///   the manifest's `metadata.embedder`.
985    /// - The result cache is copied as compaction copies it; a semantic
986    ///   entry whose vector has the old width is dropped.
987    /// - The ANN index is rebuilt if there was one, and the keyword index
988    ///   with the current tokenizer (TOK-1). Snapshot history is dropped,
989    ///   as by compaction.
990    pub fn reembed_file(
991        path: &str,
992        passphrase: &str,
993        embedder: &dyn Embedder,
994    ) -> Result<ReembedReport> {
995        Self::reembed_file_secret(
996            path,
997            passphrase.as_bytes(),
998            SecretKind::Passphrase,
999            embedder,
1000            &crate::SharedOptions::default(),
1001        )
1002    }
1003
1004    /// [`Mnemo::reembed_file`] for a raw-key database.
1005    pub fn reembed_file_with_key(
1006        path: &str,
1007        key: &[u8; crypto::KEY_LEN],
1008        embedder: &dyn Embedder,
1009    ) -> Result<ReembedReport> {
1010        Self::reembed_file_secret(
1011            path,
1012            key,
1013            SecretKind::Key,
1014            embedder,
1015            &crate::SharedOptions::default(),
1016        )
1017    }
1018
1019    /// [`Mnemo::reembed_file`] with the lock layer's options (how long to
1020    /// wait for the lock, a cancel), as [`Mnemo::compact_file_waiting`].
1021    pub fn reembed_file_waiting(
1022        path: &str,
1023        passphrase: &str,
1024        embedder: &dyn Embedder,
1025        opts: crate::SharedOptions,
1026    ) -> Result<ReembedReport> {
1027        Self::reembed_file_secret(
1028            path,
1029            passphrase.as_bytes(),
1030            SecretKind::Passphrase,
1031            embedder,
1032            &opts,
1033        )
1034    }
1035
1036    /// [`Mnemo::reembed_file_waiting`] for a raw-key database.
1037    pub fn reembed_file_waiting_with_key(
1038        path: &str,
1039        key: &[u8; crypto::KEY_LEN],
1040        embedder: &dyn Embedder,
1041        opts: crate::SharedOptions,
1042    ) -> Result<ReembedReport> {
1043        Self::reembed_file_secret(path, key, SecretKind::Key, embedder, &opts)
1044    }
1045
1046    fn reembed_file_secret(
1047        path: &str,
1048        secret: &[u8],
1049        kind: SecretKind,
1050        embedder: &dyn Embedder,
1051        opts: &crate::SharedOptions,
1052    ) -> Result<ReembedReport> {
1053        let to = embedder.identity();
1054        if to.dimensions == 0 {
1055            return Err(MnemoError::Invalid(
1056                "reembed: the embedder's identity has 0 dimensions".into(),
1057            ));
1058        }
1059        let real = std::fs::canonicalize(path)?;
1060        let mut old = Mnemo::open_exclusive_secret(&real, secret, "reembed", opts, true)?;
1061        remove_dead_temps(&real, "reembed");
1062        let from = old.recorded_identity()?;
1063        let kdf = if kind == SecretKind::Passphrase && old.kdf.below_passphrase_floor() {
1064            old.kdf.for_passphrase()
1065        } else {
1066            old.kdf
1067        };
1068        let salt = old.header.salt;
1069        let config = MnemoConfig {
1070            dimensions: to.dimensions,
1071            kdf,
1072            format_version: 0,
1073            store_float_vectors: old.float_vectors(),
1074            ..Default::default()
1075        };
1076        let (tmp, new) = create_temp(&real, "reembed", secret, kind, config, Some(salt))?;
1077        if let Err(e) = copy_mode(old.pager.file(), new.pager.file()) {
1078            drop(new);
1079            let _ = std::fs::remove_file(&tmp);
1080            return Err(e);
1081        }
1082        let result = Self::compact_into(old, new, &real, &tmp, |old, new| {
1083            old.copy_reembedded_into(new, embedder, from.clone(), to.clone())
1084        });
1085        if result.is_err() {
1086            let _ = std::fs::remove_file(&tmp);
1087        }
1088        result
1089    }
1090
1091    /// [`Mnemo::reembed_file`]'s copy: every live memory with a new vector,
1092    /// the new identity, and the cache.
1093    fn copy_reembedded_into(
1094        &mut self,
1095        new: &mut Mnemo,
1096        embedder: &dyn Embedder,
1097        from: Option<embed::EmbedderIdentity>,
1098        to: embed::EmbedderIdentity,
1099    ) -> Result<ReembedReport> {
1100        const BATCH: usize = 64;
1101        new.turn_compress_after = 0;
1102        let want_index = self.has_index();
1103        let now = memory::now_secs();
1104        let dims = to.dimensions;
1105        let mut live: Vec<Memory> = self
1106            .memories()?
1107            .into_iter()
1108            .filter(|m| !m.is_expired(now))
1109            .collect();
1110        let mut manifests = 0;
1111        let mut thread_links = 0;
1112        for m in live.iter_mut() {
1113            if embed::read_identity_from_metadata(&m.metadata).is_some() {
1114                embed::write_identity_into_metadata(&mut m.metadata, &to);
1115            }
1116        }
1117        // R-008.3: a manifest and a thread link carry a sentinel vector, not
1118        // an embedding of their text.
1119        let is_link = |m: &Memory| {
1120            m.metadata.get(MNEMO_META_KIND).and_then(|v| v.as_str())
1121                == Some(MNEMO_META_KIND_THREAD_LINK)
1122        };
1123        let to_embed: Vec<usize> = (0..live.len())
1124            .filter(|&i| !is_manifest(&live[i]) && !is_link(&live[i]))
1125            .collect();
1126        for chunk in to_embed.chunks(BATCH) {
1127            let texts: Vec<&str> = chunk.iter().map(|&i| live[i].content.as_str()).collect();
1128            let vectors = embedder.embed(&texts)?;
1129            if vectors.len() != texts.len() {
1130                return Err(MnemoError::Invalid(format!(
1131                    "reembed: the embedder returned {} vectors for {} texts",
1132                    vectors.len(),
1133                    texts.len()
1134                )));
1135            }
1136            for (&i, v) in chunk.iter().zip(vectors) {
1137                if v.len() != dims {
1138                    return Err(MnemoError::DimensionMismatch {
1139                        expected: dims,
1140                        got: v.len(),
1141                    });
1142                }
1143                if embed::is_zero_vector(&v) {
1144                    return Err(MnemoError::Invalid(format!(
1145                        "reembed: the embedder returned an all-zero vector for memory {}",
1146                        live[i].id
1147                    )));
1148                }
1149                live[i].vector = v;
1150            }
1151        }
1152        let memories = live.len();
1153        for mut m in live {
1154            if is_manifest(&m) {
1155                m.vector = Memory::scaffold_manifest(dims).vector;
1156                manifests += 1;
1157            } else if is_link(&m) {
1158                m.vector = new.sentinel_vector();
1159                thread_links += 1;
1160            }
1161            new.put_existing(m)?;
1162        }
1163        if want_index && memories > 0 {
1164            new.build_index_with(IndexConfig::default())?;
1165        }
1166        let mut cache_entries = 0;
1167        let mut cache_dropped = 0;
1168        for pos in 0..self.cache_dir.len() {
1169            let (keep, reserved, wrong_width) = {
1170                let e = &self.cache_dir[pos];
1171                (
1172                    !e.deleted && !e.is_expired(now),
1173                    embed::is_reserved_namespace(&e.namespace),
1174                    e.vector.as_ref().is_some_and(|v| v.len() != dims),
1175                )
1176            };
1177            if !keep || (reserved && self.cache_dir[pos].namespace == embed::IDENTITY_NAMESPACE) {
1178                continue;
1179            }
1180            if wrong_width {
1181                cache_dropped += 1;
1182                continue;
1183            }
1184            new.copy_cache_entry_from(self, pos)?;
1185            if !reserved {
1186                cache_entries += 1;
1187            }
1188        }
1189        new.record_identity(&to)?;
1190        Ok(ReembedReport {
1191            memories: memories - manifests - thread_links,
1192            manifests,
1193            thread_links,
1194            cache_entries,
1195            cache_dropped,
1196            from,
1197            to,
1198        })
1199    }
1200
1201    /// Copy one live cache entry from `old` at directory position
1202    /// `pos` into `self` (a freshly-created compaction target).
1203    /// Decrypts the old record body, substitutes the old *directory*'s
1204    /// `accessed_at` + `access_count` (the record body's copies are
1205    /// stale as of v5's directory-only hit path), re-encodes, writes
1206    /// a fresh record run, and pushes a matching directory entry.
1207    ///
1208    /// Caller is expected to have already filtered out tombstoned /
1209    /// expired entries and to call this on a read-write handle.
1210    pub(crate) fn copy_cache_entry_from(&mut self, old: &mut Mnemo, pos: usize) -> Result<()> {
1211        if self.read_only {
1212            return Err(MnemoError::ReadOnly);
1213        }
1214        // Snapshot the directory fields we need, so the borrow on
1215        // `old.cache_dir` ends before we touch `old.pager`.
1216        let (start_page, page_count, len, dir_accessed_at, dir_access_count, old_gen) = {
1217            let e = &old.cache_dir[pos];
1218            (
1219                e.start_page,
1220                e.page_count,
1221                e.len,
1222                e.accessed_at,
1223                e.access_count,
1224                e.gen,
1225            )
1226        };
1227
1228        // Read + decrypt the old record body.
1229        let buf = old.read_record_run(start_page, page_count, len, old_gen)?;
1230        let mut entry: CacheEntry =
1231            rmp_serde::from_slice(&buf).map_err(|e| MnemoError::Serialize(e.to_string()))?;
1232
1233        // Overwrite the stale body copies with the directory's live
1234        // values before persisting. Everything else — `created_at`,
1235        // `ttl_secs`, `cost_hint_ms`, `key_preview`, `size_bytes`,
1236        // `content_type`, `value`, `vector`, `model` — stays as the
1237        // body already recorded it.
1238        entry.accessed_at = dir_accessed_at;
1239        entry.access_count = dir_access_count;
1240
1241        let bytes =
1242            rmp_serde::to_vec(&entry).map_err(|err| MnemoError::Serialize(err.to_string()))?;
1243        let (new_start_page, new_page_count, gen) = self.write_record(&bytes)?;
1244
1245        let dir_entry = CacheDirectoryEntry {
1246            namespace: entry.namespace.clone(),
1247            key_hash: entry.key_hash,
1248            start_page: new_start_page,
1249            page_count: new_page_count,
1250            len: bytes.len() as u32,
1251            deleted: false,
1252            accessed_at: dir_accessed_at,
1253            access_count: dir_access_count,
1254            size_bytes: entry.size_bytes,
1255            ttl_secs: entry.ttl_secs,
1256            created_at: entry.created_at,
1257            vector: entry.vector.clone(),
1258            model: entry.model.clone(),
1259            gen,
1260        };
1261        let new_pos = self.cache_dir.len();
1262        self.cache_idx
1263            .insert(&entry.namespace, entry.key_hash, new_pos);
1264        self.cache_dir.push(dir_entry);
1265        self.dirty_cache = true;
1266        Ok(())
1267    }
1268}
1269
1270/// SEC-2 (RB-001.30): overwrite every occurrence of `needle` in the file
1271/// with zeros, then sync. Rekey uses it for stale copies of the old header's
1272/// key wrap: they're plaintext (page 0 is sealed, not encrypted) and can sit
1273/// in WAL regions that were grown, relocated or freed. Every other page is
1274/// AES-GCM ciphertext, where a 76-byte random match doesn't happen.
1275pub(super) fn scrub_bytes(file: &mut std::fs::File, needle: &[u8]) -> Result<usize> {
1276    if needle.is_empty() {
1277        return Ok(0);
1278    }
1279    let len = file.seek(SeekFrom::End(0))?;
1280    let chunk = 1 << 20;
1281    let overlap = needle.len() - 1;
1282    let mut hits: Vec<u64> = Vec::new();
1283    let mut buf = vec![0u8; chunk + overlap];
1284    let mut pos: u64 = 0;
1285    while pos < len {
1286        let want = (chunk as u64 + overlap as u64).min(len - pos) as usize;
1287        file.seek(SeekFrom::Start(pos))?;
1288        file.read_exact(&mut buf[..want])?;
1289        let mut i = 0;
1290        while i + needle.len() <= want {
1291            if buf[i..i + needle.len()] == *needle {
1292                hits.push(pos + i as u64);
1293                i += needle.len();
1294            } else {
1295                i += 1;
1296            }
1297        }
1298        pos += chunk as u64;
1299    }
1300    hits.dedup();
1301    let zeros = vec![0u8; needle.len()];
1302    for off in &hits {
1303        file.seek(SeekFrom::Start(*off))?;
1304        file.write_all(&zeros)?;
1305    }
1306    if !hits.is_empty() {
1307        file.sync_all()?;
1308        crate::bump_fsync();
1309    }
1310    Ok(hits.len())
1311}
1312
1313/// RB-007.8 item 8: remove the temp files that compactions of `brain` left
1314/// when their process died: in `brain`'s directory, regular files named
1315/// exactly `<brain's name>.compact-<pid>-<n>-<8 hex digits>.tmp` (the
1316/// pattern [`temp_sibling`] makes) whose pid isn't this process and no
1317/// longer runs, and, on Unix, that this user owns: in a shared directory
1318/// another user's file of the same name is left alone. Nothing else is
1319/// touched. Best effort; returns how many went.
1320///
1321/// Call it only with `brain`'s exclusive lock held: a live compaction holds
1322/// that lock for as long as its temp file exists. Off Unix no process
1323/// check is available, so nothing is removed.
1324pub(crate) fn remove_dead_compaction_temps(brain: &Path) -> usize {
1325    remove_dead_temps(brain, "compact")
1326}
1327
1328/// [`remove_dead_compaction_temps`] for the temps of another operation
1329/// (`what`, as [`temp_sibling`] names them): 12.3's `init` uses "init".
1330pub(crate) fn remove_dead_temps(brain: &Path, what: &str) -> usize {
1331    let (Some(dir), Some(name)) = (brain.parent(), brain.file_name()) else {
1332        return 0;
1333    };
1334    let prefix = format!("{}.{what}-", name.to_string_lossy());
1335    let Ok(entries) = std::fs::read_dir(dir) else {
1336        return 0;
1337    };
1338    let mut removed = 0;
1339    for entry in entries.flatten() {
1340        let file_name = entry.file_name();
1341        let Some(pid) = file_name.to_str().and_then(|n| dead_temp_pid(n, &prefix)) else {
1342            continue;
1343        };
1344        if pid == std::process::id() || crate::shared::pid_alive(pid) {
1345            continue;
1346        }
1347        let path = entry.path();
1348        let Ok(meta) = std::fs::symlink_metadata(&path) else {
1349            continue;
1350        };
1351        if !meta.file_type().is_file() || !owned_by_us(&meta) {
1352            continue;
1353        }
1354        if std::fs::remove_file(&path).is_ok() {
1355            removed += 1;
1356        }
1357    }
1358    removed
1359}
1360
1361/// Whether this process's user owns the file (Unix). Elsewhere, true: the
1362/// caller removes nothing off Unix anyway.
1363#[cfg(unix)]
1364fn owned_by_us(meta: &std::fs::Metadata) -> bool {
1365    use std::os::unix::fs::MetadataExt;
1366    #[cfg(any(test, feature = "failpoints"))]
1367    if let Some(uid) = crate::__failpoints::euid_override() {
1368        return meta.uid() == uid;
1369    }
1370    meta.uid() == rustix::process::geteuid().as_raw()
1371}
1372
1373#[cfg(not(unix))]
1374fn owned_by_us(_meta: &std::fs::Metadata) -> bool {
1375    true
1376}
1377
1378/// The pid in a compaction temp name `<prefix><pid>-<n>-<8 hex>.tmp`, if the
1379/// name is exactly that.
1380fn dead_temp_pid(name: &str, prefix: &str) -> Option<u32> {
1381    let rest = name.strip_prefix(prefix)?.strip_suffix(".tmp")?;
1382    let mut parts = rest.split('-');
1383    let (pid, n, nanos) = (parts.next()?, parts.next()?, parts.next()?);
1384    let digits = |s: &str| !s.is_empty() && s.bytes().all(|b| b.is_ascii_digit());
1385    let hex8 = nanos.len() == 8 && nanos.bytes().all(|b| b.is_ascii_hexdigit());
1386    if parts.next().is_some() || !digits(pid) || !digits(n) || !hex8 {
1387        return None;
1388    }
1389    pid.parse().ok()
1390}
1391
1392/// The phrase REKEY-1's one exception carries; `mnemo rekey` and
1393/// [`crate::keys::rekey_stored`] look for it.
1394const EITHER_SECRET: &str = "may open with either secret";
1395
1396/// A new, locked temp file next to `dest` (`Create::New`, so nothing already
1397/// at the name, a symlink included, is ever written through), under a fresh
1398/// name for each attempt while names are taken. RB-006.10.
1399pub(crate) fn create_temp(
1400    dest: &Path,
1401    what: &str,
1402    secret: &[u8],
1403    kind: SecretKind,
1404    config: MnemoConfig,
1405    salt: Option<[u8; crypto::SALT_LEN]>,
1406) -> Result<(PathBuf, Mnemo)> {
1407    let mut attempt = 0;
1408    loop {
1409        let tmp = temp_sibling(dest, what);
1410        match Mnemo::create_secret_new(&tmp, secret, kind, config, salt) {
1411            Ok(new) => return Ok((tmp, new)),
1412            Err(e) if is_name_taken(&e) && attempt + 1 < TEMP_NAME_ATTEMPTS => attempt += 1,
1413            Err(e) => return Err(e),
1414        }
1415    }
1416}