mnemo_engine/store/open.rs
1//! Create, open (read-write, read-only, raw key) and the handle's basic accessors and per-handle settings.
2
3use super::*;
4
5/// Chunk size for the torn-header scan's streaming read of the file.
6/// We walk the file 1 MiB at a time looking for 8 KiB-aligned `MWAL`
7/// magic. There is **no** cap on how far a given candidate's frame
8/// chain may extend — a 1M-vector IVF+PQ relocation can push a single
9/// transaction well past 16 MiB — so each candidate is read frame by
10/// frame via [`wal::scan_committed_txn`] rather than slurped into a
11/// bounded buffer. (DUR-1 R-002.1 part 2.)
12const TORN_HEADER_SCAN_CHUNK: usize = 1 << 20; // 1 MiB
13
14/// The header a set of recovery frames would write to page 0, if they hold
15/// one that parses.
16fn page_zero_header(frames: &[(u64, Vec<u8>)]) -> Option<Header> {
17 frames
18 .iter()
19 .find(|(page_no, _)| *page_no == 0)
20 .and_then(|(_, bytes)| Header::from_page(bytes).ok())
21}
22
23/// Whether `h` is a v9 header sealed under `dek` (RB-007.2). The version
24/// matters: `validate_seal` passes any header older than v7, which has no
25/// seal.
26fn sealed_under(h: &Header, dek: &[u8; crypto::KEY_LEN]) -> bool {
27 h.version >= VERSION_V9 && h.validate_seal(dek).is_ok()
28}
29
30/// Walk the file for every 8 KiB-aligned page whose first four bytes are
31/// `MWAL`, treating each as a candidate WAL-region start, and return the
32/// frames of the highest-`txn_id` committed transaction whose header
33/// frame fully validates under `passphrase`.
34///
35/// A candidate is accepted only if:
36/// * `wal::scan_committed_txn` returns a committed transaction (chained
37/// CRC verifies, single `txn_id`, `KIND_COMMIT` closes it);
38/// * that transaction contains a `(0, header_page)` frame;
39/// * the header parses with `wal_start == candidate_page_no` and
40/// `wal_seq == txn_id` (so a frame cannot forge a region it was not
41/// written into);
42/// * deriving the KEK from `passphrase + header.salt`, unwrapping the
43/// wrapped DEK, and validating the header seal all succeed.
44///
45/// Returns `Ok(None)` when the scan finds no valid candidate — the
46/// caller is expected to propagate the original parse error so
47/// non-mnemo files still report `BadMagic`.
48///
49/// DUR-1 (R-002.1 part 2): there is no fixed scan window — the whole
50/// file is scanned. Reading at page granularity (not byte granularity)
51/// is sufficient because every WAL region in the format is page-aligned
52/// and every WAL commit writes its first frame at the region's start.
53fn scan_header_recovery_candidates(
54 file: &mut File,
55 secret: &[u8],
56 held_dek: Option<&[u8; crypto::KEY_LEN]>,
57) -> Result<Option<Vec<wal::Frame>>> {
58 let file_len = file.seek(SeekFrom::End(0))?;
59 if file_len < PAGE_SIZE as u64 * 2 {
60 return Ok(None); // No room for even one WAL page past the header.
61 }
62
63 let mut best: Option<(u64, Vec<wal::Frame>)> = None;
64 let page_sz = PAGE_SIZE as u64;
65 let mut offset = page_sz; // Skip page 0.
66
67 // Stream through the file in 1 MiB chunks, scanning each chunk for
68 // page-aligned `MWAL` candidates. Each candidate's frame chain is
69 // then read incrementally via `wal::scan_committed_txn`, which has
70 // no window cap — a 1M-vector IVF+PQ relocation can span tens of
71 // MiB in one committed transaction.
72 let mut chunk = vec![0u8; TORN_HEADER_SCAN_CHUNK];
73 while offset < file_len {
74 let want = (TORN_HEADER_SCAN_CHUNK as u64).min(file_len - offset) as usize;
75 chunk.resize(want, 0);
76 file.seek(SeekFrom::Start(offset))?;
77 let mut filled = 0;
78 while filled < want {
79 match file.read(&mut chunk[filled..])? {
80 0 => break,
81 n => filled += n,
82 }
83 }
84 if filled == 0 {
85 break;
86 }
87 let mut inner = 0usize;
88 while inner + 4 <= filled {
89 if &chunk[inner..inner + 4] == wal::FRAME_MAGIC_BYTES {
90 let page_no = (offset + inner as u64) / page_sz;
91 if let Some((txn_id, frames)) =
92 try_recover_candidate(file, page_no, secret, held_dek)?
93 {
94 let keep = match &best {
95 None => true,
96 Some((id, _)) => txn_id > *id,
97 };
98 if keep {
99 best = Some((txn_id, frames));
100 }
101 }
102 }
103 inner += PAGE_SIZE;
104 }
105 offset += filled as u64;
106 }
107
108 Ok(best.map(|(_, f)| f))
109}
110
111/// Try to recover a committed transaction from `candidate_page_no`,
112/// returning `(txn_id, frames)` only if the candidate's
113/// `(0, header_page)` frame fully validates: the frame chain commits
114/// (via [`wal::scan_committed_txn`]'s streaming reader), the header
115/// parses with `wal_start == candidate_page_no` **and**
116/// `wal_seq == txn_id`, the DEK unwraps under `passphrase`, and the
117/// AES-GCM seal validates. DUR-1 (R-002.1 item 2) requires the
118/// `wal_seq == txn_id` check and no fixed scan window — the streaming
119/// reader handles transactions of arbitrary size.
120fn try_recover_candidate(
121 file: &mut File,
122 candidate_page_no: u64,
123 secret: &[u8],
124 held_dek: Option<&[u8; crypto::KEY_LEN]>,
125) -> Result<Option<(u64, Vec<wal::Frame>)>> {
126 let Some((txn_id, frames)) = wal::scan_committed_txn(file, candidate_page_no)? else {
127 return Ok(None);
128 };
129
130 // Must contain a header frame at page 0.
131 let header_bytes = match frames.iter().find(|(p, _)| *p == 0) {
132 Some((_, b)) => b,
133 None => return Ok(None),
134 };
135 if header_bytes.len() != PAGE_SIZE {
136 return Ok(None);
137 }
138 let mut hbuf = [0u8; PAGE_SIZE];
139 hbuf.copy_from_slice(header_bytes);
140 let candidate = match Header::from_page(&hbuf) {
141 Ok(h) => h,
142 Err(_) => return Ok(None),
143 };
144
145 // The frame's claimed region must match where we found it, and the
146 // header frame's `wal_seq` must match the committed transaction's
147 // `txn_id` — this is what pins the frame to the region it was
148 // written into and keeps a stale checkpointed-and-copied frame
149 // from being accepted as a new region's root.
150 if candidate.wal_start != candidate_page_no || candidate.wal_seq != txn_id {
151 return Ok(None);
152 }
153 // RB-007.2: a shared handle's data key, when the candidate's seal
154 // verifies under it, without the KDF.
155 if let Some(held) = held_dek {
156 if sealed_under(&candidate, held) {
157 return Ok(Some((txn_id, frames)));
158 }
159 }
160 // DEK unwrap under this candidate's own salt + KDF params.
161 let kdf = KdfParams {
162 m_cost: candidate.m_cost,
163 t_cost: candidate.t_cost,
164 p_cost: candidate.p_cost,
165 };
166 let kek = match crypto::derive_kek(secret, &candidate.salt, kdf) {
167 Ok(k) => k,
168 Err(_) => return Ok(None),
169 };
170 let dek = match crypto::unwrap_dek(&kek, &candidate.dek_nonce, &candidate.wrapped_dek) {
171 Ok(d) => d,
172 Err(_) => return Ok(None),
173 };
174 if candidate.validate_seal(&dek).is_err() {
175 return Ok(None);
176 }
177
178 Ok(Some((txn_id, frames)))
179}
180
181impl Mnemo {
182 /// Create a brand-new encrypted database at `path`.
183 pub fn create(
184 path: impl Into<PathBuf>,
185 passphrase: &str,
186 config: MnemoConfig,
187 ) -> Result<Mnemo> {
188 Self::create_secret(path, passphrase.as_bytes(), config)
189 }
190
191 /// PERF-1 step 7 (R-004.5): create a database unlocked by a **raw
192 /// 256-bit key** instead of a passphrase, for keys held by an OS
193 /// keychain or keystore. The key must be uniformly random (see
194 /// [`Mnemo::generate_key`]); it gets no stretching. The header records
195 /// minimal Argon2id parameters ([`KdfParams::raw_key`]), so opening
196 /// costs microseconds instead of tens of milliseconds. `config.kdf` is
197 /// ignored. No format change: the key goes through the same KEK
198 /// derivation a passphrase does.
199 pub fn create_with_key(
200 path: impl Into<PathBuf>,
201 key: &[u8; crypto::KEY_LEN],
202 config: MnemoConfig,
203 ) -> Result<Mnemo> {
204 let config = MnemoConfig {
205 kdf: KdfParams::raw_key(),
206 ..config
207 };
208 Self::create_secret(path, key, config)
209 }
210
211 /// A fresh random 256-bit key for [`Mnemo::create_with_key`], from the
212 /// thread-local CSPRNG (seeded from the OS). Store it in a keychain;
213 /// it can't be recovered from the file.
214 pub fn generate_key() -> [u8; crypto::KEY_LEN] {
215 *crypto::random_dek()
216 }
217
218 pub(super) fn create_secret(
219 path: impl Into<PathBuf>,
220 secret: &[u8],
221 config: MnemoConfig,
222 ) -> Result<Mnemo> {
223 Self::create_secret_as(path.into(), secret, config, Create::Truncate)
224 }
225
226 /// [`Mnemo::create_secret`] for a temp file: `Create::New` refuses a
227 /// name that already exists (a symlink included), with `AlreadyExists`.
228 pub(super) fn create_secret_new(
229 path: &Path,
230 secret: &[u8],
231 config: MnemoConfig,
232 ) -> Result<Mnemo> {
233 Self::create_secret_as(path.to_path_buf(), secret, config, Create::New)
234 }
235
236 fn create_secret_as(
237 path: PathBuf,
238 secret: &[u8],
239 config: MnemoConfig,
240 create: Create,
241 ) -> Result<Mnemo> {
242 if config.dimensions == 0 {
243 return Err(MnemoError::Invalid("dimensions must be > 0".into()));
244 }
245 let version = resolve_format_version(config.format_version)?;
246 if version >= VERSION_V9 && veccol::per_page(config.dimensions) == 0 {
247 return Err(MnemoError::Invalid(format!(
248 "a v9 file supports up to {} dimensions for now (one fp16 vector \
249 must fit a page); vector entries that span pages are not supported yet",
250 (PAYLOAD - veccol::PAGE_HEADER - 4) / 2
251 )));
252 }
253 let flags = if version >= VERSION_V9 && config.store_float_vectors {
254 FLAG_ENCRYPTED | FLAG_FLOAT_VECTORS
255 } else {
256 FLAG_ENCRYPTED
257 };
258 // Exclusive lock covers the write path; second writer sees `Locked`.
259 let file = open_with_lock(&path, /*write=*/ true, create)?;
260
261 let salt = crypto::random_salt();
262 let dek = crypto::random_dek();
263 let kek = crypto::derive_kek(secret, &salt, config.kdf)?;
264 let dek_nonce = crypto::random_nonce();
265 let wrapped = crypto::wrap_dek(&kek, &dek_nonce, &dek)?;
266 if wrapped.len() != WRAPPED_DEK_LEN {
267 return Err(MnemoError::Crypto("unexpected wrapped DEK length".into()));
268 }
269 let mut wrapped_dek = [0u8; WRAPPED_DEK_LEN];
270 wrapped_dek.copy_from_slice(&wrapped);
271
272 let mut header = Header {
273 version,
274 page_size: PAGE_SIZE as u32,
275 flags,
276 dimensions: config.dimensions as u32,
277 created_at: memory::now_secs(),
278 write_counter: 0,
279 // Page 0 is the header; pages 1..=wal_pages are the WAL;
280 // record/catalog/index pages start after it. Initial WAL size is
281 // chosen per-config (default 8 pages = 64 KiB) and clamped to a
282 // sane floor so the first transaction can always commit.
283 next_page: 1 + config.wal_pages_initial.max(MIN_WAL_PAGES),
284 catalog_start: 0,
285 catalog_pages: 0,
286 catalog_len: 0,
287 vector_count: 0,
288 m_cost: config.kdf.m_cost,
289 t_cost: config.kdf.t_cost,
290 p_cost: config.kdf.p_cost,
291 salt,
292 dek_nonce,
293 wrapped_dek,
294 index_start: 0,
295 index_pages: 0,
296 index_len: 0,
297 wal_start: 1,
298 wal_pages: config.wal_pages_initial.max(MIN_WAL_PAGES),
299 wal_seq: 0,
300 manifest_start: 0,
301 manifest_pages: 0,
302 manifest_len: 0,
303 // Regenerated per write by `apply_seal`; `None` until we seal.
304 seal_nonce: None,
305 seal_tag: None,
306 // v8+ result-cache pointer — zero-initialized on create,
307 // populated only when `cache_put` runs at least once and
308 // the next `flush` seals a directory run.
309 cache_start: 0,
310 cache_pages: 0,
311 cache_len: 0,
312 // FMT-9 v9 fields; all zero on a fresh file (and unused on v8).
313 catalog_gen: 0,
314 index_gen: 0,
315 manifest_gen: 0,
316 cache_gen: 0,
317 free: RunPtr::default(),
318 keys: RunPtr::default(),
319 bm25: RunPtr::default(),
320 index_delta: RunPtr::default(),
321 catalog_delta: RunPtr::default(),
322 };
323
324 let mut pager = Pager::new(file, dek, 0, version);
325 let mut page = header.to_page();
326 header.apply_seal(&mut page, pager.dek())?;
327 pager.write_raw(0, &page)?;
328 pager.sync()?;
329
330 Ok(Mnemo {
331 pager,
332 header,
333 catalog: Vec::new(),
334 index: IdIndex::default(),
335 path: absolute(path),
336 dimensions: config.dimensions,
337 kdf: config.kdf,
338 dirty_catalog: false,
339 ann: None,
340 dirty_index: false,
341 index_persist: IndexPersist::default(),
342 ann_unloaded: false,
343 manifest: Vec::new(),
344 // FMT-9c: v9 keeps fewer snapshots by default: with page reuse,
345 // retention is what pins pages.
346 max_snapshots: if version >= VERSION_V9 && config.max_snapshots == DEFAULT_MAX_SNAPSHOTS
347 {
348 V9_DEFAULT_MAX_SNAPSHOTS
349 } else {
350 config.max_snapshots
351 },
352 read_only: false,
353 cache_dir: Vec::new(),
354 cache_idx: DirectoryIndex::new(),
355 dirty_cache: false,
356 cache_budgets: std::collections::HashMap::new(),
357 cache_flush_policy: CacheFlushPolicy::default(),
358 cache_batch: BatchState::default(),
359 cache_counters: CounterState::default(),
360 embedder: None,
361 fsyncs_last_flush: 0,
362 agent_set: None,
363 key_index: None,
364 superseded: None,
365 layers: None,
366 rekey_cleanup: None,
367 access_delta: None,
368 commit_point: None,
369 lock_note: None,
370 window_committed: false,
371 shared: false,
372 turn_compress_after: DEFAULT_TURN_COMPRESS_AFTER_SECS,
373 free: FreeList::default(),
374 pending_free: Vec::new(),
375 free_dirty: false,
376 key_slots: KeySlotRun::default(),
377 dirty_keys: false,
378 gen_leased: false,
379 col: VecColumn::default(),
380 recs: RecPages::default(),
381 bm25: Bm25State::default(),
382 kw: None,
383 index_layout_override: None,
384 dirty_rows: std::collections::BTreeSet::new(),
385 delta_chain: Vec::new(),
386 delta_rows: 0,
387 catalog_rewrite: false,
388 index_chain: Vec::new(),
389 index_chain_ops: 0,
390 index_rebase: false,
391 index_delta_ops: None,
392 snapshot_cap_explicit: config.max_snapshots != DEFAULT_MAX_SNAPSHOTS,
393 direct_writes: false,
394 vector_cache_budget: DEFAULT_VECTOR_CACHE_BUDGET,
395 })
396 }
397
398 /// Open an existing database. A wrong passphrase fails cleanly with
399 /// [`MnemoError::WrongPassphrase`]. Takes an OS exclusive lock for
400 /// the lifetime of the returned handle; a second concurrent open
401 /// (from any process) returns [`MnemoError::Locked`].
402 ///
403 /// For a caller that only needs to read (and can tolerate the
404 /// pending-recovery / pending-migration cases documented on
405 /// [`Mnemo::open_read_only`]), prefer that entry point — it takes a
406 /// shared lock and coexists with other readers.
407 pub fn open(path: impl Into<PathBuf>, passphrase: &str) -> Result<Mnemo> {
408 Self::open_secret(path, passphrase.as_bytes())
409 }
410
411 /// PERF-1 step 7: open a database created by [`Mnemo::create_with_key`]
412 /// (or rekeyed with [`Mnemo::rekey_with_key`]) with its raw key.
413 pub fn open_with_key(path: impl Into<PathBuf>, key: &[u8; crypto::KEY_LEN]) -> Result<Mnemo> {
414 Self::open_secret(path, key)
415 }
416
417 pub(super) fn open_secret(path: impl Into<PathBuf>, secret: &[u8]) -> Result<Mnemo> {
418 Self::open_secret_with(path, secret, false)
419 }
420
421 /// The read-write open. `allow_old` lets a file older than v9 open (and
422 /// migrate in memory to v8): only compaction uses it, to read a v4–v8
423 /// file and write the v9 copy. Every public read-write open refuses
424 /// those with [`MnemoError::FormatTooOld`] (the v8 sunset, v0.13).
425 pub(super) fn open_secret_with(
426 path: impl Into<PathBuf>,
427 secret: &[u8],
428 allow_old: bool,
429 ) -> Result<Mnemo> {
430 let path: PathBuf = path.into();
431 let file = open_with_lock(&path, /*write=*/ true, Create::No)?;
432 Self::load_locked(path, file, secret, allow_old)
433 }
434
435 /// Open read-write through 12.4's lock layer, for a long operation that
436 /// must not run alongside shared handles' calls: an import, a rekey, a
437 /// compaction ([`Mnemo::compact_file`] does this itself).
438 ///
439 /// Unlike [`Mnemo::open`], which tries the lock once and fails with
440 /// [`MnemoError::Locked`], this waits up to `opts.deadline` with writer
441 /// preference (the intent lock), so busy shared readers can't keep it
442 /// out. Once it holds the lock it writes a `lock-info` note naming `op`,
443 /// so a shared caller that times out learns what holds the file. The
444 /// handle keeps the exclusive lock until it's dropped or closed, like
445 /// any read-write handle.
446 ///
447 /// Where shared mode isn't available (a network or synced filesystem
448 /// without `opts.allow_network_fs`, or a Windows build made with Rust
449 /// before 1.85) it opens as [`Mnemo::open`] does: one try, no note.
450 pub fn open_exclusive(
451 path: impl AsRef<Path>,
452 passphrase: &str,
453 op: &str,
454 opts: crate::SharedOptions,
455 ) -> Result<Mnemo> {
456 Self::open_exclusive_secret(path.as_ref(), passphrase.as_bytes(), op, &opts, false)
457 }
458
459 /// [`Mnemo::open_exclusive`] for a raw-key file.
460 pub fn open_exclusive_with_key(
461 path: impl AsRef<Path>,
462 key: &[u8; crypto::KEY_LEN],
463 op: &str,
464 opts: crate::SharedOptions,
465 ) -> Result<Mnemo> {
466 Self::open_exclusive_secret(path.as_ref(), key, op, &opts, false)
467 }
468
469 pub(super) fn open_exclusive_secret(
470 path: &Path,
471 secret: &[u8],
472 op: &str,
473 opts: &crate::SharedOptions,
474 allow_old: bool,
475 ) -> Result<Mnemo> {
476 let path = absolute(path.to_path_buf());
477 let (file, note) = lock_exclusive(&path, op, opts)?;
478 let mut db = Self::load_locked(path, file, secret, allow_old)?;
479 db.lock_note = note;
480 Ok(db)
481 }
482
483 /// The read-write open's work on a file its caller already holds
484 /// exclusively: heal a torn header, replay a pending WAL, derive the key
485 /// from `secret`, and build the handle. A shared-mode handle (12.4)
486 /// reloads through it with the call's locked file.
487 pub(super) fn load_locked(
488 path: PathBuf,
489 file: LockedFile,
490 secret: &[u8],
491 allow_old: bool,
492 ) -> Result<Mnemo> {
493 Self::load_locked_with(path, file, secret, None, allow_old)
494 }
495
496 /// [`Mnemo::load_locked`], with a data key the caller already holds (a
497 /// shared handle's, RB-007.2). The torn-header scan accepts a candidate
498 /// whose seal verifies under it, and a pending WAL is replayed with it
499 /// when the header the replay writes verifies under it, without running
500 /// the KDF: a rekey changes the secret, never the data key. Anything
501 /// that doesn't verify under it is checked with `secret`, as an open
502 /// does.
503 pub(super) fn load_locked_with(
504 path: PathBuf,
505 mut file: LockedFile,
506 secret: &[u8],
507 held_dek: Option<&[u8; crypto::KEY_LEN]>,
508 allow_old: bool,
509 ) -> Result<Mnemo> {
510 let mut hbuf = [0u8; PAGE_SIZE];
511 file.seek(SeekFrom::Start(0))?;
512 file.read_exact(&mut hbuf)?;
513 let mut header = match Header::from_page(&hbuf) {
514 Ok(h) => h,
515 Err(e) => {
516 // DUR-1 (R-002.1 part 2+3): page 0 did not parse — any
517 // error is treated as a torn-header candidate, not just
518 // `HeaderChecksum`. The whole 294-byte header lives in
519 // the first 512-byte sector, so the realistic failure is
520 // the sector zeroed or garbled, which returns `BadMagic`
521 // or `UnsupportedVersion` *before* the CRC gate.
522 //
523 // Scan the file for valid committed header-frame WAL
524 // regions, pick the highest `txn_id`, and replay it. If
525 // nothing valid is found, return the *original* error so
526 // non-mnemo files still report `BadMagic`.
527 let scan = scan_header_recovery_candidates(&mut file, secret, held_dek)?;
528 let frames = scan.ok_or(e)?;
529 // RB-003.3: the scan has checked the key; refuse an old
530 // format before healing anything.
531 if !allow_old {
532 if let Some(h) = page_zero_header(&frames) {
533 if h.version < VERSION_V9 {
534 return Err(MnemoError::FormatTooOld { version: h.version });
535 }
536 }
537 }
538 for (page_no, bytes) in &frames {
539 if bytes.len() != PAGE_SIZE {
540 return Err(MnemoError::Invalid("WAL frame is not page-sized".into()));
541 }
542 file.seek(SeekFrom::Start(page_no * PAGE_SIZE as u64))?;
543 file.write_all(bytes)?;
544 }
545 file.sync_all()?;
546 crate::bump_fsync();
547 file.seek(SeekFrom::Start(0))?;
548 file.read_exact(&mut hbuf)?;
549 Header::from_page(&hbuf)?
550 }
551 };
552
553 // Crash recovery: a committed-but-uncheckpointed transaction is
554 // replayed below. Each frame is a finished page image; the header
555 // frame, replayed to page 0, supersedes the header just read.
556 let pending = wal::recover(
557 &mut file,
558 header.wal_start,
559 header.wal_pages,
560 header.wal_seq,
561 )?;
562
563 // RB-003.3: settle the key and the format before writing anything,
564 // so a wrong secret or a refused old format leaves the file
565 // byte-identical. Both are checked against the header the file will
566 // have once the WAL is replayed (its page-0 frame, else the current
567 // header), since that header is the one the secret must open. No write
568 // the engine makes leaves a pending page-0 frame whose key wrap differs
569 // from page 0's (a rekey logs its new page 0 into a fresh WAL region,
570 // so a cut-off rekey leaves the old secret or the new one, never a
571 // mix), but checking against the replay's header is the right
572 // defence if one ever does.
573 let target = match pending.as_deref().and_then(page_zero_header) {
574 Some(h) => h,
575 None => header.clone(),
576 };
577 let kdf = KdfParams {
578 m_cost: target.m_cost,
579 t_cost: target.t_cost,
580 p_cost: target.p_cost,
581 };
582 let dek = match held_dek {
583 Some(held) if sealed_under(&target, held) => zeroize::Zeroizing::new(*held),
584 _ => {
585 let kek = crypto::derive_kek(secret, &target.salt, kdf)?;
586 crypto::unwrap_dek(&kek, &target.dek_nonce, &target.wrapped_dek)?
587 }
588 };
589 if !allow_old && target.version < VERSION_V9 {
590 return Err(MnemoError::FormatTooOld {
591 version: target.version,
592 });
593 }
594
595 if let Some(frames) = pending {
596 for (page_no, bytes) in &frames {
597 if bytes.len() != PAGE_SIZE {
598 return Err(MnemoError::Invalid("WAL frame is not page-sized".into()));
599 }
600 file.seek(SeekFrom::Start(page_no * PAGE_SIZE as u64))?;
601 file.write_all(bytes)?;
602 }
603 file.sync_all()?;
604 crate::bump_fsync();
605 // Re-read the now-current header from the replayed page 0.
606 file.seek(SeekFrom::Start(0))?;
607 file.read_exact(&mut hbuf)?;
608 header = Header::from_page(&hbuf)?;
609 }
610
611 Self::from_locked(path, file, header, dek, kdf, allow_old)
612 }
613
614 /// Build a read-write handle from a locked file whose page 0 is
615 /// `header`, with the data key already unwrapped. The shared-mode reload
616 /// (12.4) comes here directly when the seal verifies under the key it
617 /// holds, so it pays no KDF.
618 pub(super) fn from_locked(
619 path: PathBuf,
620 file: LockedFile,
621 mut header: Header,
622 dek: zeroize::Zeroizing<[u8; crypto::KEY_LEN]>,
623 kdf: KdfParams,
624 allow_old: bool,
625 ) -> Result<Mnemo> {
626 // Validate the v7 header seal now that we have the DEK. Returns
627 // `MnemoError::HeaderTampered` if any of the sealed mutable fields
628 // were rewritten after the last legitimate flush. No-op for v6 and
629 // older — pre-v7 files don't carry a seal.
630 //
631 // DUR-1 (Planner decision 2026-11-01, Option B): a seal failure is
632 // **always** reported as tampering, never auto-healed. R-002.1
633 // part 3's seal-mismatch WAL fallback is withdrawn: the seal
634 // exists for tamper-evidence; silently loading a stale committed
635 // header from the WAL would erase that evidence. A real crash
636 // cannot produce "CRC valid + seal invalid" either, because the
637 // whole 294-byte header lies inside sector 0 which storage
638 // writes atomically — a torn sector breaks magic, version or
639 // CRC first, and *those* are healed by the scan above.
640 header.validate_seal(&dek)?;
641
642 // Build the pager at the file's *current* version so initial reads
643 // use the matching AAD scheme — `Pager::page_aad` returns empty AAD
644 // for v4/v5 (no page binding) and `page_no.to_le_bytes()` for v6+.
645 let on_disk_version = header.version;
646 // The v8 sunset (v0.13): a read-write handle is v9. Older files open
647 // read-only (v8) or through `compact_file`, which converts them.
648 if !allow_old && on_disk_version < VERSION_V9 {
649 return Err(MnemoError::FormatTooOld {
650 version: on_disk_version,
651 });
652 }
653 let mut pager = Pager::new(file, dek, header.write_counter, on_disk_version);
654 let dimensions = header.dimensions as usize;
655
656 // === v4→v5 catalog migration =====================================
657 // v6 reads directly; v5 reads directly; v4 replays the catalog into
658 // the v5 shape by reading each memory's body to populate the new
659 // `accessed_at` / `access_count` fields. All page reads here still
660 // use the v4/v5 AAD scheme (none).
661 let (catalog, migrated_from_v4) = if on_disk_version >= 5 {
662 (
663 load_catalog(
664 &mut pager,
665 header.catalog_start,
666 header.catalog_pages,
667 header.catalog_len,
668 header.catalog_gen,
669 )?,
670 false,
671 )
672 } else if on_disk_version == 4 {
673 let v4 = load_catalog_v4(
674 &mut pager,
675 header.catalog_start,
676 header.catalog_pages,
677 header.catalog_len,
678 )?;
679 (migrate_v4_catalog(&mut pager, v4)?, true)
680 } else {
681 return Err(MnemoError::UnsupportedVersion(on_disk_version));
682 };
683 // FMT-9c.2: replay the v9 delta chain onto the base catalog.
684 let (catalog, delta_chain, delta_rows) =
685 apply_delta_chain(&mut pager, catalog, header.catalog_delta)?;
686 let index = build_id_index(&catalog);
687 // PERF-1 step 3: the ANN index loads on first use, not at open. A
688 // v5→v6 page migration is the exception: it must read the index
689 // now, under the old page AAD, to rewrite it.
690 let ann = if on_disk_version < 6 {
691 load_index(
692 &mut pager,
693 header.index_start,
694 header.index_pages,
695 header.index_len,
696 header.index_gen,
697 dimensions,
698 )?
699 } else {
700 None
701 };
702 let ann_unloaded = ann.is_none() && header.index_pages > 0;
703 let mut manifest = load_manifest(
704 &mut pager,
705 header.manifest_start,
706 header.manifest_pages,
707 header.manifest_len,
708 header.manifest_gen,
709 )?;
710
711 // === v5→v6 page-encryption migration =============================
712 // Pre-v6 data pages were encrypted under no AAD; v6 binds the home
713 // `page_no` as AAD so a page-transplant attack fails authentication
714 // on read. Re-stage every live record page through `write_page` so
715 // the next flush re-encrypts it under v6 AAD at the same home slot.
716 // Tombstoned entries skip — they're reclaimed by `compact_file`.
717 // Old catalog / ANN / manifest pages stay as orphan ciphertext;
718 // the next flush writes the new runs at fresh slots.
719 let migrated_pages_to_v6 = on_disk_version < 6;
720 if migrated_pages_to_v6 {
721 for e in &catalog {
722 if e.deleted {
723 continue;
724 }
725 for i in 0..e.page_count as u64 {
726 let payload = pager.read_page(e.start_page + i, 0)?;
727 pager.write_page(e.start_page + i, 0, &payload)?;
728 }
729 }
730 pager.set_version(6);
731 }
732
733 // === v6→v7 header-seal migration =================================
734 // v7 adds an AES-GCM seal at the tail of the header page that
735 // authenticates every mutable header field. Pre-v7 files have no
736 // seal — we just bump the version in memory and mark `dirty_catalog`
737 // so the next flush writes a sealed v7 header. Pages are untouched
738 // (page crypto is unchanged from v6).
739 let migrated_header_to_v7 = on_disk_version < 7;
740 if migrated_header_to_v7 {
741 pager.set_version(7);
742 }
743
744 // === v7→v8 cache-directory migration =============================
745 // v8 adds three u64 header fields (cache_start/pages/len) for the
746 // result-cache directory (Phase 10.1). Pre-v8 files have those
747 // three bytes as zero (the on-disk region beyond the v7 seal is
748 // zero-initialized), which is the legal "empty cache directory"
749 // state — no data migration required. We bump the version in
750 // memory and mark dirty_catalog so the next flush writes a
751 // v8-sealed header with the new AAD shape covering the cache
752 // pointer. Pages are untouched.
753 let migrated_header_to_v8 = on_disk_version < 8;
754 if migrated_header_to_v8 {
755 pager.set_version(VERSION);
756 }
757
758 let migrated = migrated_from_v4
759 || migrated_pages_to_v6
760 || migrated_header_to_v7
761 || migrated_header_to_v8;
762 if migrated {
763 // Stamp the version forward in memory; the next flush persists
764 // it via the WAL-committed header frame (and the v7 seal).
765 header.version = VERSION;
766 // Snapshots written under an older format reference page runs
767 // encrypted under the old AAD scheme; this build can't decrypt
768 // them under the new pager. Drop the manifest so the next
769 // flush records a fresh snapshot of the migrated state.
770 // Point-in-time recovery into the pre-migration past is
771 // sacrificed for migration simplicity. Old pages remain on
772 // disk until `compact_file`. (For v6→v7 alone — which doesn't
773 // change page crypto — the old snapshots would technically
774 // still be readable, but dropping them uniformly keeps the
775 // migration policy simple.)
776 manifest.clear();
777 }
778
779 // The ANN index pages on disk were sealed under the *old* AAD scheme
780 // if we migrated pages to v6; force a rewrite at fresh v6-encrypted
781 // slots on the next flush so a future reopen doesn't try to read
782 // them under the new AAD and fail. (Catalog and manifest already
783 // get rewritten via dirty_catalog and the manifest.clear() above.)
784 let dirty_index = migrated_pages_to_v6 && ann.is_some();
785
786 // Load the cache directory (v8+ only). Pre-v8 files have
787 // cache_pages == 0 by header layout, so this trivially yields
788 // an empty directory without touching the disk.
789 let cache_dir = load_cache_directory(
790 &mut pager,
791 header.cache_start,
792 header.cache_pages,
793 header.cache_len,
794 header.cache_gen,
795 )?;
796 let cache_idx = DirectoryIndex::build(&cache_dir);
797 // FMT-9 (a)/(b): the free list and key-slot run (empty pointers on
798 // v8, so these don't touch the disk there).
799 let free = load_free_list(&mut pager, header.free)?;
800 let key_slots = load_key_slots(&mut pager, header.keys)?;
801 let col = VecColumn::from_catalog(&catalog);
802 let recs = RecPages::from_catalog(&catalog);
803 let bm25_manifest = load_bm25_manifest(&mut pager, header.bm25)?;
804 let max_snapshots = if header.version >= VERSION_V9 {
805 V9_DEFAULT_MAX_SNAPSHOTS
806 } else {
807 DEFAULT_MAX_SNAPSHOTS
808 };
809
810 Ok(Mnemo {
811 pager,
812 header,
813 catalog,
814 index,
815 path: absolute(path),
816 dimensions,
817 kdf,
818 dirty_catalog: migrated,
819 ann,
820 dirty_index,
821 // A v5→v6 page migration leaves the index sealed under the old
822 // page AAD, so it must be rewritten on the next flush.
823 index_persist: IndexPersist {
824 forced: dirty_index,
825 ..IndexPersist::default()
826 },
827 ann_unloaded,
828 manifest,
829 // `Mnemo::open` has no [`MnemoConfig`] to read; default the cap
830 // to [`DEFAULT_MAX_SNAPSHOTS`]. Override with
831 // [`Mnemo::set_max_snapshots`] right after open if you need
832 // unlimited retention or a different cap for this handle.
833 max_snapshots,
834 read_only: false,
835 cache_dir,
836 cache_idx,
837 dirty_cache: false,
838 cache_budgets: std::collections::HashMap::new(),
839 cache_flush_policy: CacheFlushPolicy::default(),
840 cache_batch: BatchState::default(),
841 cache_counters: CounterState::default(),
842 embedder: None,
843 fsyncs_last_flush: 0,
844 agent_set: None,
845 key_index: None,
846 superseded: None,
847 layers: None,
848 rekey_cleanup: None,
849 access_delta: None,
850 commit_point: None,
851 lock_note: None,
852 window_committed: false,
853 shared: false,
854 turn_compress_after: DEFAULT_TURN_COMPRESS_AFTER_SECS,
855 free,
856 pending_free: Vec::new(),
857 free_dirty: false,
858 key_slots,
859 dirty_keys: false,
860 gen_leased: false,
861 col,
862 recs,
863 bm25: Bm25State {
864 manifest: bm25_manifest,
865 ..Bm25State::default()
866 },
867 kw: None,
868 index_layout_override: None,
869 dirty_rows: std::collections::BTreeSet::new(),
870 delta_chain,
871 delta_rows,
872 catalog_rewrite: false,
873 index_chain: Vec::new(),
874 index_chain_ops: 0,
875 index_rebase: false,
876 index_delta_ops: None,
877 snapshot_cap_explicit: false,
878 direct_writes: false,
879 vector_cache_budget: DEFAULT_VECTOR_CACHE_BUDGET,
880 })
881 }
882
883 /// Open an existing database in **read-only** mode. Takes an OS shared
884 /// lock; multiple read-only handles can coexist, but any attempt to
885 /// open the same file read-write elsewhere returns
886 /// [`MnemoError::Locked`]. All mutating methods on the returned handle
887 /// return [`MnemoError::ReadOnly`], including `recall` with
888 /// `track_access(true)`.
889 ///
890 /// Two conditions block a read-only open and return
891 /// [`MnemoError::NeedsWriteOpen`]:
892 ///
893 /// 1. The file has a committed-but-uncheckpointed WAL transaction that
894 /// would ordinarily be replayed on open. A read-only handle
895 /// cannot write the recovered pages back.
896 /// 2. The on-disk format version is older than the current
897 /// `format::VERSION`, requiring an on-disk migration.
898 ///
899 /// In either case, open the file read-write once (which recovers /
900 /// migrates in place), close the handle cleanly, and reopen
901 /// read-only. This is documented in the README's Durability section.
902 pub fn open_read_only(path: impl Into<PathBuf>, passphrase: &str) -> Result<Mnemo> {
903 Self::open_read_only_secret(path, passphrase.as_bytes())
904 }
905
906 /// PERF-1 step 7: [`Mnemo::open_read_only`] for a raw-key database.
907 pub fn open_read_only_with_key(
908 path: impl Into<PathBuf>,
909 key: &[u8; crypto::KEY_LEN],
910 ) -> Result<Mnemo> {
911 Self::open_read_only_secret(path, key)
912 }
913
914 /// The embedding width the file at `path` records, read from its header
915 /// (page 0 is sealed, not encrypted): no key, no lock, no KDF. A caller
916 /// that embeds before it takes the lock uses it to build its embedder's
917 /// identity (12.4: slow work first, unlocked). A torn header is an
918 /// error here (a read-write open heals it).
919 pub fn peek_dimensions(path: impl AsRef<std::path::Path>) -> Result<usize> {
920 let mut file = File::open(path.as_ref())?;
921 let mut hbuf = [0u8; PAGE_SIZE];
922 file.read_exact(&mut hbuf)?;
923 Ok(Header::from_page(&hbuf)?.dimensions as usize)
924 }
925
926 /// Whether `key` opens the raw-key file at `path`, decided from its
927 /// header alone: no catalog load, no lock, microseconds. `Ok(false)` is
928 /// "not this key". Unlike a read-only open, it answers for a file that
929 /// has a WAL transaction to recover (the state a kill between a commit
930 /// and its checkpoint leaves), so a caller trying each of its keys in
931 /// turn isn't stopped by [`MnemoError::NeedsWriteOpen`]. A torn header is
932 /// judged from the WAL frame that would heal it. Other failures (not a
933 /// `.mnemo` file, a tampered header, an I/O error) are errors. Added in
934 /// v0.12.2 (RB-001.23).
935 pub fn key_opens(
936 path: impl AsRef<std::path::Path>,
937 key: &[u8; crypto::KEY_LEN],
938 ) -> Result<bool> {
939 let mut file = File::open(path.as_ref())?;
940 let mut hbuf = [0u8; PAGE_SIZE];
941 file.seek(SeekFrom::Start(0))?;
942 file.read_exact(&mut hbuf)?;
943 let header = match Header::from_page(&hbuf) {
944 Ok(h) => h,
945 // A candidate frame only counts when it validates with `key`.
946 Err(parse_err) => {
947 return match scan_header_recovery_candidates(&mut file, key, None)? {
948 Some(_) => Ok(true),
949 None => Err(parse_err),
950 }
951 }
952 };
953 let kdf = KdfParams {
954 m_cost: header.m_cost,
955 t_cost: header.t_cost,
956 p_cost: header.p_cost,
957 };
958 let kek = crypto::derive_kek(key, &header.salt, kdf)?;
959 let dek = match crypto::unwrap_dek(&kek, &header.dek_nonce, &header.wrapped_dek) {
960 Ok(dek) => dek,
961 Err(MnemoError::WrongPassphrase) => return Ok(false),
962 Err(e) => return Err(e),
963 };
964 header.validate_seal(&dek)?;
965 Ok(true)
966 }
967
968 pub(super) fn open_read_only_secret(path: impl Into<PathBuf>, secret: &[u8]) -> Result<Mnemo> {
969 let path: PathBuf = path.into();
970 let file = open_with_lock(&path, /*write=*/ false, Create::No)?;
971 Self::read_only_from_locked(path, file, secret)
972 }
973
974 /// [`Mnemo::open_read_only`] through 12.4's lock layer: it waits up to
975 /// `opts.deadline` for a writer (a plain handle, a shared write, a
976 /// compaction) instead of failing at once with `Locked`, and runs
977 /// alongside other readers. The handle then holds the shared lock for
978 /// its life, as any read-only handle does; close it promptly. A `Locked`
979 /// after the wait names what holds the file when it left a note.
980 ///
981 /// Like `open_read_only`, it never recovers anything: a pending WAL or
982 /// a torn header gives [`MnemoError::NeedsWriteOpen`]. Where shared
983 /// mode isn't available it tries once, as `open_read_only` does.
984 pub fn open_read_only_waiting(
985 path: impl AsRef<Path>,
986 passphrase: &str,
987 opts: crate::SharedOptions,
988 ) -> Result<Mnemo> {
989 Self::open_read_only_waiting_secret(path.as_ref(), passphrase.as_bytes(), &opts)
990 }
991
992 /// [`Mnemo::open_read_only_waiting`] for a raw-key file.
993 pub fn open_read_only_waiting_with_key(
994 path: impl AsRef<Path>,
995 key: &[u8; crypto::KEY_LEN],
996 opts: crate::SharedOptions,
997 ) -> Result<Mnemo> {
998 Self::open_read_only_waiting_secret(path.as_ref(), key, &opts)
999 }
1000
1001 fn open_read_only_waiting_secret(
1002 path: &Path,
1003 secret: &[u8],
1004 opts: &crate::SharedOptions,
1005 ) -> Result<Mnemo> {
1006 let path = absolute(path.to_path_buf());
1007 let file = lock_shared(&path, opts)?;
1008 Self::read_only_from_locked(path, file, secret)
1009 }
1010
1011 /// The read-only open's work on a file its caller holds the shared lock
1012 /// on.
1013 fn read_only_from_locked(path: PathBuf, mut file: LockedFile, secret: &[u8]) -> Result<Mnemo> {
1014 let mut hbuf = [0u8; PAGE_SIZE];
1015 file.seek(SeekFrom::Start(0))?;
1016 file.read_exact(&mut hbuf)?;
1017 // A torn header on a read-only open cannot be healed (that requires
1018 // writing recovered frames back). Surface as `NeedsWriteOpen` so the
1019 // caller knows to reopen read-write once. DUR-1 (R-002.1 part 4):
1020 // any `Header::from_page` failure (`BadMagic`,
1021 // `UnsupportedVersion`, `HeaderChecksum`) qualifies — but only
1022 // when the file-wide scan actually finds a valid committed
1023 // header frame to recover from, so a non-mnemo file still
1024 // reports `BadMagic` cleanly. A `validate_seal` failure is NOT
1025 // routed here (Option B: tampering is never auto-healed).
1026 let header = match Header::from_page(&hbuf) {
1027 Ok(h) => h,
1028 Err(parse_err) => {
1029 let recoverable = scan_header_recovery_candidates(&mut file, secret, None)?;
1030 return Err(match (parse_err, recoverable) {
1031 // RB-004.5 item 2: as for a pending WAL (RB-003.3), a
1032 // pre-v9 file can't be opened read-write, so the advice
1033 // names the step that does heal it.
1034 (_, Some(frames))
1035 if page_zero_header(&frames).is_some_and(|h| h.version < VERSION_V9) =>
1036 {
1037 MnemoError::NeedsWriteOpen {
1038 reason: "torn header page in a pre-v9 file; run `mnemo compact \
1039 <file>` (with `--key-file <key>` for a raw-key file), \
1040 which recovers it and converts the file to v9",
1041 }
1042 }
1043 (_, Some(_)) => MnemoError::NeedsWriteOpen {
1044 reason: "torn header page needs recovery from WAL",
1045 },
1046 (e, None) => e,
1047 });
1048 }
1049 };
1050
1051 // Refuse to migrate on read-only opens. v6→v7 is header-only
1052 // (no page writes required), but even that bumps `version` on
1053 // disk and re-seals the header — which needs write access. Keep
1054 // the policy uniform across all sub-version bumps.
1055 // A v4–v7 file needs migrating, which a read-only handle can't do,
1056 // and since v0.13 no read-write handle opens it either: compaction
1057 // converts it to v9.
1058 if header.version < VERSION {
1059 return Err(MnemoError::FormatTooOld {
1060 version: header.version,
1061 });
1062 }
1063
1064 // Refuse if there is a committed-but-uncheckpointed transaction
1065 // waiting for `wal::recover` to fold it into home pages.
1066 if wal::recover(
1067 &mut file,
1068 header.wal_start,
1069 header.wal_pages,
1070 header.wal_seq,
1071 )?
1072 .is_some()
1073 {
1074 // RB-003.3: a pre-v9 file can't be opened read-write any more,
1075 // so the usual advice would send the user in a circle.
1076 // Compaction recovers the WAL and converts the file.
1077 return Err(MnemoError::NeedsWriteOpen {
1078 reason: if header.version < VERSION_V9 {
1079 "pending WAL recovery in a pre-v9 file; run `mnemo compact <file>` \
1080 (with `--key-file <key>` for a raw-key file), which recovers it and \
1081 converts the file to v9"
1082 } else {
1083 "pending WAL recovery"
1084 },
1085 });
1086 }
1087
1088 let kdf = KdfParams {
1089 m_cost: header.m_cost,
1090 t_cost: header.t_cost,
1091 p_cost: header.p_cost,
1092 };
1093 let kek = crypto::derive_kek(secret, &header.salt, kdf)?;
1094 let dek = crypto::unwrap_dek(&kek, &header.dek_nonce, &header.wrapped_dek)?;
1095 header.validate_seal(&dek)?;
1096
1097 let mut pager = Pager::new(file, dek, header.write_counter, header.version);
1098 let dimensions = header.dimensions as usize;
1099
1100 let catalog = load_catalog(
1101 &mut pager,
1102 header.catalog_start,
1103 header.catalog_pages,
1104 header.catalog_len,
1105 header.catalog_gen,
1106 )?;
1107 // FMT-9c.2: replay the v9 delta chain onto the base catalog.
1108 let (catalog, delta_chain, delta_rows) =
1109 apply_delta_chain(&mut pager, catalog, header.catalog_delta)?;
1110 let index = build_id_index(&catalog);
1111 // PERF-1 step 3: the ANN index loads on first use, not at open.
1112 let ann: Option<IvfPqIndex> = None;
1113 let ann_unloaded = header.index_pages > 0;
1114 let manifest = load_manifest(
1115 &mut pager,
1116 header.manifest_start,
1117 header.manifest_pages,
1118 header.manifest_len,
1119 header.manifest_gen,
1120 )?;
1121 let cache_dir = load_cache_directory(
1122 &mut pager,
1123 header.cache_start,
1124 header.cache_pages,
1125 header.cache_len,
1126 header.cache_gen,
1127 )?;
1128 let cache_idx = DirectoryIndex::build(&cache_dir);
1129 let free = load_free_list(&mut pager, header.free)?;
1130 let key_slots = load_key_slots(&mut pager, header.keys)?;
1131 let col = VecColumn::from_catalog(&catalog);
1132 let recs = RecPages::from_catalog(&catalog);
1133 let bm25_manifest = load_bm25_manifest(&mut pager, header.bm25)?;
1134 let max_snapshots = if header.version >= VERSION_V9 {
1135 V9_DEFAULT_MAX_SNAPSHOTS
1136 } else {
1137 DEFAULT_MAX_SNAPSHOTS
1138 };
1139
1140 Ok(Mnemo {
1141 pager,
1142 header,
1143 catalog,
1144 index,
1145 path: absolute(path),
1146 dimensions,
1147 kdf,
1148 // No mutation is possible; leave both dirty flags false so
1149 // `close()` is a genuine no-op even if we later add an
1150 // auto-flush on drop.
1151 dirty_catalog: false,
1152 ann,
1153 dirty_index: false,
1154 index_persist: IndexPersist::default(),
1155 ann_unloaded,
1156 manifest,
1157 max_snapshots,
1158 read_only: true,
1159 cache_dir,
1160 cache_idx,
1161 dirty_cache: false,
1162 cache_budgets: std::collections::HashMap::new(),
1163 cache_flush_policy: CacheFlushPolicy::default(),
1164 cache_batch: BatchState::default(),
1165 cache_counters: CounterState::default(),
1166 embedder: None,
1167 fsyncs_last_flush: 0,
1168 agent_set: None,
1169 key_index: None,
1170 superseded: None,
1171 layers: None,
1172 rekey_cleanup: None,
1173 access_delta: None,
1174 commit_point: None,
1175 lock_note: None,
1176 window_committed: false,
1177 shared: false,
1178 turn_compress_after: DEFAULT_TURN_COMPRESS_AFTER_SECS,
1179 free,
1180 pending_free: Vec::new(),
1181 free_dirty: false,
1182 key_slots,
1183 dirty_keys: false,
1184 gen_leased: false,
1185 col,
1186 recs,
1187 bm25: Bm25State {
1188 manifest: bm25_manifest,
1189 ..Bm25State::default()
1190 },
1191 kw: None,
1192 index_layout_override: None,
1193 dirty_rows: std::collections::BTreeSet::new(),
1194 delta_chain,
1195 delta_rows,
1196 catalog_rewrite: false,
1197 index_chain: Vec::new(),
1198 index_chain_ops: 0,
1199 index_rebase: false,
1200 index_delta_ops: None,
1201 snapshot_cap_explicit: false,
1202 direct_writes: false,
1203 vector_cache_budget: DEFAULT_VECTOR_CACHE_BUDGET,
1204 })
1205 }
1206
1207 /// FMT-9e: compress v9 thread-turn records once they're older than
1208 /// `secs` (by their `created_at`). The default,
1209 /// [`DEFAULT_TURN_COMPRESS_AFTER_SECS`], is 7 days: recent turns stay
1210 /// plain, so a live session reads them at full speed, and each commit
1211 /// compresses up to 256 turns that have aged past the cutoff. `0`
1212 /// compresses every turn as it's written. Compaction always compresses
1213 /// every turn, and so does a vacuum that moves one. Applies to this
1214 /// handle only (not persisted); a no-op on v8 files, which never
1215 /// compress. Reading never depends on it.
1216 pub fn set_turn_compress_after(&mut self, secs: u64) {
1217 self.turn_compress_after = secs.min(i64::MAX as u64) as i64;
1218 }
1219
1220 /// Override the manifest snapshot cap on this open handle. `0` disables
1221 /// the cap; any positive value keeps the most-recent `max` snapshots and
1222 /// drops the rest on the next flush. See [`MnemoConfig::max_snapshots`].
1223 pub fn set_max_snapshots(&mut self, max: usize) {
1224 self.max_snapshots = max;
1225 // FMT-9c: an explicit count is a hard cap on v9 (see
1226 // `apply_retention`).
1227 self.snapshot_cap_explicit = true;
1228 }
1229
1230 /// Embedding dimensionality this database expects.
1231 pub fn dimensions(&self) -> usize {
1232 self.dimensions
1233 }
1234
1235 /// Number of live (non-deleted) memories.
1236 pub fn len(&self) -> usize {
1237 self.catalog.iter().filter(|e| !e.deleted).count()
1238 }
1239
1240 /// True if the database holds no live memories.
1241 pub fn is_empty(&self) -> bool {
1242 self.len() == 0
1243 }
1244
1245 // --- internal page/record helpers -------------------------------------
1246}