add support for reviews

.gitignore

@@ -1,4 +1,5 @@
 anki/
+fsrs-rs/
 _docs/
 _proc/
 

DEV.md

@@ -9,9 +9,11 @@ This is an nbdev project. The notebooks in `nbs/` are the source of truth, and `
 - `00_schema.ipynb` (`schema.py`): the sqlite file format. Schema 18 DDL, the `unicase` collation, timestamps and id allocation, the default deck and deck config, the Basic and Cloze notetypes, and `create_collection`.
 - `01_collection.ipynb` (`collection.py`): the `Collection` class. Notes, cards, decks, card generation, search, due counts, and the checksum, guid, and HTML stripping utilities Anki uses to fingerprint notes.
 - `02_syncer.ipynb` (`syncer.py`): the AnkiWeb client. Login, the delta sync state machine, schema 11 conversions, full upload and download, and auth storage.
-- `03_core.ipynb` (`core.py`): the functional API (`add_card`, `find_notes`, `sync`, etc). Each function opens the default collection, does its work, and closes. `anki_tools` exposes these as LLM tools.
+- `03_fsrs.ipynb` (`fsrs.py`): the FSRS-6 memory model, a pure port of the scheduling half of the `fsrs-rs` crate Anki embeds: parameter upgrading/clipping (17/19/21-length sets), the forgetting curve, memory-state updates, and the SM-2 approximation for truncated histories. No card or collection knowledge.
+- `04_scheduler.ipynb` (`scheduler.py`): reviewing. The v3 scheduler state machine (new/learning/review/relearning, SM-2 and FSRS), interval fuzz, `answer_card` with its revlog/counter/leech bookkeeping, `answer_buttons`, and a simplified `next_card` queue with daily limits and sibling burying.
+- `05_core.ipynb` (`core.py`): the functional API (`add_card`, `find_notes`, `sync`, etc). Each function opens the default collection, does its work, and closes. `anki_tools` exposes these as LLM tools.
 
-`fastanki/_proto/` is not generated from a notebook. It contains protobuf modules copied from the `anki` wheel with imports rewritten, and can be refreshed the same way from a newer wheel if needed. `tests/test_sync.py` is the end to end sync test, described below. `anki/` (gitignored) is a clone of the Anki source, kept for reference. When a protocol question comes up, the answer is in `anki/rslib/src/sync/`.
+`fastanki/_proto/` is not generated from a notebook. It contains protobuf modules copied from the `anki` wheel with imports rewritten, and can be refreshed the same way from a newer wheel if needed. `tests/test_sync.py` is the end to end sync test, described below. `anki/` (gitignored) is a clone of the Anki source, kept for reference. When a protocol question comes up, the answer is in `anki/rslib/src/sync/`; for scheduling, `anki/rslib/src/scheduler/` and the `fsrs-rs/` clone (also gitignored).
 
 ## Design
 
@@ -50,11 +52,14 @@ Hard-won facts worth keeping. When a protocol question comes up the answer is in
 - `config.curModel` is a per-collection notetype id (a timestamp), so exclude it when comparing the `config` table against the oracle.
 - The AnkiWeb sync client lives in `syncer.py` (module `fastanki.syncer`), named so it doesn't collide with core's top-level `sync` function; `from fastanki import *` exposes both, and `SyncServer`/`sync_collection`/etc. come from `fastanki.syncer`.
 - `apsw` (not stdlib `sqlite3`): connections have no `.commit()`/`.rollback()` (autocommit unless inside an explicit transaction), `execute` runs multiple statements so there is no `executescript`, and double-quoted strings are disabled, so SQL string literals must use single quotes. Mutations go through `Collection._tx()`; sync uses explicit `BEGIN IMMEDIATE`/`COMMIT`/`ROLLBACK`. The busy timeout is set before the WAL pragma, else concurrent opens hit `BusyError` on the journal-mode switch.
+- Scheduling gotchas, learned porting `rslib/src/scheduler/` for `04_scheduler.ipynb`: `ANKI_TEST_MODE=1` (set before the `anki` package first computes states) disables the oracle's interval fuzz, which is what makes exact scheduling comparisons possible. Fuzz is seeded per `(card_id + reps)`, but Anki seeds Rust's ChaCha12 and we seed Python's Mersenne twister, so the two pick different (equally valid) points inside identical fuzz bounds; everything else is deterministic. Anki computes in f32: keep the protobuf floats raw (1.3 is really 1.2999999523) and round half-away-from-zero (`_round`), or intervals drift by a day at rounding boundaries.
+- fsrs-rs version matters: Anki 26.05 pins fsrs 6.6.1, whose same-day (short-term) stability floors the multiplier at 1 for Good/Easy only; 6.6.2+ floors Hard too. We match 6.6.1 -- the twin-collection oracle test caught this, and it's the arbiter if the pin moves.
+- The `cards.data` JSON carries `pos` (original new-queue position), `lrt` (last review time), and under FSRS `s`/`d`/`dr`/`decay`, rounded to 4/3/2/3 decimal places respectively. Anki clears `s`/`d`/`dr` when answering with FSRS off but leaves `decay` alone.
 
 ## Not implemented
 
 Media sync is a separate protocol and is skipped entirely. That is safe, because collection sync moves note rows and media sync moves files, and the two don't interfere. The one rule is to never create notes that reference media files we can't upload, so the API has no way to attach media.
 
 Filtered decks aren't supported. The schema 11 conversions assert if they meet one, so syncing such a collection fails rather than corrupting it.
 
-There is no scheduler. We count due cards but can't answer them. If reviewing is ever wanted, FSRS is available as a pure python package, so the Rust scheduler wouldn't need porting.
+The scheduler covers normal decks only: answering a card in a filtered deck raises. The `next_card` queue is a simplified port of Anki's v3 builder -- daily limits come from the named deck's preset rather than a per-subdeck limit tree, gathering is by due order, reviews precede new cards, and there is no display-order matrix. Those simplifications affect session ordering, never scheduling state, so they can be extended without compatibility concerns. The FSRS optimizer is deliberately absent: parameters arrive through deck-config sync, and users can optimize from Anki desktop (scheduling with synced-or-default parameters is exactly what AnkiWeb's own study feature does).

fastanki/__init__.py

@@ -1,4 +1,5 @@
 __version__ = "0.0.4"
 from .collection import *
 from .syncer import *
+from .scheduler import *
 from .core import *