Sync
The previous design put the merge inside SQL: timestamp-stamping triggers,
a tombstone table, per-device SQLite snapshots, a device id in the database,
and row merges written as correlated-subquery UPDATEs. All of that schema
was inherited by desktop, and the merge was very hard to verify.
Replace it with a single JSON document plus a three-way merge in plain Rust.
Storing the last-read remote document as the merge base makes "who changed
what" and "what was deleted" derivable, which removes the tombstone table,
the updated_at columns and all eight triggers. It also makes a single shared
file safe: if a publish is lost the remote is unchanged, so the next merge
re-detects the local edits and republishes rather than reading their absence
as a deletion. That removes the need for per-device files and for device
identity entirely. Schema impact drops to one column, decks.uuid.
Conflicts resolve to "theirs", which is required for convergence rather than
chosen for taste — preferring ours has each device re-asserting its own value
forever. Every such choice is reported to the user.
SQLite becomes invisible: the working copy lives in the app data dir and the
user only ever names a .json file. There is always a document, defaulting to
private storage, so edits are never stranded with nowhere to be saved.
Pointing at a file that doesn't exist yet moves the current library there;
pointing at one that does adopts it. A converter brings a pre-document
library.db across.
Android reaches the folder through the Storage Access Framework rather than
MANAGE_EXTERNAL_STORAGE, so no storage permission is declared at all.
Filters
Track card origin in reference.db: Universes Beyond, digital-only and
Un-set cards, plus numeric power/toughness. Each origin flag is derived as
"every printing", not "any printing" — ordinary cards get reprinted into
those sets, and Lightning Bolt appearing in Marvel Commander must not hide
it. Non-numeric power ("*", "1+*") stays NULL so it drops out of range
filters instead of being compared as zero.
Also wire up filters over data already present: price range, format legality
(restricted counts as legal) and "can be a commander". All of these degrade
silently against an older reference.db rather than failing every query.
The shared criteria live in one component used by both filter panels, as a
compact row of icon toggles.
Card Search gains the Library tab's bulk selection: checkboxes in both views,
select-all, and a bulk add-to-deck with a single undo for the whole batch.
The Rust core was already portable (caller-provided db paths, rustls rather
than openssl) and the shell already had the mobile entry point, so this is
mostly toolchain, provisioning, and platform-gating.
Toolchain:
- scripts/android-setup.sh installs what Android Studio doesn't: SDK
cmdline-tools, a pinned stable NDK, and the Rust Android targets. The NDK is
pinned rather than "latest" because sdkmanager's newest is often a beta
(30.0.15729638 is r30-beta2) -- not something to compile SQLite with. Prefers
Android Studio's bundled JDK 21, since Gradle rejects newer JDKs.
- Track src-tauri/gen/android (excluding build output): AndroidManifest.xml is
hand-edited and Android permissions have no tauri.conf.json equivalent.
- vite: serve on TAURI_DEV_HOST so a device/emulator can reach the dev server.
Card data: downloading streams ~740MB to build a ~73MB db, which is rough on a
phone. Add import_reference_db to adopt a desktop-built file instead, validated
by db::inspect_reference_db first so picking the wrong file is a clean error
rather than a broken install. Installs via copy-then-rename so an interrupted
copy can't leave a partial db.
Ingest hardening (helps desktop too):
- Give ureq connect/read timeouts. Cancellation is only polled per parsed
record, so a stalled socket previously hung forever with no way to cancel.
- Drop the remove_file before rename; POSIX rename already replaces atomically,
and the delete-first window could leave no reference.db at all if the process
was killed -- much likelier on Android.
UI: hide Open/New Library on mobile (scoped storage has no arbitrary-path
equivalent, and SQLite can't open a content:// URI), reported via a cfg!(mobile)
flag on Status rather than user-agent sniffing. Use 100dvh so the soft keyboard
doesn't break the shell.
Verified: 70/70 core tests, desktop build clean, and the aarch64 APK builds
(bundled SQLite C and ring both cross-compile under NDK 28). Not yet run on a
device.
