v1.8.0

Scanning is a different pipeline. The perceptual-hash + bag-of-visual-words
recall stage and the ORB verify stage are gone, replaced by two small ONNX models:
cornelius finds the card's corners, milo embeds the dewarped card as a 128-d
unit vector, and a brute-force cosine sweep over a prebuilt catalog names it. On
the same 100-card noisy MTG sample the old stack scored 78.0% exact printing /
88.0% right card in 1187 ms; this scores 76.0% / 90.0% in 310 ms — and it replaces
~2.6 GB of per-set indexes and whole-game rollups with two model files and one
catalog per game and language, at roughly 5 MB per 10k cards. There is no index
build in the scan path at all, so scanning a set needs no "preparing set" wait.

Both models are AGPL-3.0 (milo,
cornelius) while Bindarr is MIT, so
shipping them enabled is a licensing decision and not only a technical one.

A card in another language is a card, not an English card with a label. The
language of a copy now decides which printing it references, so it carries that
printing's name, artwork and price rather than the English row's; the name shows as
the card is printed everywhere it appears, while the English name stays searchable
alongside it. Prices for those printings improved in the same pass — Cardmarket
quotes are read where TCGplayer has no listing, the currency shown is the one the
marketplace quoted, and a price standing in from the English printing says so.

Added

• TCGdex is the Pokémon provider on a new install, and choosing is now possible.
  app_settings.pokemon_provider has existed since 1.7.2 with no way to set it, and
  only half of the app read it — the card lookups followed the setting while the set
  catalogue always came from pokemontcg.io, so an install switched by hand browsed a
  set list numbered by a provider none of its cards came from. The setting now drives
  the set sync too (218 English sets with release dates and series, against
  pokemontcg.io's 174), and there is a selector in Admin → Instance Settings that
  re-syncs the set list and rebuilds the TCGplayer product map behind it. Measured
  per card lookup: 57-206 ms via TCGdex against 971-1963 ms via pokemontcg.io, which
  also answers 5xx often enough to need a retry policy. TCGdex needs no API key and
  is the only one of the two that carries non-English cards.

  Upgrades keep pokemontcg.io. The two providers number the same sets
  differently (sv1/sv01, pgo/swsh10.5, me1/me01), and every cached card, scan catalog
  and collection row was built against one of those numberings — so only a database
  with no Pokémon sets and no Pokémon cards starts on TCGdex. Switching an existing
  install is an explicit admin action, and the set rows its cards still reference are
  kept even when the new provider does not list them.

• First run creates its own owner account. A new install no longer generates an
  admin password and prints it to a log nobody reads (and few people then change).
  The users table starts empty, the login screen turns into a set-the-owner-password
  form, and POST /api/auth/bootstrap creates the account with admin rights, refusing
  the moment any account exists. DEFAULT_ADMIN_PASSWORD still seeds it up front for
  scripted deploys. Either way the account is named admin — the name is fixed, not
  the owner's to pick, so it is the same one whichever way the install started and the
  one the orphan-row adoption looks up. That adoption now runs from the bootstrap route
  too, so a database carrying pre-multi-user rows (user_id IS NULL) does not show an
  empty collection to an owner account created through the UI. Existing installs are
  untouched: they already have accounts.

• The setup wizard covers the whole app, not just scanning. Six steps: language,
  which games you collect, the scanner's models and catalog, the optional provider
  keys, your binders and boxes, then a one-page tour of what each tab is for. The
  language comes first and is translated like the rest of the app, so the steps that
  explain a decision arrive in the language the admin reads; it is the same setting as
  Settings → Language.
  Completion is stored on the server (app_settings.setup_complete) rather than in
  localStorage, so closing it halfway resumes at the next login on any device.
  Every install is offered it once, upgrades included — the wizard is where
  catalogs, games and the provider keys are explained, and an install that predates
  it has never seen that tour. "Skip setup" dismisses it permanently.

• Catalogs replace scan indexes (Admin → Catalogs). A catalog is one game and
  language, and building it does two things: walk every set the provider lists and
  pull its cards into card_cache, then embed each one's artwork. Those were never
  one job before, and the first half was the half nobody ran — card caching used to
  happen only as a side effect of indexing, so Pokémon held 7,118 of 20,460 English
  cards (35%), with 104 of 174 sets holding just the handful their owner happened to
  have. Both phases resume, a stop still writes the partial catalog, and a set with
  no data in the chosen language counts as a gap rather than a failure.

• Coverage is shown with a denominator, because "built, 3,297 cards" reads as
  complete and is not. English counts against the set catalogue; other languages
  count against the provider's own set list for that language, so a Japanese
  Pokémon catalog reads 3,297 of 16,192.

• Corner detection runs in the browser, on a worker thread, using the same
  cornelius model the server dewarps with — so the outline on screen is by
  construction the crop that gets matched, and the client uploads one rectified
  448×448 square instead of a JPEG per preview frame (~2.7 MB per minute of pointing
  the camera at a card).

• scripts/measure-scan-floor.js — measures the scan gate instead of guessing
  it. For each sampled card it searches the catalog twice, once normally and once
  with the card's own row masked out (which is exactly what a card missing from the
  catalog looks like), under two degradation regimes, and prints what every
  candidate threshold would cost in wrong answers accepted and right answers
  rejected.

Fixed

• Non-English Magic cards were mostly unpriced. Scryfall quotes usd from
  TCGplayer and eur from Cardmarket, and only the first was read — but TCGplayer
  lists English Magic almost completely and non-English barely at all, so whole
  languages sat at 0.00 while a Cardmarket price existed: measured against the real
  cache, 241 of 1,205 Spanish rows priced, 53 of 194 Italian, 11 of 61 Simplified
  Chinese, against 96,090 of 103,656 English. A printing with no USD price now takes
  its EUR one and records price_currency accordingly. Never a mix inside one row —
  a normal price in USD beside a foil price in EUR is a pair that cannot be compared.
  Owned cards pick this up on the next daily sweep.

• Prices in another currency were shown with a dollar sign. Stored prices are not
  converted (an exchange rate is a live number this app has no source for), so a
  Cardmarket-priced card was quoted at €4.50 and displayed as $4.50. The symbol now
  follows the row's price_currency everywhere a card price is shown, including the
  price-history axis and tooltip. Amounts the owner typed or paid still show in the
  app's default.

• A German or Korean Pokémon card was priced as if it were English. TCGplayer
  sells Pokémon in two catalogues, English and Japanese, so every other language is
  matched to the English product for the same set and number. That is the closest real
  number and better than 0.00, but it is not this printing's price — a German card
  usually trades below the English one, and a Korean one is often not listed at all.
  Such rows are recorded as their own price source and the card inspector now labels
  the price "TCGplayer (English printing)" instead of presenting it as a quote for the
  card in hand.

• A card added in another language kept its English name. The copy's language and
  the printing it points at were two separate things: the localized name lives on the
  card_cache row (printed_name), so picking a card from an English search and then
  setting Language to Japanese filed the copy against the ENGLISH printing — English
  name, English art, English price, permanently. Only the camera scan resolved the
  localized printing. Every add now routes through the same swap (search, Quick Add,
  bulk add, rapid add, slab add, scan), for all eleven languages in
  shared/languages.json. A card never printed in the language asked for keeps the
  printing that was picked, which is still what the user owns.

• Names showed in English outside the collection grid. Deck lists, the deck
  picker and preview, the checkout pull list and its "Missing …" warning, filing
  recommendations, the slab picker, the scanner's toasts and its debug candidates all
  read the English name column directly. They show the printed name now; name
  stays English where logic depends on it, so the four-copy deck rule still counts a
  Japanese and an English copy as the same card and CSV exports are unchanged.

• A Japanese Pokémon card could only be found by typing Japanese. TCGdex
  publishes one name per language, so a localized row carried the localized name in
  both name and printed_name and nothing English was left to search. A card being
  added now learns its English name from its own English printing when one exists (one
  request, cached), and the TCGdex price sweep backfills rows added before this.
  Display still reads printed_name; search reads both columns.

• Re-caching a card silently reset columns no provider writes. card_cache was
  written with INSERT OR REPLACE, which deletes the row and inserts a new one — so
  every column outside the provider's own list went back to its default, taking
  price_1st_edition (written only by tcgcsvApi) with it on every price sweep. It is
  an upsert now, which also protects the learned English name above.

• Korean Pokémon scans found nothing at all. The same photo scanned as English
  named the card immediately. A candidate from the ready-made Pokémon catalog is a
  TCGplayer product id, so it is resolved by set and collector number — and that set
  id is always an English one, while Korean, Japanese and Chinese Pokémon sets are
  their own releases (SM1M, S12, SV2a) rather than localised editions of the English
  ones. Asking TCGdex/ko about base6 therefore returned nothing for every
  candidate, and a scan that had identified the card correctly reported "no confident
  match". The lookup now falls back to English when the scanned language has no such
  set, in the route and in the client's own retry, and the English card it hands back
  says so: the picker, the auto-add window and the add drawer all print a note that
  no card data exists in that language and this is the English printing standing in
  for it. The copy is still filed in the language being scanned, which it already
  was.

• The scanned card now reads as the card. Its name and collector number were the
  smallest text on screen in both the match picker and the add drawer, under a
  picture and a title that say nothing about which printing this is; they are now the
  largest, with the set name demoted beneath them. On a phone the add drawer was
  tuned to fit inside 80vh by shrinking everything — 65px of card art, 0.7rem labels,
  40px controls — which left a cramped strip at the bottom of a mostly empty screen;
  it now takes 92dvh (dvh, so the address bar cannot cut it off), with 104px of art
  and 44px touch targets. "Different printing" is gone from the drawer — it asked
  the provider for every printing of the name, which is a search, not a scan
  correction. "Other matches" stays: it is the escape for the scanner picking the
  wrong card.

• Japanese Pokémon scans returned the wrong card. Three causes, all of which
  had to go:

  • A cosine sweep never returns nothing. TCGdex serves card records for 28 of
    the 177 Japanese Pokémon sets it lists, so a Japanese catalog holds ~3.3k of
    ~16k cards — and every card outside those 28 sets was answered with the nearest
    of the wrong 3.3k, sometimes at a similarity high enough to auto-add. A scan in
    a non-English language now also sweeps the English catalog, because the artwork
    is identical across languages and English has a row for nearly every card the
    other catalog is missing; the answer is then re-expressed in the scanned
    language by set and collector number. The right card in the wrong language beats
    a wrong card in the right one. Verified on Japanese cards matched against the
    English catalog alone: スズナ → Candice, モンジャラ → Tangela, ダブラン → Duosion.
  • A card the catalog has never seen now says so. The response carries
    notInCatalog when the winner does not stand clear of ranks 2–11 of its own
    catalog, the client refuses to auto-add on it, and the candidate list is still
    shown — a bad photo and a missing card look identical from the server, and one
    of the candidates is right often enough to be worth the glance. The gate is that
    gap and not an absolute cosine because absolute cosine tracks photo quality:
    over 60 cards per catalog, a 0.65 cosine floor let 31–41 of 60 strangers through
    on clean input but 15–19 on degraded input, while a 0.10 gap held at 11–12 in
    both, rejecting no correct answer in any run.
  • Non-English artwork was embedded at thumbnail resolution. TCGdex's cached
    image url is /low.png (245×337, deliberately small so card grids stay cheap),
    so every TCGdex row was embedded from an upscaled blur while the camera handed
    over a sharp 448 crop. Builds now embed /high.png (600×825) without changing
    what the UI loads, and resume compares the url actually embedded, so raising
    the resolution rebuilds the old vectors instead of silently keeping them.
    Rebuild a Pokémon catalog to pick this up.

• A set-scoped scan no longer widens itself. Set ids do not survive a
  language — SV4a names no row in the English catalog — so a catalog with nothing
  in scope is dropped from the sweep rather than searched unscoped, which would have
  reintroduced the very wrong-card answer the scope exists to prevent. The filter is
  only ignored when no catalog has rows in scope, since "no match" for a card
  that is plainly there is the worse failure.

• Scanning could not work in Docker at all. The image never carried the two
  ONNX models and nothing fetched them, while CV_MODEL_DIR defaulted to a path
  inside the image — so a container had no models, and any catalog an admin did
  manage to build was discarded by the next image update. The image now sets
  CV_MODEL_DIR=/app/database/models on the persisted volume, scripts/fetch-models.mjs
  provisions the models (and optionally the published fallback catalogs) with their
  sizes verified, and startup names that command when they are missing instead of
  leaving the first scan to fail. SETS_DIR and INDEX_DATA_DIR are gone with the
  indexes they pointed at.

• Removed the "let members build individual set indexes" setting. There are no
  per-set indexes to build; catalog builds are admin-only because they walk an
  entire provider. The toggle, its API field and its strings are gone. The column
  stays unread — dropping it would mean a SQLite table rebuild for nothing.

Changed

• opencv-wasm is gone from the scan path; the runtime is onnxruntime-node
  on the server and onnxruntime-web in a worker on the client. Name the execution
  provider explicitly in the browser: WebGPU measured 1075 ms per frame for
  cornelius against 80 ms on wasm.
• The Scan Detail slider's recall knobs are inert. Every scan is one 448px
  embed and one cosine sweep per catalog, so recallK/orb tune a pipeline that no
  longer exists; upload resolution and the auto-add confirm window still do what
  they say. The request keeps sending both fields so an older backend behaves.