Fix: replace api.pokemontcg.io with api.tcgdex.net (#20)

core/AGENTS.md

@@ -8,7 +8,7 @@
 - `include/ccm/ports/` — interfaces (`IHttpClient`, `IFileSystem`, `ICollectionRepository<T>`, `ISetRepository`, `IImageStore`, `ICardPreviewSource`, `IPreviewByteCache`). All seams the services depend on. Add new ports here when adding new external concerns.
 - `include/ccm/infra/` — concrete adapters: `CprHttpClient`, `StdFileSystem`, `JsonCollectionRepository<T>` (header-only template), `JsonSetRepository`, `LocalImageStore`, `LocalPreviewByteCache`.
 - `include/ccm/services/` — high-level operations: `ConfigService`, `CollectionService<TCard>` (header-only template), `SetService`, `ImageService`, `CardPreviewService`, `CardSorter` (free functions; per-column sort comparators that mirror established table sorting behavior — UI-agnostic so they can be unit-tested directly), `CardFilter` (free functions; case-insensitive substring row matcher restricted to each game's `tableFields` valueKey list), `YuGiOhSetCompletion` / `DigiBattle99SetCompletion` / `PokemonSetCompletion` (pure set-completion / checklist helpers), `YuGiOhSetCatalogService` (`yugioh/set-catalog.json`), `DigiBattle99SetCatalogService` (`digibattle99/set-catalog.json`), `PokemonSetCatalogService` (`pokemon/set-catalog-west.json` / `set-catalog-asia.json`). They depend only on ports / domain.
-- `include/ccm/games/` — `IGameModule` + per-game modules. `IGameModule` consolidates the per-game seams: every module owns an `ISetSource` (required) and may own an `ICardPreviewSource` (optional, default `nullptr`). `magic/`, `pokemon/`, `yugioh/`, `digibattle99/`, and `pokemonjp/` are the reference implementations — all five expose a fully working set source + card preview source. `YuGiOhSetSource`, `DigiBattle99SetSource`, `PokemonSetSource`, and `JapanesePokemonSetSource` also expose `parseCatalog` / `fetchAllWithCatalog` (or Asia equivalents) for set-completion checklists. `pokemonjp/` is the **Asia region backend** for the unified Pokemon UI (set cache at `pokemon/sets-asia.json`, same data dir as West; TCGdex JA previews); it is registered for sets/previews but is not a separate Game menu entry. Japanese Pokémon also loads an optional EN name catalog (`JapanesePokemonEnCatalog`) for display/auto-detect / Asia set-completion gap-fill.
+- `include/ccm/games/` — `IGameModule` + per-game modules. `IGameModule` consolidates the per-game seams: every module owns an `ISetSource` (required) and may own an `ICardPreviewSource` (optional, default `nullptr`). `magic/`, `pokemon/`, `yugioh/`, `digibattle99/`, and `pokemonjp/` are the reference implementations — all five expose a fully working set source + card preview source. `YuGiOhSetSource`, `DigiBattle99SetSource`, `PokemonSetSource`, and `JapanesePokemonSetSource` also expose `fetchAllWithCatalog` (and related catalog parsers) for set-completion checklists. `pokemonjp/` is the **Asia region backend** for the unified Pokemon UI (set cache at `pokemon/sets-asia.json`, same data dir as West; TCGdex JA previews); it is registered for sets/previews but is not a separate Game menu entry. Japanese Pokémon also loads an optional EN name catalog (`JapanesePokemonEnCatalog`) for display/auto-detect / Asia set-completion gap-fill.
 - `include/ccm/util/` — `Result.hpp` (the sum type), `FsNames.hpp` (filename munging ported from `util/fs.rs`), `YuGiOhPrintingSlot.hpp` / `YuGiOhSetLookup.hpp` (Yu-Gi-Oh! print-slot helpers and cached-set **set code** lookup for the edit dialog; both header-only, unit-tested).
 - `src/` mirrors `include/ccm/` for non-template implementations.
 
@@ -27,7 +27,7 @@
 8. **HTTP query strings must be percent-encoded** before they reach `IHttpClient::get`. `cpr::Url` does **not** encode the URL string we hand it. See `MagicCardPreviewSource::buildSearchUrl` for the canonical pattern (RFC 3986 unreserved-set encoder). `IHttpClient::get` accepts arbitrary bytes back — `Result<std::string>` is a binary buffer, not text, so callers can use it for image payloads directly.
 9. **Yu-Gi-Oh! preview uses Yugipedia, not YGOPRODeck.** `YuGiOhCardPreviewSource::fetchImageUrl` queries Yugipedia's MediaWiki API with a batched list of deterministic file names (`<Slug>-<SET>-<REGION>-<RARITY>-<EDITION>.<png|jpg>`) so per-printing reprints with shared passcodes (LOB Blue-Eyes vs SDK Blue-Eyes, …) resolve to genuinely different scans. Region candidates are **always English** (`EN`/`NA`/`EU`/`AU`) regardless of `card.language`; localized scans are not queried. YGOPRODeck remains as a last-resort fallback (see `parseFallbackImageUrl`) for cards Yugipedia hasn't scanned yet, and as the source for `detectFirstPrint` / `detectPrintVariants` (`parsePrintVariants` enumerates distinct printings for the edit dialog). **Do not** restore a YGOPRODeck-only image path: that endpoint's `card_images` array is keyed by art-treatment passcode, not by physical printing, and adding `cardset=` only reorders the same passcode list (alt-art often gets promoted) without ever surfacing the per-printing scan. The YGO source therefore needs the printed edition flag to be plumbed through; `YuGiOhSelectedCardPanel::previewKey()` packs it into the third tuple slot as `<setNo>||<rarity>||<1E|UE>` so the candidate list can prioritize the correct edition without changing the generic `ICardPreviewSource` interface.
 10. **Preview byte cache (`CardPreviewService`) is by `(game, name, setId, setNo)` across two tiers, with classified failure caching and a single update mechanic.** Successful `fetchPreviewBytes` results and successful `fetchImageBytesByUrl` results are stored first in a bounded in-memory LRU (`kCacheCapacity` entries, mutex-protected — the panel calls into the service from a worker thread) and then in an optional persistent byte cache (`IPreviewByteCache`, normally `LocalPreviewByteCache` rooted at `<exeDir>/.cache/preview-cache/` — next to the executable, **not** under `dataStorage`, so previews don't follow the user's collection when the data-storage path is reconfigured). **`fetchAndCache` rejects empty response bodies** (returns error, no tier write) so a degenerate HTTP 200 cannot fill the LRU with unusable entries. Lookup order is **memory → disk → source/HTTP**, and a disk hit (positive *or* negative) is promoted into the in-memory tier on its way to the caller so the next selection of the same row stays decode-only. **Failures are split by `PreviewLookupError::Kind`**: `NotFound` is negative-cached in both tiers (memory `CacheEntry::negative=true`, disk `<hash>.neg` marker) so the user gets an instant card-back on every subsequent click for cards whose printing genuinely has no upstream image; `Transient` (HTTP/network/parse failures) is **never** cached so a brief outage cannot permanently disable previews. Per-game `ICardPreviewSource::fetchImageUrl` implementations must classify their errors honestly — `NotFound` only when the upstream answered cleanly with no match / no image variants; anything that could be the network or a schema deviation is `Transient`. **The cache update mechanic is entirely key-driven and has no side-channel API:** (a) the user editing any lookup-relevant field of a card record changes the cache key, so the next selection misses both tiers and re-runs the source — this is how a stale negative entry gets dislodged after the user fixes the record, with no manual invalidation call needed; (b) a same-key resolution that flips between positive and negative outcomes overwrites the existing entry in both tiers (`store` removes any `.neg` for that hash; `storeNegative` removes any `.bin`) so `.bin` and `.neg` for the same hash are never co-resident; (c) eviction handles passive aging (LRU on the in-memory tier; oldest-by-mtime `.bin` files on the disk tier; `.neg` markers don't count against the size cap and are not actively evicted). **Do not add a `clearCache(...)` / `invalidate(...)` method** to `CardPreviewService`: the cache invariants depend on memory and disk staying aligned through the same write paths, and any side-channel API would just be a new way for future code to forget the disk tier. If you add a new lookup disambiguator (for example a future `editionTag` slot), pack it into one of the existing key fields (see `YuGiOhSelectedCardPanel::previewKey()`'s `||`-separated trailing fields) so editing the field continues to invalidate cached entries automatically. The persistent tier is **fire-and-forget**: the adapter swallows I/O errors so a flaky or full disk degrades the experience to a fresh-install warm-up, never to a broken preview path.
-11. **`CprHttpClient` keeps one persistent `cpr::Session` for the app's lifetime.** All callers (set sources, preview sources, fallback URL fetch, auto-detect) share the same libcurl easy handle so connections to repeat hosts (`api.scryfall.com`, `api.pokemontcg.io`, `api.tcgdex.net`, `assets.tcgdex.net`, `product-images.tcgplayer.com`, `db.ygoprodeck.com`, `yugipedia.com`, `ms.yugipedia.com`, `digimoncard.io`, `images.digimoncard.io`) are reused with TLS keep-alive. Default request headers use **`Accept: */*`** so JSON endpoints and binary image downloads share one session without pinning every GET to `application/json`. The session is not thread-safe — every `get(...)` is serialized through an internal mutex. **Do not** construct a new `cpr::Session` (or `cpr::Get(...)`) per call: that throws away the connection cache and re-pays the TLS handshake every time. If you need richer behavior on the port (POST, headers per call, …) extend `IHttpClient` and the adapter while keeping the single-session ownership intact.
+11. **`CprHttpClient` keeps one persistent `cpr::Session` for the app's lifetime.** All callers (set sources, preview sources, fallback URL fetch, auto-detect) share the same libcurl easy handle so connections to repeat hosts (`api.scryfall.com`, `api.tcgdex.net`, `assets.tcgdex.net`, `product-images.tcgplayer.com`, `db.ygoprodeck.com`, `yugipedia.com`, `ms.yugipedia.com`, `digimoncard.io`, `images.digimoncard.io`) are reused with TLS keep-alive. Default request headers use **`Accept: */*`** so JSON endpoints and binary image downloads share one session without pinning every GET to `application/json`. The session is not thread-safe — every `get(...)` is serialized through an internal mutex. **Do not** construct a new `cpr::Session` (or `cpr::Get(...)`) per call: that throws away the connection cache and re-pays the TLS handshake every time. If you need richer behavior on the port (POST, headers per call, …) extend `IHttpClient` and the adapter while keeping the single-session ownership intact.
 
 ## Adding a new game