CardVision catalog gaps: missing new prints + EN/JP twins with different names never grouped #517

New Issue

2026-08-29T16:10:09Z

gambit-admin commented

2026-08-29 16:10:09 +00:00

Found during Kris's 84-card live session (2026-08-29), from the persistent scan log:

Mightyena 243/217 (SV Black Bolt/White Flare secret rare) is absent from the index — set-mates 237/245/248/217 all present and matching. The scan correctly returned low_confidence but the real card could never be offered. Audit how the index build/incremental add missed it (card_cache row missing? image fetch failed silently?) and add a coverage check: for sets we carry, alert on gaps in the numbering run.
EN/JP art twins with different localized names are never variant-grouped. Scan matched JP "Misty's Spirit 108/081" at 97 inliers with EN "Misty's Vitality 111/084" as runner-up at 47 — same art, but the picker only offered the JP group, so the user could not pick the English print they were holding. The geometric art-identity twin pass groups same-name prints; cross-language pairs with translated names fall through. Likely also stale for incrementally-added rows (twin groups computed at full build only).

Both are index-build lane, not scan-time fixes. Related: #515 (session also produced the blur false-accept findings, fixed server-side same night).

Filed from a Claude Code session

Found during Kris's 84-card live session (2026-08-29), from the persistent scan log: 1. **Mightyena 243/217 (SV Black Bolt/White Flare secret rare) is absent from the index** — set-mates 237/245/248/217 all present and matching. The scan correctly returned low_confidence but the real card could never be offered. Audit how the index build/incremental add missed it (card_cache row missing? image fetch failed silently?) and add a coverage check: for sets we carry, alert on gaps in the numbering run. 2. **EN/JP art twins with different localized names are never variant-grouped.** Scan matched JP "Misty's Spirit 108/081" at 97 inliers with EN "Misty's Vitality 111/084" as runner-up at 47 — same art, but the picker only offered the JP group, so the user could not pick the English print they were holding. The geometric art-identity twin pass groups same-name prints; cross-language pairs with translated names fall through. Likely also stale for incrementally-added rows (twin groups computed at full build only). Both are index-build lane, not scan-time fixes. Related: #515 (session also produced the blur false-accept findings, fixed server-side same night). --- Filed from a Claude Code session

gambit-admin added the bug api labels 2026-08-29 16:10:09 +00:00

gambit-admin added the 01 · scan & recognition label 2026-09-04 13:36:54 +00:00

gambit-admin referenced this issue

2026-09-21 03:14:37 +00:00

Review follow-ups: PR #517 — feat(cardvision): on-device CardVision + embed on iPhone, search on the server (Gitea #632) #762

gambit-admin referenced this issue

2026-09-25 04:40:02 +00:00

CardVision: embed on the phone, final match on the CardVision server #824

gambit-admin referenced this issue

2026-09-25 04:40:08 +00:00

EPIC: On-device CardVision — recognize cards on the iPhone without Tower #632

gambit-admin referenced this issue