[plainprogrammer]

Scanner: the selection pane (spec 088)

Closes #345.

After a capture the collector sees their capture beside what was read, a status band that says how sure the scanner is, a row of card images (swipe, or Previous and Next) with the best match centred, one details block for the centred card, and one add button per finish. A confident match is one tap; the band says in an icon and words when the scanner is less sure. The interim scrolling body, "What the scanner read", the candidates grid and the Other printings frame are gone. Nothing about how a card is read or ranked changed (ruled 2026-11-13).

Spec: docs/specs/088-scanner-selection-pane/spec.md (1.0.3, approved at Gate A; the maintainer ruled the seven scope questions as recommended). Plan: plan.md (approved at Gate B). Verification and the implementation review: verification.md.

What it does

• Band, three states from today's evidence (no change to the ranking; one new predicate, Candidate#card_sure?): success with the evidence ("Matched by its collector line", "…, one digit corrected", "Matched by its artwork"); "Printing not confirmed. Swipe to check."; "Not sure of this card. Check it against your capture."; "Not the printing that was read" on any other printing. One persistent role="status" element, replaced in place from the centred panel's <template>; its hidden tail names the centred card so every swipe is announced.
• Row: sure of the card: the matched printing first, then the card's other printings in Other printings order, 20 at a time ("1 of 14 printings", a "‹n› more printings" item with "Show 20 more", capped at 200); not sure: the candidate cards ("1 of 3 possible cards") with "All printings" per card. One Turbo Frame (scanner_selection) renders the row and its panels; a small selection Stimulus controller notes the centred card; Previous, Next, arrow keys and a tap on a neighbour do what a swipe does.
• Strip: the capture's card box beside the name and collector line read ("nothing read" when blank), written as text.
• Outcomes: nothing read and no match are said on the live camera's status line (no pane, retake at once); the catalog not ready and refusals are visible alerts in the pane's region; a send failure shows the strip with Send again / Sign in. "Done." plus the band's sentence is a visually hidden paragraph written once the pane shows.
• Layout (decision 13, a visible design change the maintainer was told of): the images' floor is 96px (spec 080 had 160px) and a shell under 600px high has a 64px strip with a 52px picture whose label is visually hidden; the add buttons are in view for a confident match at 390 × 844, 360 × 640 and 360 × 560. The numbers live in two places only (--scan-image-min and one container block in additions.css), proved by the pane-fit examples. This supersedes spec 080 decision 7 and AC-4.7 for those two numbers (recorded both ways in the specs index).
• Measured at rest (confident / not sure): 390 × 844 strip 144, images 288; 360 × 640 strip 96, images 160 / 122; 360 × 560 strip 64, images 112 / 96 (the not-sure body scrolls about 6px, the accepted limit).

Supersedes (specs index, both ways)

Spec 009 (the candidates grid, the tiles' badges, the Other printings frame, "What the scanner read" on the page), spec 011 AC-7.1 to 7.5 as the page shows them, spec 080 decision 7 / AC-4.7's two numbers, spec 082's interim body and spoken "Done" sentence. Area doc: docs/architecture/scanner.md; design system: Scanner.md, StatusBand.md.

Screenshots (factory data; card images are placeholders: a tile with the name and set number)

Phone 390 × 844, light and dark: match · dark · swiped light dark · printing not confirmed light dark · not sure light dark · nothing read light dark · failure light dark · after an add light dark

Short phone 360 × 560: match light dark · not sure light dark

Desktop (a 480px column): light · dark

Flow (capture, result, Next, Previous, Add foil, after-add line):

The repository is private, so the links are pinned to the commit and need a signed-in reader. Every shot was looked at before the specs were trusted; the one defect found (a cramped 64px strip) is fixed.

Verification

• bin/ci ran as the pre-push hook on 57504170 and passed; the suite has 5,205 examples (bin/rspec --dry-run). Scanner suites run green one at a time after each task; bin/ci checks passed after merging origin/main (spec 091's licence check made me list the two new icon files in config/third_party_licences.yml).
• Implementation review (Opus 5.5): SPEC-ALIGNED WITH FINDINGS (0 blocking, 2 major, 7 minor), all fixed and re-proved; the 200 cap and "Send again" clearing the strip were the majors.
• The scanner CSP pin (spec/requests/footer_spec.rb) and spec/scanner_sitting_core_spec.rb pass; every Turbo Stream string is escaped or from a partial; measurement mode's run-sheet row logic (Measurement mode: a retake files under the wrong card #419) is untouched.

Models

Spec and plan: main session (Sonnet 5.5; Opus 5.5 for a stretch the harness reported, trailers truthful). Spec reviews x2 and plan reviews x2: Fable 5.2, read-only. Implementers: Sonnet 5.5 (batches for Tasks 1 to 10, the screenshots, the strip fix, the review fixes). Implementation review: Opus 5.5.

Deviations from the plan

Red windows by design (named in each commit): Task 3 request specs until Task 5; the old scanner system specs from Task 5 until Task 8's migration. The spec was patched while building: 1.0.1 (a possible-cards row drops the year and lets its position wrap to fit 360px), 1.0.2 (AC-5.2's measured 6px scroll, not 22px), 1.0.3 (the compact strip hides its label). The plan's tests-first is visible in the history only where a task had its own test commit (58539866); others carry spec and code together and name the red run in the review fixes' bodies.

Owed to the maintainer

• The screen-reader pass on a phone (a result opening, a swipe, Previous and Next, an add, a refused add; each sentence said once, a fling announcing once). The examples check the text and the element.
• A look at the pane over the live camera with real card art (the shots use the substituted camera and placeholder images). No measured sitting is asked.
• Decision 13's numbers were a design change they were told of; a different ruling is a change to the two places above and their examples.
• Headless Selenium reads no console here, so "no console error" is not claimed; script/scanner/detector_parity.rb needs the spike corpus (the detector and photo path are unchanged).

🤖 Generated with Claude Code