A stone importer's trade catalog: ten slabs on the racks at two New Jersey yards, browsed by facet, and then one slab's specification page.
Direction
| Lever | Decision |
|---|---|
| Archetype | Index, faceted. A sticky rail of native checkbox facets against one list that takes the rest of the width. |
| Data stance | Table-first. The slab inventory table is the product, and the photograph is a column in it rather than an illustration beside it. |
| Density | Default. --mc-space-md inside a row, --mc-space-xl between groups, --mc-space-3xl between sections. |
| The one bold thing | Stock carries the minute it was counted, and the minutes differ by row. No specimen frame holds anything but the material. A sold slab keeps its page and routes to the two sequential cuts behind it. |
Why this product suits this system
material-catalog is the only system here written for a business whose product is a physical object photographed under one lighting setup, and whose photograph is a claim somebody can send back on a truck. Its non-negotiables are all about not getting in the way of that: a fixed neutral surround identical in both themes, no radius on a specimen, no tint over a material, and a scale solver that decides how wide an image has to be before a 4 mm vein exists in it.
A slab yard pushes every one of those. Each slab is unique, so the index cannot be a grid of one hero shot per material; stock goes stale in hours, so a figure with no timestamp is a lie; and the buyer's four questions (finish, thickness, slab dimensions, lot) are exactly what a table answers and a card does not.
Generator
mat_tokens.py --both > tokens.css
The brief recorded mat_tokens.py with no flag, which emits light only. --both was run instead so the page can meet CONTRACT.md §5 on themes; its :root block is byte-identical to the flagless output, and the flag adds [data-theme="light"], [data-theme="dark"] and a prefers-color-scheme block. The generator's own --self-test holds the specimen tokens identical across the two themes, which is the reason a dark rail beside a light specimen grid is legitimate here.
The scale, solved rather than chosen
mat_tokens.py --scale
- A row's detail plate is 168 px of a declared 300 mm field of view, 1.79 mm per pixel, which puts a 4 mm feature across 2.2 px and clears the system's
--mc-px-per-feature: 2floor. It is what makes slab 01 tell itself apart from slab 02 in the index. - A whole slab at that width would be 18 mm per pixel, so the full-piece view is not in the table at all. It is the stage of the viewing bay, drawn from a 1440 px file at the width the screen allows; the page computes the millimetres per pixel from that width and prints it beside the picture, so the figure is always true for the screen in front of you. The bay's own close crop is 560 px for 300 mm.
Deviations from CONTRACT.md
- §2, images. Photographs are image files under
img/(AVIF for the whole-slab and large detail images, WebP for the table plates) rather than inlinedata:URIs. Base64 inside the document would add roughly a third to about 1.1 MB of imagery and put it all in front of the first paint on a page whose subject is a photograph. Every<img>still carries explicitwidth,heightand a realalt. - §1, tokens.
tokens.csswas regenerated with--bothrather than patched. See the Generator section.
The uplift
What changed, in the order a visitor meets it, and why.
- A viewing bay opens the page. One whole slab, large, on the neutral grey, with its serial, lot, rack, a stock line carrying its count time, the one action that fits its state (hold, ask to be next, or see the cut beside it), a 300 mm close crop and a rack of all ten slabs under it. The slab named in the list, the rack and the address are the same slab. The material is the hero; the table is no longer the first thing a buyer meets.
- The index answers to the facets and the search. Counts are computed with each group's own selection left out, so a count says what ticking it would give. A filter that empties the list names the sentence it failed on and offers the one filter whose removal restores the most slabs, with the count.
- A layout studio lays a job on the photograph. Four jobs (L-shaped kitchen, waterfall island, one-piece island, vanity) as outlines on the actual slab at its own scale. Yield, offcut, seams and slabs needed are computed from integer square millimetres; pieces keep 20 mm from the edge and 6 mm apart; a piece that does not fit moves to a second-slab tray and the page says which slab in the lot can take it. Pieces drag, or move with the arrow keys (Shift for 50 mm, R to turn).
- Every request restates what was chosen. Hold, ask-to-be-next, quote and samples open a native dialog that shows the slab, size, lot, place, hold end time and, when it came from the studio, the pieces and the yield. A confirmed hold changes the row, the rack, the ledger and the filters.
- The slab page is built from the slab on show. Specification rows carry a unit and a test method; values the yard has no result for read "Not tested" and say what not to do on that basis. The formats matrix is read from the list, so it cannot disagree with it. A sold slab keeps its page and points at the cuts behind it.
- Every figure checks against the others. Gross areas are shown in hundredths of a square metre, rounded half up once, and every total on the page is a sum of figures that are shown (all ten 46.13 m2; racks without the sold slab 41.39 m2).
- On a phone the filters become a bar at the foot of the screen while the list is on screen, rows become blocks of specifications, and the studio and dialog re-flow to one column and a bottom sheet.
Two choices that depart from the first brief. Studio pieces are drawn as outline boxes over the photograph, in a layer that is not inside the specimen frame, so the frame still holds nothing but the material. The page's date is Friday 11 September 2026, 09:15, and every as-of time agrees with it.
Gates
catalog_lint.py, audit_page.py, slop_scan.py, craft_lint.py, motion_lint.py, layout_lint.py, copy_lint.py, direction_check.py, contrast.py and aria_lint.py. Output is in the handover.
Round four
- On a phone the sticky header is one 56px row (the way back, the name, a search icon, the theme icon). The search field and the downloads link open under it when the icon is pressed, and open by themselves when a search is already in the address. From 48rem nothing changed.
- The table's caption, left as a table-caption inside the stacked block table, shrank to its longest word (58px, one word a line). It is a block there now, and the stacked roles restate it.
- The two scales disagreed because they described two drawings of one crop: the plates in the list (166px, 1.81 mm per pixel) and the larger detail in the viewing bay (about 0.94). Both are read from the image on the page, and the list's note now says the bay draws the same crop larger.
Round eight
- 390: the stock list ran about 14,000px of repeated rows (real: the page was 13,941px, the list 4,253 of it). Cause: each slab stacked as one 421px card, plate beside name, then five label-and-value rows, so ten slabs were ten screens. Below 48rem the cards now sit two across: the plate fills its half (
.spec-thumbiswidth: 100%with a square aspect, no longer the fixed 168px), serial and stock state under it, and the facts as label-over-value pairs (class beside finish, lot beside yard, size across). The list is 2,605px, page 12,293px, and nothing scrolls sideways at 390 or 320. The plate scale line reads off the rendered plate, so it follows (1.69 mm per pixel at 390). The labels stay as::beforetext, so the restated table roles and their generated labels are unchanged. - 390: rack cards sit right of the viewport edge, plates to x=456 (not a defect). Those are the viewing bay's rack, not the stock list: ten radio plates in a scroll-snap strip (
#rack,overflow-x: auto, scrollWidth 1120 in a 366px box). The fourth plate begins at 352px and is cut by the strip's own edge, which is the affordance that it scrolls; the document itself is 390px wide, so nothing is clipped or pushed out of the page. Measured withgetBoundingClientRecton.rack__plate(16 to 120, 128 to 232, 240 to 344, 352 to 456, ...) andscrollWidthof the document (390). Left as is.
Round nine
Method: a fabricator's buyer in a browser (filter the yard, pick a slab, compare, place a hold, lay out a job), then real-phone emulation at 390 and 360, 200% zoom, dark mode and a keyboard pass. Figures checked against each other: gross areas sum to 46.13 m², less the sold 4.74 m² gives 41.39 m² on the racks; 8 of 10 serials available; every weight equals density x length x width x thickness.
- Viewing bay recomposed. The facts (finish, thickness, size, gross area, weight, rack) moved from the right column to a strip under the rack, so the grey no longer ends in a void below the plates and the rack serials no longer wrap between 1024 and
- Right column is 22rem: name, state, the two actions, then the detail crop. Below 64rem the bay flattens and the parts are ordered by
order. - Full resolution. The detail crop opens a dialog with the 560px file at its own scale ("0.54 mm per pixel", "a 4 mm vein is 7 pixels wide here"), from the button under the crop or a click on the crop. Escape returns focus to the button.
- Card layout to 88rem. The list table did not fit below about 1410px and scrolled sideways, cutting off Status at 1024 to 1366. Rows now stack into cards from 88rem down (auto-fill columns of at least 11.5rem, two across under 48rem). The picked card carries a 2px outline on its plate. Plates carry
srcset(336 and 560), so the earlier 168px, 1.81 mm per pixel claim holds only for the 336 file; with the 560 file the plate reads about 1.3 to 1.7 mm per pixel, and the printed scale reads off the rendered width. - Phone tables. The specification table stacks (value level with the property, test method beneath) and the formats matrix stacks with labelled cells; before, the method and Honed columns were clipped. Table roles are restated per layout.
- Header. The phone header applies up to 64rem (it wrapped to 121px at 768 to 1100); 57px at 320 and 360, down from 105. Nav shows from 88rem.
- Fixes. The Privacy section sat outside the page gutter at half width, now inside it. The ticket hairline in the dialogs had a gap.
[id]scroll margin double-counted thehtmlscroll padding and was removed. The slab-name link has a 44px hit area. The pieces table no longer scrolls at 320. Em dashes removed from this file and from CREDITS.md.
Round ten
A second, colder read as a fabricator's buyer placing a hold on a one-of-a-kind slab. Score on the ten heuristics, 27 before and 32 after; user control and freedom was 1 (the hold had no way back on the page), error prevention 2 and help 2 (the hold's end time and cost sat only inside the form).
- A hold had no way back (severity 3). The only release was "cancel by phone or by replying". Once a hold is confirmed, the hold button becomes "Release my hold on VS-...", in the bay, on the slab page and in the studio. It opens a dialog that says first what ends (the slab goes back on sale at once, anyone can hold or buy it, the hold cannot be brought back), then the serial, end time, reference and who it is held for. "Keep the hold" is the primary button and has focus; Escape and a click on the backdrop also keep it. "Ask to be next" got the same treatment: it now shows "You are next in line" and a "Leave the queue" button.
- The terms were in the wrong place (severity 2). A line beside every hold button now says it is free, when it ends and how to release it, before the press. In the hold form the ticket opens with slab, end time, cost, how to release it and what happens if nothing is done, then the stone details. On a phone the form no longer scrolls past those rows when it focuses the name field.
- The page did not change after the act (severity 2). The hold used to rebuild the whole slab view, which reset a layout the buyer had just made in the studio (and the studio hold button was a dead disabled one). It now updates only what the hold touches: the row, the rack plate and its accessible name, the ledger, the filter counts, the formats matrix, the lot strip and every button and note for the slab. Held-by-you reads "Held for you" everywhere the slab's state is shown.
- Focus (severity 2). After "Done", focus returned to a button that had gone disabled. The opener is now the release button, so focus stays on it.
- Serials in dialog headings no longer break at their hyphens (severity 1).
No copy-lint.ignore exists for this page, so there was no comment to check.
Inventions, named because the brief gives no rule for them: a hold ends by itself at its end time and the slab goes back on sale; releasing is immediate and unrestored; "Nobody rings you; the confirmation is emailed" and "to keep it longer, ring the yard"; a released slab reads "Released 09:15 today" (the page's own time); leaving the queue closes its reference.
Round eleven (2026-11-10)
An independent reviewer found that "Turn piece" in the layout studio dropped the Slabs figure from 2 to 1 and left "Request a quote for this layout" enabled for a layout that cannot be cut (severity 2).
- What went wrong. The layout was genuinely invalid (a turned 2800 mm piece does not fit a 1570 mm width, or lands on a neighbour), and the page did say "Check: ..." but nothing else followed. Two defects sat under it. A piece had one status, and a clash overwrote "second slab", so B and C on the second slab stopped being counted and Slabs fell to 1. And nothing tied the quote button to the verdict.
- Fix. Where a piece lies (slab 1 or 2) is now kept apart from what is wrong with it (outside the edge margin, larger than the clear area of the slab, closer than the saw kerf to another piece). Slabs counts the pieces on the second slab whether or not they clash. A piece that is wider or taller than the room inside the margin now says so, with the slab's own clear size (the slab length and width less the 20 mm margin on each side), instead of only naming the neighbour it overlaps. While the layout cannot be cut, the message beside the figures opens with "This layout cannot be cut as it stands", lists each piece and why, and says what to do; "Request a quote for this layout" is disabled and described by that message. A valid turn (for example the waterfall end in the tray, or the vanity splashback moved clear of the top) keeps its yield, offcut, slab count and an enabled quote.
- Held back from the figures. Yield and offcut show an en dash while the layout cannot be cut, because a percentage computed from the pieces that happen to fit reads as a result. The hold dialog no longer carries the layout rows when the layout is invalid, since it would print a yield for a layout that does not exist; a hold is still allowed, it is about the slab.
- Invented, named. The sentence "Yield and offcut are counted once every piece fits; move or turn the pieces to get there" and the en dash placeholder. The 20 mm margin, 6 mm kerf and slab sizes are the page's own. Not touched: the "1 slabs" plural and the phone first fold, both severity 1 notes outside this brief.
Round thirteen (2026-11-10)
An independent reviewer found that choosing a slab in the list scrolled the page back up to the viewing bay, so a reader comparing three rows lost their place each time (consistency, severity 2 in effect).
- Reproduced at 1440, 390, 360 and 1920: a name in the list was an anchor to the bay, and the click landed the page 1,400 to 2,100 px above the row (scroll position 1655 to 227 at 1440). Focus was also lost to the body after the jump.
- Fix. A name in the list no longer navigates. The bay, the rack and the page of the slab still update, and a bar that stays with the list (
.peek, sticky under the header, above the table) shows the slab on show: its rack plate, serial, name, class, finish and thickness, and its stock state with the minute it was counted. The chosen row keeps its place: its top edge is measured before and after and put back if the tag that moves from the old row to the new one changed the height of the cards above it (on a phone it did, by 26 px). Focus stays on the name that was chosen, so a keyboard reader can press Enter on the next row. On a phone the bar drops the plate and the class line and keeps name, state, as-of and the button, so it stays between 82 and 104 px tall. - The way up and the way back are both on request. "Open in the bay" in the bar scrolls to the bay and focuses the slab name there; "Back to the slab list" beside "Lay out a job" in the bay returns to the row that is on show, centred, with focus on its name (or to the list heading if no row is shown). Links from below the list (the sold slab mates, the bundle, "Lay out on") still go up to the bay as before. The rack radios do not scroll.
- Invented, named. The words "In the bay" (the row tag already says it), "Open in the bay" and "Back to the slab list"; the two sentences in the list note that say choosing a name keeps you in the list and where the slab then shows. No product fact was added. The bar has no hold button: the hold stays in the bay, one copy of that logic.
- Driven. 1440 light, 390 dark and 360 light (touch, DSF 3), 1920 dark: row top identical before and after a click or tap and after Enter on a second row; bar updates, focus on the name, no hash change, no overflow, no page errors; up and back round trip returns to the row with focus visible by keyboard.
Round fifteen (2026-11-10)
An independent reviewer found that a search with no match showed nothing near the box (severity 2): typing "zzzz" and pressing Enter left the viewport unchanged, and the answer, "Nothing in the yard answers to zzzz", sat about 1,000 px below the pinned header.
- Reproduced at 1440 light, 390 dark, 360 light and 1920 dark: after Enter the scroll position was 0, no status existed near the field, and the list heading read "0 slabs" and, for one match, "1 slabs".
- Fix. A status note now hangs under the search box from the pinned header (
#q-status,.searchnote), floating over the page so the header does not grow. Its message sits in arole="status"paragraph that is always in the document, so each change is announced; the box points at it witharia-describedby. No match: "Nothing in the yard answers to “zzzz”." (adding "with those filters." when facets are ticked) and a "Clear the search" button that empties the box and returns focus to it. Matches: "2 slabs answer to “bianco”." (or "1 slab answers to ...") with "Go to the list" and "Clear the search". "Go to the list" puts the phone search field away, scrolls to the list and focuses its heading. The note shows while focus is in the box or the note and is put away (the sr-only way) once focus leaves, so it never sits over the list the reader went to read. The list's own live line stops announcing while a search is typed, so the same sentence is not read twice. - Plurals. The list heading is now "1 slab" / "2 slabs" (
#n-unit); every count built in script (sheet summary, table summary, live line, the drop-a-filter button, the note) goes through one helper,slabsOf. The page's other counts were already singular-aware. - Severity 1 also fixed. The Sort select's value sat at the top of its 40 px box under base-select; it is centred (
align-items: centerinside theappearance: base-selectblock). - Not touched. The table caption "Slab inventory, Newark and Bayonne" showing on a phone.
- Invented, named. The words "Go to the list", "Clear the search" and "answer(s) to"; the rule that the note shows only while focus is in the search box or the note. No product fact was added.
Round seventeen (2026-11-10)
An independent reviewer reserved two things: after choosing a slab at 1440 the sticky "In the bay" bar covered the first line of the table caption "Slab inventory, Newark and Bayonne" mid-scroll (severity 1), and the hold dialog said "Nobody rings you" and "ring the Newark yard" in one table (severity 1).
- Reproduced at 1440 light, 1920 dark, 390 dark and 360 light: after tapping the second "Onice Cascata", scrolling put the caption under the stuck bar for up to 44 px of overlap (the first line of the caption) at three or four scroll positions at every width. The caption sat below the bar in the page, so a bar that stays with the list had to pass over it. A scroll margin only helps when the page scrolls to a target; it does nothing for plain scrolling, so it was not the fix.
- Fix. The list's title line and the note about the plates now come before the bar, so the bar has nothing above the rows to cover. The title is a visible
.tablecapline (aria-hidden); the table's own<caption>stays in the table with the same words for assistive technology and for the sideways-scroll label, set to.sr-only. A.peek + .tablewrapmargin gives the column header the space the caption's padding used to give. Measured after the edit at the four widths, across 22 scroll positions each including the position the selection lands on: caption rect and bar rect overlap 0 px everywhere. - Hold dialog. "Free. Nobody rings you; the confirmation is emailed." is now "Free. The confirmation is emailed; the yard does not ring you." The two release lines now say "or you ring the Newark yard" and "you ring the yard before then", so each line says who calls whom: the yard does not call you, you call the yard.
- Invented, named. Nothing new: the yard not ringing you and the reader ringing the yard to release or extend were both stated in earlier rounds (round ten). The "next in line" ticket's "We call you within the hour" is a different case (a released slab with a queue) and is untouched.
Round nineteen (2026-11-10)
Ship note from the independent score of round eighteen (36/40): at 1920 the hero's left edge (240) and the list column's (272) stepped by 32 px.
- Reproduced at 1920 dark by measurement:
h1left 240,.catalogueand.railleft 272. At 1440 and 1280 the two already agreed (both 32); at 1024 both 30.7; at 390 and 360 both 16. - Cause.
.mastheadcarried the gutter as its own padding outside the 1440 px column, while.shell__in,.pageand.bay__incarry it inside the column. Above 1440 px the hero's text started a gutter further out than every other block. - Fix. The gutter moved from
.mastheadto.masthead__in(padding-inlineinside the column; the outer padding is now vertical only). Measured after:h1left equals.catalogueleft at 1920 (272), 1440 (32), 1280 (32), 1024 (30.7), 390 and 360 (16); the ledger's right edge now meets the list column's right edge at 1920 (1648). Phone layout is unchanged (same values at 390 and 360, no overflow). - Not touched. The reviewer's sev 1 (b), the terms table above the Name field in the phone Hold dialog.
Inventions: none.