Distribution substation load operations. Thirty-two feeders on three transformer banks at Northgate 138/12.47 kV, live, with a replay of the 03:41 RTU outage.
Direction
- Archetype: Index. The table takes the page width with nothing drawn around it; a card would take width from the only thing that needs it.
- Stance: Table-first, with a station board above it. The board is the only chart that is not a sparkline: one 24-hour line for the station and one per bank, each labelled directly (the rating line and the end value are written on the chart, there is no legend). The tiles are also the table's bank filters, so the board and the table are one control.
- Density: Condensed. 32px rows and 32px controls, everything on Carbon's 8px mini-unit (
--cds-spacing-*). - The grid: Carbon's condensed 2x grid, drawn: 16 columns with a 1px gutter, switched on from the header (
aria-pressed). At 82rem and wider the table's columns are set as multiples of one sixteenth, so the overlay lines up with the column edges rather than with nothing. - The one bold thing: The numbers. IBM Plex Mono,
tabular-nums, right-aligned with fixed decimal places so the decimals align down the column, and three visually distinct marks for a measured zero (0.00, text-secondary), a value that does not apply (an em dash, text-helper) and no reading at all (a dashed rule, text-helper). Load is carried by bar length; colour only confirms, and the surplus at a wide viewport goes to that bar rather than being spread across the gaps between columns. The three marks are three sort outcomes as well as three glyphs: a measured zero sorts as zero, while no reading and does-not-apply hold the bottom whichever way a column is pointed.
What each surface answers
Every column that carries a comparable value sorts, in both feeds; the caret is drawn only for the column that is actually sorted, and a third click on one returns the table to feeder order. The Polled column is the last second each RTU answered on, which is why it is the only column with a figure in it during the outage: the eleven feeders on T3 read 03:41, and the twenty-one that answer read 03:51. The notification that names those eleven carries the control that shows them.
Interactions added in the uplift
- Row to panel. A feeder id opens a side panel (non-modal
show()at 66rem and wider, pushing the page at 82rem and wider; a modal bottom sheet below 42rem). It carries that feeder's 24-hour chart, the same slider the station has, and its readings. Escape and the close button return focus to the row. - Selection and batch actions. Row checkboxes, a header checkbox with an indeterminate state, and a bar that appears when anything is chosen: draft a switching order, mark for site visit, clear. Marks show as tags in the row.
- Search, bank filter, and a real empty state with its own drawing and a way back out.
- Over-rating list. Feeders above 100 % of rating are listed beside the board and open the panel. The count, the list, the alarm menu and the rows all derive from the same feed, so they cannot disagree.
- Charts are sliders.
role="slider"with arrows, Home, End and Page keys plus pointer scrubbing; the readout is anaria-livestatus. They are drawn as SVG at their displayed width and redrawn by aResizeObserver. - Motion. Tiles rise once on load; figures tick for 240ms when the feed changes. Every duration is a motion token, which the reduced-motion block collapses to 1ms, and the ticking is skipped outright.
Figures that must agree
Ratings are an assumption of the specimen: 50 MW per bank, 150 MW for the station. Live feed: station 123.4 MW; T1 43.5, T2 41.4, T3 38.5 (sum 123.42). Replay (RTU outage): 21 feeders answer, 46.9 MW, so the station chart there has no rating line and an automatic domain. Two feeders are over rating in the live feed (NG-1103 at 104.3 %, NG-1208 at 101.3 %), which is also the alarm count. The series colour on the charts is a purple token, chosen so that "over rating" red is never also the colour of an ordinary line.
Generator
The token file was produced by, verbatim:
carbon_tokens.py --all-themes
--all-themes emits colour only. The type sets, the 8px spacing scale and the motion tokens in the page's own <style> were taken verbatim from the same generator:
carbon_tokens.py --type
carbon_tokens.py --layout
carbon_tokens.py --motion
The theme audit quoted in the report is:
carbon_tokens.py --contrast --theme g100
Shipped theme: g100. The header control offers g10, and a reader whose system asks for light and who has not chosen gets g10; with no script at all the page renders the generated :root theme, which is Carbon white.
Measured contrast
Measured with contrast.py.
| Pair | Role | Ratio | Needs |
|---|---|---|---|
#f4f4f4 on #262626 | text-primary on layer-01, g100, every figure in the table | 13.76:1 | 4.5:1 |
#c6c6c6 on #262626 | text-secondary on layer-01, g100, the load bar | 8.86:1 | 3:1 (SC 1.4.11) |
#fa4d56 on #262626 | support-error on layer-01, a bar over 100 % of rating | 6.38:1 | 3:1 |
#161616 on #ffffff | text-primary on layer-01, g10 | 18.10:1 | 4.5:1 |
#ffffff on #0f62fe | text-on-color on button-primary, both themes | 5.00:1 | 4.5:1 |
APCA reads the first pair at Lc −97.6 and still asks for 16px at weight 400. Carbon sets dense UI text at 14px ($body-compact-01); the system wins, and the disagreement is recorded on the page.
Deviations from CONTRACT.md
- §1, tokens the generator did not emit.
--all-themescarries colour alone, so the spacing, size, motion and type-set values are held in the page's own:root. They are the verbatim output of--layout,--motionand--typeof the same generator, not hand-typed, andtokens.cssis untouched. - §5, the theme attribute. The generated file switches on
data-carbon-themeand its dark media query is guarded as:root:not([data-carbon-theme]), notdata-theme. The page uses Carbon's own attribute so the generated file stays authoritative; the in-page control wins in both directions and the OS preference is honoured on first paint by a blocking<head>script. - §4, no raw colour anywhere, including the favicon. The favicon is an inline
data:SVG drawn incurrentColorundercolor-scheme: light dark, so it carries no hex. Noapple-touch-iconand no manifest ship, because both need a second file and §2 asks for one.
Kept findings
slop_scanside-tab (warn): the 3px stripe on the inline notification is Carbon's own inline-notification anatomy; the icon and the first sentence carry the same fact, so colour never travels alone.craft_lintno-primary-action ×2 (warn): the two header icon buttons and the two pagination arrows are genuine peers, whichcraft.mdnames as the exception to the rule.copy_lintlong-paragraph ×21 (info): table rows read as prose blocks by the scanner.
Round two (2026-10-29)
A cold review scored the first screen 6/10: dense, correct, nothing set large. The board now sits on the 16-column grid with 1px gutters (station 1-10, banks 11-16), and its one expressive figure is the station load, 123.4 MW in Plex Sans Light sized in cqi from its tile, over the three banks drawn end to end against the 150 MW rating (43.5 + 41.4 + 38.5, 26.6 free). The 24-hour line is 3px with a dashed threshold band at the rating; the two feeders over rating are error-striped tiles. Table headers are heading-compact-01 at 14px.
The bank series are one hue, the purple support token, stepped toward --cds-text-primary (100%, 78%, 58%), not toward the layer: stepped toward the layer the lightest step measured 2.78:1 on g10 and 2.35:1 on g90, under SC 1.4.11's 3:1; stepped toward the text every step is 3.45:1 or more on every theme's layer. tokens.css carries no Carbon data-viz palette, so the steps are a deviation from Carbon's categorical palette, and every segment is directly labelled.
Round four (2026-10-29)
Three defects from a cold critic. The lone "Apply" button beside the Live and 03:41 replay switch is gone: the two radios already apply on change, so the button had nothing to submit. The station chart's y-axis now ticks up to 160, so the dashed 150 MW rating line and its band sit inside the drawn axis instead of above its last label (the 03:41 replay keeps its own range and draws no rating).
Table rows, their controls and any :target carry a scroll-margin-block-start of the shell header plus the table header plus a spacing step, so a row or anchor scrolled to lands clear of both fixed bars. It sits on those elements, not as scroll-padding on html, because root padding also moves every scrollIntoView and flipped the bank select's panel upward with room below (the picker probe caught it). A row passing under the sticky table header mid-scroll is what a sticky header does and is left as it is.
Round eight
- Design notes were flat white boxes (real). Cause:
.spec-gwas four equal padded sections of small type, with none of the board's figure scale or tile language. They are now laid out as the board is: sixteen columns, 1px gutters,layer-01tiles. The one bold decision is set in Plex Sans Light athead-06beside the direction lines; each measured ratio is a tile with its figure at the same size; colour and kept-from-the-system close the board. - T1, T2 and T3 differ only slightly in lightness in the dark theme (real). Cause: three tones of one hue cannot all sit 3:1 from one another on a dark layer (the third step would need a luminance above 1). Rather than move them apart, they gained a second channel: T1 solid, T2 hatched, T3 dotted, the pattern cut out of the fill in the layer colour (
--p2,--p3, tokens only). The stacked bar and each bank's own bar carry the same pattern, and the segments stay directly labelled. - 390: a long undifferentiated stack (real). Cause: the Bank column is hidden on a phone, so 32 rows ran together. In feeder order, the first visible row of each bank now names it ("Bank T2, 11 feeders") under a 2px rule (
paintGroups, run after sort and filter, and absent in any other order, where no seams exist). - Declined: no accordion over the table. The page is the table; folding it would hide the figure the phone user came for.
Round nine (2026-11-09)
Scored against the ten heuristics before and after: 31 to 34 of 40.
- The phone table needed a sideways scroll to show the load (real). Cause: the feeder column took 11rem and the numeric columns were sized by their content, so the percentage and the over-rating triangle sat off the right edge at 360 and 390. Now the feeder column takes the slack (
inline-size:99%in auto layout,min-inline-size:9rem), MW is 5rem and Load is 5.5rem. The table'sscrollWidthequals itsclientWidthat 360 (328px) and 390 (358px). The header says "Loading %" and keeps the full name for assistive tech through the.hd-xclip pattern. - Between 42rem and 66rem names wrapped to three lines (real). The same band rules now stack the name over the tag and set the feeder column at 36%.
- Touch size (real). Under
(pointer:coarse)the buttons, selects, inputs and checkboxes step from 32px to Carbon's 48px--cds-size-large. Controls inside a 32px row keep the row height and gain a 48px::afterhit area. - The bar and its rating (real). Table bars run 0 to 120 % of rating, so 100 % sits at 83.333 % of the track. A 1px tick marks it in each bar and the key names it ("100 % of rating").
- Fields on a layer (real). Search and bank select sit on layer-01, so they use
field-02(field-01 equals layer-01 in every theme and the box vanished). Radius is 0 as on every Carbon control. - Placeholder contrast (real).
--cds-text-placeholdermeasures 2.4:1 on field-02, so the placeholder is set at text-secondary instead. - A destructive colour on a routine action (real). "Mark for site visit" in the row menu was red. It is a plan, not a hazard; the danger style is gone.
- The peak label met the rating line (real). On a feeder that peaks just under 100 %, the label sat on the dashed line. It now carries a halo in the layer colour (
paint-order:stroke, tokens only). - Header figures in mono caps looked uneven (taste). Numeric column headings use Plex Sans; the figures below stay Plex Mono.
Deviations, in addition to those above:
- field-02 on layer-01, and the placeholder at text-secondary, as measured above.
- The 0 to 120 % table bar scale with a 100 % tick.
- The row's overflow menu is hidden under 42rem; the side sheet carries the same actions, so a phone has one way to each, not two.
- The 48px step under a coarse pointer rather than at a width, because a laptop with a touch screen and a phone both need it and a narrow desktop window does not.
Checked: station, bank and alert counts, the sums and the over-rating list all derive from one feed and agree live and held (03:41). Tested widths: 320, 360, 390, 720, 1440, 1920.
Round ten (2026-11-09)
A second, harsher read. Scores before and after: 31 to 35 of 40 (visibility 2 to 3, user control 2 to 4, error prevention 2 to 3, help 2 to 3). Still below 3: none.
- The over-rating list read as an all-clear while Bank T3 was dark (real, severity 3). In the 03:41 view it said "No feeder is over its rating" and the bell said "none unacknowledged", with eleven feeders unreadable. It now says "None of the 21 answering feeders is over its rating", adds a row for the dark bank ("Bank T3 stopped answering at 03:41. Loading on its 11 feeders is not known.", which filters the table to it), and the bell counts it. The bell and its menu no longer say "unacknowledged": there is no acknowledge action, so the word promised one.
- A feeder with no reading went into a switching order without a word (real, severity 3). The draft is now a list with one line per feeder, the reading beside it (no reading, over rating, or the percentage), the blind ones first, a warning above the list when any have no reading, and a remove button on every line. The message after an add repeats it.
- No way back (real, severity 3). "Clear the draft", a batch add, a row-menu add or flag, and a per-line remove now answer with a toast: what changed, and Undo (restores the marks, the draft and, for a batch, the selection). Held ten seconds, held for good while the pointer or focus is on it, and Escape dismisses it. After Clear or a batch act, focus moves to Undo because the pressed control is gone. From the feeder panel the message is written in the panel above its buttons, where the button already flips back.
- The result was out of sight (real, severity 2). On a phone the draft sits 700px above the fixed batch bar. The toast carries "Show the draft" when the draft is off screen; it scrolls there and focuses it. A batch act now consumes the selection, so the bar goes and the rows carry their tag.
- A feeder hidden by a filter stayed selected (real, severity 2). A switching order could take rows the operator could not see. Filtering now drops hidden rows from the selection and the bar's count follows.
- The panel said the last answer was in the table (real, severity 2). The table leaves those figures blank. It now says no figure is shown.
- The row menu always read "Add" (real, severity 1). It reads Remove when the feeder is already in the draft or flagged.
- "BankLoad" at 82 to 100rem with the panel open (real, severity 1). Three columns are hidden then, and the rest are re-cut in sixteenths (feeder 5, bank 1, load MW 2, loading 4, trace 2, polled 2).
Inventions, since the brief gives no rule: a bank that stops answering is an alarm of its own (so the bell reads 1 in the 03:41 view and 2 live); the toast timing; selection pruned by the filter; one draft number, SO-2041, that is never issued (the page still only drafts).
Left: the 03:41 view is a replay and still lets the operator draft an order from it (severity 1: the brief does not say whether a replay is read-only). No acknowledge action on alarms (severity 1: needs a rule the brief does not give). No explanation of the 0 to 120 % bar scale beside the table header; the key at the foot holds it (severity 1).
Driven after the last edit at 1440, 1920, 390 and 360 (touch, 3x), both schemes: batch add, Show the draft, Clear and Undo, per-line remove and Undo, panel add and remove, the alarm menu with a dark bank, live to replay with a draft held. No horizontal overflow, no console error.
Round eleven (2026-11-10)
Two severity 2 defects from the independent score, both reproduced before editing.
- The batch bar scrolled out of sight on desktop (real, severity 2). At 1440, with NG-1208 ticked at scrollY 745,
.batchwasposition: staticat top -157px, so the reader had chosen a row and saw no action. From 42rem up it is nowposition: stickyunder the shell header (top 48px, bar 40px), and the column heads sit below it (top 88px) instead of behind it. Carbon replaces the toolbar with the batch bar in the same place; this keeps that and pins it. Below 42rem it is still fixed to the foot. Checked at 676, 1440 and 1920: the bar does not wrap, and Add to switching order works from the pinned bar. - The replay and the live feed contradicted each other (real, severity 2). Both sets of facts in this brief are right on their own (live: station 123.4 MW, NG-1103 at 104.3 %, over rating since 01:24; replay: 21 feeders, 46.9 MW, NG-1103 at 54.8 %). They cannot be the same morning: the live trace passes 03:51 at about 120 MW, the replay reads 46.9 MW at 03:51. The replay is now a recording of an earlier day, so nothing is compared across them. The clock reads Mon/Tue in the replay (the live feed keeps Wed/Thu), the toggle reads "Tue 03:41 replay", and the notification says "stopped answering at 03:41 on Tuesday. This is the recording of that outage, not today's feed." NG-1103 is now one story per day: Thursday, over its rating since 01:24 for 2 h 48 min; Tuesday, 54.8 % at 03:51, peak 58.3 % at 18:51. Every figure and every state is unchanged; only the day names and those two strings moved.
Invention, since the brief gives no rule: the outage was on a Tuesday. It is the minimum that separates the recording from the live Thursday; no date is given, and none is shown.
Left: the same T3 sentence appears in the notification and in the over-rating box in the replay (severity 1); the row menu ignores Arrow keys (severity 1); the Polled column is dropped on phones, so the eleven T3 rows show dashes with no poll time there (severity 1).
Driven after the last edit at 1440 light, 1920 dark, 676 light, 390 dark and 360 light (touch, 3x): replay on, its toggle, note and readout; live; a lower row ticked and the bar position; Add to switching order from the bar. No horizontal overflow, no page error.
Round thirteen (2026-11-10)
- A row-level order landed its draft out of sight (real, severity 2, reproduced). At 1440 and 1920, NG-1209 "Add to switching order" from its row menu put the draft 127 to 227px above the viewport with focus left on the menu button, by mouse and by keyboard. On a phone the row menu is not drawn (the feeder panel carries its actions), so the same order from the panel left the draft above the sheet.
- Fix. A row-menu add now brings the draft into view (only if it is not already) and moves focus to that feeder's line in it, which is marked with the selected-layer token so a pointer user can find it too. The draft header has a Close button, and Escape inside the draft does the same: it closes the draft (the marks stay, the New switching order button reopens it) and focus returns to the row's overflow button, or on a phone to the feeder id that opened the panel. The panel shows "Show the draft" under its message after an add; it closes the sheet, reveals the draft the same way and Close returns to the feeder id. The row keeps its Order tag. On a phone the draft scrolls to just under the shell header, not under the desktop toolbar offset.
- Invention. Close on the draft header, Escape to close it, and the panel's "Show the draft" button. No product facts added. The toast no longer offers "Show the draft" after a row-menu add, because the draft is already shown.
Round fifteen (2026-11-10)
- A recorded outage could be turned into a switching order (real, severity 2, reproduced). In the Tue 03:41 replay, the batch bar's "Add to switching order", the row menu's, the feeder panel's and "New switching order" all worked, and the draft then held a feeder at its recorded value.
- Fix. The replay is read-only for adding to a draft. Each of the four controls stays where it is, drawn disabled with
aria-disabled(so it stays focusable and is not a dead control) and says why when pressed: "Ordering is off while the recording plays." with a Return to Live button, in the toast for the batch bar, the row menu and New switching order (on a phone the toast lifts above the batch bar, which can be two rows tall), and as a notice in the feeder panel above its buttons, where focus goes to its Return to Live. Return to Live switches the feed and puts focus on the control that was refused (the batch bar's order button when feeders are ticked, the panel's order button, otherwise the Live switch). The replay notification also says ordering is off and carries Return to Live. A refused batch keeps its ticks. Taking a feeder out of a draft stays live (it is not ordering): the draft's own remove buttons, Clear the draft, and a menu or panel "Remove from switching order" on a feeder already in it. - Replay panel poll time had no weekday (severity 1). "Last polled" now reads "Tue 03:51:50" in the replay; the table's Polled column is unchanged.
- Draft chip clipping "104.3 % of rating, over" at 1440 (severity 1): not reproduced. NG-1103 and NG-1208 ordered from the batch bar, at 1280, 1366, 1440 and 1920 with the feeder panel open and shut: the status ends 8px or more before the remove button and nothing overflows, so no edit.
- First row of a bank group on a phone (severity 1). The cells of the first row are pushed down by the bank label's height, so Load and the % sit where they do in any other row. At 360 and 320 the label ("Bank T2, 11 feeders") wrapped to two lines and put the figures a line off; it is now one line (
white-space:nowrap). Measured at 390, 360 and 320: the figures sit 29.5 and 35.5px below the feeder id's top in a first row, the same as a one-line row. - Invention. The sentence "Ordering is off while the recording plays." and the Return to Live buttons; the rule that removing from a draft stays allowed. No product facts added.
- The table did not reflow at 320 (WCAG 1.4.10, pre-existing, severity 1). The frozen feeder column, Load and Loading came to 309px in the 288px the page leaves, so the % column was clipped and the table scrolled sideways. Below 400px the two figure columns, their header padding and the feeder cell's padding give a little, and the Load and Loading headers drop to the 12px label size (the words and units stay). Measured on every table cell's text and bar, live and replay: the table is 288 of 288px at 320, 328 of 328 at 360, 358 of 358 at 390, the rightmost text ends inside the table, and the page does not scroll sideways. The replay notification's Return to Live pushed the page 17px wide at 320, so its buttons now wrap. The shell header's "Grid North" touched the grid-overlay icon at 320 (the name box was 56px for a 70px name): below 360px the way-back link and the name take less padding (8px and 4px either side of "Portfolio", 8px before the name), so the name ends 14.9px clear of the icon at 320. Measured name-to-icon gap: 14.9px at 320, 22.9px at 360, 52.9px at 390; 1440 is unchanged.
Round twenty-three (2026-11-10)
- A sorted numeric header drew its caret over the label (severity 2, reproduced). At 1440 light, sorting Load, Current or Bus put the 16px caret on the label (13.7, 16 and 16px of overlap), and at 1920 dark 4 and 8.3px; at 390 and 320 Load and Loading overlapped by the full 16px. The caret was absolutely placed at the start of the button, over padding the column was too narrow to keep.
- Fix. The sortable header is the Carbon row: label, unit, then caret, in one flex row, with the numeric headers right-aligned, so the caret takes its own room and can never sit on the words. The absolute placement is gone. Load, Current and Bus at 82rem and up are 6.5, 6.75 and 5.5rem (the row's measured width plus 8px padding), back to a sixteenth of the table from 112rem when the feeder panel is shut; the rebalance below sets the final widths. At 320 (below 22.5rem) the unit of Load and Loading drops under the label so the table still fits 288px. Measured label rect against caret rect, every sortable header, ascending and descending, at 1440, 1920, 1312, 1024, 768, 390, 360 and 320, and with the panel open at 1440, 1312, 1700 and 1920: a 4px gap everywhere, no clipped text, rows 33px at desktop as before, no sideways scroll.
- The feeder chart scrubbed with no visible value (severity 2, reproduced). Opening NG-1103 and pressing ArrowLeft three times moved the crosshair to Thursday 01:12 while the headline stayed 104.3 and only the hidden status read "99.9 %".
- Fix. While the panel chart is scrubbed (keyboard, pointer or touch) the headline figure reads the scrubbed value and its line reads "% at Thu 01:12", as the station chart's readout does; when the scrub ends (blur, pointer leaving, Escape, which closes the panel) it returns to "104.3" and "% of the 400 A rating". The headline keeps its size, so nothing below it moves while scrubbing.
- Invention. The wording "% at Thu 01:12" for the scrubbed line. Two small rules in the shared chart: a press on the chart places the crosshair where it came down (a tap has no pointer move), and a press anywhere else ends a scrub that touch left standing (the chart keeps focus after the finger lifts). Both apply to the station chart too. No product facts added.
- The 24 h and Polled columns were never shown (pre-existing, found and fixed). A stray
}after the 25rem media block made the rule that shows them from 42rem up invalid, so both columns were display:none at every width although every note above describes them. The brace is gone and they show from 42rem, which changed every table width, so the table was rebalanced. - Rebalance, 82rem and up. A calc() on a table column is read as auto, so the old
12.5% - 48pxPolled width let the column be squeezed to 50px at 1312 and clip its time. Now: Load, Current and Bus are 6.5, 6.75 and 5.5rem, Polled 6.5rem, Loading 17%, 24 h 9%, Bank and PF 6.25%, and the feeder column takes what is left. From 112rem with the panel shut every column is a whole sixteenth again (feeder 4, Loading 3, 24 h 2, the sort columns one each, Polled the rest of its two). With the panel open between 82 and 100rem the 24 h column goes too, with Current, PF and Bus (the panel carries that trend), and the feeder column takes the rest. - Measured after the brace fix. Caret-to-label gap on every sortable header, both directions, panel shut and open, at 1312, 1440, 1700 and 1920, live and replay: 4px or more, no clipped header or cell, no sideways scroll, rows one line (NG-1204's long name wraps at 1312 as it did before), 24 h sparklines 86 to 208px wide with 25 points, Polled times whole. At 1024 and 768 with the panel shut, the same. Phones at 390, 360 and 320 are unchanged (the table is 358, 328 and 288 of the same). At 1024 an open panel docks over Loading and Polled, as it did before; those headers cannot be pressed until it closes.
Round twenty-five (2026-11-10)
- The bank tile unit ran into its caption with the feeder panel open (severity 2, reproduced). At 1312x900 light with NG-1103 open, the three banks column is 311px, so the tile's 3fr value column (67px) was narrower than "43.5 MW" (91px) and the unit drew 7.3px into the "86.9 %" and "of rating, 10 of 10 answering" caption; 1.8px at 1366; clear at 1440 only because the column was 81px against a 91px value and the 16px gap swallowed it. Measured by Range rect, the same overlap showed in all three tiles.
- Fix. A container query on the banks column (not the page): below 28rem, at 42rem and up, the tile is two columns. Name and feeder count on the left, value and unit on the right of the same row, the bar and its caption across the full row beneath. The three-column tile is unchanged above 28rem, so panel shut at 1280 and up, and panel open from 1700, look as before. Value, unit and caption rects measured against each other (name, count, value, unit, percentage, caption, bar, and the dash of a bank that is not reporting) at 700, 768, 900, 1024, 1056, 1100, 1150, 1200, 1240, 1280, 1312, 1330, 1366, 1400, 1440, 1500, 1600, 1700, 1800 and 1920, panel open and shut, live and replay: no overlap, no text outside its tile, no sideways scroll. Open-panel tiles at 1312 to 1440 are about 63px taller (169px against 108px).
- The station chart peak label covered the 140 tick at 320 (severity 2, reproduced). At 320 the label "Peak 132.2 MW at 18:12" anchored to the left of its point and began 7px left of the plot, over "140" (7.4px overlap).
- Fix. After the chart is drawn, the label is measured and held inside the plot, 4px past the axis on the left and no further right than the plot edge. Applies to the feeder panel chart too. Label rect against every other text in the chart (value ticks, time ticks, rating label, end value) at 320, 340, 360, 390 and 1440, resting and scrubbed, live and replay: no overlap, label inside the chart.
- Invention. The 28rem threshold and the two-row tile. No product facts added.