The launch manifest for a spaceport, on the week two vehicles hold overlapping windows on the same downrange corridor and the range can only clear one.
The fifth and last brand specimen, and the one that forced the generator's final mechanism.
Where the style comes from
Transcribed from awesome-design-md/design-md/spacex/DESIGN.md, read 2026-09-16, and emitted by:
python3 tools/brand_tokens.py spacex-brand > systems/spacex-brand/tokens.css
python3 tools/brand_tokens.py spacex-brand --check
python3 tools/brand_tokens.py --self-test
| Decision | Transcribed |
|---|---|
| Ramp | #ffffff #f0f0fa #e0e0e8 #5a5a5f #3a3a3f #0a0a0a #000000 |
| Chromatic | none |
| Type | 80 · 60 · 48 · 16 · 13 · 12 px |
| Tracking | positive at every size: 1.6px at 80, 1.17px at 13, 0.32px at 16 |
| Weights | 400 and 700 |
| Radii | none. Nothing that is not a circle carries one |
| Depth | none. A near-black panel on black is the only lift |
| Modes | night first, with a full daylight palette beside it |
| Motion | not published |
It is the only monochrome brand in this family, and that is not a simplification of the reference: it is the reference. Its primary is #000000, its link on dark is #ffffff, and the single chromatic value anywhere in the file is the browser's own default link blue, listed as a fallback. What marks a primary action here is an INVERSION rather than a hue.
That broke an assumption in the generator, which had treated the accent as one value shared by every theme. accent, on-accent and accent-text are now pinned per theme, and (the part that actually failed) the accent tint is solved from the pinned value rather than the declared one, because mixing black into a black canvas produces no visible wash and the solver correctly raised rather than returning something invisible.
The finding worth carrying. This palette has no legible mid-grey on dark. Its published ink-mute #5a5a5f reads 2.91:1 against the near-black panel, under what even a control outline owes, let alone body text, so the tertiary text tier on the night setting is the pale hairline colour instead. The dark page therefore has two text weights where the light one has three, and the design-notes panel says so.
The direction
- Archetype. Index.
- Data stance. Tile-first.
- Density. Wide, display-led.
- The one bold decision. The conflict is marked by inverting the tile, not by colouring it. With no red, no amber and nothing else available, the loudest move this language has is to swap the ink and the ground, so the contested missions and the band explaining them are all inverted, and the word "Contested" is in each tile as well, so the inversion is a second signal rather than the only one.
The second state
The corridor conflict. An index's promise is that its entries are independent; this is the case where two of them are not, and neither is wrong. Both windows are valid, both vehicles are ready, and the range clears one vehicle at a time on that heading, so the page states the overlap, the next clear slot and the hour by which somebody owes a decision, rather than showing two rows that each look fine.
The pair it makes with ferrari-brand
Both references set an 80px display step. Ferrari closes the tracking to −0.028em; this one opens it to +0.02em and sets the whole thing in capitals. Same size, opposite decision, and the two pages do not look remotely alike because of it. It is the clearest thing this family says about what a type decision is actually worth.
Deviations from CONTRACT.md
- §2, the typefaces. D-DIN is proprietary. Saira Condensed and Saira are substituted: a squarish industrial grotesque that holds at 700 in capitals with the tracking opened.
- §5, motion. The reference publishes no duration or easing. The two rungs in
tokens.cssare the portfolio's, and the emitted file says so.
Findings kept, with reasons
craft_lint·disabled-unstyled: styled by.btn[disabled], witharia-describedbynaming what would make it live; the rule looks for utility-class variants.craft_lint·hover-no-transitionon.btn-ghost: it always rides with.btn, which declares the transition, and the rule reads one class at a time.audit_page·apple-touch-icon-missing,manifest-missing: portfolio-wide.copy_lint·name-drift: seecopy-lint.ignore.
What the gates changed
craft_lint's no-press-state caught a mission tile that lit up on hover and did nothing when clicked, which is a card telling the reader a lie. The mission name is now a real link with its hit area stretched over the tile, so there is one link per mission rather than six, the hover belongs to something, and the focus ring is drawn on the tile rather than on four words inside it.
The uplift round (2026-10-29)
The page stopped being a list of tiles with a banner above it and became a range-control console. The archetype stays an index; the conflict is now an instrument the reader operates rather than a paragraph they read.
- The head sets "Six launches, five slots" at the 80px step beside a plan view of the 048 degree corridor: two pads, one shared line, 180 to 1,400 km. It is the only group that moves on load (the corridor line draws, the title rises), under 640ms, and reduced motion shortens it to 1ms rather than deleting it.
- The console is a timeline of 24 and 25 September with one lane per vehicle. The range needs 21 hours to reset after a window closes, so the first vehicle's window and its reset are one hatched block across both lanes, and the second vehicle has to open after it. The conflict is a shape: a solid column joins the two bars where the windows overlap, and a dimension line counts the time left. The reader moves a window by dragging the bar (wide layouts), with the slider, or with the keyboard, and chooses which vehicle moves. When the hold ends the console, the contested tiles and the headline all invert back to the page ground, "five slots" becomes "six", and Publish the week turns on.
- Corrected figures. The old copy gave the overlap as 02:40 to 04:15; the two windows actually overlap from 03:10 to 04:15 (65 minutes). The next clear slot is derived from the rule (25 Sep 01:15 Z if Corvid 9 keeps its window) and no longer a typed constant. T-minus times on every tile are measured from range time 04:12 Z, and 31:08 to the first window checks against it.
- Phone. The timeline turns vertical, the controls become a dock pinned to the bottom edge while the timeline is on screen, and the bar cannot be dragged (a vertical drag is a scroll); the slider is the control there.
- Tabular figures. Saira Condensed has no tabular figures (1111 and 0000 differ in width), so each digit of a live figure sits in a fixed cell.
- Notes. The measured pairs are swatches whose ratios are computed in the page from the colours it is actually painting, so they follow the theme. A theme toggle was added to the nav.
- Nav now points at three real sections.
copy-lint.ignorestill names name-drift and is harmless.
Boson findings read and left: target-size warnings on nav links, theme button and tile names (real boxes measure 46 to 51px tall; the probe measures a text run), the rotated "Shared corridor" label reported as overflowing (it is rotated 42 degrees), and keyboard_probe's focus-not-visible on the tile names and footer link, which paint correct focus rings when tabbed to (checked by screenshot and computed style). Launch-policy findings are site-level.
Round four
- The timeline's two windows are now two thin bars, one per lane, 14px deep. The overlap is an outlined column across both lanes rather than a filled block, so it reads as a region and not as a third shape. The phone version draws the same thing on the vertical axis.
- The headline figure is labelled with whose wait it is ("Hanwell II waits", 22 H 05 MIN) and carries a line tying it to the telemetry card ("Set off by 1 h 05 min of overlap"). Both figures stay: one is the conflict, the other what the conflict costs, and the label says which is which. The line follows the slider (5 min of overlap leaves 21 H 05 MIN).
- At 390 the nav is a grid: wordmark and the two controls on the first row, the three section links under it. Nothing wraps unpredictably or clips.
Round eight (2026-10-29)
- "PORTFOLIO" cut off at the right edge (390). Not a clip. The word is whole (its right edge is at 365 against a container edge at 366, measured and checked on a zoomed crop of the critic's own screenshot): the row is exactly full, with no slack, so it reads as cut. Left as drawn, since the nav's two controls and the wordmark share one row by design and the section links sit under it.
- Footer rule wider than the column (1440). Real. The rule was the footer box's top border, and the box carries the padding that insets its text. The rule is now the top border of the footer's paragraph, so it runs 128 to 1312 like the rules above it.
- Overlap bars tiny beside the "Range reset, 21 H" block (1440). Real, and the one move. The axis ran evenly over 29 hours, so a 1 h 05 min overlap was 37px. It is now enlarged where the argument is: 02:00 to 08:00 takes 54% of the length (the overlap is 98px), and the 21-hour reset is squeezed either side, with a double slash on the grid at each join and "enlarged from 02:00 to 08:00 Z" in the chart's accessible name. Dragging maps the pointer through the same axis (
axis().inv), so the bar still follows the hand: 81px moved 60 minutes in the enlarged stretch. The phone's vertical timeline uses the same axis. The 03:00 tick after midnight was dropped, since it sat on "25 SEP".
Round nine (2026-11-09)
Judged as the range controller: move a window, read the manifest, resolve the conflict, by pointer and by keyboard; phone emulation at 360 and 390 (coarse pointer, 3x), 200% zoom, both display settings, 1920. No horizontal overflow at any width from 320 to 1920 in 10px steps. Heuristics before to after: 27 to 29 of 40.
Fixed, most serious first:
- Two tiles drew the wrong day (severity 2). Orrey 5 (26 Sep 09:45) and Brack (27 Sep 22:15) carried stored offsets that put their week-rail marks a day off, against their own printed windows and T-minus figures. The offsets now agree with the printed times; every T-minus on the page checks against range time 04:12 Z on Monday 21 September (31:08, 49:48, 70:28, 70:58, 125:33, 162:03).
- The 21 hour rule contradicted the manifest (2). Saltmarsh (LC-1, 23 Sep 07:30) closes 19 h 10 min before Corvid 9 opens, which "after any window closes" would have made a second conflict. The rule now reads "after a window closes on this corridor", which is what the plan view shows (LC-2 and LC-4 only).
- Phone timeline labels collided with marks (2). The axis-break slashes ran into "02:00" and "08:00"; the overlap label sat on the hatch edge; the dimension line had no label at all. Labels are moved clear of the breaks, the overlap label is two lines inside the block, and the dimension line is labelled along its length.
- A contested tile now opens the console (1). Its name linked to its own section. Corvid 9 and Hanwell II now go to the console with that vehicle chosen to move.
- The focus ring on the slider cut the label above it (1). The label has room now.
- Plan view labels were crossed by the corridor boundary (1). Tick labels carry a ground halo.
- Widows and orphans (1). "II" and "Z" no longer wrap alone, the kicker breaks at its separators, the verdict no longer repeats the line under the figure, headings balance.
- Uneven letter gaps (1). Opened tracking on the hinted face landed on whole pixels ("SHA PE"); geometric precision gives even spacing.
- Phone fact table stacks label over value; Reset is 44px wide; the weekday is on the week and on the decision time (21 to 27 September is Monday to Sunday).
Left: the moved bar shrinks to 26px once it lands in the squeezed 21 hour stretch (the squeeze is the point of the axis break; severity 1), the phone manifest indents plain tiles by their padding (1), and the display button names the current setting rather than the action (1; its accessible name says both).
Round ten (2026-11-09)
Scored fresh on the ten heuristics, with every state driven (move, hold, clear, publish, reopen) by pointer, keyboard and 360 and 390 phone emulation (coarse pointer, 3x), in both settings, at 320 to 1920. Before: 3, 3, 2, 2, 2, 3, 3, 3, 3, 2, which is 26 of 40 on a stricter read than round nine's 29. After: 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, which is 30.
Fixed, most serious first:
- Publishing could be undone by touching the slider (severity 3). "Publish the week" sent six customers a manifest, then moving the slider or the vehicle chooser quietly set it back to unsent. Publishing now has three real steps: ask (the button becomes "Send to six customers" and "Cancel", Escape cancels, the copy says it cannot be recalled), sent (the console locks, its badge reads Published, the week is on screen as sent), and an explicit "Reopen the week" that says customers keep what was sent. Moving anything while the question is open withdraws it.
- Four tiles were links that went nowhere (2). Cleared and weather-hold tiles lit up on hover and, clicked, scrolled to the top of their own section. Only the two contested tiles open something, so only they are links, and each now says what it does ("Move it in the console"). Names are headings.
- The way to the next step was 700 px away (2). Once the corridor cleared, Publish sat below the manifest with nothing pointing at it. The console's primary action becomes "Review and publish" when the corridor is clear (it takes the reader to the publish button and focuses it), and the publish button can be reached by keyboard while it is off (aria-disabled, with its reason read out), where a disabled button was skipped.
- The moved window vanished into a sliver (2). Once Hanwell II lands in the squeezed 21 hour stretch it is 26 px, and on the phone nothing said what or when it was. Its times are now written beside it, on both layouts. Round nine left this as severity 1; the phone made it 2.
- The phone's dock did not say the result (2). The clock is a screen above the slider. The dock now carries the answer ("22 h 05 min to clear", "Corridor clear", "Week published") under the slider, and the chart no longer scales down to 88% at 360 (it was drawn 300 wide in a 264 box, so labels were 10.6 px).
- Labels met marks (2 on tablet, 1 on phone). "24 Sep" ran into "02:00" from 640 to 900 px wide (it moves into the gutter there); the long "clear of the hold" label ran past the plot edge; on the phone "Clear from" sat on the 25 Sep line and the duration label on the hatch edge, and at 320 the two lane time labels met.
- Nobody said the bar could be dragged (1). The left gutter now says "Drag the bar or use the slider" where dragging exists, and not where it does not.
- The deadline had no clock (1). "Decision owed by Tue 22 Sep, 18:00 Z" now reads ", in 37 h 48 min" from range time, which checks against T-minus 31:08 to the first window.
- Phone manifest indent (1, left from round nine). Plain tiles sat 24 px in from the page column; they now sit on it, and the inverted tiles run to the screen edge with their text on the same column.
Left: the theme button names the current display, not the action (1; its accessible name says both); the "Hold" verdict is explained by one line under the clock, not defined on its own (1); the sliver's label sits on the hatch behind a halo, busy but legible (1).
Round eleven (2026-11-10)
Defects, from an independent score of 34 of 40:
- The phone dock answered 700 px from the thing it moved (severity 2). At 390 and 360 the chart is 660 px tall and the "Hanwell II waits" figure sits above it, so with the slider in view only the bottom 276 px of the chart showed. Dragging the slider or pressing "Take the next clear slot" changed nothing the reader could see but one line of small print. Reproduced at 390 dark and 360 light (isMobile, 3x). Fix: the dock now leads with a readout strip (the vehicle, how long it waits, and the state tag) above the window chooser. It is the same element that carried "x to clear" under the slider, moved to the top, set in the button size and bold, with the verdict tag beside it. It is in the dock, so it travels with the controls whether the dock is pinned over the chart or in flow, and it never covers a control (the dock grew by one 20 px line plus its gap). Hidden from 60rem up, where the clock is on screen beside the chart, and hidden from assistive technology because the live region and the slider's value text already say it. States: "Hanwell II waits 21 h 35 min" with Conflict or Hold, "Corridor clear, no wait" with Clear, "Sent to six customers" with Published.
- The intro paragraph did not follow the state (severity 1). It said "Two vehicles hold overlapping windows" after the fix. It now follows the state as the hero headline does: overlap keeps the original sentence; Hold (windows apart, reset running) says the two no longer overlap but the second still waits out the range reset; Clear says they now fly a range reset apart.
Invention: the strap wording for Hold and Clear, and the strip text for the clear and published states, all restating facts already on the page ("Corridor clear, no wait", "Week sent to six customers"). No new numbers, rules or products.
Left: the conflicted console is ink on a light page and ground on a dark one by design (the severity 1 note); the bold dock strip does not change that.
Round thirteen (2026-11-10)
Defects, from an independent score of 35 of 40 (ship verdict yes, with one note):
- The count tiles disagreed with the manifest badges (severity 1, recognition). The tiles read "5 clear slots / 2 in conflict" while the manifest badges read 3 Cleared, 2 Contested and 1 Weather hold; after the corridor cleared the tiles read 6 / 0 against 5 Cleared and 1 Weather hold. Reproduced at 360 light (initial, after +30 min, after "Take the next clear slot"). Fix: the tiles are now "Cleared" and "Contested", the badge words, and a
tally()reads them from the badges in the manifest after every update (a drag, the slider, the chooser, Reset, Take the next clear slot), so there is one source. The Cleared tile carries "of 6, 1 on weather hold", counted from the same badges. The headline word ("five" or "six" slots) is derived from the same count (launches minus all but one contested vehicle), which gives the same five and six as before. No fact in the brief changed: "Six launches, five slots" and the live-region sentence are as they were. - The bars left the screen while dragging on a phone (severity 1). At 360 light, slider at +30 min, the vehicle bars were 323 px above the top edge and only the hatch end showed above the dock. On a phone, a move (slider change, keyboard step, chooser, Reset, Take the next clear slot, or an input while the dock is pinned) now scrolls the bars into the band between the top edge and the dock, only when they are outside it. When the moved bar is a day from the fixed one and both cannot fit, the moved bar wins. It never scrolls while the dock is in flow, so the slider does not leave the reader thumb, and it scrolls instantly under reduced motion.
Invention: the wording "of 6, 1 on weather hold" (restates the manifest) and the 64 px and 12 px margins of the scroll band. No new numbers, rules or products.
Left: the fixed bar scrolls off the top when the moved one is dragged to the far end of the day (severity 1, by the rule above).
Round twenty-eight (2026-11-10)
Defects, from an independent score of 36 of 40 (ship yes, with one note):
- The moved bar's label overprinted "Range reset, 21 h" on a phone (severity 2). At 390 the slider at 645 drew "13:55 to 15:45" over the reset caption and erased "Range reset". Reproduced as reported, and wider: a sweep of every slider value (0 to 1440 in 5 minute steps, both movers) at 390, 360, 1440 and 1920 measured each label's ink box against every other and found the reset caption hit at about 47 values per mover on the phone widths, "Clear from" hit by the moved bar's times at 23 more, and at 360 and 320 the left lane's "+1" times met the right lane's. None at 1440 or 1920. Fix: after the chart is drawn, the reset caption and (on the phone chart) "Clear from" step along the axis, in 6 px steps nearest first, to the first spot where no other label touches them and, first choice, no bar either; the reset caption stays inside the hatch. The left lane's times switch to the en dash form when they would meet the right column, and at 320 lose "+1" to the line under them ("LC-2 +1"). After the fix the same sweep reports zero overlaps at 390, 360, 1440, 1920, and also at 320, 768 and 1024.
- Pressing the held Publish gave no response (severity 1). It now writes why, in ink instead of grey, in the caption beside it and in the live region: "Not published. Hanwell II still waits 11 h 20 min for the corridor. Take the next clear slot, then publish". The wait is the same figure the dock and clock show. Any change to the week restores the caption.
- The nav kept "Notes" after an instant jump to the top (severity 1). The observer band misses everything above the console, so a scroll listener sets "Corridor" when scrollY is 0 and nothing is current.
- Bar past the lane box after Snap (severity 1): not changed. Measured: the hatch box ends at the hold end (x 1192) and the Hanwell bar starts there and runs to 1222, inside the chart (right edge 1263). That is the meaning of Snap: the window opens the minute the hold ends, so it sits outside the hatch by design. Reproduces as a measurement, not as a defect.
Inventions: the wording of the refusal sentence (restates the wait and the Snap action already on the page); the "+1" moving under the pad name at 320; the 6 px step. No new numbers, rules or products.
Round thirty (2026-11-10)
Defects, from an independent score of 35 of 40 (ship yes, with one note):
- Reopen left focus on a control a screen above (severity 1). At 390, 360 and 320, after Send then Reopen,
document.activeElementwas the slider with its top 2,400 to 2,800 px above the viewport and the page did not follow (at 1440 and 1920 it was 900 px up). Reproduced on the previous file at all five widths. Reopen now focuses "Publish the week", the button that comes back in the row where the Reopen link was, so focus stays in the viewport with no scroll. Driven at 390 dark, 360 light, 320 light, 1440 light and 1920 dark through Snap, Reset, Snap again, Review and publish, Publish, Send and Reopen: every focused control is inside the viewport and is the element at its own centre (none under the sticky panel); no page errors, no horizontal overflow. - The phone dock took two thirds of a short screen (severity 1). At 320x568 the dock was 377 px (66%) and covered the bars when the chart arrived (52 and 88 px of the two bars at chart top). A new rule for viewports under 700 px tall (phone width) tightens the dock: 12 px padding, 8 px gaps, the label and slider closer, the Reset link beside the snap button. Dock height now 236 to 264 px: 46% at 320x568 (was 66%), 37% at 360x640 (was 45%); 390x844 is untouched (34%). With the chart at the top of the screen no bar is covered at 320x568 or 360x640 (was 52 and 88 px at 320x568). Every control stays 44 px tall or more.
Invention: the 700 px height threshold and the spacing values; no new words, numbers or states.
Left: at 320 wide the status tag and the snap label each wrap to a second line (the content is 222 px), so the dock there is 264 px, not the 236 px of 360. A slider sweep to 24 h with Corvid 9 still leaves 15 px of the moved bar under the dock at the far end at every phone height; it is identical on the previous file.
Round thirty-two (2026-11-10)
Defects, from an independent score of 35 of 40 (ship nearly, with one reservation):
- Keyboard focus on the two contested tiles was nearly invisible (severity 2). Reproduced as reported at 1440 light: ten Tab presses land on the Corvid 9 name, the tile drew a 2 px inset ring in the on-accent colour on its own edge, and on the white page the only change was the black edge moving from x=918 to x=920. Dark was the same with black on black. Checking every other tab stop found a second case the reviewer did not name: the console button (Take the next clear slot, Review and publish) drew its ring in the page focus colour on the console ground, which is the same colour (black on black in light, white on white in dark), a ring of ratio 1:1. Fix: the tile ring is now 3 px, drawn 6 px inside the tile edge (offset -9 px), in the tile ink opposite (white on the black tile in light, black on the white tile in dark; the page focus colour once the tile has cleared); the console button takes the console ink for its ring. On a phone the tiles with a link keep the full-bleed 24 px padding after Snap as well, so the inset ring never meets the text. Measured ring against the surface it sits on, all tab stops at 1440 light and 1920 dark: 21:1 in both themes (white on black, black on white), and against the unfocused state the same 21:1 because the ring pixels were the tile fill before. 3 px is above the 2 px minimum and 3:1 is cleared seven times over.
- After Snap on a short phone the slider thumb ran into the CLEAR label (severity 1). Reproduced at 320x568 and 360x640 (viewports 700 px tall or less; the label top was 7 px inside the thumb box, about 3 px into the glyphs). The tall phone and desktop were already clear (gap 5 px). The short-phone rule now gives the range 20 px of bottom padding instead of 12, with a 6 px tick and a one-line label; the gap to the thumb is 4 px box to box, the dock is 244 px instead of 236 (43% of 568, was 42%).
- After Send only a small caption changed (severity 1). The row where the Publish button was now shows a filled block, "Sent to six customers", in the page inversion, the caption below it keeps the facts it already had (range time 04:12 Z, a reopened week does not recall it) in ink, and the Manifest heading reads "Six vehicles, sent to six customers" until Reopen restores "earliest window first". The console already said "Published" and "Week sent to six customers"; no new fact.
Inventions: the 6 px inset and 3 px width of the ring; the 6 px tick and 20 px padding on short phones; the filled "Sent" block and the Manifest caption wording (both restate "sent to six customers" and the 04:12 Z already on the page). No new numbers, rules or products.
Round thirty-four (2026-11-10)
Defects, from an independent score of 37 of 40 (ship nearly, with one reservation):
- Cancel and Reopen the week had no focus ring (severity 3). Reproduced on the previous file at 1440 light: both took focus with
outline-style: nonewhile:focus-visiblematched..link:focus-visibledrew its ring invar(--i), which is defined on.consoleonly, so inside.publishthe declaration was invalid at computed time. Fix:var(--i, var(--sx-focus)), so a link inside the console keeps the console ink and a link anywhere else takes the page focus colour. Every focusable element was then focused after a keypress in six states (before Snap, after Snap, Review, Publish confirm, Sent, Reopened) at 1440 light and dark, 1920 dark, 390 dark, 360 light and 320 light: every one has a solid outline of 2 px (tile links 3 px, inset), and the ring against the ground it sits on is 21 to 1 in each case. - The "Sent to six customers" status looked like the button it replaced (severity 1). It was a solid ink block, 44 px tall, the shape and colour of the primary button. It is now a status: a drawn tick and the words in ink, no fill, no border, no underline, so beside the underlined Reopen link it reads as a result and not as something to press.
- The Hanwell bar overhung its lane box after Snap (severity 1). Reproduced: the bar ran x 1191.9 to 1221.8 against a hatch box ending at 1191.9. The earlier answer (round twenty-eight) was that the window opens the minute the hold ends, so it sits outside the hatch by design; that is still true, so the hatch is unchanged and each lane is now drawn as its own hairline box the width of the axis. The bar sits inside the Hanwell lane box at every width (1440: bar 1191.9 to 1221.8 in a lane box 353 to 1253; on the phone chart the bar is inside its column box).
- A stray tick stayed at the right edge of the phone chart (severity 1). After Snap the hold end and the window open are the same point, so the dimension bracket was drawn with zero length: two overlapping 10 px end ticks. The bracket is now drawn only when it spans 2 px or more, on the phone chart and the wide one (where it left a 12 px tick under the hold line). The "Opens the minute the hold ends" label stays.
- The "048" label was crossed by the dotted corridor edge (severity 1). It now ends at x 126, y 174, clear of the edge (136 at that height) and of the arc, at 390, 360, 320, 1440 and 1920.
Inventions: the lane boxes (hairline, the grid stroke) and the tick drawn with two borders on the sent status; the 2 px threshold for the bracket; the label position. No new words, numbers or states.
Round thirty-six (2026-11-10)
Defects, from an independent score of 37 of 40 (ship yes, with one note):
- A tile's name link un-published a sent week (severity 2). Reproduced on the previous file at 1440 light by keyboard: Snap, Review and publish, Publish, Send, then Shift+Tab from Reopen to the Corvid 9 link and Enter: the verdict went from Published to Conflict, the sent notice hid and the shift went to 0, with no Reopen. The cause was as reported: the tile link set the mover and fired the
changeevent, which zeroes the shift, andupdate()clearspublishedas soon as the model is not clear; nothing gated on the sent state. The same hole was open in the confirm state (Publish pressed, Send not yet): a tile click there dropped the ask and reset the shift, and the slider, the mover toggles, Snap, Reset and a drag on the bar were all live under "Six customers receive this manifest". - Fix: one
frozen()test, true from Publish to Send or Cancel and from Send to Reopen. While it holds the plan cannot change: the slider, both mover toggles, Snap and Reset are disabled;set(), the mover change handler, the tile-link handler and the bar drag each return early; the drag target and the "Drag the bar" hint are not drawn. Choice for the tiles: plain names, no move prompt. The two contested tiles lose theirhrefand their "Move it in the console" line while frozen, so they are not in the tab order, do not answer a pointer and cannot be mistaken for something to press; Cancel or Reopen restores both. A link that moves focus to Reopen with an explanation was rejected: it adds a second place that says "cannot be recalled" and makes a stray tap steal focus. Escape now cancels the ask from anywhere on the page (the listener moved from the publish row to the document), because a press on a locked tile leaves focus on the page and Escape would otherwise do nothing. The confirm and Cancel paths now go throughupdate()so the console, the tiles and the chart redraw together. - Console on the first fold at 390x844 (severity 1): not changed. Measured at 390x844: console top y 742, its heading y 775 to 846, the controls at y 1676 in flow. The heading and the first control cannot both reach the fold without either shrinking the plan below legible label size (it is 261 px tall and its labels are already about 8 px) or docking the controls from first paint over the headline. Both cost more than the finding, so the fold is as it was.
Inventions: the rule that the ask (Send pressed for) freezes the plan as the sent week does, and the document-level Escape. No new words, numbers or states. The "Move it in the console" prompt on tiles that have already cleared, before Send (severity 1 in the score), is unchanged: a cleared week can still be re-planned until it is asked for.
Round thirty-eight (2026-11-10)
Defects, from an independent score of 34 of 40 (ship yes, with one note):
- A repeated tap or Enter sent the unrecallable manifest (severity 3 and 2). Reproduced on the previous file: Publish tapped, then the same point tapped again at gaps of 120, 250, 400 and 700 ms, at 390, 360 and 320: 12 of 12 sent (the reviewer saw 320 safe; here it sent too). Two Enters on Publish at 1440 and 390 gave Published; a held Enter walked Publish, Send, Reopen and back to Publish. After the fix: the same taps at the same gaps and widths sent 0 of 12 from Publish and 0 of 12 from Review; double Enter and held Enter at 1440 and 390 sent nothing, from Publish and from Review.
- Fix, four rules at once. (1) Layout: in the confirm step Cancel, a ghost button, stands where the pressed button stood and Send waits on its own line below it, at every width, so a second press at the same point lands on Cancel. (2) Focus goes to Cancel, the safe answer, not to Send. (3) Send, and Reopen after it, carry
aria-disabledand the disabled look for 500 ms after they appear, ignore a press in that time, and the live region says "Send is ready." or "Reopen is ready." when they arm. (4) A key repeat on Enter or Space is dropped on the publish row and on Review. Reopen follows the same rule because Send hands it focus. - Review smooth-scrolled about 3,200 px (still there). Measured 2,927 px at 390 and 1,383 px at 1440. Review now raises the confirm where it was pressed: the publish row moves into the console in place of the Snap and Reset actions (one set of controls, so one set of rules), and on a phone the dock drops the mover toggle and slider while it asks, since the plan is frozen. Travel is now 0 at both widths. Send, the sent state and Reopen all happen there; Cancel, Escape and Reopen return focus to Review. Publish pressed in the manifest still confirms in place, in the manifest.
- "Move it in the console" shown on cleared tiles after Snap (still there). It is now hidden on a tile that is not contested; it shows only while the window is contested.
Inventions: the 500 ms arming delay (the rule said about 500); focus to Cancel rather than Send; the row living in the console while a Review-raised confirm is open; on a phone the sent status in the console is dropped because the dock's first line already says "Sent to six customers". No new words except the two live announcements ("Send is ready.", "Reopen is ready."). A double Enter on Publish or Review now cancels, by design: the second press lands on Cancel.
Round forty (2026-11-10)
Defect, from an independent score of 36 of 40 (ship nearly, with one reservation):
- At 1024, 1440 and 1920 the console's Send stood exactly under Review (severity 2). Reproduced on the previous file at 1440 light, mouse: Take the next clear slot, press Review at its centre, press the same point again at 550, 700, 1000 and 2000 ms: Send published in every run (4 of 4 and the chain). Round thirty-eight put Cancel first and Send below it on a phone, but on a wide screen the raised row (Cancel, Send, caption, 137 px) was taller than the controls row (100 px), so the grid grew 37 px, everything in it moved, and Send ended on Review's old point. The 500 ms arm covered 0 to 450 ms only.
- Fix: the phone arrangement, made to hold on a wide screen without moving anything. While the console asks or has sent (
.controls.is-asking, 60rem and up) the mover toggle and the slider stay in their cells butvisibility: hidden, so they hold the row's height and the console does not grow or shrink; the confirm row spans the controls and lays out as caption, Send, Cancel from left to right, bottom-aligned where Review stood. Cancel is 16rem wide (Review plus Reset is 256 px) and ends at the right edge, so it covers Review's centre; Send sits to its left, never above or below it. After Send the same cells hold the sent status (where Send was) and Reopen (where Cancel was). Phone and tablet (under 60rem) are untouched. Lines: the@media (min-width: 60rem)block after the phone.is-askingrule, about lines 510 to 519 of index.html. - Proof, by script, not by eye. At 320, 360, 390, 768, 1024, 1280, 1440 and 1920 (touch up to 768, mouse from 768 up, 1024 in both), on four paths (console Review, manifest Publish, Review then Send, Publish then Send): record the pressed control's centre, read
document.elementFromPointthere after the layout changes, then press the same point again at 550, 700, 1000 and 2000 ms (fresh page each) and once as a chain. Result: 0 sends in every run (before the fix: 5 of 5 at 1440 on the Review path). The point lands on Cancel for Review and Publish at 1024 and up and for Publish at every width, on nothing (a wrapper or the caption) on a phone Review, and on nothing after Send at every width. - A repeated tap reopened the week at 320 (found by the proof, then fixed). After Send, on a phone, the Reopen link stood 10 px above the old Send point (x 49 to 150, Send's centre at 152).
elementFromPointreported nothing there, but a touch tap 500 ms or later was snapped to the link by the browser's touch adjustment and the week reopened. The rule covers any state change from a repeated gesture, not only a send. Fix: under 60rem Reopen takesmargin-left: auto, so it stands at the right edge in the console and in the manifest (one rule, just before the wide-screen block). Slop case, a script (slop.mjs) that measures from the old Send centre to every visible enabled control, requires more than 16 px, then taps the point at 550, 700, 1000 and 2000 ms: nearest control is Reopen at 18.1 px (320 console), 68.1 (320 manifest), 58.1 (360 console), 108.1 (360 manifest), 88.1 (390 console), 138.1 (390 manifest); every repeated tap hits the caption, the row or the manifest and the week stays sent (24 of 24). The 320 console margin is 1.5 px over the slop, because the content is only 222 px wide. The earlier four-path proof was re-run at 320, 360 and 390 after this edit: 0 sends, 0 reopens.
Invention: the 16rem width of Cancel and its place at the right edge; the hidden-but-present plan controls while asking or sent on a wide screen (the phone already drops them). No new words, numbers or states.
Round forty-two (2026-11-10)
Defects, from an independent score of 35 of 40 (ship nearly, with two severity 2 defects and three cheap ones):
- A cleared tile threw the resolved plan away (severity 2). Reproduced on the previous file at 390 touch and 1280 mouse: Snap (Clear, +22 h 05 min), tap the Corvid 9 tile's Window line, and the console went back to Conflict with the shift at 0 and Corvid 9 as the mover; the same after Reopen. The round thirty-eight fix had only hidden the "Move it in the console" line on a cleared tile; the name stayed a link with its hit area stretched over the tile (and two extra tab stops). Fix: a tile is a link only while the corridor is contested.
tiles()removes thehrefwhenever the corridor is clear or the week is frozen,.name a:not([href])::afterdrops the stretched area, and the click handler returns early on a link with nohref, so a cleared tile is plain content: nohref, no pointer, no tab stop, in clear, confirm, sent and reopened alike. Nothing on a tile can reset a resolved shift now; Reset and the slider are the way back. Driven fresh for each of the five states, tapping each of the six tiles in turn at 390 touch: 0 of 24 taps changed a clear, confirm, sent or reopened state (before: 2 of 24, Corvid 9 in clear and in reopened); in conflict a tap on Corvid 9 still selects it. The Hanwell II tile (the mover) and the four plain tiles were checked in every state and change nothing. - A thumb 8 to 14 px below Cancel sent the week (severity 2). Reproduced at 390x844 touch: taps 8, 10, 12 and 14 px below Cancel's bottom edge sent (4 of 56 trials). Chromium snaps a tap to the nearest control within about 9 px (measured: a blank tap sends at 1 to 7 px from Send's edge, not at 10), and Send stood 16 px under Cancel. Fix: Send now stands 48 px clear of Cancel (the row gap plus 32 px,
.publish.is-ask #pub-send; on a wide screen the column gap is 48 px in place of 24 and Send keeps its row, so Cancel still ends the row and covers Review's centre). The 30 px a slipped tap can reach past Cancel's edge are all more than 9 px from Send, and Cancel's 44 px box does not touch Send's. - Proof, by script (
slop.mjs). Touch emulation; taps below, above, left and right of Cancel and of every other control within 90 px of Send, offsets 0 to 30 in 2 px steps (the first, at 0, just inside the control), skipping any tap inside Send's own rect, state restored after each; console Review and manifest Publish paths. Result: 0 sends at 390x844, 360x740, 320x568 (console, 64 trials each; manifest, 60) and 1024x768 (64 each); Cancel is the only control within 90 px of Send and 48 px from it. Before the fix: 4 sends at 390 console. - The clock read "0 H 00 MIN" under Published (severity 1). A sent week has no wait to count. The label now reads "Manifest" and the figure "Sent" (the verdict stays "Published", the sub-line "Week sent to six customers"); Reopen restores "Hanwell II waits" and the duration.
- "Reopen is ready." replaced the published message (severity 1). The live region now says "Week published. Six launches, six slots, sent to six customers. Reopen is ready." once Reopen arms: both, published first. The text is one constant.
- After Reopen the manifest caption read as if the week was never sent (severity 1). A
reopenedflag is set by Reopen. The Manifest caption reads "Six vehicles · sent to six customers, then reopened", and the line by the Publish button reads "Reopened after it was sent to six customers, who keep the manifest already sent" (with "Settle the corridor conflict to publish" when the corridor is contested again). Sending again shows the sent captions; the flag is not cleared, because the week was sent.
Inventions: the 48 px separation; the figure word "Sent" and the label "Manifest" on the clock; the two reopened captions (they restate the existing "customers keep the manifest already sent" and "sent to six customers"). The reopened flag lasts until the page is reloaded. No new numbers or states.
Round forty-four (2026-11-10)
Defects, from an independent score of 36 of 40 (ship with a note). Rounds thirty-eight, forty and forty-two each patched one width where a layout change put a control under the finger, and each re-score found another width or height. This round fixes the class, not the width.
- A repeated press reached a control the transition had moved under it (severity 2, twice). Reproduced on the previous file with a sweep (below), touch, 320 wide: after Reopen the second tap at 60 to 2000 ms landed on the slider and un-settled the plan (Hold, shift 1155; 320x640 and 320x800); at 320x800 Send stood under Review's centre, and a tap at 1200 and 2000 ms sent; and a case the reviewer did not name, 320x800, a press at Cancel's point after Cancel moved the slider (shift 315). On the previous file, 63 of 180 cells in three viewports (320x640, 320x800, 1440x900) moved state at 700 ms or earlier (57) or did harm later (6).
- Fix 1, the structural guard (script, near the key-repeat guard).
settle()opens an 800 ms window whenever a press changes the console: Snap, Review, Publish (when it raises the ask), Cancel and Escape, Send, Reopen, Reset. In the window a capture-phase listener on the document (pointerdown,pointerup,mousedown,mouseup,touchstart,touchend,click, non-passive) callspreventDefaultandstopImmediatePropagationon any press whose target is a control (button, input, label, a[href], or the bar's drag target) inside the console or the manifest. A click with no detail is a key press and always passes; key repeat is dropped as before; nothing is announced for a swallowed press. A press the browser retargets (touch adjustment snaps a tap near a control to it) is caught at theclick, whose target is the control. - Fix 2, the placement rule, kept as the backstop (CSS). Under 60rem the dock is a stage of four rows that it keeps while the console asks or has sent: the plan's controls stay in their rows with
visibility: hidden(the actions row too), so the dock does not grow or shrink. The raised row dissolves into the stage (display: contents): Send in the window's row at the top, the caption in the slider's row, Cancel in the actions row at the full width of the Review button it replaces, so a press at Review's centre lands on Cancel and a press at Send's old centre lands on the dock. After Send, Reopen stands at the left of the actions row where Review stood, never over Reset or the slider. Snap and Reset share one row at every phone width (no wrap), because a label that wraps Reset onto its own line moved Snap against the rows above it when the label changed (found by the sweep at 330x800 and 330x844). In a sent week the Snap label in the hidden row stays Review's and the dock's first line drops the repeated "Published" so it stays one line. On a wide screen nothing moved: the 16rem Cancel and the hidden-but-present cells of round forty already satisfy the rule. - Proof, one sweep script (
sweep.mjs). Widths 300, 320, 330, 360, 390 (touch, heights 568, 640, 700, 800, 844), 768 (touch and mouse, heights 768 and 900), 1024, 1280, 1440, 1920 (mouse, heights 768 and 900): 37 viewports. Ten paths: Snap, Review (console), Publish (manifest), Cancel after Review, Cancel after Publish, Send after Review, Send after Publish, Reopen after each, Reset. Each cell is a fresh page: set up, a real press on the control, a second real press at the same point at 60, 250, 450, 700, 1200 and 2000 ms; it records what is under the point, and whether state changed (verdict, shift, mover, ask, sent). 2,220 cells; 0 page errors; 0 cells where the first press did not move state. Cells with any change at 700 ms or earlier: 0. Cells with a harmful change at 1200 or 2000 ms (Send, Reopen, the slider, a mover, Reset): 0. What remains, by design, at 1200 and 2000 ms only: 408 cells where the second, deliberate press lands on Cancel after Review or Publish (the safe outcome, 138) or on Review or Publish after Cancel, Snap or Reopen (it opens the confirm, which asks and sends nothing, 270). Before the changes, on the same three viewports: 57 early and 6 harmful late; with the guard alone, 0 early and 8 harmful late (the placement defects), and with the stage but before the no-wrap rule, 4 (330x800 and 330x844, Snap). - After Send the console still stated a live deadline (severity 2). The facts row "Decision owed by Tue 22 Sep, 18:00 Z, in 37 h 48 min" is now replaced, while sent, by "Decision: Sent to six customers, at range time 04:12 Z" (the existing sent wording and the existing range time); Reopen restores the deadline. "Raised by Range safety, automatic" stays, since it is still true.
- A Tab to a console control left it flush against the viewport (severity 1). At 1024x900 Snap ended at 900 of 900.
.controls :where(button, input)and.publish :where(button)takescroll-margin-block: 1.5rem; Snap now ends 24 px clear, and Tab stops at 390x844, 360x740, 320x568, 768x1024, 1280x800 and 1440x900 stay in view.
Inventions: the 800 ms window (the rule said about 800); the guard also swallows the first press on Snap or Review when it lands inside the window, which includes the same control whose label changed (Snap becomes Review in place); the stage of four rows on a phone and Cancel at full width; Send at the top of the dock (the DOM order stays Cancel then Send, so on a phone Tab goes from the bottom button up to the top one, the safe answer first); Reopen at the left; no wrap for Snap and Reset on a phone; the sent facts row words (all existing). No new numbers or states. Not done: the 8 px by which a sent dock's first line is shorter than a clear one when the dock is in flow rather than pinned (Reopen still stands inside Review's old box).
Round forty-six (2026-11-10)
Defects, from an independent score of 36 of 40 (ship, with one reservation, severity 2, and two severity 1):
- The round forty-four settle guard swallowed every pointer or touch press for 800 ms (severity 2). Reproduced on the previous file, 1280x800 mouse, 390x844 and 320x568 touch: Snap, then at 100, 300, 500 and 700 ms Review (in place of Snap), Cancel after Review, and the Corvid 9 mover segment: 36 of 36 dead, no cue; at 900 ms all three act (checked at 1280). The guard protected the accident case but made a deliberate fast user's tap vanish, including Cancel.
- Fix: the guard is removed, not narrowed.
swallow(),settle(),SETTLE_MSand the seven capture-phase listeners are gone; no pointer, touch or key press is dropped for any control. Only Send and Reopen, the two irreversible steps, refuse a press, and only through the armed state they already had (aria-disabled, the disabled look, "Send is ready." and "Reopen is ready." announced after 500 ms). The key-repeat drop stays. Safety is placement, built in rounds forty to forty-four, and the proof is the same sweep run on the final file, with the guard absent. - Proof, the round forty-four sweep unchanged (
sweep.mjs). 37 viewports, 10 paths, 6 delays, a fresh page per cell, 2,220 cells: 0 page errors, 0 cells where the first press missed. Harmful changes (a send, a reopen, any slider or mover change, an un-settled plan): 0 at every delay. By delay (60, 250, 450, 700, 1200, 2000 ms) each column is identical: 135 cells open the confirm (Review after Snap, Review or Publish after Cancel, Snap after Reopen), 69 cancel it (Cancel under Review's or Publish's point), 166 change nothing. Reversible and safe, so allowed at any delay. The second press after Send hits the dock, the caption or the row, never Reopen; after Reopen it hits Snap or Review. - Deliberate use acts (
deliberate.mjs). Real presses on different controls at 100, 300 and 700 ms, 1280 mouse, 390 and 320 touch: Snap then Review, Review then Cancel, Snap then the Corvid 9 mover segment, 27 of 27 acted (before the fix 0 of 27 below 900 ms). - Send above Cancel but Tab ran Cancel then Send (severity 1). In the console
place()now puts Send before Cancel in the DOM (and Cancel before Send again when the row returns to the manifest, where Cancel stands above Send), so Tab order is the visual order in both homes and on a wide screen (Send left of Cancel). Focus still lands on Cancel; Shift+Tab goes to Send, Tab comes back to Cancel (driven at 1440, 1280, 1024, 768, 390, 360, 320 by keyboard: Esc, Send by key after the arm, Reopen by key all work). - A 85 px hole between the Sent line and its caption at 320x568 (severity 1). The phone dock keeps four rows so Reopen stands where Review will. In a sent week the dockread (aria-hidden) stays in row 1, hidden, to hold the height, and the existing
#pub-sentstatus (tick and "Sent to six customers") is shown in the slider's row, top-aligned, with the caption bottom-aligned just above Reopen: status, 30 px, caption, Reopen. The dock is the same height as before (327 to 568 at 320x568) and Reopen did not move (511 to 556). The hole is now blank air above the status (about 100 px of dock top) instead of between the status and its caption; it cannot be closed without shrinking the dock, which would move Reopen when the dock is in flow.
Inventions: none in words, numbers or states. The two status elements swap which one is shown on a phone in a sent week (both existing). The slider's row now holds the status and caption together; check that two lines never overlap there at very large text sizes (the row is 83 px at 320 and the pair is 52 px).
Round forty-eight (2026-11-10)
Defects, from an independent score of 37 of 40 (ship, with a note: one severity 2 and one severity 1):
- Publish pressed low in the viewport raised a confirm whose Send and warning fell below the fold (severity 2). Reproduced on the previous file, Publish pressed with its bottom edge 24 and 80 px above the viewport bottom: Send top 824 of 800 (1280x800), 868 of 844 (390x844), 592 of 568 (320x568), 792 of 768 (1024x768), 664 of 640 (360x640), both themes; at 200 px the group fits. Scrolling the group in is not safe: Cancel stands at the pressed point with Send 92 px below its top, so a scroll of 104 to 130 px leaves Send under the point. Fix: the group opens upward.
raise(false)measures the in-flow group; if its bottom is past the viewport less 8 px,state.upis set and the row takes.is-up(CSS beside.publish.is-ask #pub-send). The row keeps the 44 px height of the button it replaces, so Cancel stays at the pressed point, and the stack overflows upward: Send 48 px above Cancel, the warning above Send, on a card in the page ground with a 1 px stroke so the tiles behind do not show through.place()puts Send before Cancel in the DOM while up, so Tab order is the visual order (as in the console). Nothing stands below Cancel, so the pressed point holds Cancel or the card, never Send.raise()then callsreveal()for any group still outside the viewport (the console, a tiny viewport): an instant scroll (nothing moves in reduced motion either), 8 px margin, flush if the margin would put Send within 16 px of the pressed point, none at all if even flush would. Measured, manifest, Publish pressed 24 and 80 px from the bottom, light and dark, group top to bottom against the viewport: 1280x800 608 to 776 and 552 to 720; 390x844 651 to 820 and 595 to 764; 320x568 358 to 544 and 302 to 488; 1024x768 575 to 744 and 519 to 688; 360x640 431 to 616 and 375 to 560. All inside;elementFromPointat the pressed point is Cancel in all 20 cells; no horizontal overflow; no page errors. At 200 px the layout is unchanged (in flow). Console Review pressed 24, 80 and 200 px from the bottom, 48 to 200 px from the top, and in its natural place, at 1280, 1024, 768, 390, 360 and 320: 0 of 48 outside the viewport. - A tap just inside a plain tile fired the neighbouring contested tile's link (severity 1, touch). Reproduced at 390x844, 320x568, 360x640, 768x1024 and 1024x768 with touch emulation, in conflict: taps 0 to 12 px inside each edge of every plain tile; 146 of 1,248 activated Hanwell II or Corvid 9 (a page jump of 1,000 to 2,400 px and the mover switched). Fix: under
(pointer: coarse)the contested tiles' stretched link area is inset 16 px (.name a::after), so the neighbour is more than 16 px from the link and the browser's snap radius (about 13 px) cannot reach it. A mouse keeps the full tile. The same 1,248 taps after: 0 hit. The cost: the top and bottom 16 px of a contested tile no longer start the link on touch (they were padding); the middle of the tile and its side edges still do (a tap on the middle of Corvid 9 selects it at 1440, 390, 360 and 1920). - Proof, the round forty-four sweep unchanged (
sweep.mjs), on the final file. 37 viewports, 10 paths, 6 delays, 2,220 cells: 0 page errors, 0 cells where the first press missed, 0 harmful changes; by delay the columns are identical (135 open the confirm, 69 cancel it, 166 change nothing). Run again with Publish pressed 24 px from the bottom, so the upward layout is the one tested (LOW=1): 2,220 cells, 0 errors, 0 misses, 0 harmful; a second press at Publish's point lands on Cancel in 222 of 222 cells.
Invention: the upward layout and its card (one stroke, the page ground, no new words); the 16 px inset. Not done: Review pressed with its top less than about 50 px from the top edge of the viewport at 768x1024, while the dock is in flow, leaves Send above the fold, because the scroll that would show it puts Send under the finger; the point is kept safe and the group is not shown. At 390, 360 and 320 the same press is brought into view flush (Send top 0 to 8).
Round fifty (2026-11-10)
Defects, from an independent score of 38 of 40 (ship, with a note: one severity 2, one severity 1):
- A touch on the phone dock's slider track moved the plan before any scroll began (severity 2). Reproduced on the previous file with CDP touch events: touch-start at 10, 40, 60 and 80 percent of the track set the value at once (390x844: 0 to 100, 565, 875, 1185), then a vertical swipe of 200 px scrolled the page and left the value moved; a touch held and released moved it as well. 40 of 40 track starts failed across 390x844, 360x740, 320x568, 768x1024 and 1024x768 (40 of the 55 cells of the table below). The dock is sticky over the whole console, so the track sits under the thumb of an ordinary scroll. A native range sets its value at touch-start, and CSS alone cannot stop that.
- Fix: where a finger can be the pointer the input does not take the press; the thumb is dragged by a script. CSS, under
(any-pointer: coarse), beside the range rules:.rangeand the input taketouch-action: pan-yand the input takespointer-events: none(the keyboard still reaches it; role, name,aria-valuetextand the label are untouched). A pointer handler on.range(the range handler, after the input listener): a touch or pen that starts within 22 px of the thumb centre arms a drag; nothing changes until the pointer has travelled 8 px and the travel is more sideways than up or down; a start that turns vertical first is dropped, so the browser scrolls (pointercancelends the rest). Once on, the value is the start value plus the distance dragged, not the finger position, and it goes through the existinginputandchangeevents (the sameset(), the sameshowBars()). A track touch that is not on the thumb never changes the value. A mouse on such a screen acts as a native range (a click on the track sets the value, a drag follows). A fine pointer is untouched: the input takes the press itself and the handler returns. A dragged thumb takes the pressed look through.range.is-drag. - Proof, touch-start at 10, 40, 60 and 80 percent of the track (
touch.mjs, CDP touch, DSF 3). At 390x844, 360x740, 320x568, 768x1024 and 1024x768: (a) a 200 px vertical swipe, (b) no movement and release. The value before, at touch-start, after the swipe and after release was unchanged in 40 of 40 (the page scrolled 250 to 263 px in every swipe, 0 on a still touch). A 60 px thumb drag changed the value in 10 of 10, by the amount the native input gave (390: 0 to 320 and 700 to 1020; 360: 0 to 355; 320: 0 to 430; 768: 0 to 135; 1024: 0 to 450), and a vertical swipe that starts on the thumb scrolled the page with the value unchanged in 5 of 5. Total 55 of 55, 0 failures. Mouse click at 40 percent, mouse drag to 70 percent, ArrowRight (+5), PageUp, Home and End, driven at 1280 and 1024 (fine pointer) and at 390 and 1024 (touch emulation with a mouse): all act as before; no page errors. - The strap still said "Move a window" in the sent state (severity 1). While the week is sent the strap now ends "Sent to six customers. Reopen to move a window." in place of "Move a window to see what clears."; Reopen restores the original sentence (driven at 1440, 390, 360 and 1920). The words are existing ones ("Sent to six customers", Reopen), set in the same
update()line. The confirm state, where the plan is also frozen, keeps the old sentence: Cancel is on the same screen and answers it. - Sweep on the final file (
sweep.mjs, the round forty-four one, unchanged). 37 viewports, 10 paths, 6 delays, 2,220 cells: 0 page errors, 0 cells where the first press missed, 0 harmful changes; by delay each column is identical (135 open the confirm, 69 cancel it, 166 change nothing), so the only kinds are cancel, none and open-ask.
Invention: the 8 px intent threshold and the 22 px grab radius (the thumb is 20 px wide, so the radius gives a 44 px target); the relative, not absolute, drag on touch; (any-pointer: coarse) as the condition, so a hybrid screen with a mouse is covered by the handler's mouse branch. No new words beyond the sent strap sentence. Cost, named: on a touch screen the slider can no longer be set by tapping the track. The taps that still move the plan are Snap, Reset and the tile links; the bar drag works at 768 and up and the keyboard works everywhere. WCAG 2.5.7 asks for a non-drag way to do what a drag does, and Snap and Reset are the only fixed-amount taps.
Round fifty, addendum: the hour steppers (WCAG 2.5.7)
The touch rule above made a drag the only way to set the slider on a touch screen. A drag needs a single-pointer alternative, so the slider now has one.
- What it is. Two buttons, one each side of the track (
.slide-rowwraps them and.range): minus, "earlier", and plus, "later". A press moves the shift one hour (HOUR = 60) and clamps it to 0 and 1440, through the sameset()andshowBars()as the slider. They are 44 px squares, a 2 px ink stroke on the ground, filled with ink on press (and on hover, only where hover exists), a 2 px focus ring, disabled at half strength like the slider. Glyphs are 16 px SVG strokes incurrentColor, so there is no second hue and no new token. - Why an hour. The slider's own keyboard step is 5 min, 288 presses across the range. The page already counts in hours (the hour ticks on the bar's axis, the 21 h reset, every duration read as hours and minutes), and the range is 24 of them. An hour is also what the readout moves by in whole figures.
- Names. The accessible name says what the press does and the value it reaches, refreshed in
update(): "Open Hanwell II 1 hour later, to 2 hours later", "Open Hanwell II 1 hour earlier, to 0 minutes later". At an end the button is disabled and says so: "Open Hanwell II earlier: already at the earliest, 0 minutes later", "Open Hanwell II later: already at the latest, 24 hours later". The value itself is announced by the existing output ("Opens later by"). From a non-multiple, such as the snap value 1325, a press goes to 1385 and then clamps at 1440. - Where. At every width, not only where the touch rule applies. One layout and no media branch to keep honest, and the steppers help a mouse as much as a thumb (24 presses against a 1 px drag at 1440). Cost: at 320 the track is about 110 px wide, so a drag moves the shift about 13 min a pixel there; the steppers, Snap and Reset are the precise way, and the drag is the coarse one.
- Frozen plan.
frozen()(confirm, sent) disables both and the click handler returns early. In the console on a phone and on a wide screen the whole slider row is hidden while the console asks or has sent (the existing rule), so the disabled steppers are hidden with it; in a manifest confirm the dock stays and shows them disabled beside the dimmed slider. Reopen and Cancel enable them again. When a press disables the focused button (the end of the range) focus passes to the other stepper, or to the slider if both are off, so the keyboard is never left on a dead control. They are not armed: a press is reversible by the other button, Reset or Snap. - Proof.
steps.mjsat 390x844, 360x740, 320x568, 768x1024, 1024x768 (touch taps), 1440x900 and 1920x1080 (mouse): targets 44x44, no overlap with the track, no horizontal overflow, 0 to 60 to 0, clamp from 1325 in three presses, 1440 to 1380, Enter on the earlier button at 60 hands focus to the later one, Shift+Tab from later reaches the slider, both disabled in the console confirm, in the sent state and in a manifest confirm, a forced click on each does nothing in all three, a real tap on the visible disabled "later" in the manifest confirm does nothing, both enabled again after Reopen: 0 failures, 0 page errors. Thetouch.mjstable is 55 of 55 again. The sweep on the final file: 2,220 cells, 0 page errors, 0 first presses missed, 0 harmful, columns identical to before (135 open the confirm, 69 cancel it, 166 change nothing).
Invention: the one-hour step, the 12 px gap, and the disabled-at-the-end wording of the two names (no new product facts).
Round fifty-two (2026-11-10)
Defects, from an independent score of 35 of 40 (ship, with a note: four severity 1). Each was reproduced on the previous file before the edit (a, c and d at 320x568 and 1024x768 as reported; b at 1280x800).
- Reopen below the fold at 320 wide, with focus on it (a). Publish pressed in the bottom 100 px, then Send: Reopen top 559 of 568 (320x568) and 791 of 800 (320x800), the caption below it out of view. Fix:
revealSent(), called after Send's focus move. It scrolls the page, instantly (nothing moves under reduced motion either), so Reopen and the caption under it are inside the viewport with an 8 px margin; if both cannot fit, Reopen alone; if the scroll would put Reopen within 16 px of the point where Send was pressed, it does not scroll at all (the round forty placement rule outranks the view). Measured, manifest, Publish pressed low: Reopen 465 to 510 of 568 (320x568), 697 to 742 of 800 (320x800), 758 to 803 of 844 (390), 538 to 583 of 640 (360x640); caption in view in all;elementFromPointat Send's old centre is the sent status or a tile, never a control, in all four; a repeated touch tap there at 700 ms left the week sent (390, 360, 320x568, 320x800, 768). The round forty-four placement rules are untouched. - Tab left the console confirm (b). Focus starts on Cancel; Tab went to "Back to the portfolio" and scrolled from 554 to 3095 with the confirm open. Fix: one keydown listener on the document. While the confirm is open and focus is on Send or Cancel, Tab and Shift+Tab move to the other of the two (wrap), and Escape still cancels. A focus that is already elsewhere keeps the browser's own Tab. DOM order was already the visual order (round forty-six put Send before Cancel in the console and when the manifest confirm opens upward); verified by script in each layout: Send first in DOM and first by position in the console (left of Cancel on a wide screen, above it on a phone), Cancel first in the manifest. Driven at 1440, 1920, 1280, 1024, 768, 390, 360, 320x568 and 320x800: Cancel, Tab to Send, Tab to Cancel, Shift+Tab to Send, in both confirms, scroll position unchanged, focus in view; Escape cancels and returns focus to Snap or Publish.
- The track was 100 px wide at 1024 (c). From 60rem to 72rem (960 to 1151 px) the three columns (movers, slider, actions) are now two: movers and slider share the first row and the actions take a row of their own at the right edge (CSS after
.controls). Measured track: 435 px at 1024 (100 before), 467 at 960, 562 at 1151; 228 at 1152 and 356 at 1280 as before; 558 at 768 (the phone dock). The DOM order (movers, slider, actions) is still the visual order. While the console asks, the confirm takes the actions' row (grid-row: 2), the hidden movers and slider hold the row above, so Cancel stands over Review's old centre (Review 671 to 859, Cancel 671 to 927 at 1024) and the console is 201 px at rest and 200 asking, so nothing moves. - The first arrow after a Tab onto the slider jumped the page (d). At 320x568, scrollY 1450, Tab onto the slider showed no bars, the first arrow scrolled 388 px. Fix: a Tab onto the slider or either hour stepper now makes the scroll an arrow press would have made, at the Tab (
showBarson focus, one frame later). Only a Tab: the key is remembered by a capture-phase keydown and cleared by a pointerdown, so a touch drag (which focuses the slider) and Snap or Reset (which focus it from a script) scroll as before and nothing scrolls twice. Instant under reduced motion becauseshowBarsalready is. Measured: 320x568 1450 to 1062 at the Tab, 1062 after the arrow; 360x740 1450 to 1055, 1055; 390x844 1450 to 1081, 1081; reduced motion, one instant scroll event; 768x1024 bars already in view, none.
Inventions: the 72rem upper edge of the two-row band (the 3-column track is 228 px at 72rem and falls below 200 under it); the margin and the 16 px guard in revealSent() (the same numbers reveal() uses); stepper focus treated like slider focus. No new words, numbers or states.
Proof. sweep.mjs, the round forty-four one, unchanged, on the final file: 37 viewports, 10 paths, 6 delays, 2,220 cells: 0 page errors, 0 first presses missed, 0 harmful; every delay column is 135 open-ask, 69 cancel, 166 none, so the only kinds are cancel, none and open-ask. Some wide-viewport cells were lost to a crashed tab under memory pressure from other sessions and were re-run at lower concurrency until none were lost. touch.mjs (CDP touch, 55 cells, unchanged): 55 of 55 in most runs; about one run in four has one miss, always the same cell (390x844, thumb at 700, drag +60 px: value unchanged, page scrolled 45 px). The harness sets 700, which starts the page's own 145 ms smooth scroll, waits 100 ms and then touches while the page is still moving; the browser takes the touch for the scroll and cancels the pointer (logged: pointerdown at 23 ms, scroll events to 143 ms, pointercancel at 105 ms). With the harness waiting 600 ms after the set the same cell is 8 of 8 clean, and in 12 isolated repeats at the original 100 ms it is 12 of 12 on this file and on the previous one. The previous file showed no miss in 8 runs of the full script, so the rate may be higher here; the cause I could find is timing, not the drag handler, and I did not find a change of mine on that path. The slider is 435 px at 1024, so the touch.mjs thumb drag of 60 px now moves 210 min there (450 before).