Cloud security posture across three AWS accounts, built as the page an operator trusts at 3am: side navigation, breadcrumbs, a flashbar, the posture figures, one collection and a panel that answers the row selected.
Direction
| Lever | Decision |
|---|---|
| Archetype | Resource console. AppLayout idiom: top navigation, side navigation, breadcrumbs, a notifications bar, a content column that scrolls, and a right-hand panel (split panel at 1280px and wider, a bottom sheet below). |
| Data stance | Evidence-first. A verdict, one figure, one chart and a measures matrix sit above the table, and every one of them is computed from the 248 rows in the table. |
| Density | Cloudscape compact by default, comfortable on request. space-scaled-* shrinks with the setting; space-static-* does not. |
| The one bold thing | A fix is a change set you read before it runs. Review fix runs preflight checks, shows a before and after, states the cost and the permission, and needs an acknowledgement when real traffic would be cut off. The legend still prints the contrast ratio the browser measures. |
Why the archetype changed
The first version was a ten-tile wall display that never scrolled. It answered "is anything wrong" from six metres and nothing else: there was no way to find a finding, read one, filter, sort or act. The owner's brief for this round asks for the console idiom Cloudscape is actually used for, so the wall display went and its best parts stayed: the verdict, the contrast legend, the degraded state, the typed confirmation.
Structure, in the order a visitor meets it
- Top navigation. Brand, organization scope, a link back to the portfolio, Display settings, the signed-in user.
- Flashbar. On load: three change sets wait for approval, with an action that filters to them. Later: applying, verifying, resolved, suppressed (with Undo), copied, feed stopped. Each updates in place.
- Page head. "Security posture", an Info link that opens the help panel, a Refresh button that shows a loading state, and a ticking "Refreshed 14:02:36 UTC, 28 seconds ago".
- Verdict, figure, meter, chart. Computed. Twenty-one critical findings against a limit of ten; the meter draws the limit tick and the overage; the chart draws two series by shape and colour, a dashed threshold, end labels, and a cursor with a text readout the arrow keys drive.
- Accounts and measures. Stacked bars (each a button that filters the table) and a matrix of seven measures by account. Any count that is not zero links to the findings behind it.
- The collection. Property filter with operators and free text, tokens, sortable columns, a select-all checkbox, pagination with ellipsis, and preferences for page size, wrap, compact and striped rows, and which columns show. On a phone the rows become cards.
- The panel. One finding in full, a summary of several, or help (status vocabulary, service levels). Resizable by pointer and by arrow keys.
The two states
Reporting. Every feed is live; the trend is solid to today; detectors read 9 of 9.
One feed stopped (Display settings). Security Hub in eu-west-1 goes quiet at 13:41 UTC. The trend stops at the last real reading with no marker; the hero reads 19 verified critical findings and says two more are not verified; sandbox-data's bar, counts and matrix cells become an em dash or are marked; its rows say "Not verified since 13:41 UTC"; detectors read 8 of 9 with an error; and a fix for a sandbox resource is refused at preflight. Zero and no data are different marks.
Figures that check
Everything derives from one seeded set of 248 findings whose quotas are written down: 126, 78 and 44 by account; 21, 47, 108 and 72 by severity. A script confirmed the severity-by-account matrix, the category counts, and the trend against the rows' opened times before the page was built. The numbers that changed from the first version, because they did not reconcile:
- The line crossed the limit on 8 Sep, not 7 Sep. The first reading above 10 is the 8th.
- The projected September spend was $908. At $30.80 a day for 11 days the month to date is $338.80, and 30 days projects to $924.00.
Applying a fix removes a row, and the nav count, hero, meter, matrix, chart's today point and verdict all follow, with a brief highlight on each number that moved. Nothing counts up.
Generator
cloudscape_tokens.py --both > tokens.css # the linked file, never hand-edited
cloudscape_tokens.py --charts # the 155 data-visualisation tokens
cloudscape_tokens.py --density # comfortable vs compact, 24 of 140 differ
tokens.css is generated output. Every colour and every radius on the page is a var() from it.
Status vocabulary, used exactly
| State | Where it appears |
|---|---|
| success | A measure on target; a check passed; resolved |
| error | A measure over threshold; a service level missed; a check failed; a feed not reporting |
| warning | Off target but not over; a service level due soon; real traffic would be cut |
| info | A new, unassigned finding; a neutral fact |
| pending | Awaiting approval |
| in-progress | In progress; applying; verifying |
| loading | Waiting on an answer |
| stopped | Defined in the help panel; nothing on this page has stopped for good |
Measured contrast
Every pair sits on color-background-container-content (#ffffff light, #161d26 dark) and every number comes from web-design/scripts/contrast.py.
| What | Token | Light | Dark | Floor |
|---|---|---|---|---|
| A value read off a panel | color-text-body-default | 18.50:1 | 9.98:1 | 4.5:1 |
| Critical line, 2px | color-charts-red-500 | 5.01:1 | 6.21:1 | 3:1 |
| High line, 2px | color-charts-orange-500 | 5.00:1 | 6.23:1 | 3:1 |
| Critical bar, filled | color-charts-red-400 | 4.01:1 | 4.99:1 | 3:1 |
| Low bar, filled | color-charts-blue-1-400 | 4.03:1 | 4.96:1 | 3:1 |
| Primary button label | color-text-button-primary-default | 4.97:1 | 8.11:1 | 4.5:1 |
In the light theme the number in a chart token's name is its ratio to the second decimal. In the dark theme it is a floor. The legend prints what the browser measures, so a reader sees the difference rather than taking it on trust.
color-charts-red-300 is 2.998:1 on white and misses the 3:1 floor for a graphic by a thousandth, so the filled bars take the -400 step.
Deviations
- Spacing and motion are authored.
cloudscape_tokens.pyemits colour, radius, border, type and shadow and stops there. The page declares the compact and comfortable space columns and the motion table at the top of its own stylesheet. - The modal scrim is a
color-mix(). The token set has no modal-overlay token, so the scrim is 62% ofcolor-background-home-header. - One panel at a time. Cloudscape lets the tools panel and the split panel coexist; here details and help share one dock and the rail buttons are exclusive, so a phone has one sheet to manage.
- Menus use the native
popoverattribute and dialogs use<dialog>, both positioned and closed with a short fade that the reduced-motion block shortens to a single frame. AppLayoutis a React component. Its regions are present as markup.<meta name="theme-color">carries two literals. A meta attribute cannot hold avar().
Kept warnings
copy-lint.ignore keeps operator-copy, which fires on 0.0.0.0/0. On a cloud security console the CIDR block is the finding itself.
Round two: the count outranks the flashbar
An independent critic, judging the first fold cold, scored this page 6.5: the blue flashbar outweighed the "21" it existed to explain.
- The count is the focal moment. "21" is set up to 208px in the error colour, under its label, beside an error status indicator ("Two measures are over threshold"), the +3 delta and the limit meter. The chart runs full width beneath it, so the line crossing the limit on 8 Sep is the next thing read.
- Deviation, stated. 208px is past Cloudscape's largest display step (64px). Traded: system fidelity in one figure for a first screen whose one question is one count. Every other size stays on the scale.
- The flashbar is the size Cloudscape gives it. One line at 1440; at 390 the wait flash's body is visually hidden but still read, so the bar is two lines rather than five.
- Checks: 21 − 10 = 11 over; the critical series runs 7, 10, 12, 15, 18, 21 over the last five days, so the first day above the limit is 8 Sep and "+3 since 10 Sep" is 18 to 21.
Round four
- Finding titles wrap instead of being cut: the title is the primary link, so it shows whole on two or three lines (the Resource column keeps its ellipsis).
- The measured contrast left the customer's dashboard. The chart legend is now two keys (Critical, High) and the accounts card has no footer note. The live browser measurement stays, printed once, in the design notes under "Readability, measured". This replaces the earlier decision that "the legend still prints the ratio": the contract keeps the design system as the subject of the notes panel only, and a founder reading a security dashboard has no use for "a drawn mark needs 3:1".
- The side navigation follows the 1280px breakpoint in both directions: resizing down shuts the drawer, resizing up opens the column. (The 30,931px height did not reproduce: 390px is 6,675px tall fresh and after a resize from 1440.)
Round nine
Found below the fold, in driven states and at 360px; each fix is one line.
- The findings table was 23px wider than its card at 1440, so the Assignee names were cut at the edge, and wider still with the panel open. Finding and Resource columns were re-sized so it fits at 1440, and the scroller is now a focusable region with an edge shadow on whichever side has more, so a row never ends mid-word without saying so.
- At 360 and 390 the "Crossed the limit on 8 Sep" annotation ran off the chart. It now measures itself and says "Over from 8 Sep" when the long form will not fit.
- The change-set diff on a phone broke
falseandBlockPublicBucketsmid-word. Rows become a setting line over a before and after pair. - The phone measures list was six lines per measure. Total and Status now share one row; at 360 the account labels are 11px so
prod-paymentsstays on one line. - A void under the accounts card at 1440 and 1920. It now holds a "By category" list (nine rows, each a button that filters the table), computed from the same rows as the matrix.
- The in-progress status icon spun forever on every row. It is state the data holds, not work this tab is doing, so it is static; only loading spins.
- The design notes ran each term straight into the previous paragraph. Terms now have a clear space above them.
- The filter placeholder was cut at 360 ("...property or va"); it is shorter below 480px.
Round ten
A second, colder read. Everything below was found by driving the page at 1440, 1920, 390 and 360 (touch, 3x), both schemes.
- A suppressed exposure read as fixed. Suppress the S3 finding and the hero said "No resource is reachable" once all four were gone, while the dialog said they stay reachable. Suppressed exposures now keep the measure red: the note, the matrix row ("1 suppressed, still reachable") and the verdict count them, and the suppress dialog and notification say so.
- Resolved had no way back. The change set promised "turn it off again with another change set" and nothing offered it. The Resolved notification now carries Roll back, and it stays until used or dismissed. Notifications with an action are the last to be evicted when the bar is full.
- Assign to me had no Undo and overwrote an assignee silently. It now has one, restoring assignee and status.
- The result landed out of sight on a phone. The bar scrolls with the page below 768px, so Suppressed, Resolved, Assigned and Copied appeared screens above the reader. They now scroll into view and take focus (the notification itself, so Enter cannot press Undo). On a phone the empty detail sheet closes after the act, so it does not sit over the notification. The ARN copy button shows a check as well.
- Focus was lost after a suppress. A closing dialog hands focus back to its opener, which had been redrawn; the move now waits for the close.
- The act was below ten rows of facts. In the details panel, Review fix, Assign and Suppress now sit under the title; on the phone sheet they were four screens down.
- Help only in the side nav. Info links now sit beside the critical figure (the limit), the Measures heading, and the Service level column header. Two new help topics, "The policy limit" and "Measures against target". The help panel's term lists were never styled; they are now.
- Text that did not follow the rows. The chart heading said "risen for five days" after a suppress made today lower; it now reads from the series (and says four on a stopped feed). The "rising for five days" clause follows it. "Three change sets are waiting" now counts down as they are approved. Dates in prose were "4 September" and "8 September" beside "8 Sep" in the chart; all are "8 Sep". "Nine of nine" is "9 of 9".
Inventions, named
- Roll back is one press with no preflight and restores the recorded before: the finding returns with the status it had, nothing is asked. The brief gives no rule for undoing an applied change set.
- A suppressed exposure still counts as reachable. The brief says a suppressed finding stops counting toward every measure; for the one measure that describes the resource, not the finding, it does not.
- The text of the two new help topics. Definitions only, from numbers already on the page.
Left
- The seven-row measures table still ends with an ingestion-cost caption that belongs to no measure (severity 1, a fact the brief puts there).
- The four exposed resources are not marked in the findings table; "Show them" filters to them (severity 1).
Round eleven (2026-11-10)
Two defects from an independent read at 390 and 360 on a touch phone. Both reproduced before the edit (64 targets under 44px; one approval threw the page from y 3234 to y 39 and left a Resolved notice with only Roll back).
- A touch phone opened Compact, with 64 targets under 44px. On a coarse pointer the page now opens Comfortable (Cloudscape's own density for touch); the Display settings radio shows it, and choosing Compact afterwards still wins. A
(pointer: coarse)block also holds a 44px floor in either density for buttons, the flashbar action and dismiss, Info and matrix-count links, the filter field, sort buttons, category rows, pager, breadcrumbs and the nav controls. The 24px row checkbox stays 24px and its cell is the 44px target (a press beside the box ticks it); the token remove x keeps its look and borrows 12px around it. After the fix 11 targets measure under 44: the hidden skip link (40px, shown on focus) and the ten row checkboxes whose cells are 44. Desktop (fine pointer) is unchanged, Compact. - Approving threw the reader out of the queue. "Show the 3" no longer removes the waiting notice. It stays, counts down as change sets are approved ("Show the 2"), and is how the reader gets back to the table, since the Resolved notification is revealed at the top of the page on a phone and sits directly above it. It leaves when none wait, and a roll back that puts one waiting again brings it back. When the bar is full the waiting notice is evicted after every other notification, because its count is the queue. The design notes line on spacing now says Compact with a pointer, Comfortable on touch.
Inventions, named
- Coarse pointer means Comfortable by default. No product fact; a presentation rule, and the reader's choice overrides it.
- The waiting notice persists after use. It previously went on first use; the brief gives no rule for either.
Found, left
- Three change sets are "waiting", but only the Critical one has a change set to review. The two High rows (an inbound rule on tcp/3389, an IAM user without MFA) show Assign to me and Suppress only, so after the first approval "Show the 2" leads to rows that cannot be approved here. The brief gives no change set for them.
- On a phone the row's date still wraps onto its own line beside the 44px checkbox cell.
Follow-up: what waits is what this reader can approve
The "Found, left" item above is resolved by deciding the fact one way. The brief gives a change set only for an exposure (the preflight, the before and after, the permission and the cost are all written for revoking 0.0.0.0/0 or turning on Block Public Access, and Review fix is the one bold thing on the four exposed resources). It gives nothing for an inbound rule open to internal addresses on tcp/3389 or an IAM user without MFA: no permission, no before and after, no cost. So those two High rows have no change set awaiting anyone.
- The two High rows were generated as "Awaiting approval" only to make the count three. That block is gone; they keep the status and assignee they had (In progress). Only an exposure can wait for approval, and one does: the Critical rule on tcp/22, submitted by Søren Lindqvist.
- This supersedes Structure item 2. On load the notice reads "One change set is waiting for approval." with "Show it", the Awaiting approval filter returns one row, and the same count drives the notice head, its hidden sentence ("submitted it ... approve and apply it") and the button. No figure elsewhere repeated three; the access note and "three accounts" are unrelated.
- Approving it takes the count to none and the notice leaves; Roll back puts the finding back as Awaiting approval and the notice returns. Both driven at 1440, 1920, 390 and 360.
Invention: none. The count and the notice now only report what the rows hold.
Round twelve (2026-11-10)
Two severity 2 defects from an independent read, both reproduced at 1440 light before the edit.
- Review and approve stayed live while the change applied. With the panel open, a second press within the 3 s of "Applying" opened a second dialog for the same change set; its Apply, enabled after the checks, closed it with no message and no change. The panel's Review button now takes Cloudscape's loading pattern while the finding is Applying or Verifying: a spinner,
aria-disabledandaria-busy(it stays focusable, the press is ignored), the primary-disabled colours, and a hidden sentence it is described by ("Change set cs-... is applying, so it cannot be reviewed again.").openChangealso returns for a finding that is applying or already closed, so no path opens a dialog that can do nothing. It comes back with the finding if it is rolled back. - Focus was lost after Approve and apply. The dialog closed, the panel behind it was redrawn, and focus fell to the page. Focus now rests on the "Applying change set" notice and stays on it when that same notice becomes "Resolved" (Roll back is the next Tab stop; the notice itself is focused so Enter cannot press Roll back, as on a phone since round ten). A reader who has moved on during the apply keeps their place. Phones already did this through
reveal. - Found while driving a double press. The second press of a double click on Review landed on the backdrop that had just opened under it and closed the dialog. A backdrop press is ignored for 500 ms after a dialog opens (all three dialogs share the helper).
Driven at 1440 light, 390 dark, 360 light and 1920 dark, by mouse and by keyboard only, with a double press on Review and on Approve and apply: one dialog, one change set applied, Review inert while applying, focus on the notice at Applying and at Resolved, Roll back one Tab away, no page errors, no horizontal overflow.
Invention: the wording of the hidden reason, a presentation sentence from the change set's own id and state; no product fact.
- Opening a finding dropped focus (found in the same pass, fixed on request). Enter on a row's title redrew the row, so the pressed button was gone and focus fell to the page before the reader reached Review. Focus now rests on the row's replacement button, which is also kept as the panel's opener so closing the panel returns to the same row. On a phone, where the panel is a sheet, focus goes to the sheet's title (the next Tab stops are Close, then Review and approve). The panel does not take focus on a wide screen; from the row, Tab passes the remaining rows, the pager, the resize grip and Close before Review and approve (four stops with the table filtered to one row).
Driven again at 1440 light, 1920 dark, 390 dark and 360 light by keyboard only, with the whole approval path: open the finding, Tab and Shift+Tab, Enter again, Review, double Approve, Applying, Resolved, Roll back one Tab away. No page errors, no horizontal overflow.
Round twenty-eight (2026-11-10)
One severity 2 and two severity 1 defects from the round 27 read, all reproduced before the edit (1440x900, 1280x720, 1920x1080: after a row 10 open or the Show it path the panel head sat 140 to 645 px above the viewport; 390x844 dialog Apply at top 840 against a dialog ending at 828).
- The detail panel opened with its title and actions off-screen. Beside the page (1280px and up) the panel sticks only while its column lasts; a row low in the table puts the column end above the fold and pushes the panel up and out. Opening a panel (a row, an Info link, the rail) now scrolls the page up by what is missing (
revealDock), so the title and the actions under it sit at the top of the viewport; below 1280px the panel is already a fixed sheet. Measured after the fix at all five widths from row 1, row 10 and Show it: dock top 48 (or the sheet top), title in view, Review and approve in view where the finding has one. - Focus after opening a row now lands on the panel title (as it already did on a phone), not on the row button. The row button is still the opener, so closing the panel returns to it. This supersedes the round twelve choice to leave focus on the row; the actions are one Tab from the title.
- A bare "Checking…" in the preflight. Each check now names what it is checking ("Checking clients outside the organization…") until its result replaces it.
- Phone dialog buttons below the fold. Below 768px the dialog footer is sticky to the bottom of the dialog with a divider, so Cancel and Apply are on screen when it opens (390x844 Apply 739 to 783 in a dialog ending at 821; 360x740 666 to 710 in 724).
Inventions, named
- The wording of the nine "what" labels on the checks (presentation text from the check each row already holds); no product fact.
- Scrolling the page when a panel opens is a presentation rule; the brief gives none.
Left
- Esc still does not close the panel (severity 1, not asked); the display-settings Esc focus and the stopped flashbar wrap are untouched.
Round thirty-two (2026-11-10)
One severity 2 and two severity 1 defects from the round 29 read, all reproduced before the edit.
- The trend chart could not be reached by keyboard (severity 2). The page said "focus it and use the arrow keys", but
#chart-hosthad notabindex: 40 Tab presses never landed on it and the keydown handler never fired. It is now a tab stop (tabindex="0",role="application"so a screen reader passes the arrow keys through,aria-roledescription="line chart", anaria-labelnaming the series and the 29 Aug to 11 Sep range,aria-describedbythe sentence under the heading). It takes the system's focus ring, drawn inside the box so it does not cut the line of text above. Left and Right move one day, Home and End go to the first and last day, and each press announces the day's values ("10 Sep: 18 critical (8 over the limit), 44 high") through the page's polite status region. The readout is no longer itself a live region: a mouse sweep would otherwise read out every day it crosses, and leaving the chart would read the instruction. Reaching the chart now takes 22 Tab presses from load (it sits after the Refresh button). Every one of the 14 days was read by keyboard alone. - "Hover the chart" on a touch phone (severity 1). Where the primary input cannot hover (
(hover: none)) the line reads "Tap the chart to read a day." Tapping did not work before: the finger lifting firedpointerleave, which cleared the day before the tap's focus arrived. Only a mouse leaving now clears it; a touch keeps its day until the reader taps elsewhere. Tapping the middle of the plot read 7 Sep, and a tap near the right edge read 11 Sep, at 390 and 360. - Esc did not close the panel below the wide frame (severity 1). Below 1280px the panel is a sheet over the page; Esc now closes it and focus returns to the row (or Info link) that opened it. At 1280px and up the panel is part of the layout and Esc leaves it. A dialog, an open menu, the filter's list and the open navigation drawer take Esc first (menu and drawer driven at 390 and 360: one Esc closes only the menu or drawer, the sheet stays, a second Esc closes the sheet).
Inventions, named
- The chart's accessible name and the "Tap the chart to read a day." wording are presentation text built from the chart's own series and range; no product fact. The sentence under the heading gained "Home and End read the first and last."
- Esc closing the sheet only below 1280px is a presentation rule; the brief gives none.
Left
- The sentence "Left and right arrow keys read each day" still shows on a touch phone, where it is true of a keyboard but not of the reader's hand; the instruction line under the chart is the one that now says "Tap". The 360px x-axis labels "29 Aug" and "2 Sep" sitting close together (severity 1 (b)) and the two stacked notices (c) were not asked and are untouched.
Round thirty-four (2026-11-10)
Four findings from the round 33 read, all reproduced before the edit (probe at 320, 360, 390, 1440, 1920).
- The 360px x-axis labels touched (severity 1). "29 Aug" and "2 Sep" were 2.9px apart at 360 (3.2 at 320, 7.3 at 375). The axis now tries the four labels (29 Aug, 2 Sep, 6 Sep, 11 Sep), then three (29 Aug, 6 Sep, 11 Sep), then the first and last day only, and keeps the widest set whose labels are all at least 10px apart (
TICK_GAP, measured withgetBBoxafter the chart is drawn). The first and last day are always labelled. Measured minimum gap in a sweep every 5px from 320 to 440 and at 480 to 1920: 10.3px (at 340 and 385, the widths where a set is about to give way), 18.0 at 360, 12.0 at 390, 389 at 1920; two labels at 320 to 335, three at 340 to 380, four from 385. - The Actions dropdown took no keys (the consistency heuristic, scored 3). Focus now moves to the first enabled item on open; Up and Down move (and wrap), Home and End jump; Esc closes and the popover returns focus to the trigger; Tab closes and goes on from the trigger. Down or Up on the trigger opens it. The panel is
role="menu"withrole="menuitem"buttons, as the trigger'saria-haspopup="true"already promised. - "Arrow keys" on a touch phone (severity 1). The sentence under the chart title is set from the device, as the instruction line already was: where hover is unavailable it reads "Tap the chart to read a day."; otherwise the keyboard sentence stays. The markup keeps the keyboard sentence, so it is also what shows without script.
- "Awaiting approval" clipped (severity 1). The cause was
max-width: 20chwithoverflow: hiddenon every cell when the table does not wrap (it measured 172px of text in 160px). The status column now has no cap and no clip, so the label fits on one line. The table still scrolls sideways at 1440 by about 60px with no dock; that was not the status (it measured 1180 before and 1181 after) and is the width of the columns together.
Inventions, named
- The 10px floor between axis labels and which days a thinner axis keeps (the first, 6 Sep, the last) are presentation rules; the brief gives none.
- Up and Down wrapping at the ends of the menu is the WAI-ARIA menu button convention, a choice where the brief is silent.
- Two fixes the findings did not name, both so the page does not overflow at 320px: the top bar gets 4px gaps and padding below 360px (it was 38px wider than a 320px screen), and the chart's smallest width drops from 280 to 220px (it was drawn wider than its 238px box).
Left
- At 320px the "Policy limit, critical: 10" label meets the rising line near its end (the chart is 238px wide there); at 360px it ends at the line, as before.
- Esc still does not close the dock at 1440 (persistent there, not asked).
Round thirty-six (2026-11-10)
Three findings from the round 35 read, all reproduced before the edit (1440x900 for the notices and the keys, Chromium, light).
- Two stacked notices held 162 of 900px (severity 1). After one fix the bar measured 48 to 210 (Resolved 96px, waiting notice 54px) and after two fixes 262px (three notices), covering the table. The bar is now Cloudscape's stacked Flashbar (
stackItems): from two notices on, the newest shows whole, the others are hidden behind it with up to two slim layers in their own colours, and one toggle under them names the count ("View all 3 notifications", chevron down; "Collapse 3 notifications" when open, chevron up,aria-expanded). Each notice keeps its own Dismiss. Measured at 1440 after one fix: 162 to 140px; after two fixes: 262 to 144px (open, the reader's choice, it is 294px). The toggle takes the live region's announcement of the count; one notice left hides the toggle and collapses the stack for next time. An update to a notice (Applying to Resolved, a feed message re-posted) moves it to the top, so news is never behind an older notice. Focus: dismissing a notice passes focus to the next one; a notice hidden by a newer arrival while it held focus hands it to the toggle. - The table header sits below the bar. The findings table header is not itself sticky (the table scrolls sideways in its own region, which a sticky header cannot cross), so what covered it was the sticky bar over scrolled content. The bar's height is now published as
--flash-h(a ResizeObserver; 0 where the bar scrolls with the page, below 768px) and added toscroll-padding-top, so a header, row or focused control scrolled into view stops below the bar: header top 208 against a bar ending at 192 at 1440 after two fixes, 144 against 128 at 1920. - Home and End on the panel resize separator scrolled the page. Home now sets the minimum width (320) and End the maximum (640), the window-splitter pattern, with preventDefault; Left and Right are as before. Driven at 1440 and 1920: 640, 320, 640, 624, 320, all defaultPrevented.
- Focus left Next page for the "2" button. The redraw replaced every button and focus went to the current page. It now returns to the control pressed: Next stays on Next, Previous on Previous, a page number on its (now current) number. Where the pressed control is now disabled (Next onto the last page, Previous onto page 1) focus goes to the current page button beside it, the nearest enabled control. Driven at 1440, 390, 360, 1920 and 320.
Inventions, named
- The toggle wording ("View all N notifications", "Collapse N notifications") and the grey bar from the notification-grey token: Cloudscape names a count and a toggle but the page has no string for them. Presentation text from the count; no product fact.
- Two layers at most, whatever the number, and the stack resetting to collapsed when one notice is left: presentation rules the brief does not give.
- Moving an updated notice to the top, and moving focus on dismiss: presentation rules; the brief gives none.
Left
- With the panel open at 1440 the Service level, Opened and Assignee columns still scroll out of view (not asked). On a phone the bar still scrolls away with the page, and the stack shortens it to the newest notice (216px at 390 with two notices, 256px at 320).
Follow-up: the one-fix case at 64px
The reviewer's case is one fix with the panel open: Resolved over the waiting notice. At 140px the stack was barely smaller than 162. Two causes, both cut.
- The Resolved notice repeated the page. It read "Security Hub confirmed it at 14:03:12 UTC. 19 critical findings remain open, against a limit of 10.", and the hero beside it already shows 19, the limit and "9 over". The notice now reads "Security Hub confirmed it at 14:03:12 UTC." Head and text and Roll back sit in one 32px row at 1440 with the panel at its default 400px (the action inline on the right). The spoken line ("Resolved ..., 19 critical findings remain.") is unchanged, so a screen reader still hears the count. No type was shrunk.
- The stack cost 44px of chrome. The layers were 4px boxes and the toggle 28px with 4px padding. The layers are now box-shadows on the newest notice (no height; two only from three notices), the toggle is 24px (the WCAG 2.5.8 floor; 44px on touch as before) and the bar has no padding under it.
Measured at 1440 and 1920 (panel open, compact), flashbar height, with the table header's top after the header is scrolled into view (bar bottom plus 16):
| State | Round 35 | First pass | Now | Header top, scrolled in |
|---|---|---|---|---|
| One fix | 162 | 140 | 64 | 128 (bar ends 112) |
| Two fixes | 262 | 144 | 68 | 132 (bar ends 116) |
Identical at 1440 and 1920. With the panel dragged to its widest (640px) the item wraps to a header line over one content line with Roll back inline on the right, and the bar is 86px. At 390 dark and 320 light, with two fixes, the collapsed bar is 184 and 204px (216 and 256 before); the bar scrolls with the page there. Driven again at 1440, 1920, 390 dark and 320 light: expand, collapse, dismiss, focus, no page errors, no horizontal overflow.
Invention: none beyond the earlier ones. The cut text states no new fact; it removes two.
Round thirty-eight (2026-11-10)
Five findings from the round 37 read, each reproduced before the edit (Chromium; 390 dark touch for the stack, 1280, 1366, 1440 and 1920 for the rest) and re-run after the last one.
- D4, severity 3: a second act evicted the first act's Undo or Roll back at phone width. The stack capped at 2 below 768px (3 above) and, when every notice carried an action, removed the oldest. Two fixes left one Roll back; Suppress then Copy left no Undo. A notice that carries an Undo or Roll back now has
data-keep(set byflash({ keep: true })for Assign, Suppress and Resolved) and the cap never removes it, at any width, nor the waiting notice (its count is the queue). The cap still trims news (Copied, Refreshed, feed messages), oldest first. Past the cap the kept notices stay behind the stack's toggle, which counts every one: measured at 390, 360, 320, 1440 and 1920, two fixes give "View all 3 notifications" (both Roll backs and the waiting notice), Suppress then Copy gives "View all 3", Assign then Copy gives "View all 3"; at 390 and 1440, Suppress, Assign, Copy, Assign gives "View all 4" (the Copied notice was the one trimmed) and pressing the Suppress Undo from the open stack restores the finding (Findings 248, 248) and the toggle reads "Collapse 3". The stack is corrected, not removed: still newest whole, layers, one toggle. - Is there another way back for a suppressed finding? No. The Suppress notice's Undo is the only path: the status filter has no "Suppressed" value and the table lists only open findings. Dismissing that notice with its X (the reader's own act) or reloading the page ends it. Not changed; the brief gives no suppressed list.
- D1, severity 2: the Info link beside "Critical findings, open" was dead to a mouse. The figure's 208px line box overlapped the 24px link, and
elementFromPointreturned the figure's span..figure__rowis nowposition: relative; z-index: 1, above the figure. Five points on the link all return the link at 1280, 1366, 1440, 1920, 390, 360 and 320. - D2, severity 2: Status words split inside a word.
.status__textcarriesoverflow-wrap: anywhere, which lets the table's min-content collapse to one letter. The matrix now setsoverflow-wrap: normalon it, so a status wraps only at its spaces (the column floor is its longest word), and the matrix cell padding is 6px a side instead of 8 (20px of the table) so the table still fits its card at 1280 after a suppression. Range probe over every word in the last column: no word on two lines at 1280, 1366, 1440 (first load and after suppressing, where "threshold" and "reporting" split before), 1920. - D3, severity 2: End on the panel grip pushed the matrix 49px (1440) and 227px (1280) outside its card. The table sits in
.mxwrap(overflow-x: auto, a labelled focusable region, as the findings table's wrapper is) and the table has a floor of 34rem. With the panel at 640 the matrix scrolls inside its card; the wrapper stays inside the card at every width and each state of Home, End and Left. Card at 480 (1440) scrolls 558 inside it; at 1280 card 320 scrolls 558. No tall blank rows: the rows wrap at spaces on a 34rem table. Below the stacked breakpoint the floor is 0 and the layout is as before. - D5, severity 2: "Show it" left focus in the flashbar, 51 Tab stops from the result.
setFilter(the notice buttons, account bars, category rows, matrix counts, side-nav accounts) now moves focus to the matches line (#matches,tabindex="-1", a polite live region), which reads "1 match", after scrolling to the section. The next Tab goes to the table region, 13 stops to the first "Open details" button at 1280 to 1920 (was 51). The heading was not used: its "(248)" is the unfiltered total. The spoken line now says "1 finding." and not "1 findings."
Inventions, named
keepas the rule for what the cap cannot take, and the waiting notice joining it; "the cap trims news only" is presentation, the brief gives no number.- 6px matrix cell padding and the 34rem table floor: presentation values.
- Focus on the matches line after a filter from elsewhere: Cloudscape moves focus to the results region; the brief names no target.
Left
- A suppressed finding still has only its Undo notice to come back by (above).
- At 1280 with the panel at 640 the matrix is a 558px table in a 320px card, so it scrolls; its right columns are reached by scrolling or Tab into the region.
Round forty (2026-11-10)
Three findings from the round 39 read. D2 and D3 reproduced before the edit with a keyboard audit that presses every keyboard-reachable action and records document.activeElement (100 presses at 1280: 16 left it on BODY, among them Roll back, Undo on an assign, select-all, Copy from the Actions menu and from the panel, Assign from the panel, Dismiss of the last notice and Refresh, more than the reviewer named). D1 reproduced with a Range probe over the matrix at 49 widths and panel sizes (24 of the 49 clipped words before; at 1280 with the panel at its default, a 669px table in a 520px card).
- D2 and D3, severity 2: a keyboard act left focus on BODY. The cause was one class: the control pressed was removed, redrawn, disabled or moved, and nothing named where focus goes. Each now has a target, and a net sits under them.
- Roll back (a notice redrawn in place and moved to the top): focus rests on that notice (
tabindex="-1"), the next Tab is its Dismiss. - Dismiss, Undo, a notice's own timer (
dismiss(node, region)): the region it affected when the act has one (Undo of an assign or a suppress: the title of the first finding it brought back, else the first title, else the filter); else the next notification; else the page heading. - Select all, a row checkbox, a sort header, the pager, a filter token (
renderTablewrapsdrawTable): the same control after the redraw, found by its row, class and name. Where it is gone (a finding that left, a last page), the nearest survivor: the first finding, the current page, the filter. - The panel's Assign and Copy (
renderDockwrapsdrawDock): the button of that name after the redraw, else the panel title. Assign to me stays where the reader is. - Suppress by keyboard: the first finding left, no longer the select-all box one Space from selecting ten.
- Copy by the scratch-field fallback (
document.execCommand, used when the Clipboard API is absent): selecting the text took focus into the scratch field and removing it dropped to BODY; focus goes back to the button pressed. - Refresh: it set
disabled, which drops focus; it now setsaria-disabledandaria-busy(the Review fix loading pattern) and stays focusable. - Apply: the panel's Review fix, now kept focused through its redraw, is still not where the reader should rest;
holdFocustakes it from a busy button to the notice, as before. - The net: on any key press the control pressed and its ancestors are noted; if what the key did removed or hid that control and focus fell to BODY, focus goes to the nearest ancestor still on the page, under 1.5 s. Tested by removing the focused Assign button after a key: focus lands on its
.actsrow. - Result of the audit after the edit: 0 on BODY, 0 on a disabled or hidden control, at 1280, 1366, 1440 light, 1920 dark, 390 dark, 360 and 320 light (103 presses on a wide screen, the phone runs lack the select-all, sort, rail, panel-multi and grip controls, which a phone does not render).
- D1, severity 2: the matrix Status column was cut mid-word at 1280 with a finding open. The table had a 34rem floor and the card is 518px with the nav open and the panel at its default 400. Two changes. The account names in the head (
prod-payments) may break after their hyphen, which takes the table's narrowest width from 558px to 456px; the floor is removed. Below that, the matrix reflows to one block per measure (the phone layout) by a container query on its own card (@container mx (max-width: 30rem)), not the viewport, so a card narrowed by the panel reflows too. Probe over 1280, 1366, 1440, 1920 at panel 320, default, 480, 640, both densities, and 1100 to 320 without a panel, plus the nav closed: no word clipped, no word split except those hyphens, no scroll inside the card, no page overflow. 1280 default is a 518px table, 1280 at 480 and 640 and 1366 at 640 reflow. The card still scrolls (overflow-x: auto) if a width ever leaves the table too wide, but no width measured does, so no scroll cue was added.
Inventions, named
- Where focus rests: the page heading when the last notice goes with nothing to go back to; the notice itself after Roll back; the first finding title after Suppress. The brief names none.
- The 30rem reflow threshold and the hyphen break in the matrix head: presentation values.
- Viewports of about 520 to 767px, where the matrix's card is 480px or more, now keep the matrix as a table; the 767px media query reflowed it to blocks from 767 down. At 390 and narrower the layout is as before.
Left
- A hyphen inside a status note (
eu-west-1in the degraded notice) can still wrap at its hyphen in a narrow cell; it is a wrap, not a clip. - On a phone the table has no select-all, no sort header and no multi-select panel (the card list does not render them); the audit marks them n/a. Not changed.
Round forty-two (2026-11-10)
Three findings from the round 41 read. All three reproduced before the edit (D1 by the double-press rows, D2 and D3 by the focus audit: Info opened the panel with focus left on the link, and the Undo notice was 51 to 78 Shift+Tab stops away by the reviewer's count).
- D1, severity 2: Assign to me pressed on a finding already yours posted a second notice whose Undo changed nothing.
assign()now acts only on the findings not already yours. When nothing would change it posts nothing and says "Already assigned to you." in the status region. Undo gives back exactly what that press changed, once (a second press is ignored), and only for findings still as the press left them: a fix that has run since, or a later status, is not overwritten, and the spoken line says how many were restored. - D3, severity 2: the Undo after Suppress or Assign was only reachable by walking back through the page. After either act, focus rests on the notice that carries Undo (
restOn,tabindex="-1"on the notice, not its button, so a second Enter cannot press Undo), at every width. The next Tab is Undo. A phone still scrolls the notice into view. Suppress's confirm button disables itself on first press, so Enter twice posts one notice, and its Undo is single-use and restores only what that press suppressed. - D2, severity 2: an Info link opened the help panel and left focus on the link, with nothing announced and the panel 70 Tab stops away. Every Info link (the page, the policy limit, the measures, the Service level column header, and the drawer's Status indicators and Service levels items) goes through
openHelp(): focus moves to the panel heading, which is what a screen reader announces, and Close is the next Tab. Close and Esc return focus to the link that opened it. Esc now closes the help panel at 1280 and wider too, from inside it (the drawer, a column at that width, no longer blocks it); the detail panel beside the page still does not answer Esc.
Every undoable act and where focus goes
| Act | Focus after | Way back | Distance |
|---|---|---|---|
| Review fix, Apply (Resolved) | the Resolved notice | Roll back | Tab 1 |
| Review and approve, Apply | the Resolved notice | Roll back | Tab 1 |
| Suppress (panel, menu, ten) | the Suppressed notice | Undo | Tab 1 |
| Assign to me (panel, menu, ten) | the Assigned notice | Undo | Tab 1 |
| Roll back | the Rolled back notice (round 40) | none: it is the way back; Review fix runs it again | Dismiss, Tab 1 |
| Info link, Enter | the panel heading | Close | Tab 1 |
| Close, Esc | the Info link | 0 |
Review fix, Review and approve, Apply and Copy had no double-press hole: Apply twice posts one Resolved and one Roll back, a second Roll back press posts one Rolled back, Copy twice leaves one notice. No change.
Audit
focus-audit.mjs (copied from round 40, extended) records the Tab and Shift+Tab distance from the focus after each act to its Undo or Roll back, a double-press row per act (keyboard, double-click, from the Actions menu, Undo twice, out-of-order Undo) and each Info link's focus after open, Close and Esc. At 1280 and at 390 (touch): 0 rows on BODY, every distance 1 by Tab (the other direction is 48 to 99), no double press makes two notices, every Undo returns the rows to the state before. Page errors: none. Two audit faults were fixed on the way, not page faults: the row snapshot compared the selection count, which Suppress and a fix clear by design, and one scenario pressed the rail button with the panel already open, which closes it.
Inventions, named
- No Unassign. The brief gives no rule for what a finding's status returns to, so Undo is the only way back.
- "Already assigned to you." is spoken, not posted. A visible notice for a no-op would be a second thing to dismiss.
- Below 1280, closing the panel opened from the drawer's Status indicators or Service levels item returns focus to the menu button, because opening the panel closes the drawer. The Service level column's Info link is not drawn on a phone (the table is cards); the drawer item reaches the same panel.
- Esc at 1280 and wider closes the help panel only when focus is inside it.
Round forty-four (2026-11-10)
Two severity 2 findings and one severity 1 from the round 43 read. D2 and D1 reproduced before the edit; the stacked flashbar, the never-evict rule (data-keep), restOn and the focus audit are all kept.
- D2, severity 2: Tab onto any control in the sticky flashbar scrolled the page about 400px. Reproduced at 1280x800 by a Tab walk over the bar's controls (scroll-walk.mjs, scrollY before and after each press): after Assign, Tab to Undo 1472 to 1076 (-396), Dismiss -396 more, the toggle -364; after Suppress -385, -407, -353; after Resolved, Roll back -385, Dismiss -407, the toggle -118. Cause: round 36 added
--flash-htohtml { scroll-padding-top }so a row scrolled into view stops below the sticky bar. A control inside that bar sits inside the padding it describes, so Chromium scrolled the page to lift it out, and the sticky bar never moved, so every press scrolled again. Fix: while focus is in the bar (html:has(#flashbar:focus-within)), at 768px and wider where the bar is sticky, the padding stops at the top bar (--topnav-h). The bar is on screen, so nothing scrolls; every other target (a finding, the table header, a row, the panel) keeps the full padding. Measured after: 0px change on every press, Tab and Shift+Tab, across Undo, Roll back, Dismiss and the toggle, after Assign, Suppress, Resolved and a three-notice stack, at 1280x800 and 1440x900. At 390x844 the bar is not sticky and the padding is unchanged; the walk shows 0px there too once the phone's own smooth scroll to a new notice has settled. - D1, severity 2: a Copy after an Assign hid the Assign's Undo behind "View all 3 notifications". Reproduced: Assign, then Copy, left the Assign notice
hiddenwith the Undo 94 Shift+Tab stops away, and focus on Actions. Two changes, one rule: plain news (a notice with no Undo or Roll back) never covers one that carries them. - Stack.
syncStackshows the newest notice; when that notice is plain news it also shows the newest notice with a way back (data-keep), whole, under it. The rest stay behind the toggle, which still counts every notice; the layers sit under the last notice shown. When two notices are all there are and both show, nothing is behind, so the toggle is hidden (a button that does nothing is worse than none) and the stack is shut for the next arrival. Plain news that stands alone, or over plain news, stacks as before. Copy's 12 s timer returns the bar to the Undo notice alone. - Focus. After Copy (the Actions menu, the panel's ARN button, the panel's multi Copy), focus rests on the newest notice with an Undo or Roll back (
restOnWay, the samerestOnas Assign and Suppress), so the Undo is one Tab from where focus lands, however many plain acts came after. With nothing to undo, focus stays where it was. - With the waiting notice already dismissed, Assign then Copy shows both and no toggle: 76px at 1440, 208px at 390. Measured at 1440 and 1920 the bar after Assign then Copy is 100px with the waiting notice behind the toggle (it was 64 with the Assign alone); 252 at 390 and 272 at 320, where the bar scrolls with the page.
- Severity 1: Assign on findings already yours showed nothing to a sighted reader. It posts no notice (round 42) and the spoken line stays. A brief cue now appears beside the control pressed: an info icon and "Already assigned to you." (or "All N findings are already assigned to you.") left of Actions in the Findings head, or after Assign to me in the panel, gone after 4 s or at the next press.
aria-hidden, becausesay()has already spoken it. It is not a stacked notice and adds no Dismiss.
Audit
focus-audit.mjs (copied from round 42) now records the page's scroll between two flashbar controls in every Undo and Roll back distance row, fails a row whose Tab distance is more than 1 (it failed on the nearer of the two directions before), and waits for a phone's smooth scroll to settle first. New rows: Assign then Copy, Assign, Copy, Copy again, Suppress then Copy, Assign then Copy ARN from the panel, Suppress, Assign, Copy, Resolved then Copy, plus Assign twice from the menu and from the panel (the cue is beside the control, no notice is added, it goes after 4 s). The scroll is read on the Tab pass only: the Shift+Tab pass arrives at the bar from the far end of the page, which a phone scrolls smoothly, so it measured the animation and not a jump. Plus a row for Assign then Refresh. Result: at 1280, 179 presses, 0 bad; at 390 dark touch, 162 presses with 0 on BODY, then every distance row run again with the Tab-only scroll metric, 18 rows, 0 bad (the 390 control rows that a phone does not render are n/a, as in round 40). Every latest Undo and Roll back is at Tab distance 1 (Shift+Tab 48 to 100), page scroll between flashbar controls 0px, no page errors. Before the edit, the same Assign then Copy (1366x768) left the Undo hidden and focus on Actions.
Inventions, named
- Showing the newest notice with a way back under plain news, whole, and hiding the toggle when nothing is behind: presentation rules; the brief gives none.
- Copy moves focus to that notice. A reader who copies from the Actions menu or the panel now lands on the Undo of their last act; the audit's rule (the latest Undo is at Tab distance 0 or 1 after any later act) is what asks for it. Refresh is plain news too and does the same when focus is still on the Refresh button. A feed message (degraded, recovered) comes from the page, not from a press, and moves no focus.
- The 4 s length of the cue and its wording (the spoken line, unchanged): presentation values.
Left
- On a phone a one-line notice is 52 to 60px tall with its text pinned to the top, because the 44px Dismiss sets the height. Not changed.
- At 320 the organization id still breaks after its hyphen in the page sub-line, and Preferences and Display settings still use similar glyphs. Not changed.
Round forty-six (2026-11-10)
Two severity 2 findings and one severity 1 from the round 45 read. Both severity 2 reproduced before the edit. Round 44's flashbar fix, the never-evict rule (data-keep), the stack and the cue are kept; the focus rule of rounds 42 and 44 is replaced.
- D1, severity 2: focus inside the sticky side panel scrolled the page behind it. Reproduced at 1280x800 by
scroll-walk.mjs, now extended to Tab and Shift+Tab through the top bar, the notification bar, the side navigation, the panel and the rail, and across their boundaries: opening the first row, Tab to Close panel -406, Review fix -16, Assign -16, out to the rail -412, Help -376; the same from the pager to the rail (-665, then -355); the top bar from a page scrolled 1500px: Perimeter -439, Portfolio -439, Display settings -183; the side nav -6 and -357. 23 rows scrolled the page. Cause as in round 44, for every sticky region:scroll-padding-topholds the top bar, the notification bar and 8px, a control in a sticky region sits inside that padding, so Chromium scrolled the page to lift it out, and the region never moved. Fix: while focus is in the top bar the padding is 0, while it is in the side nav, the panel or the rail it stops at the top bar (html:has(.topnav:focus-within)and the like, beside the round 44 flashbar rule).:focus-withinis current when the browser scrolls, so the Tab that enters a region is covered; afocusinscript is not (tried, measured: the scroll comes first, and the top bar scrolled -400). Measured after: 0px on every press, both directions, at 1280x800, 1366x768 and 390x844, in each region. The one scroll that remains is the right one: Tab from the last pager button to the rail at the foot of the page moves the page up (-626 at 1280) because the rail is pushed out of view by the end of its column there; the Help button after it, and every control after, 0px. That is "what brings the newly focused control into view". The walk judges a row bad when the page scrolls while focus stays in one region, or crosses a boundary and scrolls although the control was already fully visible before the key (its position read from its region's top before the key). - D2, severity 2: every act sent focus to the notice at the top, so the next act was 37 to 81 Tab stops away, and Copy after Assign moved focus off Copy. One rule replaces rounds 42 and 44:
- When the control pressed still exists, focus stays on it. Notifications are announced (the notification bar is
aria-live) and take no focus. This covers Assign (panel, panel with ten, Actions menu), Copy (the ARN button, the Actions menu) and Refresh. - Each undoable act that leaves its control also shows its Undo beside it, Tab 1 from where focus is. Assign from the panel: "Assigned to you." (a success status) and an Undo button directly after Assign to me,
aria-label"Undo assignment" or "Undo assignment of N findings". Assign from the Actions menu: the same line, with Undo, directly after the Actions button, on its own line at the end of the Findings head. It lives as long as its notification: Undo (either one), or Dismiss of the notification, removes it; a later Assign replaces it; the panel draws it again after every redraw for as long as the panel shows the finding or findings that act was about. - The notification keeps its own Undo. Both do one single-use act (
undoIt): whichever is pressed first restores the findings, the other finds it used. Pressing the older act's notification Undo leaves a newer act's inline Undo where it is. - Where focus goes after an Undo: after the inline Undo, back to the control that did the act (Assign to me, Actions: distance 0); after the notification's Undo, the first finding it brought back, as before (the button pressed is gone).
- Only when the control pressed is gone does focus go to the notice or the nearest survivor, as before: Suppress (its rows and its dialog go) and Apply (its dialog goes) rest on their notice, the next Tab is Undo or Roll back.
- A phone no longer scrolls to the Assign notice and no longer takes focus to it: the sheet shows "Assigned to you." and Undo beside the button, 44px high. Copy from the Actions menu still scrolls its notice into view there, without focus. Suppress and Resolved still bring their notice in and rest on it.
restOnWay(Copy and Refresh moving focus to the newest Undo) is deleted.- Severity 1: pagination. A page change is now spoken through the existing status region: "Page 2 of 25. Showing 11 to 20 of 248." Focus stays on the pager button (round 40).
Every undoable act and where focus goes
| Act | Control pressed | Focus after | Way back | Tab distance |
|---|---|---|---|---|
| Assign to me (panel, one or ten) | stays | Assign to me | Undo after it, and the notice's | 1 (back to the control: 0) |
| Assign to me (Actions menu) | stays (Actions) | Actions | Undo after it, and the notice's | 1 (0) |
| Copy (ARN button, Actions menu) | stays | the control pressed | none to give | (0) |
| Refresh | stays | Refresh | none to give | (0) |
| Page change | stays | the pager button | none to give | (0) |
| Suppress (panel, menu, ten) | gone | the Suppressed notice | Undo | 1 |
| Review fix, Review and approve, Apply | gone | the Resolved notice | Roll back | 1 |
| Roll back | gone | the Rolled back notice (round 40) | Review fix runs it again | Dismiss, 1 |
| Undo (inline) | gone | the control that did the act | none | 0 |
| Undo (notice), Dismiss | gone | the finding it brought back, else the next notification, else the page heading | none | |
| Info link, Enter | stays | the panel heading | Close | 1 |
| Close, Esc | the Info link | 0 |
After a later plain act the Undo is where it was: Assign then Copy from the Actions menu leaves focus on Actions with the Undo still one Tab on; Assign then Copy ARN in the panel leaves focus on Copy and the Undo two Shift+Tab back (beside Assign) or in the notification. Suppress then Copy: focus on Actions, the Suppress Undo in its notification only (57 Tab stops), because Suppress left no control to carry one.
Audit
focus-audit.mjs (copied from round 44) now measures, after each act, two things. The Tab distance back to the control pressed: expected 0 (the focus after the act is that control; a row that left it is bad), for Assign (panel, ten in the panel, Actions menu), Copy after Assign (menu, ARN in the panel), Copy twice, Assign then Refresh, Suppress then Copy, Resolved then Copy. The Tab distance to the nearest Undo or Roll back, in a notification or beside the control: expected 1 (0 or 1) after Assign, Suppress and Resolved, and after Copy where the Undo's control is still the one holding focus; also the distance to the notification's own button, which must exist. New rows: the inline line and Undo sit directly after the button; Tab 1 is the Undo; inline Undo returns focus to the control and rows to the state before; notification Undo removes the inline one; Actions Assign, same; Copy and Copy again keep focus; Refresh keeps focus; both Undos together restore once; an older act's Undo leaves the newer inline Undo; Next page speaks the page and keeps focus. Result at 1280 (light): 197 presses, 0 on BODY, 0 bad. At 390 dark touch the first run was 182 presses with 5 bad rows, all one fault the audit found: on a phone the Actions menu Copy still sent focus to its notice (reveal takes focus). Copy from a menu now only scrolls the notice into view there (reveal(node, true), focus stays on Actions); the Copy rows were run again at 390 and 1280: 0 bad. Scroll walk at 1280x800, 1366x768, 390x844 and 320x844 touch: 0 bad rows, the only moves being a wrap from the last control to the top and the rail entry above. Page errors: none.
Inventions, named
- The inline "Assigned to you." line and its Undo, its wording (the notification's own head), its placement (after the button; the Head's on its own line at the end), and that it lives with its notification: presentation rules; the brief gives none.
- Focus stays on the control for Assign, Copy and Refresh at every width, so a phone no longer scrolls to the Assign notice. Cloudscape practice, not a brief rule.
- A reader whose Undo has no control to sit beside (Suppress, Resolved) reaches it in the notification, one Tab from the notice that has focus; there is no inline Roll back, because Review fix and Apply leave no control to put it beside.
Left
- The page's Tab order, 45 stops to the end of the matrix and 55 to the first row checkbox, is unchanged (severity 1 in the round 45 read).
- The notification bar wrapping to three lines at 1280 with the panel open is unchanged.
- A phone's Copy after Assign with the sheet open still shows its "Copied" notice at the top of the page under the top bar; only the icon swap on the Copy button confirms it. Unchanged.
Round forty-eight (2026-11-10)
One severity 2 and three severity 1 findings from the round 47 read. All four reproduced before the edit (1280x800, keyboard): root and three Enters gave three "root" chips; prod-pay, ArrowDown, Enter, Enter gave Account = prod-payments and a stray "prod-pay"; Tab from Cancel in Review fix with Apply disabled landed on BODY; Esc on the grip of the help panel closed it (panel width 0); Tab from Next page to the rail moved the page from 2202 to 1576 with the rail 226px above the viewport. Rounds 42 to 46 (focus rests, the flashbar, the scroll-padding-top regions) are kept.
- D1, severity 2: Enter in an empty filter box added the last text again. Cause as the reviewer read it:
closePF()hid the list but leftPF.items, and Enter picked the stale "Use: ..." option. NowclosePF()clearsPF.items; Enter with nothing highlighted and an empty box does nothing; Enter with text and a closed list (after Esc) builds its options fresh from the box.addTokrefuses a token identical to one already shown (same property, operator and value) and says "That filter is already applied." in the status region; it still empties the box and closes the list. A highlighted option is still picked by Enter on an empty box (ArrowDown on an empty box lists the properties). - Severity 1: Tab left a modal while its button was disabled. One rule for every modal, in
wireDlg: Tab on the last reachable control goes to the first, Shift+Tab on the first goes to the last; disabled, hidden and unchecked radio controls are skipped. It covers Review fix (Apply is disabled for about 1 s, so Cancel was the only stop), Suppress (disabled until the phrase is typed) and Preferences. The buttons staydisabled: a button that cannot be pressed is not offered as one, and the trap, not aria-disabled, keeps focus in. - Severity 1: Esc on the panel grip closed the panel. The grip now handles Esc: it cancels a pointer drag in progress (width back to where the drag began), else does nothing, and
preventDefaults so the page's Esc handler leaves the panel open. Esc still closes the help panel from its heading, Close and its body, and the sheet below 1280. - Severity 1: Tab from the last pager button to the rail scrolled the page 626px. The rail was a sticky column that ended with the table, so by the pager it had scrolled off the top and Tab brought it back. The column now stretches the full height of the frame and its two buttons sit in a sticky wrapper (
.rail__stick), so they stay under the top bar while any of the table is on screen. Measured at 1280x800 and 1440x900: pager to rail 0px, rail to Help 0px, and back 0px. The one visible change is that the rail's background and divider now run the full height of the frame instead of one viewport.
Filter test table (filter-test.mjs, 1280x800, keyboard unless stated)
| Path | Chips after |
|---|---|
root, Enter | "root" |
root, Enter x3 | "root" (was three) |
prod-pay, ArrowDown, Enter, Enter | Account = prod-payments (was plus "prod-pay") |
Account = prod-pay, ArrowDown, Enter, Enter | Account = prod-payments |
| empty box, Enter x2 | none |
| ArrowDown, Enter (property), ArrowDown, Enter (operator), ArrowDown, Enter (value) | Severity = Critical |
root Enter, root Enter | "root" once, "That filter is already applied." spoken |
prod-pay, ArrowDown, Enter, done twice | Account = prod-payments once |
root Enter, prod-pay ArrowDown Enter | both |
root, Esc (list closes, text stays), Enter x2 | "root" once |
root, Esc, Esc (clears the box), Enter | none |
| Backspace on an empty box, then Enter | last chip removed, nothing added |
| Backspace x2 with two chips | none |
| ArrowDown on an empty box, Enter | the box reads "Severity ", no chip |
root, ArrowUp, Enter | "root" |
root Enter, Tab away, Shift+Tab back, Enter | "root" |
pointer: type root, press the option; click the box, Enter x2; type root, press the option again; remove the chip | one chip, one chip, one chip, none |
20 rows, 0 bad.
Audit
focus-audit.mjs and scroll-walk.mjs (copied from round 46) re-run after the last edit. Focus audit at 1280: 197 presses, 0 bad; at 390 dark touch: 182 presses, 0 bad; no page errors. Scroll walk at 1280x800, 1366x768, 390x844 and 320x844 touch: 0 bad rows; pager to rail is 0px at both desktop sizes. dialog-test.mjs: 8 Tab and Shift+Tab presses in Review fix with Apply disabled stay on Cancel; with Apply enabled they alternate Cancel and Apply; Suppress with its button disabled and enabled, and Preferences (20 controls), never leave the dialog; PageDown with Apply disabled scrolls nothing (1710 before and after).
Inventions, named
- "That filter is already applied." is spoken only. A visible cue would be one more thing to dismiss; the empty box and the closed list already show that nothing was added.
- A Tab that wraps inside a modal rather than leaving to the browser: Cloudscape's modal does the same. No visible change.
- The rail column spans the full height of the frame.
Left
- With a panel open, Tab from the last pager button goes to the panel's grip first, and the page still moves 250 to 330px, because the panel is a sticky column that also ends with the table and is partly off screen at that point. Not the rail path the reviewer named; unchanged.
- With the panel open at 1280 the table still keeps 558 of 1127px and the notification bar is still three lines (severity 1 in the round 47 read).
Round fifty (2026-11-10)
Four severity 1 findings from the round 49 read (ship note: the rail). All four reproduced before the edit at 1280x800 and 390x844 touch (repro.mjs). Rounds 42 to 48 (focus rests, the flashbar, the scroll-padding-top regions, the filter and modal rules) are kept.
- Severity 1, the ship note: the rail is
role="toolbar"and its keys scrolled the page. Reproduced: ArrowDown on the Details button moved scrollY 0 to 40 to 79, Home to 1, End to 2573, focus never left the button.wireRail()now implements the vertical toolbar pattern: one Tab stop (rovingtabindex, the button last focused holds it, Details at load), ArrowDown and ArrowUp move between the two buttons and wrap, Home and End jump, and each callspreventDefault, so the page never scrolls. Enter and Space still press the button. Tab into the rail lands on the roving button; Tab out leaves in one press (it was two). Other composite roles checked for the same gap: the Actions menu (role="menu": ArrowDown, ArrowUp, Home, End, Escape, no scroll, as before), the filter list (role="listbox", driven from the input byaria-activedescendant: no scroll), no tablist, grid or radiogroup (the radios in Preferences are native inputs). The chart (role="application") and the panel grip (role="separator") are single widgets, not composites; the chart keeps Left, Right, Home and End and lets Up, Down, PageUp and PageDown scroll the page, as before. - Severity 1: the property filter left
aria-activedescendanton an option that was not highlighted. Reproduced: after choosing a property or an operator, or typing, the input namedpf-o-0while every option readaria-selected="false".refreshPF()draws the list with nothing highlighted, so it now removes the attribute;movePF()sets it, andclosePF()already removed it. The attribute names the highlighted option only: after a pick it is absent, after the next ArrowDown it ispf-o-0witharia-selected="true". - Severity 1: Tab onto the table region scrolled the page. Reproduced at 1280x800 (the region is a tab stop 777px tall, its top 505px down a 800px view): Tab moved scrollY 1361 to 1741 (1361 to 1777 with a filter), the browser aligning the region top under the top bar although its head and rows were on screen (the reviewer measured 679px). Cause: the browser scrolls the whole focus target into view and a region taller than the view cannot fit.
watchTable()sets a negativescroll-margin-bottomon the region (its height less 160px, kept current by aResizeObserveras rows change), so the area the browser tries to show is the first 160px. Measured after: a Tab from the controls above moves the page 0px at 1280x800, 1440, 1920 and at 390 and 320 (where the region is the whole list of cards); a region fully below or above the view still scrolls to show its head (1280x800: scrollY 0 to 1484 and 2641 to 1484). - Severity 1: on a phone, Apply left the Applying notice screens above the reader. Reproduced at 390 and 320: Apply from the sheet (tap and keyboard) scrolled 3854px up to the notice by smooth scroll, so it was off screen (1,000px or more above) for the first second or more of the run, then arrived. The sheet already shows "Applying change" and
say()already announces "Applying change set cs-...", both unchanged. The notice now jumps into view at once (reveal(node, scrollOnly, now),nowset for a notice withloading), so it is at 56px under the top bar 120ms after Apply, above the sheet, with focus on it as before. The Resolved notice still scrolls smoothly (it is 8px away by then). I could not reproduce a 3s absence; I measured about 1s of smooth scroll, which is what the change removes.
Rail key test (composites.mjs, 1280x800 and 1440x800, keyboard)
| Start | Keys | Focus | Page scroll |
|---|---|---|---|
| Details | ArrowDown, ArrowDown, ArrowUp, ArrowUp, End, Home, End, ArrowDown | Help, Details, Help, Details, Help, Details, Help, Details | 0px on all 8 |
| Help (pressed with Enter) | Enter | the panel opens, aria-pressed true | |
| Help | Shift+Tab, Tab | the panel Close button, Help |
Roving tabindex is 0 on the focused button and -1 on the other after every key.
Audit
focus-audit.mjs and scroll-walk.mjs (copied from round 48) re-run after the last edit. Focus audit at 1280: 197 presses, 0 bad; at 390 dark touch: 182 presses, 0 bad; no page errors. Scroll walk at 1280x800, 1366x768, 390x844 and 320x844 touch: 0 bad rows. tabtbl.mjs (Tab onto the table region) and phone-apply.mjs (the notice position every 400ms after Apply, five findings) are new. Screenshots at 1440 light, 1280 light, 1920 dark, 390 dark, 360 light and 320 light: no page errors, scrollWidth equals clientWidth in every state.
Inventions, named
- The rail wraps from the last button to the first and back. The toolbar pattern leaves wrapping open; with two buttons it is the shorter path.
- The 160px that the table region asks the browser to reveal: a presentation value, enough for the top border and the column headings.
- The Applying notice jumps rather than glides. A motion rule, not a brief rule: an in-progress state is shown when it starts.
Left
- Chart ArrowUp, ArrowDown, PageUp and PageDown and the grip's ArrowUp and ArrowDown scroll the page: neither is a composite role and neither uses those keys. Unchanged.
- The sticky side navigation under the top bar on an empty filter at 1366x768 and the 104px notice with the panel open (the other two severity 1 items in the round 49 read) are not in this round's brief and are unchanged.
Round fifty-two (2026-11-10)
Two severity 2 findings and one severity 1 from the round 51 read (ship note: the typed phrase). Both severity 2 reproduced before the edit. Rounds 42 to 50 (focus rests, the flashbar, the scroll-padding-top regions, the filter, modal and rail rules) are kept.
- D1, severity 2: the typed-confirmation input sat behind the dialog's sticky footer on a short phone. Reproduced at 667x375 (
dlg-rects.mjs, every focusable control of every dialog against the footer's top): the Suppress input at 296 to 328 against a footer top of 288, so it was entirely covered while it held focus. The same walk found a second case the reviewer did not name: Preferences,pr-striped, 286 to 302 against 288, half covered when Tab reached it. The reviewer's 320x568 figures (416 to 448 against 429) were not measured before the edit: that run was cut off by its time limit after 667x375 and 568x320, so only the after is recorded. Cause: below 768px the footer is sticky (round 40), the dialog is its own scroll container, and the browser scrolled a focused field only as far as the dialog's edge, not the footer's. Fix:scroll-padding-bottomon the dialog equal to the footer's height plus 8px (--foot-h), inside the samemax-width: 767pxrule that makes the footer sticky.padFoot()sets--foot-hwhen a dialog opens (before the focus that follows opening, which is the one that mattered: a ResizeObserver alone ran after it) andwireDlgkeeps it current with a ResizeObserver, since the footer wraps at 320. While focus is in the footer the padding is 0 (dialog:has(.dlg__foot:focus-within)), because its buttons are already in view and the padding would scroll the dialog to lift them out, the round 44 fault in a new place. Measured after, every focused control against the footer top: Suppress input 216 to 248 against 264 (667x375), 161 to 193 against 209 (568x320), 389 to 421 against 429 (320x568), 449 to 481 against 497 (360x640), 521 to 553 against 569 (390x844), 280 to 312 against 328 (844x390, where the footer is not sticky). Preferencespr-striped141 to 157 against 288 at 667x375. Review fix showed only its two footer buttons as focus stops, both clear. 0 bad rows in 186 focus positions (Tab and Shift+Tab, 93 at each of two halves of the sizes) over the three dialogs and the popover and six viewports, touch contexts. - Display settings, found on the way: it is a popover, not a dialog, and has no footer, but it is 318px tall and ran 6px off a 320px-high screen (top 8, bottom 326). It now has
max-height: calc(100dvh - 16px)and scrolls inside that; its radio rows measure inside the viewport at all six sizes. - D2, severity 2: a phone could not sort or select all. Reproduced at 390x844:
#theadisdisplay: nonebelow 768px (the table is cards), so the eight sort buttons and the select-all box measured 0x0. A block now sits above the cards on a phone (#phonesort, hidden from 768px up so a screen reader never meets two sets): a "Select all 10 on this page" checkbox with "N selected" at its right, a "Sort by" select of the visible columns and an "Order" select (Ascending, Descending). They are the same state as the header: the box sets and clearsS.selfor the page (indeterminate when some are ticked, as the header's is), the selects setS.sortand reset the page to 1, and the hidden header'saria-sort, the pager and the count follow from the one redraw (syncPhoneSort). The controls are static, so a redraw does not take focus from them. The change is spoken ("Sorted by Opened (UTC), descending." / "Selected 10 findings on this page.") through the status region. A column hidden in Preferences that is still the sort key stays in the list. Below 360px the two selects stack (the card is 254px wide at 320 and the two need 256px). 44px targets on a coarse pointer, in the.seland.phonesort__checkadditions to the existing floor..selis new: the page had no select style, so it takes the text input's tokens. - Severity 1: Enter on an unmatched property-filter value was silent. Enter now says why through the status region, and the box keeps its text: "Nothing added. No account matches “nosuch”." (the list's own hint), "Nothing added. Type part of a finding." (an operator with no text), and "Nothing added. Choose a suggestion with the arrow keys, then press Enter." (a partial value with several matches and none highlighted). A free word still becomes a text filter, as before.
Phone sort test (phonesort.mjs)
For each of the eight columns, both directions, at 390, 360 and 320, once with selectOption (the touch path) and once by keyboard (focus the select, ArrowDown or ArrowUp to the value): the first page's order equals the order the desktop header gives at 1280 for the same column and direction, the header's aria-sort reads the same, and focus stays on the sort control. Select all by tap and by Space: 10 rows ticked, the header box ticked, "10 selected"; one row unticked gives indeterminate and "9 selected"; select all again ticks all ten, again clears. Paging keeps the sort, and select all follows the page. 446 checks, 0 bad, no page errors.
Inventions, named
- The phone's block is one of Cloudscape's own patterns by shape (a select and a direction beside a collection) but the brief gives no phone sort rule; the two-select form is the smallest that is reachable by touch and keyboard with a native control.
- The sort key keeps its direction when the column changes on a phone (the header resets to ascending on a new column); with two selects, the reader's chosen direction is the less surprising state.
- The 8px added to the footer height in the dialog's scroll padding: a presentation value, so the focus ring is not cut by the footer.
- The popover's
max-heightis one viewport less 16px, the 8px marginplacePopalready keeps.
Audit
focus-audit.mjs and scroll-walk.mjs (copied from round 50) re-run after the last edit: focus audit at 1280: 197 presses, 0 bad; at 390 dark touch: 182 presses, 0 bad; scroll walk at 1280x800 and 390x844 touch: 0 bad rows; page errors none (the first 1280 run had parts 4 to 6 crash the page on load, under memory pressure, and they were run again clean). Screenshots at 1440 light, 1280 light, 1920 dark, 390 dark, 360 light, 320 light, 667x375 and 568x320: scrollWidth equals clientWidth in every state. New: dlg-rects.mjs, phonesort.mjs, pf-popover.mjs, selfit.mjs (does the widest option fit each select at 320 to 767).
Round fifty-four (2026-11-10)
One severity 2 and two severity 1 findings from the round 53 read (ship note: dismissing a notification from mid-page). The severity 2 and the chart reproduced before the edit; the second Assign did not reproduce as silent. Rounds 40 to 52 (focus rests, the flashbar, the scroll-padding-top regions, the filter, modal and rail rules) are kept.
- D1, severity 2: dismissing a notice from mid-page threw the reader to the top. Reproduced at 1280x720 (
dismiss.mjs): scrollY 1615 (#findings), press the X on "One change set is waiting for approval" by mouse or by Enter, scrollY 0 and focus on the h1; the same from the chart (294), the table middle (1918) and the pager (1975). Cause as the reviewer read it: round 40'sdismiss()fell back to region, else next notice, else$("h1"), andfocus()scrolls to its target. New rule indismiss(node, region): dismissing a notice never scrolls. The next notice when one remains and is in view (the bar is sticky from 768px, so it is); elseviewFocus(), the first heading or finding title fully on screen under the bars, else the first control;focus({ preventScroll: true })in every case, for a pointer press as for a key (focus is never left on the removed node).inView()counts the sticky bar's bottom as the top of the visible content. - What scrollY does. The code no longer moves the page. But the sticky bar holds its own space in the flow, so when the last notice goes the document shrinks by the bar's height (34 to 62px on a desktop, 66 to 140px on a phone), and Chromium's scroll anchoring lowers scrollY by that amount to keep the reader's content where it is: 1615 to 1553 at 1280x720, with the focused heading at the same pixel row (shift 0px) before and after. Literally unchanged scrollY would move the content up by the bar's height at the 250ms removal, which is the jump this round removes, so the metric the matrix holds to is the focused element's position in the viewport; scrollY is printed beside it. On a phone the notice bar scrolls away with the page, so the X is off screen mid-page; the matrix focuses it without scrolling and presses it, and the same holds.
- Kept. A notice's own Undo still passes
region(the first finding it brought back) and so still scrolls to it, as rounds 40 to 46 decided; that is an act with a destination, not a dismissal. Only Dismiss and the notice's timer take the no-scroll path. - Severity 1: the dotted cursor line ran through the "Over from 8 Sep" annotation. Reproduced at 390 dark (cursor on 10 to 12 Sep) and at 1280 (cursor on 11 and 12 Sep, "Crossed the limit on 8 Sep"). The note already had a surface-coloured halo, but the cursor is drawn above it and a halo does not fill the gaps between words. After the last text is chosen, the cursor line gets a clip path: the whole chart less a 3px-padded box around each
.note, evenodd (cur-clip). The line stops short of the words and resumes after them; the series, the grid and the rings behind a note are untouched, which a surface-coloured rectangle would have hidden. No raw colour. The degraded notes ("No reading on 11 Sep.", the stopped line) get the same box. - Severity 1: a second Assign to me reported silent. Not reproduced as missing. Pressed twice by mouse, by key at gaps of 0, 300 and 1500ms, from the panel button and from the Actions menu item, at 1280x800 and 390x844 touch: the cue "Already assigned to you." appears in both places (
.actsbeside the button,.container__headbefore Actions), in view, for 4s, and#sr(polite, role status) says the same words; focus stays on the control. Nothing edited. Two things the reviewer may have read as silence: no notice is posted (round 42), and in the panel the cue stands beside the first press's "Assigned to you. Undo", small and in the info colour.
Dismiss matrix (dismiss.mjs)
Eight notice types (waiting, waiting then Assign so one remains, Assign, Copy, Refresh, Suppress, the feed warning, Resolved) at five scroll positions (top, chart, findings, table middle, pager), by mouse and by keyboard: 80 rows at each of 1280x720, 1440x900 and 390x844 dark touch. A row is bad when focus is on BODY or out of view, or the focused element moves more than 1px in the viewport (except from the top of the page, where the bar leaving moves the page up by itself). 0 bad at all three; the first run before the edit at 1280x720 gave 8 bad of 10 for the waiting notice alone (1615 to 0 by mouse and key).
Inventions, named
- The in-view order: the next notice, else a heading or finding title, else any control. The brief names none.
- The 3px padding of the cursor's cut-out.
- Scroll anchoring is left on (see above).
Left
- (Superseded by the last item of "Chart labels" below.)
- The focus ring on the Findings heading, where focus now rests after a mid-page dismissal, spans the heading's full width, as the h1's does.
Chart labels stay inside the chart box (coordinator's addition)
Found on the way, then asked for: the bold end label "21 critical" ran 6px past the chart box's right edge at 390, 360 and 320 (box 349, text 355 at 390). Two measured fixes in drawChart, no new text:
- End labels. After the svg is in the page, the widest
.endlabelis measured withgetBBox. If it ends pastW - 2, the right margin grows by the overhang (padR) and the chart is drawn once more. The 64px and 80px margins stay the starting point; only a label that would leave the box moves anything. - Annotations. Each
.noteis measured after the short-text swap; one that starts left of 2px is moved to start at 2, one that ends pastW - 2is moved to end there. This was needed for the stopped-feed line "eu-west-1 stopped at 13:41 UTC." at 320, which ran 11px past the left edge. The cursor's clip box (above) is built after this, from the final positions. - Check (
labels.mjs): every<text>in the chart (value ticks, day ticks, threshold label, end labels, annotations) against the host's rect, on all four sides, at 320, 360, 390, 768, 1280, 1440 and 1920, light and dark, with the feed degraded, and with the detail panel open at 1280, 1440 and 1920. Before the second fix: worst overhang 11px (320, degraded). After: worst -2px (every label inside the box; 2px is the clamp). Before the first fix the default chart's "21 critical" was +6px at 390, 360 and 320. - At 320 in the stopped-feed state the two notes sit over the series line near the bottom left (a halo keeps them legible, as before for the first note); they are inside the box, which is what was asked.
Audit
focus-audit.mjs, scroll-walk.mjs and dlg-rects.mjs (copied from round 52) re-run after the last edit: focus audit at 1280: 197 presses, 0 bad; at 390 dark touch: 182 presses, 0 bad; scroll walk at 1280x800 and 390x844 touch: 0 bad rows; dlg-rects.mjs over six sizes: 186 positions, 0 bad; page errors none. The chart-label edit came after those runs; it touches only drawChart, so after it I re-ran labels.mjs, states.mjs (320, 390, 1280x720) and check.sh, not the focus audit or the walks. New: dismiss.mjs, chart.mjs, cue.mjs, cue2.mjs, states.mjs, labels.mjs, shot320.mjs. Screenshots at 1440 light, 1280x720 light, 1920 dark, 390 dark, 360 light and 320 light (chart with the cursor on the annotation, the page after a mid-page dismissal): no page errors, scrollWidth equals clientWidth in every state.
Round fifty-six (2026-11-10)
Two severity 2 findings and two severity 1 from the round 55 read (ship reservation: the filter's free text). All four reproduced before the edit (d1.mjs, misc.mjs, filter-table.mjs; 1280x720, 1366x768, 1440x900). Rounds 40 to 54 (focus rests, the flashbar, the scroll-padding-top regions, the filter, modal and rail rules, the dismiss rule, the chart labels) are kept.
- D2, severity 2: the property filter's free text ignored the Severity, Service level and Opened columns. Reproduced as reported:
criticalgave 2 matches against 21 critical,high4 against 47,past by,due inandsepgave the empty state,eu-west-1gave none. Cause as the reviewer read it:matchTokbuilt the free-text haystack from title, resource, category, control, assignee, account and status only. It now readsrowText(r): the severity word, title, resource, account, status label, the service level wording (sla(r).text, "Past by 2 h", "Due in 19 h"), the opened date as drawn (fmtDT, "11 Sep, 09:14"), assignee ("Unassigned" when none, as the cell says), region, category, control and finding id, lower case, with the narrow no-break space of "2 h" and any run of white space read as one space, on both the row and the query, so "past by 2 h" finds the cell that reads "Past by 2 h". Property tokens (Severity = Critical and the rest) are unchanged. - D1, severity 2: opening a finding, or dragging the panel wider, pushed the opened row off the screen. Reproduced at 1440x900, 1366x768 and 1280x720: the panel takes 400px from the main column, the cards and the chart above the table wrap again, the page grows about 1,000px on open and 1,900 to 2,500px at the widest, and the row ended hundreds of pixels below the screen (row 10 at 1440x900: 464 to 1735 as the reviewer read it, 464 to 1117 in my first run; End on the grip at 1280x720: row 1 298 to 1691). Two causes, both fixed: drawing the table again moves the page before the panel opens (scroll 2181 to 1528 in one click handler), and focus on the panel's title scrolled the page 460px after a first correction. Fix:
anchorNow()records where the row is on screen before anything is redrawn;holdView()puts it back withscrollByas the layout settles (aResizeObserveron the body, the end ofdrawChart, which changes the height as the chart is drawn again, and 40 animation frames) and ends at once on a wheel, a touch, a key or a press anywhere but the grip. For a drag, the anchor is taken at the start and held to the end; for End, Home and the arrow keys on the grip, taken at each press. The click handler andopenHelpfocus the panel title withpreventScrollwhile a hold is on, so focus moves exactly where it did and the page does not. For the help panel the anchor is the Info link pressed. Below 1280px the panel is a sheet over the page, nothing moves, and no anchor is taken. - One limit. The panel is sticky but ends with the frame, so a row near the end of the table held exactly where it was leaves the panel's head, Close included, above the top bar (row 10 at 1280x800: panel top -165). Found by the scroll walk, not by the reviewer. In that case the row is held lower, by the least that puts the panel's head under the top bar, and never so low that the row is not whole on the screen (
holdFix). Rows 1 and 5 stay exactly where they were. - Severity 1: Tab onto the measures matrix scrolled the page. Reproduced at 1280x720 (region 662px tall, head 194 to 126, scrollY 844 to 912; 8px at 1366x768; none at 1440 and 1920 where the region fits).
watchTable()now trims the matrix region'sscroll-margin-bottomas it does the findings table's, kept current by aResizeObserver. Measured after: 0px at all four sizes; a Tab or Shift+Tab from a screen away still brings its head into view (1280x720: head at 343 from above, 126 from below), as the table's does. - Severity 1: Escape on the help panel's grip did nothing. Reproduced at 1280 to 1920. The grip's Escape (round 48) cancels a drag in progress first; with no drag it now closes the help panel and returns focus to the Info link or the rail button that opened it, as Escape does from the heading, Close and the body. On the finding panel the grip's Escape still does nothing, because that panel answers no Escape beside the page.
Free-text table (filter-table.mjs, 1280 and 1440, every column on, 50 per page, 248 rows read from the rendered cells)
For each word the filter count, typed into the box and entered with nothing highlighted, against the count of rendered rows whose text contains it (both read from the page; neither calls matchTok):
| Column | Query | Filter | Rows with it |
|---|---|---|---|
| Severity | critical | 21 | 21 |
| Severity | high | 47 | 47 |
| Severity | medium | 108 | 108 |
| Severity | low | 87 | 87 |
| Service level | past by | 188 | 188 |
| Service level | due in | 60 | 60 |
| Service level | due in 1 | 21 | 21 |
| Opened | sep | 64 | 64 |
| Opened | 8 sep | 12 | 12 |
| Status | awaiting approval | 1 | 1 |
| Assignee | unassigned | 72 | 72 |
| Region | eu-west-1 | 44 | 44 |
| Finding | bucket | 48 | 48 |
| Account | prod-platform | 78 | 78 |
| first row, every cell | 14 words | equal | equal |
26 queries, 0 bad. Against the page: the headline reads 21 critical and 47 high, and critical and high give 21 and 47; Past by 188 and Due in 60 make the 248.
Row position table (d1.mjs, row top in px, before to after, the row's own Enter and click; End and a pointer drag to 640px)
| Viewport | Row | Open | End | Drag |
|---|---|---|---|---|
| 1440x900 | 1, 5 | 276, 584 (same) | same | same |
| 1440x900 | 10 | 464 to 738 (whole, panel head 48) | 738 | 738 |
| 1366x768 | 1, 5 | 276, 584 (same) | same | same |
| 1366x768 | 10 | 398 to 606 | 606 | 606 |
| 1280x720 | 1, 5 | 298, 606 (same) | same | same |
| 1280x720 | 10 | 385 to 558 | 558 | 558 |
36 cases: before, 36 of 36 cases moved (64px for rows 1 and 5 on open, up to 1,330px on resize) and 26 left the row not whole on the screen; after, 0 not whole on the screen, 0 page errors, 6 moved (row 10, the limit above). The panel's top is at 48 (under the top bar) in all 36. Also: Info link opening help, 200 to 200; a second row clicked with the panel open, 529 to 529 and 507 to 507; Esc during a drag restores 400 and the row; a wheel during a hold ends it.
Inventions, named
- Free text reads every column's text, hidden-by-default columns included (Region, Category, Control, Finding ID), so what a reader types is not decided by their Preferences; the old haystack already held category and control. "Unassigned" is searched because the cell shows it.
- The hold lasts 40 frames or until the reader acts; about 650ms is a presentation value. The row is held lower, not exactly, when the panel's head would otherwise be hidden: the brief's "or at least fully in view" is the rule.
- Escape on the help panel's grip closes it; Escape on the finding panel's grip does nothing, as before.
Audit
focus-audit.mjs, scroll-walk.mjs and dlg-rects.mjs (copied from round 54) re-run after the last edit. The first run of the scroll walk, before the panel-head limit above, had one bad row (open the last row, Tab to Close panel: the page scrolled 559px, the panel's head being above the top bar); that is how the limit was found. After it: focus audit at 1280: 197 presses, 0 bad; at 390 dark touch: 182 presses, 0 bad; scroll walk at 1280x800, 1280x720 and 390x844 touch: 0 bad rows; dlg-rects.mjs over six sizes: 186 positions, 0 bad; page errors none. New: d1.mjs (row position), filter-table.mjs, misc.mjs (matrix Tab, grip Escape), extras.mjs, extras2.mjs (a region off screen still scrolls to its head), dockgeo.mjs, shots.mjs. Screenshots at 1440 light, 390 dark, 360 light, 1920 dark, 1280x720 light and 320 light (the free-text filter, the 10th finding opened, the panel dragged to its widest, the help panel): no page errors, scrollWidth equals innerWidth in every state.