Brief #20. The house system: the firm's own internal console: the pipeline, the work list, the week and the ninety days behind them.
The four direction lines
| Lever | Decision |
|---|---|
| Archetype | Shell. A collapsible rail against one work area. The rail is 232px expanded, 56px collapsed, and a tab bar under 600px. |
| Data stance | Table-first. The pipeline table is the product; the stat strip, the sparklines, the bar list and the twelve-week line chart are all drawn to the table's grammar rather than a charting library's. |
| Density | The console profile: 11 / 12 / 13 / 14 px, 28px controls, 32px rows, 10×6 cell padding. |
| The one bold thing | The four-step reading ladder used as intended, with datum (13px) reserved for the size a value is read at, and absence rendered as three visibly different marks. |
Why this product suits this system: Centauri was written for exactly this screen. Its premise is reading numbers, its first-class object is the dense table that survives a phone, and the four-step ladder exists so the size a value is read at has a name. A pipeline of thirty-eight rows read at a desk and again at a kerb is the case the system was measured against.
The five declarations
- The ground:
surface. The cool near-white of a drawing sheet. Rules divide it; a box appears only around a genuine container (the scrolling table box and the code block are the only two on the page). - The hue budget: three, the firm's own. A navy neutral carries the ground, the rule, the edge and all three text steps. One blue carries the one action per screen, the sort tick and the current destination. One red carries what a reader must not miss. Status marks and person tints sit outside the budget and encode nothing a reader has to decode.
- The worst surface:
surface, in both themes.strongandproseclear 7:1 there,mutedandsubtleclear 4.5:1,edgeclears 3:1. Measured:mutedonsurfaceat 11px is 6.84:1 light and 8.43:1 dark;proseonsurfaceat 13px is 12.17:1 light and 13.98:1 dark. - How absence is rendered: three marks, never one:
- zero is written
0, with its dimension drawn at hairline length (City of Fairview: a pilot that carries no fee); data-absent="null"is an em dash: nobody gave a value (Parker County MUD 3's value; Fairview's next action);data-absent="pending"is a hollow dashed box: nobody has worked one out yet (Trinity Basin's value and its eight-week touches). Each carries a visually-hidden sentence, so the three are three states to a screen reader as well as to an eye. A date is read against today rather than on its own: the one row whose next action has run out carries acriticalmark and the word, and the one that falls today carries the word alone.
- One signature object: the dimension line: a quantity drawn to a stated domain of 150,000 USD with a tick at each end. The same drawing device ends the sorted column head, marks the current navigation destination, and is the bar of the channels list.
No lit surface is declared. console+lit is refused by ADR-010 and this page is a table, a form and a list.
Generator
The token layers are generated output, linked in the mandated order and never edited:
node generate.mjs
Palette: preset ink, register deep, palette.py 1.0.0, the shipped default since ADR-018, and the firm's own budget. The page carried quartz until 2026-09-16, when the vendored layers were refreshed from the system and the preset moved with them. Six links, in order: centauri.spine.css, centauri.console.css, centauri.base.css, centauri.nav.css, centauri.table.css, centauri.chart.css.
The three companion pages swap the profile and that system's layers and change nothing else: field.css + surface.css, catalog.css + shelf.css + table.css, marketing.css + page.css + expressive.css. The palette, the focus ring, the marks, the corners and the motion rungs are the same files in all four.
The second state
The work list under an active filter that matches nothing: the heading and the filter bar stay, the empty result names what the filter excludes and what clearing it would show, and the control that clears it is the primary action of that region and carries the same words as the one in the heading. Both states are live: the filter chip and the button swap, and focus lands on the one control that survives both.
The other hard case is in the product itself: a 51-character organization name (North Central Texas Municipal Water Authority Board) in the identity column, which wraps rather than widening the table, and is the record heading at 375px.
Deviations
- Themes. CONTRACT §5 asks for tokens redefined under a guarded
prefers-color-schemeblock and again under[data-theme="dark"]. Centauri's base layer instead setscolor-scheme: light darkon:rootand resolves every role withlight-dark(), with[data-theme]narrowing the scheme. The system wins under CONTRACT's own preamble; both directions were verified in a browser (OS-dark context, and the in-page toggle in both directions). - Form controls. Centauri Console ships no field layer. The text field, the option menu and its chevron are authored in the page from spine tokens; the chevron is inline SVG in
muted, withappearance: noneon the menu. - Parity.
parity_check.pyis run with--allow "Collapse navigation". The nav layer hides the rail's collapse button under 600px by design; the bar collapses on a downward scroll instead and returns on any upward scroll or on focus landing inside it.
A note for anything that measures state
Every interactive state in Centauri is one implementation: a ::after overlay at 8% hover, 10% focus, 10% pressed and 12% selected of currentColor. The element's own background-color and border-color therefore never move, in any state, by design. A probe that samples only the element's computed properties reads every control on this page as inert; sample ::after beside them. Measured under a real pointer at 1280px: brand, rail toggle, every destination, the primary action, the theme button, the three sort controls and both filter buttons go ::after opacity 0 to 0.08.
The two that genuinely had no state layer were fixed here rather than in the layers, because both were the page's class use: the primary action is a <button> on this screen (it opens the quick-add row rather than navigating), so the nav layer's .centauri-nav a::after could not reach it, and the sort controls carried no centauri-state. Both now do, and the sort control fills its head so the wash lands where the layer already puts the pointer's cursor.
The 2026-10-29 uplift
The page stopped being a still of a console and became one that works. Every figure on it traces to one of 38 rows (29 carry a value, 412,700 USD in all) and the generator that wrote the rows asserts each total before the file is written.
- The first screen. The stat strip leads with the open value at display size; the other four figures (due today 6, ran out 4, kept slipping 3, replied 41% as 16 of 39) each carry their denominator. The "38 open rows" count in the masthead was removed: the stat strip says it once, with its basis.
- Saved views. Five views as tabs with counts. The Save control appears only when the question has moved, a custom view can be deleted with an undo, and names are checked at the field.
- The table. Search, stage and owner filters, a sentence that says what is selected ("All 38 prospects are selected."), keys j, k, x, Enter, Esc and
/, and Ctrl+Z. Parking a selection that includes an agreed pilot gives the partial outcome "31 of 33 prospects done, 2 refused." with each reason, and the undo bar stays until it is dismissed. - The record. From 1360px a sticky column beside the table (the Touches, Owner and Source columns step aside while it is open); under that, a native dialog, a bottom sheet under 600px.
- Add a prospect. Field errors at the field; a successful add moves every figure and the tab counts.
- The week and the chart. The ten day bars filter the table. The twelve weeks (Mondays, 22 June to 7 September: 54 added, 16 replies) have a scrubber and an equivalent plain table with a totals row.
- Work. Two toggles on a list whose default filter matches nothing, so the empty state shows its reason and offers to clear, never to create.
Deviations recorded:
- Phone records. Under 600px a pipeline record keeps the name, Stage, Value and Next action; Touches, Owner and Source live in the sheet. CONTRACT's "no column dropped" is met by the sheet, not by the row.
- "This week" on the work list is Monday 7 to Sunday 13 September; the chart's weeks are Monday-based and the last one is drawn dashed because it runs to Friday.
- Absence inside the record panel re-declares the table layer's two rules (a long dash for null, a dashed box for pending), because that layer styles them only inside
.centauri-table. - Boson. Run on a copy without the linked layers, so its colour, overflow and focus findings on the chart, table and nav are artefacts of the missing CSS; the real page measures clean in
a11y-probe. Itspolicy-missingfinding is a demo with no real form to submit.
Gates
import_lint.py systems/centauri/ clean (the spine, the declared profiles, base, nav,
and only each system's own layers; every page declares
@centauri-multi-profile, because a product picks one)
nav_lint.py index/field/catalog/marketing no findings on any of the four
parity_check.py --allow "Collapse navigation" aria-375 aria-1280 parity holds,
23 controls at both widths
screen_probe.py systems/centauri/ clean: 872 focusable elements over 5 pages,
7 viewports, reduced motion and forced colors
audit_page.py index.html 2 info (touch icon, manifest)
slop_scan.py index.html clean
craft_lint.py index.html clean
motion_lint.py index.html clean
layout_lint.py index.html clean
copy_lint.py index.html 15 findings: 13 warning, 2 info: title-case on
the organization names in the table, and passive
voice in the design notes
aria_lint.py index.html 0 findings
a11y-probe clean in both themes, both the screen and its four
companions
direction_check.py --ledger 1 warn, kept: hero-undecided over all five pages.
A shell has no first screen, which the skill says
in as many words; the tool reads five documents
and asks each of them anyway
tools/check.sh centauri PASS: every gate ran and every gate is clean
Re-run 2026-10-29 after the uplift above: every gate in tools/check.sh centauri is clean; the figures in the block are from 2026-09-16, and copy_lint is now clean.
Re-run 2026-09-16, after the vendored layers were refreshed from the system and the page took the account block, the ninety-day chart and the phone's way out.
Widths measured in Chromium at 320, 375, 600, 700, 760, 800, 860, 900, 950, 1000, 1024, 1280, 1440 and 1920: the document is never wider than the window, and no console errors at any width.
The companion page: motion.html
The screen shows what the system looks like at rest. Twelve of the twenty-one pages here have no keyframes at all, and a plate cannot carry a duration, so the one layer this portfolio could never show is the one every reviewer asks about. motion.html is that layer, in the same system, reached from the ledger row ("The motion, live") and from the console screen's masthead.
Its four direction lines
| Lever | Decision |
|---|---|
| Archetype | Document. A reference read top to bottom, with a specimen wherever a sentence would have to describe movement instead of showing it. The console screen beside it is a shell; this deliberately is not. |
| Data stance | Prose-first, every duration drawn. |
| Density | The console profile, unchanged. |
| The one bold thing | The dimension line, the system's own signature object, measuring time. A duration is a quantity against a stated domain, which is what that object draws. |
What is on it
Six timing rungs on one 320 ms domain, grouped Arriving and Leaving; the seven verbs the Console layer ships, each with a specimen you can trigger; the frequency ladder with the same panel at three rungs; four counter-cases, each running the same event twice; a switch that applies the shortened reduced-motion path so it can be watched on a machine where the setting is off; and four systems' published timings side by side.
Why it is not a twenty-second entry
The contract reserves the spec panel as the only place a page's subject is the design system itself. This page's subject is motion throughout, so it is a companion to the console screen rather than a member of the set. It is not counted in the twenty-one, not in the similarity test, and has no ledger plate; it gets a second link on the console screen's row instead, off a generic companion field in build-index.mjs so the next system to ship one needs no code change.
Every timing on it was measured in the browser, not transcribed
| Verb | Printed | Chromium |
|---|---|---|
| enter | 200 ms | 200 ms |
| exit | 160 ms | 160 ms |
| expand | 200 ms | 200 ms, on ::details-content |
| collapse | 160 ms | 160 ms |
| select | 120 ms | 120 ms |
| enter, shortened | 0.01 ms | 0.01 ms |
The six dimension lines are drawn to scale, and the check is that each bar's width back-computes to its own printed label: at 1440 px the track is 375 px and the six bars return 120 / 200 / 320 / 90 / 160 / 240 exactly; at 375 px the track is 174 px and they return the same six numbers.
The four cross-system values came from the generators, never from recall:
carbon_tokens.py --motion Carbon: 70/110/150/240/400/700 ms, six easings
primer_tokens.py --scales Primer: enter 300 ms easeOut, exit 200 ms easeIn
md3_spring_css.py --table M3 standard: spatial-default 317 ms, effects-default 231 ms
centauri_tokens.py --list Centauri: three enter rungs, three exit rungs, three easings
The two values on the page that are not tokens
The last counter-case is this firm's own scroll reveal, and it is the only place the page leaves the token set. It runs 600 ms, longer than every rung here, and one half of it is linear, which check.mjs refuses inside a layer. Both are deliberate and both are read out of computingusa's shipped reveal.tsx (transition-opacity duration-600 ease-linear) and globals.css rather than invented: that specimen is a marketing surface's motion, not a console verb.
It earns its place because it is the one finding on the page that is a scar rather than doctrine. The firm shipped a moving entrance four times (32, then 48, then 64 px of travel, at 1000, 800, 700 and 900 ms, on two curves) and every version was reported as arriving in place while measuring as running correctly. An element entering a page the reader is already scrolling competes with the scroll, which moves the same element further and faster in the same moment, and no duration or curve fixes that.
The demonstration proves its own claim: 60 ms into a 600 ms fade the eased pane measures 0.30 opacity and the linear pane 0.05, so the eased one is six times further along at the same instant. Both settle at 1.
The critique round, and what looking at it found
Six full-page renders (390 / 1440 / 1920, both schemes) after every source gate was already green. Seven defects, and only one of them is something a source gate reports:
- The ruler ran the full width of the document. At 1440 px the 320 ms line spanned the page and the set read as six rules rather than as a measurement. Bounded to 34 rem, and the two sets grouped under
ArrivingandLeavinglabels instead of separated by colour; the accent is the product's one action per screen, and six exit bars is not that. - The ladder rendered as a band of nothing. Its three panels sat at
opacity: 0until pressed, so the section read as a broken page. They are present now and the control replays the arrival. - The stagger blocks were nearly invisible in both themes: the subject of the demonstration was a faint rule on a near-ground fill.
- The reveal specimen and the ladder were invisible at rest, for the same reason and with the same fix as (2) above: two labels over nothing.
- Neither replayed. Removing the attribute and putting it back inside two
requestAnimationFrames replays no transition at all: the element transitions towards the rest state for about 30 ms and then back, so nothing visibly moves. Measured at 0 ms of travel. The start state has to be committed with the transition suppressed and a reflow forced:is-resetting, on for exactly one frame. The keyframe verbs (enter, exit) need the different remove-reflow-add form and already had it. - The whole script was dead for one revision. The transition helper was named
replay, which the keyframe demos already used, andSyntaxError: Identifier 'replay' has already been declaredtakes the entire<script>with it: every control on the page, silently. Every source gate stayed green.drive-probe.mjsis what catches this, and the lesson is that it has to run after the last change rather than once. - The tab bar collided at 375 px. The brand kept its label (the rule said
display: none, under a comment saying the opposite, and lost to the layer anyway) andMisappliedwas cut overFour systems. Measured: a tab label gets 51 px at 375 and 42 px at 320,Misappliedneeded 60, and the layer clips withtext-overflow: clipso a word that does not fit loses a letter in silence. Fixed by the console screen's own icon-square brand rule, five one- and two-syllable labels, one step down the reading ladder below 360 px, and an ellipsis so the remaining 320 px overflow says so. Clean at 360, 375, 414 and 599.
Gates
audit_page.py 2 info (touch icon, manifest), as every page here
slop_scan.py 2 warn: the string "scroll" in two event names, as index.html
craft_lint.py 2 warn (peers: two toggles, three tabs), 1 info (disabled styled by the layer)
motion_lint.py clean layout_lint.py clean
type_lint.py clean ui_lint.py clean
copy_lint.py 1 warning: "Parker County MUD 3", a proper noun
aria_lint.py 0 findings unsafe_patterns.py clean
import_lint.py clean nav_lint.py no findings
direction_check.py 1 warn: hero-undecided, which index.html reports identically;
the ledger reads literal values and this contract forbids writing any
screen_probe.py clean: 275 focusable elements, 7 viewports, both themes,
reduced motion and forced colors
a11y-probe.mjs clean at 375, 1280 and 1920 in both themes, control probe caught
drive-probe.mjs 27 controls driven; nothing thrown, emptied or focus-dropped
tools/check.sh centauri PASS: every gate ran and every gate is clean
screen_probe.py found the one defect no source gate could: overflow: clip on the specimen stages cut 4 px off every focus ring inside them (SC 1.4.11), 108 findings across the viewports. The clip was bounding nothing: the enter and exit verbs travel inside the row, the rail's labels clip themselves, so it was removed rather than worked around. overflow-clip-margin was the first fix and is correct in the browser, but the probe compares rects and cannot see it, so a page relying on it would gate clean and read as unmeasured.
Three of the portfolio's own tools were hardcoded to index.html and so could measure no companion page at all. a11y-probe.mjs, drive-probe.mjs and review-shots.mjs each take --page <file> now; the default is unchanged.
Round two (2026-10-29)
A cold review scored the first screen 6/10: small mono type, a wide table, nothing set large. The console now opens on the morning question. "Needs a move today" is the one datum-scale figure, 10, beside its breakdown drawn as a dimension line over the 38 open prospects (4 ran out, 6 due today, 28 not due yet or undated). A "Show only these" toggle (aria-pressed) opens that view. Open value, kept slipping and replied drop to a quieter three-figure strip. Body text in the pipeline band is 13px or more; tracked-caps labels and key caps stay at 12px.
Round four
- Work opened on an empty state: the default filter (Blocked, due this week) matched nothing, because the three blocked tasks carry no date. It now opens on Doing, due this week, five tasks, with both toggles pressed. The empty state is still one click away (Blocked plus due this week).
tr[hidden]is now also honoured by the phone card layout for the task table, which the empty default had been hiding. - The theme button names the theme it switches to and follows the
data-themeattribute however it was set. - At 390 the Motion and Portfolio links moved behind the account control with the field app; the masthead is title, date, theme, account on two lines with nothing under the title.
- The four rows that ran out are tinted across the row in the same status red the bar above uses, so "4 ran out" and the table are the same four things. The mark and the word stay; a checked row keeps its own state.
Round eight
- "The week" and "Channels" did not align at their tops, and Channels' rows spread into empty bands (real). Cause:
.bandis a grid whose auto rows stretch when its neighbour is taller, so the shorter band's heading row grew and centred its title lower..pair > .bandnow declaresauto minmax(0,1fr) auto auto: only the list takes the slack, so the headings share a baseline and the five channels sit on the same rhythm as the week's ten rows, down to the same foot. - The rail was mostly blank above "Add a prospect" (real). Cause: the system's
.centauri-nav__actionismargin-block-start: auto, which pins it to the foot and leaves the empty run between the destinations and it. The action now follows the destinations and the account block takesmargin-block-start: auto, so the blank is below the work. This is a page-level override of the spine's rail anatomy, made here and not incentauri/(generated, never edited). - 390: the pipeline is 38 rows with no seams (real; the page is 8,989px, the list about 3,200 of it). Cause: a phone has no Bank or Stage column to read down. The list is now divided by whatever it is ordered by (
paintGroups): value bands under the default value sort, Ran out / Today / To 17 Sep / After 17 Sep / No date under the next-action sort, the first letter under name, each with its count. It follows the filters and the sort, and is drawn only below 600px. - Declined: a "Today" block in the rail. It would repeat the morning figure that opens the page and add controls to the one
<nav>thatnav_lintholds to four destinations.
Round nine
- Width sweep, 880 to 960. The
.centauri-vhtext in empty table cells was placed against the page, not the table scroller, and sat past the edge (+82px at 880 down to +2px at 960);.centauri-tableis nowposition: relativeso the scroller clips it.
Read below the fold, in driven states, at 200% zoom (720px), at 390 and 360 and in dark. Heuristics scored 0 to 4, before then after: status 3 to 4, match with the world 3 to 3, control 3 to 3, consistency 2 to 3, error prevention 3 to 3, recognition 3 to 3, flexibility 4 to 4, minimalism 3 to 3, recovery 3 to 3, help 3 to 3. Total 30 to 32. The moments behind the scores of 2 before: the Channels figures disagreed with the open rows, and a touch target on a phone was 24 to 28px against the system's 44.
- Vendored layers refreshed 0.4.0 to 0.7.0 from the skill's assets after its check passed (235 checks). The measured pairs are unchanged: muted on surface 6.84 and 8.43, prose on surface 12.17 and 13.98, re-measured from the token values.
- Channels disagreed with the rows (real, 2). Four open rows come from a site visit, the figure said three, and field day said 22 for 21. Now 21 and 4, still 54 in all.
- Touch targets (real, 2). The page used no
centauri-target. Every button, sort control, tab and row button has it, and under a coarse pointer the inputs, selects, account menu, day buttons, week range and the row checkbox cell reach 44px. A tap on the checkbox cell now toggles the box. - The pipeline at 600 to 860px, which is 200% zoom (real, 2). The table layer's container query made six-line cards with wrapped labels (the page was 8,843px). The record compaction now keys on the table's own container, 5,927px at 720, and the stat strip folds to two columns below 760.
- Phone group header tinted by a ran-out row (real, 2). The header band is opaque.
- Row-state seam (1). Hover and selection mixed each cell's own ink, so the identity cell read a step darker than its neighbours. One ink for the whole row.
- Detail sparkline (1). The end dot was an ellipse under a non-uniform scale and the eight numbers sat off their points. A round-capped zero-length stroke, points centred on their slots.
- Week scrubber (1). The thumb now stands under its week's point, inset to the plot.
- Smaller (1). The phone tab bar action was flush to the bar's foot; tracked labels collided at line-height 1; a separator wrapped to a line end; the keyboard list is a two-column grid; a long prospect name wraps beside its label at 360.
- field.html. The column is centred (it hugged the left at 1440 and 1920), the recertification dates sit in 2027, "Not a door" says 1, and the title names Rockwall.
- catalog.html. The duration bar is four edges thick so a 1 week bar reads against the 16 week domain.
Left, taste: the owner select sits alone on a second row at 720; the Channels value is far from its bar at 1920.
Round ten (2026-11-09)
A second cold read of the pipeline as Dana Reyes working the week. The reviewer stalled before it reported, so this record is of the fixes it had in progress, finished, driven and gated by the session that sent it. No heuristic score was taken this round; the fixes are rated by severity instead.
- The page did not change after an add (severity 2). A new prospect counted in the table and the stat strip, but the week chart, its twelve-week total, the figure caption, the Channels bar for its source and the chart's own annotation all kept their old figures. Each now moves with the add and moves back on Undo; the chart's scale and its peak annotation are computed from the rows instead of written in. The sentence under the week ("add up to all 38 open prospects") follows the count too.
- An agreed pilot offered Park and nothing said why it could not be parked (severity 2). Its record now offers "Back to Proposal" and says, in front of the buttons, that an agreed pilot cannot be parked.
- The result of an act in the phone sheet landed behind it (severity 2). The page behind the modal sheet is inert, so the undo strip was out of reach. It is now shown in the sheet, above the record, takes focus, and moves back to the page when the sheet closes.
- Adding a prospect on a phone opened a modal over the result (severity 2). Under the split width the new row is scrolled into view and the undo strip says where the add was counted; the record opens beside the table only where there is room for it.
- The selection bar scrolled away from the rows being chosen (severity 1). It sticks to the top of the pipeline while rows are selected.
Invention: none. "Back to Proposal" is the stage before "Pilot agreed" on the page's own ladder.
Driven after the last edit at 1440 (light), 390 (dark) and 360 (light), the phones with touch at 3x: add then Undo, the pilot record, Back to Proposal in the sheet, the selection bar. No page error and no horizontal scroll at any width.
Cold score (a second reader, same day)
A reviewer who had not seen the page scored it before any edit, driving every action at 1440 light, 390 dark and 360 light (touch, 3x) and at 1920. Heuristics 0 to 4, before then after: status 3 to 4, match with the world 3 to 3, control 3 to 3, consistency 2 to 3, error prevention 2 to 3, recognition 3 to 3, flexibility 4 to 4, minimalism 3 to 3, recovery 3 to 3, help 3 to 3. Total 29 to 32. The moments behind the 2s: blue blocks drawn over the saved-view tabs on every phone, and an organization that could be added twice.
- Blue blocks over the saved-view tabs on a touch screen (severity 3). The tabs carried
centauri-target, whose overlay is a::before, and the tab layer draws its tick of selection on the same::before; on a coarse pointer the tick became a 44px accent block over the label. The tabs no longer carry the class and take the coarse target as their height instead. Found by sweeping every painted pseudo-element at 390, not by looking at a still. - A day with nothing on it left the reader out of sight of the result (severity 2). Tapping Sun 13 emptied the table, the page got 3,000px shorter and the reader was left among the design notes on a phone, or on the week with the empty state above. The page now scrolls to the filter bar, where the chip, the count and the empty state are, and focus lands on "Clear the filters". The reason now reads "Without “Next action Sun 13 Sep”" and not a lowered "sun 13 sep".
- The undo bar covered the rail's account links at 1440 and 1920 (severity 2). It read the rail width from a variable that exists only inside the shell. It now reads the shell's state and sits beside the rail, expanded or collapsed.
- Undo inside the phone sheet dropped focus to the page (severity 2). Focus now stays on the record it was acting on.
- A prospect could be added twice (severity 2). An organization already in the pipeline is refused at the field, ignoring case and accents, and names it.
- A filter left a record open for a row that was not on screen (severity 2, wide split only). A question that excludes the open row now closes the record; a stage change made from the record still leaves it open, as the confirmation.
- The partial park message promised a list that was not there (severity 1). "The rest are listed with the reason for each" is gone; each refusal is named.
- A checked row on a phone was patchy grey (severity 1). The tone is on the card.
Left: a parked prospect whose next action has run out still counts in "4 ran out" and in "Needs a move today" (severity 1; the figures are the brief's own and moving them is the owner's call). Channels says "Sweep" where the source is "Register sweep" (severity 0, a deliberate short label).
Inventions: one. The duplicate-name refusal and its wording ("is already in the pipeline. Search for it instead of adding it again.") are not in the brief.
Driven after the last edit at 1440 light and dark, 1920 light and dark, 390 light and dark, 360 light and dark (phones with touch at 3x): tab strip, row selection, reopen then Undo, a zero day, the move-only view with the record, the duplicate refusal, a unique add then Undo, a partial park. No page error and no horizontal scroll.
Round eleven (2026-11-10)
A cold read found Park and Move leave the "Needs a move today" strip and the tab counts as they were, and that parked rows are still counted in them. Not changed: the figures (10, 4 ran out, 6 due today, 28, 38 open) are the ones this brief pins in Round two, the uplift and Round four, and Round ten already left them to the owner. Three questions are open for the owner:
- Does a parked prospect owe a next action? If not, Needs a move today falls from 10 to 7 and ran out from 4 to 1 at load.
- Then does Parked become a fourth segment of the 38 (1 ran out, 6 due today, 5 parked, 26 not due), or does the strip count the 33 that are not parked? The day bars and the red tint on the three parked rows that ran out would need the same ruling.
- Should Move clear a row? The brief gives no rule that advancing a stage settles the next action, so none was invented.
- The five Parked rows looked unparked until acted on (severity 1). The markup carries
data-stage="Parked"and nodata-parked, which onlypaintStageset, so Ellis County ESD 2, Oakmont Regional Library and Elm Fork Transit showed the strong name colour at load and the muted one after a Park, Reopen or Undo. The start-up now paints every row once throughpaintStagewith its own stage, so a row parked in the file and a row parked on the page are the same. No figure moved (10, 4, 6, 28, 412,700; tabs 38/14/10/3/9).
Invention: none.
Driven after the last edit at 1440 light, 390 dark, 360 light (touch, 3x) and 1920 dark: load, Park, Undo, Reopen, Undo. At each step the parked rows and the rows whose stage is Parked were the same set (5, 6, 5, 4, 5), the figures were unchanged, no page error and no horizontal scroll.
Round twelve (2026-11-10)
A reviewer found the filter selects on a phone drawn 44px high with the value stuck to the top and an empty lower half (severity 2). Reproduced at 390 dark, 360 light and 768 light under a coarse pointer: the value sat 8 to 10px above the control centre in every select (#f-stage, #f-owner, #bulk-owner, #np-owner, #np-source) while the chevron was centred. At 1440 and 1920 (28px) the offset was already under 1.5px. The text inputs (#q, #np-org, #np-value, #view-name) were already centred in every case.
- Cause. The coarse-pointer rule raises
min-block-sizeto the system 44px target. Underappearance: base-selectthe select lays its value out from the top of its box, withalign-items: normaland 3.25px of block padding, so extra height is empty space below the value. - Fix. One rule in the page stylesheet,
.select > select:padding-block: 0; align-items: center. The value is centred at every height the field takes, with the base-select and the fallback appearance. The chevron was already centred by its own rule. No system asset was touched.
Invention: none.
Measured after the last edit, every select and text input on the page (the add-prospect form and the view-name field forced visible, the bulk bar shown): the ink of the value against the control centre, from an element screenshot, at 1440 light, 1920 dark, 390 dark, 360 light (touch, 3x) and 768 light (touch). Selects moved from 8 to 10px off at 360, 390 and 768 to 0.81px or less at 360 and 390 and 1.14px at 768 (#np-owner, "Dana Reyes", a word with a descender, so its ink centre sits off its line box); every other select and input is within 1px, chevrons at 0.00. No page error and no horizontal scroll at any width.
Round fifteen (2026-11-10)
A reviewer (round 14) found two severity 1 defects in the add-a-prospect form at 1440, both still present from round 12: the Source select's chevron overlapped its own value ("Field day"), and the ORGANIZATION label touched the focus ring of its input. Both reproduced before editing (1440 light and 1920 dark).
- Source overlapped the chevron: two causes. (1) The form is five items (organization, value, owner, source, the actions) in a grid of four tracks,
2fr 1fr 1fr max-content, so Source fell in the max-content track and measured 73px at 1440 and 1920, narrower than its longest value "Register sweep" (93px of text); Save and Cancel wrapped to a second row. (2) The reserve for the chevron,padding-inline-endon.select > select(20px), was beaten by the form layer's.field > .select > select { padding-inline }(8px), at a higher specificity, so in every select the value could run to 8px from the edge with the chevron at 4 to 18px. Every select on the page was affected (#np-owner,#np-source,#f-stage,#f-owner,#bulk-owner); only Source was narrow enough to show it. - Fix, the select rule.
.select > selectis split: appearance and centring stay on it, and the padding is stated for.select > select, .field > .select > selectaspadding-inline-end: scaled-6 + scaled-2(24px), so it holds at the form layer's specificity.appearanceis deliberately not repeated at that specificity: it would beat the@supportsbase-select rule and drop the styled picker (seen once during the work, reverted). - Fix, the grid. From 760px
.addrowisrepeat(auto-fit, minmax(11rem, 1fr)), the organization spanning two tracks. Tracks follow the form's own width (the rail takes 264px of it), and 11rem holds the longest select value, the chevron's reserve and the value field's placeholder. - Label crowded by the focus ring.
.fieldgap was 2px; the ring is 2px wide at a 2px offset, so it reached 4px above its input into the label. The gap is now focus width + offset + 4px (8px), which also clears the error line under the input.
Not changed: how "Needs a move today" counts parked rows (the open owner question in Round eleven).
Invention: none (the 11rem track minimum and the 4px of breathing room beyond the ring are measures, not product facts).
Measured after the last edit: for every select on the page (the add form shown, the bulk bar unhidden), the end of its longest option's text (canvas-measured in the select's own font, plus the option's 8px padding) against the chevron's left edge, with the label-to-input gap and the placeholders of the two text inputs, at 17 widths from 320 to 1920 in light and dark, mouse and touch: every value clear of the chevron by 8px or more (the narrowest, Source at 1440, 55px), no placeholder cut, label gap 8px, no horizontal scroll, no page error. The same measure on the unedited file gave Source at 73px wide and a clearance of minus 55px at 1440 and 1920.
Added later in Round fifteen, the two the reviewer's severity 1 list still carried.
- The view tabs clipped "Kept slipping" on a phone, with no hint. The nav layer's
.centauri-tabsis one scrolling row (overflow-x: auto) and Centauri has no overflow pattern for it:navigation.md§ In-page tabs sets 2 to 5 tabs and says nothing about overflow; the rail's tab bar is for destinations, not views. A fade or an arrow would be a gradient or an invented control. The page's own.viewbar .centauri-tabsnow wraps (flex-wrap: wrap; overflow-x: visible), so every one of the five tabs is whole on screen at every width and nothing is hidden or needs a gesture to find. It costs one control height per extra row: two rows at 390 and 360, three at 320, one from the width where all five fit. - The undo bar covered the lowest rows while it showed. Script (
measureUndo) sets--undo-coveredon the root while the fixed bar is showing: its height, its offset from the viewport floor and 8px of air (the page's own tab bar is inside that on a phone).htmltakes it asscroll-padding-block-endand.centauri-mainaspadding-block-end: max(its own bottom padding, --undo-covered), so the end of the page scrolls clear of the bar and a scrolled-to element is stopped above it. It is re-measured when the bar's size changes (a second line of message, a resize) and cleared when the bar is dismissed, undone or carried into the record's sheet, where it is not fixed.focusRow(j, k, the return after Undo or Esc) now scrolls the whole row into view, not only its button: a phone's row is a card that runs below the button, and the button alone was clear while the card's last 15px was under the bar. - Invention: none. The 8px of air is a measure.
- Measured after the last edit, at 1440 light, 390 dark, 360 light and 320 light (touch, 3x) and 1920 dark, with the bar up after parking a row: all five tabs inside the viewport and the strip not scrollable at every width;
jheld to the last row ends with the row 8px above the bar at every width (before the scroll padding the last row ran under it by up to 15px, and at 360 and 320 two of the last four did); the end of the page 8px above the bar; the properties gone after the bar is dismissed; no horizontal scroll and no page error.
Round seventeen (2026-11-10)
Two defects from the round 16 reviewer (36/40, no severity 2 or 3). Both reproduced before editing.
- "Moved: Never" on a row that had just moved stage. The record's fourth fact was labelled "Moved" and counts next-action slips (
data-slips), so a row just moved to Pilot agreed read "Moved: Never", a stage move that never happened. It is now "Next action slipped" with the same values ("Never", "1 time", "3 times", and the quiet "Kept slipping" under two or more). The stat strip's third figure said "moved more than once" for the same count; it now says "next action slipped more than once", so the record, the strip and the "Kept slipping" view name one thing one way. - The undo bar covered the lowest visible rows (severity 1, sent back twice). Round 15 only kept the page's last row clear (padding and scroll padding at the end of the page); at any other scroll position the fixed bar still sat over rows (1440 after Move to Pilot agreed: 4 rows partly under it; 1920: 2; 390 and 360 after a bulk Park: 1 to 2 at five scroll positions). Padding cannot fix that, since a row scrolling past the bottom of the viewport always passes under a fixed bar. While the bar is fixed, the page now stops being scrolled by the window:
html[data-undo]hides the window's overflow and the body becomes the scroller,--undo-coveredtall less than the viewport (the bar, the gap under it, and 8px of air; the phone's tab bar is inside that). The bar sits in the strip below the scroller, so no row scrolls under it at any position, and the last row of the page still reaches the top of the strip. The rail, the shell and the record column take the same height off their100dvh, so none of them is cut or scrollable into the strip. The scroll position is carried between the window and the body when the strip appears and goes (no jump), the phone's collapsing tab bar and the rail's current-section mark listen to both scrollers. The round 15 rules (htmlscroll padding,.centauri-mainbottom padding) are removed, since the strip replaces them. The bar in the record's sheet (static, not fixed) is unchanged and sets nothing. - Invention: none. The strip is the bar's own measured height; 8px of air is a measure.
- Not changed: the opening figures, and how "Needs a move today" counts parked rows (the open owner question in Round eleven).
- Measure: a row partly above the strip's edge is cut flush at the edge, like any scroller's last visible row; its rectangle still extends below the edge in
getBoundingClientRect. Rows were therefore measured clipped to the scroller (the body's rect), and the toast against every row of every table at five scroll positions at 1440, 1920, 390 and 360: no row is visible in the toast's rectangle; the toast sits 8px below the scroller's edge at every width.
Round nineteen (2026-11-10)
Three reservations from the round 18 reviewer (36/40). Reproduced before editing: the first and third; the second did not reproduce as an overlap.
- "Show only these" gave a phone no visible result (severity 2). At 390x844 and 360x800, touch: scrollY stayed 0, the label stayed, and only a faint outline changed; the view tabs sat at y 807 under the phone's bottom tab bar and the filtered rows below them. Now, on a phone, the press brings the view tabs to the top of the viewport (the list follows by 260px, the tab bar clear of it), the way a day in The week already carries the reader to the list. The button's state reads in words: it is "Show all prospects" while the view is on and "Show only these" otherwise. A control whose name changes does not also carry (a screen reader would say both), so
aria-pressedis gone and the outline is keyed ondata-on; the polite status line still says "10 of 38 prospects.". One painter,paintMoveShow, is called from the tab state and from the aggregates, so choosing the tab by hand and pressing the button agree. - The undo toast over the lowest Work rows (severity 1): not reproduced. Measured in the real viewport (1440x900 and 1920x1080, after Reopen as Qualifying) at five scroll positions, with every row, stat and heading clipped to the scroller: the scroller ends at 830 and the toast starts at 838 (1010 and 1018 at 1920), zero rows in the toast's rectangle. The reviewer's picture shows the Work table cut flush at the strip's edge, which is the round seventeen design, not an overlap. No edit.
- ORGANIZATION's label 25px above the other three with an error (severity 1). The row used
align-items: end, so the field carrying an error line (taller) rose above the others. From 760px the row now aligns at the start, so the four labels share one top (464 for all four with no error, an organization error, and a value error), and the actions carry an empty,aria-hiddenlabel of their own so Save sits level with the controls (484 against the input's 484). - Invention: the words "Show all prospects". No figure, date or policy added. Not changed: how "Needs a move today" counts parked rows.