Changed#
The web app got the layer the palette never had. It carried two greys —
--ink and --ink-soft — so a paragraph a reader is meant to read and a
footnote they are meant to skip had the same weight. Three tiers is the
smallest number that lets a page say read this, read this if you want and
this is here for the record. A second border tier came with it, for edges
that separate regions rather than parts.
And state became a token. Hover and pressed were reached for ad-hoc —
bg-muted at one call site, bg-accent at another — so two controls doing the
same thing looked different, and changing what hover feels like meant finding
every one. --layer-hover and --layer-active are defined once, and
redefined in the dark block: a token whose only definition lives in the light
palette is one the dark theme silently inherits, which is how a page ends up
drawing one theme's hover on the other theme's surface.
The application's chrome became a sidebar. Wordmark, tagline, account and
language all sat at the same altitude as the work, so nothing said "this is the
chrome and the rest is the product". It was first pulled into a full-width bar;
that bar became a rail, because the thing that was missing was not altitude but
a place for navigation to live. 236px expanded, 60px collapsed, the preference
kept in localStorage so it survives a visit. Below lg the same rail arrives
as a drawer with a scrim, and closes three ways — Escape, the scrim, and
choosing a mode, which is what it was opened for.
One rail, not one per breakpoint. The first draft rendered it twice, inside
an lg:hidden block and again inside a hidden lg:flex one. Hidden is not
absent: that put two tablists on the page controlling the same panels, two
role="tab" elements per mode for a screen reader to read out, and two account
components each fetching the session. Visibility is a style; duplication is a
bug.
A phone had 261 pixels of a 390 pixel screen. The tabs root is a flex
container and orientation="vertical" leaves it a row, so the mobile bar became
a 128px column standing beside the content. It is a column below lg and a
row at it now. Invisible in every screenshot, because the only ones taken had
the drawer open over it, and obvious the moment the bounding boxes were printed.
The three modes read as navigation. They were dressed as a settings toggle: a small pill group in the sunken grey, the weight of a rule-level switch. They are not one control with three positions — Optimise shortens a prompt, Compare judges two of them, Your bill reads a usage log, which are three different jobs on three different inputs.
Collapsed, their labels are hidden rather than removed. sr-only, not
unrendered: a title attribute alone would rest a tab's whole accessible name
on the weakest source the accessibility tree has, on the one control a reader
has no other way to identify.
The account control became a menu. A badge, an avatar, a name and a Sign out button in a row fitted a full-width header and fits neither width the rail has. The identity is the trigger now and the actions live behind it, opening upward at 236px and sideways at 60px so it can never open off the screen. Signing out is rare and destructive and no longer sits one stray click from the navigation. The "temporary session" warning moved inside, where there is room for the sentence that explains it, and sits above that sentence rather than run into it: "temporary session This deployment keeps sessions in memory" is not a sentence in either language.
The rail has two columns, and every number in them is derived. A nav row is
8px of the list's padding, 1px of transparent border and 10px of its own — so
every glyph sits at 19 and every label at 46. The wordmark carried its own 12px
and hung seven pixels left of the column it heads; the account row's gap is
5px rather than 10, which is 10 minus the 5 its 22px avatar exceeds a 17px
icon, so both columns hold across two different glyph sizes. Collapsed, all five
glyphs centre on the same half-pixel.
A cascade the class list hid. The collapsed mode icons measured 12.5px left
of centre while every other glyph in the rail measured half a pixel.
TabsTrigger carries its own group-data-[orientation=vertical]/tabs:justify-start,
and a [&_button]:justify-center from the list is the same declaration at lower
specificity — so tailwind-merge did its job, the class list was correct, and the
computed value was still flex-start. The override belongs on the trigger, in
the same variant.
Three waits, each in the shape of what is coming. Optimise, Compare and
Library shared one defect: the button changed its label and the panel that will
hold the answer went on showing its empty state, so the only sign of life was at
the far side of the screen from where the reader is looking. Each draws the real
rows of its own report now — the caveat and the headline figure for Compare,
three list rows for Library — so nothing jumps when the answer lands. Each
carries role="status" and aria-live="polite"; Library keeps its sentence for
anyone listening rather than looking.
The empty result panel names what the report will contain. It was one italic sentence centred in a grey box — the shape of a panel apologising for being empty — and it is the first thing every reader sees, before they have typed anything. It now says that nothing has happened yet, what to do, and the three things pressing the button actually gets you.
Fixed#
A guard stopped guarding, twice, and both times it was this work that found
it. ui.test.mjs required the Library trigger to be spelled
<TabsTrigger value="library"> with nothing after it — so giving it a title
broke a test with nothing to say about titles, and widening it to allow
attributes still required them on one line, so wrapping the tag over four lines
broke it again. It matches the open tag and any whitespace now, and is handed
three ungated triggers it must reject rather than one. What is guarded is the
gate, not the spelling.
Every README transcript was run, and twenty of them were not transcripts.
The page is headed "Real output, transcribed". Ten commands were checked by
executing them and comparing line for line — rank, blame, diff,
baseline, check, route, prune, quality, eval, profile — with each
difference handed to an independent reader who had to reproduce it before it
counted. Twenty-nine survived, and all but four are one shape: a transcript
that stops early and is not marked abridged, so a reader takes a partial
output for a whole one. The README already had the convention — a bare … line
inside the fence — and simply had not used it.
Four were not that.
--max-growth is a token count, not a percentage. The page said
--max-growth 10 fails "a prompt that grew more than 10%". It compares
tokenDelta > limit. Measured both ways: a prompt that grew 50% but five
tokens passes --max-growth 10, and one that grew 3% but thirty tokens
fails it. This is a gate people put in CI.
blame's headline figure was six times too high. The block priced +759
tokens at 50,000 calls as +$1,138.50 a month. Run on a constructed history,
+500 tokens at 50,000 calls prints +$125.00 — a rate of $5 per million, which
is what the catalogue lists for Claude Opus 5 and what the sibling diff block
on the same page already agrees with. The figure is +$189.75, and the two
blocks had been contradicting each other.
profile handed the reader the wrong homework. The block's whole point is
the command it gives you to check its own claim, and it printed
trazum eval <prompt> --cases <cases> --model claude-sonnet-5 where the tool
prints trazum route <log> --prompt-file <prompt> --cases <cases> --yes.
check's report had been re-laid-out around its own transcript. The status
column moved from the end of the row to the front, a header line and a summary
line appeared, and the token count no longer matched the snippet printed
directly above it — 34, not 43, for the very code the page shows. The snippet
now carries both prompts it reports on, so every number in the block is
derivable from the page itself.
And the caveats had been quietly falling out of the transcripts. quality
prints two — that a before-and-after is not an experiment, and that it cannot
see what else you deployed that day — and the block had replaced the first with
a blank line and dropped the second. blame's "±10%" estimate note, rank's
two definitions, prune's yardstick sentence (cut mid-clause and ended with a
full stop the tool never prints), route's "agreement is not correctness",
eval's "read the cases below before shipping this", and baseline's "if the
growth is intended, re-record" were all absent. Every one is the sentence that
bounds the claim above it. A refusal never arrives bare, and neither does a
measurement.
trazum watch reads the store, and the section that documents it never said
so. It shows trazum watch --once under the words "what a cron entry runs",
and on a fresh checkout that command refuses: "Nothing has been measured yet:
the store at .trazum/store is empty." The refusal is exemplary — it names the
store, names trazum connect <provider> --store as what fills it, and says why
watching nothing would be wrong. The page just never mentioned the prerequisite,
so the reader met it after writing the cron entry rather than before. One
sentence, and it borrows the tool's own words.
The transcript below it cannot be reproduced here — filling the store needs a
provider credential this session does not have — and that is said rather than
worked around. What WAS checked and holds: trazum connect with no argument
answers "Available: anthropic, openai", says the credential comes from the
environment and is never stored, and points at --dry-run, exactly as the
prose claims; --dry-run names both the prefixed and unprefixed variables for
each provider and the scope each one wants, also exactly as claimed; and
trazum gateway with no argument names anthropic, openai, deepseek and google
while the page states no count beside them, which is the rule that has been
broken thirteen times elsewhere.
And trazum serve's response was missing the one field the page insists on.
The block is presented as a complete JSON document; answerCost returns
schemaVersion: 1 from every one of its four branches and the type declares it
non-optional, so a consumer copying that shape would not know the field is
there — on a page that elsewhere calls it "the only thing you must branch on".
reason and call.basis were absent too. It says which fields it trims now,
and does not trim that one.
Checked and clean in the same pass, and recorded as such: docs/format.md's
three conform invocations all run exactly as written — a plain file,
--contract profile against a real report, and - with --json over stdin —
and it has no example documents to parse, only those commands. optimize --reorder --diff reproduces its four lines word for word on a prompt built to
the shape the flag exists for. optimize --suggest calls the model and could
not be run here, which is said rather than guessed at.
The --json contract was checked, and it holds — but its harvest could have
stopped reading without saying so. docs/json-output.md documents
thirty-five top-level fields and json-contract.test.js enforces them in both
directions against a real run: nothing emitted is undocumented, nothing
documented goes unemitted, and the two fields a plain run does not produce are
covered by the flags that produce them. That is a clean result and it is
recorded as one.
The weakness was in how the test read the table. Its field pattern was
[a-zA-Z]+, and no field is spelled with a digit today, so it read all
thirty-five rows and nothing was wrong. The day somebody adds p95Usd and
documents it, the name would stop being harvested, the "emits every field it
documents" check would stop covering it, and nothing would say so — the shape
this repository has been caught by more than a dozen times. The harvest counts
the rows now and fails on any it cannot read, which was proved by planting a
nested path in the real document and watching the suite name it.
And one figure was a string the tool cannot produce. profile's
conversation-shape line wrote a median of $0.02; it interpolates
formatUsd(shape.medianUsd), and formatUsd gives four decimals under a
dollar and five under a cent — precisely so a figure that rounds to nothing at
two decimals does not print as $0.00 beside a column of real money. It is
$0.0200. A guard now takes the rule from that function rather than from any
list typed beside it, and checks every dollar amount in every fenced block on
the page; the sweep found this one and nothing else.
A completeness pass asked what every lens had missed, and found four more. The frames during a transition, the dark theme's overlay layer, Windows High Contrast, and the requests the page issues on load — four places nobody had looked because every other check measured a settled state in a default theme.
In dark mode the scrim brightened the page instead of dimming it. It was
bg-foreground/40, and --foreground in dark is a near-white, so the layer
whose whole job is to recede became a 40% white wash: the page behind it
measured a relative luminance of 0.0157 closed and 0.1681 open — eleven times
brighter — while the drawer in front of it sat at rgb(38,36,32). The thing you
were meant to be looking at read as a dark hole punched into a bright field.
The scrim has a token of its own now, defined once and deliberately not
redefined for dark. Measured after: light dims from 0.9007 to 0.2731, dark from
0.0156 to 0.0108.
Expanding the rail threw its contents across the page. Only width is
animated, so every label re-renders at full size on the first frame while the
box is still 60px — and with nothing clipping it at z-50, the language toggle
was drawn and answered hit-tests 70px into the main column for the first ~50ms
of every expand. Proved by releasing the clip again afterwards:
elementFromPoint(130, 842) returns the "Español" button at rail widths of 60,
66 and 91px without it, and nothing with it.
Under Windows High Contrast the current mode became untellable. The rail
signals it with a background tint and nothing else — the line variant's
marker is switched off, because in a vertical list it sticks out of the rail —
and a tint is exactly what forced-colors replaces. Active and inactive rows
both measured rgb(232,232,232); all that survived was 31 terracotta pixels
inside a 17px glyph. The selected tab takes a Highlight outline there: a
shape rather than a surface.
The session was fetched twice on every load, by two components reading different fields of one answer — and two answers to one question can differ. A session expiring between them puts the Library tab on the page for a reader the account control has already decided is signed out. One fetch now, passed down; the account control takes the session as a prop and still renders nothing until it arrives.
Two documents described a header the app no longer has. README.md and
docs/accounts.md both said the Sign in button and the "temporary session"
warning live in the header. They live in the sidebar and in the account menu.
Three guards, each proved by breaking it — and each cried wolf first.
Styling a shadcn primitive from its parent, which loses to the primitive's
own variant-prefixed declarations and says nothing. The guard reads what the
trigger defends out of tabs.tsx rather than a list typed beside it. Its first
version flagged [&_button:hover]:bg-layer-hover, which collides only in a
state the trigger does not define and is measured winning in both themes; its
second flagged [&_button[data-state=active]_svg]:text-terracotta, which lands
on the icon rather than the button. State and element are each half the rule.
And its extractor could not see the two worst offenders it was written for,
because [^\]]* stops at the first ] and both of them carry their own
brackets.
A focus indicator that computes to nothing — outline-none with only a
focus-visible:outline-<n> beside it. Shown firing on the pair that shipped
and staying silent on the ring that replaced it.
A desktop preference read inside the rail. Its first version flagged
<Account collapsed={railCollapsed} />, the child's own prop name; it ignores
an attribute name now and still catches collapsed={collapsed}.
And CodeQL found one in the guard itself. The extractor's alternation was
ambiguous — both branches could begin with [ — so every [[] in the input
had two readings and the star multiplied them: 0.2ms at ten repetitions, 93.6
at twenty-two, 6,051 at twenty-eight. Excluding [ from the plain-character
branch leaves one branch able to consume a bracket, and the same input takes
0.01ms.
The test written to prove that fix would have passed against the vulnerable
version. It fed the pattern a plain run of sixty [, which the ambiguous
form chews through in 0.08ms — the recurring trap of this repository, an
assertion that only ever sees inputs it cannot fail on, this time inside the
proof for the defect itself. A test file is not exempt from being a guard.
An audit of the shell found nine more, and every one was measured. Five lenses swept the running app — geometry, keyboard, colour, Spanish, breakpoints — and each finding was handed to an independent reviewer who had to reproduce the numbers before it counted.
The active row had no surface, in either theme. The same cascade as the
collapsed icons, twice more: the trigger carries
group-data-[variant=line]/tabs-list:data-[state=active]:bg-transparent and a
dark: copy of it one step higher, so bg-layer-active from the list computed
to rgba(0,0,0,0) and the collapsed rail signalled the current mode by icon
hue alone. Fixing light left dark untouched, because the dark rule outranks the
light one.
A closed drawer kept five controls in the tab order. Off-screen is not gone: at 390px the second Tab landed on a Close button at x=-248, and five consecutive presses changed zero pixels on screen.
The drawer claimed modality and did not enforce it. It is a dialog with
aria-modal now, focus moves in on open, Tab wraps inside, and Escape returns
it to the hamburger.
Two contrast floors. --ink-faint measured 2.90:1 light and 3.60:1 dark
everywhere it was used; inactive tab labels 3.77:1 in light only, where the
dark theme had a token and the light one had shadcn's /60 opacity. The
faint tier is now the lightest warm grey clearing 4.5:1 on both the rail and
the raised menu — a narrow band, because a legibility floor outranks a visual
tier.
Focus indicators. The account trigger had outline-hidden and nothing in its
place, and every other ring in the rail was at half alpha — 2.03:1 against a
floor of 3:1. All five controls paint a full-strength ring now, verified by
pixel diff rather than by reading the classes.
Spanish, and a stacked toggle. The Spanish tagline ran 13px under the
drawer's close button and lost three characters to it; the collapsed language
buttons rendered 18px against a declared 28, because flex-1 in a column makes
two buttons share the group's height.
A desktop preference reached a phone drawer, producing a 248px overlay
containing a 60px rail — no wordmark, every label sr-only, and no expand
control, because the only one is lg:flex. And crossing to a desktop with the
drawer open left a full-screen scrim and a locked body with no visible control
to lift either.
Four of the eleven were introduced while fixing the others, and each was
found by running rather than by reading: a focus trap that read its container
once and silently did nothing, a retry that stopped at finding a candidate
rather than at moving focus, an outline-none/outline-2 pair computing to
outline: none 0px, and a comment claiming controls had no focus indicator on
the strength of a probe that read the computed style three percent into a
transition.
Nothing about what the app computes changed. The web suite goes from 351 to 367 — the Library gate assertion was rewritten twice, and three guards were added below.