…ady the rails

The edit modal said "This task… / Is blocked by" and never named the other
task — a sentence missing its object, and the one remaining comprehension gap
three reviewers agreed on. Both call sites now pass the counterpart's title.

The relation dropdown was set in bodyLarge, so a relation phrase outranked the
task titles it describes and the card header above it. It drops to bodyMedium,
its menu rows lose ~20% of their pitch, and the panel's ceiling is now a whole
number of rows, so the last visible option is no longer sliced through its
text.

The trailing rail on a linked-task row was a different width in manage mode
than in browse mode, so toggling the mode re-wrapped every title on the card.
Both states now occupy the same box.

Colour and rank, all one-line token changes: the blocked-by chip keeps amber
for its border and glyph but takes an ordinary label, so an explanatory chip
stops out-shouting the status pill it explains; section headers move to the
overline style at high emphasis while row status drops to low, breaking a tie
where relationship semantics ranked equal to the word "Open"; the overflow
glyph stops being the heaviest mark in the card header.

The unlink confirmation was a raw Material AlertDialog inside an otherwise
tokenised feature — the last surface that dropped out of the design system.
It uses the shared confirmation modal now.

Also: icon sizes and one spacer literal onto tokens, and the empty-state
action gets the same trailing chevron every other navigable row has.

Round 9 caught two round-8 fixes that never landed, and four reviewers
measured them independently rather than taking the changelog at its word:

- The dropdown panel ceiling was derived from bodySmall while the rows
  render bodyMedium, so the panel stood 6.4 rows tall and sliced the
  seventh option through its glyphs.
- Browse mode reserved one action box in the trailing rail while manage
  mode occupied two, so toggling the mode still re-wrapped every title.

Both now carry regression tests that measure the laid-out result — panel
height against row pitch, and the title's box across both modes — and each
was verified to fail when its fix is reverted. The earlier tests passed
either way, which is exactly how the defects shipped.

Also in this round:

- Section eyebrows drop to medium emphasis and row status to low, so
  title, relationship and status are three distinct levels rather than
  three roles tied at #FFFFFF.
- The blocked-by chip is down to two amber marks; its chevron is neutral,
  since "go here" is not part of the blocked semantic.
- The chip says "Blocked by N tasks", matching the section it navigates to
  and the option that created the link. "Waiting on" solved two red pills
  shouting the same word — a visual problem, since solved visually by the
  outline treatment.
- "Create new linked task" moves off a raw Material AlertDialog onto the
  shared modal, the last surface in the feature outside the design system.
- The edit modal's counterpart line loses the leading ellipsis that read as
  truncation, becoming a caption label over the title at high emphasis.
- Manage mode keeps the row status: curating links is when knowing a
  blocker is already done matters most.

Round 10's hierarchy review found the emphasis ladder only held on desktop.
The wide layout styles the status label directly, but the narrow one drops it
into the list item's subtitle slot, which paints medium emphasis — the exact
tie with the section eyebrow the change was meant to break. Overridden via
spans so both layouts rank the same three roles.

Two of the geometry tests were measuring the wrong thing: WidgetTestBench
hands its child tight constraints, so a SizedBox width is silently ignored
and every case rendered at the full 800pt surface. The rail test still caught
its regression by luck of the layout; the emphasis test would have passed
against either implementation. Both now constrain through Align and were
checked to fail when their fix is reverted.

Also:

- Picker result rows carry a trailing link glyph. They are one tap from the
  card's rows and read identically — same glyph, title and status — while
  tapping here creates a link and tapping there opens the task. Nothing on
  the row said which.
- Manage mode ranks its verb above its escape hatch: edit takes medium
  emphasis and unlink drops to low. The confirmation modal is what makes
  unlinking safe; painting it as the brightest mark on the row only drew the
  eye toward the destructive action.
- Section eyebrows clear the following row's own top padding and line
  leading, so proximity binds each label to its rows instead of stranding it
  between two groups.

Opening a blocker from the multi-blocker sheet navigated while the sheet was
still up, so on a desktop layout the pane changed behind the barrier and the
user was left looking at an unchanged sheet. The only tap in the feature that
produced no visible result. The sheet now records the pick, closes, and
navigates from the caller's context; LinkedTaskRow takes an onOpen override
for surfaces that must dismiss themselves first.

The trailing rail had width parity but not height: the action pair is step8
tall while the browse chevron reserved no height at all, so desktop rows grew
20pt and the card jumped taller whenever Manage turned on. Both axes are
reserved now, the actions are pinned to the box rather than merely bounded by
it, and a height regression test covers the axis the width test could not
see — verified to fail on revert.

Also: the two modal bodies that reserve the glass footer no longer add a
spacing step on top of it, which had left the edit modal bottom-heavy against
its own top inset.

Not done, deliberately. Constraining the header's overflow button to the same
rail as the row actions is a real alignment win, but Material derives the
popup menu's width from the button's box, so the menu's own items overflow.
Not worth trading a working menu for 8pt of alignment.

The search field painted Material's default lilac caret and selection — the
only colour on any of these surfaces that belonged to no token. Set from the
interactive token, which fixes it everywhere the field is used.

Amber now marks only the ⊘ glyph, never the eyebrow label. The label's orange
sat one hue away from an On Hold row's own orange: close enough to read as
the same thing while meaning something else. The glyph carries the semantic
in a shape no status glyph uses.

Also:

- One colour for the link glyph. Emptying the card recoloured its own icon,
  because the empty state and the header painted the same Icons.add_link from
  two different tokens.
- Three hand-rolled Material Dividers become DesignSystemDivider, the picker's
  CircularProgressIndicator becomes DesignSystemSpinner, and the last raw
  `size: 16` moves onto the spacing scale.
- Modal titles drop the ellipsis their opening control keeps. "…" promises
  further UI; the modal is that UI, so the title should not promise it again.
  The link modal's top third had stacked three ellipses where only one — the
  "This task…" sentence stem — was doing real work.
- The link modal is bounded at the picker's own measure, so opening the
  relation list takes space from the results rather than pushing the sheet to
  the full height of the screen. It does not stop the sheet resizing outright:
  the shared dropdown expands inline rather than over its host, and changing
  that is a change to every dropdown in the app.

…dings

Full-text matches were intersected against the 200 open tasks the picker had
prefetched, so on a larger backlog a task you searched for by name came back
as "No tasks found" — the picker stating confidently that it isn't there.
Matches outside the window are now fetched by id and filtered to open tasks.
Two tests cover it, and the first was verified to fail against the old code.

Also from round 11:

- The card's add-link action is a labelled button. As a bare glyph it was the
  card's only creative action and its least legible control, and it gave a
  low-vision reader nothing to read once the card had any links.
- Save in the edit sheet is inert until the relation actually changes. An
  identity round-trip bumped the vector clock and enqueued a sync message for
  an edit nobody made. The tests that asserted that write now assert the
  direction mapping in both directions instead, which is what they were really
  for.
- The blocker picker carries the same amber ⊘ as the section and header chip
  that read the relationship back. It was the only surface in the feature
  stating its meaning with a sentence and nothing else, on the highest
  consequence tap in the set.
- The picker's link glyph is no longer dimmer than the status text beside it;
  it is the only thing distinguishing a row that commits from one that
  navigates.
- French and Romanian task statuses were all caps while every other language
  was sentence case, which shouted the quietest rank in the row.
- The edit sheet's two field labels share one ink, its read-only value drops
  below the editable one, and the link modal's inter-field gap matches the gap
  around it.

Round 11 made full-text matches from outside the prefetch window visible, but
the pick handler still resolved ids against the prefetched list alone, so
tapping one of those rows threw StateError. The test I wrote for that fix
asserted the row rendered and never tapped it — which is exactly how a fix
for "the picker denies a task that exists" shipped as "the picker crashes on
that task instead". Rows and picks now read one candidate pool, and the test
taps the row and asserts the link is written. Verified to fail on revert.

Also from round 12:

- The blocker picker's title is back on the shared modal chrome. Putting the
  ⊘ in the title bar made it compete for a width the close button dictates,
  clipping "What's blocking this?" mid-word on phone in English, before any
  longer locale. The mark moves into the body as a section header, where it
  also names the relation the pick will write.
- The disclosure chevron leads the title it discloses. Sitting against the
  Link button it read as that button's caret, so a tap aimed at a relation
  menu collapsed the card — and it disappeared in manage mode while the
  collapse gesture stayed live.
- DesignSystemListItem gains a subtitle-emphasis contract, plumbed through
  the selection row and picker item. That removes this feature's only raw
  TextStyle and settles the status metadata at one ink whether it is read on
  the card or in the picker one tap later.
- The empty card's link glyph takes the interactive accent the populated
  card's Link button carries; emphasis should be highest where the user has
  nothing yet. The comment claiming they already matched was wrong.
- Two measured proximity ties: the tight-top eyebrow gap and the edit sheet's
  group gap.

`getTasks` short-circuits an empty `categoryIds` to `WHERE 1 = 0`, and the
picker passed `categoryIds: []` meaning "no filter". So the browse list
returned nothing at all in production, and every test passed because they
mock `getTasks` and never exercise the query builder. JournalQueryRunner
already expands the same way for the same reason; the picker now does too,
with a test that pins the contract and fails against the old call.

Also from round 13:

- A task that has links no longer shows "you have no links" on cold open.
  The provider resolves two DB round trips, and treating "no value yet" as
  "no links" put a definitive empty CTA in front of every populated card for
  the length of that load.
- Finished tasks can be linked. "Follows up on", "Duplicates", "Fixes" and
  "Supersedes" all routinely reference work that is already done, and the
  modal reported "No tasks found" for a task the user can see elsewhere in
  the app. The blocker picker keeps the open-only filter, since a finished
  task cannot block anything.
- Undo is awaited and reports failure, matching the unlink path it mirrors.
  Fire-and-forget made a failed undo indistinguishable from a successful one,
  on the one control whose whole job is taking something back. Its window is
  also longer than the 4s default: it is the only recovery from a link that
  commits on a single tap, and it arrives at the bottom of the window while
  the eye is still on the modal that was dismissed.
- Manage-mode actions are 48pt, the Material minimum and above Apple's 44.
  They were 40, in an adjacent pair, above a row that is itself tappable —
  and a comment claimed 40 was the design system's minimum interactive
  target, which it is not. Pinned by a test.

A plain link reads the same from either end, but `inverse` still took part in
DirectedRelation's identity, so an existing incoming plain link never matched
the one the picker offers. The picker re-offered a task already listed on the
card, and taking it wrote a second identical row, inflated the count badge,
and cost two confirmations to undo one relationship. `inverse` is now
normalised away on symmetric relations, with tests for both halves: the plain
link is excluded from either side, while "blocks" and "is blocked by" stay
independently offerable, since a pair may legitimately hold both.

Failure messages also stopped blaming a cause that cannot apply. Only a
blocking edge can fail the cycle guard, but every rejected link — "Relates
to", "Duplicates", anything — reported "This would create a blocking cycle"
and offered a remedy that would not help.

Also from round 14:

- Failures no longer wear the success colour. A rejected cycle, a failed
  unlink, a failed retype and a failed undo all arrived on the same Material
  slab as the "link created" confirmation, tinted by whichever scheme the user
  picked. One shared helper puts them on the alert colour; it takes resolved
  tokens rather than a context because every caller reports after awaiting the
  write that failed.
- Manage mode keeps an add path. Its header slot is taken by Done, which left
  no way to add a link while curating them — the state where wanting one is
  most likely.
- The blocker picker's relation statement sits on the picker's own rail
  instead of 25pt left of the rows it labels, and has room above the search
  field.
- The empty card's explanation is quieter than the action it explains and
  survives three lines, so the longer locales stop ellipsizing away the
  examples that teach what "link" means here.

Both the confirmation and the four failure messages were hand-built SnackBars.
They now go through DesignSystemToast: success tone with the undo action and a
countdown strip that makes the remaining window visible, error tone for
failures. That also retires a bug I introduced last round — a hand-picked ink
on the alert fill rendered near-black on dark red in light theme, on the one
message the flow shows when it refuses. The component owns its own on-colour.

A rejected retype no longer says "please try again" for a cause that makes
retrying pointless. The cycle guard is the only thing that rejects a retype
and it can only reject a blocks edge, so that case now names the cycle. Same
defect class as the create path, closed on the last surface holding it.

Modal titles fit one line at phone width in every shipped locale. German ran
33 characters against a 230pt title box — half again the English length — and
no screenshot review sees a non-English string, so this was invisible for
fifteen rounds. A new test measures all four titles in all six locales against
the real font and fails on a second line; it caught one more overflow after
the first pass of shortening.

Deliberately not done: halving the browse-mode trailing reserve. It would give
back 48pt of measure on phone rows, but it re-creates the width mismatch
between browse and manage mode that rounds 10, 12 and 13 each independently
flagged, and that a regression test now pins.

At text scale 1.3 through 2.0 the header Row overflowed by up to 158pt on a
390pt phone, clipping the Link action and dropping the overflow menu entirely
— which in manage mode is one of only two ways out. Large text is a supported
configuration, not an edge case. The card's own name is now the part that
gives: flexible and ellipsized, so the controls survive.

The test that should have caught this was measuring an 800pt card. The bench
hands its child tight constraints, so the width it was given was ignored — the
same harness flaw that made two geometry tests no-ops in round 10. The helper
now constrains properly, and the scale cases were checked failing on revert.

Also:

- The undo toast dismisses itself. It removed the link but left the
  confirmation, its live Undo control and a draining countdown on screen for
  another ten seconds, asserting a link that no longer existed and offering to
  remove it again.
- The linked task's name moved out of the toast's one-line title into its
  description slot. Relation plus a real title ran past one line, so the
  message ellipsized away the one fact it exists to record — on a flow that
  commits without confirming.
- Modal titles are bounded to one ellipsized line in the shared component.
  A title needing a second line silently reflowed a fixed-height bar; now it
  degrades predictably, and the route semantics still announce it in full.
- The title-fit guard was measuring the wrong thing: it omitted the shipped
  letterSpacing, under-reporting by ~3pt on ~5pt of headroom. Corrected, and
  it promptly failed until the remaining long titles were shortened.

Last round's accessibility fix made the header title Flexible while leaving the
Spacer beside it. Both carry flex 1, so they split the free space evenly: the
title gave way to "Linked T…" at normal text size while most of the row sat
empty. Four reviewers found it independently. The title and its badge now share
one Expanded and the Spacer is gone, so the group takes whatever the trailing
controls do not need and gives way only when there is genuinely nothing left.

The guard added last round asserted the header does not overflow and nothing
about whether the title survived, which is how this shipped under a green
suite. It now also asserts the title is not truncated at default text size, in
both modes — and it needed two corrections before it meant anything: the suite
renders with a fallback font far wider than Inter, and it measured the card at
the full viewport width rather than the ~357pt it actually gets inside the
page's padding. Verified failing on revert with the real numbers.

Also: the unlink action is `Icons.link_off`. It was `Icons.close_rounded` —
which is StatusGlyph's own icon for Rejected, so a rejected row in manage mode
drew the same mark twice with two different meanings. The edit action already
avoids exactly this collision for Groomed.

Last round's fix was verified in English and shipped truncated in German and
Spanish, whose card title and action label are both markedly longer. The action
now gives up its word before the card gives up its name: the header measures
both strings and falls back to an icon-only link button, still named by tooltip
and semantics, where the worded one will not fit. Which locales collide is a
property of the strings, so it is measured rather than guessed at a breakpoint.

The guard loops over all six shipped locales now. It catches exactly de and es
when the worded button is forced — the two locales the review named — and this
is the third guard in this feature that only meant something after being made
to fail first.

Also:

- The empty card words both creative actions. Creating a linked task was
  reachable only through an unlabelled overflow glyph holding a single item,
  on the one screen whose job is to teach what the card is for.
- The unlink confirmation names the task. The rows it is reached from carry
  two faint glyphs each, so "this task" could not say which link was hit — on
  the feature's only irreversible action.

Its chevron sat bare, on a 40pt target this feature elsewhere calls
under-minimum, and the card's trailing rail jumped 18pt the instant it gained
its first link. Both rows now use the same reserved rail every populated row
uses, extracted so the two cannot drift apart again. Spacing called this the
one thing it would hold a ship for.

The empty card's action titles also take two lines, like the task titles beside
them elsewhere on the card. A one-line cap ate the verb in verb-final German —
on labels I promoted out of the overflow menu last round precisely so they
could be read.

Also:

- The blocked-by accent drops once every blocker is closed. A closed blocker
  releases the dependent, so amber there asserted something the header had
  already retracted, on the feature's only semantic colour.
- The header's fit allowance comes from DesignSystemButton's own size spec
  rather than a round number roughly double the button's real chrome — a
  guess that was taking the label away from locales with room for it.
- The icon-only fallback is a DesignSystemButton rather than a raw Material
  IconButton, so a user's locale no longer decides whether they get a
  design-system control.
- The empty card has no overflow menu: its only item is now a worded row in
  the body, and an unlabelled glyph was the loudest mark on the screen that
  has to teach the feature.

The change was written but never saved: the script that applied it hit a
failing assertion on a later edit and threw before writing the file, and I
re-ran only the other half. The commit message and the tracker note both said
it was done. It was not — the 96pt stand-in survived, taking the Link label
away from Romanian and Czech, which have room for it.

The allowance now derives from DesignSystemButtonSize.small's own spec
(2*step3 + lineHeight.subtitle2 + step2 = 40), so ro and cs keep their word
while de and es stay correctly icon-only.

The locale guard now asserts both directions — that the worded label is
present where it fits and absent where it does not. Asserting only that the
title survives is what let an allowance 2.4x too large pass through two rounds
and one "fixed" claim. Verified failing on revert, naming exactly ro and cs.