v1.5.3

v1.5.3 (2026-10-09)

• A current screenshot replaces the v0.1.0-era one: the welcome tour
  as it ships today - checklists, spoken and reverse percents, the
  breathing answers column - now embedded in the README and shipped
  in the AppStream metadata.

v1.5.2

v1.5.2 (2026-10-08)

• Fixed: pressing Enter above a block of line references could
  leave them stale or drag one backward. On the Mortgage template,
  two Enters above the headline turned &25/&26/&27/&28 into
  &26/&27/&28/&27 instead of &27/&28/&29/&30 - the summary lines
  then answered with the wrong derivation rows. Root cause: when
  references target CONSECUTIVE lines, the renumbering plan chains
  (each relabel's new label is the next mark's existing key), and
  the plan was applied sequentially into a label-keyed table, so
  every step silently evicted the next mark. The plan is now applied
  as the simultaneous rename it always meant to be, through a shared
  helper both the app and the tests exercise. Found by Brandon on
  the Mortgage template the same night it shipped; diagnosed by a
  three-agent research wave that converged on the same four lines of
  page.rs from three directions.

v1.5.1

v1.5.1 (2026-10-08)

• Typing and answering land in the same frame now. The evaluation
  idle ran at a priority a full frame behind the echo paint (the
  structural cause of a slight split between typing and seeing),
  and per-burst cost had crept to ~90 ms on the welcome sheet.
  Three fixes: evaluation moved onto the view's frame clock (update
  phase, so answers paint with the text), knows_unit probes are
  memoized per process (the stress sheet's 24 prose words cost
  33 ms in master-clone probes alone; now 0.12 ms for the whole
  sheet), and the re-pass flag only fires when a missed reference
  points at a line that can still answer (dangling refs stop
  forcing a wasted confirm pass). The outline menu is no longer
  rebuilt per keystroke, and error highlights skip rebuilds when
  the error set is unchanged.

v1.5.0

v1.5.0 (2026-10-08)

• The template menu grows into a small library: a Mortgage
  walkthrough whose top-of-sheet headline cites its own derivation
  through forward references (monthly payment, everything you pay,
  what the bank keeps, the payoff date - all resolved from the
  algebra below), and a Freelance invoice where the relative
  percent forms do the discount and the VAT (subtotal - 5%,
  then * 21%).
• The welcome sheet is a richer first impression: a compound-
  interest line worth keeping, and a Where-to-go-next section
  pointing at the template menu (plain text by design - the menu is
  one click away, and clickable cross-document links would drag in
  file resolution for no calculation gain).
• The debug mode: evaluate_sheet_stats and quire-cli --stats
  report the fixed-point pass count and whether a sheet ever
  exhausted the pass cap; QUIRE_DEBUG=1 makes the app eprint
  per-burst diagnostics (lines, passes, cap, elapsed time, column
  width, over-cap answers, error count). The full run over all 21
  shipped sheets: everything settles, zero cap hits, fastest path
  everywhere forward references are absent.
• Template-authoring lessons now pinned by the suite: two-word
  names never bind as assignments, total cannot be an assignment
  name (it is the keyword), and a template's forward references
  must be counted against the final layout.

v1.4.0

v1.4.0 (2026-10-08)

• The caret's line carries a subtle band: a full-height background
  tag in the palette's raised card tone, following every caret move
  and repaint on theme flips. Implemented as a plain TextTag after
  the reserved snapshot-layer hook turned out to be unexposed in
  sourceview5 0.11 - zero custom paint, and tag applications sit
  outside the undo stream, so moving around never pollutes undo
  history.
• quire-cli joins the workspace: a sheet in, its answers column
  out. Plain text mirrors the sheet's shape with the answer column
  right-aligned; --json emits one object per line with the raw
  text and the answer. Stdlib-only by house rules, and it rides the
  same engine, so sentences-with-units work at the terminal too
  (the stock measures 2 mol/L answers 2 molar).
• The Error Zoo's Units that fight section now teaches what it
  exhibits: adding Mass to Length fails with the clash named, while
  multiplying units composes (2 kg * 3 m is 6 kg*m - torque
  territory, not a bug).

v1.3.1

v1.3.1 (2026-10-08)

• The answers column breathes: it sizes to its widest visible
  answer (between 168px and 320px) instead of a fixed width that
  clipped every date and any large number. Measured with the same
  widget-font layout the cells draw with, recomputed once per
  evaluation burst; past the 320px ceiling answers ellipsize and
  the hover tooltip carries the full text, as before. Verified on
  the desktop: the tour's date cells render in full.

v1.3.0

v1.3.0 (2026-10-08)

• Forward line references resolve: &67 on line 65 answers with
  line 67's result. The spec always promised references to "any
  line, above or below", but the engine evaluated one strict
  top-down pass, so anything pointing downward stayed blank
  forever. The sheet now re-evaluates - whole-sheet passes seeded
  with the previous pass's results, stopping when the rendered
  answers stop changing (a cap of 8 bounds it) - so chains resolve
  one link per pass and a summary line at the top of a section can
  cite the derivation at the bottom. Cycles never ignite: no line
  in a circular chain is ever seeded, so a cycle stays blank
  rather than showing a wrong number, and sheets without forward
  references still cost exactly one pass. Unit lines resolve
  forward too (&2 kg * 2 over 3 kg answers 6 kg). The starter
  sheet's Line references section demos it live: a backward ref and
  a forward ref, both answering.
• A research wave (three agents: engine design, reference-app
  semantics, app blast radius) drove the design. The interesting
  negative result: iteration-as-cycles (a Newton's-method guess
  referencing itself) does NOT work under re-pass evaluation -
  cycles never ignite and seeds reset each pass - so converging
  iterations stay with recursive functions, where they are already
  sound.

v1.2.0

v1.2.0 (2026-10-08)

• Sentences carry units through the engine: the mixed-line skeleton
  evaluates through the unit engine when it holds one - the stock solution measures 2 mol/L answers 2 molar, the sample weighs 0.101 g answers 0.101 g, trailing periods and commas are glue.
  The surviving words must sit the way numbers and units sit in
  math: a number may lean on its unit, units may stand together,
  and a bound name next to anything stays prose. This is the
  feature that makes a lab notebook read like a lab notebook.
• The Complex Sample is rewritten as that lab notebook: sentences
  carry the numbers, derivations keep their own lines when they
  deserve the room, and everything referenced later is bound with
  = in the paper's notation. Building it surfaced two name
  collisions (numbat claims mean, std, and the word second
  is the seconds unit) and the notebook words around them.

v1.1.0

v1.1.0 (2026-10-08)

• Two dev torture sheets join the template menu, built to stress and
  confuse the renderer on purpose. The Stress Test loads layout
  edges: wrapped long lines, a 45-character answer next to short
  ones, region-alignment padding, tab-led lines, blank forward
  references, very long names and tag runs, and unicode prose. The
  Error Zoo loads every contained failure family at once - division
  by zero, out-of-range powers, malformed numbers, unbalanced
  parens, unknown functions and names, dead references, recursion
  depth, arity clashes, fighting units, keyword misuse, and hostile
  text - each error caged to its own line while the sheet keeps
  computing past the last cage. If a cell overlaps, an error leaks,
  or anything misbehaves, the sheet names the finding.
• Two better error messages fell out of building the zoo: unit
  dimension clashes now surface the engine's own wording (3 kg + 5 m names Mass versus Length instead of "unknown name kg"), and
  a bare period word in dead algebra reads as the unit it is, with
  numbat's suggested fix attached.
• Finding, documented: the recursion cap (200 levels) costs about
  10 KB of stack per level in debug builds - fine on the app's
  8 MB main thread, over the budget of Rust's default 2 MB test
  threads, so heavy template computes in the suite run under a
  big-stack harness.

v1.0.0

v1.0.0 (2026-10-08)

• 1.0: Phases 6 through 9 are complete and the Soulver-depth surface
  is whole - arithmetic with implicit multiplication, every
  percentage form including the reverse questions in words,
  functions with recursion and dimension-checked unit paths,
  self-updating line references, dates, recurring amounts, currency,
  units, math functions, tags, mixed lines, completion, and the
  template menu.
• Ctrl+click a math line and it explains itself: a popover shows the
  line's operations bottom-up, each as operands = value, evaluated
  in full sheet context so variables, references, and function
  bodies resolve. A relative percent reads in the sheet's words
  (200 + 15% (of 200) = 230). Unit lines and prose do not break
  down.
• Task lists: a list line may open with a GFM checkbox
  (- [ ] plan the trip); Ctrl+clicking the box toggles it as a
  plain user edit undo reverses. Checkboxes never enter the math.
• The region model: within each heading region, scalar answers
  align on the decimal point - integer answers pad so their implied
  dot sits where the region's fractional dots are. Quantities,
  errors, hex and bin keep their shapes.
• Definition sheets read as a table: the Outline popover lists every
  bound name beneath the headings; picking one lands on its line,
  with the answers column as the values and trailing comments as
  the notes.
• The template menu grows to four: budget (now with budget-vs-
  actual paired *_actual variables and a variance section where
  positive means under budget), portfolio, trip (dates for the
  countdown, rates for per-day costs, currency for the spend), and
  a Complex Sample - a scientist's worksheet running the ideal gas
  law, Arrhenius activation energy, triplicate statistics with a
  sample standard deviation, a dilution, Henderson-Hasselbalch pH,
  projectile range, and combinatorics through recursive functions.
• Two engine primitives the new sheets demanded: the prelude's math
  functions route to the unit engine (sqrt(144) is 12,
  sin(30 deg) is 0.5 - before, log(100) silently answered 100
  through the mixed skeleton), and quantity totals finish inside
  expressions (whole = total over money answers 990 EUR). Date
  words call their functions mid-line too: today + 30 days.