# Changelog — numodel-bundle

All notable changes to the bundle are documented in this file. It
covers all modules: entries are grouped per release and, within a
release, per module. Every component of the bundle carries the same
version and is released together, so a module that did not change in a
given release is listed as a version-sync entry.

The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
and the bundle adheres to [Semantic Versioning](https://semver.org/).

## [Unreleased]

## [0.12.0] — 2026-10-28

numodel-coach can write Coach's graphical model, in the edu or tight
style of `\graphicmodel`, with switching to the text view chosen by
`switch`. `\graphicmodel` now draws every flow the Coach export makes,
and `\textmodel` translates more l3fp functions. One change in
behaviour: a sum in parentheses, `(A + B)*dt`, is one flow in the
diagram; only a difference is split.

### numodel

#### Added
- `\textmodel` and the Coach export translate `trunc(x)` (as
  `Teken(x)*Entier(Abs(x))`), `round(x)`, `round(x,n)` (as
  `Round(x*10^n)/10^n`; Coach's `Round` has one argument), `fact(x)`
  (`Fac`), `logb(x)` (`Entier(Log(Abs(x)))`) and the constant `deg`
  (π/180); they used to appear untranslated in the rule table. New
  keywords `round`, `fact` and `logten` in the language files; a
  drop-in `numodel-<LANG>.def` without them falls back to `ROUND`,
  `FACT`, `LOG10`. Regression test `m014-functions`.
- The Lua engine computes `fact` and `logb` itself (as l3fp: `fact`
  for whole numbers, `logb` the decimal exponent); a model using them
  no longer falls back to `\fp_eval`.

#### Changed
- Development only: `l3build check` also runs the Lua unit tests
  (`tests/run.lua`) and stops when one fails. Eight layout tests had
  been failing unnoticed since the placement fix of 25 May 2026
  (constants on row 1 when a model has no auxiliary variables); their
  expectations are updated, the code is unchanged.

#### Fixed
- `\graphicmodel` draws the flows it used to leave out, the same
  flows the Coach export makes: an outflow whose rate is a stock
  (`h := h - v*dt`, the falling ball) gets a valve after the stock
  with a causal arrow from `v` -- before, `v` was even placed as if it
  were that valve; a flow without a variable (`V := V + 0.5*dt`) or
  with several (`(r + i)*dt`, `k/m*x*dt`) gets a valve without a name
  and an arrow from each variable. The stock itself counts: growth
  `N := N + g*N*dt` and decay `A := A - ln(2)/t_half*A*dt` get an
  unnamed valve with arrows from `g` and from `N` (the feedback loop);
  before, the constant was drawn as the valve, as if the flow were `g`. A sum in parentheses is one flow;
  only a difference `(A - B)*C` is still split into an inflow and an
  outflow. The valve of an inflow whose rate is a stock
  (`s := s + v*dt`) has no name any more: labelled with the stock it
  read as if that stock were the valve. Lua tests extended.
- `numodel.nice_range` (the axis range numodel-coach gives a variable
  no diagram has drawn) failed with a Lua error on a series that ran
  away to infinity; such a variable now keeps Coach's default range.
- `\textmodel` inside a box (minipage, tcolorbox, exam's framed
  `solution`) no longer splits its table with "Continued on next page"
  in the middle of the box. The default `longtblr` cannot break a page
  there but still measured the space left on the page; inside a box
  `\textmodel` now uses a centred `tblr`.

#### Documentation
- Manual checked against the package: `\mrule` keys listed; `\mstop`
  without stop condition is an error (not a warning), a second one
  replaces it; `\textmodel` key `blanks`; `\graphicmodel` keys
  `valve-style`, `flowarrow-style`, `flowarrow-cloud-tip`; shapes of
  `aux` (circle) and `system` (not drawn); constants on the middle row
  without auxiliaries; `\mcoords` output; the `maxiter` warning;
  `numodel-FPEVAL.def` and `tabularray` among the files and
  dependencies. The paragraph on the global `rowsep=0pt`, removed in
  0.11.0, is gone. New example for `\tablemodel`.
- "Changing and copying models" moved to the section on multiple
  models, after its introduction. The pupil's and answer version
  (`blank`) has a section of its own, "Exercises", with the alias keys
  as other replacements; it used to sit under "Generated accessors". The expression reference marks the
  newly translated functions as supported.
- `\constnode`, `\flowarrow` and the other drawing macros that
  `\graphicmodel` writes are marked internal in the source: drawing a
  Forrester diagram by hand is not supported.
- Minimum TeX Live raised to 2024 in the manuals and README:
  numodel-plot needs siunitx v3.3.8 (November 2023).

### numodel-coach

#### Changed
- `switch` takes `true`, `false` or `auto` (the default): `auto` turns
  switching between text and graphical view on for a graphical model
  and off for a text model, as before for text models.

#### Added
- `\coachmodel[mode=graphic]` (also `\coachsetup{mode=graphic}`):
  the model as Coach's graphical model, opened in the graphical view
  with switching to the text view on. The time rule gives Coach's time
  and step; a stock rule `X := X ± term·dt` a stock with one flow per
  term (`-Q·dt` in one stock and `+Q·dt` in another one flow between
  them), named `Stroom_1`, ... with the rate as formula; with
  `graphic-style=tight` a flow that is one variable used nowhere else
  (an auxiliary, also a conditional one, or a constant) is that
  variable, with its name shown, as `diagram-style=tight` draws it;
  other rules auxiliaries (`?:` as Coach's condition),
  variables without rule constants; every reference a relation.
  Objects and valves are placed as `\graphicmodel` places them in the
  matching style (`edu`, the default, or `tight`). Pupils may switch to
  the text view unless `switch=false`. A model that cannot be
  drawn (a stock rule with a term without dt, ...) is written as a
  text model with a warning. Regression test `c006-graphic`.

### numodel-plot

#### Documentation
- The manual describes the `equal` key (added in 0.11.0) and the axis
  label without unit.

## [0.11.0] — 2026-10-11

Driven by writing a full teaching module with the bundle. The default
engine is now Lua, some hundreds of times faster than `\fp_eval`. New:
`\tablemodel` for tables of computed values, pupil's and answer
versions from one declaration (`blank`), model variants with
`\copymodel`, `\mstepat`, more `\diagrammodel` keys, and angles in
degrees for Coach. Note two changes in behaviour: the default engine
(`engine=fpeval` restores the previous decimal computation), and the
row spacing of tabularray tables (numodel no longer sets it for the
whole document).

### numodel

#### Added
- `\tablemodel[keys]{Name,Name,...}`: a table of computed values, one
  column per variable (headed "symbol (unit)"), one row per step
  (`steps=0-5` or a list), rounded to each variable's significant
  figures. `row=state` shows the state of step n with the auxiliary
  values computed from it (recorded in step n+1), the row a pupil
  computes by hand. `blank` leaves the cells empty in the pupil's
  version (`blanks=open`) except the first column and the first
  `given` rows. `num={Name={...}}` passes siunitx options per column,
  `table={...}` tabularray options. Regression test `m013-tablemodel`.
- `\mstepat{<Name>}{<Var>}{<value>}` (and `\mstepatp` with an explicit
  prefix): the value of a variable at the first step where another
  variable -- usually the time -- has reached a value, so the text no
  longer depends on the time step: `\mstepat{V}{T}{1}` is v at t = 1 s
  for any Δt. Round-off of an accumulated time step is tolerated; a
  value that is never reached is an error.
- `\diagrammodel[legend=true]`: a legend for a single series, with the
  text from the keyword table of the syntax ("modelpunt" for NL,
  "model point" otherwise; a drop-in `.def` without the new
  `legend_point` keyword falls back to "model point").

- `start-format=input` (`\numodelsetup`, package option, or per
  variable as `\mvar[start-format=...]`): start values are shown
  exactly as typed in `\mvar` -- `h = 3000 m` instead of
  `h = 3,00·10³ m`, `l = 600 m` instead of `6,0·10² m` -- in the
  initial values of `\textmodel` and in `\<prefix><Name>num`/`qty`. The
  default `rounded` keeps the previous behaviour. Only a plain number
  has a typed form; a start value given as an expression stays
  rounded. The setting is read when the value is shown, so it also
  applies to variables declared before it. Regression test
  `m008-start-format`.

- `\diagrammodel` keys for teaching material: `float=false` sets the
  diagram where it stands (centred, still numbered as a figure), so it
  works in a minipage, a tcolorbox or an exam `solution`;
  `caption={...}` replaces the automatic caption and `nocaption` drops
  it (and the label); `legend={...}` gives one label per series;
  `style=marks|line|both`; `xcmmax`/`ycmmax` limit the axis size of
  this diagram. Regression test `m009-diagram-keys`.
- Pupil's and answer version from one declaration: the key `blank` on
  `\mvar`, `\mrule` (also `\mrule*`) and `\mstop` marks what the pupil
  has to supply. With `blanks=open` (`\numodelsetup`, package option or
  `\textmodel[blanks=...]`; the default) `\textmodel` shows
  `F_z = ⋯`, a conditional rule collapses to that single row, and the
  stop row reads `IF ⋯ THEN STOP ENDIF`; with `blanks=filled` all is
  shown in full. `\computemodel` always uses the full rules.
  `numodel.plaintext` follows the setting (`' F_z := ...` comments when
  open). Regression test `m010-blanks`.
- `\numodelsetup{table-style={...}}` (also a package option):
  document-wide tabularray options for `\tablemodel`, before a table's
  own `table={...}`. The values of `table-style`, `table` and `num` are
  expanded once, so a macro holding the options works (tabularray used
  to report "Unknown inner key name").
- A Lua engine for `\computemodel`: `engine=lua` (`\numodelsetup`,
  package option, or `\computemodel[engine=...]`) compiles the rules --
  with the parser of the plain-text rendering -- into Lua functions and
  runs the model loop in Lua. The H3 parachute model (8131 steps) takes
  0.03 s instead of 9.3 s; the answer version of the NLT module
  "Modelleren" compiles in 23 s instead of 45 s, with an identical
  text. Rules stay in l3fp syntax; TeX macros other than the model's
  variables are expanded once, as `\fp_eval` would see them. A model
  that cannot be translated falls back to `\fp_eval` (info message).
  Comparisons use a relative tolerance of 1e-9, so binary rounding
  (ten steps of 0.1 give 0.99999999999999989) does not move a stop
  condition by a step; final values are returned to TeX with 12
  significant digits. `engine=compare` runs both engines, keeps the
  `\fp_eval` results and reports whether they agree. All 25 models of
  the examples and tests agree. Regression test `m012-engine`.
- `\diagrammodel` plots at most `plot-points` points per series
  (`\numodelsetup`, default 1000; `points=` per diagram; 0 for all).
  A longer series is thinned by distance in the diagram, so flat
  stretches lose overlapping marks while every step of a fast change
  stays visible; no point is moved or invented, and only the plot is
  affected. The 8131-point parachute diagram takes 0.24 s instead of
  2.3 s and looks the same. Lua test `test_coords`.
- Changing and copying models. `\copymodel[Name=value,
  blank={...}]{source}{copy}` creates a model with all declarations of
  source, references renamed (`\sprongV` → `\foutV`), start values and
  blank rules adjusted; the copy becomes the current model. Changes:
  `\mrule[replace]` replaces the one rule for its target (an error when
  there is none or several); `before=`/`after=` place a new rule
  relative to another rule; `[add]` adds explicitly; `\mremoverule`
  removes the rules for a target; a second `\mstop` replaces the stop
  condition. Every model keeps a log of its declarations; a change
  rebuilds the model from it, so table, computation and exports stay
  consistent. A copy of the H3 parachute model computes exactly like
  the same model written out in full. Regression test
  `m011-copymodel`.
- `\diagrammodel[axis={...}, plot={...}, series={{...},{...}}]` passes
  pgfplots options on unchanged, after numodel's own: for the axis
  (legend position and style, ...), for every series (mark size, ...),
  and per series in series order. numodel keeps its own keys for what
  concerns the model and the layout and leaves drawing details to
  pgfplots, instead of growing a key for each of them.
- `\numodelsetup{angles=radians|degrees}` (also a package option): the
  angle unit of the program the model is written for. With `degrees`
  `\textmodel` shows `sind(x)` as `Sin(x)` and `sin(x)` as
  `Sin(180/π·x)`, `asind(x)` as `Arcsin(x)` and `asin(x)` as
  `π/180·Arcsin(x)`, as Coach in degree mode needs them; the calculation
  does not change. Set in the document body it warns, because a rule's
  text is fixed when the rule is declared.
- `\diagrammodel[equal]`: equal scales on both axes (for a path y
  against x), taken into account when the axis ranges and ticks are
  chosen. `axis={axis equal}` stretched the axis after numodel had
  chosen the ticks, so the tick labels overlapped.
- `numodel.nice_range(lo, hi, cmmax)`: the axis range `\calcplotdims`
  would give a series. `numodel.get_model` uses it for a variable that
  no `\diagrammodel` has drawn, so every computed variable has an axis
  range (`axis_source` tells `"diagram"` from `"computed"`).

#### Changed
- **The default engine is `lua`.** `\computemodel` compiles the rules
  to Lua; `engine=fpeval` keeps the previous decimal computation with
  `\fp_eval` (exact decimal steps, 16 digits, identical to `\fpeval`
  elsewhere; see the manual on when that matters). Every test and
  example gives the same output with both engines.
- `\<prefix><Name>min`/`max` of a variable without start value skip
  step 0, where the variable is still 0 (nothing has computed it yet;
  Coach's first table row shows 0 too). An auxiliary `r = sqrt(x²+y²)`
  had minimum 0. Regression test `m007-names-steps-numbers`.
- `\mstepat` follows the direction in which the variable starts: from
  above the value it finds the first step at or below it. A falling
  quantity (`\mstepat{T}{Vy}{0}`, the top of a throw) used to give step
  0. Regression test `m007-names-steps-numbers`.
- numodel no longer sets `\SetTblrInner{rowsep=0pt}` for the whole
  document. It was meant for `\textmodel` but only reached `tblr`
  tables -- the document's own and `\tablemodel` -- never
  `\textmodel`'s default `longtblr`. Tables of the document and
  `\tablemodel` get tabularray's default row spacing again (use
  `table-style={rows={rowsep=0pt}}` for the previous look).
- Numbers handed back to TeX by `\mstep`, `\mstepat`, the `min`/`max`
  accessors, `\mcoords` and table cells are written with 12
  significant digits, as the final values already were, so binary
  noise (`0.30000000000000004`, `1.0`) never reaches the document.
- `\mvar` on a name the model already has now replaces the variable in
  place (with an info message). It used to add a second row to the
  initial values and keep both entries.
- `\mrule` added after the stop condition now goes before it; it used to
  appear after the stop row in `\textmodel`.
- `\diagrammodel` builds single and multiple series the same way; a
  multi-series legend now takes the display text unexpanded.
- A `\diagrammodel` with a single series no longer has a legend by
  default. It used to show a hard-coded English "model point".
- Decimal numbers inside a model rule are shown through siunitx `\num`,
  so `\textmodel` gives them the decimal marker of the syntax and shows
  them as typed: `k = 0,24` instead of `k = 0.24` with `syntax=NL`.
  Integers, and digits inside display names such as `x_1`, are left
  alone.
- Name collisions are detected precisely. A full name that already
  belongs to a variable of another model (prefix `para` + `gX` = prefix
  `parag` + `X`) is now an error. A full name that is also another
  variable's accessor (`T` next to `Tmax`, `V` next to `Vmin`) gives a
  warning naming both and the consequence, instead of the generic
  "being redefined"; documents relying on `T` next to `Tmax` keep
  compiling.

#### Fixed
- numodel no longer writes its trace output to the log: lines such as
  `MVAR start`, `BUILD_TABLE emit rule num` and `TEXTMODEL table` --
  left-over debugging, some 1800 lines for the NLT module "Modelleren",
  two thirds of its log. `\numodelsetup{debug=true}` brings them back;
  when off, their content is not even expanded.
- `numodel.lua` set the global `debug = true`, replacing Lua's standard
  `debug` library for the whole document (any package calling
  `debug.traceback` would have failed). Removed; it was unused.
- The `maxiter` warning lacked a space ("after 5iterations").
- `\diagrammodel[float=false]` no longer makes the caption package warn
  that `hypcap=true` is ignored: outside a float hypcap is switched off
  locally (only when caption is loaded).
- `\diagrammodel[float=false]` inside a list (`enumerate`, exam's
  `questions`) gave an overfull box: the caption was as wide as the
  text, not as the list's lines. Diagram and caption now sit in a
  minipage of `\linewidth`. Regression test `m009-diagram-keys`.

#### Documentation
- The manual said that the accessors `num`, `qty` and `pre` hold the
  current value; they hold the start value (as `\textmodel` needs),
  fixed by `\mvar`. Corrected, with a pointer to `\mstep`/`\mstepat`
  for computed values.
- `\<prefix>steps` counts step 0: a model with N samples went through
  N − 1 rounds.
- New paragraph "What a step contains": a step records the values
  before that round's rules run, so an auxiliary variable in step n
  was computed from the stocks of step n − 1. A table row "t, v,
  F_w(v), a" takes t and v from step n and F_w and a from step n + 1.
- The naming caveat describes the new collision checks.
- Regression test `m007-names-steps-numbers`.

### numodel-plot

#### Added
- `\numodelplotsetup{equal}` (`\diagrammodel[equal]` in numodel): the
  same amount per centimetre on both axes; `\calcplotdims` uses the
  larger of the two tick distances for both. Regression test
  `p001-calcplotdims`.

#### Fixed
- An axis without unit and without scaling got empty parentheses
  ("X ()"); now just "X". Regression test `p005-axislabels`.
- Regression test `p001-calcplotdims` printed macros that no longer
  exist (`\xtick`, `\xwidth`); it now shows `\xtickdistance` and
  `\xcm`.

### numodel-coach

#### Changed
- A variable that is not drawn by any `\diagrammodel` now gets the axis
  range its diagram would have, computed from `\computemodel`'s series,
  instead of Coach's default 0 to 10 (which stays for a model that was
  never computed). Regression test `c003-axis-range` covers both.

#### Added
- `\coachmodel[blanks=open|filled]`: the pupil's version, with every
  `[blank]` rule, stop condition and start value as a comment
  `' F_z := ...` to complete in Coach, or the answer version in full.
  Default: numodel's `blanks` setting. Regression test `c004-blanks`.
- LaTeX accents and special letters in `coachinstruction`: `\"u`,
  `\'e`, `` \`a ``, `\^o`, `\~n`, `\c{c}`, `\ss`, `\o`, `\ae` and the
  like (also with braces and `\i`) become the Unicode letter in Coach
  instead of a stray quote. Regression test `c002-instruction`
  extended.
- Coach's angle unit ("hoekeenheid", `AngleUnits`) follows numodel's
  `angles` setting, and the trigonometric functions are exported for
  it: with degrees `sind(x)` becomes `Sin(x)`, `sin(x)`
  `Sin(180/Pi*x)`, `asind(x)` `Arcsin(x)` and `asin(x)`
  `Pi/180*Arcsin(x)`. Tested in Coach 7: in degree mode Arctan(1) = 45,
  although the guide says the inverse functions give radians.
- Greek letters in names are written without brackets (`ΔT`, `λ`, `ω`);
  Coach accepts them bare and marks a bracketed name red in its
  variable list. Brackets remain for names with other characters
  (`[t_1/2]`) and for reserved words (`[Max]`).
- A unit declared with `\DeclareSIUnit` is exported as its siunitx
  symbol (`\dieren\per\day` → `dieren/d`) without the warning
  "unknown unit macro"; only a macro siunitx does not know warns.
  Regression test `c005-angles-units`.

## [0.10.0] — 2026-10-08

Adds a third module, numodel-coach, which exports a model as a CMA Coach 7
activity, and the numodel Lua interface it builds on.

### numodel

#### Added
- Lua export API `numodel.get_model(prefix)`: returns a copy of the
  complete model, so that other packages can export it without
  touching numodel's TeX internals. It lists the variables in
  declaration order (full and short name, display text, type, start
  value both evaluated and as written, siunitx unit as written,
  significant figures) and every row in `\textmodel` order: rules
  (`calc`, `ternary`, with the star of `\mrule*`), `\mruletext` rows
  and the `\mstop` condition. It also gives the `maxiter` setting, which
  exporters use as the iteration count of the target software, and per
  variable the axis range `\diagrammodel` drew for it (`axis_min`,
  `axis_max`, after `\calcplotdims` rounding; the union over several
  diagrams).
  `\mvar`, `\mrule`, `\mruletext`, `\mstop`, `\diagrammodel` and the
  `maxiter` key now pass this data to Lua; what numodel typesets is
  unchanged. `numodel.dump_model(prefix)` prints it for debugging. First
  step towards exporting models to CMA Coach 7. New regression test
  `m005-model-api` and Lua test `test_model_api`.
- Plain-text rendering of a model in Coachtaal (or the English
  syllabus notation): `numodel.plaintext(prefix, {dialect = "NL"|"EN"})`
  returns the model rules and the initial values as text that can be
  typed or pasted into modelling software; `numodel.write_plaintext`
  writes them to a file. Expressions are parsed with l3fp's operator
  precedence and printed with Coachtaal's, adding parentheses only where
  the two differ: `-x^2` becomes `-(x^2)` (Coach gives unary minus and
  `^` the same priority), `a^b^c` becomes `a^(b^c)`, `1/2pi` becomes
  `1/(2*Pi)`, and comparisons joined by `&&`/`||` get the parentheses
  Coachtaal requires: `(x > 1) EN (x < 2)`. Decimal comma and `;` as
  argument separator in Coachtaal; functions without a Coachtaal
  counterpart are rewritten (`cot`, `ceil`, `round(x,n)`, degree
  variants such as `sind`). Names are derived from the display text
  (`F_{\text{res}}` → `F_res`, `\Delta t` → `Δt`); names Coachtaal does
  not allow bare, and reserved words, are put in brackets (`[ω]`,
  `[Max]`). siunitx units are translated with a built-in table
  (`\m\per\s\squared` → `m/s^2`, `\mole\per\litre\per\second` →
  `mol/(L*s)`) and written as comments after the initial values.
  Anything that cannot be translated exactly is reported as a warning.
  New regression test `m006-plaintext` and Lua test `test_plaintext`.

#### Fixed
- `\diagrammodel` broke when the unit of a y-variable is a **robust
  command with an optional argument**, such as the unit macros of
  `physunits` (loaded by `physconst`): `\DeclareRobustCommand{\V}[1][ ]{…}`.
  The y-unit went through full expansion (`\tl_set:Ne` for the
  unit-compatibility check and `\edef\ylabelunit{\tl_use:N …}` for the
  axis label), which expanded `\V` down to its `\kernel@ifnextchar` and
  failed with `Use of \reserved@a doesn't match its definition`,
  followed by some twenty follow-up errors. The x-axis was never
  affected, because `\xlabelunit` is set with `\def`. The unit tokens
  are now copied verbatim (`\tl_set:Nv`, `\exp_not:V`), so the axis
  label is unchanged and the unit comparison now compares the
  unexpanded unit text. The `unit-mismatch` warning passed the same
  tokens through `\tl_use:N` into a fully expanded message argument and
  would have failed in the same way; it now uses `\tl_to_str:N`. New
  regression test `m004-diagrammodel-robust-unit` emulates such units
  (single series, two series sharing a unit, and a unit mismatch) and
  fails on 0.9.1.

### numodel-plot

#### Added
- `scale-format=input`, a third way for a scaled axis to present its
  power of ten. The existing two rewrite the unit: `prefix` folds the
  factor into an SI prefix and `exponent` pulls *every* prefix out —
  including those the user wrote — and shows one power of ten on the
  bare unit. `input` leaves the unit exactly as written and puts the
  axis factor in front of it. For a $y$-range up to $5\times10^7$ in
  `\mega\joule\per\kilo\gram` the three read `E (TJ/kg)`,
  `E (10¹² J/kg)` and `E (10⁶ MJ/kg)`; for `\kilo\metre` scaled by
  $10^3$, `input` gives `10³ km` where `exponent` gives `10⁶ m`. Maps
  straight onto siunitx's own `prefix-mode=input`. The default stays
  `prefix`. Regression test `p005-axislabels` now runs the same
  prefixed unit through all three values.

#### Changed
- The feasibility check behind `scale-format=prefix` — "can 10^k be
  folded into a prefix on this unit?" — now has two ways to obtain
  the parsed unit and picks one at run time. On siunitx 3.6 and newer
  it calls the new public `\siunitx_unit_parse:nN` (added for siunitx
  issue #864); on 3.5 and older it keeps reading
  `\l__siunitx_unit_parsed_prop` through the internals it always
  used. Needed because 3.6 also drops the internal
  `\__siunitx_unit_parse:n`: without the second path the `\cs_if_exist`
  guard would have shut on 3.6 and every `prefix` label would have
  silently come out in the 10^k form — `s (10³ m)` instead of
  `s (km)` — the moment 3.6 reached MiKTeX. Verified against real
  builds of siunitx 3.3.8, 3.4.14, 3.5.10 and the 3.6 development
  branch: all four render every label identically. The prefix ↔
  power-of-ten mapping is still read from siunitx's two internal
  props, which remain the package's last internal references.

### numodel-coach

#### Added
- New module **numodel-coach**: writes a numodel model as a CMA Coach 7
  modelling activity (`.cma7`) in text mode, so pupils open the model
  of the text directly in Coach instead of retyping it from the PDF.
  `\coachmodel[options]{name}` writes the model with the current
  prefix (or `prefix=`); `\coachsetup` sets defaults. Options: `dir`
  (output directory, created when missing), `attach` (also embed the
  file in the PDF with embedfile) and `switch` (allow switching to
  Coach's graphical view; off by default, because Coach derives the
  text view from the graphical model and switching would lose the
  model). The model rules and initial values are numodel's Coachtaal
  rendering (`numodel.plaintext`), all variables go into Coach's
  variable list with their units, and the iteration count is numodel's
  `maxiter`. Everything that cannot be exported exactly is reported as
  a package warning naming the model and the file.
- The `.cma7` format is not documented by CMA. The module starts from
  an empty activity saved by Coach 7.0 in text mode, stored as Lua data
  (`numodel-coach-template.lua`, generated by `tools/gen_template.py`)
  without the webcam name and the sample instruction text it
  contained. Regression test `c001-coachmodel` reads written files back.
- Instruction text with one source for the PDF and Coach: the
  environment `coachinstruction` defines a model's instruction text
  (current prefix or `prefix=`) without typesetting it;
  `\coachinstructiontext` typesets it wherever and as often as wanted,
  and `\coachmodel` writes it to Coach's instruction window, converted
  to the HTML Coach shows (paragraphs, bold/italic, lists, inline math
  with sub- and superscripts, Greek, `\qty`/`\num`/`\unit` with a
  decimal comma). Unsupported commands give a warning and keep their
  text. One instruction per model; redefining it warns. Regression test
  `c002-instruction`.
- The axis range of each variable in Coach's variable list (used for a
  new graph or meter) is the range of its graph in the document, taken
  from `\diagrammodel`; variables without a diagram keep Coach's
  default 0 to 10. Regression test `c003-axis-range`.

### Bundle

#### Added
- New module **numodel-coach** (see above); it ships in the same flat
  CTAN archive as the other two.

#### Changed
- `README.md` is now written for users of the packages (what the bundle
  does, installation, a minimal example per module, requirements).
  Building, testing and releasing moved to `DEVELOPMENT.md` in the
  repository, which is not shipped to CTAN. The minimal numodel example
  gained the missing time rule `\mrule{T}{\ballT + \ballDt}`, and the
  manuals are referred to by their actual names
  (`numodel-manual.pdf`, ...).

## [0.9.1] — 2026-09-07

The first published release carrying the 0.9.0 work listed below it:
the 0.9.0 archive was prepared but never published, so everything
under both headings reaches users with this release.

### numodel

Version-sync release; no functional changes.

### numodel-plot

#### Fixed
- `\qtyPlain` produced wrong output on **siunitx 3.5.0 and newer**
  (released 2026-03-14). siunitx renamed the variable holding the
  current `prefix-mode` from `\l_siunitx_quantity_prefix_mode_tl` — a
  documented public variable through v3.4.14 — to
  `\l_siunitx_quantity_prefix_mode_str`, and left the old name
  declared but never assigned. 0.9.0 read only the `tl`, so it
  resolved an empty mode, matched no prefix-mode handler and fell
  through to plain `\qty`. Nothing errored: every scaled axis label
  silently gained a spurious `1`, `t (1 ks)` instead of `t (ks)`,
  which made `scale-format=prefix` — the 0.9.0 default — wrong on any
  current TeX distribution. The mode is now read from the `str` when
  it carries a value and from the `tl` otherwise, so both siunitx
  generations work. New regression test `p006-prefixmode` emulates
  both namings, so it pins the resolution order regardless of which
  siunitx is installed.

### Bundle

#### Changed
- The bundle now ships **one README and one CHANGELOG**, at the
  bundle root, on CTAN maintainer request. The per-module
  `README.md` / `CHANGELOG.md` pairs have been merged into these two
  files and removed; module-specific entries live under a per-module
  heading within each release above.

  This also settles the packaging problem that held 0.9.0 up.
  l3build's `copyctan()` copies `textfiles` flat into the package
  directory, so both modules and the bundle root wrote to the same two
  filenames and the bundle root — copied last — won; the module
  documents never reached the archive at all. The 0.9.0 archive worked
  around that with per-module subdirectories, which CTAN declined.
  With a single pair of files there is no collision left to resolve,
  so the `ctansetup.lua` helper that arranged the workaround has been
  dropped again.

## [0.9.0] — 2026-09-06

Prepared but never published: CTAN declined the archive over its
layout. Everything below ships as part of 0.9.1.

### numodel

Version-sync release; no functional changes.

### numodel-plot

#### Changed
- `\qtyPlain` no longer patches **siunitx** internals. It used to
  override `\__siunitx_number_output_integer:nnn` and
  `\__siunitx_quantity_print_unit:n` and read four internal
  variables; it is now built entirely on siunitx's documented
  code-level interface (`\siunitx_number_parse:nN`,
  `\siunitx_unit_format_extract_prefixes:nNN`,
  `\siunitx_number_adjust_exponent:Nn`, `\siunitx_number_process:NN`,
  `\siunitx_quantity_print:`, `\siunitx_print_unit:`) — the
  conclusion of the investigation behind siunitx issue #864, which
  established that the whole `prefix-mode` pipeline is already
  public. Axis labels are unchanged; the rewrite also fixes three
  cases the override got wrong:
  - `\qtyPlain{-1}{\metre}` and `\qtyPlain{<1}{\metre}` printed
    `-m` and `<m`, dropping the mantissa behind a sign or
    comparator; they now print `-1 m` and `<1 m`.
  - `\qtyPlain{1(1)}{\metre}` printed `(1)m`; a number carrying an
    uncertainty is now handed to `\qty` untouched.
  - The mantissa suppression fired only when `\qtyPlain` was reached
    through `\edef`, so a direct call under
    `prefix-mode=combine-exponent` printed `1 km` where the same
    call inside an axis label printed `km`. All three `prefix-mode`
    values now behave identically either way.
- `\pzuIfUnitNonEngTF` decides from the net power of ten siunitx
  extracts from the unit rather than from a hard-coded list of prefix
  control sequences — a list of which roughly half (`\cg`, `\dam`,
  `\hPa`, …) named control sequences siunitx does not define at all.
  Two units change verdict, both towards scaling:
  - `\centi\metre\per\centi\second` — the prefixes cancel ($10^0$),
    so the axis now scales instead of staying in `cm cs⁻¹`.
  - `\centi\metre\cubed` — `cm³` is $10^{-6}$ m³, a multiple of
    three, so a `cm³` axis now scales and is labelled `10⁻³ m³`
    instead of keeping four-digit ticks under `cm³`.
- **siunitx** is now requested with a date (`2023-11-06`, v3.3.8).
  Every function used has been public since v3.0.0, but v3.3.7 and
  v3.3.8 fixed the empty-exponent case of
  `prefix-mode=combine-exponent` and the printing of a bare `1` under
  `print-unity-mantissa=false`, both of which the scaled labels rely
  on. An older siunitx warns rather than errors.
- A scaled axis now presents its power of ten as an **SI prefix on
  the unit** by default: `s (km)` instead of `s (10³ m)`, `E (TJ/kg)`
  instead of `E (10⁶ MJ/kg)`, `m (Mg)` instead of `m (10³ kg)`. The
  fold respects unit powers (`10⁶ m²` → `km²`) and existing prefixes,
  and silently falls back to the power-of-ten form whenever no
  engineering prefix fits (`10³ m²` needs 10^1.5 per metre; `10³ m³`
  would need the non-engineering deca). Implemented with siunitx's
  `prefix-mode=combine-exponent` plus an own feasibility pre-check
  through the siunitx unit parser, because siunitx raises an error
  for non-convertible exponents and offers no public parse API yet
  (siunitx issue #864); the check is guarded so that a future siunitx
  rename degrades gracefully to the power-of-ten form.

#### Added
- New setup key `scale-format` controlling the above: `prefix` (the
  new default) folds the power of ten into the unit, `exponent`
  restores the previous `s (10³ m)` presentation.
- New setup key `halo-color` for the tick-label halo colour. The
  default `halo-color=auto` matches the halo to the surrounding
  background at every `\drawplot`: inside a `tcolorbox` it takes the
  box's `colback` (detected via the `tcblayer` nesting counter), on a
  beamer slide the colour theme's `bg`, on a page coloured through
  the `pagecolor` package `\thepagecolor`, and white otherwise. Any
  other value pins the halo to that colour
  (`\numodelplotsetup{halo-color=yellow!30}`); the resolved colour
  remains exposed as the named colour `numodelhalo`.
- Regression tests `p003-qtyplain` (the `\qtyPlain` contract across
  all three `prefix-mode` values, including the sign, comparator and
  uncertainty cases), `p004-unitnoneng` (the `\pzuIfUnitNonEngTF`
  classification, abbreviations, unit powers and cancelling prefixes
  included) and `p005-axislabels` (the label pipeline end to end:
  both `scale-format` values, a `\gram` axis, a `\centi` axis, a
  compound unit with cancelling prefixes and each of the five
  `axis-label-format` values).

## [0.8.0] — 2026-07-06

### numodel

#### Added
- `syntax=FPEVAL` (aliases `fpeval`, `l3fp`): a third rule-table
  rendering that shows each `\mrule`/`\mstop` body verbatim in l3fp
  (`\fpeval`) syntax — lowercase function names, literal `*`, `^`,
  `&&`, `||`, `?:` — typeset monospace, instead of translating it
  into XMILE/CoachTaal syllabus notation.  Variable macros still
  resolve to their `\<name>text` display names (wrapped in
  `\ensuremath` so subscripts and Greek render); the ternary is not
  expanded into `IF`/`THEN`/`ELSE` and the stop row reads
  `stop: <condition>`.  The decimal mark defaults to `point`.  Ships
  as `numodel-FPEVAL.def`.

### numodel-plot

#### Changed
- The semi-transparent white backing behind tick labels (added in
  0.7.0) is now applied **only to an axis that is drawn through the
  middle of the plot**, where curves pass behind the numbers. An axis
  placed on an edge (the default bottom/left, or the `top`/`right`
  placements used for single-sign data) leaves its labels on the white
  margin and gets no backing. In a quadrant I+IV plot, for example,
  the x-axis (middle) keeps its backing while the left-hand y-axis no
  longer carries one, so the backing no longer stands out where it
  served no purpose.
- The backing is no longer a filled rectangle but a **halo that
  follows the glyph outlines** of the scale numbers. The label is
  emitted twice: first stroke-only — a fat white pen (1.6pt, round
  joins, 80% opacity) traces the character outlines via the
  `pdfrender` package — and then the normal label on top. Between and
  around the digits the grid and curves now run through untouched
  instead of being cut off by a rectangular patch. The styles
  `numodel/xticklabel halo` and `numodel/yticklabel halo` apply this
  conditionally from `\calcplotdims`; the halo colour is the named
  colour `numodelhalo` (white by default, adjustable with
  `\colorlet{numodelhalo}{<color>}` for coloured backgrounds). New
  dependency: `pdfrender` (supported on pdfTeX and LuaTeX).
- The halo covers the plotted curves as well as the grid. Previously
  the backed tick labels were drawn before the curves, so a curve
  passing through a scale number ran *over* its backing, which only
  masked the dotted grid lines. When `\calcplotdims` places an axis
  through the middle it now activates PGFPlots layered graphics
  (`set layers`) and lifts the haloed labels onto the
  `axis descriptions` layer: the grid stays below the curves, the
  halo covers both, and the number sits on top of its own halo.

## [0.7.0] — 2026-05-30

### numodel

#### Added
- Multi-prefix `\diagrammodel`: the `prefix=` key now accepts a
  comma-separated list of prefixes, so the same set of y-variables
  can be plotted across several models in one diagram.  Series order
  is prefix-major (all y-variables of the first prefix come before
  any of the second).  When more than one prefix is supplied the
  prefix is appended in parentheses after each y-variable's display
  text in the legend, e.g. `$U_3$ (ptc)` vs `$U_3$ (pw)`.  Axis
  ranges reduce over the union of all (prefix, yvar) pairs; the unit
  filter is applied against the first prefix's first y-variable, so
  the same short names are assumed to share unit and display text
  across prefixes.  Single-prefix and bare-prefix invocations behave
  exactly as before; existing test files render byte-identical
  legends and mark layouts.
- `examples/test-multi-prefix.tex` — exercises the new feature
  alongside the legacy single-prefix paths (single yvar, multi yvar,
  multi prefix × single yvar, multi prefix × multi yvar).
- Display translation for more expression functions inside `\mrule`
  bodies, complementing the eight added in v0.4.0 (`sqrt`, `exp`,
  `ln`, `sin`, `cos`, `tan`, `asin`, `acos`):
  - `atan`, `min`, `max`, `floor`, and `ceil` (rendered as
    `ARCTAN` / `Arctan`, `MIN` / `Min`, `MAX` / `Max`, `INT` / `Entier`
    in the english / coachtaal syntax; `ceil(x)` as the l3fp identity
    `-INT(-x)` / `-Entier(-x)`).
  - Reciprocal trig `cot`, `csc`, `sec` — rendered as the slashed
    fraction `\sfrac{1}{tan(…)}`, `\sfrac{1}{sin(…)}`,
    `\sfrac{1}{cos(…)}` (l3fp has no direct token for them).  Arguments
    containing nested parentheses fall back to the older inline
    `1/tan(…)` form.
  - Inverse reciprocal trig `acsc`, `asec`, `acot` — rendered as
    `ARCSIN(\sfrac{1}{…})`, `ARCCOS(\sfrac{1}{…})`,
    `ARCTAN(\sfrac{1}{…})`.
  - Degree forward trig `sind`, `cosd`, `tand` — rendered as
    `SIN(\sfrac{\pi}{180}·…)` etc.
  - Degree inverse trig `asind`, `acosd`, `atand` — rendered as
    `\sfrac{180}{\pi}·ARCSIN(…)` etc.
  - Degree reciprocal trig `cotd`, `cscd`, `secd` — rendered as
    `\sfrac{1}{TAN(\sfrac{\pi}{180}·…)}` etc.
  - Degree inverse reciprocal trig `acscd`, `asecd`, `acotd` —
    rendered as `\sfrac{180}{\pi}·ARCSIN(\sfrac{1}{…})` etc.

  The functions whose conversion lives *inside* the parentheses
  (`acsc`/`asec`/`acot`, the `…d` reciprocals, the `…d` inverse
  reciprocals, and `ceil`) capture their argument, so a call whose
  argument itself contains parentheses is left untranslated rather
  than mis-rendered.  The two-argument `atan(y,x)` / `acot(y,x)` /
  `atand(y,x)` / `acotd(y,x)` (`ARCTAN2`) forms are not translated.
- `tests/expr-newfns-test.tex` — renders the new functions through
  `\textmodel` in both the english and coachtaal syntaxes.

#### Changed
- `\textmodel` now renders reciprocal trig functions (`cot`, `csc`,
  `sec`) as a slashed `\sfrac` fraction instead of an inline `1/…`,
  so the numerator no longer jams against a preceding term (e.g.
  `a + \sfrac{1}{sin(x)}` rather than `a + 1/sin(x)`).  Requires the
  new `xfrac` package dependency.

#### Fixed
- `\textmodel` rendering of multi-character numeric exponents.  The
  display pipeline in `\__numodel_vars_to_display:N` previously
  emitted raw `^1.2` (also `^12`, `^-3`, …) into the math-mode
  string assembled for the rule cell, so TeX math mode bound only
  the first token after `^` as the superscript: `T^1.2` typeset as
  `T¹.2` (the `.2` falling out to the baseline), `T^12` as `T¹2`.
  `\fp_eval:n`-side execution was unaffected — `\computemodel`
  always saw the full expression — so the bug was purely cosmetic
  but masked the actual rule a reader was meant to follow.  Numeric
  exponent literals (optional sign, digits, optional decimal part)
  are now wrapped in `{…}` before being typeset, so `T^1.2` reads
  `T^{1.2}` and `T^-3` reads `T^{-3}`.  Single-token literals like
  `T^4` were already correct and are now also wrapped (`T^{4}`,
  rendered identically by TeX), keeping the regex uniform.
  Explicitly braced (`^{…}`) and parenthesised (`^(…)`) exponents
  are left untouched because the regex requires a digit or sign
  immediately after `^`.

### numodel-plot

#### Added
- New example `examples/numodel-plot-example-quadrants.tex` that
  renders all nine quadrant configurations (four single-quadrant,
  four half-plane, and the full four-quadrant case) side by side, so
  the axis-line and axis-label placement rules can be inspected in
  one figure.

#### Changed
- Tick labels on both axes now render with a semi-transparent white
  background (`fill=white, fill opacity=0.8, text opacity=1, inner
  sep=1pt`) so the numbers remain optically separated from the grid
  and from plotted curves passing behind them.
- Tick labels at axis crossings are no longer hidden. The
  `numodel/axis` style sets `hide obscured x ticks=false` and
  `hide obscured y ticks=false`, so every scale number stays visible
  even where an axis line would otherwise obscure it; the new white
  background provides the visual separation from the crossing axis.
- Axis labels (quantity + unit) are now placed in the extension of
  the axis arrow rather than at the tick-label margin, in all five
  configurations where the axis line is moved off the plot edge
  (4Q, I+II, II+III, III+IV, IV+I).  The label sits exactly on the
  axis line, offset by `1em` past the arrow tip.  Previous placement
  used `xticklabel cs:1.05` / `yticklabel cs:1.05`, which positioned
  the label at the tick-label height rather than on the axis itself,
  giving an awkward diagonal offset relative to the arrow.

#### Fixed
- Tick-scale label (the `·10^n` factor that appears when a magnitude
  is extracted into the axis label) is now actually suppressed.
  The previous implementation hid it via `opacity=0` on
  `every {x,y} tick scale label`, which leaked through the new
  `text opacity=1` on the tick-label style and made the factor
  reappear near the origin.  Suppression now uses pgfplots' own
  `{x,y}tick scale label code/.code={}`, which prevents the node
  from being drawn at all (and keeps it out of the PDF text layer).

## [0.6.0] — 2026-05-26

### numodel

#### Changed (build / packaging)
- The user manual has moved out of `numodel.dtx` into a stand-alone
  `numodel-manual.tex`.  The `.dtx` is now pure docstrip source for
  `numodel.sty` / `numodel-EN.def` / `numodel-NL.def` and is no
  longer typesettable on its own.  `l3build doc` continues to build
  the manual but now compiles `numodel-manual.tex` instead of the
  `.dtx`.  No change to the package's installed files or user-facing
  API; only the build-time layout of the bundle changed.
- The `examples/` directory is no longer shipped to CTAN.  Local
  development examples stay in the repo but the CTAN upload now
  contains only the package source (.dtx/.ins/.lua/-manual.tex),
  the extracted .sty/.def, the typeset PDF, README and CHANGELOG.
  Same exclusion already in effect for `testfiles/`.

#### Added
- Per-prefix accessor `\<prefix>steps` set by `\computemodel`,
  expanding to the iteration count $N$ (= number of recorded
  samples).  Previously the count had to be derived by hand from
  e.g. `Tmax/dt`, which could be off by one depending on how
  `\mstop`'s inequality rounded.  Undefined before the first
  `\computemodel` call for that prefix.  Picking `steps` as an
  `\mvar` name now collides with this accessor (noted in the
  naming caveat of the manual).

#### Fixed
- `\textmodel` no longer raises "Undefined control sequence \DN@" /
  "\extrap@" / "\@cdots" warnings (and earlier silently scrambled
  cell contents) when an `\mvar`'s `alias`, `aliasleft`, or
  `aliasright` key holds `\cdots`, `\ldots`, or any other amsmath
  dots macro (`\dotsb`, `\dotsm`, `\dotsc`, `\dotsi`, `\dotso`,
  `\dots`).  Two interacting causes:
  - The startcell builder used `\tl_if_blank:eTF { \use:c { ... } }`
    to test whether each alias key was set.  The `:e` variant
    fully expanded `\cdots`, triggering its `\futurelet` lookahead
    outside a math context.  Replaced by
    `\str_if_eq:eeTF { \tl_to_str:c { ... } } { }`, which
    detokenizes the accessor's body after exactly one expansion
    and compares the resulting string to the empty string.  Two
    near-misses on the way to this: `\cs_if_eq:cNTF ... \c_empty_tl`
    and `\tl_if_empty:cTF`; both turn out to be `\if_meaning:w`
    comparisons against `\c_empty_tl`, and `\cs_gset:cpe` produces
    `\long\xdef` macros whose meaning never matches the
    non-`\long` `\c_empty_tl` — so empty-body accessors compared
    as not-empty.  The string-comparison route is prefix-agnostic.
  - When the alias value was then typeset inside the tabularray
    cell, `\cdots`'s lookahead still ran past the `$`-closer into
    document tokens (a regression introduced in v0.4.0 when
    `\textmodel` moved from plain `tabular` to `tabularray`'s
    `longtblr` — plain `tabular` did not pre-process cell contents,
    so the lookahead stayed contained).  Each alias substitution is
    now wrapped in an explicit `{...}` group so the lookahead is
    bounded to the cell.

#### Changed
- `\mstep{<Name>}{<i>}` now raises a `numodel` error when `<i>`
  falls outside the recorded range (`0..\<prefix>steps - 1` or, for
  Python-style negative indices, `-\<prefix>steps..-1`).  The
  previous behaviour was to silently expand to nothing, which let
  typo'd indices and off-by-one mistakes propagate into surrounding
  PGFPlots expressions as empty arguments.
- `\graphicmodel` auto-layout no longer leaves an empty row between
  the stocks row and the constants row when the model has no
  auxiliary variables to fill the middle row.  Constants drop down
  from `gridy=2` to `gridy=1` in that case, so simple
  stock-plus-constant models render without a vertical gap (visible
  in the `forrester` and `edu` examples of the manual's "Diagram
  styles in practice" section).
- Manual: assorted clarifications and corrections.

#### Removed
- `stockwidth` setup key and the `\halfstockwidth` macro it wrote
  to.  The value was never consumed — stock rectangles have used
  fixed `em`-based dimensions for a while — so the key only ever
  accepted input without effect.  No replacement; node sizes scale
  with the surrounding font size.

### numodel-plot

#### Changed (build / packaging)
- The user manual has moved out of `numodel-plot.dtx` into a
  stand-alone `numodel-plot-manual.tex`.  The `.dtx` is now pure
  docstrip source for `numodel-plot.sty` and is no longer
  typesettable on its own.  `l3build doc` continues to build the
  manual but now compiles `numodel-plot-manual.tex` instead of the
  `.dtx`.  Manual now uses Arial/Lete Sans Math/Fira Mono fonts
  (matching sibling `numodel`) and a tcolorbox+listings
  `plotexample` environment so each example body is shown verbatim
  and executed from a single source.  No change to the installed
  `.sty` or to the user-facing API.
- The `examples/` directory is no longer shipped to CTAN.  Local
  development examples stay in the repo but the CTAN upload now
  contains only the package source (.dtx/.ins/-manual.tex), the
  extracted .sty, the typeset PDF, README and CHANGELOG.

## [0.5.0] — 2026-05-23

### numodel

#### Added
- Multi-series `\diagrammodel`: the second argument now accepts a
  comma-separated list of y-variables.  Every entry whose `unitraw`
  matches the first one's is plotted in the same diagram as
  discrete model points (mark-only); the y-axis is scaled to the
  joint min/max of the kept series.  Entries with a non-matching
  unit are dropped and a `numodel/unit-mismatch` warning is issued.
  Series cycle through a colour-blind safe palette (Okabe & Ito,
  yellow omitted) ordered so consecutive colours differ in
  luminance as well — so they stay distinguishable on a greyscale
  printout.  The legend lists each kept variable's display text;
  the legacy single-entry form keeps the historical "model point"
  mark-only rendering unchanged.
- Shared-inflow Forrester diagrams: when one auxiliary variable is
  the inflow of more than one stock, `\graphicmodel` draws a single
  valve next to the first such stock, places the additional stocks
  side by side, and threads a curved branch (`to[bend left=30]`,
  matching the curved causal arrows) from the valve over the
  primary stock into each extra target.  Layout stays compact —
  the involved stocks share one row, no vertical stacking.
- Shared-outflow Forrester diagrams: mirror of the above for the
  outflow side.  When one variable drains more than one stock,
  the valve is placed to the right of the last-declared source
  stock and every earlier source attaches with a curved branch
  arcing back over the intervening stocks into the shared valve.
- `\flowarrowbent` / `\flowoutarrowbent` TikZ macros driving the
  curved inflow / outflow branches; both use the same
  `bend left=30` style as the existing curved causal arrows.
- New colour-blind safe palette `numodelseriesa`…`numodelseriesg`
  (Okabe & Ito with yellow omitted), ordered by luminance for
  greyscale legibility.  Reused as the multi-series cycle in
  `\diagrammodel`; available to user code as `\definecolor`-style
  names.
- `examples/test-multi-series.tex`,
  `examples/test-shared-inflow.tex`, and
  `examples/test-ptc-heuristic.tex` — exercise the new features
  (multi-series filtering with a unit-mismatch warning, shared
  inflow, and the PTC two-filament stocks with shared radiative
  outflow).
- Manual subsection §"Shared valves" describing the side-by-side
  layout and the curved-branch rendering for shared inflow and
  shared outflow.

#### Changed
- Flow-variable detection prefers `aux` and `stock` variables over
  `constant`s when both appear in the same inflow term.  The
  previous heuristic ("first declared wins") would mislabel an
  inflow valve as a scaling constant when that constant happened
  to be declared before the rate-bearing helper; the new rule
  picks the helper, so the valve label matches what the physics
  treats as flowing.
- Inflow term parsing now distributes top-level parenthesised
  sub-expressions of the form `(+A - B) * C` into `+A*C, -B*C`
  before flow classification.  This lets a rule such as
  `\T + (\P - \k * \T^4) * \Dt / \mc` surface `P` as the inflow
  valve and `k` as the outflow valve, instead of treating the
  whole bracket as one inflow term that masks the radiative
  `k * T^4` outflow.

#### Internal
- New Lua maps `flows.valve_extra_targets` and
  `flows.outvalve_extra_sources` track the extra stocks attached
  to a shared valve.  Exported to TeX through
  `\l__numodel_valve_extras_prop` and
  `\l__numodel_outvalve_extras_prop`.
- `auto_layout` now defers shared-valve placement to the
  designated primary source — first-declared for inflows
  (so the valve lands at the left of the stock row), last-declared
  for outflows (so it lands at the right) — and skips it for the
  other sharing stocks.  Their flow attaches with a curved branch
  emitted by `\__numodel_emit_valve_extras:n` /
  `\__numodel_emit_outvalve_extras:n`.
- `last_outflow` gap-detection in `auto_layout` now keys off
  whether the chain actually placed an outflow valve, so shared
  outflows no longer leave an empty grid cell between rows.

### numodel-plot

Version-sync release; no functional changes.

## [0.4.0] — 2026-05-19

### numodel

#### Added
- Expression reference table in the manual (`\subsection{Expression reference}`)
  listing all l3fp operators and functions with their XMILE and CoachTaal
  equivalents.
- Display translation for eight more `\fp_eval` functions inside
  `\mrule` bodies: `sqrt`, `exp`, `ln`, `sin`, `cos`, `tan`, `asin`,
  `acos`
- `\textmodel` now renders through `tabularray`'s `longtblr` (was
  plain `tabular`).  The table breaks across pages when long, the
  column header repeats on each continuation page, and continuation
  markers ("(Continued)" at the top, "Continued on next page" at the
  bottom) are localised to the current `syntax` (CoachTaal:
  "(Vervolg)" / "Wordt vervolgd op volgende pagina").
- `tblrenv` key for `\numodelsetup`, `\usepackage[tblrenv=...]{numodel}`,
  and `\textmodel[tblrenv=...]`.  Values: `tblr`, `longtblr` (default),
  `talltblr`.  Picks which `tabularray` environment wraps the table.
  Use `tblr` or `talltblr` when `\textmodel` sits inside an enclosing
  environment that already suppresses page breaks (`subfigure`,
  `minipage`, …); the default `longtblr` would otherwise emit
  spurious continuation markers in that setting.  Like `units`, the
  per-call form overrides the global only for that one render and the
  global state is restored afterwards.

#### Changed
- `numodel` now issues `\SetTblrInner{rowsep=0pt}` on package load so
  the `\textmodel` rule listing renders compactly.  This applies to
  *every* `tabularray` table in the document; users who want the
  default tabularray spacing back in their own tables can issue
  `\SetTblrInner{rowsep=2pt}` (or any other value) anywhere in their
  document.

### numodel-plot

Version-sync release; no functional changes beyond the rename below,
bumped to keep bundle component versions aligned (per CTAN's
bundle-versioning convention).

#### Changed
- Example files renamed to avoid generic filenames (CTAN maintainer
  feedback):
  - `examples/numodel-plot-simple.tex` →
    `examples/numodel-plot-example-basic.tex`
  - `examples/numodel-plot-scaled.tex` →
    `examples/numodel-plot-example-scaled.tex`

## [0.3.0] — 2026-05-17

### numodel

Version-sync release; no functional changes.

### numodel-plot

#### Added
- Rendered example plots in the manual: a minimum working example and a
  scaling demonstration in §Usage, three axis-label-format variants and
  three `grid` variants in §Keys, and three PGFPlots-style overrides in
  §PGFPlots styles.

#### Changed
- Moved `legend pos=outer north east` from `\drawplot`'s hard-coded
  axis options into the overridable `numodel/axis` style, so callers
  can change the legend position via
  `\pgfplotsset{numodel/axis/.append style={legend pos=...}}`.

#### Fixed
- `\calcplotdims` no longer emits stray `\par` tokens from blank lines
  in its source. When the macro was expanded inside a horizontal box
  (e.g. `\sbox`, a `subcaption`/`subfigure` cell), those `\par`s
  switched mode and added ~3.4 cm of empty space to the picture's
  bounding box on the left. `\drawplot` now produces a tight bounding
  box and tiles correctly in tabular/`subcaption` layouts.
- `\xlabelbuild`/`\ylabelbuild` skip the scaled-emit path (`\qtyPlain`)
  when the data magnitude lies in `[1e-2, 1e4]` and no factor-of-ten
  is needed, falling back to the simpler `\si{...}` rendering.
- Custom `grid` values passed to `\numodelplotsetup` now work as
  documented. The `grid .unknown` fallback was previously declared
  with the wrong `l3keys` syntax for a `.choice:` key (`grid .unknown`
  instead of `grid / unknown`), so any value other than `mm-dots` or
  `none` raised an "accepts only a fixed set of choices" error.

## [0.2.0] — 2026-05-16

Initial CTAN release of the bundle.

### numodel

#### Added
- `\mstep` (and `\mstepp`) accept negative indices, Python-style:
  `-1` returns the last recorded step, `-2` the penultimate, etc.
  Out-of-range indices return nothing as before. This makes
  end-of-simulation accessors writable without knowing the step
  count, e.g. `(\mstep{T}{-1}, \mstep{Y}{-1})` for the final
  $(t, y)$ pair.
- Three keys that fine-tune the Forrester diagram appearance,
  available globally via `\numodelsetup`, per-render via
  `\graphicmodel[...]`, and (for the cloud key) per-stock via
  `\mvar[...]`. Empty value resets to "follow `diagram-style`".
  - `flowarrow-style = hollow | filled` — `hollow` renders the
    classic Forrester double-line pipe with an open arrow head;
    `filled` renders a thick solid arrow. Default tracks
    `diagram-style` (`forrester` → `hollow`, otherwise `filled`).
  - `valve-style = valve | circle | edu` — `valve` draws the
    bow-tie/butterfly icon (Forrester); `circle` draws an empty
    circle; `edu` draws a labelled circle. Default tracks
    `diagram-style` (`forrester` → `valve`, otherwise `edu`).
  - `flowarrow-cloud-tip = true | false` — anchors the open end of
    inflow/outflow pipes to a cloud node (model-boundary marker).
    Default tracks `diagram-style` (`forrester` → `true`, otherwise
    `false`). Per-stock override on `\mvar` wins over the global
    default.
- `examples/test-flowarrow-11bg.tex` — exercises the three keys
  and the per-stock override.
- `examples/` directory with `free-fall.tex`, `projectile.tex`, and
  `oscillator.tex` (single-stock, two-stock, and harmonic-oscillator
  models).
- Package-time key processing: `\usepackage[syntax=english]{numodel}`
  is now wired up via `\ProcessKeyOptions`. Previously the option
  was documented but only accessible through `\numodelsetup`.
- `diagram-style` key for `\graphicmodel` and `\numodelsetup` (and
  package options), with three values:
  - `tight` (default) — current behaviour: the valve carries the
    label of the direct inflow/outflow helper or constant, and that
    helper/constant is not drawn as a separate node.
  - `forrester` — Forrester/Sterman convention: valve has no label;
    the helper/constant remains as a separate node connected to the
    valve by a causal arrow.
  - `edu` — didactic dual form: valve carries the label *and* the
    helper/constant is drawn as a separate node with a causal arrow
    to the valve.
  Passing `diagram-style=...` directly to `\graphicmodel` overrides
  the global setting for one render only; the global state is
  restored afterwards, so successive `\graphicmodel` calls can each
  pick a different style without re-issuing `\numodelsetup`.
- `examples/test-diagram-styles.tex` and
  `examples/test-diagram-styles-inline.tex` demonstrate the three
  styles side-by-side and the per-call key override.
- `units` boolean key for `\textmodel` and `\numodelsetup` (and as a
  package option), default `true`. When true, the *startwaarden*
  column of `\textmodel` renders each initial value via
  `\<prefix><Name>qty` (number + SI unit); when false, via
  `\<prefix><Name>num` (number only). Passing `units=false` directly
  to `\textmodel[units=false]` overrides the global setting for one
  render only; the global state is restored afterwards. No extra
  column is added; the unit appears inline in the existing cell.
- `examples/test-units.tex` demonstrates the default-on behaviour, the
  per-table override, the global override, and the restore.
- Localised column titles in `\textmodel`: the *startwaarden* header
  becomes "initial values" under `syntax=english` and stays
  "startwaarden" under `syntax=coachtaal`. The "model" header is
  identical in both. Implemented via two new keyword-table entries
  (`th_model`, `th_initvals`) in `\__numodel_kw:n`.
- `decimal-separator` key for `\numodelsetup` (and as a package
  option), values `comma` or `point`. Controls the decimal mark used
  in the *startwaarden* column of `\textmodel` (siunitx
  `output-decimal-marker`) and in the tick labels of `\diagrammodel`
  (pgfplots `/pgf/number format/use {comma,period}`). Default tracks
  `syntax`: `coachtaal` picks `comma`, `english` picks `point`. An
  explicit `decimal-separator=…` locks the choice and overrides any
  later `syntax` change. The override is scoped to a TeX group around
  the renderer body so a document-wide `\sisetup` is left untouched.
- `examples/test-i18n.tex` exercises both the column-title localisation
  and the decimal-separator override across the syntax/explicit-key
  matrix.

#### Changed
- `\diagrammodel` no longer needs an explicit `\calcplotdims` call
  before its internal `\drawplot` invocation. `numodel-plot` now
  performs the dimension calculation inside `\drawplot`.
- Default `syntax` is now `english` (was `coachtaal`).
- `\textmodel` *startwaarden* column now shows units by default
  (`units=true`). Previously it always rendered numbers only via
  `\<prefix><Name>num`. To restore the old number-only behaviour
  document-wide, add `\numodelsetup{units=false}`; to do it per
  render, use `\textmodel[units=false]`. The `aliasright` per-variable
  override is unaffected.
- `\textmodel` right-column header is now syntax-dependent (was
  hard-coded "startwaarden"). Documents using `syntax=english` will
  see "initial values" instead. Pass `\numodelsetup{syntax=coachtaal}`
  to keep the Dutch header.
- Default decimal mark in `\textmodel` and `\diagrammodel` now tracks
  `syntax`: a freshly loaded `\usepackage[syntax=english]{numodel}`
  document renders `0.05` and `-9.81`, where it would previously have
  inherited the user's `\sisetup`. `syntax=coachtaal` continues to
  produce `0,05` and `-9,81`. Override either with
  `\numodelsetup{decimal-separator=point|comma}`.
- `\__numodel_build_graphic:` now resets every variable's
  `gridx`/`gridy` from per-variable `gridxinit`/`gridyinit` snapshots
  recorded at `\mvar` time. This lets a model be rendered multiple
  times with different `diagram-style` values without auto-placed
  positions from a previous render being mistaken for manual
  placement.

#### Fixed
- Flow detection is now factor-aware (term-based instead of
  variable-based). A between-flow or stock-as-flow now requires the
  inflow term in one stock to be matched by a symmetric outflow term
  in the source stock — the conserved-quantity reading: stock B
  increases only because stock/source A decreases at the same rate.
  Previously, any flow variable appearing as `+fv` in one stock-rule
  and `-fv` in another was classified as a between-flow regardless
  of the surrounding factors, and any inflow whose flow var was
  itself a stock was classified as stock-as-flow regardless of
  whether the source stock actually decreased. `numodel.lua` now
  tokenises each rule into top-level additive terms and matches the
  full normalised term-string. Visible effects:
  - showcase's `EZ`/`Esys`/`dMGV` triple now renders `dMGV` as the
    between-valve from EZ to Esys (previously two unrelated valves).
  - `free-fall.tex`, `projectile.tex` and `oscillator.tex` no longer
    render their drive-stock (V, V_y, V) as a stock-as-flow into the
    integrated stock (Y, Y, X), because the drive stock does not
    decrease as the integrated stock grows. The relation is now a
    regular causal arrow.
- `\graphicmodel` no longer draws a slanted inflow arrow when a
  constant is the direct inflow of a stock. Constants now receive
  the same valve treatment as helper variables in
  `\__numodel_auto_layout:`: when detected as inflow, outflow, or
  between-flow they count as a stock for layout purposes and are
  placed adjacent to the stock on `gridy=0` instead of remaining on
  `gridy=2`.
- `\RequirePackage{float}` to support the `[H]` placement used by
  `\diagrammodel`.
- `\RequirePackage{amssymb}` for the `\leqslant`/`\geqslant` glyphs
  produced when a rule body contains `<=`/`>=`.
- Numeric output of `\<prefix><Name>num` / `\<prefix><Name>qty` no
  longer relies on the user's project-level `\sisetup` for
  `evaluate-expression`, `round-mode=figures`, or
  `exponent-mode=threshold`. These options are now passed explicitly
  to every `\num`/`\qty` call so that a standalone document with no
  `\sisetup` renders correctly. The bodies are assembled by non-expl3
  helper macros (`\NumodelDefNumCs`, `\NumodelDefQtyCs`,
  `\NumodelDefPreCs`) so that the option-string `:` separators retain
  their normal `other` catcode and siunitx's option parser can read
  them.

### numodel-plot

#### Changed
- `\drawplot` now invokes `\calcplotdims` internally; user code no
  longer needs an explicit `\calcplotdims` call before `\drawplot`.
  `\calcplotdims` remains public for advanced cases (overlay TikZ,
  custom `axis` environment).
- Default `axis-label-format` is now `ieee` (the former `parens`
  option, renamed for consistency with engineering convention).
  The `parens` key has been removed.

## [0.1] — 2026-04-24

### numodel

#### Added
- Initial release, extracted from internal project sources.
- Public API: `\newmodelprefix`, `\switchmodelprefix`, `\mvar`,
  `\mrule`, `\mruletext`, `\mstop`, `\mcoords`, `\mcoordsp`, `\mstep`,
  `\mstepp`, `\textmodel`, `\computemodel`, `\graphicmodel`,
  `\diagrammodel`.
- Bilingual rule syntax via package option `syntax=english`
  (default, XMILE-style ALL CAPS) or `syntax=coachtaal` /
  `syntax=dutch` (CoachTaal).
- Lua iteration backend (`numodel.lua`); per-prefix isolation of
  variable lists, step counters, and recorded time series.
- TikZ stock-and-flow diagram style: `stock`, `valve`, `aux`,
  `const`, `flowpipe`, `causal`, `gridscale`; helper macros
  `\flowarrow`, `\flowoutarrow`, `\flowbetweenarrow`, `\constnode`.
- Numeric output of `\textmodel` / `\computemodel` matches `siunitx`
  formatting via the user's `\sisetup` defaults.
- Optional integration with project-local `\ifinworksheet` and
  `\g_defqty_names_seq`; package functions normally without them.

### numodel-plot

#### Added
- Initial release, extracted from internal project sources.
- `\calcplotdims`, `\drawplot` — auto-sized PGFPlots rendering.
- `\numodelplotsetup` — key-value configuration for axis-label format,
  grid style, maximum axis dimensions.
- Five axis-label formats: `ieee`, `iso`, `brackets`, `qty-only`,
  `unit-only`. Automatic factor-of-ten extraction for tick magnitudes
  outside `[1e-2, 1e4]`.
- Quadrant-aware axis-line and label placement (1-, 2-, and
  4-quadrant graphs).
- Helper macros `\qtyPlain`, `\pzuIfUnitNonEngTF` exposed for reuse.
- PGFPlots styles `numodel/grid`, `numodel/ticks`, `numodel/axis`
  as project-level customisation hooks.

Pre-1.0; breaking changes may still occur before 1.0.0.