The axes are the product

The parts other libraries get bored of are the product. Ticks are computed. Labels are exact. If the axes are wrong, nothing else matters.

Why#

Terminal plotting libraries compete on marks and lose on axes. The usual axis is an afterthought. Ticks sit on arithmetic intervals and land on ugly values. Labels go through a float formatter and come out 0.30000000000000004. The y-axis changes width from row to row. Log scales and calendar time get no real answer. The body of the chart can be beautiful and the chart is still unreadable, because the axes are how a reader attaches numbers to ink.

Label lies are the quiet version of the same failure. A label that rounds differently from its neighbor, or that shows a value the tick does not sit on, puts the wrong number in someone's head. They believe it.

The idea#

Treat the boring parts as the product. Tick placement is the extended Wilkinson algorithm (Talbot, Lin, Hanrahan 2010). Candidate steps are scored for simplicity, coverage, density, and legibility. That is the placement the visualization literature settled on, not a heuristic that happens to work at one size.

Labels are exact decimals. An integer mantissa times a power of ten, written so every label parses back to exactly its value. Float artifacts are structurally impossible, not filtered out afterward. One fraction width and one SI prefix per axis (2.5M, 100µ), so a column of labels lines up and the reader carries one unit, not a new one each row. Log axes get superscript decades. Calendar axes get multi-scale labels, and they say 14:05, or Aug 2, or 2027, as the span demands. Band axes fit category labels to their bands.

Layout gets the same care. Labels are measured in display cells, CJK included. Gutters are computed. Collisions are resolved by shedding furniture, not by overlapping ink.

Consequences#

• Ticks are never supplied as strings. There is no API for hand-placed ticks, and no tick callback, because computed placement is a feature, not a limitation.
• Every label round-trips. Parse it and you have the tick's exact value.
• Axis quality is testable. Placement and formatting are pure functions, with snapshot coverage, not a judgment call about the picture.
• Log, time, and band scales are first-class axis types, not label formatters bolted onto a linear scale.
• The formatter is shared. Colorbars, legends, and the notes that disclose quantization speak the same exact decimals as the axes.

Not this#

• format!("{}", 0.1 + 0.2) does not get near a label.
• Tick counts do not ignore the space they have, and labels do not overlap.
• Caller-supplied tick strings are not the escape hatch for bad placement, and neither is a tick callback.
• Small charts do not get a second, cheaper formatter.

See The full draw is the oracle for the ink half of honesty, and Vision rule 3.

Witness#

Ticks stepping by 0.2, a value with no exact binary form. That is the classic float-artifact trap. Spliced here by the doc generator. Every label is an exact decimal, and each axis shares one fraction width:

             every label an exact decimal
0.6 ┤     ⢀⠔⠉⠉⠉⠑⠤⡀                           ⡠⠒⠉⠉⠒⠒⡄
    │    ⡔⠁      ⠈⢆                        ⡠⠊      ⠈⠑⢄
    │  ⡠⠊          ⠑⢄                     ⡜          ⠈
0.4 ┤ ⡜             ⠈⢆                  ⢀⠎
    │⠜                ⠣⡀               ⡰⠁
0.2 ┤                  ⠱⡀            ⢀⠔⠁
    │                   ⠈⢆          ⢀⠎
    │                    ⠈⠢⣀      ⢀⠔⠁
0.0 ┤                       ⠣⢄⣀⣀⡠⠔⠊
    └┬───────┬───────┬────────┬───────┬───────┬───────┬
     0      10      20       30      40      50      60

Spelled today#

scale::Ticks is extended-Wilkinson placement, with Ticks::log10 decades and Ticks::time calendar ticks over unix seconds; scale::Scale is the axis specification (Linear | Log | Time | Bands). Exact-decimal formatting and per-axis SI prefixes live in the tick formatter; label measurement is display-cell aware via unicode-width. This section may rot; the rest must not.