feat(layout): add the layout engine, generators and fill

Bolt commented

2026-08-18 13:08:38 +01:00

Collaborator

Summary

bundled Helvetica AFM widths as JSON, measured in pure functions, with the baseline computed once in layout and emitted on the primitive
widths carry no kern pairs: pdf-lib kerns widthOfTextAtSize but draws a single Tj string no viewer kerns, so the preview is told the same
auto, fixed and grow across stack, row and box, with padding, gap, per-edge borders, alignment and the three box shapes
the resolve step the pipeline names - evaluate expressions, expand repeats, flatten widget nodes - plus range, list and timeSequence
fill carries its subtree and scope into layout, because its iteration count is a layout result rather than an input to one
fill under an auto ancestor is rejected at parse and flagged again at resolve, never measured as zero
layout emits the full mark set and both renderers take it; the parity matrix, overflow band, click-to-select and pages -> document stay M4
closes #11, closes #12, closes #13

Test plan

pnpm test green: 253 tests, 140 of them new across metrics, measure, layout, resolve, sheet, generators and fill
every glyph in the bundled table checked against pdf-lib's own AFM data, with the kerning divergence pinned on the AV pair
inline snapshots over the sizing and container combinations, and over repeat over, count and fill
fill packed end to end through the pipeline: fixed and grow ancestors, $i bound per iteration, packing across a row and ahead of a grow sibling
an iteration with no size is flagged rather than repeated to the cap
the M2 notes fixture now sizes its chain, which is the new fill rule doing its job rather than a change of intent
pnpm lint, pnpm build and prettier green

## Summary - bundled Helvetica AFM widths as JSON, measured in pure functions, with the baseline computed once in layout and emitted on the primitive - widths carry no kern pairs: pdf-lib kerns widthOfTextAtSize but draws a single Tj string no viewer kerns, so the preview is told the same - auto, fixed and grow across stack, row and box, with padding, gap, per-edge borders, alignment and the three box shapes - the resolve step the pipeline names - evaluate expressions, expand repeats, flatten widget nodes - plus range, list and timeSequence - fill carries its subtree and scope into layout, because its iteration count is a layout result rather than an input to one - fill under an auto ancestor is rejected at parse and flagged again at resolve, never measured as zero - layout emits the full mark set and both renderers take it; the parity matrix, overflow band, click-to-select and pages -> document stay M4 - closes #11, closes #12, closes #13 ## Test plan - `pnpm test` green: 253 tests, 140 of them new across metrics, measure, layout, resolve, sheet, generators and fill - every glyph in the bundled table checked against pdf-lib's own AFM data, with the kerning divergence pinned on the AV pair - inline snapshots over the sizing and container combinations, and over repeat over, count and fill - fill packed end to end through the pipeline: fixed and grow ancestors, $i bound per iteration, packing across a row and ahead of a grow sibling - an iteration with no size is flagged rather than repeated to the cap - the M2 notes fixture now sizes its chain, which is the new fill rule doing its job rather than a change of intent - `pnpm lint`, `pnpm build` and prettier green

Bolt added 3 commits

2026-08-18 13:08:38 +01:00

feat(layout): measure text from the bundled helvetica width table c98f76e66f

- widths only, no kern pairs: pdf-lib kerns widthOfTextAtSize but draws
  a single Tj string that no viewer kerns, so the number would be the
  width of nothing that reaches paper
- the preview is told the same thing, or the browser kerns and drifts
- a character outside WinAnsi is flagged on its node here rather than
  left to throw out of pdf-lib at export
- the baseline moves into the metrics module, so layout is the only
  place that computes one

Closes #11

feat(layout): add the sizing model and container properties eb8c3646c7

- auto from content, fixed in mm, grow sharing leftover by weight, over
  stack, row and box nested without limit
- align is the cross axis only: spreading along the main axis is what
  grow and a spacer are for, which is how the day circles distribute
- four identical edges collapse to one rect, so the common bordered
  section stays one primitive; anything else is a line per edge
- the full mark set of decision 19 lands, so both renderers take line,
  circle and the rounded and elliptical box shapes
- an intrinsic size is memoised per node: a parent asks twice, and an
  unmeasurable mark must reach the outline once

Closes #12

feat(layout): expand repeat, generators, when and fill 378ef75583

- the resolve step the pipeline names: evaluate, expand, flatten, and
  hand layout a tree with no spec language left in it
- fill is the exception and has to be: its iteration count is a layout
  result, so a fill node keeps its subtree and the scope to resolve it
  in, and layout asks for one iteration at a time as it packs them
- fill under an auto ancestor is rejected at parse and flagged again at
  resolve; measuring it as zero is the empty notes section decision 5
  exists to prevent
- range, list and timeSequence join the same whitelist the arithmetic
  is in, so the schema and the evaluator cannot drift
- timeSequence is end-exclusive: a row is the slot that starts at its
  label, and an empty range is no rows rather than an error
- a number or boolean in a text slot is written out as it reads - the
  one coercion in a language with no string operations
- the M2 notes fixture now sizes its chain, which is the new rule doing
  its job rather than a change of intent

Closes #13

Bolt added 2 commits

2026-08-18 13:40:45 +01:00

fix(layout): keep a fill out of the sized nodes 1fd721c455

- isSized claimed every kind but text and circle carries width and
  height, so measuring a container holding a fill threw a TypeError and
  took the whole sheet down with it
- list the sized kinds rather than subtract them, so a kind without a
  size of its own is left out by default

fix(layout): stop measuring a fill along its container 0802818bc6

- what a fill packs is the space the container was given, not an
  iteration measured up front, and an auto container has none to give
- latent today: the spec rejects a fill under an auto parent, so the
  figure this drops was always discarded

TJC requested review from TJC

2026-08-18 13:41:13 +01:00

Bolt commented

2026-08-18 13:41:41 +01:00

Author

Collaborator

Review of this branch at medium effort, five findings. The two that matter are fixed here.

Fixed

isSized claimed every kind but text and circle carried width and height, and ResolvedFill carries neither. Measuring any container holding a fill threw a TypeError out of intrinsicExtent; not a SpecError, so attempt rethrew it and layoutTemplate took the whole sheet down - the white screen decision 6 exists to prevent. Reachable from a parse-valid template: row > stack (auto width, stretched height) > rect { repeat: fill }. Regression test added. (1fd721c)
measureContainer added a fill's sample extent along the container's main axis, which is not what a fill contributes there. Latent, since the spec rejects a fill under an auto parent and the figure was always discarded, but it disagreed with expandFills, which already skips fills. (0802818)

Left for a follow-up

edgesOf and num (resolve.ts:297, :323) accept negative padding, gap and radius, while theme lengths go through mmSchema (positive()). padding: -50 draws children off the page with an empty errors array.
rootOf (sheet.ts:49) reuses template.root.id for the synthetic wrapper stack, so under decision 21 a wrapper-charged cap error is indistinguishable from an iteration's.
'align' in parent ? parent.align : 'stretch' (fill.ts:59) misreads an explicitly-undefined align; parent.align ?? 'stretch' is the robust form. Same shape at fill.ts:45 and resolve.ts:288.

Test plan

pnpm test 254 passed, tsc --noEmit clean, eslint . clean

Review of this branch at medium effort, five findings. The two that matter are fixed here. ## Fixed - `isSized` claimed every kind but `text` and `circle` carried `width` and `height`, and `ResolvedFill` carries neither. Measuring any container holding a `fill` threw a `TypeError` out of `intrinsicExtent`; not a `SpecError`, so `attempt` rethrew it and `layoutTemplate` took the whole sheet down - the white screen decision 6 exists to prevent. Reachable from a parse-valid template: `row > stack (auto width, stretched height) > rect { repeat: fill }`. Regression test added. (1fd721c) - `measureContainer` added a fill's sample extent along the container's main axis, which is not what a fill contributes there. Latent, since the spec rejects a fill under an auto parent and the figure was always discarded, but it disagreed with `expandFills`, which already skips fills. (0802818) ## Left for a follow-up - `edgesOf` and `num` (`resolve.ts:297`, `:323`) accept negative padding, gap and radius, while theme lengths go through `mmSchema` (`positive()`). `padding: -50` draws children off the page with an empty `errors` array. - `rootOf` (`sheet.ts:49`) reuses `template.root.id` for the synthetic wrapper stack, so under decision 21 a wrapper-charged cap error is indistinguishable from an iteration's. - `'align' in parent ? parent.align : 'stretch'` (`fill.ts:59`) misreads an explicitly-undefined align; `parent.align ?? 'stretch'` is the robust form. Same shape at `fill.ts:45` and `resolve.ts:288`. ## Test plan - `pnpm test` 254 passed, `tsc --noEmit` clean, `eslint .` clean

TJC merged commit 091c11d4ee into main

2026-08-18 13:42:50 +01:00

TJC referenced this pull request from a commit

2026-08-18 13:42:52 +01:00

Merge pull request 'feat(layout): add the layout engine, generators and fill' (#30) from feat/m3-layout-engine into main

TJC deleted branch feat/m3-layout-engine

2026-08-18 13:42:55 +01:00

Rows
Columns

feat(layout): add the layout engine, generators and fill #30

Summary

Test plan

Fixed

Left for a follow-up

Test plan