PSTricks semantic layer + conformance harness against real PSTricks

[pyramation]

Two commits: the semantic-layer rewrite, and a harness that checks it against genuine PSTricks.

Semantic layer

Parser rebuilt around an ordered AST with diagnostics and source-order rendering. Command coverage roughly doubles — pscustom, psgrid, psdots, psellipse, psbezier, pscurve, psecurve, psccurve, pswedge, multido — plus text macros, header environments, and a new expression parser in @latex2js/utils for algebraic plot bodies.

Adds a Vite playground with a headless render page, a Playwright suite over every corpus example, and CI running both.

pnpm test is green across all 10 projects.

Conformance harness (tools/pstricks-conformance/)

Renders the same source twice — LaTeX2JS and real PSTricks in pyramation/pstricks-latex — and ranks pairs by disagreement. Ground truth is the spec made visual.

• render-examples.mjs — splits the 24 example files into 57 pictures, renders each. 53/57 succeed in ~30s.
• fuzz-corpus.mjs — 336 systematic cases: every command × style axis, plus draw-order and degenerate-input probes. 334/336 render.
• compare.mjs — scores pairs on ink / colour histogram / layout occupancy.

The score is deliberately a triage ordering, not a gate — SVG and Ghostscript output cannot match pixel for pixel.

Why this is needed: the existing corpus cannot measure coverage. \psline appears 256 times across all examples while every command added here appears once, and \psecurve appears zero times despite being implemented.

What the harness already found

Not fixed in this PR — recorded here so they are tracked.

1. Interactive re-render discards draw order. psgraph.ts pspicture()/userEvent removes and re-appends .userline/.psplot, so they jump to the top layer and reorder among themselves on first pointer move. The initial draw respects source order; the interactive path does not. A diagram is correct until the pointer moves over it.
2. \pswedge draws concave arcs. 17-pie renders as a petal/star rather than a filled pie. Source is pure PSTricks, no shims.
3. Only fillstyle=solid is implemented, and the fallbacks disagree — hlines becomes a solid fill on pspolygon/psarc but no fill on psellipse/pswedge.
4. \section renders as literal source text, and the proof environment emits an undefined \qed.
5. Starred forms (\psframe*) honour fillcolor; PSTricks fills with linecolor and ignores it.
6. A malformed macro copy-pasted into 3 corpus files — \userline...(2,2)(sin(x)}{-y}, where ( should be {. LaTeX rejects it; LaTeX2JS parses it silently.

Dialect divergences needing a decision

LaTeX2JS accepts input PSTricks rejects: pow(a,b), infix bodies without algebraic=true, variable plot bounds ({alpha-3}), a bare algebraic key, plotpoints=1 (PSTricks requires at least 2), implicit multiplication (0.5x^3). Each is keep-the-extension or conform.

🤖 Generated with Claude Code

https://claude.ai/code/session_01HBYcp3Go9naDWgFThChaDf

[pyramation]

Fixes applied since the PR opened

Nine defects found and fixed by comparing against real PSTricks. All 177 tests green; every fix carries a regression test over the interaction that produced it, because each of these rendered visibly wrong while the suite stayed green.

Rendering

• \pswedge and \psarc swept the wrong way. Y inverts the axis, so PSTricks' counter-clockwise sweep is SVG sweep-flag 0; both passed 1 and traced the complementary arc. The pie chart rendered as a five-pointed star. Sweep span is now normalised into [0, 2π) so a reversed end angle takes the long way round, and a full turn is emitted as two half-turns.
• Draw order was discarded on interaction. The pointer handler removed .userline/.psplot and re-appended them, putting them above every later shape and regrouping by command type. A diagram was correct until the pointer first crossed it. Drawing now happens into a layer the handler replaces wholesale, so the interactive path uses the same source-order walk as the initial one.
• Only fillstyle=solid existed, and the fallbacks disagreed — hlines became a solid fill on pspolygon/psarc and no fill on psellipse/pswedge. All twelve sites route through one resolver; hlines, vlines and crosshatch now render as real SVG patterns honouring hatchwidth/hatchsep/hatchangle/hatchcolor.
• psaxes never drew tick labels, and ticks/labels were parsed then discarded, so ticks=none still drew ticks. Adding labels exposed two more: ticks stepped from the end of the axis (marking an axis spanning -3.5 to 3.5 at half-integers), and because Y inverts, the y loop's condition was false immediately — no y tick had ever been drawn.
• xcolor tints were passed through raw. fillcolor=gray!40 reached the browser unusable, and an unparsable fill falls back to black — example 10's light grey plane rendered as a solid black shape covering the vectors drawn over it.

Text

• \section rendered as literal source. Not unimplemented — silently skipped. Text transforms matched the pristine line but replaced inside the already-mutated contents, so once -- became an en dash the replace was a no-op. This was a whole class: \subsection, \textbf, \textit, \footnote over dashes or quotes all failed the same way.
• Every proof ended in a MathJax error box. \end{proof} emitted the literal string $\qed$, and MathJax defines no \qed. Now the open square amsthm uses, set flush right.
• List items broke at their first macro. EnvContent matches Command before Line, so one source line arrived as several nodes. The same text outside a list was fine, which is why it read as a styling quirk.
• SVGSelection.text threw a ReferenceError in any DOM not exposing SVGTextElement as a global, jsdom included.

Corpus

• A malformed \userline...(2,2)(sin(x)}{-y} — ( where a brace belongs — copied into three files, including a verbatim block publishing it as documentation. It was the only thing keeping four pictures from compiling; all 57 now rasterize.

Harness corrections (it was wrong twice, in ways that looked like renderer bugs)

• The splitter dropped \psset above a picture, so references rendered at the wrong scale — 16-curves came out looking rotated.
• \uservariable{alpha}(0.1,0){x} declares its starting value; the harness pinned every such variable to 1, producing a faithful reference of a different diagram.
• Ink is now measured against the drawing's bounding box rather than the canvas, since the two renderers pad completely differently.

Where it stands

18 pairs ranked, 3 listed separately as structurally incomparable (slider chrome, whole-document typesetting). The two lowest ranked are the starred-fill divergence — \psframe*[fillcolor=...] fills with fillcolor here and with linecolor in PSTricks — where the lenient reading plainly matches author intent (the bar chart wants blue bars). Their gap is entirely colour; ink and layout sit at 92% and above. Left unchanged and documented, along with document numbering and CSS colour names, in tools/pstricks-conformance/README.md.

🤖 Generated with Claude Code

[pyramation]

CI was failing on every run — fixed, then extended

Root cause: pnpm/action-setup refuses to install when the version is named both in its own version: input and in package.json's packageManager. Every run died in setup, before installing a single dependency — which is why they all failed in ~7 seconds. Dropping the input leaves packageManager as the single declaration, which is also what keeps local and CI on the same pnpm.

Both original jobs now pass in ~80s.

Third job: PSTricks conformance

Renders every example with genuine PSTricks in pyramation/pstricks-latex, renders the same sources with LaTeX2JS through Playwright, and scores them. The runner is amd64, so the image runs natively there rather than emulated as it does on Apple silicon.

What gates and what doesn't: a picture that stops compiling under PSTricks fails the job — that's deterministic and a definite regression. The similarity score is a heuristic between two different rasterizers and is published, not enforced; SVG and Ghostscript output can never match pixel for pixel, so a threshold there would only produce noise.

158s total, and the breakdown says where it goes:

step	time
Pull the PSTricks image	67s
Install Playwright browsers	35s
Render 66 pictures with real PSTricks	13s
Render 33 with LaTeX2JS	19s
Score them	1s

The image pull is the bulk. If that becomes annoying it's worth a paths filter so the job only runs when the renderer or corpus changes — happy to add it, but I'd rather see it run on everything first.

Operational fixes

• Superseded runs are cancelled instead of queueing. This branch had eight failed runs stacked up today.
• Job timeouts so a hang costs minutes rather than the default six hours.
• Playwright browsers cached against their exact version, so an upgrade misses the cache rather than restoring browsers the new Playwright then refuses.
• e2e publishes one browsable gallery page beside the loose PNGs — a self-contained HTML file with every rendering inlined, which is far easier to review from an artifact download than 33 separate images.

Artifacts per run

• example-renderings (2.5 MB) — every LaTeX2JS render plus gallery.html
• pstricks-conformance (0.8 MB) — the side-by-side comparison plus every PSTricks reference

🤖 Generated with Claude Code

[pyramation]

Filed #27 with a full plan for the dialect work and the divergences found while building the conformance harness.

The short version: LaTeX2JS accepts a dialect of PSTricks that is partly a deliberate extension and partly accident, and nothing currently distinguishes the two. #27 proposes a \psset{dialect=latex2js} key so a document declares which language it is written in, fixes the accidents underneath it, and records the rest as decisions.

Three things from it are strictly fixes with no design question attached, and could land before any dialect work:

• X/Y coerce NaN to 0, so a broken plot renders as a flat line on the axis instead of failing. This is the mechanism behind most of what this PR fixed.
• arrowscale (28 uses), plotstyle (3) and showorigin (6) are parsed and ignored. plotstyle=dots is why the tangent-point dots are missing from every picture in graph.tex while the PSTricks reference shows them.
• Pi is undefined — the table has pi, PI and π but not pst-plot's own spelling.

Nothing in #27 blocks this PR.