feat(ui): render markdown math

[ishowman]

Summary

Close #831, and extend Plannotator math rendering to support the common KaTeX/LLM-friendly delimiters \(...\) and \[...\], in addition to the existing $...$ and $$...$$ forms.

What changed

• Render inline math with both $...$ and \(...\).
• Render display math with both $$...$$ and \[...\].
• Keep escaped dollars and non-math text literal.
• Preserve accurate line tracking for display-math blocks so annotation exports stay correct.

Scope

This stays within the UI/parser layer and does not require server API changes.

Validation

• bun test packages/ui/components/InlineMarkdown.test.ts packages/ui/utils/parser.test.ts
• PATH=/Users/xxq/Documents/codes/plannotator/packages/ui/node_modules/.bin:$PATH bun run typecheck
• git diff --check

Notes

This update is aimed at markdown generated by LLMs and math rendered by KaTeX-style delimiters.

[XxxXMil]

Would it be possible to also support \(...\) and \[...\] delimiters? They’re common in LLM-generated Markdown and KaTeX auto-render supports them by default.

[ishowman]

Would it be possible to also support \(...\) and \[...\] delimiters? They’re common in LLM-generated Markdown and KaTeX auto-render supports them by default.

I've submitted new commit for it.Could you please verify if the result matches your expectations when you have spare time?

[XxxXMil]

Thanks for working on this! I tested the PR locally in Pi by installing the PR branch.

Some parts work well:

• Inline math with \(...\) renders correctly.
• Block math with \[...\] and $$...$$ renders correctly.

I also found a few issues during manual testing:

1. Inline math using $...$ is not very stable in my testing. Some $...$ expressions render, but others remain as raw text.
2. Highlighting anchoring doesn't shown on rendered math.

I attached screenshots showing these cases:

Raw text for the last screenshot:

In a linear Gaussian model, the prediction \( \hat{y} = \mathbf{x}^\top \boldsymbol{\beta} \) is compared against the target $y$ using the loss \( \mathcal{L}(\boldsymbol{\beta}) = \frac{1}{2N}\sum_{i=1}^{N}(y_i - \mathbf{x}_i^\top\boldsymbol{\beta})^2 \), while the noise term satisfies $\varepsilon \sim \mathcal{N}(0,\sigma^2)$.

In this case, $...$ works well.

[XxxXMil]

Small follow-up: I found the cause of the inconsistent $...$ behavior.

Testing inline `$...$` with spaces inside the delimiters:

$ \varepsilon \sim \mathcal{N}(0,\sigma^2) $

Testing inline `$...$` without spaces inside the delimiters:

$\varepsilon \sim \mathcal{N}(0,\sigma^2)$

$\varepsilon \sim \mathcal{N}(0,\sigma^2)$ renders correctly, but $ \varepsilon \sim \mathcal{N}(0,\sigma^2) $ stays as raw text because of the spaces just inside the delimiters.

This might be avoidable by prompting LLMs to emit $...$ without inner spaces?

Other test: