fix: detect static text direction per block

[CodeinScrubs]

Description

Static mode currently resolves dir=auto once for the complete Markdown string. A mixed Hebrew/Arabic/Persian and English response therefore gives every block the same base direction, while code can inherit RTL.

This change keeps static content as one Markdown document but adds a rehype pass that assigns direction to each semantic block. Keeping one parse preserves cross-block references and footnotes. The public detector now uses strong-character content majority, with first-strong as the tie-breaker, so RTL-majority technical prose such as an answer beginning with an English identifier is not forced LTR. Rendered code-block containers are explicitly LTR.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality not to work as expected)
• Documentation update
• Performance improvement
• Refactoring (no functional changes)

Related Issues

Fixes #509

Changes Made

• Detect static Markdown direction per heading, paragraph, list item, quote, definition, caption, and table cell.
• Use content-majority direction with first-strong tie-breaking and exclude inline/fenced code from prose evidence.
• Keep code LTR at both HAST and custom code-block rendering layers.
• Add Hebrew, Persian/English, English, and fenced-code regression coverage.

Testing

• All existing tests pass
• Added new tests for the changes
• Manually tested the changes

Test Coverage

• pnpm --filter streamdown test: 73 files, 985 tests passed
• pnpm --filter streamdown build: ESM and declaration build passed
• Ultracite check on all six touched TypeScript/TSX files: passed
• git diff --check: passed

Screenshots/Demos

Not included. The regression is asserted against the rendered DOM: no document-level auto direction in static mode, per-block dir values, and an LTR code-block container.

Checklist

• My code follows the project's code style
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings or errors
• I have added tests that prove my fix is effective
• New and existing unit tests pass locally with my changes
• I have created a changeset (pnpm changeset)

Changeset

• I have created a changeset for these changes

Additional Notes

I maintain the MIT-licensed BidiLens project and used its per-semantic-block/content-majority model as design input. This PR intentionally implements the small policy natively instead of adding BidiLens as a dependency: Streamdown supports Node 18, while BidiLens 0.3 declares Node 22.12+, so a direct dependency would unnecessarily raise the host's runtime floor.

The Vercel preview check requires authorization from a Vercel team member because this PR comes from an external fork. Socket security and Vercel Agent Review passed; the preview authorization state is not a build or test failure in this patch.

[vercel]

Someone is attempting to deploy a commit to the Vercel Team on Vercel.

A member of the Team first needs to authorize it.