Add vl-convert-fontsource crate and explicit font install

[jonmmease]

Summary

Add support for downloading, caching, and registering fonts from the Fontsource catalog (which includes Google Fonts and other open-source font families). Fonts are fetched on demand, cached to disk with LRU eviction, and loaded into fontdb for use in Vega/Vega-Lite rendering.

Motivation

vl-convert renders charts to static images using server-side text layout, which requires the actual font files. Previously, users had to manually download font files and point vl-convert at the directory. This PR adds automatic font downloading from Fontsource so users can register any of ~1,800 open-source font families by name.

Usage

CLI

# Use Roboto when converting a Vega-Lite spec to PNG
vl-convert vl2png -i chart.vl.json -o chart.png --fontsource-font "Roboto"

# Multiple fonts
vl-convert vl2svg -i chart.vl.json -o chart.svg \
  --fontsource-font "Roboto" \
  --fontsource-font "Playfair Display"

Python

import vl_convert as vlc

# Register a font globally (persists for the session)
vlc.register_fontsource_font("Roboto")

# Register specific variants only (smaller download)
vlc.register_fontsource_font("Roboto", variants=[(400, "normal"), (700, "normal")])

# Configure disk cache size (default 512 MB)
vlc.configure(fontsource_cache_size_mb=256)

Rust

use vl_convert_rs::text::register_fontsource_font;

// Register globally (persists for the session)
register_fontsource_font("Roboto", None).await?;

Per-request fonts via VlOpts/VgOpts:

use vl_convert_rs::converter::{VlOpts, VgOpts, FontsourceFontRequest};
use vl_convert_fontsource::{FontStyle, VariantRequest};

// Vega-Lite conversion with specific font variants
let svg = converter.vegalite_to_svg(vl_spec, VlOpts {
    fontsource_fonts: Some(vec![
        FontsourceFontRequest {
            family: "Roboto".into(),
            variants: Some(vec![VariantRequest { weight: 400, style: FontStyle::Normal }]),
        },
    ]),
    ..Default::default()
}).await?;

// Vega conversion with all variants of a font
let png = converter.vega_to_png(vg_spec, VgOpts {
    fontsource_fonts: Some(vec![
        FontsourceFontRequest {
            family: "Playfair Display".into(),
            variants: None, // downloads all available variants
        },
    ]),
    ..Default::default()
}, None, None).await?;

Per request fonts don't remain in memory after the conversion, so are a better solution for long running server processes. This is also the foundation that auto font detection will use in the following PR.

Architecture

New crate: vl-convert-fontsource

Self-contained crate with no dependencies on other vl-convert workspace crates. The core library (API client, disk cache, variant resolution) has no fontdb dependency — fontdb integration is behind an optional fontdb feature flag. This design means the crate could be spun out of the vl-convert workspace as an independent Fontsource client if there is future interest.

Core functionality (no fontdb required):

• Metadata fetch: Downloads font metadata from the Fontsource API (/v1/fonts/{id}), cached to {cache_dir}/metadata/
• Blob download: Fetches TTF files with parallel downloads (default 8 concurrent), cached to {cache_dir}/blobs/
• LRU eviction: File-based LRU using mtime, with cross-process file locking (fs4)
• Variant filtering: Download all variants or a specific subset of (weight, style) pairs
• Cache integrity: Magic-bytes validation on cached blobs with self-healing (corrupt files are deleted and re-fetched)

With fontdb feature:

• FontsourceDatabaseExt trait for batch register/unregister on fontdb::Database
• RegisteredFontBatch for tracking registered face IDs

Public API: FontsourceClient, ClientConfig, LoadedFontBatch, VariantRequest, FontStyle, FontsourceError. With fontdb feature: FontsourceDatabaseExt, RegisteredFontBatch.

Integration: worker font overlay system

The converter worker pool gains a font overlay mechanism:

• FontBaselineSnapshot: Shared RwLock snapshot of the resolved fontdb::Database, cloned by workers at startup.
• Per-request overlays: Vec<LoadedFontBatch> flows through the VlConvertCommand enum. Workers call register_fontsource_batch to temporarily load request-scoped fonts, then unregister_fontsource_batch after rendering. Fonts specified via VlOpts/VgOpts fontsource_fonts field are transient.
• Persistent registration: register_fontsource_font() adds fonts to the global FONT_CONFIG and bumps the baseline version, making them available to all subsequent requests.

Other changes in converter.rs

• New VlConvertCommand variants (VgToJpeg, VgToPdf, VlToJpeg, VlToPdf, SvgToPng, SvgToJpeg, SvgToPdf) route all conversion paths through the worker command channel. This was necessary so per-request fontsource fonts can be applied consistently before any render.
• SVG-to-raster/PDF conversions that previously used free functions now go through the worker pool.

Environment variables

Variable	Effect
VL_CONVERT_FONT_CACHE_DIR	Override cache directory (set to "none" to disable disk caching)
VL_CONVERT_FONTSOURCE_API_URL	Override API base URL (for mirrors or proxies)

Default cache location: {dirs::cache_dir()}/vl-convert/fontsource/ (e.g., ~/Library/Caches/vl-convert/fontsource/ on macOS, ~/.cache/vl-convert/fontsource/ on Linux)

Testing

Integration test suite in vl-convert-fontsource/tests/load_fontsource.rs with a custom TestServer that tracks per-endpoint hit counts and max inflight concurrent requests.

Coverage includes:

• Cache behavior: cache hits skip network, corrupt metadata triggers re-fetch, corrupt blobs self-heal via magic-bytes check, LRU eviction keeps current request's fonts
• Download correctness: variant filtering (empty/unavailable errors), all-variants-when-none-specified
• Concurrency: parallel download bounding, in-process download deduplication
• Lifecycle: register/unregister via FontsourceDatabaseExt, idempotent unregister, async/blocking parity

Existing vl-convert-rs end-to-end tests continue to pass.

Known Limitations

• No metadata TTL: Cached font metadata is served until manually deleted. Fonts rarely change on Fontsource.
• TTF downloads: Downloads TTF files (larger than WOFF2) because fontdb requires TTF/OTF. Fontsource provides TTF for all fonts, so no fonts are excluded by this.
• CLI downloads all variants: --fontsource-font downloads all available variants. The Python API supports variant selection via the variants parameter.