TeX Live 2026 + more

[adam-weber]

Upgrades the engine to TeX Live 2026, and bundles three workstreams that grew alongside it:

1. on-demand package normalization for build-required packages
2. a normalized-artifact (prebuilt) cache
3. data-driven font-file resolution w/ memory-management improvements

TeX Live 2026 upgrade

• Bump the busytex submodule to the TL2026 build (WASM engines + TLPDB).
• Regenerated bundle metadata: packages/bundles/file-manifest.json, bundles.json, file-to-package.json.
• Updated package.json (description → TeX Live 2026).

Package normalization (build-required packages)

A provider chain (packages/providers/) that turns fetched packages into ready-to-use runtime files:

• prebuilt / tds-zip (no build), ins-docstrip / dtx (run docstrip via a headless busytex engine).
• build-engine.ts + busytex-engine.ts: TeX build step on the same WASM the compiler uses; lazy-imported
  so the provider core stays browser-safe.
• Declarative overrides.json registry; structured unbuildable errors instead of crashes.
• Wired into processZipData (ctan-core.ts) so both the proxy and dev server pick it up.
• Fixes the fflate-under-Bun unzip crash (unzipSync).

Normalized-artifact (prebuilt) cache

• prebuilt-cache.ts: serve a pre-normalized artifact for build-required packages (e.g. acrotex/insdljs)
  instead of building in-request — the only way these work on the (FS-less) Cloudflare Worker.
• scripts/prebuild-packages.js: offline prebuild → packages/bundles/prebuilt/<pkg>.json (+ _index.json manifest);
  deterministic, atomic writes.
• Both proxies read it from disk; the Worker reads from R2 (gated by the manifest, loaded per isolate like the font index).
• Artifacts/manifest are generated, not committed (gitignored; shipped to R2/releases).

Font-file resolution

• font-index.ts builds a stem→package index from the TLPDB (committed font-file-to-package.json),
  so TeX font requests by file stem (phvr8t, …) resolve to their package on every runtime.
• TDS-routing fix in processExtractedFiles (anchored matchers so fonts land where kpathsea searches).

CTAN proxy / fetch improvements

• Shared ctanDownloadUrls() in ctan-core.ts — one source of truth for CTAN URL conventions
  (used by ctan-proxy.ts, serve-local.ts, and the prebuild).
• Resolve fontspec fonts by name; fix CTAN file vs. package routing.
• serve-local.ts: configurable port (SIGLUM_DEV_PORT).

Memory management & storage

• src/worker.js, src/storage.js, src/compiler.js: memory-cleanup improvements and format-cache handling,
  including capture & replay of preamble side-effect files so cached .fmt reuse doesn't drop runtime-written files.
• Lazy bundle loading + decompress-lz4.ts.

Tests

• New suite: tests/unit, tests/integration, tests/worker, tests/browser, shared tests/setup mocks/fixtures, vitest configs.
• Normalization: tests/unit/providers.test.ts, tests/integration/normalize-corpus.test.ts,
  tests/unit/prebuilt-cache.test.ts, and scripts/providers-bun-smoke.ts (Bun parity).
• All green locally: 395 vitest tests, tsc clean, bun-parity OK.

Docs

• docs/self-hosting.md: how to run the engine on any host (proxy contract, reference servers, where artifacts come from).

[adam-weber]

Closes #10