#117 Suppress pdfjs Canvas/DOM warnings in text-only PDF extraction

[Anonymous]

Originally created by: yablokolabs

Summary

os.fs.read_document can extract PDF text successfully while pdfjs-dist emits optional Canvas/DOM polyfill warnings in a runtime where optional canvas support is absent, including packaged/SEA-style deployments. The warnings look like extraction failures even though Atomic deliberately does not render fonts, images, or graphics.

This is finding §15 from Yabloko Labs' evaluation.

Current behavior

The extractor dynamically imports the pdfjs legacy build and deliberately uses text-only settings:

• pdf-extractor.ts (github.com)
• pdf-extractor.ts (github.com)

verbosity: 0 is passed to getDocument, but optional-dependency warnings are emitted during module import, before that per-document setting applies. With pdfjs-dist@4.10.38 and @napi-rs/canvas unavailable, the warning family is:

Warning: Cannot load "@napi-rs/canvas" package: ...
Warning: Cannot polyfill `DOMMatrix`, rendering may be broken.
Warning: Cannot polyfill `ImageData`, rendering may be broken.
Warning: Cannot polyfill `Path2D`, rendering may be broken.

A normal npm installation may include the optional canvas package and hide the symptom. The relevant regression environments are a fresh install with optional dependencies omitted and the packaged SEA/bundle where the native optional package is not shipped.

The tool already has a structured warnings field for genuine page-extraction problems:

• pdf-extractor.ts (github.com)

Proposed behavior

Provide a quiet text-only loading path that avoids these known, non-actionable rendering warnings when canvas support is absent. Do not require a native canvas package just to extract glyph text, and do not hide genuine parse or page-extraction errors.

The implementation must not intercept or replace process-global console methods; that is unsafe under concurrent reads and can hide unrelated diagnostics.

Acceptance criteria

• The regression test runs the PDF module import in a fresh child process, not the test runner's already-imported module cache.
• It reproduces a deployment without @napi-rs/canvas—for example npm ci --omit=optional or the actual SEA bundle—and fails on the current warning strings above.
• A successful text-only PDF read in that environment emits none of the known optional Canvas/DOM warnings to stdout or stderr.
• Missing graphics/rendering support does not require installing @napi-rs/canvas when only text is requested.
• Corrupt PDFs and page-level extraction failures remain visible through errors or details.warnings.
• Warning handling is scoped to the pdfjs loading strategy; process-global console interception is prohibited.
• Concurrent PDF reads cannot interfere with process-wide logging.
• CI covers at least one optional-dependency-absent or packaged-bundle target in addition to a normal install.