Linux CLI follow-up: distributable archive and remaining gaps #59

Context

This follows the Linux CLI proposal in #5 and the first implementation in #6. Thanks to that work, the direction is clear. Current main has since gained a different portable engine and CLI path for Linux and Windows, but the README still says Linux is on the way: there is no downloadable Linux CLI release or Linux desktop app. On Linux, legacy .doc files are recognized for indexing but lack a reader.

Incremental work proposed here

• Build and test the current CLI on Ubuntu 24.04 x86_64 in CI.
• Produce a checksum-verified CPU tarball with the pinned llama.cpp and PDFium libraries and license notices. Keep the model as a separate verified download.
• Read legacy .doc text through optional LibreOffice Writer on Linux and document the supported formats and runtime dependencies.

The accompanying PR is based directly on current main. It follows #5 and credits #6, but does not require #6 to merge: that branch targets an older architecture and currently conflicts with main. This smaller step also does not carry over #6's presentation-deck support or .deb/.rpm packaging.

Implementation PR: #60 (erlonL:linux-cli-doc-archive → ARahim3:main).

Verification and known failures

On Linux x86_64, Swift 6.2.4, the CLI release build and archive smoke check passed. swift test --skip ImageTests passed 88 tests in 11 suites, including the new .doc test. The full test run reported five failures in existing HEIC/TIFF/WebP image fixtures with Ubuntu 24.04's FFmpeg 6.1; those fixtures need FFmpeg 8.1. The model-backed reference-vector test and macOS tests were not run for this change.

Remaining Linux work

• No Linux desktop shell or release channel yet; the tarball is a CI artifact, not a published release.
• No .deb/.rpm or aarch64 package. The tarball targets the Ubuntu 24.04 ABI and still needs system glibc, C++/OpenMP libraries and FFmpeg for media.
• No presentation-deck or RAW-image reader, and no OCR for exact words in scans or screenshots.
• Bring the image fixture suite green on the supported Linux media stack, then verify model-backed indexing and search end to end on Linux.

This issue tracks those gaps without treating the CLI tarball as feature parity with the macOS app.