[upgrade] Baysor: migrate Julia v0.7.1 → C++ cpp-0.8.2 (eliminates PARQUET_TO_CSV for Baysor paths)

[an-altosian]

Current state

• Pinned: Baysor v0.7.1 (Julia) at modules/local/baysor/run/main.nf:5
• Source: https://github.com/kharchenkolab/Baysor master branch
• Distribution: Julia binary, containerised via community.wave.seqera.io
• Modules: baysor/{preprocess, create_dataset, run, preview, segfree}
• Pipeline modes: image (cellpose → baysor), coordinate (proseg/segger → baysor), preview, segfree

Proposed upgrade

• Target: Baysor cpp-0.8.2 (C++ port) — released 2026-04-30
• Source: https://github.com/kharchenkolab/Baysor cpp branch — https://github.com/kharchenkolab/Baysor/releases/tag/cpp-0.8.2
• Distribution: Native C++ binary via CMake, or Docker (image TBD)
• Project description: "a faithful C++ implementation of the current Baysor algorithm with a more efficient runtime and broader modern I/O support"
• Release notes for 0.8.2: "improves cross-platform builds, adds some optimizations for large runs (e.g. Xenium 5K)"

Why upgrade

1. Direct parquet support — the C++ port natively reads parquet (no Julia Parquet.jl Zstd issue). This would let us delete the PARQUET_TO_CSV step in BAYSOR_GENERATE_PREVIEW and BAYSOR_RUN_TRANSCRIPTS_PARQUET (tiled). That step OOMs on the Atera (Xenium WTA 18k-target) at default memory — see the Atera compatibility report (docs/2026-05-28_REVIEW_atera-on-spatialaxe-compatibility.md).
2. Performance — C++ rewrite is "much more efficient" per project description. Optimizations explicitly target Xenium 5K-panel scale, which is directly relevant to Atera ~18k-target workloads.
3. Direct experiment.xenium input — could simplify bundle-staging logic for image- and coordinate-mode subworkflows.
4. Algorithmic continuity — cpp port preserves the v0.7.1 algorithm, so segmentation behavior should match.

Migration plan

• Identify or build official container for cpp-0.8.2 (binary build vs Docker)
• Update containers in baysor/run, baysor/preview, baysor/segfree, baysor/preprocess, baysor/create_dataset
• Audit CLI argument compatibility — cpp may differ from Julia (e.g., column flag names, --scale)
• Drop or refactor PARQUET_TO_CSV calls in baysor_generate_preview and baysor_run_transcripts_parquet subworkflows
• If experiment.xenium direct input works, simplify BAYSOR_PREPROCESS_TRANSCRIPTS accordingly
• End-to-end tests on a Xenium v1 bundle (XOA 4.x) and Atera Cell Pellet

Risks

• CLI compatibility: needs per-command verification
• Output format compatibility: cpp may emit different polygon / transcript-assignment file structures; downstream XR import needs validation
• Container ecosystem: no Wave container yet known — may need a request to Seqera or a custom build

Cross-links

• Related (paired): Punkst / Ficture C++ migration — together these eliminate PARQUET_TO_CSV entirely from the pipeline.
• Triggered by: Atera compatibility session 2026-05-28.