Imported from SakshamDiwan/DistributedPanGA (src/AGENTS.md). Install upstream with npx skills add SakshamDiwan/DistributedPanGA --skill src. Copyright stays with the author.

src/ — implementation

Full symbol-level map: docs/agent/architecture.md. Constraints summary: root AGENTS.md.

Layering, innermost first: constants.h → record/tables/pack/work_split → extract → sort → dsort (MPI) → kmer_adapter → contig_assignment → fastga_pipeline → drivers (pga-*.c).

• One driver per phase; pga-mpi-merge.c is the final one and defines the FastGA externs (NTHREADS, FREQ, IBYTE, NCONTS, RankForPair, …) that fastga_pipeline.c declares extern. Adding an extern means updating every driver that links the pipeline (pga-merge.c too).
• fastga_pipeline.c is lifted FastGA code, kept close to upstream on purpose. Match its brace and indentation style there rather than the surrounding code, and keep edits minimal so it stays diffable against FastGA.
• kmer_adapter.h structs are cast to Kmer_Stream / Post_List. Field order and widths are load-bearing — never reorder, insert, or change a type.
• #include "constants.h" defines KMER 40, which collides with the runtime int KMER in the merge drivers; pga-mpi-merge.c #undef KMER after its includes. Preserve that if you add includes.
• Buffers get +1 byte for msd_sort's sentinel; Get_Contig_Piece callers pass buf+1 for its buffer[-1] sentinel. Both are easy to lose in a refactor and fail silently.
• sort_records hardcodes 4 threads; the merge drivers hardcode NTHREADS=1 and FREQ=10. No env var overrides these.
• Errors in MPI code paths call MPI_Abort with a distinct code (2 = local sort check, 3 = boundary check). Keep codes distinct.

Headers are not Makefile prerequisites — after editing any .h here, rebuild with make -B or the binary will be stale. See docs/agent/build.md.