[DylanCosto]

Summary

• New command, pgsum rows: builds GenomeOS's per-term PGS evidence batches, byte for byte what GenomeOS's Python builder (pgs_pgsum_rows._build) writes:

  • the rows, merging the evidence rows with each term's source row, description and orientation from a GenomeOS catalog pack;
  • each batch document's digest;
  • the exact checkpoint delta.

  src/rows.rs lists every Python function it mirrors (canonical JSON with Python's float repr, decimal.Decimal arithmetic, describe, orientation, weight, the reducer).

• Non-ASCII source lines: exit with status 3, so GenomeOS builds that run in Python. Python's text rules differ from ASCII ones only there.

• evidence output unchanged: its rows now come from for_each_row, and a record met twice is recognised by its line instead of by re-serializing it. About 40% faster.

• Allocator: mimalloc. With glibc malloc, rows built in one thread and freed in the other were about 3× slower.

• Version: 0.6.0-dev. Tagging and releasing 0.6.0 is a separate step.

Checks (HG002, run D's inputs)

• 354 sample batches (first two and last of all 126 report models) are identical to GenomeOS's Python builder: digest, delta and whole document bytes.
• 417 batches (830,602 terms, two whole models) rebuilt natively match the digests in the case's signed batch chain.
• evidence output is byte-identical to this branch's base (main 74ecb0a) on 100,000-term ranges of three models.
• cargo test --release: 134 passed, 0 failed.
• Speed: 200,000 terms in 4.4 s against about 28 s in Python (VM under load). A full 61 M-term rebuild against every signed batch runs tonight.

The GenomeOS side is DylanCosto/GenomeOS branch pgs-rows-rust-20261022. It uses rows when the pinned pgsum has it, rebuilds each run's first batch in Python as a check, and falls back to Python for older pgsum.

🤖 Generated with Claude Code

[DylanCosto]

Added 498358a for review finding 3 (round 2): Task::parse reads every number token of the task back and requires it to print to the same text Python wrote; otherwise pgsum rows exits 3 and GenomeOS builds that run in Python. This covers a float serde_json reads 1 ULP off, or an integer beyond 64 bits. There is no crate-wide serde_json feature change. Tests: 135 passed. HG002: 354 of 354 sample batches still identical to Python.