taxonomy: offline NCBI taxonomy table, embedded in the Docker image

[hyphaltip]

Implements spec item A1 (offline taxonomy; docs/superpowers/specs/2026-10-04-packaging-service-and-reports-design.md), per curator request "embed/use the slim table (all taxa) in the container".

What

• matpredict curate-db build-taxonomy --taxdump <dir|taxdmp_*.zip|taxdump.tar.gz> --snapshot DATE --out ncbi_taxonomy.tsv.zst: slim all-taxa table (taxid, parent, rank, nuclear genetic code, scientific name; merged ids). Dump files may be plain, .gz or .zst.
• src/MATPredict/db/local_taxonomy.py: lookups with efetch semantics (lineage root-first without root/self; phylum = phylum-rank ancestor; own genetic code; merged ids followed).
• db/taxonomy.py: default_lineage_taxids, default_lineage_phylum_name, default_genetic_code use the table first when $MATPREDICT_TAXONOMY is set. $MATPREDICT_OFFLINE=1: never call NCBI; a missing taxid raises and the router records it in routing_error.
• Reports gain taxonomy_source (e.g. local NCBI taxonomy snapshot 2026-10-01).
• Dockerfile: builds the table from taxdmp_2026-10-01.zip (SHA-256 pinned; NCBI publishes no checksum), sets MATPREDICT_TAXONOMY and MATPREDICT_OFFLINE=1. CI smoke test checks a lookup with --network none.
• README: configuration table + "Offline taxonomy" section. CHANGELOG: Unreleased.

Measured

• Table: 3,016,750 taxa, 101,411 merged ids, 23.7 MB; build 53 s; load 1.96 s, 76 MB RSS (streamed into int arrays; a first version used 1.45 GB).
• vs 8,185 cached NCBI efetch answers: phylum 8,185/8,185, genetic code 8,185/8,185, lineage 8,179/8,185 (the 6 were re-parented by NCBI between fetch and snapshot, e.g. C. gattii VGV 2268382); 21 taxids deleted in the snapshot.
• End to end, offline, empty cache (P. tannophilus, --taxid 669874): 0 NCBI calls; genetic code 26 from the table; calls identical to the online run (results/2026-10-04_offline_taxonomy_check/).
• Tests: 1017 passed (10 new in tests/db/test_local_taxonomy.py, on Python 3.12 without compression.zstd).

Not measured yet: image size with the table, and the CI image build (runs on this PR).

🤖 Generated with Claude Code
Claude-Session: https://claude.ai/code/session_01JkhYFEPaa1WenE5MtA8Qya

[hyphaltip]

CI fix (018a549): the image build failed in build-taxonomy with No such file or directory: 'zstd'. pixi.lock pins Python 3.13, which has no compression.zstd (3.14+), and the fallback called the zstd binary, which is not on PATH in that RUN step. My earlier local checks ran on an older Python 3.14 environment, which hid this.

Fix: zstandard (conda-forge 0.25.0) is a dependency for Python < 3.14 (pixi.toml, pyproject, environment.yml; relock adds only zstandard, cffi, pycparser); the table is streamed through it.

Checked on the locked default environment (Python 3.13.15) with an empty PATH: build 87 s (2.3 GB peak), table content identical to the 3.14 build, load 1.72 s / 74 MB. Unit tests: 1017 passed on the 3.12 test env, which used the zstd-binary fallback path (no zstandard there); pytest is not in the default environment, so the tests did not run on 3.13.