refactor(traits): build the traits table in Rust and drop the Python it replaces

.github/workflows/guard.yml

@@ -279,6 +279,29 @@ jobs:
           grep -E 'planted_failure.*FAILED' "$log" || { cat "$log"; echo "ERROR: the run failed, but not on the planted test"; exit 1; }
           echo "Control: the planted test failed, as it must."
 
+      - name: Traits controls (diplotype order, REF check and missing phase rows)
+        env:
+          CARGO_BUILD_JOBS: 3
+        run: |
+          src=crates/og-steps/src/traits.rs
+          saved="${RUNNER_TEMP}/traits.rs"
+          log="${RUNNER_TEMP}/traits-controls.log"
+          cp -f "$src" "$saved"
+          trap 'cp -f "$saved" "$src"' EXIT
+          perl -0pi -e 's/(fn allowed\b.*?)(\n    out\n)/$1\n    out.reverse();$2/s; s/if c\[2\]\.chars\(\)\.next\(\) != s\.row\["ref"\]\.chars\(\)\.next\(\)/if c[2].is_empty()/; s/(fn phase_vcf\b.*?)(    let mut out = AtomicFile::create\(path\))/$1    sites.clear();\n$2/s' "$src"
+          cmp -s "$saved" "$src" && { echo "ERROR: no traits control was planted"; exit 1; }
+          if cargo test --locked -p og-steps traits::tests --no-fail-fast > "$log" 2>&1; then
+            cat "$log"
+            echo "ERROR: traits accepted the planted faults"
+            exit 1
+          fi
+          cat "$log"
+          for test in apoe_grid_keeps_haplotype_order_and_missing_calls stream_prefers_first_snv_but_checks_every_ref_and_bounds_memory phase_requires_one_block_and_only_multiple_hets_are_exported; do
+            grep -E "traits::tests::${test} .*FAILED" "$log" || { echo "ERROR: $test did not catch its fault"; exit 1; }
+          done
+          cp -f "$saved" "$src"
+          trap - EXIT
+
       - name: cargo test
         run: cargo test --workspace --locked --no-fail-fast
 
@@ -351,7 +374,7 @@ jobs:
                    tests/test_results_index.py \
                    tests/test_pgx_outside_calls.py tests/test_hla_second_caller.py tests/test_paraphase_report.py \
                    tests/test_copy_number.py tests/test_hba_copy_number.py \
-                   tests/test_lpa_kiv2.py tests/test_traits_report.py tests/test_callability.py; do
+                   tests/test_lpa_kiv2.py tests/test_traits_report.py tests/test_traits_table.py tests/test_genotype_rules_indels.py tests/test_callability.py; do
             rc=0
             echo "::group::${t}"
             case "$t" in
@@ -363,6 +386,43 @@ jobs:
           done
           exit "$fail"
 
+      - name: Traits independent expectations reject changed reference labels and marker metadata
+        working-directory: .
+        env:
+          OG: ${{ github.workspace }}/bin/og
+          OG_REQUIRED: 1
+        run: |
+          python3 - <<'PY'
+          import contextlib, io, pathlib, runpy, tempfile
+          test = runpy.run_path("tests/test_traits_report.py")["main"].__globals__
+          original = pathlib.Path(test["RULES"]).read_text()
+          with tempfile.TemporaryDirectory() as work:
+              for rid, key, label in [("APOL1", "G0/G0", "G0/G0"), ("CAD9p21", "nonrisk/nonrisk", "no risk copy")]:
+                  path = pathlib.Path(work) / "rules.tsv"
+                  old = f"combination\t{rid}\t{key}\t{label}\t"
+                  assert old in original
+                  path.write_text(original.replace(old, f"combination\t{rid}\t{key}\tWRONG_REFERENCE\t"))
+                  test["RULES"], test["FAILS"] = str(path), 0
+                  log = io.StringIO()
+                  with contextlib.redirect_stdout(log):
+                      rc = test["main"]()
+                  print(log.getvalue())
+                  assert rc == 1 and test["FAILS"] == 1
+                  assert f"[FAIL] {rid}, reference blocks only: {label}" in log.getvalue()
+                  print(f"Control: {rid} WRONG_REFERENCE failed its independent expectation")
+          marker = test["tr"].MARKERS["SMN1_27134"]
+          rs, chrom, pos, ref, alt = marker["dbsnp"]
+          marker["dbsnp"] = rs, chrom, pos + 1, ref, alt
+          test["FAILS"] = 0
+          log = io.StringIO()
+          with contextlib.redirect_stdout(log):
+              test["marker_checks"]()
+          print(log.getvalue())
+          assert test["FAILS"] == 1
+          assert "[FAIL] marker SMN1_27134: the pinned dbSNP record agrees" in log.getvalue()
+          print("Control: retained SMN1 off-copy metadata failed its independent expectation")
+          PY
+
       # The fake-docker job builds no og, so its cn-engine-calls case says it
       # is skipped there; here bin/og is built and the case must run.
       - name: The fake-docker case cn-engine-calls (steps 35, 41 and 46 through og cn) with og

CHANGELOG.md

@@ -4,6 +4,9 @@ Each release's notes. From 1.0.0 on, a major or minor release has hand-written n
 
 ## Unreleased
 
+- Step 42 reads traits and genotype rules with `og traits table`, exports WhatsHap sites with `og traits phase-sites`, and checks the catalogs with `og traits check`. The replaced Python evaluator is removed. Report columns and readings stay the same. Genotypes outside the catalog are streamed past, and write errors now report a controlled error with the path instead of a traceback. Both tables are staged before either is published. A staging failure or a directory destination leaves both previous tables intact. Each file is then replaced by a separate atomic rename. The pair is not a transaction: a failure of the second rename can leave the first table updated. Catalogs and output remain plain TSV; quotes never hide a tab or bypass validation.
+- The unused Python copy-number engine is removed. Its historical catalog tests materialise a pinned test oracle and keep testing the Rust commands.
+
 - The Rust summary's CHIP reader marks a depth table with NaN unreadable instead of sorting it into a usable sensitivity estimate. Its unindexed APOL1 G2 reader keeps only records overlapping the deletion span and stops past that span on sorted input.
 
 - The Rust `og` binary now reads the rest of the summary's own sections: `og summary collect --only c`. It reads 32 sections of summary.json (couple, parascopy, paralogs, hla, prs, ancestry_fine, manta, delly, cnvpytor, sv_consensus, annotsv, expansions and their readings, telomere, roh, haplogroup, y_haplogroup, y_tree, y_str, consumer_export, bloodtype, lpa, vntyper, locityper, mito, traits with the APOL1 G2 call, archaic, ancient, viral, exomiser, chip, mei and mosaic). Nothing calls it yet: the reports still read the summary that `bin/collect_summary.py` writes. The sections are the same as the Python's on 15 new parity cases, compared as JSON with key order. The readers run on a thread pool, one per section. On a synthetic sample folder with every section and a 1,000,000-record VCF it takes 0.33 s and 34 MB, where the Python readers took 0.33 s and 41 to 59 MB. What differs from the Python: