dna-decode 0.5.1__tar.gz

dna_decode-0.5.1/.claude/settings.local.json

@@ -0,0 +1,16 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(powershell -NoProfile -File /tmp/bundle_repo.ps1)",
+      "Bash(mkdir -p /c/Users/Farshad/AppData/Local/Temp/repo_verify)",
+      "Bash(rm -rf /c/Users/Farshad/AppData/Local/Temp/repo_verify/*)",
+      "Bash(powershell -NoProfile -Command \"Expand-Archive -Path 'C:\\\\Users\\\\Farshad\\\\Downloads\\\\dna_decode_repo.zip' -DestinationPath 'C:\\\\Users\\\\Farshad\\\\AppData\\\\Local\\\\Temp\\\\repo_verify' -Force\")",
+      "Read(//c/Users/Farshad/AppData/Local/Temp/**)",
+      "Bash(mkdir -p \"C:/Users/Farshad/AppData/Local/Temp/claude_bundle_for_codex_2026-05-26\")",
+      "Bash(rm -rf *)",
+      "Bash(mkdir -p plans wiki scripts dna_decode/data dna_decode/eval tests)",
+      "Bash(cp __TRACKED_VAR__/plans/Trait_Decoding_Roadmap.md plans/)",
+      "Bash(cp __TRACKED_VAR__/plans/EP_4_Non_AMR_Phenotype_Candidates.md plans/)"
+    ]
+  }
+}

dna_decode-0.5.1/.claude/testable-modules.md

@@ -0,0 +1,30 @@
+# Testable Modules
+<!-- Auto-maintained by /test-epilogue — do not edit manually -->
+
+- scripts/audit_cohort.py — Phase 2.5 cohort audit report generator; cohort overview + clade composition + metadata completeness + assembly QC quantiles + AST method breakdown + GO/WARN/NO-GO verdict; argparse + cohort-loader + AST-loader integration tests
+- scripts/build_mini_cohort.py — Gate B mini-cohort selector; picks N highest-assembly-quality strains per R/S class from a source cohort parquet (sort: contig_count asc, n50 desc); argparse + ValueError surface for missing class
+- scripts/build_validation_report_card.py — Anchor-4 report-card classify() 7-state cell machine (SCORED/POWERED_UNSCORED/UNDERPOWERED/ABSTAINS_BY_DESIGN/NOT_CENSUSED/LABEL_CONFOUNDED/NO_FREE_PHENOTYPE_SOURCE) + surface-precedence; main() read-only roll-up emit (JSON+MD; WIKI/ROOT redirectable for test); lineage-disclosure layer: load_lineage_metrics (canonical-keyed, malformed->empty) + build_lineage_block (scored/incomplete/not_computed) + _assert_weighted_renderable C3 guard (no weighted point without Wilson CI + eff-N)
+- scripts/cipro_curated_baseline.py — EP1 audit-infrastructure module-level constants + ABLATION_FEATURE_SETS contract + _load_amrfinder_features JSON parsing. 2-layer verdict (original_condition_4 + amended_condition_4 gated on no_POINT >= 0.773 OR mechanism_only >= 0.80). LR + XGB orchestration over multi-block ablation sets skipped.
+- scripts/cipro_mechanism_audit.py — EP1 audit-infrastructure pure-logic: _classify_symbol (gene → mechanism class with tolerant prefix match incl. acrR + bla* + qnr*), _is_synonymous_point (first AA == last AA detection), _parse_amrfinder_outputs (main.tsv POINTX → kind=mutation routing; mutations.tsv class+synonymous filter; cross-TSV dedupe). Docker AMRFinder invocation skipped.
+- scripts/cipro_mechanism_phenotype_merge.py — EP1 audit-infrastructure pure-logic: _classify_noise per-row mapping with strict PRIMARY_CIPRO_MECHANISMS = QRDR | plasmid_protect_modify; co_resistance_modifiers separated; mechanism_opacity_flag separates AMRFinder-incomplete from labels-noisy. Pre-curated-baseline gate orchestration skipped.
+- scripts/cipro_mic_audit.py — EP1 audit-infrastructure pure-logic: _parse_mic (strips <>= operators, handles NA variants) + _confidence_tier (HIGH_R/HIGH_S/DECISIVE/BORDERLINE/AMBIGUOUS/CONFLICT/NO_MIC classification under CLSI 2024 + EUCAST 14.0 breakpoints). Raw AST CSV load skipped.
+- scripts/compute_lineage_metrics.py — per-cohort lineage recompute pure helpers: parse_cohort_dir / find_artifact (latest) / read_selected / _fasta_ok (M1 genome gate) / ensure_cohort_genomes / reconcile_raw_metrics (M4 raises on mismatch; skips no-run-dir accs) / graded_lineage_bucket / build_threshold_results / build_cell (scored vs partial/unreconciled no-tier) / load+upsert+write sidecar (canonical-key idempotent; malformed->fresh) / process_cohort no-artifact FileNotFoundError. Docker Mash in process_cohort/main skipped.
+- scripts/leaderboard.py — formatting, markdown grouping, bundle-reading helpers (subprocess paths mocked)
+- scripts/ncbi_pd_provenance_census.py — census_result_to_sidecar_row group->organism normalizer (refuses error/capped/missing-count rows) + upsert_census_result idempotent powering sidecar (one row per organism×drug; skips degraded; preserves header). Network metadata streaming (census_one/main) skipped.
+- scripts/provenance_disjoint_validate.py — main() fail-closed manifest gate (incomplete manifest + no --allow-incomplete-manifest -> exit 2; override proceeds DEGRADED) + prior_accessions exclusion set wiring into select_disjoint (exact-self cohort not excluded). Network selection + Docker AMRFinder skipped.
+- scripts/pipeline.py — CLI dispatcher with ingest/train/predict/attribute subcommands; argparse validation + exit codes
+- scripts/populate_cache.py — Phase 2 standalone embedding-cache driver; strain_id ↔ assembly_accession mapping; --allow-mock gate; end-to-end mock populate
+- scripts/quantize_fidelity_check.py — top-K intersection + Spearman comparison, GO/NO-GO aggregation, markdown report writer
+- scripts/smoke_gate_12strain_cipro.py — 2026-05-17 patches: --drug-templated output strings (write_packet heading + cohort description + default filename slug) + NT-XGBoost runner fallback to ast_labels iteration when cohort.per_drug_strain_ids[drug] missing. Full LOSO + NT cache run is orchestration (skipped).
+- scripts/smoke_pipeline.py — end-to-end synthetic-fixture smoke run; AUROC + top-1 attribution assertions
+- scripts/tmp_smx_external_validate.py — NON-FROZEN TMP-SMX external scorer pure-logic: score_cohort (4-stratum sul/dfr routing, _binary R/S/None gap, strata_reproduced gate incl. sul-only-empty `so is None` branch) + build_artifact (external-validation-v1 schema + EXPERIMENTAL_SCORED/scorer_local branding + SCORED/INDETERMINATE headline) + frozen-leak guard. _mic/_binary/load_sci234/load_oxford/main I/O skipped.
+- scripts/build_external_validation_report.py — external-validation roll-up pure-logic: load_external_artifacts (run-scoped; refuses glob-all; skips hard_fail/degraded/corrupt) + _publishable + cluster_weighted_with_ci (clonality math + Wilson CI + eff-N) + build_cell (experimental branding passthrough) + render_json/render_md (strict→binary fallback incl. both-empty default-to-strict) + _fmt_ci. Docker Mash (compute_lineage_for_cohort/main) skipped.
+- dna_decode/genome_map/build_map.py — genome-map assembler; build_genome_map per-feature tier precedence (AMR determinant-phenotype > virulence-determinant > function) + phenotype wall + virulence wall (_virulence_field) + M2-isolated virulence metric keys; build_feature_table flattener. PURE (no Docker/IO).
+- dna_decode/genome_map/virulence_overlay.py — VF overlay pure-logic: parse_virulence_hits (per_hit→DeterminantHit, uncalled dropped, tandem kept), join_virulence (VF-namespaced counts + _count_ambiguous_contig_hits M1 + all_virulence_symbol_fallback), genome_pathotype_call (full resolve_call contract; insufficient_context only on unavailable VF/no contigs), cluster_pathotype_context, _pergene_cov_from_per_gene, virulence_organism_in_scope (E.coli/Shigella gate)
+- dna_decode/pathotype/vf_runner.py — canonical-VF blastn caller; parse_blastn_outfmt6 PURE parse (minus-strand normalization, qlen=0 + malformed-line guards, per_gene/per_cluster best-hit + per_hit every-called-HSP) + _interval_dedup (overlap-merge keeps higher-coverage HSP attrs; disjoint copies retained) + _db_sha256 + build_vf_diff side-by-side concordance. Live blastn subprocess (run_canonical_vf) skipped.
+- dna_decode/eval/clonality.py — clonality clustering + cluster-weighted metrics; greedy_representative_clusters_from_matrix (C2 chaining-resistant, NOT single-linkage) / cluster_class (C1 DISCORDANT never majority-voted) / cluster_weighted_confusion (one vote per lineage; missing-pred->ABSTAIN) / wilson_ci (well-behaved at small/extreme N) / effective_lineage_n (normalizes int/str cls) / cluster_members. Docker Mash in greedy_representative_clusters skipped.
+- dna_decode/data/cell_key.py — canonical (organism, drug) join key (strip+lowercase); canonical_cell_key tuple + cell_key_str 'org|drug' — single source of truth shared by report card + lineage sidecar + scored JSONs
+- dna_decode/eval/cohort_manifest.py — accession-manifest leakage registry; build_manifest() scans data/raw/*/selected.tsv + data/processed/*.parquet (incomplete=True fail-closed on any load failure); prior_accessions() exact-self exclusion (not substring); role/organism/drug parsing
+- dna_decode/data/shipped_decoder_surface.py — SHIPPED_DECODER_SURFACE deployed-claim registry; shipped_decoder_rows/surface_index/all_surface_drugs; phenotype_source_status enum (ncbi_pd/label_confounded/no_free_source); CLI-drug coverage anti-drift guard
+- dna_decode/data/bvbrc_genome.py — Phase 2 BV-BRC Genomes-tab CSV adapter; header normalization (Title Case → lowercase_underscore); organism filter (Species → Genome Name fallback); duplicate-key warning; tolerant int parsing
+- dna_decode/viz/browser.py — matplotlib attribution plot rendering + TSV export

dna_decode-0.5.1/.gitignore

@@ -0,0 +1,76 @@
+# Runtime data + caches (large, gitignored per Step 1 of technical plan)
+# Anchored to repo root — doesn't match dna_decode/data/ subpackage
+# Ignore data/ contents per-entry (NOT the dir itself) so specific subdirs can be re-included
+/data/*
+# Exception: committed public reference alleles for the fungal-AMR real-data test (tiny GenBank CDS)
+!/data/fungal_ref/
+# Exception: committed Arabidopsis Path-B phenotype labels (small public AraPheno CSVs)
+!/data/arabidopsis/
+# Exception: committed public influenza N1 NA reference + real field alleles for the antiviral (4th-kingdom)
+# real-data test (tiny GenBank CDS). Same class as fungal_ref.
+!/data/antiviral_ref/
+# Exception: committed public HIV-1 HXB2 RT CDS reference for the genome-mode caller self-check + planted-
+# mutation test (tiny NCBI CDS, K03455.1:2550-4229). Same class as fungal_ref/antiviral_ref.
+!/data/hiv_ref/
+# Exception: committed public SARS-CoV-2 Mpro (nsp5) reference for the genome-mode caller integrity gate
+# (tiny NCBI CDS, NC_045512.2:10055-10972). Same class as fungal_ref/antiviral_ref/hiv_ref.
+!/data/sarscov2_ref/
+*.h5
+*.parquet
+*.pkl
+*.pt
+*.bin
+
+# Python
+__pycache__/
+*.pyc
+*.pyo
+*.egg-info/
+.pytest_cache/
+.mypy_cache/
+.ruff_cache/
+.coverage
+.coverage.*
+htmlcov/
+
+# Environments
+.venv/
+venv/
+env/
+.env
+.env.local
+
+# IDE
+.vscode/
+.idea/
+*.swp
+.DS_Store
+
+# Build outputs
+build/
+dist/
+*.egg
+
+# Model checkpoints (large, fetched from HuggingFace)
+*.safetensors
+*.ckpt
+checkpoints/
+
+# Logs
+*.log
+logs/
+
+# Claude execute-plan scaffolding — recreated by /execute-plan runs
+.claude/execute-plan-state/
+
+# Generated audit / analysis reports (re-runnable via scripts/audit_cohort.py etc.)
+/reports/
+
+# AMRFinder runs (generated artifacts; tens of MB per strain)
+/data/amrfinder_runs/
+
+# Project-state scratch/working copies (stale duplicates of the live ledger; not history)
+/project_state/*-scratch.md
+
+# genome-map spike per-genome feature maps (large, regenerable via scripts/genome_map_spike.py)
+wiki/genome_map_spike_*/

dna_decode-0.5.1/.python-version

@@ -0,0 +1 @@
+3.11

dna_decode-0.5.1/2026-05-30-1200-ep4-pathotype/approvals.md

@@ -0,0 +1 @@
+<!-- approvals.md for 2026-05-30-1200-ep4-pathotype -->

dna_decode-0.5.1/2026-05-30-1200-ep4-pathotype/audit-trail.md

@@ -0,0 +1 @@
+<!-- audit-trail.md for 2026-05-30-1200-ep4-pathotype -->

dna_decode-0.5.1/2026-05-30-1200-ep4-pathotype/intent-contract.md

@@ -0,0 +1 @@
+<!-- intent-contract.md for 2026-05-30-1200-ep4-pathotype -->

dna_decode-0.5.1/2026-05-30-1200-ep4-pathotype/recommendation.md

@@ -0,0 +1 @@
+<!-- recommendation.md for 2026-05-30-1200-ep4-pathotype -->

dna_decode-0.5.1/2026-05-30-1200-ep4-pathotype/result.md

@@ -0,0 +1 @@
+<!-- result.md for 2026-05-30-1200-ep4-pathotype -->

dna_decode-0.5.1/CHANGELOG.md

@@ -0,0 +1,455 @@
+# Changelog
+
+All notable changes to `dna_decode`. Format loosely follows [Keep a Changelog](https://keepachangelog.com/);
+this is a solo research-tool repo so the granularity is per-release-theme, not per-PR.
+
+## [Unreleased] — Anchor-4: standing decoder-suite validation report card (2026-06-10)
+
+- **NEW `dna_decode/eval/cohort_manifest.py`** — data-driven accession-manifest registry. `build_manifest()`
+  scans EVERY `data/raw/*/selected.tsv` + EVERY `data/processed/*.parquet`; `prior_accessions(exclude_cohort
+  =name)` excludes prior accessions by **EXACT-self cohort identity (not substring)**, and `Manifest.incomplete
+  =True` on any load failure. REPLACES the hardcoded `_FLAGSHIP_PARQUET_COHORTS` list (3 cohorts) in
+  `scripts/provenance_disjoint_validate.py` — leakage exclusion now covers all 8 parquet cohorts (744
+  accessions vs ~175 before).
+- **`scripts/provenance_disjoint_validate.py` now FAILS CLOSED on an incomplete manifest** (`INCOMPLETE_MANIFEST`,
+  exit 2) unless `--allow-incomplete-manifest` is passed (which stamps degraded independence into the artifact).
+  Artifact carries `manifest_complete` / `manifest_degraded`.
+- **`scripts/ncbi_pd_provenance_census.py` now SELF-PERSISTS its powering verdict** to
+  `wiki/provdisjoint_census_results.json` (group→organism normalizer; idempotent upsert per `(organism, drug)`;
+  REFUSES to persist error/row-capped rows so a degraded run can't overwrite a good powering verdict).
+  Previously stdout-only.
+- **NEW `dna_decode/data/shipped_decoder_surface.py`** — the authoritative DEPLOYED-CLAIM surface registry
+  (organism, drug, engine, organism_scope, phenotype_source_status, census_group).
+- **`scripts/build_validation_report_card.py`** — rows now = shipped surface ∪ observed cells (a new decoder
+  can't ship invisibly); added a 7th cell-state `LABEL_CONFOUNDED` (oxacillin×S.aureus, unreliable mecA
+  surrogate); `NO_FREE_PHENOTYPE_SOURCE` is now surface-driven. Honest per-cell tier + no aggregate headline
+  preserved. Writes `wiki/decoder_validation_report_card.{md,json}`. Current card: 25 cells (6 SCORED /
+  4 NOT_CENSUSED / 1 UNDERPOWERED / 2 ABSTAINS_BY_DESIGN / 1 LABEL_CONFOUNDED / 11 NO_FREE_PHENOTYPE_SOURCE).
+
+## [Unreleased] — intron-aware multi-HSP codon mapping (engine)
+
+- **`observed_substitutions` (the shared target-site codon-mapper) is now INTRON-AWARE.** It stitches the
+  query(CDS-reference)-position → subject-nucleotide map across ALL HSPs on the gene's best contig (not just
+  the single best HSP). Because codon numbering is by *query* position (contiguous CDS), a codon that spans
+  an exon-exon boundary — its 3 nts in two different HSPs (exons separated by an intron in the genome) — is
+  still translated correctly. This generically improves **every** genome-mode caller (fungal ERG11, K13)
+  for multi-exon / split-across-contigs genes, and **unblocks intron-containing targets** (pfcrt has 13
+  exons; GenBank deposits are ~2471 bp genomic). Intronless genes are the single-HSP special case →
+  identical prior behavior (guarded by the existing fungal/K13 tests; 967 passed, 0 regressions). Validated
+  on the real (non-repetitive) 3D7 K13 CDS artificially split into exons: exon1 + deep-exon2 mutations both
+  detected across the intron; a mid-codon boundary assembles to WT with no spurious call. 2 tests
+  (`tests/test_intron_aware_mapping.py`). Known limit: a periodic CDS self-aligns at its period (use a real
+  non-repetitive reference — real CDSs are).
+- **pfcrt genome mode FLIPPED ON** (same run) — committed the 3D7 pfcrt CDS reference
+  (`data/antimalarial_ref/Pf3D7_pfcrt_cds.fna`, NCBI RefSeq XM_001348968, 424aa, WT Lys@76) so
+  `dna-amr --drug chloroquine --genome-fasta X.fna` works (`--pfcrt-ref` to override). **Validated on REAL
+  13-exon genomic pfcrt alleles** (the 2471 bp GenBank field isolates): the intron-aware mapper recovered
+  **K76T across the introns** + the full canonical CQ-R haplotype (A220S/Q271E/I356T/R371I = CVIET-type) on
+  all 6 tested; WT 3D7 → S. Genome-mode ref is picked by target gene (K13 → `--k13-ref`, pfcrt →
+  `--pfcrt-ref`); the intron guard is replaced by a ref-existence check. Real genomic fixture committed
+  (`Pf_pfcrt_MN419894_K76T.fna`). This is the first REAL multi-exon end-to-end validation of the engine.
+
+## [Unreleased] — antimalarial vertical: P. falciparum K13 (the 3rd kingdom, protozoan)
+
+- **`dna_decode/data/antimalarial_amr.py`** + **`scripts/pf_kelch13_caller.py`** — extends the proven
+  deterministic target-site method (bacterial AMRFinder → fungal ERG11) to *Plasmodium falciparum*
+  artemisinin partial resistance via the WHO-validated **Pfkelch13 (K13)** propeller markers (C580Y, R561H,
+  R539T, I543T, A675V, R622I, …). Hand-curated catalog (no AMRFinder-equivalent for Plasmodium); the caller
+  REUSES the fungal caller's gene-generic `observed_substitutions` (BLAST K13-CDS-vs-genome → gap-aware
+  codon-map → catalog), so the only new surface is the catalog + a thin wrapper. Artemisinin partial
+  resistance is a clearance phenotype, not an MIC — the validated K13 marker IS the genotypic call; an S
+  call surfaces non-K13 / partner-drug blind spots. Real 3D7 K13 reference committed
+  (`data/antimalarial_ref/Pf3D7_K13_cds.fna`, NCBI XM_001350122.1, 726aa, WT Cys@580); G0-completion test
+  validates C580Y numbering on the **real** reference. Offline-safe (absent BLAST+ → INDETERMINATE).
+  6 tests. **3rd kingdom decoded** (bacteria → fungi → protozoa).
+- **Wired into the unified `dna-amr` CLI** (`--drug artemisinin|artesunate|dihydroartemisinin`) — routes to
+  the K13 engine (`--genome-fasta` real-BLAST, or `--observed K13:C580Y` wheel-only), mirroring the fungal
+  productization; emits the same `amr-mechanism-call-v1` record; `--organism` default relabels to
+  `Plasmodium_falciparum`. Shared `_target_site_record` + `_emit_target_site` now back both the fungal and
+  antimalarial branches (extracted, not duplicated). 4 CLI tests. So `dna-amr` now spans **3 kingdoms**.
+- **+ chloroquine (pfcrt K76T)** — extends the antimalarial vertical to the iconic chloroquine-resistance
+  marker (`dna-amr --drug chloroquine --observed pfcrt:K76T` → R). `gene_for_drug` routes drug→target
+  gene; K76T → CQ-R is unambiguous (pfmdr1 partner-drug markers deliberately omitted — their direction
+  flips between amodiaquine and lumefantrine). Shipped first as `--observed`-only; **genome mode was then
+  flipped on** once the intron-aware mapper + committed pfcrt CDS reference landed (see the intron-aware
+  entry above) — `dna-amr --drug chloroquine --genome-fasta X.fna` now works on real 13-exon genomic pfcrt.
+  5 tests.
+
+## [Unreleased] — self-calibrating AMR rule (`calibrate_organism`)
+
+- **`dna_decode/eval/calibrate_organism.py`** — auto-selects the per-organism AMR rule config from a
+  ≥~15R/15S labeled cohort: chooses the determinant COUNTER (`qrdr_point` vs broad drug-class) and the
+  count THRESHOLD by leave-one-out balanced accuracy, and auto-excludes INTRINSIC gene families (≥90% of
+  both R and S, grouped at gene-family granularity so polymorphic intrinsics like blaOXA-51-family are
+  caught). Returns a `CalibratedRule` (`.predict()` applies it); ABSTAINS with verdict `EXPRESSION_FLOOR`
+  when no presence config clears the 0.70 LOO floor (expression-driven R that gene-presence cannot decode).
+  Motivated by the wider-AMR boundary taxonomy (CONTENT/TUNING/EXPRESSION) — the counter, not just the
+  threshold, is organism-specific (the Klebsiella-vs-Salmonella cipro contrast). Validated on cached cohorts
+  (`wiki/calibrate_organism_validation_2026-06-08.md`): Campylobacter→1, Klebsiella→2 (+oqxAB excluded),
+  Salmonella→broad@1 (deployed 0.567→1.0) all LOO 1.0; Acinetobacter + Pseudomonas meropenem → abstain.
+  16 unit tests.
+- **Wired into `call_resistance(..., organism=...)`** (opt-in) via a committed registry
+  `dna_decode/data/calibrated_amr_rules.json` (built by `scripts/build_calibrated_registry.py`). When an
+  organism is passed AND has a registry entry: a CALIBRATED entry applies its counter/threshold/intrinsic
+  exclusions; an EXPRESSION_FLOOR entry returns **`prediction: "ABSTAIN"`** (refuses to predict an
+  expression-driven organism×drug rather than over-call). `organism=None` (or unknown organism, or an
+  explicit `resistance_threshold`) keeps the unchanged `DRUG_RULE` default — backward-compatible. Registry
+  is IN-SAMPLE (N≈30) and opt-in by design; abstain verdicts are conservative. 9 wiring tests.
+- **Design-review hardening (2026-06-09):** (a) `INSUFFICIENT_EVIDENCE` verdict + `MIN_CLASS_COUNT` guard —
+  a one-class/under-powered cohort no longer yields a bogus `EXPRESSION_FLOOR` (fixed a degenerate
+  Pseudomonas registry entry); (b) `loo_balanced_accuracy` now truly balanced (was plain accuracy), `None`
+  when a class is absent; (c) deployed config is the deterministic full-cohort `_select_best_config` pick
+  (removed the modal/tie-break ambiguity), LOO separately estimates the selection procedure; (d)
+  `build_calibrated_registry` resolves runs via the validator `reuse_glob` (Pseudomonas under-load fix —
+  all 5 entries now on valid 15R/15S); (e) promotion gate adds a **specificity floor** + min-10/class and
+  treats config-match as a **flag, not a gate** (non-inferior OOS perf suffices) — all 3 cipro configs
+  `promotion_eligible=True`. M2 (AMRFinder `Method`-column propagation) documented + deferred. 947 tests.
+
+## [Unreleased] — cross-decoder analyses (concordance + profile + co-localization)
+
+Three ANALYSES that compose the shipped decoders (no new DB) — variety roadmap Waves 1-2.
+
+- **`dna-coloc`** (`dna-decode coloc`, Wave 2) — links each acquired resistance gene to plasmid replicon(s)
+  on the SAME assembly contig → "is *this* AMR gene plasmid-borne?". Enabled by a new opt-in engine
+  positions-mode (`call_alleles(..., with_positions=True)` returns each hit's subject contig+coords;
+  default-off, every existing caller unchanged). Same-contig is suggestive, not proof (caveat shipped).
+  Pure core + real-BLAST e2e (blaNDM-1 on plasmid-contig → plasmid-borne; sul1 on chrom-contig → not).
+
+- **`dna-concordance`** (`dna-decode concordance`) — compares the two independent acquired-gene callers,
+  AMRFinder (`dna-amr` main.tsv) vs ResFinder (`dna-resfinder` blastn), at the gene-family level (allele
+  variant stripped; `sul1`≠`sul2`, `blaNDM-1`≈`blaNDM-19`) + Jaccard agreement. The cross-check `resfinder`
+  was built to enable.
+- **`dna-profile`** (`dna-decode profile`) — runs every assembly-FASTA decoder (pathotype + serotype +
+  plasmid + resfinder) on one genome → a single unified report; each section degrades independently.
+- Kept out of the `TRAITS` decoder registry (new `ANALYSES` dict; disjoint namespaces). 15 new tests.
+
+## [Unreleased] — MLST sequence-typing decoder (the blocked one, unblocked via PubMLST)
+
+- **`dna-mlst`** (`dna-decode mlst`) — 8th decoder. 7-gene MLST: exact per-locus allele (blastn 100/100 on
+  the shared engine) → profile → Sequence Type via the PubMLST profiles table. v0: E. coli Achtman
+  (adk/fumC/gyrB/icd/mdh/purA/recA, 16,242 STs). The earlier blocker was DB-discovery (CGE `mlst_db` raw
+  paths 404); resolved by using **PubMLST's REST API** (`pubmlst_ecoli_achtman_seqdef` scheme 4).
+  `dna-mlst --fetch-db` installs the scheme (DB on demand, gitignored). Novel/incomplete profiles report
+  "ST not called", never guessed. **Validated end-to-end: K-12 MG1655 → ST10** (real DB + real genome) +
+  synthetic real-BLAST e2e + pure-core tests. New shared `mlst/core.py` (profiles parse + ST lookup). 5 tests.
+
+## [Unreleased] — DisinFinder biocide-resistance decoder (roadmap W4) + profile completion
+
+- **`dna-disinfinder`** (`dna-decode disinfinder`) — 7th decoder. Acquired biocide/disinfectant resistance
+  genes (quaternary-ammonium qac* + formaldehyde formA) via the DisinFinder DB on the shared engine (reuses
+  resfinder's CGE gene parser). Hospital infection-control relevant; qac genes often share plasmids with AMR
+  → pair with `dna-coloc`. Offline-safe; DB on demand. 3 tests.
+- **`dna-profile` now also runs `pointfinder`** — the run-all covers all 5 assembly-FASTA decoders
+  (pathotype/serotype/plasmid/resfinder/pointfinder).
+- MLST (roadmap remainder) stays deferred: DB raw-paths 404 + it needs exact-allele/profile→ST semantics
+  (a distinct batch, not the presence/codon pattern).
+
+## [Unreleased] — PointFinder chromosomal point-mutation decoder (roadmap W3)
+
+- **`dna-pointfinder`** (`dna-decode pointfinder`) — 6th decoder. Chromosomal AMR point mutations via the
+  PointFinder DB: blastn each reference gene CDS vs the assembly, map the subject amino acid at each
+  catalogued codon (new shared `typing/codon_map.py` — gap-aware codon→subject-AA, the proven fungal-ERG11
+  pattern, now in-package), call a mutation when the subject AA matches a `Res_codon` in
+  `resistens-overview.txt`. v0 scope: E. coli FQ QRDR (gyrA/parC/gyrB/parE). An INDEPENDENT point-mutation
+  caller (`caller_is_independent_baseline: true`) complementing `amr` (AMRFinder POINT) + `resfinder`
+  (acquired only). Epistasis (`Required_mut`) recorded, not enforced. Offline-safe; DB on demand.
+  Validated on synthetic (S83L) + the real committed E. coli DB (gyrA codon83=S/87=D). 5 tests.
+
+## [Unreleased] — typing-decoder family (plasmid + serotype + resfinder on one shared engine)
+
+Three new deterministic curated-DB decoders — the tool grows from 2 traits to 5, all on one engine.
+
+- **Shared engine** `dna_decode/typing/blast_caller.py` (`call_alleles`) — the generic best-HSP-per-allele
+  blastn core (reuses pathotype/vf_runner's resolvers). The CGE curated-DB pattern (pathotype + plasmid)
+  is now config-per-decoder, not a from-scratch build. Plasmid refactored onto it (DRY).
+- **`dna-serotype`** (`dna-decode serotype`) — E. coli O:H serotyping via SerotypeFinder allele DB
+  (best O-antigen + best H-antigen → `O25:H4`, partial → `O104:H?`). A genuinely new trait.
+- **`dna-resfinder`** (`dna-decode resfinder`) — acquired-AMR-gene detection via ResFinder allele DB,
+  per-class. Deliberately an **independent** caller vs the AMRFinder-based `dna-amr`
+  (`caller_is_independent_baseline: true`) — the cross-tool concordance check the AMR decoder lacked.
+- All offline-safe (status `unavailable`, exit 3); DBs downloaded on demand (not committed). ~16 new tests.
+
+### plasmid replicon-typing decoder (earlier in this Unreleased cycle)
+
+New deterministic trait decoder `dna-plasmid` (`dna-decode plasmid`) — the tool grows beyond AMR.
+
+- **New capability:** plasmid incompatibility (Inc) replicon typing from a genome assembly via the curated
+  PlasmidFinder allele DB + real blastn (identity 95 / coverage 60, PlasmidFinder defaults). Reports the Inc
+  replicons present (IncF/IncH/IncI/IncX/IncN/…) — composing with `dna-amr`: AMR says *what* resistance,
+  plasmid typing says whether it likely rides a known mobile element.
+- Sibling architecture to `dna-pathotype` (curated-DB blastn caller; reuses `pathotype/vf_runner`'s blastn
+  resolvers — DRY). Offline-safe: no blastn / no DB → `status: "unavailable"` (exit 3), never a crash.
+  `caller_is_independent_baseline: false` (faithful to PlasmidFinder's own method, not an independent check).
+- DB downloaded on demand (not committed), like the VirulenceFinder DB. 7 tests
+  (`tests/test_plasmid_decoder.py`); cli-dispatch registry contract updated to the 3-decoder set.
+
+## [0.5.0] — 2026-06-08 — Fungal AMR decoder (the kingdom jump)
+
+`dna-amr` now decodes **fungal** azole/echinocandin resistance, not just bacterial — the determinant-scan
+method validated across the bacteria→fungi kingdom boundary.
+
+- **New capability:** `dna-amr --drug fluconazole|voriconazole|caspofungin|micafungin` routes to a
+  BLAST-ERG11/FKS1 target-site engine (vs the AMRFinder engine for bacterial drugs — there is no
+  AMRFinder-for-fungi). Two source modes: `--genome-fasta` (BLAST the committed C. auris ERG11 reference)
+  and `--observed GENE:SUB[,...]` (pure, wheel-only, no BLAST). Emits the same `amr-mechanism-call-v1`
+  record as the bacterial path (uniform tool surface); S calls surface the efflux/aneuploidy blind spots.
+- **Validation (Gate G1, `wiki/fungal_ep7_g1_closeout_2026-06-08.md`):** on a de-confounded C. auris
+  WGS+MIC cohort (S. Africa bloodstream, PRJNA737309 + AraPheno-style Table S1 MICs), the deterministic
+  caller found the catalogued ERG11 mutation in **100% of fluconazole-MIC-R isolates across two clades**
+  (clade I Y132F, clade III F126L/VF125AL) — sensitivity 1.0. Specificity is label-limited (reduced-
+  susceptibility F126L carriers fall below the CDC tentative breakpoint), the documented "suspect the
+  label" pattern; the genotype is the trustworthy output.
+- **Infra shipped:** `dna_decode/data/fungal_amr.py` (hand-curated catalog + CDC tentative breakpoints),
+  `scripts/fungal_erg11_caller.py` (BLAST→codon-map→catalog), `scripts/build_fungal_cohort.py` (cohort
+  validation + within-clade de-confound + LABEL_LIMITED_FAILURE verdict), `scripts/assemble_sra_cohort.py`
+  (targeted ERG11 read-mapping, ~4 min/isolate vs ~45 min full assembly). Committed real C. auris reference
+  + 3 public allele fixtures (`data/fungal_ref/`).
+- **Eukaryotic Path B (Arabidopsis flowering-time embedding test, Gate G2)** pre-staged + brainstorm-revised
+  + CPU-only dry-manifest gate coded (`scripts/g2_dry_manifest.py`); GPU run deferred to the workhorse.
+- ~43 new tests (fungal catalog/caller/cohort/CLI + dry-manifest). Bacterial path unchanged.
+
+## [0.4.0] — 2026-06-07 — Multi-Organism AMR Decoder (capstone)
+
+Milestone release consolidating the AMR arc. No new code — a capstone over v0.3.x.
+
+`dna-amr` is a deterministic, interpretable AMR R/S decoder validated across **6 drugs × 4 organisms ×
+4 mechanism classes, spanning the gram divide** (E. coli, K. pneumoniae, P. aeruginosa, S. aureus),
+deployed as `dna-amr` / `dna-decode`. Every per-drug rule beats naive AMRFinder on independent data.
+
+- **One engineering principle** held across every organism/mechanism: count the drug's SPECIFIC
+  resistance determinants (target point-mutations / drug-specific Subclass / acquired gene-family), not
+  the broad drug-class bag — intrinsic chromosomal genes (efflux) are the cross-organism gotcha.
+- **Honest limits, named in every output:** `undetectable_mechanisms` (efflux/porin/regulatory expression
+  phenotypes) + label-quality caveats (oxacillin → use cefoxitin). The recurring binding constraint is the
+  de-confounded, reliably-labeled substrate — not the method.
+- Capstone: `wiki/amr_multiorganism_capstone_2026-06-07.md`. 108 tests green.
+
+**This is the milestone.** Further organism/drug breadth is diminishing-returns (re-confirms the same two
+findings). The genuinely-different next leaps (cross-lab validation, a non-AMR sampling-independent
+substrate, multimodal/eukaryotic) require resources beyond autonomous code work.
+
+## [0.3.7] — 2026-06-07
+
+1st Gram-positive: S. aureus oxacillin (MRSA/mecA) — genotype transfers; honest label finding.
+
+### Added
+- **oxacillin** (6th drug, 1st Gram-positive): mecA-based MRSA rule (threshold 1 + METHICILLIN-subclass,
+  excludes blaZ penicillinase). Added to mic_tiers (breakpoints, classes, mec loci, primary mechanism) +
+  DRUG_RULE. `supported_drugs()` now 6. +2 tests (104 → 106 → 108 green).
+- **S. aureus oxacillin validation (4th organism, 1st Gram-positive):** N=30.
+  `wiki/staphylococcus_aureus_oxacillin_validate_2026-06-07.md`.
+
+### Finding (the honest result)
+- **mecA genotype detection TRANSFERS to Gram-positive: sens 1.000** (all 15 R strains carry mecA). The
+  acquired-gene + Subclass-refinement approach works on a Gram-positive, as on the gram-negatives.
+- **spec 0.333 is oxacillin-LABEL noise, NOT a rule defect:** 10/15 "oxacillin-susceptible"-labeled strains
+  carry full-length mecA — far above genuine OS-MRSA (<5%). Oxacillin direct AST is the documented unreliable
+  comparator for mecA; CLSI/EUCAST recommend **cefoxitin** as the surrogate. The proper-label re-test is
+  **substrate-blocked** (cefoxitin = only 3R on this NCBI dataset).
+- **Terminal:** Gram-positive mecA detection generalizes; phenotype-label validation is the limit — the
+  project's recurring "substrate/label is the binding constraint" lesson, now confirmed on a Gram-positive.
+
+## [0.3.6] — 2026-06-07
+
+3rd organism (Pseudomonas) + cross-organism shipped in the CLI.
+
+### Added
+- **`dna-amr --organism <O>`** (genome mode) — passes through to AMRFinder `-O` (organism-specific QRDR
+  point-mutation detection); recorded in `provenance.amrfinder_organism`. Default Escherichia. Closes the
+  gap where cross-organism support lived only in validation scripts, not the shipped CLI. +2 CLI tests.
+- **Pseudomonas aeruginosa cipro VALIDATED** (3rd organism): N=30 acc 0.867 / sens 0.80 / spec 0.933
+  (beats naive AMRFinder 0.767). The QRDR-POINT rule transfers UNCHANGED to a *less-similar* gram-negative
+  (MexAB-OprM efflux, intrinsic AmpC — no oqxAB). 3 FN = efflux-mediated cipro-R (expected blind spot).
+  `wiki/pseudomonas_aeruginosa_ciprofloxacin_validate_2026-06-07.md`.
+- **`scripts/organism_drug_validate.py`** — generalized any-NCBI-organism × any-drug validator
+  (auto-discovers latest PDG metadata, reuses cached runs). Every future organism is now a one-command run.
+
+### Result
+dna-amr validated across **3 organisms** (E. coli, K. pneumoniae, P. aeruginosa). The "count the drug's
+specific determinants, not the broad class bag" principle holds across all three — strong evidence it is
+organism-general, not E. coli-specific.
+
+## [0.3.5] — 2026-06-07
+
+Klebsiella full drug matrix complete — dna-amr validated 5 drugs × 2 organisms.
+
+### Added
+- **Klebsiella cef + gent + tet validated** (rules applied unchanged from E. coli):
+  - ceftriaxone: acc 0.800 / sens 1.0 / spec 0.6 ✅
+  - gentamicin: acc 0.867 / sens 0.867 / spec 0.867 ✅
+  - tetracycline: acc 0.800 / spec 1.0 / **sens 0.600** ⚠️ PARTIAL (efflux blind spot — see below)
+  - `scripts/klebsiella_drug_validate.py` (drug-agnostic; reuses cached Klebsiella runs across cohorts).
+  - Consolidated: `wiki/klebsiella_drug_matrix_2026-06-07.md`.
+- **`gene_prefixes` rule refinement** (`amr_rules.py`): tetracycline now counts only acquired `tet*` genes,
+  excluding intrinsic K. pneumoniae OqxAB efflux (AMRFinder-tagged TETRACYCLINE but present in susceptible
+  isolates). Same cross-organism principle as cipro QRDR-POINT. **Also improved E. coli tet 0.833 → 0.917.**
+
+### Findings
+- **tetracycline / Klebsiella is the honest PARTIAL:** the acquired-`tet*` rule is precise (spec 1.0) but
+  sens 0.600 — 6/15 R strains are efflux-mediated (oqxAB overexpression), undetectable by ANY
+  curated-determinant rule (an expression phenotype, surfaced in `undetectable_mechanisms`). A documented
+  biological limit, not a rule defect.
+- **Cross-organism principle confirmed 3× (cipro/tet/the gent+cef+mero Subclass refinements):** count the
+  drug's specific resistance determinants, not the broad drug-class bag; intrinsic chromosomal determinants
+  (efflux) are the organism-specific gotcha.
+
+### Result
+dna-amr spans **5 drugs × 2 organisms × 4 mechanism classes**; 4/5 Klebsiella drugs clear the bar
+zero-tuning, all beat naive AMRFinder. +3 tet tests (104 green).
+
+## [0.3.4] — 2026-06-07
+
+Klebsiella meropenem — 2nd organism, NEW mechanism class (carbapenem). Phase 3 slice 2.
+
+### Added
+- **meropenem decoder** (5th drug): acquired-carbapenemase rule (threshold 1 + CARBAPENEM-subclass
+  refinement — blaKPC/NDM/OXA-48). **Klebsiella N=30 acc 0.867 / sens 1.0 / spec 0.733** (vs naive
+  AMRFinder 0.533; the CARBAPENEM-subclass refinement lifts spec 0.067→0.733 by excluding ESBL/AmpC that
+  raise meropenem MIC without hydrolyzing it). `wiki/klebsiella_meropenem_validate_2026-06-07.md`.
+  Carbapenem is the defining K. pneumoniae clinical threat — a mechanism class E. coli AMR never covered.
+- meropenem added to `mic_tiers.py` (breakpoints CLSI R≥4/S≤1 + EUCAST; AMRFinder classes;
+  carbapenemase loci catalog; primary mechanism). `supported_drugs()` now returns 5.
+- `scripts/klebsiella_meropenem_validate.py` (reuses cached cipro-cohort AMRFinder runs on strain overlap).
+- +2 tests (carbapenemase counted / ESBL excluded). 102 green.
+
+### Honest scope
+The meropenem rule is blind to porin-loss-mediated carbapenem resistance (no carbapenemase gene) — the
+expected FN mode; surfaced in `undetectable_mechanisms`. 4 FP (susceptible strains carrying a carbapenemase
+gene — likely low-expression / borderline MIC).
+
+## [0.3.3] — 2026-06-07
+
+Cross-organism: Klebsiella + the QRDR-POINT cipro rule (roadmap Phase 3, slice 1).
+
+### Added
+- **Cross-organism transfer (Klebsiella pneumoniae cipro):** N=30 NCBI cohort, **acc 1.000** with the
+  deployed rule. The method generalizes across organisms. `scripts/klebsiella_cipro_transfer.py`,
+  `wiki/klebsiella_cipro_transfer_2026-06-07.md`. `_run_amrfinder` gained an `organism` param
+  (`-O Klebsiella_pneumoniae`).
+- `qrdr_point_count` / `qrdr_point_determinants` in `amr_rules.py` — count fluoroquinolone QRDR
+  target-alteration POINT mutations (gyrA/parC/parE) only.
+
+### Changed (ratified — changes the deployed cipro number)
+- **cipro `DRUG_RULE` switched to the QRDR-POINT rule globally** (`counter='qrdr_point'`): count QRDR
+  target POINT mutations ≥2, not the broad QUINOLONE-class determinant bag. Rationale: the broad rule
+  FAILED on Klebsiella (acc 0.5 — intrinsic chromosomal OqxAB efflux, absent in E. coli, saturates the
+  count). The canonical target-mutation count is cross-organism-robust. Net effect:
+  - Klebsiella cipro 0.5 → **1.000**
+  - E. coli cipro in-cohort 0.939 → 0.925 (−1.4pp; dropped cases were qnr/efflux-mediated)
+  - E. coli cipro **cross-source (NCBI) 0.955 → 1.000** (+4.5pp — QRDR-POINT generalizes better; the
+    in-cohort −1.4pp was tuning-cohort overfit).
+- **Platform finding:** cross-organism transfer requires counting the drug's TARGET-alteration mutations,
+  not the broad drug-class bag; intrinsic chromosomal determinants are the organism-specific gotcha.
+- Tests: cipro tests updated to QRDR-POINT (point-mutation rows); +4 new (qrdr helpers + intrinsic-exclusion);
+  cohort regression re-pinned to 0.925/0.875/0.973. 24 total green.
+
+## [0.3.2] — 2026-06-06
+
+Trust-hardening + honesty taxonomy (adversarial-review follow-through). No new science — makes the
+shipped decoders auditable + honest about their blind spots.
+
+### Added
+- **Blind-spot honesty:** every SUSCEPTIBLE `dna-amr` call now carries `undetectable_mechanisms`
+  (`efflux` / `porin_loss` / `regulatory` — expression/regulatory resistance absent from AMRFinder's
+  curated determinants) + a caveat that a negative means "no curated determinant found", not "definitely
+  susceptible". `UNDETECTABLE_MECHANISMS` in `amr_rules.py`.
+- **Discordance taxonomy:** `discordance_bucket(prediction, true_label)` + `evaluate_cohort` now emit a
+  failure-mode breakdown — `FN_undetected_mechanism` (R missed → the blind spots) vs
+  `FP_determinant_without_phenotype` (called R but susceptible → label noise / silent-or-low-expression /
+  borderline MIC). The "failure-tolerant tool" deliverable: names where it fails.
+- **Provenance pin:** output JSON `provenance.amrfinder_image` records the pinned AMRFinderPlus image
+  (`ncbi/amr:4.2.7-2026-03-24.1`; tag encodes the DB date) so an R/S verdict is reproducible against a
+  known determinant source. `AMRFINDER_IMAGE_PINNED` in `amr_rules.py`.
+- **Value-add headline** in `wiki/dna_amr_multidrug_validation_2026-06-06.md`: explicit naive-AMRFinder
+  vs dna-amr table (cef +0.28 acc/+0.50 spec; gent +0.43/+0.57) — proves the per-drug policy adds value
+  over vanilla "any determinant → R", not just re-prints AMRFinder hits.
+- 5 new tests (blind-spots on S/R calls, discordance taxonomy, cohort discordance breakdown). 20 total.
+
+### Fixed
+- Stale gentamicin "NOT yet cohort-validated" claims (`amr_rules.py` docstring + wiki caveat) reconciled
+  with the N=128 acc 0.945 validation that DRUG_RULE/README/CHANGELOG already recorded (claim-hygiene).
+
+### Validated (cross-source — closes the same-database gap)
+- **Independent NCBI Pathogen Detection validation** (`scripts/xsource_validation.py`,
+  `wiki/dna_amr_xsource_validation_2026-06-07.md`): 22 E. coli, balanced ~11R/11S per drug, **zero
+  accession overlap** with the 176 BV-BRC cohort accessions (enforced at selection). Result: cipro 0.955,
+  cef 0.864, gent 1.000, tet 0.909 — comparable to / better than in-cohort. **Answers the product
+  question:** dna-amr beats naive "any-determinant→R" AMRFinder on UN-tuned data by +0.09 (cipro) /
+  +0.23 (cef) / +0.41 (gent) / +0.0 (tet) accuracy — the per-drug threshold + Subclass refinement IS the
+  value-add, not determinant discovery. Sensitivity 1.0 on all 4 drugs (FN=0); all errors are FP
+  (determinant-present-but-susceptible). Honest scope: NCBI Pathogen Detection is a different
+  source/curation, not a controlled different-lab study.
+
+## [0.3.1] — 2026-06-06
+
+The "one coherent tool" consolidation (after the v0.3.0 build settled the embedding question).
+
+### Added
+- **Unified `dna-decode` console entry** (`dna_decode/cli.py`) — single tool that dispatches to the
+  trait decoders: `dna-decode amr …`, `dna-decode pathotype …`, `dna-decode list` (capability +
+  validation status), `dna-decode --version`. Thin pass-through (argv delegated verbatim); the
+  per-decoder entries (`dna-amr`, `dna-pathotype`) stay independently usable + unchanged. 6 dispatch tests.
+
+### Changed
+- Project ledger (`project_state/dna-decode-2026-05-11.md`) updated to record the strategic inflection:
+  embedding frontier closed (0-for-3), deterministic mechanism-feature decoders are the product.
+
+## [0.3.0] — 2026-06-06
+
+The "deterministic decoders win" release. The frozen-genome-embedding (NT mean-pool) thesis was tested
+to a decisive conclusion and the project committed to deterministic, interpretable mechanism-feature
+decoders as the shipping path.
+
+### Added
+- **Multi-drug deterministic AMR caller** (`dna-amr`). Extended from cipro-only to **all 4 drugs**, with
+  per-drug validated rules baked into `dna_decode/eval/amr_rules.py::DRUG_RULE`:
+  - ciprofloxacin — threshold 2 (QRDR point-mutations) — N=147 acc 0.939.
+  - ceftriaxone — threshold 1 + **extended-spectrum subclass refinement** (CEPHALOSPORIN/CARBAPENEM;
+    excludes intrinsic blaTEM-1/blaEC that are ampicillin-R not ceftriaxone-R) — N=60 acc 0.933.
+  - tetracycline — threshold 1 (acquired tet genes) — N=12 acc 0.833 (small N, provisional).
+  - gentamicin — threshold 1 + **GENTAMICIN-subclass refinement** (excludes aph/aadA
+    streptomycin-kanamycin genes that don't confer gentamicin-R) — N=128 acc 0.945.
+  - General fix: a broad AMR class over-calls (cef spec 0.41, gent spec 0.39) because it counts genes for
+    OTHER class members; AMRFinder's Subclass field is the drug-specific discriminator. One-line refinement.
+  - `call_resistance(tsv, drug)` now auto-selects the per-drug rule; explicit threshold still overrides.
+  - Validation: `wiki/dna_amr_multidrug_validation_2026-06-06.md`.
+- **De-confound gate** (`dna_decode/eval/cohort_deconfound.py`) — within-lineage label-contrast
+  precondition (3-state DE_CONFOUNDED/WARN/CONFOUNDED + promotability) for any embedding-vs-classical
+  falsifier. The reusable guard against study==class confounding.
+- **AMR embedding falsifier** (`scripts/amr_falsifier.py`) + **QRDR-POINT knowledge baseline**
+  (`dna_decode/eval/point_baseline.py`) + within-lineage diagnostic.
+- **dna-amr external validation** — held-out N=29 acc 0.862 / sens 0.882 / spec 0.833
+  (`wiki/dna_amr_external_validation_2026-06-05.md`).
+- **EP-6 carbon-utilization substrate infra** — `dna_decode/data/bacdive.py` loader +
+  `scripts/bacdive_carbon_util_feasibility.py` census + `scripts/bacdive_li2023_to_long.py` adapter.
+
+### Findings (recorded, not code)
+- **AMR embedding thesis FALSIFIED on the cleanest substrate.** Cipro N=147 (de-confounded): NT-XGBoost
+  0.914 beats k-mer (+8.9 pp) but **LOSES to QRDR-POINT 0.943**; NT within-lineage concordance = chance
+  (0.605, p=0.365) → it learned lineage, not mechanism. Decision: `plans/AMR_embedding_niche_decision_2026-06-05.md`.
+- **Carbon-utilization (EP-6) E. coli-INFEASIBLE.** Data acquired (Li et al. 2023 OSF jwkr7); E. coli
+  slice = 27 strains, 0 carbon sources clear the ≥100 floor. The embedding-niche test needs a THIRD
+  requirement beyond sampling-independent-label + no-catalog: **organism-specific depth at scale**.
+  `wiki/bacdive_carbon_util_feasibility_2026-06-06.md`.
+
+### Changed
+- `dna-amr` CLI `--resistance-threshold` now defaults to the per-drug validated config (was hard-coded 2).
+- README decoder table + `plans/Trait_Decoding_Roadmap.md` Phase 2/4 updated.
+
+## [0.2.0] — 2026-06-05
+
+- In-package `dna-amr` console entry (deterministic AMR mechanism caller, cipro-validated).
+- `dna-pathotype` console entry (deterministic VirulenceFinder-marker pathotype resolver + abstention +
+  canonical-VF diff; ExPEC recall 0.917).
+- Packaging: both decoders ship in the wheel (`[project.scripts]`).
+
+## [pathotype-v0] / [phase-1-shipped] — 2026-05/06
+
+- Phase-1 closeout: NT-frozen-pooling characterized (passes concentrated-signal mechanisms, fails
+  distributed). v0 cipro decoder + pathotype v0 resolver shipped. See git history + `wiki/`.
+
+## [expression-context-v0] — 2026-06-10
+
+- `dna_decode/eval/expression_context.py` — deterministic IS-element-upstream-of-target detector (blastn ISAba1+OXA-51 vs assembly; all-hits, no truncation; same-contig + strand-aware upstream proximity; offline-safe). Reads regulatory CONTEXT, not gene presence, to cross the EXPRESSION abstain floor.
+- `amr_rules.call_resistance` gains an optional `genome_fasta` + a GATED EXPRESSION_FLOOR ABSTAIN->R override (registry `expression_context.enabled`; ships **off/experimental** — opt-in only, never default-on; default decoder behavior unchanged).
+- Independent-cohort validation (15R/15S disjoint from the cached 30): **HOLD — UNDERPOWERED, not a falsification**. A mechanism audit found 14/15 R carry strong acquired carbapenemases (non-target); only 1 strain is intrinsic-only-R (the signal's target subset), detector-negative. The validator was stratified by acquired-carbapenemase and now gates on `target_R_rescues>=1 AND n_target_R>=10` — the cohort cannot test the signal (intrinsic-only carbapenem-R Acinetobacter is rare; acquired OXA-23 dominates). Override remains disabled (opt-in/off). 18 new tests, 0 regressions.