Skip to content

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[0.0.33] - 2026-08-25

Honest-uncertainty release. Every headline number the report shows now carries an interval that accounts for how the cohort is actually structured, and the report states which negatives it was scored against — the choice that moves the headline more than any modeling decision. Also ships the pipeline DAG as a drawn object and the report's visual rebuild.

Added

  • The verification-bias ladder as a panel (#136). Anchor choice moves the headline 13.1 AUC points and the report stated it in the third clause of a caption. Four rungs now, scored on the same model and the same positives with only the negative class changing: verified within-patient negatives 0.845, all negatives pooled 0.856, unpaired negatives 0.865, healthy donors 0.976 — each with its n, patient count and a patient-clustered interval. The unpaired rung existed in the campaign record but never in the report. The donor rung is labelled as a different question (between people rather than within a patient), not as an unreliable one: at 55 negatives its interval is only ~1.7× wider than the verified rung's, checked against a stratified two-sample bootstrap and Hanley–McNeil before the claim was drawn.
  • The Nextflow DAG drawn into the report (#127), in Methods (two densities, structure only) and Run diagnostics (this run's task counts above the tables they summarise). Layouts are computed at build time from one declaration and injected as finished SVG; the page never positions a node. tests/test_pipeline_diagram.py asserts the declaration against the workflow in both directions, so a stage added to the DAG cannot leave the diagram describing the old topology.
  • Report visual rebuild (#126): a hero zone carrying the primary endpoint, the burden-response curve and the stacking lift; findings that title the chart they are derived from; a 16-term glossary behind info icons.

Changed

  • The primary endpoint carries an interval (#132). Sensitivity at 98% specificity against verified true negatives had shipped as a bare point estimate — _wilson existed but was wired only to the LOD bins. It is now 0.483 [0.424, 0.519] from a patient-clustered bootstrap with the threshold re-estimated inside every resample, because the threshold is itself an estimate from a finite negative set. Decomposing the width separates two things that had been conflated: clustering inflates the variance 1.26×, estimating the threshold inflates it 11.3×.
  • AUC intervals resample patients, not samples (#132). _bootstrap_auc resampled rows, so every evaluator and holdout interval in the scoreboard treated a patient's repeated timepoints as independent. Measured design effect 1.21 on the stacking AUC — intervals widen ~10%. CI values from earlier releases are not comparable to these.
  • Subgroup AUCs gain a per-class floor, an interval and tier composition (#134). The panel printed retinoblastoma's AUC on 72 positives at the same visual weight as NSCLC's on 2,254. The floor now counts the scarce class, rows carry patient-clustered intervals (primary evaluator only — every other subgroup table is exploratory), and each row reports the share of its positives that are tumour-confirmed, which ranges from 51% to 90% across histologies.
  • Dual-anchor operating points and the one-sided donor bound (#125, closing #123); per-sample depth metadata as first-class label columns (#122).

Fixed

  • Chart axis overrides no longer erase axis titles — PL_LAYOUT merged layouts with a shallow Object.assign, so a partial {xaxis:{…}} replaced the whole axis object and dropped its title. Three charts had silently lost a label (#126).
  • Category ticks are forced one per row: plotly thinned them on short plots, and the row it dropped first was the best evaluator at the top of the chart (#126).
  • Two mypy errors that reproduce under the pinned toolchain but not in CI. Neither was type noise — the CLI one guarded a pivot_table result that unpacks the characters of a column name when pandas returns a flat Index (#130).
  • mkdocs build --strict passes again: four unannotated parameters and one generator without a return annotation had been failing the documented release gate.

Documentation

  • The label taxonomy is six tiers, and the docs said five — in eleven files, including AGENTS.md, which agents read on every turn. Worse, .agents/rules/labeling-hierarchy.md and the ctdna-labeling skill both stated that a CH-only Possible ctDNA+ is "demoted to Possible ctDNA−". The code demotes it to Undetermined (LABEL_UNDETERMINED), which is outside _MODEL_LABELS and therefore excluded from modelling — following the documented behaviour would have put CH-only samples in the negative class. docs/biology/ctdna-labeling.md contradicted itself, prose against its own code block three lines below.
  • The report interpretation guide described a report that no longer exists — four tabs, no Methods tab, none of this release's panels. It now covers five tabs, the primary endpoint and its interval, the verification-bias ladder, the run map, the subgroup floor and tier composition, and why the interval width is mostly threshold estimation rather than clustering. Its #page-5-cohort-qc link, a leftover from the retired Quarto dashboard, is repointed.
  • The API reference covered 5 of 13 modules. report_data — which decides what every number in the report means — and selection — which decides which features reach a model — were both undocumented while the scoreboard displaying their output had a page. All thirteen now have one, with cli on mkdocs-typer.
  • The release guide described a flow no release has used. It said to land the version bump on develop and PR develop → main; v0.0.32, v0.0.31 and v0.0.28 each merged a release/vX.Y.Z branch straight into main and back-merged afterwards. The documented path is also riskier — with develop as the staging area, anything merged there between the bump and the PR rides into the release unreviewed.
  • README claimed Boruta-SHAP as a multimodal selection option two bullets above its own table naming grootcv the default; Boruta is now a legacy extra that cannot be installed alongside arfs at all.

Infrastructure

  • Notebook hygiene tests (#130): a notebook written as nbformat 4.4 carries no cell ids, so nbdev's export hash moves on every run and the sync gate fails on an unrelated PR.
  • An sbatch wrapper for GPU research jobs outside the Nextflow DAG, and an isolated-env recipe that refuses to install unless the target really is isolated (#130).
  • Research scripts for the review campaign: the grouped-CV leakage delta, a two-measure depth re-check, the reporting-precision battery, and the matched-stratum histology analysis (#128, #132, #133).

[0.0.32] - 2026-08-18

Scoreboard-contract release (#108, closing #107): the scoreboard stops fabricating values for inputs it never saw. Cut back-to-back with v0.0.31 so the #96 confirmation run on iris exercises the GrootCV migration and the honest scoreboard in a single production run.

Changed

  • The scoreboard is rebuilt on an explicit, versioned contract (#108, found via #107: the shipped v0.0.29 scoreboard stamped selection_method=legacy_cohens_d, n_selected_features=0 and selection_overlap_pct=0.0 on all 26 evaluators — fabricated defaults for inputs the process never saw). Every field is now sourced from a named artifact or explicitly unknown (NaN / None / "unknown") and recorded in a per-row missing_fields column (schema_version=2.0 stamped on every row; fixed column set — an AUC column for every configured model, NaN when it didn't run). KREVIEW_SCOREBOARD now stages the selection-QC sidecars (the same channel REPORT consumes), so selection columns are real (mrmr, true counts, total inputs). The previously silent n_features=0-when-a-GPU-model-wins pathology is now visible as a named missing field. Dead monolithic-era code (embedded selection_qc probing, "stacking" branches, the legacy label, selection_overlap_pct) is deleted. extract_evaluator_summary is the single place that knows the model-results key shapes; a drift-canary test fails if a producer renames a key without the extractor learning it. scoreboard.py is now nbdev-generated from nbs/07_scoreboard.ipynb — the standalone-module exception list shrinks to reproducibility.py.

[0.0.31] - 2026-08-18

GrootCV migration release (#96): the multimodal all-relevant selector moves from the unmaintained BorutaShapPlus to arfs 3.0 GrootCV, validated head-to-head on the real v0.0.29 cohort before a line of pipeline code changed. run_hpc.sh now runs grootcv — the next production run doubles as the tabicl-inclusive confirmation before the library default flips.

Changed

  • Multimodal all-relevant selection migrated to arfs 3.0 GrootCV (#96, literature-backed and validated on the real v0.0.29 cohort). grootcv is the recommended --multimodal-selection for production (run_hpc.sh now uses it): measured selection stability Nogueira 0.85 vs 0.65 for boruta_shap (pairwise Jaccard 0.77 vs 0.53) with equivalent-or-better downstream AUC and sensitivity, and boruta's full-train selection a strict subset of grootcv's. New evidence-based knobs, exposed CLI→Nextflow: --selection-cutoff/multimodal_selection_cutoff (default 3.0 — the #96 sweep winner: dominates boruta_shap on stability, AUC and sens@100spec-healthy simultaneously) and --selection-n-iter/multimodal_selection_n_iter (default 10 — 0.96 selection agreement with n_iter=50 at ~6× less runtime; ~40 s on the production 944-feature super-matrix). GrootCV's LightGBM threads are pinned to task.cpus in the Nextflow module (oversubscription guard). Both container images now ship the arfs extra.
  • boruta_shap is deprecated and BorutaShapPlus left the core dependencies: it is mutually uninstallable with arfs 3.0 (BorutaShapPlus pins numpy<=2.0.0, arfs 3.0 needs numpy>=2.0.2). Reproducing pre-#96 runs requires the new [legacy-boruta] extra (conflicts with [arfs] by construction); using the strategy logs a deprecation warning. The mi default of --multimodal-selection is unchanged; flipping the default to grootcv is deferred to the tabicl-inclusive confirmation run (#96 rider).

Fixed

  • arfs strategies no longer crash on genomic-coordinate feature names (#96, found on the real v0.0.29 super matrix): LightGBM rejects feature names with JSON-special characters (FsdGenomewide__fsd_gw_chr8:46838888-…_ratio), so Leshy/GrootCV now fit under sanitized names and map selections back to the originals (_lgbm_safe_columns, collision-guarded). The docker smoke test exercises exactly this case, and the leshy/grootcv strategy tests are real passing tests again (the #91/#92 xfails are gone; the boruta_shap test skips visibly unless [legacy-boruta] is installed).

[0.0.30] - 2026-08-17

Report + rigor release: the reporting layer is rebuilt as one data-driven, PHI-guarded, self-contained page (#79, Quarto path deleted); the train/test split is patient-grouped (#101 — found by the new report's own integrity check); and the two terminal failures from the first production v0.0.29 iris run are fixed (#97 multimodal-ablation collapse, #98 stranded dashboards). Holdout metrics from this release onward are not directly comparable to earlier runs (see the split entry below).

Fixed

  • The train/test split is now grouped by patient (#101, found by the #79 report's own integrity check on the v0.0.29 run: 1,266 patients had samples in BOTH train and test, optimistically biasing every holdout metric). _assign_train_test_split uses StratifiedGroupKFold on PATIENT_ID — all timepoints of a patient travel together, every timepoint stays usable, label-tier stratification is approximately preserved (verified on the real 16,272-sample labels: 0 leaked patients, test fraction 0.203, per-tier drift ≤ 0.4 pp, deterministic). A post-split assertion makes a leaked patient a hard error. Healthy donors (no cBioPortal record, one sample per donor) now carry PATIENT_ID = SAMPLE_ID, keeping the column complete for all grouped consumers. The eval-stage CV folds deliberately stay ungrouped — measured on the real v0.0.29 matrices, patient-mixing inflates CV AUC by ≤ ~0.001 (an order of magnitude below fold noise); the grouped holdout is the honest headline. Full measurement tables and the documented decision are on #101. Note: holdout numbers from runs ≤ v0.0.29 are not comparable to grouped-split runs — expect shifts of ~0.005–0.02 from test-set re-composition (the leaked repeat-timepoint patients were the harder cases, so the old numbers were dragged down, not inflated).

Changed

  • The reporting layer is one data-driven page (#79). kreview report now takes the pipeline --outdir and renders a single self-contained HTML (plotly inlined from the installed package — zero CDN, air-gapped-safe) via the new kreview/report_data.py data layer: sortable scoreboard with per-evaluator deep-dive modals (per-model AUC+CI, fold AUCs, OOF-computed ROC/PR, calibration, decision curves, subgroup AUCs with an n≥30 floor, feature-group ablation stability), multimodal stacking, cohort composition with the train/test split (and a loud patient-leakage warning — see the split issue), and run diagnostics from the execution trace. Aggregates only: assert_no_phi refuses to emit anything sample-identifying. The #98 manifest semantics carry over (always written, failure recorded). Validated end-to-end on the real iris v0.0.29 outputs: 26 evaluators, 5 MB page, rendered in ~3 s.
  • KREVIEW_REPORT (Nextflow) stages channel inputs into the canonical outdir layout and runs the new renderer — process_low, 8 GB, no more 64→192 GB ladder. The separate KREVIEW_REPORT_MULTIMODAL process is gone (its content is the report's multimodal tab); multimodal/ablation inputs are optional sentinels (#97 pattern), so a failed optional stage degrades the tab instead of starving the report.

Removed

  • The Quarto render path (#79): both .qmd templates (~4,000 lines, 1,009 of them duplicated), _render_quarto_report/_find_quarto, render-time SHAP, and the quarto-cli, papermill and itables dependencies. kreview/feature_cards.py (used only by the deleted templates) is gone too. Report-only params --shap-samples, --shap-features and --cvd-safe removed from the CLI and Nextflow (eval-stage SHAP is untouched).

Fixed

  • A failed dashboard no longer strands the successful ones (#98, found on the iris v0.0.29 run where 17/26 rendered dashboards published nothing). kreview report still exits 1 when any dashboard fails (fail loud), but the KREVIEW_REPORT wrapper now applies the retry-then-degrade pattern from #59: non-terminal attempts fail (the retry ladder gets a chance), and the terminal attempt exits 0 so everything that rendered publishes. The CLI also always writes reports/report_manifest.json (succeeded/failed evaluators + pointers to per-failure *_render.log Quarto debug logs, which now publish too) — a partial reports/ directory is loud, never silently incomplete.

Changed

  • Multimodal ablation runs in the GPU container when GPU models are in play (#97 follow-up). KREVIEW_MULTIMODAL_ABLATION now routes its container, queue, memory, time and cluster options on params.multimodal_gpu_models — when GPU stacking models are requested, the ablation gets the GPU image and a GPU slot instead of silently falling back to a CPU model. The #97 loud CPU fallback remains for environments with no GPU.
  • Lint toolchain fully pinned: mypy==2.3.1 and ruff==0.15.4 join the exact nbdev/black pins in the [dev]/[all] extras — unpinned linter releases twice broke CI on untouched code.

Added

  • Vendored the official migrate-nextflow-code skill from nextflow-io/agent-skills (Apache-2.0, attribution in VENDORED.md) into .agents/skills/ — detect → fix → verify procedures for the strict-syntax migration class that produced #80 and the #97 config gotcha. Its detection tool, nextflow lint (26.04+), reports the tree at 0 errors / 36 style warnings.

Fixed

  • Multimodal ablation no longer collapses when the best stacking model is GPU-only (#97, found on the first production v0.0.29 iris run). The LOO ablation stage runs in the CPU container, but picked the best stacking model unconditionally — when that was tabicl (GPU-only), _build_model returned None, all 39 evaluator ablations failed with a baffling sklearn error, and the terminally-failed stage starved MULTIMODAL_MERGE and REPORT_MULTIMODAL, silently losing the merged multimodal results and dashboard. Three fixes, per the terminal-failure policy (degrade AND surface loudly):
  • multimodal_ablation now collects every model's stacking AUC, probes candidates best-first, and falls back loudly to the best model the environment can build — re-baselining deltas against the used model's own full-matrix AUC (deltas against an unbuildable model's baseline would be scientifically wrong). The substitution is recorded in ablation_results.json (ablation_model_fallback: requested/used/AUCs/reason).
  • If no candidate is buildable, it raises one clear RuntimeError naming the models — not 39 repeats of estimator ... Got None instead.
  • The workflow now passes ifEmpty(file('NO_ABLATION')) on the merge input — the multimodal_merge module already handled that sentinel; the workflow just never sent it. Verified behaviourally with a forced-fail ablation stub: merge + multimodal dashboard now survive a terminal ablation failure (and the pre-fix tree provably loses them). Guarded by a structural assertion in scripts/nextflow_stub_test.sh.

[0.0.29] - 2026-07-22

Hardening release: closes the recurring release-breakage classes from the 2026-07 review (fail-loud invariants, Nextflow v25–v26, centralized HPC env) and validates the shipped container end to end. The boruta_shap multimodal workflow (used by run_hpc.sh) is now smoke-tested inside the built CPU image on every CI run.

Added

  • Test coverage for the CLI, report.py, and the multimodal strategies (#63). Added a report.py smoke test (0% → 57% — covers the matrix-not-found guard) and CliRunner --help tests across the full nested command tree (eval ablate *, eval multimodal *), which guard command registration/wiring. A new test-strategies CI job installs the arfs extra and runs the multimodal strategy tests, which previously skipped in every environment — the same "looks protected, isn't" pattern that hid the original BorutaShap outage. Un-skipping them immediately surfaced that all three are broken against modern deps (not kreview bugs — the strategy dispatch is correct): arfs 2.4 is incompatible with sklearn 1.9 (force_all_finite, breaks leshy) and lightgbm 4.x (categorical_feature, breaks grootcv) — #91; and boruta_shap hits a BorutaShap→shap xgboost-parser error on some versions — #92. All three are now xfail(strict=False) with tracked issues so the breakage is visible rather than invisibly skipped (they xpass and alert when the deps are fixed). The arfs extra also now pins setuptools<81 (arfs imports the removed pkg_resources). Overall coverage 44% → 46%; CI floor 38% → 42%. (No nf-test harness — the Nextflow -stub-run in scripts/nextflow_stub_test.sh covers the workflow wiring, per the #80 maintainer decision.)

Changed

  • Centralized the HPC env setup and fixed inconsistent hardening (#58). The Singularity read-only-/home / cache-redirect env was copy-pasted across five modules and had drifted: PYTORCH_CUDA_ALLOC_CONF (the CUDA-OOM guard) was set in multimodal_single but missing from eval_gpu_single and ablate_gpu_single, and the report modules lacked HOME/TMPDIR/MPLCONFIGDIR. It is now defined once each (params.gpu_env_setup, params.report_env_setup) and interpolated into every module's script:, so the hardening is applied identically everywhere. scoreboard.nf's python3 heredoc now fails loud with a clear message if the interpreter is missing (Singularity PATH strip) instead of a cryptic exit 127.

Two deliberate design choices, contrary to the issue's original suggestion, because the nf-core iris institutional config sets its own beforeScript, withLabel resources, errorStrategy and cache: (1) the env stays in-script, not in a beforeScript, which is a single non-additive directive that would clobber (or be clobbered by) iris's; (2) the per-withName resource ladders stay as they are — they are defensive armour against iris's withLabel override, and Nextflow 24+ forbids the def-closure hoisting that would dedup them. Rationale recorded in .agents/memory/reference-iris-config-interaction.md.

Fixed

  • Published output nested a working-directory prefix (#84). Because a process' output: path carries the working subdir (selected/, fused/, output/, prep_out/, …) and publishDir already named the destination folder, results landed at matrices/selected/selected/…, matrices/fused/fused/…, models/multimodal/prep_out/…. Added saveAs: { fn -> file(fn).name } to the seven affected modules so files publish by basename — the tree now matches the documented structure (matrices/selected/X, models/multimodal/X). Publish-path only; nothing consumes params.outdir (all inter-process data flows through channels), so the DAG is unaffected. Guarded by a new flatness assertion in scripts/nextflow_stub_test.sh.
  • DuckDB reads masked failures as empty results (#61). Both chunked-read paths (_read_parquet_chunk, run_feature_sql) retried on any exception and then returned an empty DataFrame — indistinguishable from a legitimate 0-row read, so a persistently failing feature silently dropped those samples (iter_feature_chunks skips an "empty" chunk). They now classify the error: OperationalError (I/O, resource, OOM) backs off and retries, and raises if it never succeeds; ProgrammingError / DataError (bad column, malformed SQL, type mismatch) raises immediately — no wasted retries, no masking. The legitimate "no files found" case still returns empty.
  • eval_engine masked errors as low scores (#61). Two broad excepts turned real failures into a 0.0: mutual_info_score scored a feature 0 on any exception, and the ablation inner-CV appended 0.0 for a fold that raised. Both benign degenerate cases (constant feature, <2 classes, too-few-samples, empty folds) are handled explicitly before these points, so both now raise with context instead. The four legitimate degenerate return 0.0 cases are preserved. A new shared _numeric_feature_columns helper replaces two duplicated select_dtypes(include=np.number) filters and loudly flags any non-metadata column that is unexpectedly non-numeric (an object-dtype feature that would otherwise be dropped silently), consolidating the two sites into one implementation.
  • GPU retry ladder never engaged (#59). The GPU eval/ablation/multimodal wrappers always exited 0 (emitting an error-JSON on failure) to avoid a collect() deadlock — but that also meant errorStrategy='retry' never fired, so the 64→256 GB / gpushort→gpu escalation was dead code for transient CUDA OOM. The wrappers now fail loud on non-terminal attempts (exit non-zero → the ladder climbs) and only degrade gracefully (error-JSON + exit 0) once retries are exhausted, preserving the channel-closing invariant that prevents the deadlock (see commits 698c72e / dcfe356).
  • Silent evaluator drop on ablation failure (#60). kreview_eval.nf paired each matrix with its best_subset via combine(by:0) — an inner join — so an evaluator whose ablation failed and was ignored upstream was silently dropped from both CPU and GPU eval (missing from results entirely, with no error). Now uses join(by:0, remainder:true) with a NO_BEST_SUBSET fallback (identical to the ablation-off path) and a loud per-evaluator warning, so the evaluator is degraded-but-present, never dropped. This is also strictly safer than the old combine + .ifEmpty guards when the whole ablation channel is empty.
  • Failed evaluators masked as blank rows (#59/#60). A GPU wrapper that exhausts its retry ladder emits {"error": ...}; the scoreboard rendered that as an innocent all-NaN row. The scoreboard now carries a status column (OK / PARTIAL / FAILED / NO_RESULTS) with an error_detail, logs degraded evaluators loudly, and no longer drops an evaluator that throws during metric extraction (emits a FAILED row instead). Both report templates show the status column and a "Degraded evaluators" callout, and their stale kreview run fallback text (removed in #55) was corrected to the Nextflow invocation.

Added

  • scripts/test_nextflow_join_dropguard.nf — a standalone channel-logic regression test for #60 (3 matrices, 2 best_subsets → all 3 survive with a fallback). Run by scripts/nextflow_stub_test.sh, which also structurally asserts the #59 retry guard is present in all three GPU wrappers.

Removed

  • kreview run (monolithic pipeline command) and everything that existed only to serve it: the Nextflow KREVIEW_RUN process (run.nf), the params.pipeline_mode switch, and its withName resource block. The Nextflow multistage DAG is now the only way to run the pipeline. kreview run was a second, drifting implementation of the whole pipeline — the largest single source of the fix-one-path-miss-the-other bugs (see #55) — and was neither used nor exercised by CI. Individual stages remain available as subcommands (kreview label|extract|select|eval|fuse|report) for debugging.
  • --export-duckdb. It existed only on kreview run and was unreachable from the Nextflow DAG, so it was never usable on HPC. Feature matrices are written as parquet under the output directory and can be queried directly with DuckDB or pandas.
  • Four orphaned Nextflow modules (eval_cpu.nf, eval_gpu.nf, select.nf, eval_multimodal.nf). They were included by zero workflows and encoded an older DAG that ran straight off EXTRACT, bypassing SELECT/ABLATE — re-enabling one would have produced different, incorrect results.

Fixed

  • nextflow.config would not parse on Nextflow 24 or newer (#80), so the pipeline could not start at all on a current Nextflow — an outright adoption blocker, and invisible because nothing in CI ran Nextflow. Three constructs were rejected by the modern config parser:
  • try/catch around the nf-core institutional includeConfig. Replaced with the nf-core ternary idiom, which also honours NXF_OFFLINE (an if statement is rejected too).
  • Five ${manifest.version} references in container tags — manifest is not resolvable from process/profiles scope. The version now lives in params.kreview_version, which the manifest reads back, so there is still exactly one version literal in the file.
  • ${HOME} interpolation in the Singularity cacheDir, now env('HOME').
  • Stale withName: selectors for KREVIEW_EVAL_CPU/KREVIEW_EVAL_GPU, left behind when the Gen-1 bulk modules were deleted. They matched no process and made Nextflow warn on every run; the resource settings they carried applied to nothing.

Added

  • Nextflow stub smoke test (scripts/nextflow_stub_test.sh, and a stub profile). Every process now declares a stub: block, so -stub-run -profile stub exercises the entire DAG — config parsing, all four profiles, module includes and channel wiring — in seconds without data or containers. CI runs it against both ends of the supported range. It asserts from execution_trace.txt that all 17 processes ran and every task reached COMPLETED, and fails on any withName: selector naming a process that does not exist.

Changed

  • Documentation now presents a single run path. --pipeline_mode multistage is no longer needed (or accepted); use -profile docker locally and -profile iris/slurm on HPC.
  • Supported Nextflow range is now declared and enforced: v25–v26. manifest.nextflowVersion was !>=22.10.1, an unbounded floor that let a modern Nextflow get far enough to fail with a cryptic parse error. It is now !>=25.04.0 (the floor env() requires), verified against 25.04.6, 25.10.6 and 26.04.6.

[0.0.28] - 2026-07-09

Fixed

  • Silent Report Failures: kreview report now exits with code 1 when any dashboards fail to render, enabling Nextflow to retry with higher memory allocations. Previously exited 0 even when multiple reports failed.
  • ValueError in Feature Plotting: Added a fallback in report_template.qmd when the univariate top_col (selected by eval_stats) is filtered out during feature selection or matrix fusion. Falls back to the first available numeric column instead of crashing Plotly.
  • CUDA OOM on Shared GPU Nodes: Added PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to the KREVIEW_MULTIMODAL_SINGLE_GPU Nextflow process, reducing PyTorch memory fragmentation on shared Slurm nodes. Prevents TabICL_ft failures from CUDA allocation errors.
  • IPython Home Path on HPC: Exported IPYTHONDIR="$PWD/.ipython" in report.nf, report_multimodal.nf, and multimodal_single.nf to prevent errors when /home is read-only on HPC Singularity nodes.
  • Healthy Normal Specificity = 0.0 in Multimodal Stacking: The stacking matrix parquet files did not carry the original 4-tier label text (e.g. "Healthy Normal", "Possible ctDNA-"), so evaluate_model() could not compute sensitivity at 100% healthy-normal specificity. Fixed by plumbing oof_sample_labels through the full pipeline: evaluator JSONs → baselines loader → stacking matrix alignment → parquet persistence → evaluate_model() in all paths (decomposed Nextflow + monolithic kreview run).

Added

  • _sample_label Column: The stacking_matrix.parquet and raw_features_matrix.parquet outputs from multimodal_prep() now include a _sample_label column carrying the 4-tier text labels. This enables downstream consumers (multimodal_single, dashboard) to compute per-label-tier specificity without joining back to labels.parquet.

[0.0.27] - 2026-06-25

Fixed

  • GPU Stacking random_state Duplicate: Resolved a TypeError: got multiple values for argument 'random_state' occurring during fine-tuning of TabPFN and TabICL GPU multimodal models in multimodal_eval() and multimodal_single(). This was caused by random_state being passed as both a positional argument and duplicated via unpacked gpu_kwargs.
  • Baseline Best Metric Recovery: Fixed a bug in _load_per_evaluator_baselines() where best_model and best_auc keys were resolved as None from the source model JSONs. They are now dynamically computed from the model's actual OOF AUC values.
  • Nextflow GPU Singularity Path: Added symlinks for python3 and python to /usr/local/bin inside the Dockerfile, and replaced container python sub-processes with native grep extraction in multimodal_single.nf's script blocks. This prevents python3 command not found (exit code 127) errors triggered by Singularity path-stripping behavior on HPC nodes.
  • GPU Pipeline Deadlock Avoidance: Configured KREVIEW_MULTIMODAL_SINGLE_GPU to catch non-zero exit codes during evaluation failures and output an error metadata JSON instead of terminating the pipeline immediately. This prevents Nextflow's .collect() channel from deadlocking when a GPU node experiences a runtime error.
  • Resource Limits Hardening: Bumped KREVIEW_ABLATE_CPU_SINGLE memory allocation from 32 GB to 64 GB to prevent OOM errors on large transcription factor annotation matrices, and bumped KREVIEW_EXTRACT walltime to 1.5h.

Added

  • Dashboard Readiness Column: Appended the unique _sample_id column to both the multimodal stacking_matrix.parquet and raw_features_matrix.parquet. This enables joining multimodal predictions back to labels.parquet for per-label-tier (e.g. Healthy Normal) and per-cancer-type performance comparisons in the kreview dashboard.

[0.0.26] - 2026-06-23

Fixed

  • BorutaShap Import: The pip package BorutaShapPlus provides an importable module named BorutaShap, not BorutaShapPlus. Corrected from BorutaShapPlus import BorutaShap → from BorutaShap import BorutaShap in eval_engine.py. This was the root cause of multimodal pipeline failures in v0.0.24 and v0.0.25 — previously misdiagnosed as a container build issue.
  • Test Skip Guard: Fixed test_boruta_shap_passes_through which used import BorutaShapPlus as its skip guard. Since the module name is BorutaShap, the test was always being silently skipped even when the package was correctly installed.
  • Report Memory (OOM): Bumped KREVIEW_REPORT and KREVIEW_REPORT_MULTIMODAL memory from 32 GB → 64 GB base (128 GB on retry). In v0.0.25, 8/26 reports OOM'd at SHAP summary plot computation (cell 31/53) and 3/26 OOM'd at pandoc --embed-resources HTML self-containment. Evaluators with 270+ features at --shap-samples 5000 require more memory than the 32 GB allocation.

[0.0.25] - 2026-06-22

Fixed

  • Container Dependencies: Moved papermill>=2.5.0 from [jupyter] optional extra to core dependencies. Quarto's --execute-params (-P) flag requires papermill at runtime, and the CPU Docker container did not include it — causing all 26 evaluator reports to fail on HPC with "The papermill package is required for processing --execute-params".
  • BorutaShapPlus Container Availability: Ensured BorutaShapPlus>=0.1.3 (already in core deps since v0.0.24) is present in rebuilt containers. The v0.0.24 container was built from a stale state, causing multimodal prep to fail with "No module named 'BorutaShapPlus'" and cascading failure of the entire multimodal pipeline chain.

[0.0.24] - 2026-06-18

Added

  • ARFS Feature Selection: Added Leshy (Boruta+LightGBM/SHAP) and GrootCV (cross-validated SHAP importance) as --multimodal-selection options. Install via pip install kreview[arfs].
  • Strategy Validation: _select_multimodal_features() now raises ValueError on unknown strategy names instead of silently falling through to MI ranking. Typos like --multimodal-selection borta_shap are caught immediately.
  • MI-Reduction Helper: Extracted _mi_reduce_confirmed() to deduplicate the >500-feature MI reduction logic across Boruta-SHAP, Leshy, and GrootCV strategies.
  • Tests: Added test_invalid_strategy_raises, test_valid_strategies_constant, test_leshy_passes_through, and test_grootcv_passes_through to test_selection.py.

Changed

  • BorutaShap → BorutaShapPlus: Migrated from BorutaShap>=1.0.17 to BorutaShapPlus>=0.1.3 with scipy binom_test and NumPy np.NaN compatibility shims.

Fixed

  • Quarto EROFS on Read-Only Singularity: Quarto 1.9+ creates a .quarto/ project directory next to the QMD source. Inside read-only Singularity containers, this failed with EROFS (error 30). Templates are now staged to a writable .render_{feat_name}/ directory before quarto render.
  • Report publishDir Double-Nesting: Fixed publishDir "${params.outdir}/reports" → "${params.outdir}" in report.nf and report_multimodal.nf to prevent results/reports/reports/ nesting.
  • GrootCV Constructor: Fixed invalid parameters (estimator, n_iterations, cv) → correct ARFS API (objective="binary", n_folds=5, n_iter=50, silent=True).
  • Leshy Constructor: Fixed n_iter (not a valid param) → max_iter=50; added n_estimators=1000, verbose=0.
  • Test Import: Updated test_selection.py skip guard from import BorutaShap to import BorutaShapPlus.

[0.0.23] - 2026-06-16

Fixed

  • Scoreboard Nextflow Quote-Escaping: Replaced python3 -c "..." with bash heredoc (python3 << 'EOF') in scoreboard.nf to prevent Nextflow shell interpolation from stripping escaped quotes, which caused NameError: name 'evaluator' is not defined and KREVIEW_SCOREBOARD failing all retry attempts.
  • scipy.stats Version Attribute: Fixed AttributeError in the binom_test compatibility shim by importing scipy (top-level) instead of accessing __version__ on scipy.stats (which has no such attribute). This caused KREVIEW_MULTIMODAL_PREP to crash when --multimodal-selection boruta_shap was used (scipy ≥1.12).
  • Quarto HPC Cache Permissions: Export XDG_CACHE_HOME and XDG_DATA_HOME to writable local paths in report.nf and report_multimodal.nf, preventing Quarto from failing on HPC compute nodes with read-only ~/.cache directories.

[0.0.22] - 2026-06-11

Fixed

  • WPSGenome MAD Memory Optimization: Replaced the nested Python loop that materialized ~446M rows in memory with a DuckDB-native window function query, reducing peak memory usage to under 100MB and runtime to ~7s.
  • GPU Ablation TypeError: Fixed a bug where _build_ablation_model_factories called GPUModelCVAdapter directly with invalid model_name argument instead of using _build_gpu_model builder, causing GPU ablation to silently fail and default to ALL.

[0.0.21] - 2026-06-10

Performance

  • WPSGenome SQL pushdown: Uses DuckDB native list_avg(), list_max(), list_min() to aggregate 1.9B rows inside DuckDB instead of materializing in pandas. Reduces peak memory from 96+ GB to ~17 GB.
  • DuckDB resource tuning: Default threads 4→8, memory 4GB→32GB, exposed via --duckdb-threads / --duckdb-memory CLI options.
  • Nextflow KREVIEW_EXTRACT: cpus 4→8, memory 32GB→64GB per attempt, maxRetries 7→3 (SQL pushdown eliminates the memory ladder).

Added

  • configure_duckdb() API for pre-connection resource configuration.
  • --duckdb-threads and --duckdb-memory CLI options on extract and run commands.
  • sql_pivot_column evaluator property for multi-row SQL result pivoting.
  • Exact MAD hybrid computation via second lightweight DuckDB query.
  • Nextflow params: duckdb_threads, duckdb_memory.

Fixed

  • Nextflow resilience: Added explicit errorStrategy + maxRetries to 14 processes that relied on global defaults. One failed evaluator no longer terminates the entire pipeline — it is ignored after 3 retries.
  • KREVIEW_FUSE uses terminate (not ignore) since downstream processes depend on the super-matrix.
  • Added ifEmpty channel guards on EXTRACT→SELECT and GPU ablation .combine() to prevent pipeline deadlocks when upstream processes are ignored.
  • Added missing withName config blocks for KREVIEW_MULTIMODAL_SINGLE_CPU, KREVIEW_MULTIMODAL_SINGLE_GPU, and KREVIEW_MULTIMODAL_MERGE.

Changed

  • get_duckdb_conn() defaults: 8 threads, 32GB memory (was 4 threads, 4GB).

[0.0.20] - 2026-06-09

Added

  • Feature Group Ablation: kreview eval ablate {cpu,gpu,merge} CLI commands for nested CV feature group selection. Identifies suffix-based feature groups, evaluates all subsets via inner CV, and picks the optimal subset per model per fold using sensitivity_at_100spec_healthy.
  • Nested CV Pipeline: Per-model per-fold feature subsets flow from ablation → merge → eval via --best-subset. Supports both CPU (LR/RF/XGB) and GPU (TabPFN/TabICL) models.
  • Nextflow Ablation Stages: 3 new modules (ablate_cpu_single.nf, ablate_gpu_single.nf, merge_ablation.nf) wired into kreview_eval.nf behind params.run_ablation flag. All stages run in parallel per evaluator.
  • Scoreboard Columns: nested_cv flag and best_sens_100spec_healthy (best sensitivity across ALL models) added to build_scoreboard().
  • Bootstrap CIs: _compute_oof_metrics() produces 95% bootstrap CIs for AUC, sensitivity@100%spec, and all clinical metrics.
  • GPU Subgroup Stats: gpu_models() now reports assay_stats and cancer_type_stats for subgroup analysis.
  • Test Suite: tests/test_ablation.py with 17 tests covering feature grouping, subset generation, OOF metrics, backward compatibility, and scoreboard.

Changed

  • eval cpu and eval gpu accept optional --best-subset for nested CV with per-fold feature subsets.
  • eval_cpu_single.nf and eval_gpu_single.nf accept optional best_subset input channel (backward compatible via sentinel file).
  • nextflow.config: new params run_ablation (default: false), ablation_inner_folds (default: 3).

[0.0.19] - 2026-06-08

Fixed

  • WPSGenome Extraction Failure: Removed broken SQL pushdown (extract_sql()) that used trim(col, '[]') on native list<float> arrays, silently producing NULL results. Replaced with pure-Python extract() using _to_array() helper that handles numpy arrays, Python lists, and legacy string-encoded arrays. Added _to_array() as a public utility.
  • WPSGenome Last-Row-Wins Bug: Rewrote extract() to aggregate per-region-type (TSS, CTCF) statistics across all regions instead of overwriting with the last row. Now computes mean, peak_valley, std, and MAD of per-region array means.
  • WPSPanel Last-Row-Wins Bug: Rewrote extract() to aggregate per-region-type (TSS, CTCF) statistics across all panel regions (~59-85 rows per type) instead of overwriting with the last row. local_depth now averaged across panel regions instead of using last value.
  • GPUModelCVAdapter Serialization: Added __getstate__/__setstate__ to exclude unpicklable model_factory lambda from pickle. Deserialized adapters work for inference (predict_proba) but raise TypeError on fit() — preventing silent misuse.
  • Joblib Corrupt File Detection: _save_fitted_models() now validates post-write file size (>100 bytes) and deletes truncated files. Error-level logging replaces silent log.warning. Partial files from failed joblib.dump() are cleaned up.
  • parse_array() Numpy Passthrough: Now handles numpy arrays and Python lists as input (passthrough), not just string-encoded arrays. Previously returned [] for native parquet list<float> data, causing silent data loss.

Changed

  • WPSBackground Cleanup: Removed bare try/except for consistency with WPSGenome/WPSPanel. Errors now propagate to CLI handler. Added structured logging for empty/success cases with feature counts and chromosome counts. Metrics list moved to class constant _metrics.

Removed

  • WPSGenome SQL Pushdown: Removed extract_sql(), sql_pivot_column, and CLI pivot code. See v0.0.17 entry — the SQL path was fundamentally incompatible with native array storage.
  • WPSGenome FFT Features: Dropped spectral_max_power and spectral_dominant_freq. Empirical analysis (r=0.91–1.00 with basic stats) proved FFT is redundant with mean/std/peak_valley. Nucleosome periodicity is captured by WPSBackgroundEvaluator.
  • WPSPanel spectral_max_power: Removed — real-data analysis (4 patient samples) showed r=0.90–1.00 with std after per-region-type aggregation. spectral_dominant_freq retained (38.5% CV in wps_tf across samples — potentially informative for curated TSS/CTCF loci).

[0.0.18] - 2026-06-05

Added

  • Fine-tuned GPU Model Variants: tabpfn_ft and tabicl_ft via FinetunedTabPFNClassifier and FinetunedTabICLClassifier with configurable --finetune-epochs (default: 50) and --finetune-lr (default: 1e-5). Total GPU model count: 4 (was 2).
  • GPUModelCVAdapter: sklearn-compatible wrapper in eval_engine.py that resolves AttributeError: 'FinetunedTabPFNClassifier' has no attribute 'classes_' during cross_val_predict. Exposes classes_ as a plain attribute set during fit().
  • kreview eval multimodal nested subgroup: 5 subcommands — run (monolithic backward-compat), prep, single, ablation, merge.
  • Decomposed multimodal Nextflow pipeline: 4 new NF modules (multimodal_prep.nf, multimodal_single.nf, multimodal_ablation.nf, multimodal_merge.nf) enabling per-model parallel execution. Legacy eval_multimodal.nf kept for standalone testing via kreview eval multimodal run.
  • GPU foundation model expansion: 4 GPU model variants (tabpfn, tabpfn_ft, tabicl, tabicl_ft) evaluated per-evaluator in parallel via KREVIEW_EVAL_GPU_SINGLE. Each process runs all requested GPU models and outputs a single {evaluator}_gpu_model_results.json.
  • Intentional 7-color palette: Vanilla/FT shade families (#00e5ff/#00b0ff for TabPFN, #76ff03/#64dd17 for TabICL) in report templates.
  • Stacking feature selection: MI or Boruta-SHAP on stacking matrix via --stacking-selection.
  • Decomposed NF resource configs: Per-process withName blocks for KREVIEW_MULTIMODAL_PREP, KREVIEW_MULTIMODAL_SINGLE_GPU, and KREVIEW_MULTIMODAL_ABLATION.

Changed

  • Default --gpu-models → tabpfn,tabpfn_ft,tabicl,tabicl_ft (4 models, was 2).
  • kreview eval multimodal uses nested Typer sub-commands (was flat hyphenated commands).
  • NF config: decomposed process resource configs replace monolithic withName: 'KREVIEW_EVAL_MULTIMODAL'.
  • Report _name_map includes TabPFN-FT and TabICL-FT display names.
  • multimodal_cpu_only NF param marked as legacy (decomposed pipeline routes CPU/GPU per-process).

Removed

  • --no-finetune CLI flag (use --gpu-models tabpfn,tabicl for zero-shot only).
  • Dead KREVIEW_EVAL_MULTIMODAL import from kreview_eval.nf.

Fixed

  • TabPFN AttributeError: 'FinetunedTabPFNClassifier' has no attribute 'classes_' via GPUModelCVAdapter wrapper.
  • Stale multimodal-prep (hyphen) references in CLI help strings → multimodal prep (space).
  • Stale DAG comments in eval_cpu_single.nf, eval_gpu_single.nf, report_multimodal.nf referencing removed KREVIEW_EVAL_MULTIMODAL.

[0.0.17] - 2026-06-04

Added

  • GPU Model Pre-flight Validation: _validate_gpu_models() in cli_eval.py verifies GPU model packages are installed and TABPFN_TOKEN is set before any work begins. Shared across eval gpu, eval multimodal, and monolithic run commands. Exits with code 1 and actionable instructions on failure.

Fixed

  • OOF Sample IDs Alignment (P0): oof_sample_ids scoped to train-split only in CPU, GPU, and monolithic paths to match oof_probs length from cross_val_predict. Added oof_sample_ids and oof_labels to merge skip set preventing GPU from overwriting CPU alignment keys. This was the root cause of multimodal stacking failure ("25 had IDs but no probs").
  • BorutaShap scipy Compatibility (P1): Monkey-patched scipy.stats.binom_test with binomtest().pvalue wrapper for BorutaShap compatibility on scipy ≥1.12 (removed in 1.12). Eliminates ImportError: cannot import name 'binom_test' in multimodal raw feature selection.
  • Scoreboard Crash Handling (P2): Added try/except with traceback to scoreboard.nf inline Python. Moved file writes inside try block. Added KREVIEW_SCOREBOARD withName config block (retry 3×, then ignore). Added per-evaluator try/except in build_scoreboard() so a malformed evaluator result doesn't kill the scoreboard for all evaluators.
  • WPSGenome Timeout (P3): Implemented DuckDB SQL pushdown (extract_sql()) for WPSGenomeEvaluator — parses string-encoded arrays via list_avg/list_max/list_min in C++. Added sql_pivot_column for tall-to-wide pivot in CLI. Used pivot_table(aggfunc='first') for duplicate safety with null guard on pivot failure fallback. Added time/queue escalation in nextflow.config (retries 3+ → cmobic_cpu 8h). Note: This SQL pushdown was later found to be broken — krewlyzer stores arrays as native FLOAT[], not string-encoded. Removed in [Unreleased].
  • GPU Silent Failure Logging (P4): Record per-model error keys ({model}_error) in GPU results dict. Detect all-models-failed state and set results["error"]. Added visible ⚠ WARNING stdout output in both cli_eval.py and monolithic cli.py.
  • REPORT Blocked by SCOREBOARD (P5): Guarded REPORT channel with .ifEmpty(file('NO_SCOREBOARD')) so reports render even when scoreboard fails. Templates already check scoreboard_path.exists().
  • REPORT_MULTIMODAL SLURM Rejection: Added withName: 'KREVIEW_REPORT_MULTIMODAL' config block with explicit queue = params.partition ?: 'cmobic_short'. Previously fell through to process_medium label which lacked a queue directive, causing iris institutional config to inject the inaccessible cpu partition into sbatch.
  • Nextflow Channel Deadlocks: Added .ifEmpty([]) guards on JSON collect, joblib collect, and GPU collect channels to prevent pipeline deadlock when tasks fail with errorStrategy = 'ignore'.
  • Pivot Crash on Duplicate Rows: Switched from pivot() to pivot_table(aggfunc='first') in CLI wide-pivot for WPSGenome tall-to-wide conversion, preventing ValueError: cannot reshape on duplicate region_type rows.
  • Null Guard on Pivot Failure: Added null check on feat_matrix after pivot failure fallback, preventing AttributeError: NoneType has no attribute 'columns'.
  • Queue Directive Safety Net: Added queue = params.partition ?: 'cmobic_short' to process_medium and process_high label defaults. Prevents any future process from failing SLURM submission due to missing partition.

Changed

  • Version Bump: __init__.py, settings.ini, nextflow.config → 0.0.17.
  • Version-agnostic run_hpc.sh: Auto-detects version from nextflow.config at runtime. Dynamic GPU_MODELS based on TABPFN_TOKEN availability. No manual edits needed between releases.

[0.0.16] - 2026-06-03

Added

  • 80/20 Stratified Holdout Validation: _assign_train_test_split() in CtDNALabeler.label_all() assigns a split column (train/test/exclude) to labels.parquet. Split is stratified by 4-tier label with random_state=42 for reproducibility. evaluate_holdout() in eval_engine.py provides unbiased AUC estimates on the held-out 20%.
  • Sensitivity at Fixed Specificity: evaluate_model() now computes sensitivity at 100%, 99%, and 95% specificity, plus a healthy-normal-only variant using sample_labels. 12 new JSON fields per model.
  • Intelligent GPU Feature Capping: gpu_models() accepts --max-gpu-features (default: 150) and caps features using score-based priority (MI > AUC > variance). Returns gpu_feature_cap_indices for holdout dimension consistency.
  • CH-Only → Undetermined Label: Samples with n_non_ch_variants == 0 and no SV/CNA/IMPACT match are now labeled Undetermined (was: Possible ctDNA−). Undetermined is excluded from binary classification via build_binary_target().
  • WPSGenome Streaming: extract_columns and max_chunk_rows class attributes on FeatureEvaluator enable DuckDB column projection and chunked reads. WPSGenome sets max_chunk_rows = 5_000_000 and projects to 4 columns, reducing peak memory ~80%.
  • Scoreboard Enhancements: build_scoreboard() surfaces sens_at_100spec, holdout_auc, holdout_sens_100spec, holdout_n_train/n_test, and auc_drop (CV overfit diagnostic).
  • LABEL_UNDETERMINED Constant: CtDNALabeler.LABEL_UNDETERMINED = "Undetermined" added alongside existing 5-tier constants.
  • Report Value Boxes: Both report_template.qmd and report_multimodal_template.qmd now display "Sens @ 100% Spec" (with detection count) and "Holdout AUC" (with dynamic color: green/yellow/red by AUC drop severity) in the Executive Summary.
  • Report Clinical Metrics: Model summary tables auto-discover and display Sens@100%Spec, Sens@95%Spec, and Holdout AUC per model.
  • Report Scoreboard Expansion: Scoreboard display expanded from 7 to 13 columns (best_model, holdout_auc, auc_drop, sens_at_100spec, sens_at_95spec, holdout_n_train, holdout_n_test).

Changed

  • Feature Selection Train-Only: score_features() and select_features() (mRMR + hybrid paths) now filter to split == "train" before scoring, preventing test set leakage into feature rankings.
  • LABEL_META_COLS: Added split to the 22-entry meta column set in core.py to prevent it from leaking into feature matrices.
  • Nextflow GPU Module: eval_gpu_single.nf now stages eval_stats parquet and passes --eval-stats-dir + --max-gpu-features to the CLI.
  • Nextflow Workflow: kreview_eval.nf wires SELECT.out.eval_stats → EVAL_GPU_SINGLE.
  • Version Bump: __init__.py, settings.ini, nextflow.config → 0.0.16.
  • Report meta_cols Dynamic Import: Both templates now import LABEL_META_COLS from kreview.core instead of maintaining a stale hardcoded copy. Prevents label metadata columns from leaking into feature sets.
  • Report Glossary: Added Undetermined label definition and train/test split documentation to both template glossary sidebars.
  • MODEL_LABELS / POSITIVE_LABELS Public API: Promoted from private _MODEL_LABELS / _POSITIVE_LABELS in selection.py to public exports. Backward-compat aliases maintained.

Fixed

  • GPU Holdout Dimension Mismatch: cli.py and cli_eval.py now apply gpu_feature_cap_indices to X_test before calling evaluate_holdout(), matching the capped training feature space.
  • Data Leakage in mRMR/Hybrid Selection: Both select_features() code paths now restrict MI scoring to train-only rows.
  • Unused Import: Removed pandas import from report.py (F401).
  • Boolean Comparisons: Replaced == False with ~ operator in labels.py (E712).
  • Ambiguous Variable: Renamed l → lbl in list comprehensions in eval_engine.py (E741).
  • Nested If Statements: Merged nested if blocks in eval_engine.py and registry.py (SIM102).
  • Ternary Simplifications: Converted if/else blocks to ternary in eval_engine.py (SIM108).
  • Test Lint: Removed 5 unused imports from test files and fixed 7 E712/SIM108 in test assertions.

[0.0.15] - 2026-06-01

Added

  • CPU+GPU JSON Merge Helpers: load_model_results(directory, evaluator_name) and load_all_model_results(directory) in eval_engine.py transparently discover and merge *_model_results.json (CPU) and *_gpu_model_results.json (GPU) files. GPU model keys (AUC, OOF probs, SHAP) are merged into the CPU dict. Used by report templates, scoreboard, and multimodal engine.
  • KREVIEW_SCOREBOARD: New Nextflow process (scoreboard.nf) generates scoreboard_combined__all.parquet after all CPU/GPU eval jobs complete. Uses build_scoreboard() which now calls load_all_model_results() for unified GPU+CPU scoring.
  • GPU Exit Code: kreview eval gpu now exits with code 1 if NO models produced valid AUC results (was silent success). Prevents Nextflow from treating empty GPU results as success.
  • Unit Tests: 11 new tests in test_merge_helpers.py covering CPU-only, GPU-only, CPU+GPU merge, malformed JSON, and directory scan scenarios.

Changed

  • GPU JSON Output Naming: GPU eval module now produces *_gpu_model_results.json (was *_model_results.json), preventing filename collision when CPU and GPU outputs are collected into the same Nextflow channel.
  • Report Input Signature: KREVIEW_REPORT now accepts 6 inputs (matrices, model_results, eval_stats, selection_qc, joblib_files, scoreboard) for complete dashboard rendering.
  • Workflow Wiring: kreview_eval.nf creates unified ch_all_jsons and ch_all_joblib channels that mix CPU+GPU outputs, feeds them to SCOREBOARD → REPORT.
  • Scoreboard Loading: build_scoreboard() now uses load_all_model_results() instead of manual glob+load loop.
  • Multimodal Baselines Loading: _load_per_evaluator_baselines() now uses load_all_model_results() instead of manual glob+load loop.
  • FSD Density Calculation: Added select_dtypes(include="number") filter before density computation in fsd.py and fsd_genomewide.py to prevent TypeError on non-numeric metadata columns.
  • Multimodal NF Module: Removed unnecessary staging loop — --results-dir . uses Nextflow work dir symlinks directly.
  • Dockerfile Optimization: Selective builder COPY (only pyproject.toml, kreview/, LICENSE), single pip install "${WHL}[gpu]" pass to ensure local wheel resolution, dropped python3.12-dev and wget from GPU runtime, merged OCI labels into single LABEL instruction.
  • CI Parallelism: test.yml split into parallel test (Python) and docker (matrix [cpu, gpu]) jobs with fail-fast: false. GPU build gets jlumbroso/free-disk-space cleanup (~20-30 GB freed).
  • .dockerignore: Expanded to exclude docs/, nextflow/, scripts/, tests/, .github/, .agents/ — reduces build context transfer.

Fixed

  • OOF Label Key Search: "oof_labels".endswith("_oof_labels") is False — fixed to check both exact match and suffix match. Applied to both report_template.qmd and report_multimodal_template.qmd.
  • FSD TypeError: Non-numeric columns (e.g., sample_id, filename) caused TypeError: ufunc 'divide' not supported in FSD density calculation. Now filtered to numeric columns only.
  • Multimodal Validation Logging: Now shows CPU vs GPU JSON counts separately for better debugging.
  • Docker GPU Build CI Failure: GPU image build exceeded runner disk space (~14 GB free). Fixed by adding jlumbroso/free-disk-space action and optimizing Dockerfile layers.

Removed

  • tabpfn-extensions[interpretability] from [gpu] extras — unused since v0.0.13 (SHAP computation replaced by shapiq). Eliminates transformers, wandb, huggingface-hub, tokenizers transitive dependencies (~500 MB).
  • Deprecated HPC scripts: run_hpc_0.8.3.sh, run_hpc_v0.0.10.sh, run_hpc_v0.0.11.sh, runner.sh, utils/symlinker.py — replaced by unified scripts/run_hpc.sh.

Breaking Changes

  • GPU JSON output renamed: {eval}_model_results.json → {eval}_gpu_model_results.json
  • KREVIEW_REPORT input signature expanded from 2 to 6 inputs

[0.0.14] - 2026-05-25

Added

  • Reproducibility Seed: --seed CLI flag on run, eval cpu, eval gpu, eval multimodal, and select commands (default: 42).
  • Deterministic Mode: --deterministic / --no-deterministic flag for PyTorch cudnn (default: True) on all eval commands.
  • kreview/reproducibility.py: seed_everything() utility for Python, PyTorch, and cuDNN seeding.
  • Nextflow Params: seed and deterministic params in nextflow.config, forwarded through all 8 pipeline modules.
  • Reproducibility Docs: New section in models-and-metrics.md documenting seed propagation strategy.

Fixed

  • 3 hardcoded random_state=42 in _select_multimodal_features() now accept caller seed.
  • 7 internal forwarding gaps where random_state was not passed to child functions.
  • GPU models (TabPFN, TabICL) now receive random_state parameter via _build_gpu_model().
  • shapiq.TabularExplainer and shap.PermutationExplainer now seeded for reproducible SHAP values.
  • BorutaShap.fit() now receives random_state for deterministic feature selection.
  • SLURM Partition Routing: All withName process blocks now pin queue explicitly to bypass the nf-core iris institutional config's dynamic queue closure that injects the inaccessible cpu partition into sbatch -p.
  • KREVIEW_LABEL Error Strategy: Changed from inherited 'ignore' (from iris config) to 'terminate' — pipeline aborts immediately if labeling fails instead of deadlocking.

[0.0.13] - 2026-05-24

Added

  • GPU Foundation Models: TabPFN v8.0.3 and TabICL v2.1 support via _build_gpu_model() with updated import paths (tabpfn.TabPFNClassifier, tabicl.TabICLClassifier).
  • shapiq Integration: SHAP values for GPU models now computed via shapiq.TabularExplainer (model-agnostic kernel-based Shapley values), replacing the deprecated tabpfn-extensions interpreter.
  • Unified Model Persistence: _save_fitted_models() helper in cli_eval.py provides consistent joblib saves for both CPU and GPU models. --skip-gpu-joblib flag opts out of large GPU model files.
  • Multimodal GPU Models: kreview eval multimodal now accepts --gpu-models tabpfn,tabicl for GPU-accelerated stacking and raw feature evaluation.
  • Pre-computed SHAP Fallback: Report templates render mean |SHAP| bar charts from JSON when joblib files are unavailable (e.g., --skip-gpu-joblib).
  • KREVIEW_REPORT_MULTIMODAL: New Nextflow process for rendering multimodal stacking dashboards in multistage mode.
  • HPC Script v0.0.13: Updated SLURM launch script with GPU multimodal params, Boruta-SHAP selection, and top_percentile.

Changed

  • Feature Selection: --top-percentile replaces --top-k for MI-based feature selection. Percentage-based selection adapts to varying feature set sizes across evaluators.
  • Boruta-SHAP MI Reducer: When Boruta-SHAP confirms >500 features, an MI-based reducer narrows the set to top_percentile (default 10%).
  • SLURM Hardening: KREVIEW_LABEL and KREVIEW_EVAL_GPU_SINGLE processes now set cache = 'lenient' and scratch = false to prevent institutional queue/cache interference on IRIS.
  • Dynamic Report Rendering: ROC CI reverse-map, DCA loop, and AUC deltas now dynamically discover models from DYNAMIC_MODELS instead of hardcoding LR/RF/XGB.

Fixed

  • ROC CI KeyError: Hardcoded {"Logistic Regression": "lr", ...} reverse-map in report_template.qmd replaced with dynamic lookup from DYNAMIC_MODELS, preventing crashes when GPU models are present.
  • DCA Loop: DCA now renders curves for all trained models (was hardcoded to RF/XGB only).
  • AUC Deltas: Pairwise AUC delta display now discovers all auc_delta_* keys dynamically.
  • Monolithic GPU Fitted Capture: gpu_res, gpu_fitted = gpu_models(...) now captures fitted models (was _ = ...).

Dependencies

  • tabpfn >= 8.0.3 (was >= 2.0)
  • tabicl >= 2.1 (was >= 0.0.4)
  • Added shapiq for GPU model SHAP computation.

[0.0.12] - 2026-05-24

Added

  • Nextflow publishDir: All 8 multistage modules now publish outputs to params.outdir with mode: copy. Output structure: labels/, matrices/raw/, matrices/selected/, models/cpu/, models/gpu/, matrices/fused/, models/multimodal/, reports/.
  • Label-first DAG: KREVIEW_LABEL runs once as the first step in multistage mode, producing labels.parquet shared across all extract jobs. Eliminates 26× redundant re-labeling (~13 min saved).
  • --labels CLI flag: kreview extract now accepts --labels /path/to/labels.parquet to skip internal labeling and use pre-computed labels.
  • Boruta-SHAP in HPC script: scripts/run_hpc_v0.0.11.sh now passes --multimodal_selection boruta_shap for rigorous multimodal feature selection.

Changed

  • Report DAG dependency: Report process now receives both matrices AND *_model_results.json from CPU/GPU eval, enabling complete dashboard rendering (ROC, SHAP, metrics). Report runs in parallel with FUSE + MULTIMODAL.
  • Extract module: extract.nf now accepts labels.parquet as input from upstream KREVIEW_LABEL.

Documentation

  • README: Comprehensive overhaul — added mRMR/Boruta-SHAP, GPU models, multimodal stacking, Nextflow HPC, pipeline architecture Mermaid diagram, fixed stale --workers flag.
  • Statistical Tests: Detailed mRMR documentation (optimization objective, F-statistic/Pearson explanation, algorithm steps), Boruta-SHAP documentation (shadow features, SHAP importance, flowchart, configuration table, when-to-choose guide), fixed stale Selection QC keys.
  • Nextflow Operations: Updated module list to match actual filenames, added iris profile docs, added publishDir output structure tree, added parq-cli tip.
  • Pipeline CLI: Fixed stale --workers 4, added --labels to extract example, added --ch-hotspot-maf, added label-first modular pipeline flow, added parq-cli tip.
  • Pipeline Architecture: Updated DAG (Label-first, Report parallel with Multimodal), fixed process table with actual _single.nf module names, added publishDir column.

[0.0.11] - 2026-05-24

Changed

  • Feature Selection (Default): Default single-evaluator strategy changed from hybrid_union to mrmr (Minimum Redundancy Maximum Relevance). mRMR selects features maximizing target correlation while minimizing inter-feature redundancy, preventing multi-collinearity.
  • Multimodal Selection: --multimodal-selection now supports boruta_shap (interaction-aware, SHAP-based) in addition to mi (mutual information, default).
  • CLI Log Output: kreview select and kreview run now display method-aware selection summaries (mRMR shows variance-dropped count; hybrid_union shows AUC∩MI overlap breakdown).

Added

  • mrmr-selection dependency: Added mrmr-selection>=0.2.8 to pyproject.toml.
  • BorutaShap dependency: Added BorutaShap>=1.0.17 to pyproject.toml.
  • Structured startup logging: kreview select and kreview eval multimodal now emit log.info("select_start", ...) and log.info("eval_multimodal_start", ...) at startup for machine-parseable audit trails.
  • Multimodal selection method persistence: multimodal_results.json now includes raw_features_selection_method ("mi" or "boruta_shap").
  • Selection Strategy value box: Multimodal dashboard executive summary displays the cross-evaluator selection strategy.
  • Scoreboard selection_method column: Now visible in both single-evaluator and multimodal dashboard scoreboard tables.
  • mRMR scatter disclaimer: When strategy="mrmr", the Feature Selection scatter plot displays a callout explaining that AUC/MI axes are observational (mRMR uses F-statistic + Pearson correlation internally).
  • mRMR scatter coloring: Selected features labeled "Selected (mRMR)" with cyan color, distinct from hybrid_union's 4-way AUC/MI overlap coloring.

Fixed

  • KeyError crash in cli_select.py: Fixed crash when --strategy mrmr was used — the log output tried to access hybrid-union-only QC keys (n_overlap_both, n_auc_only, n_mi_only).
  • Misleading cli.py log output: kreview run --strategy mrmr now shows mRMR-specific summary instead of zeroed-out hybrid_union counts.
  • Stale docstrings: Updated module docstrings in selection.py, cli_select.py, cli_eval.py, and eval_engine.py to reflect mRMR as default.

[0.0.10] - 2026-05-22

Added

  • Modular CLI: Extracted run into atomic sub-commands (label, extract, fuse, eval, select, report).
  • Nextflow DAG: Complete rebuild of monolithic pipeline into a multi-stage, per-evaluator parallelized workflow.
  • GPU Evaluation: Native support for PyTorch-based GPU models (TabPFN, TabICL) via the kreview eval --gpu-models target.
  • Multimodal Models: Cross-evaluator stacking capability allowing evaluation of multiple fragmentomics assays simultaneously.
  • Testing Optimizaton: Dropped unit testing execution time from >6min to <30s via module-scoped caching.
  • Docker Containers: Stabilized multi-target release action generating discrete CPU and GPU containers cleanly.

[0.0.9] - 2026-05-07

Changed

  • Feature Selection: Replaced Cohen's D --top-n selection with hybrid union: top X% by Univariate AUC ∪ top X% by Mutual Information. This captures both linear and non-linear predictors.
  • CLI: --top-n deprecated (prints warning, ignored). Use --top-percentile (default: 10%) instead. --compute-univariate-auc now defaults to True.
  • Volcano Plot: X-axis changed from Cohen's D to Univariate AUC for better alignment with model-based selection.
  • Top-20 Bar Chart: Ranked by Univariate AUC instead of Cohen's D.
  • Statistical Ledger: Sort priority changed to univariate_auc > mutual_info > cohens_d > kw_statistic.
  • Feature Cards: Display AUC + MI scores instead of Cohen's D.

Added

  • mutual_info_score(): New function in eval_engine.py using sklearn.feature_selection.mutual_info_classif(k=3) for non-linear feature relevance scoring.
  • selection_qc metadata: Saved in model_results.json with method, overlap stats, and feature counts for audit trail.
  • Feature Selection QC Scatter: New dashboard plot showing AUC vs MI with 4-color category coding (Both / AUC-only / MI-only / Dropped).
  • Scoreboard columns: selection_method, n_selected_features, selection_overlap_pct.
  • Structured logging: feature_scoring_complete, feature_selection_complete, univariate_auc_disabled, variance_guard_dropped, model_skip_insufficient_data.
  • Resume checkpoint: Warns on legacy JSONs missing selection_qc (pre-v0.0.9 Cohen's D runs).
  • Tests: 13 new tests for univariate_auc (6) and mutual_info_score (7) covering bounds, NaN handling, constant features, signal detection, and reproducibility.
  • univariate_auc logging: log.warning("univariate_auc_failed") replaces bare silent except.

Fixed

  • Duplicate label mask (GAP-1): Removed duplicate 4-tier label mask construction; reuse scoring target for model target.
  • Redundant import (GAP-3): Removed import structlog as _slog inside evaluator loop; use module-level log.
  • Silent degradation (GAP-4): Explicit warning when --no-compute-univariate-auc degrades to MI-only selection.
  • Double imputation (GAP-6): Cached _impute() result in variance guard to avoid re-computation.
  • Silent model skip (GAP-7): Added echo + structlog warning when model eligibility fails (< 20 samples or single class).

[0.0.8] - 2026-05-06

Fixed

  • AUC Consistency (D-01): ROC plots and KDE density now use out-of-fold predictions matching the official AUC value boxes. Eliminates the misleading AUC discrepancy (e.g., 0.845 vs 0.72) caused by mixing training/subsample models.
  • Dashboard Crashes (D-02/D-03): All cohens_d_true_vs_healthy references guarded with column existence checks. Statistical Ledger falls back to kw_statistic → univariate_auc when Cohen's D is unavailable.
  • Classification Labels (C-02): Confusion matrix and classification report now correctly show "ctDNA Negative" instead of "Healthy Normal".
  • Fold Variability (C-01): Now displays all three models (LR, RF, XGBoost) with per-model std in title instead of hardcoded RF-only.
  • Duplicate Imports (C-03): Removed redundant plotly import in report template.
  • Volcano Hover (GAP-1): NaN-safe hover text shows "N/A" instead of "nan" for zero-variance features.
  • JSON Bloat (GAP-6): OOF probability arrays rounded to 6 decimal places (~40% size reduction per evaluator).
  • HPC Memory (OOM): Base memory bumped 256→512GB for genome-wide evaluators (FSD, WPS).

Added

  • Structured Logging: structlog-based logging in CLI report() with per-evaluator timing, success/fail counts, and summary statistics.
  • Smart Resume (GAP-4): Resume checkpoint validates JSON freshness — warns if OOF keys are missing from pre-hardening runs.
  • SHAP Waterfall Guards (S-04): Validates array lengths before plotting to prevent IndexError.
  • SHAP Docstrings: All three SHAP helper functions now have full docstrings.
  • Variable Initialization: oof_y, oof_preds, and kde_labels initialized at template top level to prevent NameError when models fail.

Removed

  • plot_threshold_sensitivity() — dead code replaced by pre-computed threshold sweep in v0.0.7.
  • Redundant cross_val_score import and second independent CV computation.
  • Stale plot_threshold_sensitivity entry from _modidx.py.

[0.0.7] - 2026-04-13

Added

  • Dashboard Redesign: Restructured from single-page to 6-page progressive disclosure hierarchy (Executive Summary, Model Validation, Feature Explanation, Biomarker Yield, Cohort & QC, Data Explorer).
  • XGBoost Integration: Full XGBoost model evaluation alongside LR and RF in single_feature_model().
  • Bootstrap AUC CIs: 95% confidence intervals for all model AUCs using scipy.stats.bootstrap.
  • Decision Curve Analysis (DCA): Standalone decision_curve_analysis() function computing net clinical benefit across thresholds for RF and XGBoost models.
  • Precision-Recall Curves: PR curves and average precision for all three models (LR, RF, XGBoost).
  • Fold-Level AUC Tracking: Per-fold AUC tracking via cross_val_score for all three models with *_auc_std stability metric.
  • Threshold Sensitivity Sweep: 50-point sweep of sensitivity, specificity, and PPV across thresholds (0.01–0.99) for RF.
  • Feature Stability: CV cross-fold feature importance consistency scoring (0.0–1.0 scale).
  • QC Metrics: Per-feature n_missing, pct_missing, and is_zero_variance computed in evaluate_feature().
  • Feature Cards: Auto-generated metadata cards from evaluator registry with tier, category, and derived feature type detection.
  • SHAP Tabbed Interface: Unified SHAP tabs for RF and XGBoost with consistent layout.
  • Verdict Value Box: AUC-threshold verdict (Strong ≥0.80, Moderate ≥0.70, Weak <0.70) in Executive Summary.
  • Label × Cancer Type Sunburst: Interactive sunburst visualization replacing static bar chart.
  • Per-Sample Coverage Histogram: Histogram of non-null feature percentage per sample.
  • Per-Chromosome Feature Chart: Conditional chromosome-level feature bar chart.
  • Feature Importances Tab: Top-20 RF Gini importance bar chart in Model Validation.
  • AUC Deltas: RF–LR and XGB–RF AUC deltas surfaced in Performance Metrics.
  • Training Time: Model training time displayed in Performance Metrics tab.
  • VAF Scatter: Re-added tumor burden independence scatter (top feature vs max_vaf, LOWESS trendline).
  • great_tables Integration: Column-grouped data explorer using great_tables with auto-detected feature family spanners.
  • Quarto Auto-Discovery: _find_quarto() probes PATH then 7 well-known install locations (Positron, Homebrew, system, conda).
  • Methods Link: Sidebar link to API documentation site.
  • Documentation: New dashboard guide, DCA methodology doc, feature cards API reference, comprehensive CLI flag docs.
  • Tests: 24 new tests (53 total) covering DCA, PR curves, fold AUC, threshold sweep, feature stability, QC metrics, training time, AUC deltas, and feature cards.

Changed

  • JSON schema expanded from 24 to 44 fields per feature set.
  • Dashboard template: 1,555 lines, 52 balanced code blocks.
  • Language discipline applied across docs and template (removed promotional/overclaiming language).
  • custom.scss: bumped font size to 0.95rem, added QC warning class, added @media print rules.

Fixed

  • Silent except: pass in calibration code replaced with log.warning().
  • Division by len(df) without zero-guard in class balance text.
  • Scoreboard NaN values: sensitivity, specificity, n_samples, n_positive now correctly extracted from rf_classification_report dict instead of non-existent top-level keys.
  • QC tab Figure code dump: unsuppressed fig.update_traces() return displayed as raw Figure({...}) text.
  • Dashboard debug: true leaked Python output into rendered HTML; replaced with warning: false.
  • QC row heights summed to 135%; rebalanced to 100% (10+40+25+25).
  • Undefined _to_float_array in wps_panel.py and wps_genomewide.py replaced with correctly imported parse_array.
  • SHAP waterfall rendering stabilized for edge cases with zero feature contributions.
  • Typo fallsbacks → fallbacks in pyproject.toml.
  • great_tables>=0.15.0 added to pyproject.toml dependencies.

[0.0.6] - 2026-04-10

Fixed

  • Bin-Level Extractor Coverage Filtering: Fixed critical data bug where FSC and FSR on-target extractors computed median() across all 28,823 genomic bins, including the 98% with zero coverage. The zero-coverage sentinel values (-29.897 for log2, 0.0 for ratios) dominated the median, producing constant features and AUC=0.500 for all models. Now filters to total > 0 bins before aggregation.
  • SHAP Feature Shape Mismatch: Fixed TreeExplainer crashes caused by passing a feature-subset matrix instead of the full model-trained feature set. SHAP now computes on all 50 model features; --shap-features only limits visualization display count.
  • SHAP Binary Output Handling: Added robust handling for all TreeExplainer return shapes (list, 3D array, 2D array) and safe expected_value extraction for binary classifiers.
  • Error Visibility: Changed Quarto stderr capture from first 500 chars to last 1500 chars, surfacing actual Python errors instead of kernel boot messages.

Added

  • Nextflow HPC Parity: Wired all new CLI flags (--top-n, --shap-samples, --shap-features, --resume, --skip-report, --cvd-safe) into Nextflow nextflow.config and run.nf with HPC-optimized SLURM defaults (10-fold CV, 5000 SHAP samples, 64GB RAM).
  • CLI Configuration Logging: All three commands (label, run, report) now log their full parameter state at startup for reproducibility.
  • Resume Support: --resume flag skips evaluators with existing model results, enabling incremental re-runs.
  • Variance Guard: CLI automatically drops zero-variance features before model training with a logged warning, preventing wasted compute on constant columns.
  • Coverage Monitoring: All bin-level extractors now emit n_covered_bins and n_total_bins in the feature matrix for downstream QC.

Changed

  • Memory Management: Aggressive dashboard optimization — subsampled KDE plots (2000 samples), gc.collect() between SHAP runs, explicit dataframe deletion after filtering.
  • SHAP Explainer: Switched from generic shap.Explainer to shap.TreeExplainer to prevent OOM-causing KernelExplainer fallback.
  • HPC Resource Allocation: process_high memory increased from 32GB to 64GB for SHAP-heavy report generation.

[0.0.5] - 2026-04-09

Fixed

  • PyPI Release Wheel Artifact Synchorization: Enforced a permanent "Triple Bump" protocol standardizing manual version matching globally across settings.ini, nextflow.config, and kreview/__init__.py natively. This resolves a decoupling bug that structurally blocked Github Actions python -m build commands from publishing Wheels accurately tracking downstream package footprints correctly.

[0.0.4] - 2026-04-09

Fixed

  • Documentation Sync: Hardcoded the mike version provider default target alias gracefully to latest intrinsically allowing GitHub Action GH-Pages orchestration arrays to map tags seamlessly.
  • Git Flow Constraints: Implemented official mandatory 8-step PR-driven workflow structures strictly bypassing raw fast-forward terminal merges blocking CI.

[0.0.3] - 2026-04-09

Added

  • Global Config Parity: Synchronized dynamic execution limits tracking across native Python modules (v0.0.3) and HPC workflow blocks (nextflow.config 0.0.3).
  • Core CLI Flags: Engineered @app.callback() directly into kreview Typer infrastructure enabling standalone --version invocation blocks natively supporting Nextflow software_versions.yaml registry capture logging.

Fixed

  • Systemic Module Sync: Repaired native __init__.py decoupling bugs where sequential tool paths caused legacy 0.0.1 footprints to bypass metadata injections.
  • Git Asset Purge: Permanently bound dynamic CLI Quarto runtimes (html, static_plots, nohup) directly inside .gitignore blocking blob registry accumulation remotely.
  • CI Formatting: Established uniform black . compliance thresholds natively locking both nbs/ JSON source notebooks and /kreview Python exports securely for Github Actions lint workflows.

[0.0.2] - 2026-04-09

Added

  • Diagnostic Upgrades: Implemented explicit Sensitivity and Specificity clinical evaluations automatically calculating Youden's J static cutoffs and logging metrics securely to the results JSON matrix.
  • Model Expansion: Added scalable XGBoost modeling alongside native Random Forest and Logistic Regression classifiers inline.
  • Theme Selection: Introduced set_theme() support parsing --cvd-safe toggle parameters. Allows zero-friction toggling between Okabe-Ito (colorblind secure) and global Muted Neon visualization workflows.
  • Cluster Deployment Optimization: Injected [all] pip extra support resolving seamless multi-tool HPC dependencies smoothly (pip install -e ".[all]").

Fixed

  • Dashboard Standalone Render: Resolved format: dashboard clipping anomalies resulting from undocumented Quarto _quarto.yml project context bleeding. Hardcoded formatting constraints dynamically allowing pipeline usage deeply isolated in site-packages perfectly.
  • Path Resolution Hacks: Removed namespace package AST resolution assumptions and substituted secure __file__ contextual path checks, completely resolving editable pip install -e . template NoneType bugs natively.
  • CLI Options: Restored destroyed --verbose option parsing.

[0.0.1] - 2026-04-09

Added

  • Formally released the production-grade kreview Evaluation Framework for fragmentomics features.
  • Five-tier classification algorithm implemented for accurate clinical ctDNA labeling: Possible ctDNA−, True ctDNA+, Possible ctDNA+, Healthy Normal, Malignancy (Heme).
  • Scikit-learn Pipeline injection to eliminate cross-fold scaling leakage during standard ML evaluate_feature.
  • DuckDB I/O optimizations strictly mapping internal batch chunks dynamically configured for Desktop (--chunk-size 50) and HPC SLURM networks (--chunk-size 500).
  • Integrated dynamic kaleido plotting backends for clinical PDF Quarto workflows.
  • Dedicated Nextflow NF-Core DSL2 environment scaling wrapper specifically designed to ingest manifest.txt parameters recursively without crushing symmetric work/ limits.
  • GitHub Container Registry release pipeline (ghcr.io/msk-access/kreview) fully established utilizing OCI registry labels.

Fixed

  • Out-of-fold probability caching logic strictly refactored away from simplistic aggregation bias.
  • Hardcoded papermill exceptions mapped into modular pipeline architecture.
  • Replaced ambiguous test metrics with explicitly verified Benjamini-Hochberg (FDR) corrections over Mann-Whitney metrics.