Version 0.5.0

[0.5.0] - 2026-09-10

Added

• Run metadata now compares the Nextflow version and runtime environment.
  A run's Nextflow version can change between two runs and silently explain a
  behaviour difference, but it was never surfaced. RunLoader now reads the
  run's data-lineage WorkflowRun record (via a new
  LineageStore.environmentForSession()) and populates RunSnapshot with the
  Nextflow version and build, the container engine, and whether Wave and
  Fusion were enabled. RunComparator.compareMetadata() adds these as metadata
  rows (version/engine/Wave/Fusion are meaningful changes; the build number is
  treated as context via OBVIOUS_METADATA), so they render in every report
  through the existing metadata table with no renderer-specific plumbing. Rows
  appear only when a run recorded the value, so runs without a .lineage/ store
  (lineage.enabled=true, Nextflow 25.04+) are not padded with blanks.
  Per-run plugin versions are deliberately not compared — Nextflow does not
  persist them in the lineage store, the history file, or the task cache — and
  the HTML metadata section now says so explicitly.

• The task layer now compares I/O counters and reports execution hardware.
  RunComparator.TASK_FIELDS gained the disk-I/O counters (read_bytes,
  write_bytes, syscr, syscw, vol_ctxt, inv_ctxt) and the
  execution-environment fields (cpu_model, hostname, native_id). These are
  already present in every TraceRecord Nextflow writes to the cache, so no new
  data source is needed — the report simply stopped throwing them away. All nine
  are added to OBVIOUS_TASK_FIELDS, so they surface for context (notably,
  cpu_model/hostname explain a performance regression the perf layer already
  flags — run B's task landing on a slower CPU) without flipping the "identical"
  verdict or tripping --fail-on-change unless --verbose is set.

Changed

• The Parameters note now explains where non-launch params go. The HTML
  report's params section only ever lists values from the launch command and
  -params-file; params left at their defaults or set inside nextflow.config
  / an activated profile were silently absent, which read as "unset". The note
  now states those are resolved config, not launch input, and links to the
  Configuration layer where they actually appear.

• --help now documents the full exit-code contract. DiffPlugin.dispatch()
  maps outcomes to four exit codes — 0 success, 1 runtime error, 2 usage
  error, 3 --fail-on-change on a difference — but usage() only mentioned
  3 (buried in the --fail-on-change entry). A CI author reading --help had
  no way to tell 2 ("I typed the command wrong") from 1 ("the diff itself
  failed"). A new Exit codes: block in usage() spells out all four; doc-only,
  no behaviour change.

• The process-wiring section now leads with a node-link diagram of the DAG,
  not just a table of changed edges. renderDag() already had the full union
  of process→process edges tagged UNCHANGED/ADDED/REMOVED
  (RunComparator builds it), but the report threw the unchanged edges away and
  listed only added/removed rows — so the reader never saw where in the
  topology a change sat. A new dagSvg() helper lays the union graph out with a
  lightweight longest-path (Kahn) layering — columns = topological depth — and
  emits a self-contained inline SVG: unchanged edges are neutral hairlines,
  added edges solid green, removed edges dashed red, and a process appearing in
  only one run gets a matching node outline. The layout is computed in Groovy so
  the SVG needs no JavaScript or external assets (preserving the report's
  no-network-assets guarantee), and the existing per-edge table is kept beneath
  it as the precise detail and large-graph fallback. On the rich-report demo the
  ALIGN→QC edge is now visibly rerouted through the newly inserted MARKDUP
  node. Only HtmlReportRenderer and its test changed; no DiffResult
  accessors or other renderers were touched.

Changed

• HTML report summary reorganised from a flat wall of boxes into a scannable
  hierarchy. The summary section previously rendered up to 14 identical,
  equal-weight statCards in a single auto-fit grid — the reader had to read
  every box one by one to find the answer, with no cue that the first five were
  a single task distribution and the rest were per-layer change counts.
  renderSummary() now leads with a headline number (total task-level
  differences, coloured by the identical/different verdict), collapses the four
  mutually-exclusive task buckets (Changed / Only in A / Only in B / Unchanged)
  into one stacked proportion bar with a counted legend — with Recomputed
  demoted to an annotation since it is a cross-cut of changed, not a fifth
  bucket — and groups the remaining diff-layer counts under labelled
  Failures and Changes by layer bands. New statGroup() and
  dispositionBar() helpers plus supporting CSS (.summary-headline,
  .card-group, .disp*) reuse the existing colour tokens; no counts,
  DiffResult accessors, or other renderers changed.

Fixed

• Report percentages are now locale-independent. Format.signedPct() and
  HtmlReportRenderer.fmt() formatted floating-point values with
  String.format('%.1f', …) / String.format('%.2f', …), which use the JVM's
  default Locale. Under a comma-decimal locale (e.g. pt_BR, de_DE) the
  report emitted values like +12,5% and 1,50 instead of +12.5% and
  1.50, corrupting the rendered percentages and breaking any downstream
  numeric parsing that expects . as the decimal separator. Both call sites
  now pass Locale.ROOT so output is stable regardless of the host locale.

• The lineage-derived DAG now reads the lineage/v1beta1 store Nextflow
  actually writes, instead of silently falling back to the symlink heuristic.
  LineageStore parsed a pre-v1beta1 flat record shape — discriminator
  type, with sessionId/name/input at the top level. Current Nextflow
  (25.04+) instead writes a lineage/v1beta1 envelope whose discriminator is
  kind and whose payload is nested under spec. Every record therefore failed
  the type != 'TaskRun' guard, edgesForSession() returned null, and
  DagComparator fell back to reconstructing edges from work-dir input symlinks
  — so the authoritative-provenance path this class exists to provide was dead
  against any real store, with nothing logged above debug level. LineageStore
  now reads the discriminator via kindOf() (kind, falling back to type)
  and the payload via specOf() (the spec map, falling back to the record
  itself), so both the current envelope and legacy flat stores reconstruct. The
  existing unit tests were green only because they encoded the same obsolete
  flat shape; a new fixture of real v1beta1 records captured from a
  rich-report run (src/test/resources/lineage/rich-report) now pins the
  end-to-end INDEX_REF→ALIGN→MARKDUP→QC→MULTIQC reconstruction, alongside
  direct v1beta1 envelope cases. The examples/rich-report report was
  regenerated so its DAG layer reflects the authoritative lineage.

Full Changelog: v0.4.0...v0.5.0