IDSLab 0.2.0: comparison and sweeps

Visual pipeline builder; cost objectives and stage cost; greedy_fixed selection; plugins; idslab compare with paired per-seed differences and idslab sweep; fixes for chained score dependencies and external-score freshness. Runs from 0.1.0 rerun.

Changes

Finished run units from 0.1.0 are not reused: the version is part of every unit's fingerprint, and the
fixes below change what a fingerprint covers.

Added:

• Visual pipeline builder in the dashboard: drag stages onto a canvas, wire outputs to stages or
  terminals, drag score edges for features, and edit every stage setting in a panel (models from a
  catalog with parameter forms, search spaces, pipeline steps, decisions, training rows). A dataset
  panel reads a sample of the data and suggests the label, type and categorical columns. The config is
  written with the loader's rules and validated on every change; node positions are saved next to the
  config (idslab.config.dump_yaml writes configs that read back unchanged).
• Cost objectives: protocol.objective: {cost: {missed_attack: A, false_alarm: B}} chooses on VAL by
  the lowest average cost per flow; {recall_at_fpr: X} by the highest recall within a false alarm
  cap. The objective appears as its own column in every report.
• Stage cost: each run records the share of TEST flows reaching every stage and its scoring time; a
  stage may declare cost:. Reports add the expected cost per flow of the cascade against its first
  stage alone.
• protocol.search: greedy_fixed: later stages with a fixed decision stay on while earlier stages are
  judged. On the synthetic R2 example it raised VAL and TEST accuracy in each of 3 seeds.
• Plugins through the entry point group idslab.plugins: models and pipeline steps for the builder
  catalog, and named objectives.
• idslab compare: experiments side by side, and for experiments on the same data and splits a
  paired comparison per unit on the same TEST flows, with paired bootstrap intervals. Writes
  compare.md, compare.json and compare.html.
• idslab sweep: one base config with named variants or a grid of dotted-key changes; every variant
  is checked first, written as its own config, run, reported and compared.
• Dashboard: Compare and Sweeps pages; compare and sweep run as dashboard jobs.

Fixed:

• Chained score dependencies (a stage that feeds another stage while using a third stage's scores)
  are refused at config load. Their out-of-fold scores were not fully out of fold: the inner stage's
  fold refits used features from models that had seen the held-out fold.
• External score sidecars (.json) are part of the unit fingerprint, so editing a sidecar reruns the
  unit and its provenance is checked again; before, the finished unit was skipped.
• Reports check each unit's external score files against the current ones and leave out units whose
  files changed, naming the reason; before, stale results from replaced files stayed in the report.
• Dashboard on Windows: Stop ends the job's worker processes too, and a path on another drive or with
  a colon is refused with a message instead of an error.

Documentation:

• A short README with an overview figure and a table of guides; INSTALL.md with step-by-step guides
  for Windows, macOS and GNU/Linux from the download to the dashboard; docs/architecture.md,
  docs/limitations.md, docs/compare-and-sweeps.md; mini READMEs for the example designs and the
  dashboard source. Figures are drawn by docs/img/make_figures.py from a real run.
• A project site (docs/index.html, served by GitHub Pages from docs/).
• IDSLab is not on PyPI yet; the guides install it from GitHub
  (pip install git+https://github.com/hashtagRR/IDSLab.git).

Install

This version is not on PyPI; install it from its tag:

pip install "git+https://github.com/hashtagRR/IDSLab.git@v0.2.0"

The current version installs with pip install idslab. Every change is in CHANGELOG.md.