v0.5.0

Added

• Phase 39 adds the paper-faithful Frequency Consistency Metric (FCM) to
  image-evaluate and rendered-evaluate: grayscale 5x5 LoG (sigma=1,
  zero padding), normalized Frobenius error clamped to [0, 1], and the
  lower-is-better max_frequency_consistency gate for kind: image and
  kind: rendered. JSON metadata records the flat-reference policy (both
  flat = 0; candidate non-flat = 1), and reports/triage/PR comments surface
  the aggregate mean.

• ca uncertainty-evaluate and kind: uncertainty for position NEES (DoF=3),
  normalized position NEES, and chi-square coverage from GT plus validated 3x3 state
  covariance. NIS remains deferred because innovations are not available.

• Optional DreamSim perceptual distance for image-evaluate and
  rendered-evaluate, including max_dreamsim_distance config gates,
  triage, PR-comment deltas, and a [perceptual] dependency extra. The first
  real invocation may download official pretrained weights.

• Experimental reference-free ca plane-consistency evaluation and
  kind: structure CI checks. The deterministic
  plane_normal_dispersion / coplanar_offset_rmse metrics are
  PNE/CPV-inspired local-plane proxies, not claims of reproducing the
  published metrics; they surface in gates, triage, JSON, and PR comments.

• Benchmark report bundle golden path — ca benchmark eval --out <dir>
  now writes a reproducible report bundle for CI artifacts and static
  publishing: metrics.json, summary.md, report.html,
  manifest.lock.yaml, and provenance.json. The bundle records the
  suite, gate, input hashes, CloudAnalyzer version, and portable paths so
  a SLAM/LIO candidate map + trajectory can be reviewed without relying on
  machine-local state. The README now leads with this "dataset/reference +
  candidate outputs -> metrics/report/gate" workflow, and
  .github/workflows/slam-benchmark-smoke.yml dogfoods the clone-after-install
  smoke path.

• Static benchmark leaderboard builder — ca leaderboard build <bundle>... --out <site> turns one or more benchmark report bundles into
  a self-contained results.json + index.html site. Rows include dataset,
  sequence, method, gate result, CloudAnalyzer version, report links, artifact
  hashes, and metric values; incomparable rows are surfaced as warnings rather
  than silently mixed.

• Unified gate severity policy — config-driven ca check now shares
  a stable gate summary schema across artifact, trajectory, run,
  loop-closure, image, and rendered checks. Check specs can use
  severity: fail | warn | soft_fail | skip | not_applicable, and the CLI
  exposes --warn-only and --strict with consistent exit-code behavior for
  CI.

• SLAM driver conformance helper — plugin authors can call
  run_slam_driver_conformance(MyDriver) from ca.testing.conformance to
  validate metadata, deterministic output paths, trajectory/map shapes,
  stdout/stderr discipline, failure behavior, and writer compatibility before
  publishing a driver adapter.

• image config check kind (Phase 31) — wires ca image-evaluate
  into the config-driven QA layer so photometric quality can ride the
  same regression gate as every other artifact. A kind: image check
  takes rendered_dir + reference_dir, scores PSNR / SSIM per pair,
  and gates on the aggregate mean via gate: { min_psnr, min_ssim }
  (min_psnr in dB). Bit-identical pairs (PSNR = +∞) are excluded from
  the mean and trivially pass min_psnr; a gate fails when zero pairs
  match across the two directories. The kind emits Markdown/HTML reports
  (save_image_report), participates in ca's severity-weighted triage
  ranking, and surfaces PSNR / SSIM means with up/down deltas in
  ca report-pr-comment. This is the first non-geometric axis to reach
  the CI gate; LPIPS (an optional [photometric] extra, adding a
  max_lpips gate key) follows in a later phase. See docs/ci.md
  § Config-Driven QA.

• ca image-evaluate (Phase 30) — first photometric evaluation
  surface in CloudAnalyzer. Scores a set of rendered images against a
  reference (ground-truth) set on PSNR + SSIM, pairing by filename
  across two directories. Implementation is pure numpy +
  scipy.ndimage.gaussian_filter (no new heavy dep). Emits per-pair
  metrics plus a per-metric summary (mean / median / min / max). New
  module ca/core/image_evaluate.py carries the
  ImageEvalRequest / ImageEvalResult contract and the psnr /
  ssim helpers. This is the foundation for a future
  ca rendered-evaluate that would drive a 3DGS PLY into images at
  known camera poses and chain into this scoring path; see
  docs/commands/image-evaluate.md for the full reference.

• Canonical third-party SLAM driver plugin (Phase 29). The repo
  now ships
  plugins/cloudanalyzer-driver-example/,
  a complete pip-installable package that registers an Open3D
  scan-to-scan point-to-point ICP driver under the
  cloudanalyzer.slam_run_drivers entry-point group. After
  pip install -e plugins/cloudanalyzer-driver-example,
  ca slam-run --driver example works end-to-end. The plugin depends
  only on what CloudAnalyzer already pulls in (no kiss-icp / kiss-slam
  / small_gicp extras required) — so it doubles as a worked template
  for third-party authors. CI installs and exercises it via a new test
  (tests/test_slam_run_plugin_example.py), so the entry-point
  pathway introduced in Phase 28 is now validated against regressions
  by a real package install, not just by in-process monkey-patches.

• Plugin / entry-point registry for ca slam-run drivers (Phase 28).
  ca.core.slam_run now exposes register_driver(name, factory),
  get_driver(name), and list_drivers(). The three built-in drivers
  (kiss-icp, kiss-slam, small-gicp) register themselves at module
  import time via lazy factories. Third-party packages can publish their
  own driver under the cloudanalyzer.slam_run_drivers entry-point
  group and have it surface automatically as
  ca slam-run --driver <their-name> after pip install. The CLI's
  former hard-coded if/elif over driver names is gone — --driver
  now resolves through the registry. Broken third-party plugins are
  logged and skipped instead of crashing the CLI. See
  docs/commands/slam-run.md § Adding a third-party SLAM driver for
  the contract + a worked pyproject.toml example.

Changed

• Public demo/report artifacts now scrub machine-local absolute paths and
  escape embedded copy-command strings consistently, making checked-in HTML
  reports safer to publish and deterministic in CI.
• The synthetic-room 3DGS fixture check in the main Test workflow now uses
  --skip-renders, keeping CPU GitHub runners on the geometry fixture path
  while the rendered self-QA workflow continues to dogfood kind: rendered
  with checked-in reference PNGs.
• ca slam-run --driver help text now lists the three built-in drivers
  and points at the entry-point group; the validated choice set is
  computed dynamically from the registry rather than hard-coded.

Fixed

• Avoided a circular import on script entry points that imported
  ca.batch -> ca.info -> ca.core -> ca.core.checks -> ca.batch by loading
  get_info lazily inside batch_info.
• Fixed .github/workflows/leaderboard.yml parsing by replacing a YAML-hostile
  shell heredoc commit message with normal git commit -m ... -m ... flags.
• Fixed oriented 3D box IoU (ca detection-evaluate /
  ca tracking-evaluate) on NumPy 2.x, which rejects 2D inputs to
  np.cross; polygon clipping now uses an explicit 2D cross product.
• Capped open3d<0.20: open3d 0.20.0 requires libusb-1.0 at import time
  and ships stubs that reject str paths.