feat(apps): add impact-reg-konfai registration orchestrator

[vboussot]

📚 Stacked on #18 (pr/seg). The diff is against pr/seg.

Summary

Adds impact-reg-konfai, a standalone CLI app bundle (apps/impact_reg/) that orchestrates IMPACT-Reg multimodal registration presets through the konfai-apps runtime. Each preset is a self-contained KonfAI app (default repo VBoussot/ImpactReg) whose model emits, on the fixed grid, the resampled moving image and a displacement field. This wrapper adds the registration-specific glue that does not fit the generic konfai-apps pipeline — multi-preset DVF ensembling, modality-aware evaluation, and ensemble-spread uncertainty — as three composable sub-commands so a UI or CLI can drive each step independently. It mirrors the impact-seg-konfai bundle from the base PR, giving registration the same thin-wrapper packaging.

What changed

New app package (apps/impact_reg/impact_reg_konfai/)

• impact_reg.py — ImpactRegKonfAIApp with three operations mirroring konfai-apps infer/eval/uncertainty:
  • register() runs each selected preset via a separate konfai-apps infer subprocess (isolating konfai's process-global Config singleton and KONFAI_* env), averages the per-preset DVFs through _ensemble(), and writes Moved.mha, DVF.mha, and Transform.h5 per case (P%03d). Masks are optional and restrict the metric region; a whole-image mask is auto-filled via _full_mask() so every preset receives its four declared inputs. keep_dvf persists per-preset fields under Ensemble/ for later uncertainty.
  • evaluate() scores any subset of modalities — image MAE (Evaluation_with_images.yml), segmentation Dice with nearest-neighbor resampling (Evaluation_with_seg.yml), and landmark TRE (Evaluation_with_fid.yml) — delegating to KonfAIApp.evaluate. Landmarks warp the fixed fiducials forward through the fixed→moving transform and compare against the moving fiducials (no field inversion).
  • uncertainty() stacks ≥2 ensemble DVFs into one multi-component volume (samples as vector components, real fixed-grid geometry preserved on the trailing axes) and runs the preset's generic Uncertainty.yml (Norm then StandardDeviation) via konfai-apps uncertainty.
  • get_available_presets() / _app_id() resolve presets from a local directory (folders whose app.json declares task == "registration") or the HF repo, overridable via KONFAI_IMPACTREG_REPO (default VBoussot/ImpactReg).
• cli.py — impact-reg-konfai argparse front end with register / eval / uncertainty sub-commands: a shared _add_device() (mutually exclusive --gpu/--cpu, -q/--quiet, --download, --force_update), a _default_preset() fallback with an actionable error when none resolve (used by eval/uncertainty), register taking positional presets to ensemble plus --fixed-mask/--moving-mask, --tta, and --uncertainty (keep per-preset DVFs), and an eval guard requiring at least one complete modality (image/seg/fid).
• Reuses only public API: read_landmarks, write_landmarks, apply_to_data_transform (konfai), KonfAIApp, get_available_apps_on_hf_repo (konfai_apps).

Packaging & docs

• pyproject.toml declares impact-reg-konfai with a dynamic setuptools-scm version and the impact-reg-konfai = impact_reg_konfai.cli:main entry point; setup.py pins konfai==<version> and konfai-apps==<version> to the same release.
• Adds LICENSE (Apache-2.0), README.md, Logo.png, and .gitignore; the orchestrator, CLI, and test modules carry the Apache-2.0 SPDX header.

Testing

• Adds tests/test_evaluate_landmark_direction.py, a regression test locking the landmark transform direction in ImpactRegKonfAIApp.evaluate. With a known translation T: fixed → moving and m_i = T(f_i), it monkeypatches KonfAIApp.__init__/evaluate and asserts the metric's reference is the moving fiducial set while the scored points are the fixed fiducials warped by T, so a perfect transform yields TRE ≈ 0 (not doubled). It uses synthetic data and does not exercise real inference.
• No changes to the core konfai package or konfai-apps; the PR only adds files under apps/impact_reg/.

Review notes

• Stacked PR: base is pr/seg. The diff is fully additive — 10 new files, no existing file modified — so there is no core or konfai-apps behavior change and no backward-compatibility risk. Review after the base merges.
• Runtime behavior depends on network/Hugging Face access to VBoussot/ImpactReg (or a local dir via KONFAI_IMPACTREG_REPO) and shells out to the konfai-apps CLI (subprocess.run(..., check=True), annotated # nosec B603). Per the app trust model, resolving a preset copies/imports its .py and installs its requirements.txt.
• The two version pins in setup.py assume konfai, konfai-apps, and this package are released in lockstep.