Share detector postprocessing across nodes

[klangenk]

Motivation

Every detector node does the same three things with a model's raw output: drop low-confidence predictions, suppress overlapping boxes, and turn what survives into the loop's dataclasses. None of that depends on the model, yet each node repository grew its own copy — and the copies drifted. clip_point is byte-identical in three places across dfine_node and yolov5_node, while clip_box exists in those same three plus a fourth in yolov5_node's trainer that uses a centre-based convention under the same name. A fourth node would have copied it again.

This is the first of several changes moving the model-agnostic parts of node development into this library.

Implementation

• Add detector/geometry.py with clip_box (top-left anchored), clip_point, and clip_box_centered — the centre-based variant gets its own name so the two conventions can no longer be confused
• Add detector/postprocess.py with Detection, MIN_BOX_SIZE, bbox_iou, non_max_suppression, post_process and detections_from_xyxy, taken from the dfine_node implementation
• Unify the two detection containers: to_image_metadata (what a detector node reports) and to_detections (what a trainer's auto-detection pass reports) now share one _append_detections routine, so both paths clip and filter identically
• Add category_by_index / category_by_name, which raise ValueError with a helpful message instead of IndexError or a bare assert — a model/metadata mismatch is a real error, not a detection to skip quietly
• Make the scalar arguments of non_max_suppression and post_process keyword-only: transposing origin_h and origin_w was too easy
• Add learning_loop_node/tests/unit/, the first suite that runs without a Learning Loop, with 31 tests covering clipping, suppression, category resolution and trainer/detector parity
• Gate the credential-dependent CI jobs on the new unit job, which needs no secrets
• Document the new modules and the offline suite in AGENTS.md

Notes for the reviewer

• post_process keeps truncating box corners with int() so detector output stays bit-identical to what nodes reported before this change. detections_from_xyxy, which serves the trainer path, rounds instead — matching the code it replaces. The difference is deliberate and documented.
• clip_box clamps a corner only at the lower bound, so a box entirely outside the image keeps its origin and collapses to zero size. That is inherited behaviour, unreachable from post_process (which clips first), and it is captured by a test rather than changed here.
• Unifying the trainer path means auto-detections are now clipped and min-size filtered, which they were not before. This is a deliberate behaviour change on that path; the follow-up dfine_node PR is where it becomes observable.
• No node consumes this yet. The node PRs are separate and stay pinned to learning_loop_node==0.21.0 until this is released.

---

⬛ claude-opus-5[1m] · 265k tokens · $6.40

[klangenk]

Superseded by #94, which combines all three phases into one PR per repository. The commits are unchanged and still readable in order.