feat(model-lane): v5.0 Phase 5 — build-lane breadth (model-scaffold 5 tasks + architecture-zoo detection/synthesis)

[Yoojin-nam]

Phase 5 of the v5.0 model-engineering lane — build-lane breadth. Expands the existing /model-scaffold and /architecture-zoo skills; no new skills / detectors / probes (counts unchanged — 51 skills / 41 detectors / 38 guidelines), torch-free CI. Lands in [Unreleased] for the next release / v5.0.0.

What's in this PR

• /model-scaffold now generates 5 task types (was segmentation-only): --task segmentation (U-Net), classification (small multi-label CNN; swap in a timm backbone), detection (torchvision Faster R-CNN + FPN), synthesis (Pix2Pix U-Net generator + PatchGAN), ssl (SimCLR encoder + projection head, NT-Xent). Refactored scaffold.py into a TASKS registry. The patient-level seed-locked split is task-independent, and each emitted train.py / evaluate.py is hygienic by construction (all RNGs seeded via a shared seed_everything, cuDNN deterministic, train-only loader, eval() + no_grad()). The challenge + regression test now verify all 5 tasks (frozen split + valid Python + check_training_hygiene, network-free; segmentation keeps the torch forward tier).
• /architecture-zoo adds detection.md + synthesis.md family cards — R-CNN family / Faster R-CNN+FPN / Mask R-CNN / RetinaNet / YOLO / DETR; Pix2Pix / CycleGAN / SPADE / diffusion / VAE / fastMRI — each grounded in its source paper with when-to-use, medical use, reference implementation, validation setup, and the matching scaffold template. The decision-tree index now routes to them (no more "forthcoming").

Notes

• Package size (Codex Add manage-refs pre-submission gate and align self-review reference QC #13): templates are small .py / .md text — no model weights; the npm tarball audit passes and size is unchanged in practice (~19 MB, dominated by pre-existing journal profiles / checklists).
• Honest contract preserved: runnability is verified by the AST hygiene gate + (segmentation) the self-skipping torch forward tier; CI stays torch-free.

Verification

All CI-mirror gates pass locally: validate_skills.sh, every gen_*.py --check, validate_catalog_consistency, check_domain_probe_sync --strict, frontmatter / routing-assets / locale / version / npm-audit, plus the updated multi-task challenge + regression test.

🤖 Generated with Claude Code