fix(yolo-skill): don't require ultralytics on mps when framework_ok is False (backport to master)

[solderzzc]

Summary

Backport of #210 (already merged to develop) to master.

master is the branch the Aegis-AI app's skill installer actually clones from at install/reinstall time (skill-registry-service.cjs fetches skills.json off master and git sparse-checkouts the skill from there) — develop isn't in that path at all, so #210 alone doesn't reach any installed skill. This PR ports the identical fix to master so a skill reinstall actually picks it up.

• load_optimized()'s final fallback branch (taken when the optimized runtime check fails, framework_ok=False) unconditionally imported ultralytics — but mps installs deliberately never ship torch/ultralytics (requirements_mps.txt only lists onnxruntime).
• Any mps machine where _check_mps_runtime() fails (CoreML execution provider not found) crashed with No module named 'ultralytics' when loading a YOLO skill, instead of falling back to the pre-built .onnx model via plain CPUExecutionProvider.
• Reported in SharpAI/DeepCamera#207 — the reporter's exact traceback (Failed to load model: No module named 'ultralytics', backend=mps, framework_ok=False) reproduces this path exactly.

Fix

On the mps backend, that fallback branch now calls _load_onnx_coreml() (already used elsewhere in the same function for the mps happy path) against the pre-built .onnx, instead of assuming ultralytics is importable. If no pre-built .onnx exists either, it raises a clear RuntimeError instead of a confusing ModuleNotFoundError.

Applied identically to both copies of env_config.py (skills/detection/yolo-detection-2026/scripts/ and skills/lib/), matching how #194 touched both.

Test plan

• skills/lib/test_env_config_mps_fallback.py (3 tests) re-verified against current master — passes.
• Ran full skills/lib/test_env_config_*.py suite on master — 4 pre-existing failures in test_env_config_ane.py reproduce identically on master without this change (unrelated to this fix, not introduced by it; verified by temporarily reverting just the two env_config.py files back to master's version and re-running).
• Not yet verified against Brian's actual machine — this PR unblocks him reinstalling the yolo-detection-2026 skill from the Aegis-AI Skills sidebar to pick it up.

🤖 Generated with Claude Code

[solderzzc]

Closing — this repo's branch strategy only allows develop→master PRs (per CI's check-target job). Reopening the same fix as a develop→master PR instead.