Upgrade: dep bumps, daemon transport client, model export, CRDT scene sync

[cartertbirchall]

Summary

Four parallel upgrades to the Auroch Syna platform.

• Dependency bumps across the ML stack (torch, diffusers, transformers, etc.)
• Daemon transport client — real WebSocket client so callers can talk to a remote daemon instead of loading models in-process
• Model export tool — ONNX, TorchScript, and CoreML export for DA-2 and LaMa
• CRDT scene sync — LWW registers + OR-Set for offline-first, conflict-free scene editing

1 — Dependency bumps (pyproject.toml)

Package	Before	After
torch	≥2.7.0	≥2.12.0
torchvision	≥0.22.0	≥0.27.0
diffusers	≥0.33.1	≥0.38.0
transformers	≥4.48.3	≥4.57.6 (keeping 4.x; 5.x has breaking API changes)
accelerate	≥1.1.1	≥1.13.0
peft	≥0.7.1	≥0.19.1
timm	≥1.0.15	≥1.0.27
scikit-image	≥0.24.0	≥0.26.0
einops	≥0.7.0	≥0.8.2
opencv-python	≥4.12.0	≥4.13.0
websockets	≥12.0	≥14.0

New optional extra [convert] (onnx, onnxruntime, coremltools).
Note: nunchaku wheel URL still pins torch 2.7 — update that URL separately when testing CUDA.

2 — Daemon transport client (daemon/client.py)

DaemonClient — async WebSocket client matching server.py's wire protocol.

• await client.command(kind, params) — sends command, awaits ack/error
• client.subscribe(kinds) — opts into event stream
• DaemonClient.run_command(...) — blocking one-shot wrapper
• DaemonClient.ping(endpoint) — health check

DaemonModelClient — drop-in for ModelClient that routes generate_scene
to a daemon process, keeping the orchestrator process torch-free.

3 — Model export tool (tools/convert_model.py)

python tools/convert_model.py --model da2  --format onnx        -o models/da2.onnx
python tools/convert_model.py --model da2  --format torchscript  -o models/da2.pt
python tools/convert_model.py --model da2  --format coreml       -o models/da2.mlpackage
python tools/convert_model.py --model lama --format onnx         -o models/lama.onnx
python tools/convert_model.py --model lama --format coreml       -o models/lama.mlpackage

Includes ONNX graph validation, dynamic batch axes, CoreML fp16 targeting macOS 14+.

4 — CRDT scene sync (scene/crdt.py)

Primitive	Purpose
LWWRegister[T]	Last-Write-Wins register with timestamp + author tiebreak
ORSet[T]	Observed-Remove Set; concurrent adds survive removes
SceneDelta	Compact JSON diff: added/removed/changed objects + scalar fields
diff(a, b)	Compute delta from snapshot a to b
merge(base, delta)	Apply delta to produce merged snapshot
merge_snapshots(a, b)	Three-way merge of two diverged snapshots

19 unit tests in tests/test_crdt.py — all pass, zero ML deps.

Test plan

• pytest tests/ — 39 tests pass on CPU-only machine (no GPU needed)
• auroch-syna info CLI starts and prints device info
• python tools/convert_model.py --model lama --format torchscript -o /tmp/lama.pt succeeds with [ml] installed
• DaemonClient.ping("ws://127.0.0.1:8765") returns True against a running daemon

🤖 Generated with Claude Code