Hygiene: revive the format-on-edit hook, gitignore dist and lint caches

The PostToolUse hook still invoked black at the pre-restructure
venv path, silently no-opping since the ruff migration; it now
runs ruff format from the workspace venv.  Also corrects the
test-onnx group comment: torchwright_doom's onnxruntime-gpu moved
from its main dependencies to an opt-in test-onnx group.

Test Suite Status: Unknown
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Author: physicsrob

.claude/settings.json

@@ -9,7 +9,7 @@
         "hooks": [
           {
             "type": "command",
-            "command": "F=$(jq -r '.tool_input.file_path // empty'); case \"$F\" in *.py) /home/rob/workspace/torchwright/.venv/bin/black --quiet --fast \"$F\" 2>/dev/null ;; esac; true"
+            "command": "F=$(jq -r '.tool_input.file_path // empty'); case \"$F\" in *.py) /home/rob/workspace/torchdoom/.venv/bin/ruff format --quiet \"$F\" 2>/dev/null ;; esac; true"
           }
         ]
       }

.gitignore

@@ -3,6 +3,9 @@ __pycache__
 .venv
 *.egg-info
 .pytest_cache
+.mypy_cache
+.ruff_cache
+dist/
 *.onnx
 *.gif
 *.png

pyproject.toml

@@ -78,11 +78,10 @@ dev = [
 ]
 # CPU onnxruntime for the ONNX-runtime and HF-parity tests (the OnnxTokenModule
 # oracle in onnx_load.py, and everything that compares against it).  This is a
-# DEDICATED non-default group, deliberately NOT in `dev`: the umbrella's
-# `uv sync --all-packages` installs every member's dev group alongside
-# torchwright_doom's `onnxruntime-gpu` main dependency, and the CPU and GPU
-# builds collide on the same `onnxruntime` import path.  Keeping it out of `dev`
-# lets the local workspace use the GPU build while the Modal test image
+# DEDICATED non-default group, deliberately NOT in `dev`: torchwright_doom
+# carries `onnxruntime-gpu` (in its own opt-in `test-onnx` group), and the CPU
+# and GPU builds collide on the same `onnxruntime` import path.  Keeping it out
+# of `dev` lets the local workspace use the GPU build while the Modal test image
 # (modal_image.py) opts in explicitly via `--group test-onnx`.  Pinned in
 # lockstep with the gpu extra so the CPU
 # oracle the parity tests compare against matches the 1.26.0 runtime contract;