v4.2.1 - Wheel sync + worker bug fixes

v4.2.0 was tagged but the publish workflow was cancelled before any
PyPI upload because orca/engine.c (the wheel-build source) was stale
relative to the runtime-fallback engine.c at repo root. The wheels
would have shipped without board_encode_state_full and a few other
C entry points. v4.2.1 is the first 4.2 release that actually
publishes to PyPI.

Fixes

• orca/engine.c synced with root engine.c. Wheel-build now
  includes board_encode_state_full, mcts_tree_new, and the rest of
  the May 14 engine additions.
• _dr undefined in orca/data.py:950. Reference to a leftover
  variable name after a refactor; replaced with DISTANT_RANGE. Self-
  play workers no longer crash with NameError when distant-exploration
  triggers.
• PR smoke CI now also skips tests/test_c_mcts.py (stale tests
  against an older C engine ABI; cleanup tracked separately).

Everything from v4.2.0 below is included in v4.2.1.

v4.2.0 - Training Workflow & Observability

Training Workflow Polish

• Hardware profiles (--profile=mps-laptop|cuda-single|cuda-multi|cpu-only|colab-t4) - pick sensible defaults for batch size, workers, MCTS sims, and games-per-iter in one flag. Explicit CLI args still win.
• python -m orca init <name> - scaffolder that drops a templated project with config, train.sh, play.sh, plugins.py, and README. Lowers first-run friction from "read the wiki" to "run two commands."
• Atomic checkpoint and replay-buffer writes - torch.save and pickle.dump now go via .tmp then os.replace(). SIGKILL mid-write no longer corrupts the canonical file.
• Checkpoint metadata - every .pt carries a _hexbot_meta dict (schema_version, arch, iter, elo, git_sha, hexbot_version, timestamp). Loaders no longer need to infer architecture from filename.
• ETA + moving-average iteration timer - rolling 8-iter window prints projected completion at the start of every iteration.
• --auto-tuner-dry-run - preview AutoTuner decisions without applying them.
• Plateau detection wiring - the PLATEAU_* config values and --plateau-* CLI flags were stored but never read; they now actually trigger an MCTS sim boost when ELO stalls.

Observability

• TensorBoard writer (--tensorboard, opt-in) - logs loss/total, loss/policy, loss/value, elo/current, lr, time/iter_seconds, buffer/size, games/completed to runs/<id>/.
• Weights & Biases integration (--wandb, opt-in via pip install 'hexbot[wandb]') - same metric stream, same step indexing.
• Run manifest (runs/<id>/manifest.json) - CLI args, config snapshot, git sha, hostname, GPU info, hexbot/PyTorch versions, written at run start.
• Worker error log (runs/<id>/workers.log) - process pool failures now log full tracebacks with iteration, timestamp, and source site instead of being swallowed or inlined.

Community & Discoverability

• Featured Community Bots table in README, auto-regenerated from leaderboard.json by a scheduled GitHub Action (or push to leaderboard.json).
• Colab quickstart notebook (notebooks/colab_quickstart.ipynb) - one-click train + TensorBoard view on a free T4 GPU.
• PR smoke CI (.github/workflows/pr-smoke.yml) - tests + 2-iteration training + scaffolder check on every PR and push to main.

Cleanup

• orca/distributed.py - MultiGPUTrainer and RayTrainer stubs now emit a UserWarning and carry STUB labels in docstrings. They previously advertised DDP / Ray scaling but silently fell back to single-GPU OrcaTrainer.
• SealBot expert demo samples auto-enabled once _last_policy_loss < 3.0, the same threshold used for the soft MCTS target switch.

Stretch

• Optuna sweep adapter (python -m orca.sweep, requires pip install 'hexbot[sweep]') - hyperparameter sweep over lr, batch_size, mcts_sims, train_steps with final ELO as the objective.

Optional extras added

• hexbot[tensorboard], hexbot[wandb], hexbot[sweep]. hexbot[all] now pulls all three.

Backwards Compatible

All v4.1.4 API unchanged. New features are opt-in via flags; existing scripts run unchanged.