Split utils.py into a Pyodide-safe lite half and a slim CPython
wrapper, repoint env-load-path modules at the lite variant, and add
a proof-of-concept that boots predicators inside Pyodide and resets
each of the 18 envs in the dropdown.

Architecture:
- predicators/utils_lite.py is the bulk of the old utils.py minus
  imports of torch, scipy.stats.beta, imageio, and the
  pretrained-model SDKs. predicators/utils.py keeps those 11 helpers
  (DelayDistribution subclasses, save_video/save_images, the four
  beta_bernoulli_* functions, create_llm_by_name/create_vlm_by_name)
  and re-exports the lite module via `from utils_lite import *`, so
  CPython callers see no API change.
- structs.py imports utils_lite directly (not utils) to break the
  cycle that triggers when utils_lite is loaded first under Pyodide.
  torch and pretrained_model_interface are lazy imports inside the
  two make_*_process methods that need them (via a shared
  `_torch_and_delay` helper).
- 119 env-path modules (predicators/{envs, ground_truth_models,
  pybullet_helpers, perception, execution_monitoring}) flipped from
  `from predicators import utils` to `from predicators import
  utils_lite as utils` so they don't pull torch into Pyodide.
- envs/__init__.py and ground_truth_models/__init__.py call
  import_submodules(tolerate_import_errors=True) — catches plain
  ImportError too, not just ModuleNotFoundError — so optional-dep
  failures (torch / gym_sokoban / gymnasium-robotics) skip the
  submodule with a warning instead of aborting auto-discovery.
- base_env.py defers `from predicators.pretrained_model_interface
  import OpenAILLM` to inside the one method that uses it.
- setup.py: splits heavy ML/LLM deps into an `[ml]` extra; the base
  install is the env-runtime slim set that's installable under
  Pyodide via micropip. `[develop]` includes `[ml]` plus formatters.
  Strict version pins (numpy/matplotlib/pillow/pyyaml) are qualified
  with `sys_platform != 'emscripten'` so they don't conflict with
  Pyodide's pre-loaded packages.
- predicators/third_party/{__init__.py, fast_downward_translator/__init__.py}
  — empty package markers so `find_packages` actually ships the
  third_party tree in the built wheel (the env path actually imports
  `predicators.third_party.fast_downward_translator.translate`).
- Replace `assert not params` with `assert len(params) == 0` in 22
  ground-truth option policies; numpy ndarray params trigger
  "truth value of empty array is ambiguous" otherwise.

Browser POC (web/):
- web/app/{index.html, main.js} drive a Three.js renderer (via
  urdf-loader) over Pyodide+pybullet-wasm. JSON.stringify-quoted
  option names; mesh primitives with no mesh_url are skipped (they
  were rendering as 1m box stand-ins for in-memory vertex meshes
  like domino's top triangle).
- web/app/setup.py is the Pyodide-side bridge: per-env CFG
  overrides for envs that need legacy options, multi-link visual
  transforms, body/color diffs, begin_option/step_option protocol.
  begin_option wraps the NSRT grounding in try/except so a sampler
  raising KeyError or AssertionError returns {error: ...} instead
  of crashing across the Pyodide → JS boundary; step_option also
  catches the broad Exception after OptionExecutionFailure so a
  policy raising mid-rollout doesn't leave _active_option armed.
  Also guards `get_gt_options` with try/except in 3 call sites so
  envs without registered options (the domino fan/ramp/stairs
  variants) load cleanly with an empty option list.
- web/app/bundle.sh builds the predicators wheel, the gym 0.26 shim
  wheel, and packs predicators/envs/assets/ into a tarball
  (skipping the 32 MB tar rebuild when no asset is newer than the
  existing tarball). The pybullet WASM wheel is fetched from
  BasisResearch/pybullet-pyodide.
- All 18 envs in the dropdown reach `action_dim=...` when smoked
  headlessly in Chromium via web/app/browser_smoke.mjs:
  ants, balance, barrier, blocks, boil, circuit, coffee, cover,
  domino, domino_fan, domino_fan_ramp, domino_fan_ramp_stairs,
  fan, float, grow, laser, magic_bin, switch.

CI (.github/workflows/web.yml):
- One `import-check` job. Boots Pyodide in Node, installs the three
  wheels, asserts every env in web/app/index.html's dropdown
  registers as a non-abstract BaseEnv subclass. ~30 s wall. No
  Chromium, no env construction, no asset extraction.
- import_check.mjs wraps the Python import in try/except writing the
  traceback to stderr (which Pyodide's stderr handler captures), and
  Node-level unhandledRejection/uncaughtException handlers print
  the stack before exit — so a WASM abort surfaces a real message
  instead of just `pyodide.asm.js:8`.

CI (.github/workflows/predicators.yml):
- unit-tests / static-type-checking / lint steps install `[ml]` so
  they get the heavy deps slim utils.py imports at module top.
  yapf / isort / docformatter stay slim.

mypy.ini: add `web` to the top-level exclude so mypy doesn't bail
out on "Duplicate module named setup" (./setup.py vs
./web/app/setup.py).

.predicators_pylintrc: add `web` to ignore-paths — the browser
bridge monkey-patches pybullet, accesses env internals, and uses
in-function imports; those checks don't help the POC.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>