feat: make use_parallel_embedding reachable instead of deleting the code behind it

[isayev]

Resolves the decision left open by #132/#133: connect the feature rather than delete it.

The gap

Parallel conformer embedding existed as a constructor argument on the isomer engine with no route from Auto3DOptions. No main() or smiles2mols run could turn it on, so isomers/parallel_embed.py was reachable only from tests — which is why audit M53 listed it as ~138 lines of dead code. The factory already accepted all three parameters; only the config layer and the two call sites were missing.

Three fields, not one

Wiring only the boolean would have half-plumbed it. isomer_engine.py gates on use_parallel_embedding and len(smi_name_tuples) >= parallel_embedding_threshold, so the other two are read by the same code path and would have stayed at their constructor defaults — leaving the worker count and the batch-size gate untunable from anywhere.

field	default	purpose
use_parallel_embedding	False	opt in
parallel_workers	4	worker processes
parallel_embedding_threshold	10	stay serial below this many molecules — spawning processes for a handful costs more than it saves

They flow Auto3DOptions → CLIConfig → both IsomerEngineFactory.create sites. CLIConfig is not optional here: test_cliconfig_covers_all_auto3doptions_fields requires every user-facing Auto3DOptions field to be reachable from the CLI/YAML layer, which is also what makes these settable from a config file rather than Python only.

Default stays off. Enabling it changes a run's resource profile, and that should be the caller's choice rather than something they discover.

A self-review catch worth recording

I first declared the bounds as Field(default=4, ge=1) on the pydantic model. _check_bounds' own docstring warns that a second hand-maintained constraint set is exactly the drift it exists to prevent — it enforces Auto3D.config.FIELD_BOUNDS precisely so the two schemas cannot disagree. The Field constraints were removed and the single table extended instead.

Testing

The tests assert what the factory is called with, not what the dataclass holds. Auto3DOptions(use_parallel_embedding=True).use_parallel_embedding is True would pass with the plumbing still entirely missing — it tests the dataclass, not the wiring. That is the trap this codebase has hit repeatedly, so the test stubs IsomerEngineFactory.create, drives smiles2mols, and asserts the three values arrived.

Mutation-verified both ways:

• un-wiring the smiles2mols call site → the reachability test fails
• dropping the FIELD_BOUNDS entries → both bounds cases fail

Plus a test that the default is still serial, so this cannot silently become opt-out.

1297 passed, 9 skipped; ruff clean.

Consequence for the deletion list

M53's "test-only" claim for isomers/parallel_embed.py no longer holds, and the module is off that list — recorded in the plan doc. #132's diagnostics fix now protects a path a user can actually take.