Query: categorical reaction-condition recommendation

Create a reusable Python CLI at skills/retrosynthesis_planning/scenarios/openai4s_codebases/05_condition_recommendation.py. It accepts --workspace PATH, reads only the installed public boundary, and atomically writes PATH/results/intermediate_results.json.

Installed public inputs

• PATH/installation.json: require scenario_id == "reaction_condition_uspto_categorical_v1".
• PATH/public/inputs.json: rows with reaction_id, reactants, and product; condition labels are hidden.
• PATH/public/vocabulary.json: allowed categorical values for exactly catalyst1, solvent1, solvent2, reagent1, and reagent2.
• PATH/public/model_outputs.json: each row has reaction_id, ranked predictions, and error; each prediction has rank, score, and one complete five-slot conditions object.
• PATH/public/config.json: integer top_k and random_seed.

Validate one output per reaction, complete tuples rather than a Cartesian mix of marginal labels, vocabulary membership, duplicate tuples, and Top-K. Use identity molecule canonicalization only for the bundled CC0 protocol fixture. Build the standard intermediate artifact with deterministic trajectory_sha256.

Never read PATH/private_evaluator/, hidden condition sets, gt_codebase.py, or scenarios/gt_codebases/; do not import GT code. Provide --help and fail nonzero on malformed input.