Joint instruction and few-shot optimization using grounded proposals and categorical TPE.

The implementation follows DSPy MIPROv2's three stages: metric-filtered demonstration bootstrapping, predictor-specific grounded instruction proposal, and seeded multivariate Bayesian search. Minibatch trials inform the surrogate, but only full validation evaluations can select the returned program.

:init_temperature is the pinned DSPy proposal-temperature control. :proposal_response_format optionally binds each proposal to Imp's strict one-instruction JSON schema (:off, :auto, or :required).

Program-aware proposals use structural program metadata by default. Source text is included only through the explicit program_grounding: :module_source or program_grounding: {:text, context} opt-in.

:proposer_fidelity defaults to Imp's :beam_native grounded proposer. Set it to :dspy_3_2_1 to use DSPy's pinned data- and tip-aware proposal flow. Program awareness requires explicit text grounding and performs DSPy's program-description and module-description calls before each candidate; few-shot awareness uses the same ordered demo arms searched by the optimizer. Both zero-shot and joint instruction/demonstration search are supported; unsupported proposer combinations fail before any LM call. Pinned dataset grounding renders the JSON-safe values in Imp.Example with DSPy/Python spelling. Use Jason.OrderedObject when nested JSON object order is semantically significant; unsupported values fail before proposer transport with their exact example path.

:search_fidelity separately controls parameter search. The startup-only :dspy_3_2_1_optuna_4_9_0_startup mode reproduces Optuna 4.9.0's NumPy RandomState startup sequence exactly and rejects configurations that would enter modeled TPE. :dspy_3_2_1_optuna_4_9_0 continues through Optuna's multivariate categorical TPE phase with the pinned split, Parzen kernels, candidate sampling, and independent NumPy RNG streams. The default categorical Parzen search is BEAM-native and produces its own deterministic sequence for a given seed.

An explicit compile-time seed: 0 is a real seed in the default BEAM-native mode. The pinned :dspy_3_2_1 proposer mode deliberately mirrors DSPy's Python-truthiness behavior and retains the constructor seed when the compile override is zero.

Checkpoints

Pass these to Imp.optimize/5. :max_trials limits the number of new objective trials executed by one invocation. :checkpoint_fn receives a JSON-safe checkpoint after setup and after each completed trial. Pass any emitted checkpoint back as :resume_state to continue without replaying setup or completed trials. Checkpoints are trial-atomic, so an interrupted in-flight trial is retried.