Refactor trans_models_t to use mlr3: replace fit_fun/gof_fun with Learner/AutoTuner interface

[Copilot]

Replaces the ad-hoc fit_fun/gof_fun function-passing interface with first-class mlr3 integration. Learner identity, hyperparameters, and a serialized untrained spec are stored natively in DuckDB; cross-validation uses mlr3 tasks and measures throughout.

Schema (trans_models_t)

Old	New	Type	Notes
model_family	learner_id	VARCHAR	mlr3 twoclass LearnerClassif key
model_params	learner_params	MAP(VARCHAR,VARCHAR)	Atomic scalar params only
fit_call	learner_spec	BLOB	Serialized untrained Learner
goodness_of_fit	crossval_score	MAP(VARCHAR,DOUBLE)	prediction$score(measures)
model_obj_part	crossval_predictions	BLOB	Serialized PredictionClassif
model_obj_full	learner_full	BLOB	Serialized trained Learner

Primary key: (id_run, id_trans, fit_call) → (id_run, id_trans, learner_id)

API

# New signatures — no backward compatibility
fit_partial_models(self, learner, measures, sample_frac = 0.7, seed = NULL, cluster = NULL)

# fit_full_models supports two mutually exclusive modes:
# 1. Score-select mode: pick best partial model per transition by score, retrain on full data
fit_full_models(self, select_score, select_maximize = TRUE, cluster = NULL)

# 2. Direct-learner mode: train a fresh learner clone directly on full data (no partial models needed)
fit_full_models(self, learner, cluster = NULL)

# measures accepts a character vector of mlr3 measure IDs (convenience) or a list of Measure objects
db$trans_models_t <- db$fit_partial_models(
  learner  = mlr3::lrn("classif.ranger", num.trees = 500, predict_type = "prob"),
  measures = c("classif.auc"),   # or list(mlr3::msr("classif.auc"))
  seed     = 42
)

# Score-select mode
db$trans_models_t <- db$fit_full_models(
  select_score    = "classif.auc",
  select_maximize = TRUE
)

# Direct-learner mode (crossval_score / crossval_predictions will be NULL)
db$trans_models_t <- db$fit_full_models(
  learner = mlr3::lrn("classif.ranger", predict_type = "prob")
)

Worker logic

• fit_partial_model_worker: Builds as_task_classif(..., positive = "TRUE"), deep-clones and trains the learner, scores held-out split via prediction$score(measures). For AutoTuner, extracts the optimal inner learner for learner_id/learner_params/learner_spec.
• fit_full_model_worker: Operates in two modes. In score-select mode, reconstructs from learner_spec BLOB (falls back to do.call(mlr3::lrn, ...) on deserialization failure, logging the reconstructed learner via fallback$format()), and passes crossval_score/crossval_predictions through from the partial model. In direct-learner mode, clones and trains the passed learner; crossval_score/crossval_predictions are NULL. A single fetch() retrieves all MAP/BLOB columns for the best rows, eliminating a redundant second fetch.
• predict_trans_pot: Deserializes learner_full and calls learner$predict_newdata(pred_data_post)$prob[, "TRUE"] directly — no intermediate copy needed, as mlr3 drops non-feature columns automatically.

New method

get_crossval_plots(id_run, id_trans) deserializes all crossval_predictions BLOBs and returns mlr3viz::autoplot() results for visual GoF inspection. Integration tests added (guarded by requireNamespace("mlr3viz")).

Dependencies

mlr3 moved to hard Imports; mlr3viz remains in Suggests.