fit_full_models: read training data in a fixed order

[mmyrte]

Stack 6/6, on top of #63.

What

DuckDB returns the rows of trans_pred_data_v() in no fixed order. Learners such as ranger fit a different model on reordered rows even with a fixed seed. Full models, and the potentials predicted from them, therefore did not replay. On the PIE benchmark, two runs of the same seeded random forest gave potentials differing by up to 0.1.

fit_partial_models() already asked for ordered = TRUE when seeded; fit_full_models() now always does. The cost is one keyed sort of the training table.

Tests

test_integ_trans_models_t.R: a seeded classif.ranger fitted twice through fit_full_models() gives identical split values. On PIE, the random forest potentials of two runs are now bit-identical; they differed before.

🤖 Generated with Claude Code

https://claude.ai/code/session_01D1LLswXe3jLjGTFrBsxgPf

---

Generated by Claude Code

[mmyrte]

@claude At this point, I think the added complexity of keeping the argument is unnecessary; the cost of reordering is too small. Let's just drop the ordered argument (also for the partial learner), and make it the default by moving the ordering from the data.table::setkeyv to the SQL query.

[mmyrte]

Done in fac76ad. trans_pred_data.sql now ends with order by tr.id_coord, tr.id_period, and the setkeyv() in trans_pred_data_v() is gone, together with its ordered argument. The callers no longer pass an ordering argument:

• the partial and full model workers;
• get_pred_filter_score() / pred_filter_worker(), which lose ordered_pred_data.

A new check in test_db_evoland.R confirms that the rows come back sorted. The full suite passes with at_home = TRUE. I merged the change forward into #65 and #66 without conflicts; the suite passes on #66 as well (651 tests).

One caller outside this repo: evoland-experiments/2026-05-ssp-ch/050-covariate-selection.qmd still passes ordered_pred_data = TRUE. It won't break. It passes a Filter object rather than a string, so the argument lands in ... and is ignored. The line can be dropped whenever that script's evoland pin is bumped.

---

Generated by Claude Code