Records, statistics, and algorithms

1. MultiStatistics.register(..., chapters=) can target a subset of chapters, so fitness min/max and a size statistic need not share one function (DEAP#720).
2. MultiStatistics.compile materializes the input once. A generator, map, or zip is no longer exhausted by the first chapter.
3. An empty Logbook with a header prints that header from stream() instead of the word “empty”.
4. Logbook.to_json / from_json round-trip entries, nested chapters, and the header. NumPy scalars become Python numbers.
5. duplicate_count is a variety statistic: population length minus distinct keys. Hashable keys scan in linear time; unhashable but sortable keys use an adjacent-run count after sort.
6. The four ea_* algorithms accept log_time (per-generation wall time), logger (instead of only print), and fronts (append a new ParetoFront of that generation's survivors, not the cumulative hall of fame).
7. HallOfFame.update inserts into an empty archive, no-ops at maxsize=0, and replaces a similar member when the new individual is strictly better.
8. ParetoFront.update skips an individual that has no fitness instead of raising AttributeError after the front already holds a member.
9. HallOfFame.remove raises IndexError on an out-of-range index instead of desynchronizing keys and items.
10. Logbook.pop normalizes a negative index before comparing it to the stream cursor.
11. Logbook.pop removes the chapter row that shares that generation, so pop(i) and del logbook[i] stay aligned.
12. Logbook.__delitem__ removes the chapter row that shares the same generation — including a later occurrence of a repeated gen and every index in a slice — not the same list index.
13. Logbook.stream and str pair chapter cells by gen. A generation recorded without a chapter no longer shifts later values onto the wrong row or IndexErrors once the stream cursor is past the shorter chapter.
14. History.update records every member of a batch. A single individual without history_index no longer orphans the rest.
15. GridArchive tessellates behavior descriptors into a MAP-Elites grid: add keeps the best individual per cell, random_elites samples parent copies, and stats reports coverage and qd_score. ea_map_elites drives evaluate → archive → var_or and logs archive metrics each generation. Fitness stays on ind.fitness; behavior measurement stays on the caller. Semantic projection → add → sel_team_archive on occupied cells is a documented caller recipe; team scoring stays on the caller.
16. var_or mates two clones of the only parent when the pool has a single individual, so \((1,\lambda)\) / \((1+\lambda)\) with cx_prob > 0 no longer raises ValueError on sample(..., 2). ea_mu_comma_lambda with survivors=1 can run past generation one.
17. GridArchive.add rejects a non-finite first weighted objective. NaN or infinity no longer replaces a finite elite or occupies an empty cell.
18. CvtArchive and UnstructuredArchive sit next to GridArchive. CVT keeps one elite per k-means / caller centroid; unstructured adds a point that is far enough from every member or replaces the nearest neighbor. cvt_centroids builds the CVT tessellation from a behavior sample. ea_map_elites accepts either archive.
19. step_islands runs one evaluate → vary → select step on each deme, then an optional migrate (typically mig_ring). Each deme has its own toolbox, so selection pressure can differ while the topology stays a ring. Migrants keep fitness when eval_keys match; distinct keys invalidate arrivals. Append-only columnar evaluation may full-rescore with interpret_tapes after vstack, or use tape_lookback / suffix_rescore for a legal dirty suffix — DEAP has neither.
20. HallOfFame.update and ParetoFront.update skip an individual whose fitness is missing, invalid, or non-finite. An unevaluated creator individual no longer occupies a slot. NaN no longer sorts to the front of keys as if it were the best member.
21. Logbook.pop pairs a row without gen positionally when the chapter is the same length as the parent. Deleting that row no longer leaves chapter values behind or blanks the remaining cells.
22. GridArchive rejects a range whose ends are not finite. (0, inf) no longer maps every descriptor to cell 0, and (-inf, high) no longer crashes descriptor_to_index with int(nan).
23. ea_generate_update_restarts keeps the last evaluated population when generate returns empty. The empty batch is still a stop signal; it no longer overwrites a finished run with [].
24. Logbook.clear deletes every parent row through __delitem__. Chapters and the stream cursor stay aligned, matching del logbook[:]. list.clear no longer leaves chapter generations behind or drops later stream rows.
25. ea_map_elites skips stats.compile when the seed or offspring list is empty. A pre-filled archive with an empty initial, or generation zero with no individuals, no longer raises ValueError from max / numpy.max on an empty reduction. Archive metrics still record. The same guard is in record_generation for the other ea_* drivers.
26. var_or rejects each of cx_prob and mut_prob outside [0, 1]. A negative component whose sum still sits in [0, 1], or a NaN, no longer produces offspring.
27. ea_generate_update keeps the last evaluated population when generate returns empty. The empty batch is still a stop signal and does not call update([]).
28. n_evals= is an optional evaluation-budget stop on ea_simple, ea_mu_plus_lambda, ea_mu_comma_lambda, ea_map_elites, and ea_policy. Generations stay the default. EvalCache wraps evaluate / evaluate_batch by expression text (or a caller key) plus matrix identity and row count. A hit does not call evaluate again; n_evals / nevals still count the fitness assignment. promote_subtree and tune_ephemerals drop matching fitness-cache keys when they invalidate the compile cache. DEAP's ea_* drivers stop on generations only.
29. apply_policy_action maps a discrete policy token onto existing toolbox callables only: next_lexicase_cases, tune_ephemerals, promote_subtree, evaluate_invalid, interpret_tapes, and step_islands. Skip tokens are intentional no-ops; unknown tokens and missing required kwargs are rejected without raising. Fitness assignment and rescore ownership stay on the caller — not a second ea_* driver. DEAP has no policy action schema.
30. HallOfFame.to_json / from_json round-trip maxsize and archive members as genes plus fitness values. Checkpoint(..., hof_ind_cls=) stores hof as JSON instead of dill and rebuilds it on load. DEAP has no text serialization for the hall of fame. See Using checkpoints.
31. HallOfFame.insert and ParetoFront.insert skip an individual whose fitness is missing, invalid, or non-finite, matching update. from_json no longer restores a NaN member.
32. apply_policy_action charges PolicyActionGuard with the pre-dispatch evaluation estimate for step_islands. The post-step re-estimate no longer drops after invalids are already scored, so a tight n_evals cap cannot admit a second island step.
33. evaluate_invalid is public on algorithms / tools. It is the helper ea_* and apply_policy_action already use: score individuals whose fitness is invalid, prefer evaluate_batch when registered, otherwise map plus evaluate. DEAP inlines that scan in each ea* and does not export it.
34. ea_policy is the thin ea_simple loop plus one observe → decide → apply_policy_action step per generation. It calls PolicyActionGuard.begin_generation, rebuilds lexicase cases= when the policy asks, and records the action (and a generalization_gap chapter when exams are present). Fitness stays on the toolbox. Policy-action evaluations count toward n_evals and the generation nevals. When a policy step meets or exceeds that budget, the generation is recorded without variation so unevaluated offspring cannot replace the population. Not step_program_search: Slim, tune, archive, and team composition stay on the caller. DEAP has no policy driver.