[Low/Nit] Repo-wide review: minor findings tracking issue

[gratus907]

Severity: Low / Nit (tracking issue — bundles minor findings from a full repo-wide review)

These are lower-severity items from a repo-wide multi-vector review (see the Critical/High/Medium findings filed as separate issues). Grouped here since each is small in isolation; split out into its own issue before starting work on any one of them.

Spaces

• TupleSpace/RecordSpace use identity equality while every sibling space (RealSpace, IntegerSpace, CategoricalSpace, PermutationSpace, ArraySpace) is a frozen dataclass with value equality/hash — TupleSpace(RealSpace(0,1)) == TupleSpace(RealSpace(0,1)) is False. Degrades Problem equality for composite-space problems to identity. (spaces/composites/tuple_space.py:22-55, record_space.py:23-65)
• RealSpace.__post_init__ normalizes bounds via bare float(), accepting invalid definitions like RealSpace("0", "1") and RealSpace(True, True). (spaces/scalar.py:59-88)
• Compiled vs. generic geometry disagree on the permutation distance law — compiled PermutationSpaceGeometry scores 0/1 positional mismatch, generic fallback scores magnitude (|l-r|/(size-1))^2. A third-party space wrapping a PermutationSpace without a geometry provider gets materially different diversity distances. (spaces/geometry/compile.py:126-172 vs geometry/permutation.py:50-60)
• ArraySpace permits length=0 while TupleSpace/RecordSpace/PermutationSpace require non-empty declarations; a zero-length array poisons StructuredSpaceDiversityMetric.distance (raises on first pairwise distance). (spaces/composites/array_space.py:81-83)
• Duplicated/dead helpers: is_builtin_structured_space/is_builtin_child_space are byte-identical (geometry/leaf.py:174-229); require_real_candidate/require_integer_candidate/require_record_candidate exist in both composites/adapters.py and geometry/leaf.py with divergent strictness; StructuredSearchSpace.is_active_leaf_path has zero callers; CategoricalSpace.sample allocates a throwaway IntegerSpace per draw instead of using the already-imported random_state_randint.
• When a direction-free EvaluationProtocol is supplied, direction is stored but not meaningfully bound — downstream consumers read problem.direction as authoritative regardless, a silent dual source of truth. (problem.py:229-236)

Study / execution / evaluators

• Stale-async skips the max_evaluations < 0 guard the sync path and fast path both have — silently returns an empty report instead of raising. (study/stale_async.py:287-331)
• store_completion_group is duplicated verbatim across two modules; the drain-to-ordered-outcomes loop is triplicated across AsyncEvaluator.evaluate, evaluate_batch_exact_async, and StudyExactAsyncStepSession.finish; the records/refinements/budget/trace assimilation block is duplicated between execution.run and run_stale_async. (study/common.py:61-106 vs evaluators/async_evaluator/artifacts.py:194-235)
• Joblib infrastructure retry/suspend-resume silently re-executes every not-yet-consumed request on resume, including ones already completed in workers whose results were discarded — undocumented at-least-once semantics with no accounting; a wall-time/cost-metered objective double-bills. (evaluators/joblib/asynchronous.py:279-321,479-515)
• _classify_joblib_failure matches infrastructure failures by exception class name ("BrokenProcessPool", "TerminatedWorkerError"); a coincidentally-named user exception is misclassified as infrastructure and silently retried. (evaluators/joblib/asynchronous.py:56-72)

CSA

• Checkpoints silently drop trace_state on restore — a traced run resumed from checkpoint stops tracing with no error. (csa/engine/state.py:329 / optimizer.state_from_dict)
• ask() crashes on a structured seed whose active_leaf_paths() is empty (all-conditional branches inactive) instead of falling back to the plain operator path. (csa/engine/ask.py:119-130 with operators/editing.py:48-50)
• With biased potential enabled and a degenerate bank (all identical candidates → distance scale 0), _resolve_biased_sigma2 raises ValueError inside tell() instead of degrading gracefully. (csa/scoring/model_state.py:513-515)
• Dead code: module-level infer_score_gap is imported nowhere, shadowed by same-named keyword params, duplicates optimizer.infer_score_gap_for_entries (csa/banking/update/logic.py:577-595); the full-bank branch of admit_observation is unreachable from the engine and duplicates admit_full_bank_observation with divergent semantics (csa/banking/update/admission.py:64-119).
• newcomer_mask name is inverted — actually contains preserved-elite indices, masking them so newcomers seed first; misleading for future edits. (csa/engine/boundary.py:184-196)
• CSAPendingProposals.get is a linear scan over a tuple — O(p·m) tell-side validation, quadratic for large async batches. (csa/engine/state.py:79-83)

Artifacts

• RunReport/RunResult/NondominatedRunSurface/Trace direct constructors don't canonicalize sequence fields to tuples (records/observations/events, unlike refinements which is normalized) — externally mutable, unhashable "frozen" artifacts whose construction-time invariants can be invalidated after the fact. (artifacts/terminal.py:248,386-387,623-624,213)
• terminal_surface_setstate silently fills unknown/future fields with None on unpickling older payloads instead of failing loudly; unpickling then replace-ing with new refinements can raise a surprising TypeError (the _candidate_equal field is dropped by design on pickle). (artifacts/terminal.py:818-819)
• The shared require_json_* validators in json_types.py are used only by progression/*; every other codec hand-rolls isinstance checks with inconsistent semantics (e.g. accepting True as an int/float where the shared helper wouldn't). Already-drifted duplicated boundary validation.

Population algorithms

• ~350 lines of byte-identical generational scaffolding (ask/tell/_ask_initial_population/_materialize_generation/_generate_child/tournament selection/_sort_population) duplicated across ga, clearing_ga, species_ga, restricted_tournament_ga, and (partially) de optimizers — the species-GA bugs filed separately live exactly in the block that diverged. Consider extracting a shared base parametrized by member type + survival-policy hook.
• ScipyMinimizeResult.__post_init__ raises ValueError on any non-finite scipy result, aborting the whole proposal batch instead of degrading to the original evaluated proposal. (local_search/scipy/results.py:36-56)
• sample_exchange_count's documented fractional cap is never attainable due to an off-by-one exclusive randint bound (e.g. max_segment_fraction=0.5, size 10 → max segment length 4, not 5). (population/permutation/operators.py:95-99)

Packaging / docs / tests

• CHANGELOG's 0.1.0 "Added" list omits TupleSpace and PermutationSpace (both exported and documented as 0.1.0 features), and cites StructuredSearchSpace where the root facade exports SearchSpace — stability.md calls the changelog "the authoritative list" of supported surface. (CHANGELOG.md:19-22)
• tests/surface/test_packaging_metadata.py unconditionally import tomli, making the whole module a collection error (not a skip) on any env without the test extra, even on 3.11+ where stdlib tomllib would suffice; also uses a CWD-relative Path("pyproject.toml").
• No __version__ attribute on the package (only in build metadata); no PyPI classifiers/project.urls/authors in pyproject.toml.
• docs.yml only runs mkdocs build --strict — nothing ever deploys the site (no gh-pages/Pages job, no site_url). Build-only is safe but may not be intended if docs are meant to be published.
• The documented checkpoint save pattern (open("checkpoint.json", "w") + json.dump) overwrites the previous checkpoint in place — a crash mid-write leaves a truncated, unloadable file with the prior good checkpoint gone. Docs-only; recommend temp-file + os.replace.

Fix direction: triage into individual issues as capacity allows; several items above (species-GA scaffolding duplication, CSA aligned-bank-views abstraction) are natural follow-ups once their corresponding High-severity issues are fixed.