[evol assembly] Testing gaps: several exported functions lack test coverage

[dfalster]

Tracking issue for test coverage gaps, following the test suite added in #25 and extended in #26.

Current state: devtools::test() is green (79 pass, 0 fail), covering the core community + fitness primitives. Several exported functions remain untested.

Tested

• Object: community_start, trait_matrix, length.community (+ internal community_add / community_drop)
• Bounds: bounds, bounds_infinite, check_bounds, check_point
• Controls: demographic_step_control, plant_default_assembly_control (via test setup)
• Demography / fitness: community_demography (empty + single resident), community_selection_gradient, community_viable_fitness_1D, max_fitness, max_growth_rate

Untested — should work today (highest value)

• community_solve_singularity_1D — 1D evolutionary attractor (headline capability; self-contained, expected to pass)
• assembler_start / assembler_run / assembler_control — the full assembly loop (stochastic and maximum-fitness births)
• community_fitness_landscape — grid method (GP/surrogate method may be separate)
• tidy_assembly — output tidying
• mutational_vcv_proportion

Untested — blocked / deferred (test alongside the fix)

• community_plot_fitness_landscape — known not working
• plot_community, plot_community_2d — plotting (visual; lower priority)
• plant_community_check_for_inviable_strategies — has an open TODO to drop its direct plant dependency
• util_nlsolve — only exercised by the equilibrium_solve_* solvers, which are not yet reworked/verified for the community object (see CLAUDE.md "Known issues")

Notes

• When adding tests, verify each function actually works first; report breakage rather than asserting around it.
• Integration tests run the SCM and are a few seconds each — keep max_patch_lifetime small (e.g. 30) as the existing tests do.
• See CLAUDE.md (Test baseline / Known issues) for the broader status.

[dfalster]

Addressed the "Untested — should work today (highest value)" group. All five functions were verified to work first, then covered with tests. Suite is now green: 133 pass, 0 fail, 0 skip, 0 warn (up from 79).

Done

• community_solve_singularity_1D — test-solve-attractors.R. Finds the 1D attractor (≈0.0449 for lma within bounds = c(0.035, 0.045), selection gradient → ~0) and covers the non-bracketing-bounds branches (edge_ok = TRUE warns + returns the edge; edge_ok = FALSE errors).
• assembler_start / assembler_run / assembler_control — test-assembler.R. Control defaults + validation (stochastic flips run_type to single_step; rejects unknown params / missing vcv / maximum-births-with-single-step). Integration tests drive both the maximum-fitness (grid method) and stochastic assembly loops; assert history grows, residents stay within bounds, birth rates positive.
• community_fitness_landscape (grid) — test-fitness-landscape.R. Grid spans the bounds, resident is flagged with fitness ~0 at equilibrium, auto-solves demography when needed, rejects unknown methods. (GP/bayesopt method left uncovered — pulls in mlr3mbo.)
• tidy_assembly — test-assembler.R. One row per resident-step, expected columns, list-columns of tibbles for traits / fitness_landscape.
• mutational_vcv_proportion — test-assembler.R. Diagonal log-scale vcv (p * (log(upper) - log(lower))), zero off-diagonals, accepts a community, rejects infinite bounds.

Notes

• Added helper-assembly.R with shared assembly_model_support(max_patch_lifetime = 30) (previously inlined in test-community.R); integration tests keep max_patch_lifetime = 30 as suggested.
• No breakage found in any of these paths — nothing needed to be asserted-around.
• The blocked/deferred group (community_plot_fitness_landscape, plot_community*, plant_community_check_for_inviable_strategies, util_nlsolve / equilibrium_solve_*) remains untouched, to be tested alongside their respective fixes.
• CLAUDE.md test baseline updated to 133.

[dfalster]

Cleared most of the deferred group + made the suite fast (PR #36)

Following #33, the suite now uses the DD99 toy harness for model-agnostic functional coverage and keeps a minimal plant smoke set for the genuine SCM path. The mandatory feedback loop drops from a ~60s SCM-laden suite to ~1s; the plant-SCM anchor (~60s) is isolated in test-plant-smoke.R and only needs running when touching the plant path.

Deferred items — now done

• util_nlsolve — direct test (known nonlinear system + non-convergence error path).
• equilibrium_solve_* (nleqslv/dfsane) — plus equilibrium_hybrid/single_step; all five demography solvers verified against DD99's analytic multi-resident equilibrium through the community object (previously the gap).
• plant_community_check_for_inviable_strategies — the model-agnostic connector is now tested (explicit path incl. the drop branch); the plant path is exercised in the smoke set. (The original TODO to drop its direct plant dependency is effectively met by the harness connector from [evol assembly] Fast toy models with invasion fitness for dev & validation #33.)

Deferred items — still open (recommend splitting into a focused issue)

• community_plot_fitness_landscape, plot_community, plot_community_2d — these need a fix, not just a test:
  • they call is_tibble/%>%/select/aes_string/map_dbl but the package only imports ggplot2/purrr (not tibble/dplyr/magrittr);
  • plot_community* expect the old flat tidy_assembly schema (trait columns + invader), but tidy_assembly now emits list-columns (traits, fitness_landscape) + a resident flag;
  • community_plot_fitness_landscape requires a fitness_surrogate_function (only the GP method sets it) and hardcodes a log x-scale.
• util_nlsolve via the GP/bayesopt landscape and other mlr3mbo paths remain out of scope (heavy dep).

Bonus

Stochastic births are no longer hardcoded to the log scale (community_trait_transform in mutation/immigration/mutational_vcv_proportion), so stochastic assembly runs on a linear trait scale too — consistent with how the rest of the pipeline threads trait_scale.

Suite: 256 pass, 0 fail, 0 warn. PR: #36.