Fix support-vs-array index confusion in sampling and DEFM::simulate

[gvegayon]

Summary

Fixes a sampling bug where a simulated Markov/DEFM process would collapse the first time a conditioning state discovered mid-simulation was revisited. In the defm README "Example 2: A fun model" (a 10-state torus walk), the active state advances correctly through row 12 but breaks at row 13 and never recovers.

Root cause is a support-index / array-index mix-up in two places:

1. Model::sample(const Array_Type &, params) — keys2support maps a hash key to a support index, but the reuse branch assigned i = locator->second (already a support index) and then computed a = arrays2support[i], mapping it through arrays2support a second time. So reusing a known conditioning state selected an unrelated support. The two "correct" rows in the README output were coincidences (a support index that happened to be a valid array index mapping back to the same support).

2. DEFM::simulate — the mirror-image bug: the first process of each individual was sampled with sample(idx->at(model_num++), par), i.e. arrays2support[model_num] (a support index) passed to the sample(size_t, params) overload, which itself expects an array index and resolves it through arrays2support. This gives wrong first-row draws for the 2nd+ individuals in a multi-individual simulation.

Fix

Key on the support index directly in both places. The Array_Type overload now delegates probability computation to update_pset_probs(params, 1u, a), matching the existing sample(size_t, params) path and removing the duplicated recompute loop (which carried the same index confusion into its own likelihood(..., i, ...) call).

Tests

Adds tests/20-defm-simulate-torus.cpp: a deterministic 10-state torus walk (strong transition reward + strong ones penalty, so the outcome is RNG-independent and portable). It asserts every simulated row has exactly one active state advancing by one and wrapping y9→y0.

• Fails on the old code (walk collapses at row 13).
• Passes with this fix.
• Full suite: 109 assertions in 21 test cases, all passing.

Verified against the defm README Example 2 (single individual) and a two-individual case where the second individual starts in a distinct state — both now walk the torus correctly for all rows.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 89.30%. Comparing base (de83fec) to head (83b3e7c).

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #25      +/-   ##
==========================================
+ Coverage   89.04%   89.30%   +0.26%     
==========================================
  Files          46       46              
  Lines        3204     3170      -34     
==========================================
- Hits         2853     2831      -22     
+ Misses        351      339      -12     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[Copilot]

Pull request overview

This PR fixes a support-index vs array-index mix-up that could cause DEFM/Markov simulation sampling to select the wrong support when reusing a previously seen conditioning state, and corrects the analogous indexing bug in DEFM::simulate for multi-individual runs. It also adds a regression test based on a deterministic 10-state torus walk.

Changes:

• Fix Model::sample(const Array_Type&, params) to avoid remapping a support index through arrays2support and to reuse the update_pset_probs path.
• Fix DEFM::simulate to pass an array index to sample(size_t, params) rather than a support index.
• Add a deterministic torus-walk regression test and wire it into the test runner.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 2 comments.

File	Description
tests/main.cpp	Adds the new DEFM torus simulation regression test to the test aggregator.
tests/20-defm-simulate-torus.cpp	New regression test validating the simulated state advances deterministically around a 10-state torus.
include/barry/models/defm/defm-meat.hpp	Fixes DEFM simulation to sample using the correct index type for the first process of each individual.
include/barry/model-meat.hpp	Fixes support vs array index handling in Model::sample(Array_Type, params) and simplifies probability computation by delegating to update_pset_probs.

[gvegayon]

Good catch — addressed in 83b3e7c. Added a pset_probs.empty() check to the guard, matching sample(size_t, params):

if (pset_probs.empty() ||
    !first_calc_done[a] ||
    !vec_equal_approx(params, params_last[a]))
    update_pset_probs(params, 1u, static_cast<int>(a));

init() adds arrays without sizing pset_probs, and a prior likelihood() call sets first_calc_done/params_last without ever filling it, so this guards both. Note there is a deeper, pre-existing cache-coherence limitation shared with sample(size_t, params): if likelihood() populated params_last for a second support and pset_probs was sized by a first support, a same-params reuse of that second support could still read an unfilled slice (e.g. defm_mle then sim_defm on the same object). That's out of scope here and would want a dedicated per-support 'pset-probs-valid' flag — happy to open a follow-up.

[gvegayon]

Done in 83b3e7c. The test now simulates two individuals with different starting states (starts = {0, 3}) and asserts each one's walk advances by one from its own start. The second individual (starting at y3) exercises the DEFM::simulate first-row fix specifically: on the old code its walk collapses to all-zeros from its first sampled step. Confirmed the extended test fails against the pre-fix headers (Individual 2, row 3: expected single active state y5, got 0 active) and passes with the fix.