[wshlavacek]

Closes #1058.

RuleMonkey 3.12.0, vendored in #1220, gives every Arrhenius rule it runs BNG2's rate. Where expanding a rule at load is exact, it expands it. Everywhere else it counts dG at each firing, the way BNG2's network counts it per reaction (richardposner/RuleMonkey#109). The shapes it cannot run, it refuses itself, naming the rule and the reason. That leaves bngsim's own gate with nothing to catch, so this PR removes it.

What changes

• Removed from python/bngsim/_network_free_checks.py:

  • refuse_rulemonkey_energy_misreads and rulemonkey_energy_misreads;
  • the helpers only they used: _arrhenius_centers, _expanded_exactly, _bond_partners, _component_names, _plain_bond_pattern, _graph, _components and the Site alias;
  • the module docstring's paragraph on them.

  Their calls in Simulator and RuleMonkeySession.__init__ are gone too. _local and _children stay, because other checks use them.

• Tests (python/tests/test_rulemonkey_arrhenius.py): test_rm_refuses_an_energy_rule_it_would_misread is replaced by value tests on all fifteen misread/ models. Each replicate gets a fresh Model or session: 40 for rm and the session, 200 for SSA.

  • test_energy_context_fires_at_bng2s_closed_form: method='rm', a RuleMonkeySession, and network SSA on the .net, each checked against the closed form in the model's .bngl header. Each observable is binomial(1000, exp(-k t)). Two stay at 1000 because BNG2 never fires the rule there: rule_context's s~0 complexes and ringopen's open chain. The SSA case checks the closed form itself.
  • test_energy_context_matches_network_ssa: rm against network SSA, a two-sample z test at each time point.

• New model tests/data/nfsim/arrhenius/state_pattern, generated with BNG2 2.9.3:

  • A reversible binding rule with the energy pattern A(b!1).B(a!1,s~1), and 100 each of A, B(s0) and B(s1).
  • Detailed balance gives the exact stationary means: AB0 = 22.33 and AB1 = 68.20. If the state pattern were dropped, both would be 39.33.
  • rm settles on those means at t = 5 and matches network SSA along the way. Network SSA settles on them too.

• Refusal test: test_rulemonkey_names_an_energy_rule_shape_it_refuses_itself now matches RuleMonkey's own message, which names the rule: rule '_R1' is outside the shapes RM runs: it changes a state.

• Docs: the "Arrhenius (energy) rules" section of docs/user-guide/network-free.md now says:

  • what rm runs;
  • what RuleMonkey refuses;
  • one case RuleMonkey reads differently from BNG2. Two unidirectional mirror-image rules, X and _reverse_X, with the same phi and Ea0 parameters write the same XML as one reversible rule, and RuleMonkey gives the second rule 1-phi.

• changelog.d/1058.fixed.md.

Shapes RuleMonkey still refuses

None of the fifteen misread models is one of them: all fifteen load and run on 3.12.0. The shapes are:

• a state change;
• a homodimer;
• a ring closure;
• two bonds;
• include/exclude constraints;
• a bond that can reach a .-joined energy pattern.

RuleMonkey refuses each of these by name, and state_change covers that path.

Verification

• Sensitivity. On 3.12.0 I swapped each closed-form rate for the misread one, the rate with the context dropped or misapplied. The closed-form check then fails on every observable the misread changes, at |z| from 50 to 515 on rm's runs.
• RuleMonkey tests. The 18 files that drive RuleMonkey: 496 passed, 3 skipped (448 before). This is after rebasing on ff817e95 (ssa: carry what each firing rounds away, and read samples at its exact time (#1059) #1075's SSA change) and rebuilding.
• Test floors. The full suite grows by 48, well inside ci/test_floors.env's bounds, so no floor changes.