[wshlavacek]

Refresh third_party/rulemonkey with scripts/vendor_rulemonkey.py at tag v3.12.0 (richardposner/RuleMonkey@9df2984), from 3.11.0 (f11c251). The second commit records the new baseline in scripts/RULEMONKEY_VENDORING.md.

Two-repo ordering, done upstream first

This tree's expression.hpp and expression.cpp had moved past 3.11.0's ExprTk sync pin (6ce2276), so the EXPRTK_SYNC_FILES guard would have refused the tag. richardposner/RuleMonkey#120 re-synced RuleMonkey's standalone copy to bngsim 4361958 before the release was cut. --summary reports ExprTk sync pin: 4361958… (matches current BNGsim tree), and that pin is still current on main at 02bb304.

What reaches method="rm" and RuleMonkeySession

• Functions settle in dependency order (A table function whose counter is a function recurses without end (stack overflow) richardposner/RuleMonkey#111). A function that reads one listed after it in the XML no longer reads a stale value. A table function indexed by a function no longer overflows the stack. A function that reads a table function gets its whole expression.
• reactant_N() through a helper (Local-function product rate law silently yields zero propensity (no reactions fire) richardposner/RuleMonkey#59): such a rate law fires at its own rule's count, where it never fired before.
• Late t_start (A run from a large t_start loses its waiting times to rounding: keep the clock as origin + elapsed richardposner/RuleMonkey#110). RuleMonkey keeps its clock as a per-call origin plus elapsed time. bngsim already ran clock-free models on elapsed time. A model that reads the clock has run on absolute model time since Re-vendor RuleMonkey 3.11.0: lift method='rm''s time() and time-TFUN refusals, run rm on absolute model time, and pick up the RuleMonkey fixes since 3.10.1 #1022, and now keeps its waiting times there. I measured SSA, and rm on a model that reads the clock, lose waiting times to rounding at a large absolute t_start (z = -287 at 1e12, silent) #1059's decay with g() made to read time(), 10 seeds, A0 at t0 + 2 against 1e5·e⁻¹ = 36788. It gives 36733 (z = −1.1) from t0 = 0, 1e9, 1e12 and 1e13 alike; SSA, and rm on a model that reads the clock, lose waiting times to rounding at a large absolute t_start (z = -287 at 1e12, silent) #1059 reports 22743 (z = −291) at 1e12 on 3.11.0. That is the rm half of SSA, and rm on a model that reads the clock, lose waiting times to rounding at a large absolute t_start (z = -287 at 1e12, silent) #1059. The ssa half is bngsim's own engine and is not touched here.
• A session that runs out of events moves its clock to t_end. A clock-reading model that decays out now reports current_time == 10.0 after simulate(0, 10, 10), and simulate(10, 20, 10) runs.
• A non-finite t_end, a NaN segment start and a NaN in sample_times are refused.

RuleMonkey's Arrhenius work (richardposner/RuleMonkey#109) does not reach users yet. refuse_rulemonkey_energy_misreads still refuses every shape 3.11.0 misread, and lifting it is #1058.

run_ensemble() is new in RuleMonkey's API, and bngsim does not call it. RuleMonkey's include/ changed only by addition, so nothing bngsim compiles against changes signature. librulemonkey now links Threads::Threads, found by RuleMonkey's own CMakeLists.txt under add_subdirectory().

The one test change

test_rulemonkey_names_an_energy_rule_shape_it_refuses_itself matched 3.11.0's wording of RuleMonkey's own refusal of a state-change Arrhenius rule ("binding/unbinding shape"). RuleMonkey still refuses that rule by name, now as "outside the shapes RM runs: it changes a state …", and the test matches that.

Verified

• vendor_rulemonkey.py --summary, then --check before writing (it reported the expected 18-path drift and no guard failures), then the refresh, then --check again: "matches v3.12.0 @ 9df2984".
• Rebuilt with scripts/rebuild_editable.py, with no warnings from RuleMonkey's sources.
• test_rulemonkey.py 38 passed, and test_timeout.py -k rulemonkey 10 passed.
• Every test file that drives RuleMonkey (18 files): 448 passed, 3 skipped. The skips are the RuleMonkey-not-compiled cases.
• scripts/local_ci.py wheel --python 3.12: ALL PASS, including the RuleMonkey session smoke.
• The pre-push hook's full python/tests run passed, twice. The first push's SSH connection timed out during a 29-minute cold-cache run; the second passed again and pushed.

Changelog: changelog.d/+rulemonkey-3-12-0.fixed.md.

Follow-ups: #1058 (lift refuse_rulemonkey_energy_misreads, tracked upstream as richardposner/RuleMonkey#118), and #1059's ssa half.