[wshlavacek]

This PR closes the changelog's Unreleased section as 3.12.0 and moves CMakeLists.txt, pyproject.toml and CITATION.cff off 3.11.0 together. A second commit brings RELEASING.md up to date with how 3.11.0 was cut.

Minor, not patch

Models observe this release. Results change for:

• every Arrhenius model whose energy-pattern context the load-time expansion could not represent, 48 to 52 standard errors off BioNetGen's network on the issue's reproducers (Arrhenius expansion drops or misapplies energy-pattern context, and drops a rule's own reactant context #109);
• a rule labelled _reverse_X with no mirror rule X, which took the reverse half's phi (Arrhenius expansion drops or misapplies energy-pattern context, and drops a rule's own reactant context #109);
• every model whose XML lists a function ahead of one it reads, and a function that reads a table function (A table function whose counter is a function recurses without end (stack overflow) #111);
• every rate law that reaches reactant_N() through a helper function, which never fired (Local-function product rate law silently yields zero propensity (no reactions fire) #59);
• runs from a t_start late enough that waiting times rounded away, z = -290 from 1e12 (A run from a large t_start loses its waiting times to rounding: keep the clock as origin + elapsed #110);
• sessions that reach an absorbing state, which kept the time of their last event.

An embedder also gains a capability, run_ensemble() (#113). Either alone makes it at least a minor, as 3.10.0 and 3.11.0 were.

Minor, not major

include/ moved only by addition: Simulator::run_ensemble() and the EnsembleSpec and EnsembleResult structs it takes and returns. Every other line in the include/ diff is comment text. No existing declaration changed, so every existing call, bngsim's and PyBNF's included, compiles unchanged and means what it did.

librulemonkey now links Threads::Threads privately. RuleMonkeyConfig.cmake re-finds it for find_package() users, and an add_subdirectory() consumer such as bngsim gets it from RuleMonkey's own find_package(Threads).

The Tier-0 Arrhenius refusal changes scope. 3.11.0 ran several context shapes at the wrong rate with no warning. Each now runs at BioNetGen's rate or is refused with the rule and the reason named; the changelog lists the shapes still refused. The Tier-0 refusal of a local rate law that reads reactant_N() now also covers one that reaches it through a helper. One load-time warning is new: an energy pattern whose rate bound stays loose after the per-class split (#109).

Five runtime errors are new, each on an input that used to run wrong silently:

• a function that depends on itself, through the functions it reads or a table counter (A table function whose counter is a function recurses without end (stack overflow) #111);
• a non-finite t_end on run(), simulate(), step_to(), run_ensemble() or parameter_scan();
• a NaN segment start on simulate();
• a NaN in TimeSpec::sample_times;
• simulate(TimeSpec) with a t_end before the session's time.

Not listed

• The expression-layer refresh to BNGsim 4361958 (Update the vendored BNGsim expression layer to 4361958 #120), as RELEASING.md says for a contents refresh. It is what lets BNGsim vendor this release, because vendor_rulemonkey.py's EXPRTK_SYNC_FILES guard matches bngsim main again.
• The RELEASING.md update in this PR: merge commits rather than squash, the tag on the merge commit after checking its tree, ten CI checks with tsan, pre-commit from any install, and the date written twice.
• PyPI remains out of scope (Packaging: rulemonkey-harness is not a distributable Python package #77).

The BNGsim re-vendor is tracked in #118 (lanl/bngsim#1058), which also drops bngsim's refuse_rulemonkey_energy_misreads. lanl/bngsim#1059 drops bngsim's elapsed-time shift once this is vendored.

Verified

• ctest --preset release 67/67 (57 at 3.11.0). The ten new cases are ensemble_test, ensemble_scheduler_test and the three rm_driver_reps_* cases (Run replicate ensembles in parallel: run_ensemble() and rm_driver --reps #113), arrhenius_context_test and energy_match_test (Arrhenius expansion drops or misapplies energy-pattern context, and drops a rule's own reactant context #109), late_start_test (A run from a large t_start loses its waiting times to rounding: keep the clock as origin + elapsed #110), function_settle_order_test (A table function whose counter is a function recurses without end (stack overflow) #111) and reactant_count_helper_test (Local-function product rate law silently yields zero propensity (no reactions fire) #59). The bng_oracle case now also scores arrhenius_context and arrhenius_context_bind, and arrhenius_formula gains seven columns (Arrhenius expansion drops or misapplies energy-pattern context, and drops a rule's own reactant context #109).
• RuleMonkeyConfigVersion.cmake reports PACKAGE_VERSION "3.12.0" after cmake --preset release on this commit.
• mkdocs build --strict (mkdocs-material 9.5) is clean.
• pre-commit on CHANGELOG.md CMakeLists.txt pyproject.toml CITATION.cff passes.

The date is 2026-10-22. If this merges on a later day, the changelog heading and CITATION.cff date-released move with it.