[Montana]

Closes #1361.

What was wrong

The default ODE tolerances are rtol=1e-8 and a scalar atol=1e-8: BNG2.pl's, chosen for species held as molecule counts. A species that stays below atol for a whole run has its error controlled by nothing but the tolerance. Models in molar units (SBML, BioModels) are full of such species, and nothing said so. On main c0877fa:

• The issue's binding model (species near 1e-9 M), at defaults: max relative error 2.2 against a tight reference, and A = -8.7e-10 at t = 5000 for 4.1e-10.
• tests/data/BIOMD0000000002.xml (L at 1e-5 M; complexes at 3e-10 to 7e-9 M): errors of 1e-2. Its largest species is above atol, so a test of the whole run against atol passes it. The species that matter are below it.

The fix: a warning, not a new default

Simulator.run() raises the new bngsim.AtolScaleWarning (a UserWarning, exported) for an ODE run when both hold:

• The caller gave no atol: neither run(atol=...) nor set_tolerances(atol=...), and no vector or tracking atol is in force.
• Some species stays below atol for the whole run while reaching at least 1e-6 of the run's largest species. That floor is what keeps a species stuck at rounding in a count model (1e-17 of the rest) quiet, while it catches BIOMD0000000002, whose complexes reach 7e-4 of its ligand.

The message names up to five such species and their peak, and names the remedy: atol="auto" (#196), or a scalar atol well below them. docs/user-guide/solvers.md gets a section on the default tolerances and models in small units.

Why not change the default? atol="auto" as the default is the issue's first suggestion, and arguably the better end state. But it changes every default ODE trajectory, BNG2.pl parity at defaults included, so it is a maintainer's call. This PR makes the hazard visible without moving any number; a follow-up can change the default.

Measured, for false positives

corpus	runs	warned
every .net in the repo (tests/data, benchmarks, parity_checks; counts, by BNG2.pl)	176	2: nanomolar_dimer.net (species at 1.5e-9) and jac_selfcheck_amount_scaled.net (at 1.7e-10). Both are small-unit models, which is what the warning is for
tests/data/BIOMD*.xml	12	1: BIOMD0000000002
the BioModels corpus (1,323 SBML files, t = 0–100)	1,260 (63 did not load or run)	35 (2.8%)

Are the 35 right? For each warned BioModel, the default run was compared with a reference at rtol=1e-12, atol=1e-20. The error is the largest over species, relative to each species' peak:

• At the default: above 1e-3 in all 35, median 1 (100%). Three run away entirely (BIOMD0000000362, 0086 and 0085, at 1e+275 and more).
• With atol="auto", the remedy the warning names: median 6.1e-7.
• Two exceptions: BIOMD0000000609 (1) and BIOMD0000000438 (1.6e+5) stay far off with auto as well, so something beyond the tolerance is wrong in them. They are not this PR's.

So no warned model was a false alarm.

Tests (test_atol_scale_warning.py, 8)

test	on main
test_a_molar_model_at_the_default_atol_is_warned: ends negative, one warning naming 'A()' and atol='auto'	fail
test_a_mixed_scale_biomodel_is_warned: BIOMD0000000002, naming 'BL'	fail
test_the_remedy_it_names_is_right_and_quiet: control. atol='auto' is within 1e-5 of the reference, with no warning	pass
test_an_atol_the_caller_chose_is_not_second_guessed[run/set_tolerances]: controls	pass
test_a_count_model_is_not_warned: control	pass
test_a_species_at_the_noise_of_a_count_model_is_not_warned: control. A species below atol but at 1e-15 of the largest	pass
test_ssa_is_not_warned: control. A concentration of 1e-9 in a compartment of 1e10, ten molecules: SSA has no atol	fail (no AtolScaleWarning on main)

Mutations

mutant	fails
V1: no 1e-6 relevance floor	the count-model noise control
V2: a set_tolerances atol is warned as well	[set_tolerances]
V3: SSA is checked too	test_ssa_is_not_warned
V4: warn only when the whole run is below atol (the first design)	the BIOMD0000000002 test

Measured

Full suite (macOS arm64, main c0877fa + this branch): 16,424 passed, 0 failed, 137 skipped. Six existing tests now raise the warning, and each is a real case: five run a model whose only species is tiny (7.5e-14 to 3e-9; literal-parsing and rounding tests), and one runs a molar BioModel whose Bfa1P4_Tem1GDP peaks at 3.2e-10 while its largest species reaches 1.7e-4. None of them asserts on warnings. pre-commit run --all-files (new files staged) and sphinx -W with docs/requirements-ci.txt are clean.