PDep explorer cannot read any RMG-written network file (bath-gas rule enforced on the wrong file)

[alongd]

The Arkane PES-explorer path added in #179 rejects every network file RMG writes. Not an edge case — the whole feature path, for real RMG output.

The rule, and where it is enforced

write_arkane_explorer_input_file requires every bath-gas label to resolve to a species() block carrying a literal reactive=False (t3/pdep/explorer/input_file.py:703-745, ~line 748):

Bath gas label '{label}' is a species() block that does not carry a literal 'reactive=False' keyword.

The intent is correct. Arkane identifies the bath gas as the unreactive species, and a bath-gas species left reactive makes Arkane try to generate statmech for it (arkane/input.py:207-217, whose own error text says "If it is a bath gas, set reactive = False to avoid generating…").

Why no RMG file can satisfy it

arkane/pdep.py:654 PressureDependenceJob.save_input_file() — the function RMG calls at rmgpy/rmg/pdep.py:869 to write pdep/network<i>_<n>.py — writes species blocks containing exactly: label, structure, E0, modes, spinMultiplicity, opticalIsomers, molecularWeight, collisionModel, energyTransferModel, thermo.

reactive is never written, for any species. There is no code path in that writer that emits it.

Confirmed on this repo's own real fixture: in tests/data/pdep_network/iteration_1/RMG/pdep/network4_2.py, He (line 225) and Ne (line 237) are named as bathGas in the network(...) block (line 438) while their species() blocks carry no reactive keyword at all. tests/test_pdep/test_explorer_input_file.py:11-14 already concedes the fixture "cannot exercise the bath-gas-refusal … paths" — the tests pass because they use hand-built strings that real RMG never produces.

The information is present, in the other block

The same save_input_file does write the bath gas, into network(...):

f.write('    bathGas = {\n')
for spec, frac in self.network.bath_gas.items():

populated upstream from rmgpy/rmg/pdep.py:856-858:

bath_gas = [spec for spec in reaction_model.core.species if not spec.reactive]
assert len(bath_gas) > 0, 'No unreactive species to identify as bath gas'

So the bath-gas identity is authoritative and always present — in network(...), which T3's writer deliberately drops.

Suggested direction

Treat the source's network(...) bathGas as the authoritative declaration, and emit reactive=False on those species in the file T3 generates. That is a faithful translation between two encodings of the same fact in a T3-authored file, not a rewrite of user input. It would also let bath_gas default sensibly instead of the current trap where PDepExplorerConfig(bath_gas=None) passes construction-time validation (config.py:202-213) and only raises at write time (input_file.py:703-707), after the run directory has been claimed.

Design questions worth settling before implementing: whether to default config.bath_gas from the network block, inject reactive=False on write, or both; and what should happen when a source species carries an explicit reactive=True while being named as bath gas (conflict → refuse, presumably).

Related

• run_arkane_job has no timeout and no kill path (t3/pdep/explorer/arkane.py:381-399, noted in a comment there); a hung Arkane hangs the campaign.
• Separately: RMG's assert len(bath_gas) > 0 means a pdep run with no declared inert dies deep inside network generation. A T3 schema validator refusing that at input time is worth adding — this was the one surviving idea from the now-closed Pdep fixes #60.