deduce_method_type substring-matches 'am' inside 'cam', so CAM-B3LYP types as semiempirical and open-shell species run restricted

Level.deduce_method_type (arc/level.py:344) tests its method lists by substring:

semiempirical_methods = ['am', 'pm', 'zindo', 'mndo', 'xtb', 'nddo']
...
elif any(sm_method in self.method for sm_method in semiempirical_methods):
    self.method_type = 'semiempirical'

'am' is a substring of 'cam', so every Coulomb-attenuated functional types as semiempirical.
is_species_restricted (arc/job/adapters/common.py:631) early-returns True for
REFERENCE_AGNOSTIC_METHOD_TYPES, which includes semiempirical, before any multiplicity
check — so an open-shell species at a CAM functional is decided restricted.

Executed on a doublet OH ([OH], multiplicity 2, number_of_radicals=None):

b3lyp          type=dft            is_species_restricted=False
cam-b3lyp      type=semiempirical  is_species_restricted=True
camh-b3lyp     type=semiempirical  is_species_restricted=True
lc-camb3lyp    type=semiempirical  is_species_restricted=True
cam-qtp00      type=semiempirical  is_species_restricted=True
wb97xd         type=dft            is_species_restricted=False
m062x          type=dft            is_species_restricted=False

It reaches the input file. Gaussian (cam-b3lyp, data/ess_methods.yml:17) — benign, because
Gaussian selects the unrestricted method itself for multiplicity > 1:

b3lyp     #P integral=(grid=ultrafine, Acc2E=12) guess=mix ub3lyp/def2tzvp ...
cam-b3lyp #P integral=(grid=ultrafine, Acc2E=12) guess=mix cam-b3lyp/def2tzvp ...

TeraChem (camb3lyp, data/ess_methods.yml:154) is not benign — terachem.py:278 writes
the prefix explicitly:

b3lyp     charge 0 / spinmult 2 / basis def2tzvp / method ub3lyp
camb3lyp  charge 0 / spinmult 2 / basis def2tzvp / method rcamb3lyp

method rcamb3lyp with spinmult 2 is a restricted DFT calculation on a doublet radical — a
wrong number with no error raised. QChem would be the same class of failure
(qchem.py:225 writes input_dict['unrestricted'] = 'False') if it registered a CAM
functional; it currently does not.

Two other consequences of the same misclassification, since semiempirical is
reference-agnostic throughout: job_scf_reference_is_restricted returns None for these
levels, and any frequency scale factor / method-type dispatch keyed on method_type sees
semiempirical for what is an ordinary range-separated hybrid.

The narrow fix is to stop matching 'am'/'pm' as bare substrings — e.g. anchor the
semiempirical stems to a token boundary, or match am1/pm3/pm6/pm7 explicitly rather
than their two-letter prefixes. Worth auditing the other lists for the same trap while there
('ri' matches inside ripwpb95, 'ic' inside a good many names).

Found while reviewing #1014 — @calvinp0 spotted the misclassification there; the TeraChem
escalation and the executed inputs above are from verifying it. Not caused by that PR;
arc/level.py is untouched by it.

Start at arc/level.py:344 and trace how method_type feeds is_species_restricted in arc/job/adapters/common.py:631. Reproduce the listed CAM examples for an open-shell species, then inspect terachem.py:278 and the relevant data/ess_methods.yml entries. Done means CAM methods are classified as DFT and unrestricted inputs are preserved where required, without breaking genuine semiempirical names.