Custom boundary without `boundary_params` fails at `simulate()` with `KeyError: "Boundary 'my_boundary' is not registered"`

[AlexanderFengler]

Simulator("ddm", boundary=my_callable) with no boundary_params constructs without warning, because ddm ships "boundary_params": [] and the guard checks key presence rather than emptiness. make_boundary_dict then tests boundary_params for truthiness, takes the registry branch, and looks the callable's __name__ up in the boundary registry. The user is told their function is not registered and shown a list of six built-in boundary names, which points at the wrong fix — the actual fix is the one-line boundary_params=[...]. ModelConfigBuilder.add_boundary already rejects the same mistake at the point it is made.

Environment: ssm-simulators 0.13.2, Python 3.12, macOS (Darwin 25.4.0).

Reproducer

import numpy as np

from ssms.basic_simulators.simulator_class import Simulator


def my_boundary(t, theta=0.2, scale=1.0):
    return scale * np.maximum(1.0 - theta * t, 0.1)


# "ddm" ships "boundary_params": [], so the missing-params warning never fires.
sim = Simulator("ddm", boundary=my_boundary)
print("boundary_name  :", sim.config["boundary_name"])
print("boundary_params:", sim.config["boundary_params"])

sim.simulate(theta={"v": 0.5, "a": 1.0, "z": 0.5, "t": 0.3}, n_samples=10)

Output

boundary_name  : my_boundary
boundary_params: []
Traceback (most recent call last):
  File "/…/snippet2.py", line 15, in <module>
    sim.simulate(theta={"v": 0.5, "a": 1.0, "z": 0.5, "t": 0.3}, n_samples=10)
  File "/…/site-packages/ssms/basic_simulators/simulator_class.py", line 651, in simulate
    boundary_dict = make_boundary_dict(model_config_local, theta)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/…/site-packages/ssms/basic_simulators/simulator.py", line 328, in make_boundary_dict
    boundary_info = boundary_registry.get(boundary_name)
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/…/site-packages/ssms/config/boundary_registry.py", line 117, in get
    raise KeyError(
KeyError: "Boundary 'my_boundary' is not registered. Available boundaries: ['addm_collapse', 'angle', 'conflict_gamma', 'constant', 'generalized_logistic', 'weibull_cdf']"

Related cases, from a second run that catches instead of crashing (verbatim excerpts):

shipped config for ddm   boundary_params=[]
shipped config for addm  boundary_params=[]
shipped config for angle boundary_params='<no key>'

=== B: Simulator('angle', boundary=<callable>), no boundary_params ===
warnings at construction: ["Custom boundary function provided without 'boundary_params'. You may need to specify bound"]
simulate RAISED KeyError: "Boundary 'my_boundary' is not registered. …"

=== C: documented form, boundary_params=['theta','scale'] ===
warnings at construction: []
OK -> rts[:3]: [0.548798 1.370101 1.313849]

=== D: parameter-free boundary with explicit boundary_params=[] ===
config['boundary_params']: []
simulate RAISED KeyError: "Boundary 'flat_boundary' is not registered. …"

=== E: ModelConfigBuilder.add_boundary(cfg, <callable>) with no params ===
RAISED ValueError: Must provide boundary_params when using custom boundary function

Case D is the sharper edge: there is currently no value of boundary_params that expresses "my boundary takes no extra parameters" — [] fails identically to omitting it — so that shape of custom boundary is unreachable through Simulator.

Why

• ssms/basic_simulators/simulator_class.py:388-400 — callable branch of _apply_custom_boundary: sets config["boundary_name"] = fn.__name__ (392), and warns + sets config["boundary_params"] = [] only if "boundary_params" not in config (394-400). ddm and addm ship the key with value [], so no warning.
• ssms/basic_simulators/simulator_class.py:206-216 — _build_config applies config.update(config_overrides) before _apply_custom_boundary, so a user-supplied boundary_params and a base model's shipped one are indistinguishable at the point of the check.
• ssms/basic_simulators/simulator.py:303 — if callable(config.get("boundary")) and config.get("boundary_params"): — truthiness, so [] falls through to the registry branch.
• ssms/basic_simulators/simulator.py:326-328 → ssms/config/boundary_registry.py:115-120 — registry lookup on the function's __name__, raising the misleading KeyError.
• Correct behaviour already exists at ssms/config/model_config_builder.py:405-412.

Suggested fix

Two independent changes. First, dispatch on presence rather than truthiness so a callable boundary never falls through to a registry lookup on a function name, which also makes boundary_params=[] mean "no extra parameters":

-    if callable(config.get("boundary")) and config.get("boundary_params"):
+    if callable(config.get("boundary")) and config.get("boundary_params") is not None:

Second, fail fast at construction the way add_boundary does, checking the user's own config_overrides rather than the merged config so a base model's shipped [] cannot suppress it:

elif callable(boundary):
    self._validate_boundary_function(boundary)
    if "boundary_params" not in config_overrides:
        raise ValueError(
            f"Custom boundary {getattr(boundary, '__name__', 'custom')!r} requires an "
            "explicit boundary_params list, e.g. Simulator('ddm', boundary=my_boundary, "
            "boundary_params=['theta', 'scale']). Pass [] if it takes no extra parameters."
        )
    config["boundary_params"] = config_overrides["boundary_params"]

Optionally, in the else branch at simulator.py:326-328, if config["boundary"] is callable but the registry lookup fails, re-raise naming the real problem instead of the registry's "not registered" text.

Is this intended?

The documented form works cleanly — case C above constructs with no warnings and simulates — so this is not a docs bug, and config["boundary_name"] = fn.__name__ is defensible on its own as a readable metadata label. There is also already a warning on models whose config omits boundary_params (case B), which a maintainer may consider sufficient. The counter is that ddm and addm get no warning at all, and that even when the warning does fire it is emitted at construction while the failure arrives later at simulate(), misattributed to the registry.

Two caveats. The truthiness guard at simulator.py:303 looks deliberate — the comment at 304-307 says it exists so process-local custom boundaries survive multiprocessing workers that start with a fresh registry; is not None preserves that intent, whereas dropping the second clause entirely would not. And the construction-time raise is a behaviour change for code that builds a Simulator with a callable boundary and only introspects .config without simulating; that combination is already non-functional, but it is technically a break. I exercised only the Simulator path — I did not survey whether ssms.dataset_generators or hssm_support reach make_boundary_dict with a callable boundary and empty boundary_params.

[linear-code]

HSSM-706