ESN workflow migration

rd.ESN now uses the same nn.Module workflow as a custom rd.Model. The previous facade is replaced directly. There is no compatibility class.

Previous behavior	Current workflow
NumPy input produced NumPy forecasts	Supply floating Torch tensors; forecasts are Torch tensors. Trajectories use the optional dynamics integration.
A batch of size one was squeezed	The input layout is preserved: (1, T, D) forecasts stay (1, H, D).
esn.forecast(100) supplied a positional horizon	Use esn.forecast(horizon=100). The optional first argument is observation history.
.model exposed the raw graph	Use .graph; .workflow is the registered common Model child.
.feature_size was inferred on fitting	.input_size is known after construction with a width, build(width) or successful fitting.
The facade had its own conversion and fitting logic	Training-only scaling, explicit targets and forecasts all use Model's implementation.

Convert external arrays explicitly at the boundary:

import numpy as np
import torch
import resdag as rd

external_array = np.sin(np.linspace(0, 6, 100))[:, None]
series = torch.as_tensor(external_array, dtype=torch.float64)
esn = rd.ESN(reservoir_size=16, washout=5, seed=5).fit(series)
forecast = esn.forecast(horizon=10)
array_for_plotting = forecast.detach().cpu().numpy()
assert forecast.shape == (10, 1)

Automatic standardization is now the default. Fit uses training samples after washout and returns predictions in the original units. Set preprocessing="off" when supplying your own transformed data. Washout may be zero.

Build before attaching an optimizer so it sees every parameter:

trainable = rd.ESN(
    reservoir_size=12, input_size=1, washout=5,
    dtype=torch.float64, readout_kwargs={"trainable": True}, seed=6,
)
optimizer = torch.optim.Adam(trainable.parameters(), lr=1e-3)
trainable.fit(series)
parameter_ids = [id(parameter) for parameter in trainable.parameters()]
trainable.build(1)  # Same-width calls preserve the existing parameters.
assert parameter_ids == [id(parameter) for parameter in trainable.parameters()]

Without a width, first fitting stages a candidate and attaches it after success. A failed first fit leaves the preset unbuilt. Refitting preserves the existing workflow, parameters and streaming state when fitting fails.

Ordinary esn(x) is differentiable and carries state between calls. Reset it when switching independent sequences, or thread explicit states with forward_stateless. Forecasting resynchronizes on supplied history or the cached training tail and begins at the immediate unseen sample.

Use .to(...), .float() and .double() for standard module conversion. A conversion before lazy fitting fixes the current placement; subsequent data must match. There are no facade-specific .device or .dtype shadow fields. A fixed seed initializes the whole preset, including unfitted readout weights, while restoring the surrounding Torch CPU/available accelerator RNG streams. Unseeded construction follows ordinary Torch random-state behavior.

A state_dict includes the fitted workflow and width. Recreate the same preset architecture/settings before loading; device and dtype may differ. Metadata checks cannot verify the full behavior of arbitrary custom Python functions. Complete artifact loading without reconstructing the architecture belongs to the artifact roadmap phase.

The current ESN and Model checkpoint metadata version is 2. Custom preprocessing and named trajectory clocks are explicit parts of that schema. Older common-workflow checkpoints are rejected with a reconstruction/refit message; no compatibility loader or automatic converter is provided.

Custom input_transforms={"input": module} and output_transforms={"output": module} follow the same reversible preprocessing protocol as Model. Their original modules remain registered before a lazy build and remain the fitted workflow's modules afterward. See custom preprocessing and named trajectories.