Declarative DeepNetwork API, fused pipeline, and batch-synchronous sweeps

[LegrandNico]

Declarative DeepNetwork API, mixed/fused pipelines, and batch-synchronous learning

Extends the vectorised DeepNetwork backend with a declarative construction API and a way to compose PyHGF networks with frozen calculations into larger models (up to a full Transformer), trained locally without backprop or autodiff.

What's new

• Declarative construction (builder.py) — LayerConfig, resolve_coupling_fn, and DeepNetwork.from_configs / from_dict: build networks from JSON/YAML-serializable configs for reproducible sweeps.
• Weight transplant (transplant.py) — from_linear, from_feedforward, from_embedding: build a DeepNetwork that reproduces a trained Equinox module's forward pass in PyHGF's layer layout.
• Mixed pipelines (hybrid.py) — PCModule with DeepNetworkAdapter (learning) / EquinoxAdapter (frozen), PCSequential, Residual, and the linear_/layer_norm_/gelu_adapter helpers. Errors are routed part-by-part with hand-derived backward formulas; no global computation graph.
• Transformer (transformer.py) — MultiHeadAttention, HybridGPT, hybrid_from_gpt: assemble a GPT from mixed parts, any slot frozen or learning.
• Fused executor (fused.py) — FusedPipeline + step_report: run a part tree as a single compiled program per training step, with all state threaded explicitly.
• New belief-propagation entry points (vectorized_belief_propagation.py, learning.py) — batch-synchronous batch_step / sample_step, split prediction_sweep / update_sweep / learn_sweep, input_prediction_error, and rank-one vectorized_weight_gradient_factors. Surfaced on DeepNetwork as prediction, update, input_error, batch_update, predict_states.

Also

• Categorical layers restricted to the output layer (softmax = single N-way choice).
• All new public symbols are documented in docs/source/api.rst.
• Tests for builder, hybrid, fused, transplant, batch learning, and the PCModule state contract.

[codecov]

Codecov Report

❌ Patch coverage is 75.90361% with 140 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
pyhgf/model/fused.py	64.48%	62 Missing and 3 partials ⚠️
pyhgf/model/transformer.py	45.90%	33 Missing ⚠️
pyhgf/model/deep_network.py	69.87%	12 Missing and 13 partials ⚠️
pyhgf/model/hybrid.py	86.74%	11 Missing ⚠️
pyhgf/utils/vectorized_belief_propagation.py	95.31%	3 Missing and 3 partials ⚠️

Files with missing lines	Coverage Δ
pyhgf/model/__init__.py	86.66% <100.00%> (+4.84%)	⬆️
pyhgf/model/builder.py	100.00% <100.00%> (ø)
pyhgf/updates/vectorized/learning.py	100.00% <100.00%> (ø)
pyhgf/utils/vectorized_belief_propagation.py	93.38% <95.31%> (-0.11%)	⬇️
pyhgf/model/hybrid.py	86.74% <86.74%> (ø)
pyhgf/model/deep_network.py	87.50% <69.87%> (-7.53%)	⬇️
pyhgf/model/transformer.py	45.90% <45.90%> (ø)
pyhgf/model/fused.py	64.48% <64.48%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.