Keras 3 Persistence, Duplicate‑Edge Graph Export, and Hyperdense Connectivity Intensity Fix

[Thunderblok]

Core Additions

Native Keras 3 .keras round‑trip (architecture + weights + optimizer state) via save_model_and_meta / load_model_safe (safe_mode).
Deterministic seeding utilities for reproducible evaluation and resume training.
Graph exporter that preserves duplicate inbound edges (no accidental dedup), enabling faithful reconstruction in future backends (Torch / ONNX).
Serialization audit script to surface unregistered custom layers early.
CI workflow (minimal deps) exercising persistence and custom layer round‑trip.
Connectivity Semantics Fix

Reinterprets p_lateral_connection as an intensity λ (expected multiplicity) rather than a Bernoulli probability.
Supports hyperdense regimes (λ > 1) discovered by prior TPE / Bayesian optimization; removes accidental clamping that collapsed search space.
Adds floor(λ * decay) + fractional Bernoulli sampling logic for duplicate lateral edges (stable, lightweight alternative to Poisson draws).
Updated docstrings and terminology to prevent future regressions.
Testing & Validation

Round‑trip persistence test: prediction parity + resume training parity within tolerance.
Custom layer safe_mode serialization test (no custom_objects needed).
Duplicate edge fidelity verified in exporter (repeated inbound edges retained in JSON spec).
New statistical expectation tests: empirical mean multiplicity ≈ λ * decay within tight tolerance.
Gating / max consecutive lateral connection behavior preserved (legacy invariant).
Added intensity sampling unit tests (fractional part, negative intensity → zero multiplicity).
Risk & Mitigation

Wide try/except in exporter isolated; future validate_spec() planned.
No shape/dtype metadata yet—intentional to keep initial spec minimal/stable.
Intensity semantics documented to prevent reintroduction of probability clamping.
Why It Matters

Deterministic, restartable training flows (optimizer state continuity).
Faithful structural export unlocks cross‑framework rebuilds and visualization.
Preserves hyperdense architectural expressivity critical to observed performance gains.
Establishes a clean, versionable contract for future tooling (spec validators, Torch builder).
Follow‑Up (Not in this PR)

Add shapes/dtypes to spec metadata.
Negative custom layer test (unregistered should fail clearly).
Idempotent save→load→save validation.
Multi‑IO + regularization layer export tests.
Optional Poisson-mode sampler (if ever needed) behind a flag.

[Copilot]

Pull Request Overview

This PR introduces comprehensive Keras 3 persistence functionality, graph export capabilities, and fixes hyperdense connectivity semantics in the Cerebros neural architecture search framework. The changes enable deterministic training workflows with complete model state preservation and faithful cross-framework model reconstruction.

• Native Keras 3 persistence with round-trip validation for architecture, weights, and optimizer state
• Graph export functionality that preserves duplicate edges for accurate model reconstruction
• Reinterpretation of p_lateral_connection as intensity λ rather than probability to support hyperdense regimes

Reviewed Changes

Copilot reviewed 25 out of 32 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
cerebros/persistence/keras_io.py	Core persistence functions for saving/loading models with metadata
cerebros/keras_export.py	Graph export utilities for cross-framework model reconstruction
cerebros/denseautomlstructuralcomponent/dense_automl_structural_component.py	Connectivity semantics fix for hyperdense lateral connections
tests/	Comprehensive test suite for persistence, export, and connectivity features
cerebros/utils/random.py	Deterministic seeding utilities for reproducible training

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}