Update documentation to match current implementation

- CHANGELOG: Add fitness_criterion fix, checkpoint/reporter timing
  changes, CTRNN double-buffer fix, aggregation validation fix.
  Remove dead GPU_DESIGN_NOTES.md reference.
- installation.rst: Remove hardcoded version numbers, add optional
  extras section documenting [gpu], [examples], [docs], [all].
- module_summaries.rst: Update Checkpointer to reflect post_evaluate
  timing, save_checkpoint signature, restore_checkpoint behavior.
- reproducibility.rst: Fix checkpoint numbering explanation (suffix N
  is the evaluated generation, not the next-to-evaluate generation).
- config_file.rst: Document time_constant_* parameters (7 params).
- migration.rst: Add v1.x to 2.0 migration section (CTRNN API change).
- innovation_numbers.rst: Fix invalid pseudocode, update checkpoint
  data format, replace dead file reference.
- customization.rst: Remove dead example links (circuits/, openai-lander/).
- faq.rst: Correct GPU availability description.
- genome-interface.rst: Add fitness_criterion parameter to crossover.
- examples/README.md: Add lorenz-ctrnn, signal-tracking-gpu,
  spike-timing-gpu examples.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: CodeReclaimers

CHANGELOG.md

@@ -16,14 +16,6 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - **Configurable node gene distance** via `compatibility_include_node_genes` in `[DefaultGenome]`. Default True (current behavior). Set to False for canonical NEAT distance formula. (NEAT paper compliance)
 - **Configurable enabled-state penalty** via `compatibility_enable_penalty` in `[DefaultGenome]`. Default 1.0 (current behavior). Set to 0.0 for canonical NEAT distance formula. (NEAT paper compliance)
 - **Canonical spawn allocation** via `spawn_method = proportional` in `[DefaultReproduction]`. Default `smoothed` (current behavior). (NEAT paper compliance)
-
-### Changed
-- **Distance function now matches genes by innovation number**, consistent with crossover behavior. Previously used tuple keys (endpoint pairs). This affects speciation when the same connection endpoints receive different innovation numbers in different generations (uncommon but possible). (NEAT paper compliance)
-- **Dangling nodes are now pruned** after `mutate_delete_node` and `mutate_delete_connection`. Hidden nodes that become disconnected from all outputs are automatically removed along with their connections. This reduces structural bloat in long evolution runs.
-
-### Fixed
-- **75% disable rule** now matches the NEAT paper specification (Stanley & Miikkulainen, 2002, p. 111). Previously, the rule was applied after random attribute inheritance, producing an effective ~87.5% disable rate. Now correctly produces 75%. This may affect evolution dynamics in existing configurations.
-
 - **GPU-accelerated evaluation** for CTRNN and Izhikevich spiking networks via optional CuPy dependency
   - `GPUCTRNNEvaluator` and `GPUIZNNEvaluator` in `neat.gpu.evaluator`
   - Batch-evaluates entire populations on GPU using padded tensor operations
@@ -32,9 +24,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   - Requires sum aggregation; other aggregation functions raise `ValueError`
   - `import neat` never loads CuPy; all GPU imports are lazy
   - Benchmark script in `benchmarks/gpu_benchmark.py`
-  - See `GPU_DESIGN_NOTES.md` for design rationale
+  - GPU comparison examples in `examples/signal-tracking-gpu/` and `examples/spike-timing-gpu/`
 
 ### Changed
+- **Distance function now matches genes by innovation number**, consistent with crossover behavior. Previously used tuple keys (endpoint pairs). This affects speciation when the same connection endpoints receive different innovation numbers in different generations (uncommon but possible). (NEAT paper compliance)
+- **Dangling nodes are now pruned** after `mutate_delete_node` and `mutate_delete_connection`. Hidden nodes that become disconnected from all outputs are automatically removed along with their connections. This reduces structural bloat in long evolution runs.
 - **CTRNN integration method** changed from forward Euler to exponential Euler (ETD1)
   - Integrates the linear decay term `-y/tau` exactly
   - Unconditionally stable regardless of `dt/tau` ratio (forward Euler required `dt < 2*tau`)
@@ -43,6 +37,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
     to the same continuous solution as `dt` decreases
   - The `time_constant_min_value` constraint is relaxed: values well below the integration
     timestep are now safe
+- **Checkpoint timing moved from `end_generation` to `post_evaluate`**. Checkpoints now save the evaluated population (with fitness values) rather than the unevaluated post-reproduction population. Restoring a checkpoint no longer re-runs the last generation's fitness evaluation. Checkpoint file `N` now means "generation N has been evaluated." Old 5-tuple checkpoint files are still loadable for backward compatibility.
+- **Reporter species output moved from `end_generation` to `post_evaluate`**. The `StdOutReporter` species detail table now appears alongside fitness statistics for the same generation, eliminating the previous mismatch where species sizes from the next generation were printed under the current generation's banner.
+
+### Fixed
+- **`fitness_criterion = min` now works correctly**. Previously, only the termination check honored this setting. Best-genome tracking, stagnation detection, elite selection, crossover parent selection, spawn allocation, and statistics reporting all hardcoded "higher is better." All fitness comparisons now use direction-aware methods on the `Config` object (`is_better_fitness`, `meets_threshold`, `worst_fitness`).
+- **75% disable rule** now matches the NEAT paper specification (Stanley & Miikkulainen, 2002, p. 111). Previously, the rule was applied after random attribute inheritance, producing an effective ~87.5% disable rate. Now correctly produces 75%. This may affect evolution dynamics in existing configurations.
+- **Two double-buffer bugs in CTRNN advance method** fixed. Incorrect buffer swapping could cause state corruption during multi-step CTRNN evaluation.
+- **Aggregation validation** for builtins and callables fixed.
 
 
 ## [1.1.0] - 2025-12-05
@@ -91,11 +93,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 - **Add-node mutation bias behavior**: Newly inserted nodes created by `mutate_add_node` now start with zero bias so that splitting a connection is as neutral as possible with respect to the original signal flow. This makes the structural mutation less disruptive while preserving the existing weight-preserving semantics (incoming weight 1.0, outgoing weight equal to the original connection).
-- **Checkpoint Generation Semantics**: Clarified and corrected how checkpoint generation numbers are labeled and interpreted.
-  - A checkpoint file named `neat-checkpoint-N` now always contains the population, species state, and RNG state needed to begin evaluating **generation `N`**.
-  - Previously, checkpoints were labeled with the index of the generation that had just been evaluated, while storing the *next* generation's population; this could make restored runs appear to "repeat" the previous generation.
-  - The NEAT evolution loop and genetic algorithm behavior are unchanged; this is a bookkeeping fix that aligns checkpoint behavior with user expectations and the original NEAT paper's generational model.
-  - New and updated tests in `tests/test_checkpoint.py` and `tests/test_population.py` enforce the invariant that checkpoint `N` resumes at the start of generation `N`.
+- **Checkpoint Generation Semantics**: Clarified and corrected how checkpoint generation numbers are labeled and interpreted. (Note: checkpoint timing was further improved in v2.1 to save after evaluation rather than after reproduction, eliminating wasted work on restore.)
 - **Population Size Drift**: Fixed small mismatches between actual population size and configured `pop_size`
   - `DefaultReproduction.reproduce()` now strictly enforces `len(population) == config.pop_size` for every non-extinction generation
   - New `_adjust_spawn_exact` helper adjusts per-species spawn counts after `compute_spawn()` to correct rounding/clamping drift
@@ -131,7 +129,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - New `InnovationTracker` class in `neat/innovation.py`
 - Comprehensive unit tests in `tests/test_innovation.py` (19 tests)
 - Integration tests in `tests/test_innovation_integration.py` (6 tests)
-- Innovation tracking documentation in `INNOVATION_TRACKING_IMPLEMENTATION.md`
+- Innovation tracking documentation in `docs/innovation_numbers.rst`
 
 ### Changed
 - **BREAKING**: `DefaultConnectionGene.__init__()` now requires mandatory `innovation` parameter

docs/config_file.rst

@@ -624,6 +624,37 @@ required for your particular implementation.
     The probability that :term:`mutation` will replace the response multiplier of a node with a newly :py:meth:`chosen <attributes.FloatAttribute.init_value>` 
     random value (as if it were a new node).
 
+.. index:: time_constant
+
+* *time_constant_init_mean*
+    The mean of the normal/gaussian distribution used to select time constant values for new nodes.
+    **Default: 1.0.** Only relevant for CTRNN networks; for feedforward and discrete-time recurrent
+    networks this attribute is unused.
+
+* *time_constant_init_stdev*
+    The standard deviation of the distribution used to select time constant values for new nodes.
+    **Default: 0.0** (all new nodes start with the mean value).
+
+* *time_constant_max_value*
+    The maximum allowed time constant value. **Default: 10.0.**
+
+* *time_constant_min_value*
+    The minimum allowed time constant value. **Default: 0.01.**
+
+* *time_constant_mutate_power*
+    The standard deviation of the zero-centered normal/gaussian distribution from which a time constant
+    mutation value is drawn. **Default: 0.0** (no mutation).
+
+* *time_constant_mutate_rate*
+    The probability that mutation will change the time constant of a node by adding a random value.
+    **Default: 0.0** (no mutation).
+
+* *time_constant_replace_rate*
+    The probability that mutation will replace the time constant of a node with a newly chosen random
+    value. **Default: 0.0** (no replacement).
+
+.. versionadded:: 2.0
+
 .. index:: ! single_structural_mutation
 .. index:: ! structural_mutation_surer
 .. index:: mutation

docs/customization.rst

@@ -76,13 +76,8 @@ New genome types
 To use a different genome type, you can create a custom class whose interface matches that of
 `DefaultGenome` and pass this as the ``genome_type`` argument to the `Config` constructor. The minimum genome type interface is documented here: :ref:`genome-interface-label`.
 
-This is demonstrated in the `circuit evolution
-<https://github.com/CodeReclaimers/neat-python/blob/master/examples/circuits/evolve.py>`_ example.
-
 Alternatively, you can subclass `DefaultGenome` in cases where you need to just add some extra behavior.
-This is done in the `OpenAI lander
-<https://github.com/CodeReclaimers/neat-python/blob/master/examples/openai-lander/evolve.py>`_ example to
-add an evolvable per-genome reward discount value. It is also done in the :py:mod:`iznn` setup, with :py:class:`IZGenome <iznn.IZGenome>`.
+This is done in the :py:mod:`iznn` setup, with :py:class:`IZGenome <iznn.IZGenome>`.
 
 .. index:: species
 

docs/faq.rst

@@ -180,11 +180,13 @@ Smaller populations:
 Can I use GPUs?
 ~~~~~~~~~~~~~~~
 
-**NEAT-Python is CPU-only.** The evolutionary algorithm itself doesn't benefit from GPU acceleration.
+The evolutionary algorithm itself (selection, crossover, speciation) runs on CPU. However, neat-python
+includes optional **GPU-accelerated network evaluation** for CTRNN and Izhikevich spiking networks via
+the ``neat.gpu`` module. Install with ``pip install 'neat-python[gpu]'``. See :doc:`ctrnn` for details.
 
-**However:**
+**Beyond the built-in GPU support:**
 
-- You **can** use GPUs in your fitness function (e.g., if running neural network simulations with PyTorch)
+- You **can** also use GPUs in your own fitness function (e.g., running simulations with PyTorch or JAX)
 - The evolved networks themselves are small and fast on CPU
 - Use ``ParallelEvaluator`` to utilize multiple CPU cores
 

docs/genome-interface.rst

@@ -40,9 +40,11 @@ Initialization/Reproduction
 Crossover/Mutation
 ---------------------------
 
-  :py:meth:`configure_crossover(self, genome1, genome2, config) <DefaultGenome.configure_crossover>`
+  :py:meth:`configure_crossover(self, genome1, genome2, config, fitness_criterion=None) <DefaultGenome.configure_crossover>`
 
-  Configure the genome as a child of the given parent genomes.
+  Configure the genome as a child of the given parent genomes. The optional ``fitness_criterion`` parameter
+  (``'max'``, ``'min'``, or ``None``) determines which parent is considered fitter for the purpose of inheriting
+  disjoint and excess genes. When ``'min'``, lower fitness is better. Defaults to ``'max'`` if not provided.
 
   :py:meth:`mutate(self, config) <DefaultGenome.mutate>`
 

docs/innovation_numbers.rst

@@ -221,9 +221,9 @@ When two genomes mate, genes are matched by innovation number:
                    parent1_genes[innovation],
                    parent2_genes[innovation]
                ]).copy()
-           elif fittest parent has it:
-               # Disjoint/excess - inherit from fittest
-               child_genes[innovation] = fittest_gene.copy()
+           elif innovation in fitter_parent_genes:
+               # Disjoint/excess from fitter parent - inherit
+               child_genes[innovation] = fitter_parent_genes[innovation].copy()
 
 Generation Tracking
 ~~~~~~~~~~~~~~~~~~~
@@ -250,10 +250,11 @@ The innovation tracker is automatically saved with checkpoints:
 
 .. code-block:: python
 
-   # Saving
-   checkpoint_data = (generation, config, population, species_set, rndstate)
-   # innovation_tracker saved as part of population.reproduction
-   
+   # Saving (in post_evaluate callback)
+   checkpoint_data = (generation, config, population, species_set,
+                      rndstate, best_genome)
+   # innovation_tracker is saved as part of config.genome_config
+
    # Loading
    population = neat.Checkpointer.restore_checkpoint('checkpoint-file')
    # innovation_tracker automatically reconnected to genome_config
@@ -263,7 +264,8 @@ The global counter state is preserved, so innovation numbers continue from where
 Implementation Details
 ~~~~~~~~~~~~~~~~~~~~~~
 
-For complete implementation details, see ``INNOVATION_TRACKING_IMPLEMENTATION.md`` in the `source repository <https://github.com/CodeReclaimers/neat-python/blob/master/INNOVATION_TRACKING_IMPLEMENTATION.md>`_.
+For complete implementation details, see the ``InnovationTracker`` class in ``neat/innovation.py``
+in the `source repository <https://github.com/CodeReclaimers/neat-python/blob/master/neat/innovation.py>`_.
 
 References
 ----------

docs/installation.rst

@@ -11,26 +11,22 @@ two methods outlined below:
 
 Install neat-python from PyPI using pip
 ---------------------------------------
-To install the most recent release (version 1.1.0) from PyPI, you should run the command (as root or using `sudo`
+To install the most recent release from PyPI, you should run the command (as root or using `sudo`
 as necessary)::
 
     pip install neat-python
 
 Note that the examples are not included with the package installed from PyPI, so you should download the `source archive
-for release 1.1.0
-<https://github.com/CodeReclaimers/neat-python/releases/tag/v1.0.0>`_ and use the example code contained in it.
-
-You may also just get the 1.1.0 release source, and install it directly (as shown below)
-instead of `pip`.
+<https://github.com/CodeReclaimers/neat-python/releases>`_ and use the example code contained in it.
 
 Install neat-python from source
 --------------------------------
 Obtain the source code by either cloning the source repository::
 
     git clone https://github.com/CodeReclaimers/neat-python.git
 
-or downloading the `source archive for release 1.1.0
-<https://github.com/CodeReclaimers/neat-python/releases/tag/v1.1.0>`_.
+or downloading the latest `source archive
+<https://github.com/CodeReclaimers/neat-python/releases>`_.
 
 Note that the most current code in the repository may not always be in the most polished state, but I do make sure the
 tests pass and that most of the examples run.  If you encounter any problems, please open an `issue on GitHub
@@ -47,3 +43,16 @@ For development (editable install with dev dependencies)::
     pip install -e ".[dev]"
 
 This installs the package in editable mode with testing tools (pytest, coverage, etc.).
+
+Optional extras
+---------------
+
+neat-python supports optional dependency groups that can be installed via pip extras::
+
+    pip install 'neat-python[gpu]'       # GPU acceleration (CuPy, requires NVIDIA GPU)
+    pip install 'neat-python[examples]'  # Dependencies for running examples
+    pip install 'neat-python[docs]'      # Documentation building tools
+    pip install 'neat-python[all]'       # Everything (includes GPU)
+
+The ``[gpu]`` extra installs CuPy for GPU-accelerated CTRNN and Izhikevich network evaluation.
+See :doc:`ctrnn` for details.

docs/migration.rst

@@ -1,7 +1,49 @@
-Migration Guide for neat-python 1.0
-====================================
+Migration Guide
+===============
+
+Migration from 1.x to 2.0
+--------------------------
+
+neat-python 2.0 includes one breaking API change:
+
+CTRNN time constants
+~~~~~~~~~~~~~~~~~~~~
+
+In v1.x, all CTRNN nodes shared a single fixed time constant passed at network creation time::
+
+    # v1.x (no longer works)
+    net = neat.ctrnn.CTRNN.create(genome, config, time_constant=0.01)
+
+In v2.0, each node carries its own time constant as an evolved gene attribute::
+
+    # v2.0
+    net = neat.ctrnn.CTRNN.create(genome, config)
+
+The ``time_constant`` parameter has been removed from ``CTRNN.create()``. Time constants are now
+configured via the ``[DefaultGenome]`` section of your config file using the ``time_constant_*``
+parameters (e.g., ``time_constant_init_mean``, ``time_constant_mutate_rate``). The defaults
+(mean 1.0, zero mutation rate) reproduce the old behavior of a fixed time constant of 1.0.
+
+For details, see `CTRNN-CHANGES.md <https://github.com/CodeReclaimers/neat-python/blob/master/examples/lorenz-ctrnn/docs/CTRNN-CHANGES.md>`_.
+
+Checkpoint format
+~~~~~~~~~~~~~~~~~
+
+Checkpoints created with v1.x are not loadable in v2.0 due to internal class changes (per-node
+time constants change the gene structure). Re-run evolution from scratch or keep the old version
+installed for loading legacy checkpoints.
+
+Other changes
+~~~~~~~~~~~~~
+
+- Feedforward and discrete-time recurrent configurations require no changes.
+- The CTRNN integration method changed from forward Euler to exponential Euler (ETD1), which
+  improves numerical stability but produces slightly different trajectories for the same ``dt``.
+
+Migration from 0.93 to 1.0
+---------------------------
 
-This guide helps you migrate from neat-python 0.93 to 1.0, which includes breaking changes to the parallel evaluation APIs.
+This section helps you migrate from neat-python 0.93 to 1.0, which includes breaking changes to the parallel evaluation APIs.
 
 Overview of Changes
 -------------------