This refactor keeps noodle.h as the public user-facing include, while splitting implementation into a small number of maintainable modules.

Public files

File	Purpose
noodle.h	Main public API. User sketches should include this.
noodle_buffer.h	Public grow-only tensor buffer helper API.
noodle_config.h	Compile-time configuration.
noodle_fs.h	Filesystem backend abstraction/configuration.

Implementation files

File	Purpose
noodle_internal.h	Private declarations shared by implementation files. Not intended for user sketches.
noodle_internal.cpp	Private shared globals and helper setup.
noodle_io.cpp	File/storage I/O helpers.
noodle_math.cpp	Reusable math primitives: dot product, max search, BN1D, BN2D, and BN+ReLU helpers.
noodle_conv.cpp	Outer Conv1D, Conv2D, ConvTranspose2D, PROGMEM Conv2D, and NoodleBuffer Conv2D wrappers.
noodle_fcn.cpp	Dense/FCN layers, including file-backed, memory-backed, and PROGMEM-backed parameter paths.
noodle_shape.cpp	Flatten, reshape, global average pooling, and global max pooling.
noodle_dw.cpp	Depthwise convolution and NoodleBuffer depthwise wrappers.
noodle_memory.cpp	Raw buffer helpers, slicing, and global convolution scratch-buffer management.
noodle_buffer.cpp	Grow-only NoodleBuffer allocation helpers.

Existing examples should continue to use:

#include "noodle.h"

Public Noodle API for small CNN/ML inference on microcontrollers.

Do not include noodle_internal.h from sketches. It is for the implementation files only.