Fix: Implement dynamic shape for Flatten layer to replace hardcoded fallback

[jtaeyeon05]

🧠 D: DEFINE — Problem & Opportunity

What is the problem, limitation, or opportunity? Why does this matter for SKaiNET?

Currently, the flatten() layer in NeuralNetworkDslImpl and StageImpl lacks dynamic shape inference. Instead, it relies on a hardcoded fallback (lastDimension = 1568) meant for a specific MNIST test case. This causes severe ArrayIndexOutOfBoundsException errors when users build custom architectures (e.g., a 64-channel CNN), as the framework miscalculates the input size for subsequent dense layers.

---

🔍 A: ASSESS — Feasibility & Impact

Provide an evaluation of the proposal. Address the following:

✔️ Feasibility

Straightforward but requires updating all spatial layers (Conv, Pool) in the DSL to calculate and update their output shapes mathematically before passing them to the next layer.

✔️ Expected Impact

Critical fix: Unlocks the ability to build and run custom CNN architectures without hardcoded shape collisions.

✔️ Risks / Constraints

Need to ensure shape calculation formulas (handling stride, padding, dilation) exactly match the tensor operations.

✔️ Dependencies

Core sk.ainet.lang.nn.dsl module.

---

📚 R: RESEARCH — What Must Be Understood First?

Document research tasks or open questions that must be answered before implementation:

Research Tasks

• Identify the bug location (flatten() TODO block).

---

🛠️ C: CONTRIBUTE — Implementation Plan

Break down actionable steps required to deliver this feature:

Development Tasks

• Refactor DSL builders to track multidimensional shapes (not just a flat lastDimension integer).
• Add shape calculation logic to conv2d, maxPool2d, avgPool2d, etc.
• Remove the hardcoded 1568 fallback in FlattenImpl.
• Write unit tests verifying shape inference for various CNN architectures.

Acceptance Criteria

• Feature works as intended and passes all tests
• Fully documented with clear examples
• Follows SKaiNET coding style & modular design
• Reviewed and approved by maintainers
• No regressions introduced elsewhere

---

💬 Additional Notes

• Bug Lotations: sk.ainet.lang.nn.dsl.NetworkBuilder 1077, 1431

[michalharakal]

@jtaeyeon05, thanks for the clear write-up and for pinpointing the exact lines, this is a really helpful.
You're right that the hardcoded 1568 in flatten() is a real footgun for anyone building a CNN that isn't the MNIST reference. The DSL needs to track multi-dimensional shapes through conv* / pool* layers and let flatten() derive the flattened size from them.

Picking this up now on a stacked branch on top of #537, which just landed the underlying conv shape math in a shared ConvShapeUtils.

The plan is to:

• extend that helper with pool2d / upsample2d arithmetic
• add an input(IntArray) overload so users can declare CNN input shapes, for example input(intArrayOf(1, 28, 28))
• wire currentShape tracking through the spatial DSL builders
• remove the 1568 fallback

The bundled MnistCnn model will be updated to use the new input(IntArray) form so it doesn't silently rely on the magic number.

Will tag you on the PR for review. Appreciate the report.