Phase 3: Implement differentiable online-write mode (boundary-state gradients)

[georgepullen]

Purpose

Implement differentiable online-write mode so boundary-state gradients can flow through update steps.

Mandatory Reading (blocking)

First comment must summarize:

• reports/NL_IMPLEMENTATION_ORACLE.md section 6.1.4
• docs/PAPER_COMPLIANCE.md semantic contract notes
• src/nested_learning/training.py online loop
• src/nested_learning/titan/self_modifying.py update sequence
• src/nested_learning/hope/block.py update chunk functions

Required Code Anchors

• src/nested_learning/training.py
• src/nested_learning/hope/block.py
• src/nested_learning/titan/self_modifying.py
• src/nested_learning/fast_state.py

Scope

• Add mode switch: online_updates_mode=stopgrad|differentiable.
• Keep stopgrad mode as default for backward compatibility.
• Add memory-control options for differentiable mode.

Test Requirements

• Explicit test proving later-chunk loss gradients affect earlier update states in differentiable mode.
• Stopgrad mode retains previous behavior.

Deliverables

• New mode implementation + config docs + tests.

Acceptance Criteria

• 2k differentiable smoke run stable.
• Boundary-gradient test passes in differentiable mode.
• First issue comment contains mandatory reading summary.