Phase 4: Port paper-faithful online/per-layer path to FSDP with parity tests

[georgepullen]

Purpose

Bring distributed training (starting with FSDP) to paper-faithful parity with single-GPU online/per-layer behavior.

Mandatory Reading (blocking)

First comment must summarize:

• reports/NL_IMPLEMENTATION_ORACLE.md sections 5.3 and 6.1.4
• docs/PAPER_COMPLIANCE.md distributed caveats
• train_fsdp.py
• src/nested_learning/training.py distributed guards and online loop

Required Code Anchors

• train_fsdp.py
• train_dist.py
• src/nested_learning/training.py
• parity tests in tests/

Scope

• Add FSDP support for:
  • online chunk updates
  • per-layer teach signals
• Keep fail-fast behavior explicit for unsupported combinations.
• Add single-GPU vs FSDP parity harness.

Deliverables

• Updated FSDP path.
• Parity test script + report template.

Acceptance Criteria

• 1k-step FSDP faithful run completes.
• Parity drift against single-GPU baseline is within defined tolerance.
• First issue comment contains mandatory reading summary.