Test Configuration:
- 4 experts (one per chip)
- 10 tokens in context
- 4 tokens processed per expert
- Hidden dimension: 16

IMPORTANT: Current implementation tests DATA FLOW only, not actual MoE computation:
- Compute phase is a simple +1.0 operation, not expert network computation
- Focus is on verifying correct token routing and result gathering
- Can be extended to add real expert models later

Core Components:
- Kernels: dispatch (all-to-all), compute (+1.0), combine (all-to-all)
- Orchestration: end2end, dispatch-only, combine-only, dispatch+compute
- Unit Tests: test_dispatch_only, test_combine_only, test_dispatch_compute
- E2E Test: test_end2end with unique value tracing

KEY DESIGN: Use INDEPENDENT scratch_test buffer for combine phase
- Problem: Reusing scratch caused combine to read stale dispatch data
- Solution: Dispatch+Compute use scratch, Combine uses scratch_test
- Prevents corruption when combine's stage-in doesn't fully overwrite
  dispatch's data (writes 4 tokens, stride based on 10 NUM_TOKENS)

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Keep the example focused on the end-to-end dispatch, compute, and combine path
- Remove obsolete debug docs, partial tests, and unused kernel variants
- Align README, test naming, and scratch buffer handling with the current two-chip hardware test