feat: add reproducible matched-backbone control receipts

[DivyamTalwar]

Summary

• add an explicit standard-additive residual control while preserving the geometric default
• run geometric and additive arms from identical initialization and a pre-generated token stream
• emit a self-verifying JSON receipt with configuration, environment, initialization/data digests, losses, gradient norms, active-gradient counts, throughput, and limitations
• fix next-token training with non-contiguous target slices by using reshape

Competitive review

The design adapts reproducibility principles—not code—from:

• PyTorch reproducibility guidance: https://docs.pytorch.org/docs/stable/notes/randomness.html
• nanoGPT's readable baseline/benchmark pattern: https://github.com/karpathy/nanoGPT
• LM Evaluation Harness result/sample artifacts: https://github.com/EleutherAI/lm-evaluation-harness
• LM Evaluation Harness output interface: https://github.com/EleutherAI/lm-evaluation-harness/blob/main/docs/interface.md

This implementation is original and intentionally small enough to audit.

Verification

• python -m compileall -q model tests experiments
• pytest -q — 11 passed
• git diff --check
• five-step CPU smoke receipt, 427,528 serialized parameters per arm:
  • geometric final loss 5.5529; 40 active-gradient parameters; 3,860.9 tokens/s
  • additive final loss 5.5745; 24 active-gradient parameters; 5,771.0 tokens/s

These tiny synthetic-token values verify plumbing only. They are not quality or performance claims.

Research integrity

The receipt states that additive-control geometric-controller parameters are dormant: serialized parameter count is equal, active computation is not. Publishable conclusions still require real data, multiple seeds, matched compute, held-out evaluation, and target-hardware runs.

Please review only; do not merge as part of this research wave.

[DivyamTalwar]

Validation checkpoint

The latest head fdce6f180bbfc5c994723e94e9051d5c10403891 passed authoritative GitHub Actions CI run 32663148824.

Local review evidence:

• 11 tests passed
• compileall and whitespace checks passed
• self-verifying receipt records identical initialization/data digests and equal serialized parameter counts
• five-step CPU smoke explicitly reports 40 active-gradient geometric parameters versus 24 additive parameters

The PR body and receipt deliberately disclose that the controller parameters are dormant in the additive arm. This follows reproducibility principles from PyTorch, nanoGPT, and LM Evaluation Harness without copying their code or presenting synthetic-token measurements as model-quality evidence.

Per the requested workflow, this PR remains open and must not be merged.