Add NGB v4 tuned nanoGPT protocols and 4×4 model

[charlesmartin14]

Purpose

Create a separate version-4 nanoGPT experiment family under baseline/ngb without changing or overwriting the existing version-3 one-head results. NGB v4 addresses the one-epoch undertraining and seed-dependent adaptive-optimizer instability observed in the checked-in v3 comparison, and adds a distinct, more expressive small language-model architecture.

New experiment family

• baseline/ngb/configs/v4_one_head.yaml
  • one block, one head, width 128
  • two corpus-equivalent epochs
  • isolated result slug v4_one_head
• baseline/ngb/configs/v4_small_4x4.yaml
  • four blocks, four heads, width 128
  • two corpus-equivalent epochs
  • isolated result slug v4_small_4x4

Both reuse the same pinned, document-disjoint FineWeb-Edu token cache and common train/validation/test/BLEU probes. Validation cross-entropy selects checkpoints; test metrics and WeightWatcher diagnostics remain monitoring-only.

Tuned optimizer centers

One-head v4

• SGD + Nesterov: peak 0.05, floor 5e-4, 5% warm-up, momentum 0.90, weight decay 0.01
• AdamW: peak 3e-4, floor 1e-5, 2.5% warm-up, betas (0.90, 0.95), weight decay 0.10
• Muon matrices: peak 0.01, floor 1e-4, weight decay 0.02
• Muon auxiliary AdamW: peak 2e-4, floor 1e-5, weight decay 0.10

Small 4×4 v4

The adaptive centers are the same; SGD uses peak 0.03 and floor 3e-4 for the deeper model.

These are conservative qualification centers, not declarations of global optimality. baseline/ngb/QUALIFICATION.md defines the stress-seed screen, three-seed qualification, and frozen eight-seed replication protocol.

Runtime and architecture changes

• generalize the shared GPT implementation from exactly one block/head to arbitrary positive layer/head counts with width divisibility checks
• generalize WeightWatcher inventory from six matrices to 6 × n_layer
• keep Muon on all hidden two-dimensional transformer matrices while the tied embedding/head and other parameters use auxiliary AdamW
• isolate v4 roots under /tmp/rg-ngb/results/<run_slug> while retaining the v3 /tmp/rg-nanogpt-one-head contract
• retain conda-native execution; no project venv, shell wrapper, or home-directory default
• include model identity in manifests and completed-run validation so architectures cannot silently reuse each other's artifacts

Comparison notebooks

NGB adds six python3/Papermill notebooks:

1. run v4 one-head
2. compare v4 one-head optimizers
3. run v4 small 4×4
4. compare v4 small 4×4 optimizers
5. compare the two v4 architectures
6. compare v3 versus v4 one-head

The comparison path now supports arbitrary matched completed seed sets rather than hard-coding three seeds. It adds:

• run-level 95% Student-t intervals
• matched-seed paired optimizer contrasts
• validation-selected and final checkpoint tables
• best-validation step and final-minus-best drift
• clipping frequency and maximum update-to-weight ratio
• layer/block-resolved alpha, ERG gap, and trap diagnostics
• perplexity intervals obtained by exponentiating loss-space intervals rather than constructing impossible symmetric intervals directly in perplexity space

Validation

Completed before opening this PR:

• git diff --check
• compilation of shared source and all tests
• validation and Python parsing of all six notebook files
• pip check
• existing one-head suite: 27 passed
• direct NGB v4 suite: 6 passed

The full two-epoch scientific campaigns have not been run in CI; this PR supplies the tested protocols, runtime support, analysis notebooks, and qualification plan. Existing v3 result directories and checked-in output notebook are not modified.