Add NGB v4 tuned and expressive nanoGPT baselines

[charlesmartin14]

Purpose

Add a genuinely separate baseline/ngb experiment family for properly tuned small-language-model optimizer baselines. The existing one-head v3 protocol and its checked-in results remain unchanged.

Protocols

Tuned one-head v4

• 1 block, 1 head, width 128, context 256
• 2 corpus-equivalent epochs
• separate result namespace: /tmp/rg-ngb/results/ngb_v4_one_head_2epoch
• SGD: peak 0.05, floor 5e-4, 5% warm-up
• AdamW: peak 3e-4, floor 1e-5, 2.5% warm-up, decay 0.10
• Muon: matrix peak 0.01, floor 2e-4, matrix decay 0.02; auxiliary AdamW peak 3e-4, floor 1e-5, decay 0.10

Distinct small 4x4 v4

• 4 transformer blocks
• 4 attention heads
• width 128, context 256
• 7,253,248 trainable parameters
• 24 WeightWatcher/Muon hidden matrices
• 2 corpus-equivalent epochs
• separate result namespace: /tmp/rg-ngb/results/ngb_v4_small_4x4_2epoch

Implementation

• generalized GPT implementation for arbitrary positive layer/head counts
• shared verified FineWeb-Edu corpus, with protocol-scoped results and fingerprints
• dynamic 6×n_layer WeightWatcher inventory and completion validation
• conda-native direct Python/Papermill workflow; no shell wrappers
• /tmp/rg-ngb runtime default with no home-directory fallback
• restart-safe training and strict completed-run validation

Corrected comparisons

The two comparison notebooks now:

• discover the complete seed intersection shared by all optimizers
• reject unmatched requested seed sets
• report run-level two-sided 95% Student-t intervals
• report matched-seed paired optimizer differences
• report final versus validation-selected test metrics
• report best validation step, final-minus-best drift, clipping frequency, and maximum update-to-weight ratio
• derive perplexity intervals by exponentiating loss-space confidence limits, preventing impossible negative perplexity bounds
• provide full and post-transient task plots
• provide optimizer-level and block-resolved WeightWatcher alpha, ERG-gap, and trap plots

Files

• two explicit v4 YAML protocols
• generalized rg_ngb package
• two training notebooks
• two corrected comparison notebooks
• architecture, optimizer, path, notebook, matched-seed, paired-statistics, and perplexity regression tests
• dedicated NGB job in baseline CI

Safety

No v3 source notebook, output notebook, result directory, optimizer profile, or protocol fingerprint is modified. NGB installs alongside the existing one-head package and uses a separate /tmp/rg-ngb root.